Fix code style for protobuf-c header file

CMakeLists.txt

@@ -131,8 +131,8 @@ endif()
 
 if(IS_IOS)
 else()
-    list(REMOVE_ITEM PADDLE_MOBILE_H ${CMAKE_CURRENT_SOURCE_DIR}/src/ios_io/PaddleMobile.h)
-    list(REMOVE_ITEM PADDLE_MOBILE_CC ${CMAKE_CURRENT_SOURCE_DIR}/src/ios_io/PaddleMobile.mm)
+    list(REMOVE_ITEM PADDLE_MOBILE_H ${CMAKE_CURRENT_SOURCE_DIR}/src/ios_io/PaddleMobileCPU.h)
+    list(REMOVE_ITEM PADDLE_MOBILE_CC ${CMAKE_CURRENT_SOURCE_DIR}/src/ios_io/PaddleMobileCPU.mm)
     list(REMOVE_ITEM PADDLE_MOBILE_H ${CMAKE_CURRENT_SOURCE_DIR}/src/ios_io/op_symbols.h)
 endif()
 

README.md

@@ -69,7 +69,17 @@ Paddle-Mobile是PaddlePaddle组织下的项目，是一个致力于嵌入式平
 
 - **苹果设备的GPU Metal实现**
 
-    基于Metal实现的苹果设备的GPU预测库，也已经在实现中，近期也会有相应可运行版本。
+|mobilenetfssd|速度|
+|------------|-----|
+|A9(ms)|33.78|
+|A10(ms)|24.05|
+|A11(ms)|17.15|
+|||
+|genet|速度|
+|A9(ms) |3.49|
+|A10(ms)|2.54|
+|A11(ms)|1.43|
+
 
 - **FPGA**
 

metal/paddle-mobile-demo/paddle-mobile-demo.xcodeproj/project.pbxproj

@@ -8,22 +8,29 @@
 
 /* Begin PBXBuildFile section */
 		30D0ED21F392CFA3885B1002 /* Pods_paddle_mobile_demo.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 18896810981724F8A0FED62A /* Pods_paddle_mobile_demo.framework */; };
+		C2E67E5E21524E460013F575 /* LoadPointerViewController.m in Sources */ = {isa = PBXBuildFile; fileRef = C2E67E5D21524E460013F575 /* LoadPointerViewController.m */; };
 		FC013928210204A3008100E3 /* PreProcessKernel.metal in Sources */ = {isa = PBXBuildFile; fileRef = FC013927210204A3008100E3 /* PreProcessKernel.metal */; };
 		FC039B8220E11C550081E9F8 /* AppDelegate.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC039B8120E11C550081E9F8 /* AppDelegate.swift */; };
 		FC039B8420E11C550081E9F8 /* ViewController.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC039B8320E11C550081E9F8 /* ViewController.swift */; };
 		FC039B8720E11C550081E9F8 /* Main.storyboard in Resources */ = {isa = PBXBuildFile; fileRef = FC039B8520E11C550081E9F8 /* Main.storyboard */; };
 		FC039B8920E11C560081E9F8 /* Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = FC039B8820E11C560081E9F8 /* Assets.xcassets */; };
 		FC039B8C20E11C560081E9F8 /* LaunchScreen.storyboard in Resources */ = {isa = PBXBuildFile; fileRef = FC039B8A20E11C560081E9F8 /* LaunchScreen.storyboard */; };
-		FC3602C82108580600FACB58 /* MetalHelper.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC3602C72108580600FACB58 /* MetalHelper.swift */; };
-		FC918191211DBC3500B6F354 /* paddle-mobile.png in Resources */ = {isa = PBXBuildFile; fileRef = FC918190211DBC3500B6F354 /* paddle-mobile.png */; };
-		FC918193211DC70500B6F354 /* iphone.JPG in Resources */ = {isa = PBXBuildFile; fileRef = FC918192211DC70500B6F354 /* iphone.JPG */; };
-		FCD04E6320F3146B0007374F /* params in Resources */ = {isa = PBXBuildFile; fileRef = FCD04E6120F3146A0007374F /* params */; };
-		FCD04E6420F3146B0007374F /* model in Resources */ = {isa = PBXBuildFile; fileRef = FCD04E6220F3146A0007374F /* model */; };
-		FCDFD3FB211D72C3005AB38B /* ModelHelper.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCDFD3FA211D72C3005AB38B /* ModelHelper.swift */; };
-		FCDFD41B211D91C7005AB38B /* synset.txt in Resources */ = {isa = PBXBuildFile; fileRef = FCDFD41A211D91C7005AB38B /* synset.txt */; };
+		FC803BCD214D27930094B8E5 /* FPSCounter.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC803BCB214D27920094B8E5 /* FPSCounter.swift */; };
+		FC803BCE214D27930094B8E5 /* VideoCapture.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC803BCC214D27920094B8E5 /* VideoCapture.swift */; };
+		FCBCCC552122EF5500D94F7E /* MetalHelper.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCBCCC542122EF5400D94F7E /* MetalHelper.swift */; };
 		FCEBEC2C20E1391F00C0B14D /* paddle_mobile.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = FCEBEC2B20E1391F00C0B14D /* paddle_mobile.framework */; };
 		FCEBEC2D20E1391F00C0B14D /* paddle_mobile.framework in Embed Frameworks */ = {isa = PBXBuildFile; fileRef = FCEBEC2B20E1391F00C0B14D /* paddle_mobile.framework */; settings = {ATTRIBUTES = (CodeSignOnCopy, RemoveHeadersOnCopy, ); }; };
-		FCEEE7D4210627A000444BEC /* banana.jpeg in Resources */ = {isa = PBXBuildFile; fileRef = FCEEE7D3210627A000444BEC /* banana.jpeg */; };
+		FCF437E8214B6DDB00943429 /* MultiPredictViewController.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCF437E7214B6DDB00943429 /* MultiPredictViewController.swift */; };
+		FCFE9B692152858600DECA15 /* hand.jpg.zip in Resources */ = {isa = PBXBuildFile; fileRef = FCFE9B632152858600DECA15 /* hand.jpg.zip */; };
+		FCFE9B6A2152858600DECA15 /* synset.txt in Resources */ = {isa = PBXBuildFile; fileRef = FCFE9B642152858600DECA15 /* synset.txt */; };
+		FCFE9B6B2152858600DECA15 /* banana.jpeg in Resources */ = {isa = PBXBuildFile; fileRef = FCFE9B652152858600DECA15 /* banana.jpeg */; };
+		FCFE9B6C2152858600DECA15 /* hand.jpg in Resources */ = {isa = PBXBuildFile; fileRef = FCFE9B662152858600DECA15 /* hand.jpg */; };
+		FCFE9B6D2152858600DECA15 /* iphone.JPG in Resources */ = {isa = PBXBuildFile; fileRef = FCFE9B672152858600DECA15 /* iphone.JPG */; };
+		FCFE9B6E2152858600DECA15 /* paddle-mobile.png in Resources */ = {isa = PBXBuildFile; fileRef = FCFE9B682152858600DECA15 /* paddle-mobile.png */; };
+		FCFE9C512152859600DECA15 /* genet_params in Resources */ = {isa = PBXBuildFile; fileRef = FCFE9B752152859500DECA15 /* genet_params */; };
+		FCFE9C522152859600DECA15 /* genet_model in Resources */ = {isa = PBXBuildFile; fileRef = FCFE9B762152859500DECA15 /* genet_model */; };
+		FCFE9D232152859600DECA15 /* ar_model in Resources */ = {isa = PBXBuildFile; fileRef = FCFE9C4C2152859500DECA15 /* ar_model */; };
+		FCFE9D242152859600DECA15 /* ar_params in Resources */ = {isa = PBXBuildFile; fileRef = FCFE9C4D2152859500DECA15 /* ar_params */; };
 /* End PBXBuildFile section */
 
 /* Begin PBXCopyFilesBuildPhase section */
@@ -44,6 +51,8 @@
 		081C9CF10DB06C58B8B6B039 /* Pods-paddle-mobile-demo.release.xcconfig */ = {isa = PBXFileReference; includeInIndex = 1; lastKnownFileType = text.xcconfig; name = "Pods-paddle-mobile-demo.release.xcconfig"; path = "../Pods/Target Support Files/Pods-paddle-mobile-demo/Pods-paddle-mobile-demo.release.xcconfig"; sourceTree = "<group>"; };
 		18896810981724F8A0FED62A /* Pods_paddle_mobile_demo.framework */ = {isa = PBXFileReference; explicitFileType = wrapper.framework; includeInIndex = 0; path = Pods_paddle_mobile_demo.framework; sourceTree = BUILT_PRODUCTS_DIR; };
 		878829884E1A14D7044721D5 /* Pods-paddle-mobile-demo.debug.xcconfig */ = {isa = PBXFileReference; includeInIndex = 1; lastKnownFileType = text.xcconfig; name = "Pods-paddle-mobile-demo.debug.xcconfig"; path = "../Pods/Target Support Files/Pods-paddle-mobile-demo/Pods-paddle-mobile-demo.debug.xcconfig"; sourceTree = "<group>"; };
+		C2E67E5C21524E460013F575 /* LoadPointerViewController.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = LoadPointerViewController.h; sourceTree = "<group>"; };
+		C2E67E5D21524E460013F575 /* LoadPointerViewController.m */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.objc; path = LoadPointerViewController.m; sourceTree = "<group>"; };
 		FC013927210204A3008100E3 /* PreProcessKernel.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = PreProcessKernel.metal; sourceTree = "<group>"; };
 		FC039B7E20E11C550081E9F8 /* paddle-mobile-demo.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = "paddle-mobile-demo.app"; sourceTree = BUILT_PRODUCTS_DIR; };
 		FC039B8120E11C550081E9F8 /* AppDelegate.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = AppDelegate.swift; sourceTree = "<group>"; };
@@ -52,15 +61,23 @@
 		FC039B8820E11C560081E9F8 /* Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = Assets.xcassets; sourceTree = "<group>"; };
 		FC039B8B20E11C560081E9F8 /* Base */ = {isa = PBXFileReference; lastKnownFileType = file.storyboard; name = Base; path = Base.lproj/LaunchScreen.storyboard; sourceTree = "<group>"; };
 		FC039B8D20E11C560081E9F8 /* Info.plist */ = {isa = PBXFileReference; lastKnownFileType = text.plist.xml; path = Info.plist; sourceTree = "<group>"; };
-		FC3602C72108580600FACB58 /* MetalHelper.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; name = MetalHelper.swift; path = "../../paddle-mobile-unit-test/paddle-mobile-unit-test/MetalHelper.swift"; sourceTree = "<group>"; };
-		FC918190211DBC3500B6F354 /* paddle-mobile.png */ = {isa = PBXFileReference; lastKnownFileType = image.png; path = "paddle-mobile.png"; sourceTree = "<group>"; };
-		FC918192211DC70500B6F354 /* iphone.JPG */ = {isa = PBXFileReference; lastKnownFileType = image.jpeg; path = iphone.JPG; sourceTree = "<group>"; };
-		FCD04E6120F3146A0007374F /* params */ = {isa = PBXFileReference; lastKnownFileType = file; path = params; sourceTree = "<group>"; };
-		FCD04E6220F3146A0007374F /* model */ = {isa = PBXFileReference; lastKnownFileType = file; path = model; sourceTree = "<group>"; };
-		FCDFD3FA211D72C3005AB38B /* ModelHelper.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ModelHelper.swift; sourceTree = "<group>"; };
-		FCDFD41A211D91C7005AB38B /* synset.txt */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = text; path = synset.txt; sourceTree = "<group>"; };
+		FC27991121343A39000B6BAD /* paddle-mobile-demo-Bridging-Header.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = "paddle-mobile-demo-Bridging-Header.h"; sourceTree = "<group>"; };
+		FC4FD97B2140EE250073E130 /* libc++.tbd */ = {isa = PBXFileReference; lastKnownFileType = "sourcecode.text-based-dylib-definition"; name = "libc++.tbd"; path = "usr/lib/libc++.tbd"; sourceTree = SDKROOT; };
+		FC803BCB214D27920094B8E5 /* FPSCounter.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = FPSCounter.swift; sourceTree = "<group>"; };
+		FC803BCC214D27920094B8E5 /* VideoCapture.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = VideoCapture.swift; sourceTree = "<group>"; };
+		FCBCCC542122EF5400D94F7E /* MetalHelper.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = MetalHelper.swift; sourceTree = "<group>"; };
 		FCEBEC2B20E1391F00C0B14D /* paddle_mobile.framework */ = {isa = PBXFileReference; explicitFileType = wrapper.framework; path = paddle_mobile.framework; sourceTree = BUILT_PRODUCTS_DIR; };
-		FCEEE7D3210627A000444BEC /* banana.jpeg */ = {isa = PBXFileReference; lastKnownFileType = image.jpeg; path = banana.jpeg; sourceTree = "<group>"; };
+		FCF437E7214B6DDB00943429 /* MultiPredictViewController.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = MultiPredictViewController.swift; sourceTree = "<group>"; };
+		FCFE9B632152858600DECA15 /* hand.jpg.zip */ = {isa = PBXFileReference; lastKnownFileType = archive.zip; path = hand.jpg.zip; sourceTree = "<group>"; };
+		FCFE9B642152858600DECA15 /* synset.txt */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = text; path = synset.txt; sourceTree = "<group>"; };
+		FCFE9B652152858600DECA15 /* banana.jpeg */ = {isa = PBXFileReference; lastKnownFileType = image.jpeg; path = banana.jpeg; sourceTree = "<group>"; };
+		FCFE9B662152858600DECA15 /* hand.jpg */ = {isa = PBXFileReference; lastKnownFileType = image.jpeg; path = hand.jpg; sourceTree = "<group>"; };
+		FCFE9B672152858600DECA15 /* iphone.JPG */ = {isa = PBXFileReference; lastKnownFileType = image.jpeg; path = iphone.JPG; sourceTree = "<group>"; };
+		FCFE9B682152858600DECA15 /* paddle-mobile.png */ = {isa = PBXFileReference; lastKnownFileType = image.png; path = "paddle-mobile.png"; sourceTree = "<group>"; };
+		FCFE9B752152859500DECA15 /* genet_params */ = {isa = PBXFileReference; lastKnownFileType = file; path = genet_params; sourceTree = "<group>"; };
+		FCFE9B762152859500DECA15 /* genet_model */ = {isa = PBXFileReference; lastKnownFileType = file; path = genet_model; sourceTree = "<group>"; };
+		FCFE9C4C2152859500DECA15 /* ar_model */ = {isa = PBXFileReference; lastKnownFileType = file; path = ar_model; sourceTree = "<group>"; };
+		FCFE9C4D2152859500DECA15 /* ar_params */ = {isa = PBXFileReference; lastKnownFileType = file; path = ar_params; sourceTree = "<group>"; };
 /* End PBXFileReference section */
 
 /* Begin PBXFrameworksBuildPhase section */
@@ -88,6 +105,7 @@
 		7B7DED984E9EE7BFB45E24E8 /* Frameworks */ = {
 			isa = PBXGroup;
 			children = (
+				FC4FD97B2140EE250073E130 /* libc++.tbd */,
 				18896810981724F8A0FED62A /* Pods_paddle_mobile_demo.framework */,
 			);
 			name = Frameworks;
@@ -115,49 +133,82 @@
 		FC039B8020E11C550081E9F8 /* paddle-mobile-demo */ = {
 			isa = PBXGroup;
 			children = (
-				FC0E2C2020EDC03B009C1FAC /* models */,
-				FC0E2C1D20EDC030009C1FAC /* images */,
+				FCFE9B6F2152859500DECA15 /* models */,
+				FCFE9B622152858600DECA15 /* images */,
+				FC803BCA214D27920094B8E5 /* VideoCapture */,
+				FC8CFED2213519540094D569 /* Net */,
 				FC039B8120E11C550081E9F8 /* AppDelegate.swift */,
-				FC013927210204A3008100E3 /* PreProcessKernel.metal */,
 				FC039B8320E11C550081E9F8 /* ViewController.swift */,
 				FC039B8520E11C550081E9F8 /* Main.storyboard */,
 				FC039B8820E11C560081E9F8 /* Assets.xcassets */,
 				FC039B8A20E11C560081E9F8 /* LaunchScreen.storyboard */,
 				FC039B8D20E11C560081E9F8 /* Info.plist */,
-				FC3602C72108580600FACB58 /* MetalHelper.swift */,
-				FCDFD3FA211D72C3005AB38B /* ModelHelper.swift */,
+				FC27991121343A39000B6BAD /* paddle-mobile-demo-Bridging-Header.h */,
+				FCF437E7214B6DDB00943429 /* MultiPredictViewController.swift */,
+				C2E67E5C21524E460013F575 /* LoadPointerViewController.h */,
+				C2E67E5D21524E460013F575 /* LoadPointerViewController.m */,
 			);
 			path = "paddle-mobile-demo";
 			sourceTree = "<group>";
 		};
-		FC0E2C1D20EDC030009C1FAC /* images */ = {
+		FC803BCA214D27920094B8E5 /* VideoCapture */ = {
+			isa = PBXGroup;
+			children = (
+				FC803BCB214D27920094B8E5 /* FPSCounter.swift */,
+				FC803BCC214D27920094B8E5 /* VideoCapture.swift */,
+			);
+			path = VideoCapture;
+			sourceTree = "<group>";
+		};
+		FC8CFED2213519540094D569 /* Net */ = {
+			isa = PBXGroup;
+			children = (
+				FC013927210204A3008100E3 /* PreProcessKernel.metal */,
+				FCBCCC542122EF5400D94F7E /* MetalHelper.swift */,
+			);
+			path = Net;
+			sourceTree = "<group>";
+		};
+		FCFE9B622152858600DECA15 /* images */ = {
 			isa = PBXGroup;
 			children = (
-				FC918192211DC70500B6F354 /* iphone.JPG */,
-				FC918190211DBC3500B6F354 /* paddle-mobile.png */,
-				FCDFD41A211D91C7005AB38B /* synset.txt */,
-				FCEEE7D3210627A000444BEC /* banana.jpeg */,
+				FCFE9B632152858600DECA15 /* hand.jpg.zip */,
+				FCFE9B642152858600DECA15 /* synset.txt */,
+				FCFE9B652152858600DECA15 /* banana.jpeg */,
+				FCFE9B662152858600DECA15 /* hand.jpg */,
+				FCFE9B672152858600DECA15 /* iphone.JPG */,
+				FCFE9B682152858600DECA15 /* paddle-mobile.png */,
 			);
 			name = images;
 			path = ../../images;
 			sourceTree = "<group>";
 		};
-		FC0E2C2020EDC03B009C1FAC /* models */ = {
+		FCFE9B6F2152859500DECA15 /* models */ = {
 			isa = PBXGroup;
 			children = (
-				FCD04E6020F3146A0007374F /* mobilenet */,
+				FCFE9B742152859500DECA15 /* genet */,
+				FCFE9C4B2152859500DECA15 /* fluid_fssd_new_ar */,
 			);
 			name = models;
 			path = ../../models;
 			sourceTree = "<group>";
 		};
-		FCD04E6020F3146A0007374F /* mobilenet */ = {
+		FCFE9B742152859500DECA15 /* genet */ = {
+			isa = PBXGroup;
+			children = (
+				FCFE9B752152859500DECA15 /* genet_params */,
+				FCFE9B762152859500DECA15 /* genet_model */,
+			);
+			path = genet;
+			sourceTree = "<group>";
+		};
+		FCFE9C4B2152859500DECA15 /* fluid_fssd_new_ar */ = {
 			isa = PBXGroup;
 			children = (
-				FCD04E6120F3146A0007374F /* params */,
-				FCD04E6220F3146A0007374F /* model */,
+				FCFE9C4C2152859500DECA15 /* ar_model */,
+				FCFE9C4D2152859500DECA15 /* ar_params */,
 			);
-			path = mobilenet;
+			path = fluid_fssd_new_ar;
 			sourceTree = "<group>";
 		};
 /* End PBXGroup section */
@@ -195,6 +246,7 @@
 				TargetAttributes = {
 					FC039B7D20E11C550081E9F8 = {
 						CreatedOnToolsVersion = 9.3.1;
+						LastSwiftMigration = 0940;
 					};
 				};
 			};
@@ -221,14 +273,18 @@
 			isa = PBXResourcesBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
-				FCD04E6320F3146B0007374F /* params in Resources */,
+				FCFE9D232152859600DECA15 /* ar_model in Resources */,
 				FC039B8C20E11C560081E9F8 /* LaunchScreen.storyboard in Resources */,
-				FC918191211DBC3500B6F354 /* paddle-mobile.png in Resources */,
+				FCFE9C522152859600DECA15 /* genet_model in Resources */,
+				FCFE9D242152859600DECA15 /* ar_params in Resources */,
+				FCFE9B6E2152858600DECA15 /* paddle-mobile.png in Resources */,
+				FCFE9C512152859600DECA15 /* genet_params in Resources */,
+				FCFE9B692152858600DECA15 /* hand.jpg.zip in Resources */,
 				FC039B8920E11C560081E9F8 /* Assets.xcassets in Resources */,
-				FCEEE7D4210627A000444BEC /* banana.jpeg in Resources */,
-				FC918193211DC70500B6F354 /* iphone.JPG in Resources */,
-				FCDFD41B211D91C7005AB38B /* synset.txt in Resources */,
-				FCD04E6420F3146B0007374F /* model in Resources */,
+				FCFE9B6A2152858600DECA15 /* synset.txt in Resources */,
+				FCFE9B6B2152858600DECA15 /* banana.jpeg in Resources */,
+				FCFE9B6D2152858600DECA15 /* iphone.JPG in Resources */,
+				FCFE9B6C2152858600DECA15 /* hand.jpg in Resources */,
 				FC039B8720E11C550081E9F8 /* Main.storyboard in Resources */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
@@ -280,10 +336,13 @@
 			buildActionMask = 2147483647;
 			files = (
 				FC039B8420E11C550081E9F8 /* ViewController.swift in Sources */,
-				FCDFD3FB211D72C3005AB38B /* ModelHelper.swift in Sources */,
+				FC803BCE214D27930094B8E5 /* VideoCapture.swift in Sources */,
 				FC013928210204A3008100E3 /* PreProcessKernel.metal in Sources */,
+				FCF437E8214B6DDB00943429 /* MultiPredictViewController.swift in Sources */,
+				FCBCCC552122EF5500D94F7E /* MetalHelper.swift in Sources */,
+				FC803BCD214D27930094B8E5 /* FPSCounter.swift in Sources */,
+				C2E67E5E21524E460013F575 /* LoadPointerViewController.m in Sources */,
 				FC039B8220E11C550081E9F8 /* AppDelegate.swift in Sources */,
-				FC3602C82108580600FACB58 /* MetalHelper.swift in Sources */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
 		};
@@ -428,19 +487,23 @@
 			baseConfigurationReference = 878829884E1A14D7044721D5 /* Pods-paddle-mobile-demo.debug.xcconfig */;
 			buildSettings = {
 				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
+				CLANG_ENABLE_MODULES = YES;
 				CODE_SIGN_IDENTITY = "iPhone Developer";
 				CODE_SIGN_STYLE = Automatic;
 				DEVELOPMENT_TEAM = A798K58VVL;
+				ENABLE_BITCODE = NO;
 				INFOPLIST_FILE = "paddle-mobile-demo/Info.plist";
 				IPHONEOS_DEPLOYMENT_TARGET = 9.0;
 				LD_RUNPATH_SEARCH_PATHS = (
 					"$(inherited)",
 					"@executable_path/Frameworks",
 				);
-				PRODUCT_BUNDLE_IDENTIFIER = com.paddlemobile.metal;
+				PRODUCT_BUNDLE_IDENTIFIER = "com.baidu.paddle-mobile";
 				PRODUCT_NAME = "$(TARGET_NAME)";
 				PROVISIONING_PROFILE = "";
 				PROVISIONING_PROFILE_SPECIFIER = "";
+				SWIFT_OBJC_BRIDGING_HEADER = "paddle-mobile-demo/paddle-mobile-demo-Bridging-Header.h";
+				SWIFT_OPTIMIZATION_LEVEL = "-Onone";
 				SWIFT_VERSION = 4.0;
 				TARGETED_DEVICE_FAMILY = "1,2";
 			};
@@ -451,19 +514,22 @@
 			baseConfigurationReference = 081C9CF10DB06C58B8B6B039 /* Pods-paddle-mobile-demo.release.xcconfig */;
 			buildSettings = {
 				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
+				CLANG_ENABLE_MODULES = YES;
 				CODE_SIGN_IDENTITY = "iPhone Developer";
 				CODE_SIGN_STYLE = Automatic;
 				DEVELOPMENT_TEAM = A798K58VVL;
+				ENABLE_BITCODE = NO;
 				INFOPLIST_FILE = "paddle-mobile-demo/Info.plist";
 				IPHONEOS_DEPLOYMENT_TARGET = 9.0;
 				LD_RUNPATH_SEARCH_PATHS = (
 					"$(inherited)",
 					"@executable_path/Frameworks",
 				);
-				PRODUCT_BUNDLE_IDENTIFIER = com.paddlemobile.metal;
+				PRODUCT_BUNDLE_IDENTIFIER = "com.baidu.paddle-mobile";
 				PRODUCT_NAME = "$(TARGET_NAME)";
 				PROVISIONING_PROFILE = "";
 				PROVISIONING_PROFILE_SPECIFIER = "";
+				SWIFT_OBJC_BRIDGING_HEADER = "paddle-mobile-demo/paddle-mobile-demo-Bridging-Header.h";
 				SWIFT_VERSION = 4.0;
 				TARGETED_DEVICE_FAMILY = "1,2";
 			};

metal/paddle-mobile-demo/paddle-mobile-demo/AppDelegate.swift

@@ -19,7 +19,6 @@ class AppDelegate: UIResponder, UIApplicationDelegate {
 
     var window: UIWindow?
 
-
     func application(_ application: UIApplication, didFinishLaunchingWithOptions launchOptions: [UIApplicationLaunchOptionsKey: Any]?) -> Bool {
         // Override point for customization after application launch.
         return true

metal/paddle-mobile-demo/paddle-mobile-demo/Base.lproj/Main.storyboard

@@ -11,6 +11,34 @@
         <capability name="documents saved in the Xcode 8 format" minToolsVersion="8.0"/>
     </dependencies>
     <scenes>
+        <!--Multi Predict View Controller-->
+        <scene sceneID="ec4-AW-9Vs">
+            <objects>
+                <viewController id="Vwd-lt-764" customClass="MultiPredictViewController" customModule="paddle_mobile_demo" customModuleProvider="target" sceneMemberID="viewController">
+                    <view key="view" contentMode="scaleToFill" id="55D-rz-Ex6">
+                        <rect key="frame" x="0.0" y="0.0" width="375" height="667"/>
+                        <autoresizingMask key="autoresizingMask" widthSizable="YES" heightSizable="YES"/>
+                        <subviews>
+                            <button opaque="NO" contentMode="scaleToFill" contentHorizontalAlignment="center" contentVerticalAlignment="center" buttonType="roundedRect" lineBreakMode="middleTruncation" translatesAutoresizingMaskIntoConstraints="NO" id="TQt-X9-PdF">
+                                <rect key="frame" x="164" y="318" width="46" height="30"/>
+                                <state key="normal" title="Button"/>
+                                <connections>
+                                    <action selector="predictAct:" destination="Vwd-lt-764" eventType="touchUpInside" id="d4z-Cv-6jY"/>
+                                </connections>
+                            </button>
+                        </subviews>
+                        <color key="backgroundColor" white="1" alpha="1" colorSpace="custom" customColorSpace="genericGamma22GrayColorSpace"/>
+                        <constraints>
+                            <constraint firstItem="TQt-X9-PdF" firstAttribute="centerY" secondItem="55D-rz-Ex6" secondAttribute="centerY" id="bL3-wr-TcH"/>
+                            <constraint firstItem="TQt-X9-PdF" firstAttribute="centerX" secondItem="55D-rz-Ex6" secondAttribute="centerX" id="sBi-RQ-sJn"/>
+                        </constraints>
+                        <viewLayoutGuide key="safeArea" id="bsd-h4-RYZ"/>
+                    </view>
+                </viewController>
+                <placeholder placeholderIdentifier="IBFirstResponder" id="68E-SG-96s" userLabel="First Responder" sceneMemberID="firstResponder"/>
+            </objects>
+            <point key="canvasLocation" x="-559" y="686"/>
+        </scene>
         <!--View Controller-->
         <scene sceneID="tne-QT-ifu">
             <objects>
@@ -20,12 +48,11 @@
                         <autoresizingMask key="autoresizingMask" widthSizable="YES" heightSizable="YES"/>
                         <subviews>
                             <imageView userInteractionEnabled="NO" contentMode="scaleAspectFit" horizontalHuggingPriority="251" verticalHuggingPriority="251" translatesAutoresizingMaskIntoConstraints="NO" id="ZZh-fw-LwK">
-                                <rect key="frame" x="0.0" y="20" width="375" height="247"/>
+                                <rect key="frame" x="0.0" y="20" width="225" height="247"/>
                             </imageView>
-                            <label opaque="NO" userInteractionEnabled="NO" contentMode="left" horizontalHuggingPriority="251" verticalHuggingPriority="251" text="Thread:" textAlignment="natural" lineBreakMode="tailTruncation" baselineAdjustment="alignBaselines" adjustsFontSizeToFit="NO" translatesAutoresizingMaskIntoConstraints="NO" id="2EB-m2-a3L">
-                                <rect key="frame" x="10" y="538" width="68" height="24"/>
+                            <label opaque="NO" userInteractionEnabled="NO" contentMode="left" horizontalHuggingPriority="251" verticalHuggingPriority="251" horizontalCompressionResistancePriority="749" text="Platform:" textAlignment="natural" lineBreakMode="tailTruncation" baselineAdjustment="alignBaselines" adjustsFontSizeToFit="NO" translatesAutoresizingMaskIntoConstraints="NO" id="2EB-m2-a3L">
+                                <rect key="frame" x="10" y="538" width="35" height="24"/>
                                 <constraints>
-                                    <constraint firstAttribute="width" constant="68" id="Q5J-tq-JSX"/>
                                     <constraint firstAttribute="height" constant="24" id="SYv-As-Si8"/>
                                 </constraints>
                                 <fontDescription key="fontDescription" type="system" pointSize="20"/>
@@ -33,12 +60,12 @@
                                 <nil key="highlightedColor"/>
                             </label>
                             <pickerView contentMode="scaleToFill" translatesAutoresizingMaskIntoConstraints="NO" id="DlO-dk-RMr">
-                                <rect key="frame" x="88" y="510.5" width="287" height="80"/>
+                                <rect key="frame" x="55" y="510.5" width="320" height="80"/>
                                 <constraints>
                                     <constraint firstAttribute="height" constant="80" id="Sbi-05-Mwd"/>
                                 </constraints>
                             </pickerView>
-                            <pickerView contentMode="scaleToFill" translatesAutoresizingMaskIntoConstraints="NO" id="6MG-gv-hD5">
+                            <pickerView contentMode="scaleToFill" horizontalCompressionResistancePriority="749" translatesAutoresizingMaskIntoConstraints="NO" id="6MG-gv-hD5">
                                 <rect key="frame" x="85" y="401" width="290" height="80"/>
                                 <constraints>
                                     <constraint firstAttribute="height" constant="80" id="yAL-JY-G6b"/>
@@ -47,7 +74,6 @@
                             <label opaque="NO" userInteractionEnabled="NO" contentMode="left" horizontalHuggingPriority="251" verticalHuggingPriority="251" text="Models" textAlignment="natural" lineBreakMode="tailTruncation" baselineAdjustment="alignBaselines" adjustsFontSizeToFit="NO" translatesAutoresizingMaskIntoConstraints="NO" id="avL-VK-Kha">
                                 <rect key="frame" x="10" y="429" width="65" height="24"/>
                                 <constraints>
-                                    <constraint firstAttribute="width" constant="65" id="6oA-g2-Xq4"/>
                                     <constraint firstAttribute="height" constant="24" id="EwE-B3-z2R"/>
                                 </constraints>
                                 <fontDescription key="fontDescription" type="system" pointSize="20"/>
@@ -142,9 +168,14 @@
                                 <fontDescription key="fontDescription" type="system" pointSize="15"/>
                                 <textInputTraits key="textInputTraits" autocapitalizationType="sentences"/>
                             </textView>
+                            <view contentMode="scaleToFill" translatesAutoresizingMaskIntoConstraints="NO" id="Cil-py-NiA">
+                                <rect key="frame" x="225" y="20" width="150" height="247"/>
+                                <color key="backgroundColor" white="1" alpha="1" colorSpace="custom" customColorSpace="genericGamma22GrayColorSpace"/>
+                            </view>
                         </subviews>
                         <color key="backgroundColor" red="1" green="1" blue="1" alpha="1" colorSpace="custom" customColorSpace="sRGB"/>
                         <constraints>
+                            <constraint firstItem="m5L-O7-P31" firstAttribute="top" secondItem="Cil-py-NiA" secondAttribute="bottom" constant="10" id="16p-IK-b5X"/>
                             <constraint firstItem="6Tk-OE-BBY" firstAttribute="trailing" secondItem="VQn-bS-fWp" secondAttribute="trailing" constant="10" id="1Xg-0h-9SE"/>
                             <constraint firstItem="avL-VK-Kha" firstAttribute="leading" secondItem="6Tk-OE-BBY" secondAttribute="leading" constant="10" id="2t9-hS-VXa"/>
                             <constraint firstItem="R90-Yf-S6g" firstAttribute="centerY" secondItem="wUL-9N-u1V" secondAttribute="centerY" id="76b-Ny-1Og"/>
@@ -159,11 +190,12 @@
                             <constraint firstItem="XpL-9M-UOp" firstAttribute="centerY" secondItem="wUL-9N-u1V" secondAttribute="centerY" id="KWW-qT-Rzf"/>
                             <constraint firstItem="6MG-gv-hD5" firstAttribute="centerY" secondItem="avL-VK-Kha" secondAttribute="centerY" id="KZa-YZ-DEs"/>
                             <constraint firstItem="2EB-m2-a3L" firstAttribute="leading" secondItem="6Tk-OE-BBY" secondAttribute="leading" constant="10" id="Le3-TN-zOL"/>
-                            <constraint firstItem="ZZh-fw-LwK" firstAttribute="trailing" secondItem="6Tk-OE-BBY" secondAttribute="trailing" id="MeS-HQ-voE"/>
+                            <constraint firstItem="ZZh-fw-LwK" firstAttribute="trailing" secondItem="6Tk-OE-BBY" secondAttribute="trailing" constant="-150" id="MeS-HQ-voE"/>
                             <constraint firstItem="m5L-O7-P31" firstAttribute="top" secondItem="ZZh-fw-LwK" secondAttribute="bottom" constant="10" id="NUL-Ta-VI8"/>
                             <constraint firstItem="m5L-O7-P31" firstAttribute="leading" secondItem="6Tk-OE-BBY" secondAttribute="leading" constant="15" id="RFA-z1-9aB"/>
                             <constraint firstItem="wUL-9N-u1V" firstAttribute="width" secondItem="a3K-ri-NVs" secondAttribute="width" id="Rp6-Bh-BN3"/>
                             <constraint firstItem="6MG-gv-hD5" firstAttribute="trailing" secondItem="6Tk-OE-BBY" secondAttribute="trailing" id="S0W-0G-75m"/>
+                            <constraint firstItem="Cil-py-NiA" firstAttribute="top" secondItem="6Tk-OE-BBY" secondAttribute="top" id="UNc-Et-9Yv"/>
                             <constraint firstItem="w7H-Sk-Rai" firstAttribute="leading" secondItem="wUL-9N-u1V" secondAttribute="trailing" id="VBM-8b-jP0"/>
                             <constraint firstItem="VQn-bS-fWp" firstAttribute="top" secondItem="m5L-O7-P31" secondAttribute="bottom" constant="8" id="VpS-4N-mOo"/>
                             <constraint firstItem="wUL-9N-u1V" firstAttribute="top" secondItem="2EB-m2-a3L" secondAttribute="bottom" constant="35" id="VpU-j2-gaE"/>
@@ -175,10 +207,12 @@
                             <constraint firstItem="ZZh-fw-LwK" firstAttribute="top" secondItem="6Tk-OE-BBY" secondAttribute="top" id="eIC-fZ-OEE"/>
                             <constraint firstItem="976-fk-Kx2" firstAttribute="centerY" secondItem="wUL-9N-u1V" secondAttribute="centerY" id="fFg-pB-eyU"/>
                             <constraint firstItem="6Tk-OE-BBY" firstAttribute="bottom" secondItem="wUL-9N-u1V" secondAttribute="bottom" constant="40" id="fG6-0p-I0P"/>
+                            <constraint firstItem="Cil-py-NiA" firstAttribute="trailing" secondItem="6Tk-OE-BBY" secondAttribute="trailing" id="gGK-DB-ibv"/>
                             <constraint firstItem="XpL-9M-UOp" firstAttribute="leading" secondItem="w7H-Sk-Rai" secondAttribute="trailing" id="guC-Db-cA9"/>
                             <constraint firstItem="6MG-gv-hD5" firstAttribute="leading" secondItem="avL-VK-Kha" secondAttribute="trailing" constant="10" id="jNW-iC-u7V"/>
                             <constraint firstItem="4ey-Xr-U4e" firstAttribute="bottom" secondItem="6Tk-OE-BBY" secondAttribute="bottom" id="o1X-q5-P7j"/>
                             <constraint firstItem="6MG-gv-hD5" firstAttribute="top" secondItem="VQn-bS-fWp" secondAttribute="bottom" constant="8" id="tAE-ss-jlA"/>
+                            <constraint firstItem="Cil-py-NiA" firstAttribute="leading" secondItem="ZZh-fw-LwK" secondAttribute="trailing" id="teJ-PP-h2R"/>
                             <constraint firstItem="4ey-Xr-U4e" firstAttribute="top" secondItem="wUL-9N-u1V" secondAttribute="bottom" constant="10" id="udc-wT-jqd"/>
                             <constraint firstItem="ZZh-fw-LwK" firstAttribute="leading" secondItem="6Tk-OE-BBY" secondAttribute="leading" id="vXI-l2-CjL"/>
                             <constraint firstItem="VQn-bS-fWp" firstAttribute="leading" secondItem="6Tk-OE-BBY" secondAttribute="leading" constant="10" id="wtI-Dl-YPq"/>
@@ -195,11 +229,81 @@
                         <outlet property="resultTextView" destination="VQn-bS-fWp" id="306-c7-3vM"/>
                         <outlet property="selectImageView" destination="ZZh-fw-LwK" id="afR-Bv-6AW"/>
                         <outlet property="threadPickerView" destination="DlO-dk-RMr" id="Kk4-QV-b5o"/>
+                        <outlet property="videoView" destination="Cil-py-NiA" id="QY2-BP-SNS"/>
                     </connections>
                 </viewController>
                 <placeholder placeholderIdentifier="IBFirstResponder" id="dkx-z0-nzr" sceneMemberID="firstResponder"/>
             </objects>
-            <point key="canvasLocation" x="-724" y="98.50074962518741"/>
+            <point key="canvasLocation" x="-1449" y="-3"/>
+        </scene>
+        <!--Load Pointer View Controller-->
+        <scene sceneID="56v-9i-I4d">
+            <objects>
+                <viewController id="4MS-jc-i6A" customClass="LoadPointerViewController" sceneMemberID="viewController">
+                    <view key="view" contentMode="scaleToFill" id="VbZ-nk-rJR">
+                        <rect key="frame" x="0.0" y="0.0" width="375" height="667"/>
+                        <autoresizingMask key="autoresizingMask" widthSizable="YES" heightSizable="YES"/>
+                        <subviews>
+                            <imageView userInteractionEnabled="NO" contentMode="scaleToFill" horizontalHuggingPriority="251" verticalHuggingPriority="251" fixedFrame="YES" translatesAutoresizingMaskIntoConstraints="NO" id="2p5-S3-M4T">
+                                <rect key="frame" x="16" y="63" width="240" height="128"/>
+                                <autoresizingMask key="autoresizingMask" flexibleMaxX="YES" flexibleMaxY="YES"/>
+                            </imageView>
+                            <button opaque="NO" contentMode="scaleToFill" contentHorizontalAlignment="center" contentVerticalAlignment="center" buttonType="roundedRect" lineBreakMode="middleTruncation" translatesAutoresizingMaskIntoConstraints="NO" id="37q-nm-0H7">
+                                <rect key="frame" x="38" y="610" width="42" height="30"/>
+                                <constraints>
+                                    <constraint firstAttribute="height" constant="30" id="ofW-G3-KST"/>
+                                    <constraint firstAttribute="width" constant="42" id="pwd-tO-zcJ"/>
+                                </constraints>
+                                <state key="normal" title="Image"/>
+                            </button>
+                            <button opaque="NO" contentMode="scaleToFill" contentHorizontalAlignment="center" contentVerticalAlignment="center" buttonType="roundedRect" lineBreakMode="middleTruncation" translatesAutoresizingMaskIntoConstraints="NO" id="fAg-ai-yaA">
+                                <rect key="frame" x="119" y="610" width="34" height="30"/>
+                                <constraints>
+                                    <constraint firstAttribute="height" constant="30" id="IES-jf-Z1n"/>
+                                    <constraint firstAttribute="width" constant="34" id="jxK-Xn-WCE"/>
+                                </constraints>
+                                <state key="normal" title="Load"/>
+                                <connections>
+                                    <action selector="loaderButtonPressed:" destination="4MS-jc-i6A" eventType="touchUpInside" id="3cy-PD-aiE"/>
+                                </connections>
+                            </button>
+                            <button opaque="NO" contentMode="scaleToFill" contentHorizontalAlignment="center" contentVerticalAlignment="center" buttonType="roundedRect" lineBreakMode="middleTruncation" translatesAutoresizingMaskIntoConstraints="NO" id="pdS-6e-Pd1">
+                                <rect key="frame" x="185" y="610" width="49" height="30"/>
+                                <constraints>
+                                    <constraint firstAttribute="width" constant="49" id="ddY-uM-fzA"/>
+                                    <constraint firstAttribute="height" constant="30" id="yKd-YL-UML"/>
+                                </constraints>
+                                <state key="normal" title="Predict"/>
+                                <connections>
+                                    <action selector="predictButtonPressed:" destination="4MS-jc-i6A" eventType="touchUpInside" id="sOH-iT-s1w"/>
+                                </connections>
+                            </button>
+                            <button opaque="NO" contentMode="scaleToFill" contentHorizontalAlignment="center" contentVerticalAlignment="center" buttonType="roundedRect" lineBreakMode="middleTruncation" translatesAutoresizingMaskIntoConstraints="NO" id="DZa-sd-lY7">
+                                <rect key="frame" x="279" y="610" width="34" height="30"/>
+                                <constraints>
+                                    <constraint firstAttribute="width" constant="34" id="aSO-4q-PgA"/>
+                                    <constraint firstAttribute="height" constant="30" id="eAt-Uc-BxX"/>
+                                </constraints>
+                                <state key="normal" title="clear"/>
+                            </button>
+                        </subviews>
+                        <color key="backgroundColor" white="1" alpha="1" colorSpace="custom" customColorSpace="genericGamma22GrayColorSpace"/>
+                        <constraints>
+                            <constraint firstItem="vsb-FH-h7h" firstAttribute="bottom" secondItem="37q-nm-0H7" secondAttribute="bottom" constant="27" id="4Wf-Uh-gvr"/>
+                            <constraint firstItem="DZa-sd-lY7" firstAttribute="leading" secondItem="pdS-6e-Pd1" secondAttribute="trailing" constant="45" id="8dB-uI-cs9"/>
+                            <constraint firstItem="fAg-ai-yaA" firstAttribute="leading" secondItem="37q-nm-0H7" secondAttribute="trailing" constant="39" id="EAV-Oq-jeD"/>
+                            <constraint firstItem="vsb-FH-h7h" firstAttribute="bottom" secondItem="fAg-ai-yaA" secondAttribute="bottom" constant="27" id="Px0-A9-Eql"/>
+                            <constraint firstItem="pdS-6e-Pd1" firstAttribute="leading" secondItem="fAg-ai-yaA" secondAttribute="trailing" constant="32" id="ZUR-Nv-aNb"/>
+                            <constraint firstItem="vsb-FH-h7h" firstAttribute="bottom" secondItem="pdS-6e-Pd1" secondAttribute="bottom" constant="27" id="kPx-mt-ab9"/>
+                            <constraint firstItem="37q-nm-0H7" firstAttribute="leading" secondItem="vsb-FH-h7h" secondAttribute="leading" constant="38" id="trH-Fq-sSv"/>
+                            <constraint firstItem="vsb-FH-h7h" firstAttribute="bottom" secondItem="DZa-sd-lY7" secondAttribute="bottom" constant="27" id="yNJ-hq-2Qg"/>
+                        </constraints>
+                        <viewLayoutGuide key="safeArea" id="vsb-FH-h7h"/>
+                    </view>
+                </viewController>
+                <placeholder placeholderIdentifier="IBFirstResponder" id="hGb-Pb-icS" userLabel="First Responder" sceneMemberID="firstResponder"/>
+            </objects>
+            <point key="canvasLocation" x="-721" y="-427"/>
         </scene>
     </scenes>
     <resources>

metal/paddle-mobile-demo/paddle-mobile-demo/LoadPointerViewController.h

+//
+//  LoadPointerViewController.h
+//  paddle-mobile-demo
+//
+//  Created by Xiao,Haichun on 2018/9/19.
+//  Copyright © 2018年 orange. All rights reserved.
+//
+
+#import <UIKit/UIKit.h>
+
+@interface LoadPointerViewController : UIViewController
+
+@end

metal/paddle-mobile-demo/paddle-mobile-demo/LoadPointerViewController.m

+//
+//  LoadPointerViewController.m
+//  paddle-mobile-demo
+//
+//  Created by Xiao,Haichun on 2018/9/19.
+//  Copyright © 2018年 orange. All rights reserved.
+//
+
+#import "LoadPointerViewController.h"
+#import <Metal/Metal.h>
+#import "paddle-mobile-demo-Bridging-Header.h"
+
+@interface LoadPointerViewController ()
+
+@property (strong, nonatomic) id<MTLDevice> device;
+@property (strong, nonatomic) id<MTLTexture> texture;
+@property (strong, nonatomic) id<MTLCommandQueue> queue;
+@property (strong, nonatomic) PaddleMobileGPU *runner;
+@property (strong, nonatomic) ModelConfig *modelConfig;
+
+@end
+
+@implementation LoadPointerViewController
+
+- (void)viewDidLoad {
+    [super viewDidLoad];
+  
+  
+  self.device = MTLCreateSystemDefaultDevice();
+  
+  self.queue = [self.device newCommandQueue];
+  
+    // Do any additional setup after loading the view.
+//  NSString *modelPath = [[NSBundle mainBundle] URLForResource:@"genet_model" withExtension:nil].path;
+//  NSString *paramPath = [[NSBundle mainBundle] URLForResource:@"genet_params" withExtension:nil].path;
+  
+  NSString *modelPath = [[NSBundle mainBundle] URLForResource:@"ar_model" withExtension:nil].path;
+  NSString *paramPath = [[NSBundle mainBundle] URLForResource:@"ar_params" withExtension:nil].path;
+
+  long fileSize;
+  FILE *fp;
+  fp = fopen([modelPath UTF8String], "rb");
+  fseek(fp, 0, SEEK_END);
+  fileSize = ftell(fp);
+  rewind(fp);
+  void *buffer = malloc(fileSize);
+  fread(buffer, 1, fileSize, fp);
+  fclose(fp);
+  
+  long paramfileSize;
+  FILE *parmaFilePointer;
+  parmaFilePointer = fopen([paramPath UTF8String], "rb");
+  fseek(parmaFilePointer, 0, SEEK_END);
+  paramfileSize = ftell(parmaFilePointer);
+  rewind(parmaFilePointer);
+  void *parmaBuffer = malloc(paramfileSize);
+  fread(parmaBuffer, 1, paramfileSize, parmaFilePointer);
+  fclose(parmaFilePointer);
+  
+  _modelConfig = [[ModelConfig alloc] init];
+//  _modelConfig.means = @[[NSNumber numberWithFloat:128.0], [NSNumber numberWithFloat:128.0], [NSNumber numberWithFloat:128.0]];
+//  _modelConfig.scale = 0.017;
+//  _modelConfig.dims = @[[NSNumber numberWithFloat:1], [NSNumber numberWithFloat:128.], [NSNumber numberWithFloat:128.0],[NSNumber numberWithFloat:3.0]];
+  _modelConfig.means = @[[NSNumber numberWithFloat:103.94], [NSNumber numberWithFloat:116.78], [NSNumber numberWithFloat:123.68]];
+  _modelConfig.scale = 1;
+  _modelConfig.dims = @[[NSNumber numberWithFloat:1], [NSNumber numberWithFloat:160.], [NSNumber numberWithFloat:160.0],[NSNumber numberWithFloat:3.0]];
+  _modelConfig.modelPointer = buffer;
+  _modelConfig.modelSize = (int)fileSize;
+  _modelConfig.paramPointer = parmaBuffer;
+  _modelConfig.paramSize = (int)paramfileSize;
+}
+- (IBAction)loaderButtonPressed:(id)sender {
+//  _runner = [[PaddleMobileGPU alloc] initWithCommandQueue:self.queue net:GenetType modelConfig:_modelConfig];
+  _runner = [[PaddleMobileGPU alloc] initWithCommandQueue:self.queue net:MobileNetSSDType modelConfig:_modelConfig];
+  
+  [_runner load];
+}
+- (IBAction)predictButtonPressed:(id)sender {
+  [self predict];
+}
+
+- (id<MTLTexture>) createTextureFromImage:(UIImage*) image device:(id<MTLDevice>) device
+{
+  image  =[UIImage imageWithCGImage:[image CGImage]
+                              scale:[image scale]
+                        orientation: UIImageOrientationLeft];
+  
+  NSLog(@"orientation and size and stuff %ld %f %f", (long)image.imageOrientation, image.size.width, image.size.height);
+  
+  CGImageRef imageRef = image.CGImage;
+  
+  size_t width = self.view.frame.size.width;
+  size_t height = self.view.frame.size.height;
+  
+  size_t bitsPerComponent = CGImageGetBitsPerComponent(imageRef);
+  size_t bitsPerPixel = CGImageGetBitsPerPixel(imageRef);
+  
+  CGColorSpaceRef colorSpace = CGImageGetColorSpace(imageRef);
+  
+  CGImageAlphaInfo alphaInfo = CGImageGetAlphaInfo(imageRef);
+  
+  //  NSLog(@"%@ %u", colorSpace, alphaInfo);
+  
+  CGBitmapInfo bitmapInfo = kCGBitmapByteOrderDefault | alphaInfo;
+  //    NSLog(@"bitmap info %u", bitmapInfo);
+  
+  
+  CGContextRef context = CGBitmapContextCreate( NULL, width, height, bitsPerComponent, (bitsPerPixel / 8) * width, colorSpace, bitmapInfo);
+  
+  if( !context )
+  {
+    NSLog(@"Failed to load image, probably an unsupported texture type");
+    return nil;
+  }
+  
+  CGContextDrawImage( context, CGRectMake( 0, 0, width, height ), image.CGImage);
+  
+  
+  MTLPixelFormat format = MTLPixelFormatRGBA8Unorm;
+  
+  MTLTextureDescriptor *texDesc = [MTLTextureDescriptor texture2DDescriptorWithPixelFormat:format
+                                                                                     width:width
+                                                                                    height:height
+                                                                                 mipmapped:NO];
+  id<MTLTexture> texture = [device newTextureWithDescriptor:texDesc];
+  
+  [texture replaceRegion:MTLRegionMake2D(0, 0, width, height)
+             mipmapLevel:0
+               withBytes:CGBitmapContextGetData(context)
+             bytesPerRow:4 * width];
+  
+  return texture;
+}
+
+- (void)predict {
+  _texture = [self createTextureFromImage:[UIImage imageNamed:@"hand.jpg"] device:self.device];
+  NSTimeInterval startTime = [[NSDate date] timeIntervalSince1970];
+  NSInteger max = 428;
+  for (int i = 0;i < max; i ++) {
+    [_runner predict:_texture withCompletion:^(BOOL success , NSArray<NSNumber *> *result) {
+      if (success) {
+        if (i == max -1) {
+          double time = [[NSDate date] timeIntervalSince1970] - startTime;
+          time = (time/max)*1000;
+          NSLog(@"gap ==== %fms",time);
+        }
+//        for (int i = 0; i < result.count; i ++) {
+//          NSNumber *number = result[i];
+//          NSLog(@"result %d = %f:",i, [number floatValue]);
+//        }
+      }
+    }];
+  }
+}
+
+- (void)didReceiveMemoryWarning {
+    [super didReceiveMemoryWarning];
+    // Dispose of any resources that can be recreated.
+}
+
+/*
+#pragma mark - Navigation
+
+// In a storyboard-based application, you will often want to do a little preparation before navigation
+- (void)prepareForSegue:(UIStoryboardSegue *)segue sender:(id)sender {
+    // Get the new view controller using [segue destinationViewController].
+    // Pass the selected object to the new view controller.
+}
+*/
+
+@end

metal/paddle-mobile-demo/paddle-mobile-demo/MetalHelper.swift

-//
-//  MetalHelper.swift
-//  paddle-mobile-demo
-//
-//  Created by liuRuiLong on 2018/7/25.
-//  Copyright © 2018年 orange. All rights reserved.
-//
-
-import Metal
-import MetalKit
-import Foundation
-import paddle_mobile
-import MetalPerformanceShaders
-
-class MetalHelper {
-    let device: MTLDevice
-    let queue: MTLCommandQueue
-    let textureLoader: MTKTextureLoader
-    static let shared: MetalHelper = MetalHelper.init()
-    private init(){
-        device = MTLCreateSystemDefaultDevice()!
-        queue = device.makeCommandQueue()!
-        textureLoader = MTKTextureLoader.init(device: device)
-    }
-    
-    static func scaleTexture(queue: MTLCommandQueue, input: MTLTexture, size:(width: Int, height: Int), complete: @escaping (MTLTexture) -> Void) {
-        let tmpTextureDes = MTLTextureDescriptor.init()
-        tmpTextureDes.width = size.width
-        tmpTextureDes.height = size.height
-        tmpTextureDes.depth = 1
-        tmpTextureDes.usage = [.shaderRead, .shaderWrite]
-        tmpTextureDes.pixelFormat = .rgba32Float
-        tmpTextureDes.textureType = .type2D
-        tmpTextureDes.storageMode = .shared
-        tmpTextureDes.cpuCacheMode = .defaultCache
-        let dest = MetalHelper.shared.device.makeTexture(descriptor: tmpTextureDes)
-        
-        let scale = MPSImageLanczosScale.init(device: MetalHelper.shared.device)
-        
-        let buffer = queue.makeCommandBuffer()
-        scale.encode(commandBuffer: buffer!, sourceTexture: input, destinationTexture: dest!)
-        buffer?.addCompletedHandler({ (buffer) in
-            complete(dest!)
-        })
-        buffer?.commit()
-    }
-}
-

metal/paddle-mobile-demo/paddle-mobile-demo/ModelHelper.swift

-//
-//  ModelHelper.swift
-//  paddle-mobile-demo
-//
-//  Created by liuRuiLong on 2018/8/10.
-//  Copyright © 2018年 orange. All rights reserved.
-//
-
-import UIKit
-import MetalKit
-import Foundation
-import paddle_mobile
-import MetalPerformanceShaders
-
-class PreProccess: CusomKernel {
-    init(device: MTLDevice) {
-        let s = CusomKernel.Shape.init(inWidth: 224, inHeight: 224, inChannel: 3)
-        super.init(device: device, inFunctionName: "preprocess", outputDim: s, usePaddleMobileLib: false)
-    }
-}
-
-let modelHelperMap: [SupportModel : ModelHelper] = [.mobilenet : MobileNetHelper.init()]
-
-enum SupportModel: String{
-    case mobilenet = "mobilenet"
-    static func supportedModels() -> [SupportModel] {
-        return [.mobilenet]
-    }
-}
-
-protocol ModelHelper {
-    var dim: [Int] { get }
-    var modelPath: String { get }
-    var paramPath: String { get }
-    var modelDir: String { get }
-    var preprocessKernel: CusomKernel { get }
-    func getTexture(image: CGImage, getTexture: @escaping (MTLTexture) -> Void)
-    func resultStr(res: [Float]) -> String
-}
-
-extension ModelHelper {
-    func getTexture(image: CGImage, getTexture: @escaping (MTLTexture) -> Void) {
-        let texture = try? MetalHelper.shared.textureLoader.newTexture(cgImage: image, options: [:]) ?! " texture loader error"
-        MetalHelper.scaleTexture(queue: MetalHelper.shared.queue, input: texture!, size: (224, 224)) { (resTexture) in
-            getTexture(resTexture)
-        }
-    }
-}
-
-struct MobileNetHelper: ModelHelper{
-    class PreWords {
-        var contents: [String] = []
-        init(fileName: String, type: String = "txt", inBundle: Bundle = Bundle.main) {
-            if let filePath = inBundle.path(forResource: fileName, ofType: type) {
-                let string = try! String.init(contentsOfFile: filePath)
-                contents = string.components(separatedBy: CharacterSet.newlines).filter{$0.count > 10}.map{
-                    String($0[$0.index($0.startIndex, offsetBy: 10)...])
-                }
-            }else{
-                fatalError("no file call \(fileName)")
-            }
-        }
-        subscript(index: Int) -> String{
-            return contents[index]
-        }
-    }
-    let labels = PreWords.init(fileName: "synset")
-    
-    func resultStr(res: [Float]) -> String {
-        var s: [String] = []
-        res.top(r: 5).enumerated().forEach{
-            s.append(String(format: "%d: %@ (%3.2f%%)", $0 + 1, labels[$1.0], $1.1 * 100))
-        }
-        return s.joined(separator: "\n")
-    }
-    
-    var preprocessKernel: CusomKernel
-    let dim = [1, 224, 224, 3]
-    let modelPath: String
-    let paramPath: String
-    let modelDir: String
-    
-    init() {
-        modelPath = Bundle.main.path(forResource: "model", ofType: nil) ?! "model null"
-        paramPath = Bundle.main.path(forResource: "params", ofType: nil) ?! "para null"
-        modelDir = ""
-        preprocessKernel = PreProccess.init(device: MetalHelper.shared.device)
-    }
-}

metal/paddle-mobile-demo/paddle-mobile-demo/MultiPredictViewController.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import UIKit
+import paddle_mobile
+
+class MultiPredictViewController: UIViewController {
+  var runner1: Runner!
+  var runner2: Runner!
+  override func viewDidLoad() {
+    super.viewDidLoad()
+    let mobileNet = MobileNet_ssd_hand.init(device: MetalHelper.shared.device)
+    let genet = Genet.init(device: MetalHelper.shared.device)
+    runner1 = Runner.init(inNet: mobileNet, commandQueue: MetalHelper.shared.queue, inPlatform: .GPU)
+    let queue2 = MetalHelper.shared.device.makeCommandQueue()
+    
+    runner2 = Runner.init(inNet: genet, commandQueue: MetalHelper.shared.queue, inPlatform: .GPU)
+  }
+
+  @IBAction func predictAct(_ sender: Any) {
+    let success = self.runner2.load()
+//    DispatchQueue.global().async {
+      let image1 = UIImage.init(named: "hand.jpg")
+//      let success = self.runner2.load()
+//      if success {
+//        for i in 0..<10000 {
+//          print(i)
+//          self.runner2.predict(cgImage: image1!.cgImage!, completion: { (success, res) in
+//            print("result1: ")
+////            print(res)
+//          })
+//        }
+//      } else {
+//        print("load failed")
+//      }
+//      self.runner1.clear()
+//    }
+//    return
+//    DispatchQueue.global().async {
+////      sleep(1)
+//      let image1 = UIImage.init(named: "banana.jpeg")
+////      if success {
+//        for _ in 0..<10 {
+//          self.runner2.predict(cgImage: image1!.cgImage!, completion: { (success, res) in
+//            print("result2: ")
+//            print(res)
+//          })
+//        }
+////      } else {
+////        print("load failed")
+////      }
+////      self.runner2.clear()
+//    }
+  }
+}

metal/paddle-mobile-demo/paddle-mobile-demo/Net/MetalHelper.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Metal
+import MetalKit
+import Foundation
+import paddle_mobile
+
+class MetalHelper {
+  let device: MTLDevice
+  let queue: MTLCommandQueue
+  let textureLoader: MTKTextureLoader
+  static let shared: MetalHelper = MetalHelper.init()
+  private init(){
+    device = MTLCreateSystemDefaultDevice()!
+    queue = device.makeCommandQueue()!
+    textureLoader = MTKTextureLoader.init(device: device)
+  }
+  
+
+}
+

metal/paddle-mobile-demo/paddle-mobile-demo/Net/PaddleMobile.swift

+//
+//  PaddleMobile.swift
+//  paddle-mobile-demo
+//
+//  Created by liuRuiLong on 2018/9/5.
+//  Copyright © 2018年 orange. All rights reserved.
+//
+
+import Foundation

metal/paddle-mobile-demo/paddle-mobile-demo/Net/PreProcessKernel.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+using namespace metal;
+
+
+kernel void mobilenet_preprocess(
+                       texture2d<float, access::read> inTexture [[texture(0)]],
+                       texture2d<float, access::write> outTexture [[texture(1)]],
+                       uint2 gid [[thread_position_in_grid]])
+{
+    if (gid.x >= outTexture.get_width() ||
+        gid.y >= outTexture.get_height()) {
+        return;
+    }
+    const auto means = float4(123.68f, 116.78f, 103.94f, 0.0f);
+    const float4 inColor = (inTexture.read(gid) * 255.0 - means) * 0.017;
+    outTexture.write(float4(inColor.z, inColor.y, inColor.x, 0.0f), gid);
+}
+
+kernel void mobilenet_preprocess_half(
+                       texture2d<half, access::read> inTexture [[texture(0)]],
+                       texture2d<half, access::write> outTexture [[texture(1)]],
+                       uint2 gid [[thread_position_in_grid]])
+{
+    if (gid.x >= outTexture.get_width() ||
+        gid.y >= outTexture.get_height()) {
+        return;
+    }
+    const auto means = half4(123.68f, 116.78f, 103.94f, 0.0f);
+    const half4 inColor = (inTexture.read(gid) * 255.0 - means) * 0.017;
+    outTexture.write(half4(inColor.z, inColor.y, inColor.x, 0.0f), gid);
+}
+
+kernel void mobilenet_ssd_preprocess(
+                       texture2d<float, access::read> inTexture [[texture(0)]],
+                       texture2d<float, access::write> outTexture [[texture(1)]],
+                       uint2 gid [[thread_position_in_grid]])
+{
+    if (gid.x >= outTexture.get_width() ||
+        gid.y >= outTexture.get_height()) {
+        return;
+    }
+    const auto means = float4(123.68f, 116.78f, 103.94f, 0.0f);
+    const float4 inColor = (inTexture.read(gid) * 255.0 - means) * 0.017;
+    outTexture.write(float4(inColor.z, inColor.y, inColor.x, 0.0f), gid);
+}
+
+kernel void mobilenet_ssd_preprocess_half(
+                            texture2d<half, access::read> inTexture [[texture(0)]],
+                            texture2d<half, access::write> outTexture [[texture(1)]],
+                            uint2 gid [[thread_position_in_grid]])
+{
+    if (gid.x >= outTexture.get_width() ||
+        gid.y >= outTexture.get_height()) {
+        return;
+    }
+    const auto means = half4(123.68f, 116.78f, 103.94f, 0.0f);
+    const half4 inColor = (inTexture.read(gid) * 255.0 - means) * 0.017;
+    outTexture.write(half4(inColor.z, inColor.y, inColor.x, 0.0f), gid);
+}
+
+kernel void genet_preprocess(texture2d<float, access::read> inTexture [[texture(0)]], texture2d<float, access::write> outTexture [[texture(1)]], uint2 gid [[thread_position_in_grid]])
+{
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height()) {
+    return;
+  }
+  const auto means = float4(128.0f, 128.0f, 128.0f, 0.0f);
+  const float4 inColor = (inTexture.read(gid) * 255.0 - means) * 0.017;
+  outTexture.write(float4(inColor.z, inColor.y, inColor.x, 0.0f), gid);
+}
+
+kernel void genet_preprocess_half(texture2d<half, access::read> inTexture [[texture(0)]], texture2d<half, access::write> outTexture [[texture(1)]], uint2 gid [[thread_position_in_grid]])
+{
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height()) {
+    return;
+  }
+  const auto means = half4(128.0f, 128.0f, 128.0f, 0.0f);
+  const half4 inColor = (inTexture.read(gid) * 255.0 - means) * 0.017;
+  outTexture.write(half4(inColor.z, inColor.y, inColor.x, 0.0f), gid);
+}
+
+kernel void mobilent_ar_preprocess(texture2d<float, access::read> inTexture [[texture(0)]], texture2d<float, access::write> outTexture [[texture(1)]], uint2 gid [[thread_position_in_grid]])
+{
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height()) {
+    return;
+  }
+  const auto means = float4(128.0f, 128.0f, 128.0f, 0.0f);
+  const float4 inColor = (inTexture.read(gid) * 255.0 - means) * 0.017;
+  outTexture.write(float4(inColor.z, inColor.y, inColor.x, 0.0f), gid);
+}
+
+kernel void mobilent_ar_preprocess_half(texture2d<half, access::read> inTexture [[texture(0)]], texture2d<half, access::write> outTexture [[texture(1)]], uint2 gid [[thread_position_in_grid]])
+{
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height()) {
+    return;
+  }
+  const auto means = half4(128.0f, 128.0f, 128.0f, 0.0f);
+  const half4 inColor = (inTexture.read(gid) * 255.0 - means) * 0.017;
+  outTexture.write(half4(inColor.z, inColor.y, inColor.x, 0.0f), gid);
+}
+
+kernel void scale(texture2d<float, access::sample> inTexture [[texture(0)]], texture2d<float, access::write> outTexture [[texture(1)]], uint2 gid [[thread_position_in_grid]]) {
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height()) return;
+  float w_stride = inTexture.get_width() / outTexture.get_width();
+  float h_stride = inTexture.get_height() / outTexture.get_height();
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  float4 input = inTexture.sample(sample, float2(gid.x * w_stride,    gid.y * h_stride), 0);
+  outTexture.write(input, gid);
+}
+
+kernel void scale_half(texture2d<float, access::sample> inTexture [[texture(0)]], texture2d<half, access::write> outTexture [[texture(1)]], uint2 gid [[thread_position_in_grid]]) {
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height()) return;
+  float w_stride = inTexture.get_width() / outTexture.get_width();
+  float h_stride = inTexture.get_height() / outTexture.get_height();
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  float4 input = inTexture.sample(sample, float2(gid.x * w_stride,    gid.y * h_stride), 0);
+  outTexture.write(half4(input), gid);
+}

metal/paddle-mobile-demo/paddle-mobile-demo/PreProcessKernel.metal

-//
-//  PreProcessKernel.metal
-//  paddle-mobile-demo
-//
-//  Created by liuRuiLong on 2018/7/20.
-//  Copyright © 2018年 orange. All rights reserved.
-//
-
-#include <metal_stdlib>
-using namespace metal;
-
-
-kernel void preprocess(
-                       texture2d<float, access::read> inTexture [[texture(0)]],
-                       texture2d<float, access::write> outTexture [[texture(1)]],
-                       uint2 gid [[thread_position_in_grid]])
-{
-    if (gid.x >= outTexture.get_width() ||
-        gid.y >= outTexture.get_height()) {
-        return;
-    }
-    const auto means = float4(123.68f, 116.78f, 103.94f, 0.0f);
-    const float4 inColor = (inTexture.read(gid) * 255.0 - means) * 0.017;
-    outTexture.write(float4(inColor.z, inColor.y, inColor.x, 0.0f), gid);
-}
-
-kernel void preprocess_half(
-                       texture2d<half, access::read> inTexture [[texture(0)]],
-                       texture2d<half, access::write> outTexture [[texture(1)]],
-                       uint2 gid [[thread_position_in_grid]])
-{
-    if (gid.x >= outTexture.get_width() ||
-        gid.y >= outTexture.get_height()) {
-        return;
-    }
-    const auto means = half4(123.68f, 116.78f, 103.94f, 0.0f);
-    const half4 inColor = (inTexture.read(gid) * 255.0 - means) * 0.017;
-    outTexture.write(half4(inColor.z, inColor.y, inColor.x, 0.0f), gid);
-}
-
-
-
-
-

metal/paddle-mobile-demo/paddle-mobile-demo/VideoCapture/FPSCounter.swift

+
+
+import Foundation
+import QuartzCore
+
+public class FPSCounter {
+  private(set) public var fps: Double = 0
+
+  var frames = 0
+  var startTime: CFTimeInterval = 0
+
+  public func start() {
+    frames = 0
+    startTime = CACurrentMediaTime()
+  }
+
+  public func frameCompleted() {
+    frames += 1
+    let now = CACurrentMediaTime()
+    let elapsed = now - startTime
+    if elapsed > 0.1 {
+      let current = Double(frames) / elapsed
+      let smoothing = 0.75
+      fps = smoothing*fps + (1 - smoothing)*current
+      if elapsed > 1 {
+        frames = 0
+        startTime = CACurrentMediaTime()
+      }
+    }
+  }
+}

metal/paddle-mobile-demo/paddle-mobile-demo/VideoCapture/VideoCapture.swift

+
+import UIKit
+import Metal
+import CoreVideo
+import AVFoundation
+
+@available(iOS 10.0, *)
+@objc public protocol VideoCaptureDelegate: NSObjectProtocol {
+  @objc optional func videoCapture(_ capture: VideoCapture, didCaptureSampleBuffer sampleBuffer: CMSampleBuffer, timestamp: CMTime)
+    @objc optional func videoCapture(_ capture: VideoCapture, didCaptureVideoTexture texture: MTLTexture?, timestamp: CMTime)
+    @objc optional func videoCapture(_ capture: VideoCapture, didCapturePhoto previewImage: UIImage?)
+    @objc optional func videoCapture(_ capture: VideoCapture, didCapturePhotoTexture texture: MTLTexture?)
+}
+
+/**
+  Simple interface to the iPhone's camera.
+*/
+@available(iOS 10.0, *)
+public class VideoCapture: NSObject {
+    public var previewLayer: AVCaptureVideoPreviewLayer?
+    public weak var delegate: VideoCaptureDelegate?
+    public var fps = -1
+    private let device: MTLDevice?
+    private let videoOrientation: AVCaptureVideoOrientation
+    private var textureCache: CVMetalTextureCache?
+    private let captureSession = AVCaptureSession()
+    private let videoOutput = AVCaptureVideoDataOutput()
+    private let photoOutput = AVCapturePhotoOutput()
+    private let queue = DispatchQueue(label: "net.machinethink.camera-queue")
+    private var lastTimestamp = CMTime()
+    private let cameraPosition: AVCaptureDevice.Position
+    public init(device: MTLDevice? = nil, orientation: AVCaptureVideoOrientation = .portrait, position: AVCaptureDevice.Position = .back) {
+        self.device = device
+        self.videoOrientation = orientation
+        self.cameraPosition = position
+        super.init()
+    }
+
+    public func setUp(sessionPreset: AVCaptureSession.Preset = .medium,
+                    completion: @escaping (Bool) -> Void) {
+        queue.async {
+            let success = self.setUpCamera(sessionPreset: sessionPreset)
+            DispatchQueue.main.async {
+                completion(success)
+            }
+        }
+    }
+
+    func fontCamera() -> AVCaptureDevice? {
+        let deveices = AVCaptureDevice.DiscoverySession.init(deviceTypes: [.builtInWideAngleCamera], mediaType: AVMediaType.video, position: .front).devices
+        return deveices.first
+        
+    }
+    
+    func setUpCamera(sessionPreset: AVCaptureSession.Preset) -> Bool {
+        if let inDevice = device{
+            guard CVMetalTextureCacheCreate(kCFAllocatorDefault, nil, inDevice, nil, &textureCache) == kCVReturnSuccess else {
+                print("Error: could not create a texture cache")
+                return false
+            }
+        }
+        
+        captureSession.beginConfiguration()
+        captureSession.sessionPreset = sessionPreset
+
+        var oCaptureDevice: AVCaptureDevice?
+        switch cameraPosition {
+        case .back:
+            oCaptureDevice = AVCaptureDevice.default(for: AVMediaType.video)
+            break
+        case .front:
+            oCaptureDevice = fontCamera()
+            break
+        default:
+            break
+        }
+        
+        guard let captureDevice = oCaptureDevice else {
+            print("Error: no video devices available")
+            return false
+        }
+
+        guard let videoInput = try? AVCaptureDeviceInput(device: captureDevice) else {
+            print("Error: could not create AVCaptureDeviceInput")
+            return false
+        }
+
+        if captureSession.canAddInput(videoInput) {
+            captureSession.addInput(videoInput)
+        }
+
+        let previewLayer = AVCaptureVideoPreviewLayer(session: captureSession)
+        previewLayer.videoGravity = AVLayerVideoGravity.resizeAspect
+        previewLayer.connection?.videoOrientation = self.videoOrientation
+        self.previewLayer = previewLayer
+
+        let settings: [String : Any] = [
+        kCVPixelBufferPixelFormatTypeKey as String: NSNumber(value: kCVPixelFormatType_32BGRA)
+        ]
+
+        videoOutput.videoSettings = settings
+        videoOutput.alwaysDiscardsLateVideoFrames = true
+        videoOutput.setSampleBufferDelegate(self, queue: queue)
+        if captureSession.canAddOutput(videoOutput) {
+            captureSession.addOutput(videoOutput)
+        }
+
+        // We want the buffers to be in portrait orientation otherwise they are
+        // rotated by 90 degrees. Need to set this _after_ addOutput()!
+        videoOutput.connection(with: AVMediaType.video)?.videoOrientation = self.videoOrientation
+
+        if captureSession.canAddOutput(photoOutput) {
+            captureSession.addOutput(photoOutput)
+        }
+
+        captureSession.commitConfiguration()
+        return true
+    }
+
+    public func start() {
+        if !captureSession.isRunning {
+            captureSession.startRunning()
+        }
+    }
+
+    public func stop() {
+        if captureSession.isRunning {
+            captureSession.stopRunning()
+        }
+    }
+
+    /* Captures a single frame of the camera input. */
+    public func capturePhoto() {
+        let settings = AVCapturePhotoSettings(format: [kCVPixelBufferPixelFormatTypeKey as String: NSNumber(value: kCVPixelFormatType_32BGRA)])
+        settings.previewPhotoFormat = [
+            kCVPixelBufferPixelFormatTypeKey as String: settings.__availablePreviewPhotoPixelFormatTypes[0],
+            kCVPixelBufferWidthKey as String: 480,
+            kCVPixelBufferHeightKey as String: 360,
+        ]
+        photoOutput.capturePhoto(with: settings, delegate: self)
+    }
+
+    func convertToMTLTexture(sampleBuffer: CMSampleBuffer?) -> MTLTexture? {
+        if let textureCache = textureCache, let sampleBuffer = sampleBuffer, let imageBuffer = CMSampleBufferGetImageBuffer(sampleBuffer) {
+            let width = CVPixelBufferGetWidth(imageBuffer)
+            let height = CVPixelBufferGetHeight(imageBuffer)
+            var texture: CVMetalTexture?
+            CVMetalTextureCacheCreateTextureFromImage(kCFAllocatorDefault, textureCache, imageBuffer, nil, .bgra8Unorm, width, height, 0, &texture)
+            if let texture = texture {
+                return CVMetalTextureGetTexture(texture)
+            }
+        }
+        return nil
+    }
+
+    func convertToUIImage(sampleBuffer: CMSampleBuffer?) -> UIImage? {
+        if let sampleBuffer = sampleBuffer,
+            let imageBuffer = CMSampleBufferGetImageBuffer(sampleBuffer) {
+            let width = CVPixelBufferGetWidth(imageBuffer)
+            let height = CVPixelBufferGetHeight(imageBuffer)
+            let rect = CGRect(x: 0, y: 0, width: CGFloat(width), height: CGFloat(height))
+            let ciImage = CIImage(cvPixelBuffer: imageBuffer)
+            let ciContext = CIContext(options: nil)
+            if let cgImage = ciContext.createCGImage(ciImage, from: rect) {
+                return UIImage(cgImage: cgImage)
+            }
+        }
+        return nil
+    }
+}
+
+
+@available(iOS 10.0, *)
+extension VideoCapture: AVCaptureVideoDataOutputSampleBufferDelegate {
+  public func captureOutput(_ output: AVCaptureOutput, didOutput sampleBuffer: CMSampleBuffer, from connection: AVCaptureConnection) {
+    // Because lowering the capture device's FPS looks ugly in the preview,
+    // we capture at full speed but only call the delegate at its desired
+    // framerate. If `fps` is -1, we run at the full framerate.
+    let timestamp = CMSampleBufferGetPresentationTimeStamp(sampleBuffer)
+    let deltaTime = timestamp - lastTimestamp
+    if fps == -1 || deltaTime >= CMTimeMake(1, Int32(fps)) {
+        lastTimestamp = timestamp
+        self.delegate?.videoCapture?(self, didCaptureSampleBuffer: sampleBuffer, timestamp: timestamp)
+        if self.delegate?.responds(to: #selector(VideoCaptureDelegate.videoCapture(_:didCaptureVideoTexture:timestamp:))) ?? false{
+            let texture = convertToMTLTexture(sampleBuffer: sampleBuffer)
+            delegate?.videoCapture?(self, didCaptureVideoTexture: texture, timestamp: timestamp)
+        }
+    }
+  }
+
+  public func captureOutput(_ output: AVCaptureOutput, didDrop sampleBuffer: CMSampleBuffer, from connection: AVCaptureConnection) {
+    print("dropped frame")
+  }
+}
+
+@available(iOS 10.0, *)
+extension VideoCapture: AVCapturePhotoCaptureDelegate {
+  public func photoOutput(_ captureOutput: AVCapturePhotoOutput,
+                          didFinishProcessingPhoto photoSampleBuffer: CMSampleBuffer?,
+                          previewPhoto previewPhotoSampleBuffer: CMSampleBuffer?,
+                          resolvedSettings: AVCaptureResolvedPhotoSettings,
+                          bracketSettings: AVCaptureBracketedStillImageSettings?,
+                          error: Error?) {
+    var imageTexture: MTLTexture?
+    var previewImage: UIImage?
+    if error == nil {
+        if self.delegate?.responds(to: #selector(VideoCaptureDelegate.videoCapture(_:didCapturePhotoTexture:))) ?? false{
+            imageTexture = convertToMTLTexture(sampleBuffer: photoSampleBuffer)
+            self.delegate?.videoCapture?(self, didCapturePhotoTexture: imageTexture)
+        }
+        
+        if self.delegate?.responds(to: #selector(VideoCaptureDelegate.videoCapture(_:didCapturePhoto:))) ?? false{
+            previewImage = convertToUIImage(sampleBuffer: previewPhotoSampleBuffer)
+            self.delegate?.videoCapture?(self, didCapturePhoto: previewImage)
+        }
+    }
+  }
+}

metal/paddle-mobile-demo/paddle-mobile-demo/ViewController.swift

@@ -14,10 +14,34 @@
 
 import UIKit
 import MetalKit
+import CoreMedia
 import paddle_mobile
 import MetalPerformanceShaders
 
-let threadSupport = [1]
+var platform: Platform = .GPU
+let threadSupport: [(Platform, String)] = [(.GPU, "GPU"), (.CPU, "CPU")]
+
+//.mobilenet_ssd : Runner.init(inNet: MobileNet_ssd_hand.init(device: MetalHelper.shared.device), commandQueue: MetalHelper.shared.queue, inPlatform: platform),
+let modelHelperMap: [SupportModel : Runner] = [
+                                               .genet : Runner.init(inNet: Genet.init(device: MetalHelper.shared.device), commandQueue: MetalHelper.shared.queue, inPlatform: platform),
+                                               .mobilenet_ssd_ar : Runner.init(inNet: MobileNet_ssd_AR.init(device: MetalHelper.shared.device), commandQueue: MetalHelper.shared.queue, inPlatform: platform)]
+//, .genet : Genet.init()
+//let modelHelperMap: [SupportModel : Net] = [.mobilenet : MobileNet.init(), .mobilenet_ssd : MobileNet_ssd_hand.init()]
+
+let netSupport: [SupportModel : Net] = [.genet : Genet.init(device: MetalHelper.shared.device), .mobilenet_ssd_ar : MobileNet_ssd_AR.init(device: MetalHelper.shared.device)]
+
+enum SupportModel: String{
+  //  case mobilenet = "mobilenet"
+//  case mobilenet_ssd    = "mobilenetssd"
+  case genet            = "genet"
+  case mobilenet_ssd_ar = "mobilenetssd_ar"
+  
+  static func supportedModels() -> [SupportModel] {
+    // .mobilenet,
+    // .mobilenet_ssd,
+    return [.genet, .mobilenet_ssd_ar]
+  }
+}
 
 class ViewController: UIViewController {
   @IBOutlet weak var resultTextView: UITextView!
@@ -25,28 +49,40 @@ class ViewController: UIViewController {
   @IBOutlet weak var elapsedTimeLabel: UILabel!
   @IBOutlet weak var modelPickerView: UIPickerView!
   @IBOutlet weak var threadPickerView: UIPickerView!
+  @IBOutlet weak var videoView: UIView!
+//  var videoCapture: VideoCapture!
+
   var selectImage: UIImage?
-    var program: Program?
-    var executor: Executor<Float32>?
-    var modelType: SupportModel = .mobilenet
+  var inputPointer: UnsafeMutablePointer<Float32>?
+  var modelType: SupportModel = SupportModel.supportedModels()[0]
   var toPredictTexture: MTLTexture?
-    var modelHelper: ModelHelper {
-        return modelHelperMap[modelType] ?! " has no this type "
-    }
+  
+  var runner: Runner!
+  
   var threadNum = 1
   
   @IBAction func loadAct(_ sender: Any) {
-        let inModelHelper = modelHelper
-        let queue = MetalHelper.shared.queue
-        let loader = Loader<Float32>.init()
-        do {
-            let modelPath = inModelHelper.modelPath
-            let paraPath = inModelHelper.paramPath
+     runner = Runner.init(inNet: netSupport[modelType]!, commandQueue: MetalHelper.shared.queue, inPlatform: platform)
+    
+    if platform == .CPU {
+      if inputPointer == nil {
+        inputPointer = runner.preproccess(image: selectImage!.cgImage!)
+       
+      }
+    } else if platform == .GPU {
+      if self.toPredictTexture == nil {
+        runner.getTexture(image: selectImage!.cgImage!) {[weak self] (texture) in
+          self?.toPredictTexture = texture
+        }
+      }
+    } else {
+      fatalError( " unsupport " )
+    }
     
-            program = try loader.load(device: MetalHelper.shared.device, modelPath: modelPath, paraPath: paraPath)
-            executor = try Executor<Float32>.init(inDevice: MetalHelper.shared.device, inQueue: queue, inProgram: program!)
-        } catch let error {
-            print(error)
+    if runner.load() {
+      print(" load success ! ")
+    } else {
+      print(" load error ! ")
     }
   }
   
@@ -58,62 +94,124 @@ class ViewController: UIViewController {
   }
   
   @IBAction func clearAct(_ sender: Any) {
-        executor?.clear()
-        program = nil
-        executor = nil
-        
+    runner.clear()
   }
   
   @IBAction func predictAct(_ sender: Any) {
+    let max = 50
+    switch platform {
+    case .GPU:
       guard let inTexture = toPredictTexture else {
         resultTextView.text = "请选择图片 ! "
         return
       }
       
-        guard let inExecutor = executor else {
-            resultTextView.text = "请先 load ! "
-            return
+      for _ in 0..<10{
+        runner.predict(texture: inTexture) { (success, resultHolder)  in
+          resultHolder?.releasePointer()
+        }
       }
       
-        do {
-            let max = 100
-            var startDate = Date.init()
+      let startDate = Date.init()
       for i in 0..<max {
-                try inExecutor.predict(input: inTexture, expect: modelHelper.dim, completionHandle: { [weak self] (result) in
+        runner.predict(texture: inTexture) { [weak self] (success, resultHolder)  in
           guard let sSelf = self else {
             fatalError()
           }
+          if success {
+            if i == max - 1 {
+              let time = Date.init().timeIntervalSince(startDate)
+              DispatchQueue.main.async {
+//                print(resultHolder!.result![0])
+                sSelf.resultTextView.text = sSelf.runner.net.resultStr(res: resultHolder!)
+                
+                sSelf.elapsedTimeLabel.text = "平均耗时: \(time/Double(max) * 1000.0) ms"
+               
+              }
+            }
+          }
           
-                    if i == (max / 2 - 1) {
-                        startDate = Date.init()
+          DispatchQueue.main.async {
+            resultHolder?.releasePointer()
+          }
+//            print("释放")
+        }
+//        print("sleep before ")
+//        usleep(33000)
+//        print("sleep after ")
+      }
+    case .CPU:
+      guard let inInputPointer = inputPointer else {
+        fatalError( " need input pointer " )
       }
       
+      for _ in 0..<10 {
+        runner.predict(inputPointer: inInputPointer) { (success, res) in
+          res?.releaseOutput()
+        }
+      }
+      
+      let startDate = Date.init()
+      for i in 0..<max {
+        runner.predict(inputPointer: inInputPointer) { [weak self](success, res) in
+          guard let sSelf = self else {
+            fatalError()
+          }
+          if success {
             if i == max - 1 {
               let time = Date.init().timeIntervalSince(startDate)
               DispatchQueue.main.async {
-                            sSelf.resultTextView.text = sSelf.modelHelper.resultStr(res: result.resultArr)
-                            sSelf.elapsedTimeLabel.text = "平均耗时: \(time/Double(max/2) * 1000.0) ms"
+//                sSelf.resultTextView.text = sSelf.runner.net.resultStr(res: res)
+                sSelf.elapsedTimeLabel.text = "平均耗时: \(time/Double(max) * 1000.0) ms"
+              }
             }
           }
-                }, preProcessKernle: self.modelHelper.preprocessKernel)
+          res?.releaseOutput()
+        }
       }
-        } catch let error {
-            print(error)
     }
   }
   
   override func viewDidLoad() {
     super.viewDidLoad()
+    
+//    if runner.load() {
+//      print(" load success ! ")
+//    } else {
+//      print(" load error ! ")
+//    }
+//    
     modelPickerView.delegate = self
     modelPickerView.dataSource = self
     threadPickerView.delegate = self
     threadPickerView.dataSource = self
     
-        selectImage = UIImage.init(named: "banana.jpeg")
+    selectImage = UIImage.init(named: "hand.jpg")
     selectImageView.image = selectImage
-        modelHelper.getTexture(image: selectImage!.cgImage!) {[weak self] (texture) in
-            self?.toPredictTexture = texture
-        }
+    
+//    if platform == .CPU {
+//      inputPointer = runner.preproccess(image: selectImage!.cgImage!)
+//    } else if platform == .GPU {
+//      runner.getTexture(image: selectImage!.cgImage!) {[weak self] (texture) in
+//        self?.toPredictTexture = texture
+//      }
+//    } else {
+//      fatalError( " unsupport " )
+//    }
+    
+//    videoCapture = VideoCapture.init(device: MetalHelper.shared.device, orientation: .portrait, position: .back)
+//    videoCapture.fps = 30
+//    videoCapture.delegate = self
+//    videoCapture.setUp { (success) in
+//      DispatchQueue.main.async {
+//        if let preViewLayer = self.videoCapture.previewLayer {
+//          self.videoView.layer.addSublayer(preViewLayer)
+//          self.videoCapture.previewLayer?.frame = self.videoView.bounds
+//        }
+//        self.videoCapture.start()
+//      }
+//    }
+
   }
 }
 
@@ -142,7 +240,7 @@ extension ViewController: UIPickerViewDataSource, UIPickerViewDelegate{
     if pickerView == modelPickerView {
       return SupportModel.supportedModels()[row].rawValue
     } else if pickerView == threadPickerView {
-            return "\(threadSupport[row])"
+      return threadSupport[row].1
     } else {
       fatalError()
     }
@@ -152,7 +250,8 @@ extension ViewController: UIPickerViewDataSource, UIPickerViewDelegate{
     if pickerView == modelPickerView {
       self.modelType = SupportModel.supportedModels()[row]
     } else if pickerView == threadPickerView {
-            self.threadNum = threadSupport[row]
+      
+      platform = threadSupport[row].0
     } else {
       fatalError()
     }
@@ -167,11 +266,40 @@ extension ViewController:  UIImagePickerControllerDelegate, UINavigationControll
       }
       sSelf.selectImage = image
       sSelf.selectImageView.image = image
-            sSelf.modelHelper.getTexture(image: image.cgImage!, getTexture: { (texture) in
+      sSelf.runner.getTexture(image: image.cgImage!, getTexture: { (texture) in
         sSelf.toPredictTexture = texture
       })
     }
   }
 }
 
+var bool1 = false
+extension ViewController: VideoCaptureDelegate{
+  func predictTexture(texture: MTLTexture){
+    runner.scaleTexture(input: texture) { (scaledTexture) in
+      self.runner.predict(texture: scaledTexture, completion: { (success, resultHolder) in
+//        print(resultHolder!.result![0])
+        resultHolder?.releasePointer()
+      })
+    }
+  }
+  
+  
+//  @available(iOS 10.0, *)
+//  func videoCapture(_ capture: VideoCapture, didCaptureVideoTexture texture: MTLTexture?, timestamp: CMTime) {
+////    if !bool1 {
+////      DispatchQueue.main.asyncAfter(deadline: DispatchTime.init(uptimeNanoseconds: 500000000)) {
+//    self.predictTexture(texture: texture!)
+////      }
+//
+//
+////      bool1 = true
+////    }
+//
+//  }
+
+}
+
+
+
 

metal/paddle-mobile-demo/paddle-mobile-demo/paddle-mobile-demo-Bridging-Header.h

+//
+//  Use this file to import your target's public headers that you would like to expose to Swift.
+//
+
+#import <paddle_mobile/paddle_mobile.h>

metal/paddle-mobile-unit-test/paddle-mobile-unit-test/AppDelegate.swift

-//
-//  AppDelegate.swift
-//  paddle-mobile-unit-test
-//
-//  Created by liuRuiLong on 2018/8/10.
-//  Copyright © 2018年 orange. All rights reserved.
-//
+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
 
 import UIKit
 

metal/paddle-mobile-unit-test/paddle-mobile-unit-test/ViewController.swift

-//
-//  ViewController.swift
-//  paddle-mobile-unit-test
-//
-//  Created by liuRuiLong on 2018/8/10.
-//  Copyright © 2018年 orange. All rights reserved.
-//
+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
 
 import UIKit
+import Metal
+//import MetalKit
 import paddle_mobile
 
 class ViewController: UIViewController {
-
     override func viewDidLoad() {
         super.viewDidLoad()
+        let device = Metal.MTLCreateSystemDefaultDevice()!
+        let queue = device.makeCommandQueue()!
+        let test = PaddleMobileUnitTest.init(
+            inDevice: device,
+            inQueue: queue
+        )
+        test.testConcat()
+//        test.testReshape()
+//        test.testTranspose()
         print(" done ")
     }
 

metal/paddle-mobile/paddle-mobile.xcodeproj/project.pbxproj

@@ -7,7 +7,31 @@
 	objects = {
 
 /* Begin PBXBuildFile section */
+		4AA1EA862146625E00D0F791 /* BilinearInterpOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = 4AA1EA852146625E00D0F791 /* BilinearInterpOp.swift */; };
+		4AA1EA88214662BD00D0F791 /* BilinearInterpKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = 4AA1EA87214662BD00D0F791 /* BilinearInterpKernel.swift */; };
+		4AA1EA8A2146631C00D0F791 /* BilinearInterp.metal in Sources */ = {isa = PBXBuildFile; fileRef = 4AA1EA892146631C00D0F791 /* BilinearInterp.metal */; };
+		4AA1EA8C2146640900D0F791 /* SplitOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = 4AA1EA8B2146640900D0F791 /* SplitOp.swift */; };
+		4AA1EA8E2146647F00D0F791 /* SplitKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = 4AA1EA8D2146647F00D0F791 /* SplitKernel.swift */; };
+		4AA1EA90214664CD00D0F791 /* Split.metal in Sources */ = {isa = PBXBuildFile; fileRef = 4AA1EA8F214664CD00D0F791 /* Split.metal */; };
+		4AA1EA92214665D700D0F791 /* ShapeOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = 4AA1EA91214665D700D0F791 /* ShapeOp.swift */; };
+		4AA1EA942146661500D0F791 /* ShapeKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = 4AA1EA932146661500D0F791 /* ShapeKernel.swift */; };
+		4AA1EA982146666500D0F791 /* FlattenOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = 4AA1EA972146666500D0F791 /* FlattenOp.swift */; };
+		4AA1EA9E2148D6F900D0F791 /* ConcatKernel.inc.metal in Headers */ = {isa = PBXBuildFile; fileRef = 4AA1EA9D2148D6F900D0F791 /* ConcatKernel.inc.metal */; };
+		4AA1EAA02148DEEE00D0F791 /* ReshapeKernel.inc.metal in Sources */ = {isa = PBXBuildFile; fileRef = 4AA1EA9F2148DEEE00D0F791 /* ReshapeKernel.inc.metal */; };
+		4AA1EAA2214912CD00D0F791 /* FlattenKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = 4AA1EAA1214912CC00D0F791 /* FlattenKernel.swift */; };
+		4AA1EAA4214A295C00D0F791 /* Split.inc.metal in Sources */ = {isa = PBXBuildFile; fileRef = 4AA1EAA3214A295C00D0F791 /* Split.inc.metal */; };
+		4AA1EAA6214B5F6800D0F791 /* Shape.metal in Sources */ = {isa = PBXBuildFile; fileRef = 4AA1EAA5214B5F6800D0F791 /* Shape.metal */; };
+		4AA1EAA8214B7AFB00D0F791 /* BilinearInterp.inc.metal in Sources */ = {isa = PBXBuildFile; fileRef = 4AA1EAA7214B7AFB00D0F791 /* BilinearInterp.inc.metal */; };
+		4AA1EAAA214F53D800D0F791 /* BoxCoder.inc.metal in Sources */ = {isa = PBXBuildFile; fileRef = 4AA1EAA9214F53D800D0F791 /* BoxCoder.inc.metal */; };
+		4AA1EAAC214F55C800D0F791 /* Softmax.inc.metal in Sources */ = {isa = PBXBuildFile; fileRef = 4AA1EAAB214F55C800D0F791 /* Softmax.inc.metal */; };
+		4AA1EAAE214F5FD900D0F791 /* TransposeKernel.inc.metal in Sources */ = {isa = PBXBuildFile; fileRef = 4AA1EAAD214F5FD900D0F791 /* TransposeKernel.inc.metal */; };
+		4AF928772133F1DB005B6C3A /* BoxCoder.metal in Sources */ = {isa = PBXBuildFile; fileRef = 4AF928762133F1DB005B6C3A /* BoxCoder.metal */; };
+		4AF9287921341661005B6C3A /* Softmax.metal in Sources */ = {isa = PBXBuildFile; fileRef = 4AF9287821341661005B6C3A /* Softmax.metal */; };
+		4AF928822135673D005B6C3A /* ConcatKernel.metal in Sources */ = {isa = PBXBuildFile; fileRef = 4AF928812135673D005B6C3A /* ConcatKernel.metal */; };
+		4AF9288421357BE3005B6C3A /* Elementwise.metal in Sources */ = {isa = PBXBuildFile; fileRef = 4AF9288321357BE3005B6C3A /* Elementwise.metal */; };
 		D3831F70E7E0B565B9AC22DA /* Pods_paddle_mobile.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = DD2E06330A1E7129C918DB46 /* Pods_paddle_mobile.framework */; };
+		FC0226562138F33800F395E2 /* TransposeKernel.metal in Sources */ = {isa = PBXBuildFile; fileRef = FC0226552138F33800F395E2 /* TransposeKernel.metal */; };
+		FC0226582138F38D00F395E2 /* PoolKernel.metal in Sources */ = {isa = PBXBuildFile; fileRef = FC0226572138F38D00F395E2 /* PoolKernel.metal */; };
 		FC039B6F20E11C3C0081E9F8 /* paddle_mobile.h in Headers */ = {isa = PBXBuildFile; fileRef = FC039B6D20E11C3C0081E9F8 /* paddle_mobile.h */; settings = {ATTRIBUTES = (Public, ); }; };
 		FC039B9720E11C9A0081E9F8 /* Extensions.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC039B9420E11C9A0081E9F8 /* Extensions.swift */; };
 		FC039B9820E11C9A0081E9F8 /* Errors.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC039B9520E11C9A0081E9F8 /* Errors.swift */; };
@@ -35,17 +59,54 @@
 		FC0E2DBE20EE460D009C1FAC /* BatchNormKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC0E2DBD20EE460D009C1FAC /* BatchNormKernel.swift */; };
 		FC0E2DC020EE461F009C1FAC /* ElementwiseAddKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC0E2DBF20EE461F009C1FAC /* ElementwiseAddKernel.swift */; };
 		FC1B16B320EC9A4F00678B91 /* Kernels.metal in Sources */ = {isa = PBXBuildFile; fileRef = FC1B16B220EC9A4F00678B91 /* Kernels.metal */; };
-		FC1B186620ECF1C600678B91 /* ResizeKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC1B186520ECF1C600678B91 /* ResizeKernel.swift */; };
+		FC292C5421421B2F00CF622F /* PaddleMobileGPU.h in Headers */ = {isa = PBXBuildFile; fileRef = FC292C5321421B2E00CF622F /* PaddleMobileGPU.h */; settings = {ATTRIBUTES = (Public, ); }; };
+		FC292C5621421B4600CF622F /* PaddleMobileGPU.m in Sources */ = {isa = PBXBuildFile; fileRef = FC292C5521421B4600CF622F /* PaddleMobileGPU.m */; };
+		FC292C81214255BD00CF622F /* CPUCompute.mm in Sources */ = {isa = PBXBuildFile; fileRef = FC292C7C214255BC00CF622F /* CPUCompute.mm */; };
+		FC292C82214255BD00CF622F /* MobileNetSSD.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC292C7E214255BC00CF622F /* MobileNetSSD.swift */; };
+		FC292C85214257CB00CF622F /* CPUCompute.h in Headers */ = {isa = PBXBuildFile; fileRef = FC292C7D214255BC00CF622F /* CPUCompute.h */; settings = {ATTRIBUTES = (Public, ); }; };
+		FC292C872142624800CF622F /* Genet.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC292C862142624800CF622F /* Genet.swift */; };
+		FC33B0F02147659000714A93 /* MobileNet.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC33B0EF2147659000714A93 /* MobileNet.swift */; };
 		FC3602CC2108819F00FACB58 /* PaddleMobileUnitTest.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC3602CB2108819F00FACB58 /* PaddleMobileUnitTest.swift */; };
 		FC4CB74920F0B954007C0C6D /* ConvKernel.metal in Sources */ = {isa = PBXBuildFile; fileRef = FC4CB74820F0B954007C0C6D /* ConvKernel.metal */; };
 		FC4CB74B20F12C30007C0C6D /* ProgramOptimize.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC4CB74A20F12C30007C0C6D /* ProgramOptimize.swift */; };
+		FC4FD9752140E1DE0073E130 /* PaddleMobile.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC4FD9742140E1DE0073E130 /* PaddleMobile.swift */; };
+		FC4FD9792140E4980073E130 /* PaddleMobileCPU.h in Headers */ = {isa = PBXBuildFile; fileRef = FC4FD9772140E4980073E130 /* PaddleMobileCPU.h */; settings = {ATTRIBUTES = (Public, ); }; };
+		FC4FD97A2140E4980073E130 /* libpaddle-mobile.a in Frameworks */ = {isa = PBXBuildFile; fileRef = FC4FD9782140E4980073E130 /* libpaddle-mobile.a */; };
+		FC4FD97E2140F2C30073E130 /* libstdc++.tbd in Frameworks */ = {isa = PBXBuildFile; fileRef = FC4FD97D2140F2C30073E130 /* libstdc++.tbd */; };
 		FC5163F620EF556E00636C28 /* Texture2DTo2DArrayKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC5163F520EF556E00636C28 /* Texture2DTo2DArrayKernel.swift */; };
 		FC60DB8920E9AAA500FF203F /* MetalExtension.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC60DB8820E9AAA500FF203F /* MetalExtension.swift */; };
+		FC803BBF214CB65A0094B8E5 /* ConvAddPreluOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC803BBE214CB65A0094B8E5 /* ConvAddPreluOp.swift */; };
+		FC803BC1214CB77A0094B8E5 /* ConvAddPreluKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC803BC0214CB77A0094B8E5 /* ConvAddPreluKernel.swift */; };
+		FC803BC3214CB79C0094B8E5 /* ConvAddPreluKernel.metal in Sources */ = {isa = PBXBuildFile; fileRef = FC803BC2214CB79C0094B8E5 /* ConvAddPreluKernel.metal */; };
+		FC803BC5214CB8F00094B8E5 /* ConvAddPrelu.inc.metal in Sources */ = {isa = PBXBuildFile; fileRef = FC803BC4214CB8F00094B8E5 /* ConvAddPrelu.inc.metal */; };
+		FC803BC7214CBA820094B8E5 /* Macro.metal in Sources */ = {isa = PBXBuildFile; fileRef = FC803BC6214CBA820094B8E5 /* Macro.metal */; };
+		FC803BC9214CFC8D0094B8E5 /* FetchKernel.metal in Sources */ = {isa = PBXBuildFile; fileRef = FC803BC8214CFC8D0094B8E5 /* FetchKernel.metal */; };
 		FC82735920E3C04200BE430A /* OpCreator.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC82735820E3C04200BE430A /* OpCreator.swift */; };
+		FC9A19E32148C31300CD9CBF /* MobilenetSSD_AR.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC9A19E22148C31300CD9CBF /* MobilenetSSD_AR.swift */; };
 		FC9D037920E229E4000F735A /* OpParam.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC9D037820E229E4000F735A /* OpParam.swift */; };
 		FC9D038020E22FBB000F735A /* FeedOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC9D037F20E22FBB000F735A /* FeedOp.swift */; };
 		FC9D038220E2312E000F735A /* FetchOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC9D038120E2312E000F735A /* FetchOp.swift */; };
 		FC9D038420E23B01000F735A /* Texture.swift in Sources */ = {isa = PBXBuildFile; fileRef = FC9D038320E23B01000F735A /* Texture.swift */; };
+		FCA3A1632132A4AC00084FE5 /* ReshapeKernel.metal in Sources */ = {isa = PBXBuildFile; fileRef = FCA3A1622132A4AC00084FE5 /* ReshapeKernel.metal */; };
+		FCA3A1652132A5EB00084FE5 /* Common.metal in Sources */ = {isa = PBXBuildFile; fileRef = FCA3A1642132A5EB00084FE5 /* Common.metal */; };
+		FCA67B1721364EF000BD58AA /* ConvTransposeKernel.metal in Sources */ = {isa = PBXBuildFile; fileRef = FCA67B1621364EF000BD58AA /* ConvTransposeKernel.metal */; };
+		FCA67CD52138272900BD58AA /* ConvAddMetal.metal in Sources */ = {isa = PBXBuildFile; fileRef = FCA67CD42138272900BD58AA /* ConvAddMetal.metal */; };
+		FCA67CD7213827AC00BD58AA /* ConvAddBNReluKernel.metal in Sources */ = {isa = PBXBuildFile; fileRef = FCA67CD6213827AC00BD58AA /* ConvAddBNReluKernel.metal */; };
+		FCA67CD92138287B00BD58AA /* ConvBNReluKernel.metal in Sources */ = {isa = PBXBuildFile; fileRef = FCA67CD82138287B00BD58AA /* ConvBNReluKernel.metal */; };
+		FCBCCC572122F41300D94F7E /* DwConvBNReluOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCBCCC562122F41300D94F7E /* DwConvBNReluOp.swift */; };
+		FCBCCC592122F42700D94F7E /* ConvBNReluOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCBCCC582122F42700D94F7E /* ConvBNReluOp.swift */; };
+		FCBCCC5B2122F66F00D94F7E /* ConvBNReluKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCBCCC5A2122F66F00D94F7E /* ConvBNReluKernel.swift */; };
+		FCBCCC5D2122F8A100D94F7E /* DepthwiseConvOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCBCCC5C2122F8A100D94F7E /* DepthwiseConvOp.swift */; };
+		FCBCCC5F2122FB3B00D94F7E /* PriorBoxOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCBCCC5E2122FB3B00D94F7E /* PriorBoxOp.swift */; };
+		FCBCCC612122FBDF00D94F7E /* PriorBoxKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCBCCC602122FBDF00D94F7E /* PriorBoxKernel.swift */; };
+		FCBCCC632122FCC000D94F7E /* TransposeKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCBCCC622122FCC000D94F7E /* TransposeKernel.swift */; };
+		FCBCCC652122FCD700D94F7E /* TransposeOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCBCCC642122FCD700D94F7E /* TransposeOp.swift */; };
+		FCBCCC67212306B000D94F7E /* ConcatOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCBCCC66212306B000D94F7E /* ConcatOp.swift */; };
+		FCBCCC69212306D300D94F7E /* ConcatKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCBCCC68212306D300D94F7E /* ConcatKernel.swift */; };
+		FCBCCC6B2123071700D94F7E /* BoxcoderOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCBCCC6A2123071700D94F7E /* BoxcoderOp.swift */; };
+		FCBCCC6D2123073A00D94F7E /* BoxcoderKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCBCCC6C2123073A00D94F7E /* BoxcoderKernel.swift */; };
+		FCBCCC6F2123097100D94F7E /* MulticlassNMSOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCBCCC6E2123097100D94F7E /* MulticlassNMSOp.swift */; };
+		FCBCCC71212309A700D94F7E /* MulticlassNMSKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCBCCC70212309A700D94F7E /* MulticlassNMSKernel.swift */; };
 		FCD04E6620F314C50007374F /* PoolOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCD04E6520F314C50007374F /* PoolOp.swift */; };
 		FCD04E6820F315020007374F /* PoolKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCD04E6720F315020007374F /* PoolKernel.swift */; };
 		FCD04E6A20F319EC0007374F /* SoftmaxOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCD04E6920F319EC0007374F /* SoftmaxOp.swift */; };
@@ -55,15 +116,55 @@
 		FCD04E7220F343420007374F /* ConvAddOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCD04E7120F343420007374F /* ConvAddOp.swift */; };
 		FCD04E7420F3437E0007374F /* ConvAddKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCD04E7320F3437E0007374F /* ConvAddKernel.swift */; };
 		FCDC0FEB21099A1D00DC9EFB /* Tools.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCDC0FEA21099A1D00DC9EFB /* Tools.swift */; };
+		FCDDC6C6212F9FB800E5EF74 /* PreluKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCDDC6C5212F9FB800E5EF74 /* PreluKernel.swift */; };
+		FCDDC6C8212FA3CA00E5EF74 /* ConvTransposeKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCDDC6C7212FA3CA00E5EF74 /* ConvTransposeKernel.swift */; };
+		FCDDC6CA212FDF6800E5EF74 /* BatchNormKernel.metal in Sources */ = {isa = PBXBuildFile; fileRef = FCDDC6C9212FDF6800E5EF74 /* BatchNormKernel.metal */; };
+		FCDDC6CC212FDFDB00E5EF74 /* ReluKernel.metal in Sources */ = {isa = PBXBuildFile; fileRef = FCDDC6CB212FDFDB00E5EF74 /* ReluKernel.metal */; };
+		FCDDC6CF212FE14700E5EF74 /* PriorBoxKernel.metal in Sources */ = {isa = PBXBuildFile; fileRef = FCDDC6CE212FE14700E5EF74 /* PriorBoxKernel.metal */; };
+		FCDE8A33212A917900F4A8F6 /* ConvTransposeOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCDE8A32212A917900F4A8F6 /* ConvTransposeOp.swift */; };
+		FCE3A1A92153DE5100C37CDE /* ConvAddAddPreluOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCE3A1A82153DE5100C37CDE /* ConvAddAddPreluOp.swift */; };
+		FCE3A1AB2153DE8C00C37CDE /* ConvAddAddPreluKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCE3A1AA2153DE8C00C37CDE /* ConvAddAddPreluKernel.swift */; };
+		FCE3A1AD2153E8BA00C37CDE /* ElementwiseAddPreluOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCE3A1AC2153E8BA00C37CDE /* ElementwiseAddPreluOp.swift */; };
+		FCE3A1AF2153E8EE00C37CDE /* ElementwiseAddPreluKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCE3A1AE2153E8EE00C37CDE /* ElementwiseAddPreluKernel.swift */; };
+		FCE3A1B12153E90F00C37CDE /* ElementwiseAddPreluKernel.inc.metal in Sources */ = {isa = PBXBuildFile; fileRef = FCE3A1B02153E90F00C37CDE /* ElementwiseAddPreluKernel.inc.metal */; };
+		FCE3A1B32153E91900C37CDE /* ElementwiseAddPreluKernel.metal in Sources */ = {isa = PBXBuildFile; fileRef = FCE3A1B22153E91900C37CDE /* ElementwiseAddPreluKernel.metal */; };
+		FCE9D7B7214F869000B520C3 /* Net.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCE9D7B6214F869000B520C3 /* Net.swift */; };
+		FCE9D7B9214FAA4800B520C3 /* NMSFetchResultKernel.metal in Sources */ = {isa = PBXBuildFile; fileRef = FCE9D7B8214FAA4800B520C3 /* NMSFetchResultKernel.metal */; };
+		FCEB684A212F00DB00D2448E /* PreluKernel.metal in Sources */ = {isa = PBXBuildFile; fileRef = FCEB6849212F00DB00D2448E /* PreluKernel.metal */; };
+		FCEB684C212F093800D2448E /* PreluOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCEB684B212F093800D2448E /* PreluOp.swift */; };
 		FCEBC0F420F1FDD90099DBAF /* ConvAddBatchNormReluOp.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCEBC0F320F1FDD90099DBAF /* ConvAddBatchNormReluOp.swift */; };
 		FCEBC0F620F1FE120099DBAF /* ConvAddBatchNormReluKernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCEBC0F520F1FE120099DBAF /* ConvAddBatchNormReluKernel.swift */; };
 		FCF2D73820E64E70007AC5F5 /* Kernel.swift in Sources */ = {isa = PBXBuildFile; fileRef = FCF2D73720E64E70007AC5F5 /* Kernel.swift */; };
 /* End PBXBuildFile section */
 
 /* Begin PBXFileReference section */
+		4AA1EA852146625E00D0F791 /* BilinearInterpOp.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = BilinearInterpOp.swift; sourceTree = "<group>"; };
+		4AA1EA87214662BD00D0F791 /* BilinearInterpKernel.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = BilinearInterpKernel.swift; sourceTree = "<group>"; };
+		4AA1EA892146631C00D0F791 /* BilinearInterp.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; path = BilinearInterp.metal; sourceTree = "<group>"; };
+		4AA1EA8B2146640900D0F791 /* SplitOp.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = SplitOp.swift; sourceTree = "<group>"; };
+		4AA1EA8D2146647F00D0F791 /* SplitKernel.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = SplitKernel.swift; sourceTree = "<group>"; };
+		4AA1EA8F214664CD00D0F791 /* Split.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; path = Split.metal; sourceTree = "<group>"; };
+		4AA1EA91214665D700D0F791 /* ShapeOp.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = ShapeOp.swift; sourceTree = "<group>"; };
+		4AA1EA932146661500D0F791 /* ShapeKernel.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = ShapeKernel.swift; sourceTree = "<group>"; };
+		4AA1EA972146666500D0F791 /* FlattenOp.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = FlattenOp.swift; sourceTree = "<group>"; };
+		4AA1EA9D2148D6F900D0F791 /* ConcatKernel.inc.metal */ = {isa = PBXFileReference; explicitFileType = sourcecode.metal; fileEncoding = 4; path = ConcatKernel.inc.metal; sourceTree = "<group>"; };
+		4AA1EA9F2148DEEE00D0F791 /* ReshapeKernel.inc.metal */ = {isa = PBXFileReference; explicitFileType = sourcecode.metal; fileEncoding = 4; path = ReshapeKernel.inc.metal; sourceTree = "<group>"; };
+		4AA1EAA1214912CC00D0F791 /* FlattenKernel.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = FlattenKernel.swift; sourceTree = "<group>"; };
+		4AA1EAA3214A295C00D0F791 /* Split.inc.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; path = Split.inc.metal; sourceTree = "<group>"; };
+		4AA1EAA5214B5F6800D0F791 /* Shape.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; path = Shape.metal; sourceTree = "<group>"; };
+		4AA1EAA7214B7AFB00D0F791 /* BilinearInterp.inc.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; path = BilinearInterp.inc.metal; sourceTree = "<group>"; };
+		4AA1EAA9214F53D800D0F791 /* BoxCoder.inc.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; path = BoxCoder.inc.metal; sourceTree = "<group>"; };
+		4AA1EAAB214F55C800D0F791 /* Softmax.inc.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; path = Softmax.inc.metal; sourceTree = "<group>"; };
+		4AA1EAAD214F5FD900D0F791 /* TransposeKernel.inc.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; path = TransposeKernel.inc.metal; sourceTree = "<group>"; };
+		4AF928762133F1DB005B6C3A /* BoxCoder.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; path = BoxCoder.metal; sourceTree = "<group>"; };
+		4AF9287821341661005B6C3A /* Softmax.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; path = Softmax.metal; sourceTree = "<group>"; };
+		4AF928812135673D005B6C3A /* ConcatKernel.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; path = ConcatKernel.metal; sourceTree = "<group>"; };
+		4AF9288321357BE3005B6C3A /* Elementwise.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; path = Elementwise.metal; sourceTree = "<group>"; };
 		CDF58151D902A1CBAE56A0C2 /* Pods-paddle-mobile.debug.xcconfig */ = {isa = PBXFileReference; includeInIndex = 1; lastKnownFileType = text.xcconfig; name = "Pods-paddle-mobile.debug.xcconfig"; path = "../Pods/Target Support Files/Pods-paddle-mobile/Pods-paddle-mobile.debug.xcconfig"; sourceTree = "<group>"; };
 		DD2E06330A1E7129C918DB46 /* Pods_paddle_mobile.framework */ = {isa = PBXFileReference; explicitFileType = wrapper.framework; includeInIndex = 0; path = Pods_paddle_mobile.framework; sourceTree = BUILT_PRODUCTS_DIR; };
 		E2A7957C92EDA5C3BEC0FFC2 /* Pods-paddle-mobile.release.xcconfig */ = {isa = PBXFileReference; includeInIndex = 1; lastKnownFileType = text.xcconfig; name = "Pods-paddle-mobile.release.xcconfig"; path = "../Pods/Target Support Files/Pods-paddle-mobile/Pods-paddle-mobile.release.xcconfig"; sourceTree = "<group>"; };
+		FC0226552138F33800F395E2 /* TransposeKernel.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = TransposeKernel.metal; sourceTree = "<group>"; };
+		FC0226572138F38D00F395E2 /* PoolKernel.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = PoolKernel.metal; sourceTree = "<group>"; };
 		FC039B6A20E11C3C0081E9F8 /* paddle_mobile.framework */ = {isa = PBXFileReference; explicitFileType = wrapper.framework; includeInIndex = 0; path = paddle_mobile.framework; sourceTree = BUILT_PRODUCTS_DIR; };
 		FC039B6D20E11C3C0081E9F8 /* paddle_mobile.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = paddle_mobile.h; sourceTree = "<group>"; };
 		FC039B6E20E11C3C0081E9F8 /* Info.plist */ = {isa = PBXFileReference; lastKnownFileType = text.plist.xml; path = Info.plist; sourceTree = "<group>"; };
@@ -93,17 +194,54 @@
 		FC0E2DBD20EE460D009C1FAC /* BatchNormKernel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = BatchNormKernel.swift; sourceTree = "<group>"; };
 		FC0E2DBF20EE461F009C1FAC /* ElementwiseAddKernel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ElementwiseAddKernel.swift; sourceTree = "<group>"; };
 		FC1B16B220EC9A4F00678B91 /* Kernels.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = Kernels.metal; sourceTree = "<group>"; };
-		FC1B186520ECF1C600678B91 /* ResizeKernel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ResizeKernel.swift; sourceTree = "<group>"; };
+		FC292C5321421B2E00CF622F /* PaddleMobileGPU.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = PaddleMobileGPU.h; sourceTree = "<group>"; };
+		FC292C5521421B4600CF622F /* PaddleMobileGPU.m */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.objc; path = PaddleMobileGPU.m; sourceTree = "<group>"; };
+		FC292C7C214255BC00CF622F /* CPUCompute.mm */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.objcpp; path = CPUCompute.mm; sourceTree = "<group>"; };
+		FC292C7D214255BC00CF622F /* CPUCompute.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = CPUCompute.h; sourceTree = "<group>"; };
+		FC292C7E214255BC00CF622F /* MobileNetSSD.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = MobileNetSSD.swift; sourceTree = "<group>"; };
+		FC292C862142624800CF622F /* Genet.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = Genet.swift; sourceTree = "<group>"; };
+		FC33B0EF2147659000714A93 /* MobileNet.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = MobileNet.swift; sourceTree = "<group>"; };
 		FC3602CB2108819F00FACB58 /* PaddleMobileUnitTest.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = PaddleMobileUnitTest.swift; sourceTree = "<group>"; };
 		FC4CB74820F0B954007C0C6D /* ConvKernel.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = ConvKernel.metal; sourceTree = "<group>"; };
 		FC4CB74A20F12C30007C0C6D /* ProgramOptimize.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ProgramOptimize.swift; sourceTree = "<group>"; };
+		FC4FD9742140E1DE0073E130 /* PaddleMobile.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = PaddleMobile.swift; sourceTree = "<group>"; };
+		FC4FD9772140E4980073E130 /* PaddleMobileCPU.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = PaddleMobileCPU.h; sourceTree = "<group>"; };
+		FC4FD9782140E4980073E130 /* libpaddle-mobile.a */ = {isa = PBXFileReference; lastKnownFileType = archive.ar; path = "libpaddle-mobile.a"; sourceTree = "<group>"; };
+		FC4FD97D2140F2C30073E130 /* libstdc++.tbd */ = {isa = PBXFileReference; lastKnownFileType = "sourcecode.text-based-dylib-definition"; name = "libstdc++.tbd"; path = "usr/lib/libstdc++.tbd"; sourceTree = SDKROOT; };
 		FC5163F520EF556E00636C28 /* Texture2DTo2DArrayKernel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = Texture2DTo2DArrayKernel.swift; sourceTree = "<group>"; };
 		FC60DB8820E9AAA500FF203F /* MetalExtension.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = MetalExtension.swift; sourceTree = "<group>"; };
+		FC803BBE214CB65A0094B8E5 /* ConvAddPreluOp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ConvAddPreluOp.swift; sourceTree = "<group>"; };
+		FC803BC0214CB77A0094B8E5 /* ConvAddPreluKernel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ConvAddPreluKernel.swift; sourceTree = "<group>"; };
+		FC803BC2214CB79C0094B8E5 /* ConvAddPreluKernel.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = ConvAddPreluKernel.metal; sourceTree = "<group>"; };
+		FC803BC4214CB8F00094B8E5 /* ConvAddPrelu.inc.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = ConvAddPrelu.inc.metal; sourceTree = "<group>"; };
+		FC803BC6214CBA820094B8E5 /* Macro.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = Macro.metal; sourceTree = "<group>"; };
+		FC803BC8214CFC8D0094B8E5 /* FetchKernel.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = FetchKernel.metal; sourceTree = "<group>"; };
 		FC82735820E3C04200BE430A /* OpCreator.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = OpCreator.swift; sourceTree = "<group>"; };
+		FC9A19E22148C31300CD9CBF /* MobilenetSSD_AR.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = MobilenetSSD_AR.swift; sourceTree = "<group>"; };
 		FC9D037820E229E4000F735A /* OpParam.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = OpParam.swift; sourceTree = "<group>"; };
 		FC9D037F20E22FBB000F735A /* FeedOp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = FeedOp.swift; sourceTree = "<group>"; };
 		FC9D038120E2312E000F735A /* FetchOp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = FetchOp.swift; sourceTree = "<group>"; };
 		FC9D038320E23B01000F735A /* Texture.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = Texture.swift; sourceTree = "<group>"; };
+		FCA3A1622132A4AC00084FE5 /* ReshapeKernel.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = ReshapeKernel.metal; sourceTree = "<group>"; };
+		FCA3A1642132A5EB00084FE5 /* Common.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = Common.metal; sourceTree = "<group>"; };
+		FCA67B1621364EF000BD58AA /* ConvTransposeKernel.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = ConvTransposeKernel.metal; sourceTree = "<group>"; };
+		FCA67CD42138272900BD58AA /* ConvAddMetal.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = ConvAddMetal.metal; sourceTree = "<group>"; };
+		FCA67CD6213827AC00BD58AA /* ConvAddBNReluKernel.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = ConvAddBNReluKernel.metal; sourceTree = "<group>"; };
+		FCA67CD82138287B00BD58AA /* ConvBNReluKernel.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = ConvBNReluKernel.metal; sourceTree = "<group>"; };
+		FCBCCC562122F41300D94F7E /* DwConvBNReluOp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = DwConvBNReluOp.swift; sourceTree = "<group>"; };
+		FCBCCC582122F42700D94F7E /* ConvBNReluOp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ConvBNReluOp.swift; sourceTree = "<group>"; };
+		FCBCCC5A2122F66F00D94F7E /* ConvBNReluKernel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ConvBNReluKernel.swift; sourceTree = "<group>"; };
+		FCBCCC5C2122F8A100D94F7E /* DepthwiseConvOp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = DepthwiseConvOp.swift; sourceTree = "<group>"; };
+		FCBCCC5E2122FB3B00D94F7E /* PriorBoxOp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = PriorBoxOp.swift; sourceTree = "<group>"; };
+		FCBCCC602122FBDF00D94F7E /* PriorBoxKernel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = PriorBoxKernel.swift; sourceTree = "<group>"; };
+		FCBCCC622122FCC000D94F7E /* TransposeKernel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = TransposeKernel.swift; sourceTree = "<group>"; };
+		FCBCCC642122FCD700D94F7E /* TransposeOp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = TransposeOp.swift; sourceTree = "<group>"; };
+		FCBCCC66212306B000D94F7E /* ConcatOp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ConcatOp.swift; sourceTree = "<group>"; };
+		FCBCCC68212306D300D94F7E /* ConcatKernel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ConcatKernel.swift; sourceTree = "<group>"; };
+		FCBCCC6A2123071700D94F7E /* BoxcoderOp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = BoxcoderOp.swift; sourceTree = "<group>"; };
+		FCBCCC6C2123073A00D94F7E /* BoxcoderKernel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = BoxcoderKernel.swift; sourceTree = "<group>"; };
+		FCBCCC6E2123097100D94F7E /* MulticlassNMSOp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = MulticlassNMSOp.swift; sourceTree = "<group>"; };
+		FCBCCC70212309A700D94F7E /* MulticlassNMSKernel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = MulticlassNMSKernel.swift; sourceTree = "<group>"; };
 		FCD04E6520F314C50007374F /* PoolOp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = PoolOp.swift; sourceTree = "<group>"; };
 		FCD04E6720F315020007374F /* PoolKernel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = PoolKernel.swift; sourceTree = "<group>"; };
 		FCD04E6920F319EC0007374F /* SoftmaxOp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = SoftmaxOp.swift; sourceTree = "<group>"; };
@@ -113,9 +251,25 @@
 		FCD04E7120F343420007374F /* ConvAddOp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ConvAddOp.swift; sourceTree = "<group>"; };
 		FCD04E7320F3437E0007374F /* ConvAddKernel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ConvAddKernel.swift; sourceTree = "<group>"; };
 		FCDC0FEA21099A1D00DC9EFB /* Tools.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = Tools.swift; sourceTree = "<group>"; };
+		FCDDC6C5212F9FB800E5EF74 /* PreluKernel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = PreluKernel.swift; sourceTree = "<group>"; };
+		FCDDC6C7212FA3CA00E5EF74 /* ConvTransposeKernel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ConvTransposeKernel.swift; sourceTree = "<group>"; };
+		FCDDC6C9212FDF6800E5EF74 /* BatchNormKernel.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = BatchNormKernel.metal; sourceTree = "<group>"; };
+		FCDDC6CB212FDFDB00E5EF74 /* ReluKernel.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = ReluKernel.metal; sourceTree = "<group>"; };
+		FCDDC6CE212FE14700E5EF74 /* PriorBoxKernel.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = PriorBoxKernel.metal; sourceTree = "<group>"; };
+		FCDE8A32212A917900F4A8F6 /* ConvTransposeOp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ConvTransposeOp.swift; sourceTree = "<group>"; };
+		FCE3A1A82153DE5100C37CDE /* ConvAddAddPreluOp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ConvAddAddPreluOp.swift; sourceTree = "<group>"; };
+		FCE3A1AA2153DE8C00C37CDE /* ConvAddAddPreluKernel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ConvAddAddPreluKernel.swift; sourceTree = "<group>"; };
+		FCE3A1AC2153E8BA00C37CDE /* ElementwiseAddPreluOp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ElementwiseAddPreluOp.swift; sourceTree = "<group>"; };
+		FCE3A1AE2153E8EE00C37CDE /* ElementwiseAddPreluKernel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ElementwiseAddPreluKernel.swift; sourceTree = "<group>"; };
+		FCE3A1B02153E90F00C37CDE /* ElementwiseAddPreluKernel.inc.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = ElementwiseAddPreluKernel.inc.metal; sourceTree = "<group>"; };
+		FCE3A1B22153E91900C37CDE /* ElementwiseAddPreluKernel.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = ElementwiseAddPreluKernel.metal; sourceTree = "<group>"; };
+		FCE9D7B6214F869000B520C3 /* Net.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = Net.swift; sourceTree = "<group>"; };
+		FCE9D7B8214FAA4800B520C3 /* NMSFetchResultKernel.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = NMSFetchResultKernel.metal; sourceTree = "<group>"; };
+		FCEB6849212F00DB00D2448E /* PreluKernel.metal */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.metal; path = PreluKernel.metal; sourceTree = "<group>"; };
+		FCEB684B212F093800D2448E /* PreluOp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = PreluOp.swift; sourceTree = "<group>"; };
 		FCEBC0F320F1FDD90099DBAF /* ConvAddBatchNormReluOp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; name = ConvAddBatchNormReluOp.swift; path = "paddle-mobile/Operators/ConvAddBatchNormReluOp.swift"; sourceTree = SOURCE_ROOT; };
 		FCEBC0F520F1FE120099DBAF /* ConvAddBatchNormReluKernel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ConvAddBatchNormReluKernel.swift; sourceTree = "<group>"; };
-		FCF2D73720E64E70007AC5F5 /* Kernel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; name = Kernel.swift; path = "paddle-mobile/Operators/Kernels/Kernel.swift"; sourceTree = SOURCE_ROOT; };
+		FCF2D73720E64E70007AC5F5 /* Kernel.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; name = Kernel.swift; path = "paddle-mobile/Operators/Kernels/Base/Kernel.swift"; sourceTree = SOURCE_ROOT; };
 /* End PBXFileReference section */
 
 /* Begin PBXFrameworksBuildPhase section */
@@ -123,7 +277,9 @@
 			isa = PBXFrameworksBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
+				FC4FD97E2140F2C30073E130 /* libstdc++.tbd in Frameworks */,
 				D3831F70E7E0B565B9AC22DA /* Pods_paddle_mobile.framework in Frameworks */,
+				FC4FD97A2140E4980073E130 /* libpaddle-mobile.a in Frameworks */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
 		};
@@ -133,6 +289,7 @@
 		336CBE234BF5DE48658DE65F /* Frameworks */ = {
 			isa = PBXGroup;
 			children = (
+				FC4FD97D2140F2C30073E130 /* libstdc++.tbd */,
 				DD2E06330A1E7129C918DB46 /* Pods_paddle_mobile.framework */,
 			);
 			name = Frameworks;
@@ -168,10 +325,19 @@
 		FC039B6C20E11C3C0081E9F8 /* paddle-mobile */ = {
 			isa = PBXGroup;
 			children = (
+				FCE9D7B6214F869000B520C3 /* Net.swift */,
+				FC9A19E22148C31300CD9CBF /* MobilenetSSD_AR.swift */,
+				FC33B0EF2147659000714A93 /* MobileNet.swift */,
+				FC292C862142624800CF622F /* Genet.swift */,
+				FC292C7E214255BC00CF622F /* MobileNetSSD.swift */,
+				FC292C7C214255BC00CF622F /* CPUCompute.mm */,
+				FC292C7D214255BC00CF622F /* CPUCompute.h */,
+				FC292C5521421B4600CF622F /* PaddleMobileGPU.m */,
+				FC292C5321421B2E00CF622F /* PaddleMobileGPU.h */,
+				FC4FD9762140E4920073E130 /* CPU */,
+				FC4FD9742140E1DE0073E130 /* PaddleMobile.swift */,
 				FC039BAE20E11CC20081E9F8 /* Program */,
 				FC039BA320E11CBC0081E9F8 /* Operators */,
-				FC039BA120E11CB70081E9F8 /* Loader.swift */,
-				FC039B9A20E11CA00081E9F8 /* Executor.swift */,
 				FC039B9C20E11CB20081E9F8 /* framework */,
 				FC039B9320E11C9A0081E9F8 /* Common */,
 				FC039B6D20E11C3C0081E9F8 /* paddle_mobile.h */,
@@ -196,6 +362,8 @@
 		FC039B9C20E11CB20081E9F8 /* framework */ = {
 			isa = PBXGroup;
 			children = (
+				FC039BA120E11CB70081E9F8 /* Loader.swift */,
+				FC039B9A20E11CA00081E9F8 /* Executor.swift */,
 				FC039B9D20E11CB20081E9F8 /* Tensor.swift */,
 				FC039B9E20E11CB20081E9F8 /* Dim.swift */,
 				FC9D038320E23B01000F735A /* Texture.swift */,
@@ -219,6 +387,23 @@
 				FCD04E6920F319EC0007374F /* SoftmaxOp.swift */,
 				FCD04E6D20F31B4B0007374F /* ReshapeOp.swift */,
 				FCD04E7120F343420007374F /* ConvAddOp.swift */,
+				FCBCCC562122F41300D94F7E /* DwConvBNReluOp.swift */,
+				FCBCCC582122F42700D94F7E /* ConvBNReluOp.swift */,
+				FCBCCC5C2122F8A100D94F7E /* DepthwiseConvOp.swift */,
+				FCBCCC5E2122FB3B00D94F7E /* PriorBoxOp.swift */,
+				FCBCCC642122FCD700D94F7E /* TransposeOp.swift */,
+				FCBCCC66212306B000D94F7E /* ConcatOp.swift */,
+				FCBCCC6A2123071700D94F7E /* BoxcoderOp.swift */,
+				4AA1EA8B2146640900D0F791 /* SplitOp.swift */,
+				4AA1EA91214665D700D0F791 /* ShapeOp.swift */,
+				4AA1EA972146666500D0F791 /* FlattenOp.swift */,
+				4AA1EA852146625E00D0F791 /* BilinearInterpOp.swift */,
+				FCBCCC6E2123097100D94F7E /* MulticlassNMSOp.swift */,
+				FCDE8A32212A917900F4A8F6 /* ConvTransposeOp.swift */,
+				FCEB684B212F093800D2448E /* PreluOp.swift */,
+				FC803BBE214CB65A0094B8E5 /* ConvAddPreluOp.swift */,
+				FCE3A1A82153DE5100C37CDE /* ConvAddAddPreluOp.swift */,
+				FCE3A1AC2153E8BA00C37CDE /* ElementwiseAddPreluOp.swift */,
 			);
 			path = Operators;
 			sourceTree = "<group>";
@@ -243,24 +428,46 @@
 		FC086BA520E67E8500D85EF7 /* Kernels */ = {
 			isa = PBXGroup;
 			children = (
+				FCDDC6CD212FE02100E5EF74 /* Base */,
+				FCEB6837212F00B100D2448E /* metal */,
+				FCDDC6C7212FA3CA00E5EF74 /* ConvTransposeKernel.swift */,
 				FC0E2DBB20EE45FE009C1FAC /* ConvKernel.swift */,
-				FCF2D73720E64E70007AC5F5 /* Kernel.swift */,
-				FC1B16B220EC9A4F00678B91 /* Kernels.metal */,
-				FC1B186520ECF1C600678B91 /* ResizeKernel.swift */,
 				FC0E2DB920EE3B8D009C1FAC /* ReluKernel.swift */,
 				FC0E2DBD20EE460D009C1FAC /* BatchNormKernel.swift */,
 				FC0E2DBF20EE461F009C1FAC /* ElementwiseAddKernel.swift */,
 				FC5163F520EF556E00636C28 /* Texture2DTo2DArrayKernel.swift */,
-				FC4CB74820F0B954007C0C6D /* ConvKernel.metal */,
 				FCEBC0F520F1FE120099DBAF /* ConvAddBatchNormReluKernel.swift */,
 				FCD04E6720F315020007374F /* PoolKernel.swift */,
 				FCD04E6B20F31A280007374F /* SoftmaxKernel.swift */,
 				FCD04E6F20F31B720007374F /* ReshapeKernel.swift */,
+				4AA1EAA1214912CC00D0F791 /* FlattenKernel.swift */,
 				FCD04E7320F3437E0007374F /* ConvAddKernel.swift */,
+				FCBCCC5A2122F66F00D94F7E /* ConvBNReluKernel.swift */,
+				FCBCCC602122FBDF00D94F7E /* PriorBoxKernel.swift */,
+				FCBCCC622122FCC000D94F7E /* TransposeKernel.swift */,
+				FCBCCC68212306D300D94F7E /* ConcatKernel.swift */,
+				FCBCCC6C2123073A00D94F7E /* BoxcoderKernel.swift */,
+				4AA1EA8D2146647F00D0F791 /* SplitKernel.swift */,
+				4AA1EA932146661500D0F791 /* ShapeKernel.swift */,
+				4AA1EA87214662BD00D0F791 /* BilinearInterpKernel.swift */,
+				FCBCCC70212309A700D94F7E /* MulticlassNMSKernel.swift */,
+				FCDDC6C5212F9FB800E5EF74 /* PreluKernel.swift */,
+				FC803BC0214CB77A0094B8E5 /* ConvAddPreluKernel.swift */,
+				FCE3A1AA2153DE8C00C37CDE /* ConvAddAddPreluKernel.swift */,
+				FCE3A1AE2153E8EE00C37CDE /* ElementwiseAddPreluKernel.swift */,
 			);
 			path = Kernels;
 			sourceTree = "<group>";
 		};
+		FC4FD9762140E4920073E130 /* CPU */ = {
+			isa = PBXGroup;
+			children = (
+				FC4FD9782140E4980073E130 /* libpaddle-mobile.a */,
+				FC4FD9772140E4980073E130 /* PaddleMobileCPU.h */,
+			);
+			path = CPU;
+			sourceTree = "<group>";
+		};
 		FCD592FA20E248EC00252966 /* Base */ = {
 			isa = PBXGroup;
 			children = (
@@ -271,6 +478,56 @@
 			path = Base;
 			sourceTree = "<group>";
 		};
+		FCDDC6CD212FE02100E5EF74 /* Base */ = {
+			isa = PBXGroup;
+			children = (
+				FCF2D73720E64E70007AC5F5 /* Kernel.swift */,
+			);
+			path = Base;
+			sourceTree = "<group>";
+		};
+		FCEB6837212F00B100D2448E /* metal */ = {
+			isa = PBXGroup;
+			children = (
+				4AF928812135673D005B6C3A /* ConcatKernel.metal */,
+				4AA1EA9D2148D6F900D0F791 /* ConcatKernel.inc.metal */,
+				4AF9288321357BE3005B6C3A /* Elementwise.metal */,
+				FC1B16B220EC9A4F00678B91 /* Kernels.metal */,
+				FC4CB74820F0B954007C0C6D /* ConvKernel.metal */,
+				4AF928762133F1DB005B6C3A /* BoxCoder.metal */,
+				4AA1EAA9214F53D800D0F791 /* BoxCoder.inc.metal */,
+				4AA1EAA5214B5F6800D0F791 /* Shape.metal */,
+				4AA1EA8F214664CD00D0F791 /* Split.metal */,
+				4AA1EAA3214A295C00D0F791 /* Split.inc.metal */,
+				4AA1EA892146631C00D0F791 /* BilinearInterp.metal */,
+				4AA1EAA7214B7AFB00D0F791 /* BilinearInterp.inc.metal */,
+				4AF9287821341661005B6C3A /* Softmax.metal */,
+				4AA1EAAB214F55C800D0F791 /* Softmax.inc.metal */,
+				FCEB6849212F00DB00D2448E /* PreluKernel.metal */,
+				FCDDC6C9212FDF6800E5EF74 /* BatchNormKernel.metal */,
+				FCDDC6CB212FDFDB00E5EF74 /* ReluKernel.metal */,
+				FCDDC6CE212FE14700E5EF74 /* PriorBoxKernel.metal */,
+				FCA3A1622132A4AC00084FE5 /* ReshapeKernel.metal */,
+				4AA1EA9F2148DEEE00D0F791 /* ReshapeKernel.inc.metal */,
+				FCA3A1642132A5EB00084FE5 /* Common.metal */,
+				FCA67B1621364EF000BD58AA /* ConvTransposeKernel.metal */,
+				FCA67CD42138272900BD58AA /* ConvAddMetal.metal */,
+				FCA67CD6213827AC00BD58AA /* ConvAddBNReluKernel.metal */,
+				FCA67CD82138287B00BD58AA /* ConvBNReluKernel.metal */,
+				FC0226552138F33800F395E2 /* TransposeKernel.metal */,
+				4AA1EAAD214F5FD900D0F791 /* TransposeKernel.inc.metal */,
+				FC0226572138F38D00F395E2 /* PoolKernel.metal */,
+				FC803BC2214CB79C0094B8E5 /* ConvAddPreluKernel.metal */,
+				FC803BC4214CB8F00094B8E5 /* ConvAddPrelu.inc.metal */,
+				FC803BC6214CBA820094B8E5 /* Macro.metal */,
+				FC803BC8214CFC8D0094B8E5 /* FetchKernel.metal */,
+				FCE9D7B8214FAA4800B520C3 /* NMSFetchResultKernel.metal */,
+				FCE3A1B02153E90F00C37CDE /* ElementwiseAddPreluKernel.inc.metal */,
+				FCE3A1B22153E91900C37CDE /* ElementwiseAddPreluKernel.metal */,
+			);
+			path = metal;
+			sourceTree = "<group>";
+		};
 /* End PBXGroup section */
 
 /* Begin PBXHeadersBuildPhase section */
@@ -278,6 +535,10 @@
 			isa = PBXHeadersBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
+				FC4FD9792140E4980073E130 /* PaddleMobileCPU.h in Headers */,
+				FC292C85214257CB00CF622F /* CPUCompute.h in Headers */,
+				FC292C5421421B2F00CF622F /* PaddleMobileGPU.h in Headers */,
+				4AA1EA9E2148D6F900D0F791 /* ConcatKernel.inc.metal in Headers */,
 				FC039B6F20E11C3C0081E9F8 /* paddle_mobile.h in Headers */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
@@ -315,6 +576,7 @@
 				TargetAttributes = {
 					FC039B6920E11C3C0081E9F8 = {
 						CreatedOnToolsVersion = 9.3.1;
+						LastSwiftMigration = 0940;
 					};
 				};
 			};
@@ -372,53 +634,124 @@
 			buildActionMask = 2147483647;
 			files = (
 				FC9D038020E22FBB000F735A /* FeedOp.swift in Sources */,
+				4AA1EAAA214F53D800D0F791 /* BoxCoder.inc.metal in Sources */,
 				FC039B9F20E11CB20081E9F8 /* Tensor.swift in Sources */,
+				FC803BC9214CFC8D0094B8E5 /* FetchKernel.metal in Sources */,
+				FCA67CD7213827AC00BD58AA /* ConvAddBNReluKernel.metal in Sources */,
+				4AF9287921341661005B6C3A /* Softmax.metal in Sources */,
+				4AA1EA942146661500D0F791 /* ShapeKernel.swift in Sources */,
 				FC0E2DBC20EE45FE009C1FAC /* ConvKernel.swift in Sources */,
 				FC039BAA20E11CBC0081E9F8 /* ElementwiseAddOp.swift in Sources */,
+				FCDE8A33212A917900F4A8F6 /* ConvTransposeOp.swift in Sources */,
+				FCBCCC6B2123071700D94F7E /* BoxcoderOp.swift in Sources */,
+				4AA1EAAE214F5FD900D0F791 /* TransposeKernel.inc.metal in Sources */,
+				4AA1EAA4214A295C00D0F791 /* Split.inc.metal in Sources */,
+				FC803BC7214CBA820094B8E5 /* Macro.metal in Sources */,
 				FC039B9B20E11CA00081E9F8 /* Executor.swift in Sources */,
+				4AF9288421357BE3005B6C3A /* Elementwise.metal in Sources */,
 				FCD04E7020F31B720007374F /* ReshapeKernel.swift in Sources */,
+				FCE3A1B12153E90F00C37CDE /* ElementwiseAddPreluKernel.inc.metal in Sources */,
 				FCD04E7220F343420007374F /* ConvAddOp.swift in Sources */,
 				FC039BBB20E11CC20081E9F8 /* ProgramDesc.swift in Sources */,
+				FCE3A1AB2153DE8C00C37CDE /* ConvAddAddPreluKernel.swift in Sources */,
 				FC9D037920E229E4000F735A /* OpParam.swift in Sources */,
 				FC3602CC2108819F00FACB58 /* PaddleMobileUnitTest.swift in Sources */,
-				FC1B186620ECF1C600678B91 /* ResizeKernel.swift in Sources */,
 				FCF2D73820E64E70007AC5F5 /* Kernel.swift in Sources */,
+				FCDDC6CC212FDFDB00E5EF74 /* ReluKernel.metal in Sources */,
+				FC0226562138F33800F395E2 /* TransposeKernel.metal in Sources */,
+				FCDDC6C6212F9FB800E5EF74 /* PreluKernel.swift in Sources */,
+				FCA67CD52138272900BD58AA /* ConvAddMetal.metal in Sources */,
+				FCBCCC5B2122F66F00D94F7E /* ConvBNReluKernel.swift in Sources */,
+				4AA1EA8C2146640900D0F791 /* SplitOp.swift in Sources */,
+				FC292C81214255BD00CF622F /* CPUCompute.mm in Sources */,
 				FCEBC0F420F1FDD90099DBAF /* ConvAddBatchNormReluOp.swift in Sources */,
+				4AA1EAAC214F55C800D0F791 /* Softmax.inc.metal in Sources */,
 				FC0E2DC020EE461F009C1FAC /* ElementwiseAddKernel.swift in Sources */,
+				4AF928772133F1DB005B6C3A /* BoxCoder.metal in Sources */,
+				FC803BBF214CB65A0094B8E5 /* ConvAddPreluOp.swift in Sources */,
+				FC33B0F02147659000714A93 /* MobileNet.swift in Sources */,
+				FCEB684C212F093800D2448E /* PreluOp.swift in Sources */,
+				4AA1EAA8214B7AFB00D0F791 /* BilinearInterp.inc.metal in Sources */,
+				FCA67CD92138287B00BD58AA /* ConvBNReluKernel.metal in Sources */,
 				FC60DB8920E9AAA500FF203F /* MetalExtension.swift in Sources */,
 				FCEBC0F620F1FE120099DBAF /* ConvAddBatchNormReluKernel.swift in Sources */,
+				4AA1EA8A2146631C00D0F791 /* BilinearInterp.metal in Sources */,
+				FCDDC6CA212FDF6800E5EF74 /* BatchNormKernel.metal in Sources */,
 				FC1B16B320EC9A4F00678B91 /* Kernels.metal in Sources */,
 				FC039BBA20E11CC20081E9F8 /* TensorDesc.swift in Sources */,
 				FC039BA020E11CB20081E9F8 /* Dim.swift in Sources */,
 				FC039BB820E11CC20081E9F8 /* framework.pb.swift in Sources */,
 				FC039B9920E11C9A0081E9F8 /* Types.swift in Sources */,
 				FC4CB74920F0B954007C0C6D /* ConvKernel.metal in Sources */,
+				FCA3A1632132A4AC00084FE5 /* ReshapeKernel.metal in Sources */,
+				FC4FD9752140E1DE0073E130 /* PaddleMobile.swift in Sources */,
+				FCBCCC592122F42700D94F7E /* ConvBNReluOp.swift in Sources */,
 				FC039BA920E11CBC0081E9F8 /* ConvOp.swift in Sources */,
 				FC9D038420E23B01000F735A /* Texture.swift in Sources */,
+				FCE3A1B32153E91900C37CDE /* ElementwiseAddPreluKernel.metal in Sources */,
+				4AA1EAA2214912CD00D0F791 /* FlattenKernel.swift in Sources */,
+				4AA1EA982146666500D0F791 /* FlattenOp.swift in Sources */,
+				FCBCCC652122FCD700D94F7E /* TransposeOp.swift in Sources */,
+				4AA1EAA6214B5F6800D0F791 /* Shape.metal in Sources */,
 				FCD04E6E20F31B4B0007374F /* ReshapeOp.swift in Sources */,
 				FC039B9820E11C9A0081E9F8 /* Errors.swift in Sources */,
 				FC039BBF20E11CC20081E9F8 /* Attribute.swift in Sources */,
+				4AA1EA8E2146647F00D0F791 /* SplitKernel.swift in Sources */,
 				FCD04E7420F3437E0007374F /* ConvAddKernel.swift in Sources */,
 				FC039BB920E11CC20081E9F8 /* Scope.swift in Sources */,
+				FC292C5621421B4600CF622F /* PaddleMobileGPU.m in Sources */,
 				FCD04E6620F314C50007374F /* PoolOp.swift in Sources */,
+				FCE9D7B9214FAA4800B520C3 /* NMSFetchResultKernel.metal in Sources */,
 				FC039BAC20E11CBC0081E9F8 /* BatchNormOp.swift in Sources */,
+				FCBCCC6F2123097100D94F7E /* MulticlassNMSOp.swift in Sources */,
 				FC039BBC20E11CC20081E9F8 /* VarDesc.swift in Sources */,
+				FC292C872142624800CF622F /* Genet.swift in Sources */,
+				FC803BC5214CB8F00094B8E5 /* ConvAddPrelu.inc.metal in Sources */,
+				4AF928822135673D005B6C3A /* ConcatKernel.metal in Sources */,
+				FCBCCC632122FCC000D94F7E /* TransposeKernel.swift in Sources */,
+				FCBCCC71212309A700D94F7E /* MulticlassNMSKernel.swift in Sources */,
 				FCDC0FEB21099A1D00DC9EFB /* Tools.swift in Sources */,
 				FC0E2DBA20EE3B8D009C1FAC /* ReluKernel.swift in Sources */,
+				4AA1EA862146625E00D0F791 /* BilinearInterpOp.swift in Sources */,
+				FCBCCC6D2123073A00D94F7E /* BoxcoderKernel.swift in Sources */,
+				FCBCCC69212306D300D94F7E /* ConcatKernel.swift in Sources */,
+				FCDDC6C8212FA3CA00E5EF74 /* ConvTransposeKernel.swift in Sources */,
 				FC82735920E3C04200BE430A /* OpCreator.swift in Sources */,
+				FCA3A1652132A5EB00084FE5 /* Common.metal in Sources */,
+				4AA1EA92214665D700D0F791 /* ShapeOp.swift in Sources */,
+				FC803BC1214CB77A0094B8E5 /* ConvAddPreluKernel.swift in Sources */,
+				FCBCCC5D2122F8A100D94F7E /* DepthwiseConvOp.swift in Sources */,
+				FCE3A1AF2153E8EE00C37CDE /* ElementwiseAddPreluKernel.swift in Sources */,
+				FCE9D7B7214F869000B520C3 /* Net.swift in Sources */,
 				FC0E2DBE20EE460D009C1FAC /* BatchNormKernel.swift in Sources */,
 				FC039BAB20E11CBC0081E9F8 /* Operator.swift in Sources */,
 				FCD04E6A20F319EC0007374F /* SoftmaxOp.swift in Sources */,
+				FC292C82214255BD00CF622F /* MobileNetSSD.swift in Sources */,
+				FCBCCC612122FBDF00D94F7E /* PriorBoxKernel.swift in Sources */,
+				FCBCCC5F2122FB3B00D94F7E /* PriorBoxOp.swift in Sources */,
 				FC9D038220E2312E000F735A /* FetchOp.swift in Sources */,
+				FCA67B1721364EF000BD58AA /* ConvTransposeKernel.metal in Sources */,
 				FC039BBD20E11CC20081E9F8 /* Program.swift in Sources */,
 				FC039BA220E11CB70081E9F8 /* Loader.swift in Sources */,
+				FCBCCC67212306B000D94F7E /* ConcatOp.swift in Sources */,
 				FCD04E6C20F31A280007374F /* SoftmaxKernel.swift in Sources */,
+				FCEB684A212F00DB00D2448E /* PreluKernel.metal in Sources */,
+				4AA1EAA02148DEEE00D0F791 /* ReshapeKernel.inc.metal in Sources */,
+				FC9A19E32148C31300CD9CBF /* MobilenetSSD_AR.swift in Sources */,
+				FCDDC6CF212FE14700E5EF74 /* PriorBoxKernel.metal in Sources */,
 				FC4CB74B20F12C30007C0C6D /* ProgramOptimize.swift in Sources */,
+				FCE3A1A92153DE5100C37CDE /* ConvAddAddPreluOp.swift in Sources */,
 				FC5163F620EF556E00636C28 /* Texture2DTo2DArrayKernel.swift in Sources */,
+				FCE3A1AD2153E8BA00C37CDE /* ElementwiseAddPreluOp.swift in Sources */,
 				FC039BC020E11CC20081E9F8 /* BlockDesc.swift in Sources */,
+				FC803BC3214CB79C0094B8E5 /* ConvAddPreluKernel.metal in Sources */,
+				4AA1EA90214664CD00D0F791 /* Split.metal in Sources */,
 				FCD04E6820F315020007374F /* PoolKernel.swift in Sources */,
+				FC0226582138F38D00F395E2 /* PoolKernel.metal in Sources */,
 				FC039BAD20E11CBC0081E9F8 /* ReluOp.swift in Sources */,
+				FCBCCC572122F41300D94F7E /* DwConvBNReluOp.swift in Sources */,
 				FC039BBE20E11CC20081E9F8 /* OpDesc.swift in Sources */,
+				4AA1EA88214662BD00D0F791 /* BilinearInterpKernel.swift in Sources */,
 				FC039B9720E11C9A0081E9F8 /* Extensions.swift in Sources */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
@@ -550,6 +883,7 @@
 			isa = XCBuildConfiguration;
 			baseConfigurationReference = CDF58151D902A1CBAE56A0C2 /* Pods-paddle-mobile.debug.xcconfig */;
 			buildSettings = {
+				CLANG_ENABLE_MODULES = YES;
 				CODE_SIGN_IDENTITY = "";
 				CODE_SIGN_STYLE = Automatic;
 				DEFINES_MODULE = YES;
@@ -557,6 +891,7 @@
 				DYLIB_COMPATIBILITY_VERSION = 1;
 				DYLIB_CURRENT_VERSION = 1;
 				DYLIB_INSTALL_NAME_BASE = "@rpath";
+				ENABLE_BITCODE = NO;
 				INFOPLIST_FILE = "paddle-mobile/Info.plist";
 				INSTALL_PATH = "$(LOCAL_LIBRARY_DIR)/Frameworks";
 				IPHONEOS_DEPLOYMENT_TARGET = 9.0;
@@ -565,10 +900,16 @@
 					"@executable_path/Frameworks",
 					"@loader_path/Frameworks",
 				);
+				LIBRARY_SEARCH_PATHS = (
+					"$(inherited)",
+					"$(PROJECT_DIR)/paddle-mobile/CPU",
+				);
+				MACH_O_TYPE = mh_dylib;
 				MTL_LANGUAGE_REVISION = UseDeploymentTarget;
 				PRODUCT_BUNDLE_IDENTIFIER = "orange.paddle-mobile";
 				PRODUCT_NAME = "$(TARGET_NAME:c99extidentifier)";
 				SKIP_INSTALL = YES;
+				SWIFT_OPTIMIZATION_LEVEL = "-Onone";
 				SWIFT_VERSION = 4.0;
 				TARGETED_DEVICE_FAMILY = "1,2";
 			};
@@ -578,6 +919,7 @@
 			isa = XCBuildConfiguration;
 			baseConfigurationReference = E2A7957C92EDA5C3BEC0FFC2 /* Pods-paddle-mobile.release.xcconfig */;
 			buildSettings = {
+				CLANG_ENABLE_MODULES = YES;
 				CODE_SIGN_IDENTITY = "";
 				CODE_SIGN_STYLE = Automatic;
 				DEFINES_MODULE = YES;
@@ -585,6 +927,7 @@
 				DYLIB_COMPATIBILITY_VERSION = 1;
 				DYLIB_CURRENT_VERSION = 1;
 				DYLIB_INSTALL_NAME_BASE = "@rpath";
+				ENABLE_BITCODE = NO;
 				INFOPLIST_FILE = "paddle-mobile/Info.plist";
 				INSTALL_PATH = "$(LOCAL_LIBRARY_DIR)/Frameworks";
 				IPHONEOS_DEPLOYMENT_TARGET = 9.0;
@@ -593,6 +936,11 @@
 					"@executable_path/Frameworks",
 					"@loader_path/Frameworks",
 				);
+				LIBRARY_SEARCH_PATHS = (
+					"$(inherited)",
+					"$(PROJECT_DIR)/paddle-mobile/CPU",
+				);
+				MACH_O_TYPE = mh_dylib;
 				MTL_LANGUAGE_REVISION = UseDeploymentTarget;
 				PRODUCT_BUNDLE_IDENTIFIER = "orange.paddle-mobile";
 				PRODUCT_NAME = "$(TARGET_NAME:c99extidentifier)";

src/ios_io/PaddleMobile.h → metal/paddle-mobile/paddle-mobile/CPU/PaddleMobileCPU.h

@@ -17,7 +17,17 @@
 #import <CoreImage/CoreImage.h>
 #import <Foundation/Foundation.h>
 
-@interface PaddleMobile : NSObject
+@interface PaddleMobileCPUResult: NSObject
+
+@property (assign, nonatomic, readonly) float *output;
+
+@property (assign, nonatomic, readonly) int outputSize;
+
+-(void)releaseOutput;
+
+@end
+
+@interface PaddleMobileCPU : NSObject
 
 /*
     创建对象
@@ -34,13 +44,36 @@
 */
 - (BOOL)load:(NSString *)modelAndWeightPath;
 
+/*
+ * 从内存中加载模型
+ * */
+- (BOOL)LoadCombinedMemory:(size_t)modelLen
+               andModelBuf:(const uint8_t *)modelBuf
+         andModelParamsLen:(size_t)combinedParamsLen
+      andCombinedParamsBuf:(const uint8_t *)combinedParamsBuf;
+
+/*
+ *  对图像进行预处理, 需要外部开辟 output 内存, 外部释放 output 内存
+ * */
+-(void)preprocess:(CGImageRef)image
+           output:(float *)output
+            means:(NSArray<NSNumber *> *)means
+        scale:(float)scale
+        dim:(NSArray<NSNumber *> *)dim;
+
+/*
+ * 预测预处理后的数据, 返回结果使用结束需要调用其 realseOutput 函数进行释放
+ * */
+- (PaddleMobileCPUResult *)predictInput:(float *)input
+                                    dim:(NSArray<NSNumber *> *)dim;
+
 /*
     进行预测, means 和 scale 为训练模型时的预处理参数, 如训练时没有做这些预处理则直接使用 predict
 */
 - (NSArray *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim means:(NSArray<NSNumber *> *)means scale:(float)scale;
 
 /*
-    进行预测
+    进行预测, 默认 means 为 0, scale 为 1.0
 */
 - (NSArray *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim;
 

metal/paddle-mobile/paddle-mobile/CPUCompute.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#pragma once
+
+#import <Foundation/Foundation.h>
+
+
+@interface CPUResult: NSObject
+@property (assign, nonatomic) float *output;
+@property (assign, nonatomic) int outputSize;
+@end
+
+@interface NMSCompute: NSObject
+
+@property (assign, nonatomic) float scoreThredshold;
+
+@property (assign, nonatomic) int nmsTopK;
+
+@property (assign, nonatomic) int keepTopK;
+
+@property (assign, nonatomic) float nmsEta;
+
+@property (assign, nonatomic) float nmsThreshold;
+
+@property (assign, nonatomic) int background_label;
+
+@property (strong, nonatomic) NSArray<NSNumber *> *scoreDim;
+
+@property (strong, nonatomic) NSArray<NSNumber *> *bboxDim;
+
+-(CPUResult *)computeWithScore:(float *)score andBBoxs:(float *)bbox;
+
+@end

metal/paddle-mobile/paddle-mobile/CPUCompute.mm

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+
+#import "CPUCompute.h"
+
+#import <map>
+#import <vector>
+#import <utility>
+#import <algorithm>
+
+
+
+
+struct NMSParam {
+  
+  float *score_data;
+  
+  float *box_data;
+  
+  float *output;
+  
+  int output_size;
+  
+  std::vector<int> score_dim;
+  
+  std::vector<int> box_dim;
+  
+  float scoreThredshold;
+  
+  int nmsTopK;
+  
+  int keepTopK;
+  
+  float nmsEta;
+  
+  float nmsThreshold;
+  
+  int background_label;
+};
+
+
+constexpr int kOutputDim = 6;
+constexpr int kBBoxSize = 4;
+
+template <class T>
+bool SortScorePairDescend(const std::pair<float, T>& pair1,
+                          const std::pair<float, T>& pair2) {
+  return pair1.first > pair2.first;
+}
+
+template <class T>
+static inline void GetMaxScoreIndex(
+                                    const std::vector<T>& scores, const T threshold, int top_k,
+                                    std::vector<std::pair<T, int>>* sorted_indices) {
+  for (size_t i = 0; i < scores.size(); ++i) {
+    if (scores[i] > threshold) {
+      sorted_indices->push_back(std::make_pair(scores[i], i));
+    }
+  }
+  // Sort the score pair according to the scores in descending order
+  std::stable_sort(sorted_indices->begin(), sorted_indices->end(),
+                   SortScorePairDescend<int>);
+  // Keep top_k scores if needed.
+  if (top_k > -1 && top_k < static_cast<int>(sorted_indices->size())) {
+    sorted_indices->resize(top_k);
+  }
+}
+
+template <class T>
+static inline T BBoxArea(const T* box, const bool normalized) {
+  if (box[2] < box[0] || box[3] < box[1]) {
+    // If coordinate values are is invalid
+    // (e.g. xmax < xmin or ymax < ymin), return 0.
+    return static_cast<T>(0.);
+  } else {
+    const T w = box[2] - box[0];
+    const T h = box[3] - box[1];
+    if (normalized) {
+      return w * h;
+    } else {
+      // If coordinate values are not within range [0, 1].
+      return (w + 1) * (h + 1);
+    }
+  }
+}
+
+template <class T>
+static inline T JaccardOverlap(const T* box1, const T* box2,
+                               const bool normalized) {
+  if (box2[0] > box1[2] || box2[2] < box1[0] || box2[1] > box1[3] ||
+      box2[3] < box1[1]) {
+    return static_cast<T>(0.);
+  } else {
+    const T inter_xmin = std::max(box1[0], box2[0]);
+    const T inter_ymin = std::max(box1[1], box2[1]);
+    const T inter_xmax = std::min(box1[2], box2[2]);
+    const T inter_ymax = std::min(box1[3], box2[3]);
+    const T inter_w = inter_xmax - inter_xmin;
+    const T inter_h = inter_ymax - inter_ymin;
+    const T inter_area = inter_w * inter_h;
+    const T bbox1_area = BBoxArea<T>(box1, normalized);
+    const T bbox2_area = BBoxArea<T>(box2, normalized);
+    return inter_area / (bbox1_area + bbox2_area - inter_area);
+  }
+}
+
+template <typename T>
+static inline void NMSFast(
+                           const T *bbox_data,
+                           std::vector<int> bbox_dim,
+                           const T *score_data,
+                           const T score_threshold, const T nms_threshold,
+                           const T eta, const int top_k,
+                           std::vector<int>* selected_indices) {
+  // The total boxes for each instance.
+  int num_boxes = bbox_dim[0];
+  // 4: [xmin ymin xmax ymax]
+  int box_size = bbox_dim[1];
+
+  std::vector<T> scores_data(num_boxes);
+  std::copy_n(score_data, num_boxes, scores_data.begin());
+  std::vector<std::pair<T, int>> sorted_indices;
+  GetMaxScoreIndex(scores_data, score_threshold, top_k, &sorted_indices);
+
+  selected_indices->clear();
+  T adaptive_threshold = nms_threshold;
+
+  while (sorted_indices.size() != 0) {
+    const int idx = sorted_indices.front().second;
+    bool keep = true;
+    for (size_t k = 0; k < selected_indices->size(); ++k) {
+      if (keep) {
+        const int kept_idx = (*selected_indices)[k];
+        T overlap = JaccardOverlap<T>(bbox_data + idx * box_size,
+                                      bbox_data + kept_idx * box_size, true);
+        keep = overlap <= adaptive_threshold;
+      } else {
+        break;
+      }
+    }
+    if (keep) {
+      selected_indices->push_back(idx);
+    }
+    sorted_indices.erase(sorted_indices.begin());
+    if (keep && eta < 1 && adaptive_threshold > 0.5) {
+      adaptive_threshold *= eta;
+    }
+  }
+}
+
+template <typename T>
+void MultiClassNMS(const T *boxes_data,
+                   const std::vector<int> &box_dim,
+                   const T *scores_data,
+                   const std::vector<int> &score_dim,
+                   std::map<int, std::vector<int>>* indices, int* num_nmsed_out,
+                   const int& background_label, const int& nms_top_k,
+                   const int& keep_top_k, const T& nms_threshold,
+                   const T& nms_eta, const T& score_threshold) {
+  
+  int64_t class_num = score_dim[0];
+  int64_t predict_dim = score_dim[1];
+  int num_det = 0;
+  for (int c = 0; c < class_num; ++c) {
+    if (c == background_label) continue;
+    const T *score_data = scores_data + c * predict_dim;
+    
+    /// [c] is key
+    NMSFast<T>(boxes_data, box_dim, score_data, score_threshold, nms_threshold, nms_eta,
+                   nms_top_k, &((*indices)[c]));
+    num_det += (*indices)[c].size();
+  }
+
+  *num_nmsed_out = num_det;
+  if (keep_top_k > -1 && num_det > keep_top_k) {
+    std::vector<std::pair<T, std::pair<int, int>>> score_index_pairs;
+    for (const auto& it : *indices) {
+      int label = it.first;
+      const T* sdata = scores_data + label * predict_dim;
+      const std::vector<int>& label_indices = it.second;
+      for (size_t j = 0; j < label_indices.size(); ++j) {
+        int idx = label_indices[j];
+        // PADDLE_ENFORCE_LT(idx, predict_dim);
+        score_index_pairs.push_back(std::make_pair(sdata[idx], std::make_pair(label, idx)));
+      }
+    }
+    // Keep top k results per image.
+    std::stable_sort(score_index_pairs.begin(), score_index_pairs.end(),
+                     SortScorePairDescend<std::pair<int, int>>);
+    score_index_pairs.resize(keep_top_k);
+
+    // Store the new indices.
+    std::map<int, std::vector<int>> new_indices;
+    for (size_t j = 0; j < score_index_pairs.size(); ++j) {
+      int label = score_index_pairs[j].second.first;
+      int idx = score_index_pairs[j].second.second;
+      new_indices[label].push_back(idx);
+    }
+    new_indices.swap(*indices);
+    *num_nmsed_out = keep_top_k;
+  }
+}
+
+template <typename T>
+void MultiClassOutput(const T *scores_data,
+                      const std::vector<int> &score_dim,
+                      const T *bboxes_data,
+                      T *outputs_data,
+                      const std::map<int, std::vector<int>>& selected_indices) {
+  int predict_dim = score_dim[1];
+  int count = 0;
+  for (const auto& it : selected_indices) {
+    /// one batch
+    int label = it.first;
+    const T* sdata = scores_data + label * predict_dim;
+    const std::vector<int>& indices = it.second;
+    for (size_t j = 0; j < indices.size(); ++j) {
+      int idx = indices[j];
+      const T* bdata = bboxes_data + idx * kBBoxSize;
+      outputs_data[count * kOutputDim] = label;           // label
+      outputs_data[count * kOutputDim + 1] = sdata[idx];  // score
+      // xmin, ymin, xmax, ymax
+      std::memcpy(outputs_data + count * kOutputDim + 2, bdata, 4 * sizeof(T));
+      count++;
+    }
+  }
+}
+
+void MultiClassNMSCompute(NMSParam *param) {
+  assert(param->score_dim[0] == 1);
+  assert(param->box_dim[0] == 1);
+  assert (param->score_dim.size() == 3);
+  assert(param->box_dim.size() == 3);
+
+  float* outputs;
+  auto background_label = param->background_label;
+  auto nms_top_k = param->nmsTopK;
+  auto keep_top_k = param->keepTopK;
+  auto nms_threshold = param->nmsThreshold;
+  auto nms_eta = param->nmsEta;
+  auto score_threshold = param->scoreThredshold;
+
+  std::vector<int> score_dim_one_batch = {param->score_dim[1], param->score_dim[2]};
+  std::vector<int> box_dim_one_batch = {param->box_dim[1], param->box_dim[2]};
+  
+  std::vector<int> batch_starts = {0};
+  
+  std::map<int, std::vector<int>> indices;
+  int num_nmsed_out = 0;
+  
+  MultiClassNMS<float>(param->box_data, box_dim_one_batch, param->score_data, score_dim_one_batch, &indices, &num_nmsed_out,
+                       background_label, nms_top_k, keep_top_k, nms_threshold,
+                       nms_eta, score_threshold);
+  batch_starts.push_back(batch_starts.back() + num_nmsed_out);
+
+  int output_size = 0;
+  int num_kept = batch_starts.back();
+  if (num_kept == 0) {
+    outputs = new float[1];
+    outputs[0] = -1;
+    output_size = 1;
+  } else {
+    outputs = new float[num_kept * kOutputDim];
+    int64_t s = batch_starts[0];
+    int64_t e = batch_starts[1];
+    if (e > s) {
+      MultiClassOutput<float>(param->score_data, score_dim_one_batch, param->box_data, outputs, indices);
+    }
+    output_size = num_kept * kOutputDim;
+  }
+  param->output = outputs;
+  param->output_size = output_size;
+}
+
+@implementation CPUResult
+@end
+
+@implementation NMSCompute
+
+-(CPUResult *)computeWithScore:(float *)score andBBoxs:(float *)bbox {
+  NMSParam param;
+  param.box_data = bbox;
+  param.score_data = score;
+  param.background_label = self.background_label;
+  param.scoreThredshold = self.scoreThredshold;
+  param.nmsTopK = self.nmsTopK;
+  param.keepTopK = self.keepTopK;
+  param.nmsEta = self.nmsEta;
+  param.nmsThreshold = self.nmsThreshold;
+  std::vector<int> score_dim;
+  for (int i = 0; i < self.scoreDim.count; ++i) {
+    score_dim.push_back(self.scoreDim[i].intValue);
+  }
+  param.score_dim = score_dim;
+  
+  std::vector<int> box_dim;
+  for (int i = 0; i < self.bboxDim.count; ++i) {
+    box_dim.push_back(self.bboxDim[i].intValue);
+  }
+  param.box_dim = box_dim;
+  MultiClassNMSCompute(&param);
+  CPUResult *cr = [[CPUResult alloc] init];
+  cr.output = param.output;
+  cr.outputSize = param.output_size;
+  return cr;
+}
+
+@end
+
+

metal/paddle-mobile/paddle-mobile/Common/Extensions.swift

@@ -97,6 +97,21 @@ extension Array where Element: Comparable{
   }
 }
 
+extension Array {
+  public func strideArray(inCount: Int = 20) -> [(Int, Element)] {
+    if count < inCount {
+      return (0..<count).map{ ($0, self[$0]) }
+    } else {
+      let stride = count / inCount
+      var newArray: [(Int, Element)] = []
+      for i in 0..<inCount {
+        newArray.append((i * stride, self[i * stride]))
+      }
+      return newArray
+    }
+  }
+}
+
 extension String{
   func cStr() -> UnsafePointer<Int8>? {
     return (self as NSString).utf8String

metal/paddle-mobile/paddle-mobile/Common/MetalExtension.swift

@@ -56,8 +56,9 @@ extension MTLDevice {
     do {
       let pipLine = try makeComputePipelineState(function: function)
       return pipLine
-        } catch _ {
-            fatalError("make pip line error occured")
+    } catch let error {
+      print(error)
+      fatalError("make pip line error occured : \(error)")
     }
     
   }
@@ -71,6 +72,234 @@ extension MTLDevice {
     return buffer!
   }
   
+  func texture2tensor_loop<P>(texture: MTLTexture, cb: ([Int], P)->Void) -> Void {
+    let bpR = texture.width * 4 * MemoryLayout<P>.size
+    let bpI = texture.height * bpR
+    let region = MTLRegion.init(origin: MTLOrigin.init(x: 0, y: 0, z: 0), size: MTLSize.init(width: texture.width, height: texture.height, depth: 1))
+    for i in 0..<texture.arrayLength {
+      let pointer: UnsafeMutablePointer<P> = UnsafeMutablePointer<P>.allocate(capacity: bpI)
+      texture.getBytes(pointer, bytesPerRow: bpR, bytesPerImage: bpI, from: region, mipmapLevel: 0, slice: i)
+      for tx in 0..<texture.width * texture.height * 4 {
+        var k = tx
+        var xyzn: [Int] = [0, 0, 0, 0]
+        xyzn[1] = k / (texture.width * 4)
+        k %= (texture.width * 4)
+        xyzn[3] = k % 4
+        xyzn[0] = k / 4
+        xyzn[2] = i
+        cb(xyzn, pointer[tx])
+      }
+    }
+  }
+  
+  func texture2tensor_3<P>(texture: MTLTexture, dim: [Int],  transpose: [Int] = [0, 1, 2, 3]) -> [P] {
+    var tdim: [Int] = [1, 1, 1, 1]
+    for i in 0..<dim.count {
+      tdim[4 - dim.count + i] = dim[i]
+    }
+    let count = dim.reduce(1) { $0 * $1 }
+    var tensor: [P] = .init(repeating: Float32(0.0) as! P, count: count)
+    let ndim: [Int] = transpose.map { tdim[$0] }
+    assert(dim.count == 3)
+    assert(texture.width == ndim[3])
+    assert(texture.height == ndim[2])
+    assert(ndim[0] == 1)
+    assert(texture.arrayLength == (ndim[1] + 3) / 4)
+    texture2tensor_loop(texture: texture) { (xyzn: [Int], v: P) in
+      var tg: [Int] = [0, 0, 0, 0]
+      tg[1] = xyzn[2] * 4 + xyzn[3]
+      tg[2] = xyzn[1]
+      tg[3] = xyzn[0]
+      var ig: [Int] = [0, 0, 0, 0]
+      for k in 0..<4 {
+        ig[transpose[k]] = tg[k]
+      }
+      let ix = ig[0] * tdim[1] * tdim[2] * tdim[3] + ig[1] * tdim[2] * tdim[3] + ig[2] * tdim[3] + ig[3]
+      if ix < count {
+        tensor[ix] = v
+      }
+    }
+    return tensor
+  }
+  
+  func texture2tensor_2<P>(texture: MTLTexture, dim: [Int],  transpose: [Int] = [0, 1, 2, 3]) -> [P] {
+    var tdim: [Int] = [1, 1, 1, 1]
+    for i in 0..<dim.count {
+      tdim[4 - dim.count + i] = dim[i]
+    }
+    let count = dim.reduce(1) { $0 * $1 }
+    var tensor: [P] = .init(repeating: Float32(0.0) as! P, count: count)
+    let ndim: [Int] = transpose.map { tdim[$0] }
+    assert(dim.count == 2)
+    let w = (ndim[3] + 3) / 4
+    assert(texture.width == w)
+    assert(texture.height == ndim[2])
+    assert(ndim[0] == 1)
+    assert(ndim[1] == 1)
+    assert(texture.arrayLength == 1)
+    
+    texture2tensor_loop(texture: texture) { (xyzn: [Int], v: P) in
+      var tg: [Int] = [0, 0, 0, 0]
+      tg[2] = xyzn[1]
+      tg[3] = xyzn[0] * 4 + xyzn[3]
+      var ig: [Int] = [0, 0, 0, 0]
+      for k in 0..<4 {
+        ig[transpose[k]] = tg[k]
+      }
+      let ix = ig[0] * tdim[1] * tdim[2] * tdim[3] + ig[1] * tdim[2] * tdim[3] + ig[2] * tdim[3] + ig[3]
+      if ix < count {
+        tensor[ix] = v
+      }
+    }
+    return tensor
+  }
+  
+  func texture2tensor_1<P>(texture: MTLTexture, dim: [Int],  transpose: [Int] = [0, 1, 2, 3]) -> [P] {
+    var tdim: [Int] = [1, 1, 1, 1]
+    for i in 0..<dim.count {
+      tdim[4 - dim.count + i] = dim[i]
+    }
+    let count = dim.reduce(1) { $0 * $1 }
+    var tensor: [P] = .init(repeating: Float32(0.0) as! P, count: count)
+    let ndim: [Int] = transpose.map { tdim[$0] }
+    assert(dim.count == 1)
+    let w = (ndim[3] + 3) / 4
+    assert(texture.width == w)
+    assert(texture.height == 1)
+    assert(ndim[0] == 1)
+    assert(ndim[1] == 1)
+    assert(ndim[2] == 1)
+    assert(texture.arrayLength == 1)
+    
+    texture2tensor_loop(texture: texture) { (xyzn: [Int], v: P) in
+      var tg: [Int] = [0, 0, 0, 0]
+      tg[3] = xyzn[0] * 4 + xyzn[3]
+      var ig: [Int] = [0, 0, 0, 0]
+      for k in 0..<4 {
+        ig[transpose[k]] = tg[k]
+      }
+      let ix = ig[0] * tdim[1] * tdim[2] * tdim[3] + ig[1] * tdim[2] * tdim[3] + ig[2] * tdim[3] + ig[3]
+      if ix < count {
+        tensor[ix] = v
+      }
+    }
+    return tensor
+  }
+  
+  func texture2tensor<P>(texture: MTLTexture, dim: [Int], transpose: [Int] = [0, 1, 2, 3]) -> [P] {
+    if dim.count == 3 {
+      return texture2tensor_3(texture: texture, dim: dim, transpose: transpose)
+    } else if dim.count == 2 {
+      return texture2tensor_2(texture: texture, dim: dim, transpose: transpose)
+    } else if dim.count == 1 {
+      return texture2tensor_1(texture: texture, dim: dim, transpose: transpose)
+    }
+    var tdim: [Int] = [1, 1, 1, 1]
+    for i in 0..<dim.count {
+      tdim[4 - dim.count + i] = dim[i]
+    }
+    let count = dim.reduce(1) { $0 * $1 }
+    var tensor: [P] = .init(repeating: Float32(0.0) as! P, count: count)
+    let ndim: [Int] = transpose.map { tdim[$0] }
+    
+    assert(texture.width == ndim[2])
+    assert(texture.height == ndim[1])
+    assert(texture.arrayLength == (ndim[0] * ndim[3] + 3) / 4)
+    
+    texture2tensor_loop(texture: texture) { (xyzn: [Int], v: P) in
+      var tg: [Int] = [0, 0, 0, 0]
+      tg[1] = xyzn[1]
+      tg[2] = xyzn[0]
+      tg[0] = (xyzn[2] * 4 + xyzn[3]) / ndim[3]
+      tg[3] = (xyzn[2] * 4 + xyzn[3]) % ndim[3]
+      var ig: [Int] = [0, 0, 0, 0]
+      for k in 0..<4 {
+        ig[transpose[k]] = tg[k]
+      }
+      let ix = ig[0] * tdim[1] * tdim[2] * tdim[3] + ig[1] * tdim[2] * tdim[3] + ig[2] * tdim[3] + ig[3]
+      if ix < count {
+        tensor[ix] = v
+      }
+    }
+    return tensor
+  }
+  
+  func tensor2texture<P>(value: [P], dim: [Int], transpose: [Int] = [0, 1, 2, 3], inComputePrecision: ComputePrecision = .Float32) -> MTLTexture {
+    if value.count > 0 {
+      assert(value.count == dim.reduce(1) { $0 * $1 })
+    }
+    
+    var tdim: [Int] = [1, 1, 1, 1]
+    for i in 0..<dim.count {
+      tdim[4 - dim.count + i] = dim[i]
+    }
+    let ndim: [Int] = transpose.map { tdim[$0] }
+    
+    let textureDesc = MTLTextureDescriptor.init()
+    textureDesc.width = ndim[2]
+    textureDesc.height = ndim[1]
+    textureDesc.depth = 1
+    textureDesc.usage = [.shaderRead, .shaderWrite]
+    
+    if inComputePrecision == .Float16 {
+      textureDesc.pixelFormat = .rgba16Float
+    } else if inComputePrecision == .Float32 {
+      textureDesc.pixelFormat = .rgba32Float
+    }
+    
+    textureDesc.textureType = .type2DArray
+    textureDesc.storageMode = .shared
+    textureDesc.cpuCacheMode = .defaultCache
+    textureDesc.arrayLength = (ndim[0] * ndim[3] + 3) / 4
+    let texture = makeTexture(descriptor: textureDesc)!
+    
+    if value.count > 0 {
+      var rcount: Int = (ndim[0] * ndim[3] + 3) / 4
+      rcount = rcount * 4 * ndim[1] * ndim[2]
+      var nvalue: [Float32] = .init(repeating: 0.0, count: rcount)
+      
+      for i0 in 0..<tdim[0] {
+        for i1 in 0..<tdim[1] {
+          for i2 in 0..<tdim[2] {
+            for i3 in 0..<tdim[3] {
+              let ig = [i0, i1, i2, i3]
+              let ix = (i0 * tdim[1] * tdim[2] * tdim[3]) + (i1 * tdim[2] * tdim[3]) + (i2 * tdim[3]) + i3
+              
+              let jg = transpose.map { ig[$0] }
+              let k = jg[0] * ndim[3] + jg[3]
+              let jx = ((k / 4) * ndim[1] * ndim[2] * 4) + (jg[1] * ndim[2] * 4) + (jg[2] * 4) + (k % 4)
+              
+              nvalue[jx] = value[ix] as! Float32
+            }
+          }
+        }
+      }
+      
+      let region = MTLRegion.init(origin: MTLOrigin.init(x: 0, y: 0, z: 0), size: MTLSize.init(width: ndim[2], height: ndim[1], depth: 1))
+      if inComputePrecision == .Float16 {
+        let xvalue: [UInt16] = .init(repeating: 0, count: rcount)
+        let pointer: UnsafeMutablePointer<Float32> = UnsafeMutablePointer(mutating: nvalue)
+        let outputP: UnsafeMutablePointer<UInt16> = UnsafeMutablePointer(mutating: xvalue)
+        float32ToFloat16(input: pointer, output: outputP, count: rcount)
+        let bpR = ndim[2] * 4 * 2
+        let bpI = ndim[1] * bpR
+        for i in 0..<textureDesc.arrayLength {
+          let p = outputP + texture.width * texture.height * 4 * i
+          texture.replace(region: region, mipmapLevel: 0, slice: i, withBytes: p, bytesPerRow: bpR, bytesPerImage: bpI)
+        }
+      } else {
+        let pointer: UnsafeMutablePointer<Float32> = UnsafeMutablePointer(mutating: nvalue)
+        let bpR = ndim[2] * 4 * MemoryLayout<P>.size
+        let bpI = ndim[1] * bpR
+        for i in 0..<textureDesc.arrayLength {
+          let p = pointer + texture.width * texture.height * 4 * i
+          texture.replace(region: region, mipmapLevel: 0, slice: i, withBytes: p, bytesPerRow: bpR, bytesPerImage: bpI)
+        }
+      }
+    }
+    return texture
+  }
+  
   func makeFloatTexture<P>(value: [P], textureWidth: Int, textureHeight: Int, arrayLength: Int) -> MTLTexture{
     
     let textureDesc = MTLTextureDescriptor.init()
@@ -85,19 +314,25 @@ extension MTLDevice {
     textureDesc.arrayLength = arrayLength
     let texture = makeTexture(descriptor: textureDesc)!
     
-        if arrayLength == 1 && value.count >= 4{
-            let pointer: UnsafeMutablePointer<P> = UnsafeMutablePointer<P>.allocate(capacity: value.count * MemoryLayout<P>.size)
+    if value.count >= 4{
+      let counts = arrayLength * 4 * textureWidth * textureHeight
+      let pointer: UnsafeMutablePointer<P> = UnsafeMutablePointer<P>.allocate(capacity: counts * MemoryLayout<P>.size)
       for i in 0..<value.count {
         pointer[i] = value[i]
       }
+      for i in value.count..<counts {
+        pointer[i] = 0 as! P
+      }
       
       let bytesPerRow = texture.width * texture.depth * 4 * MemoryLayout<P>.size
+      let bytesPerImage = texture.height * bytesPerRow
       let region = MTLRegion.init(origin: MTLOrigin.init(x: 0, y: 0, z: 0), size: MTLSize.init(width: texture.width, height: texture.height, depth: texture.depth))
-            texture.replace(region: region, mipmapLevel: 0, withBytes: pointer, bytesPerRow: bytesPerRow)
+      for i in 0..<arrayLength {
+        let p = pointer + texture.width * texture.height * 4 * i
+        texture.replace(region: region, mipmapLevel: 0, slice: i, withBytes: p, bytesPerRow: bytesPerRow, bytesPerImage: bytesPerImage)
+      }
     } else {
       
-            
-            
     }
     
     return texture
@@ -105,7 +340,7 @@ extension MTLDevice {
 }
 
 extension MTLComputeCommandEncoder {
-    func dispatch(computePipline: MTLComputePipelineState, outTexture: MTLTexture) {
+  public func dispatch(computePipline: MTLComputePipelineState, outTexture: MTLTexture) {
     let slices = (outTexture.arrayLength * 4 + 3)/4
     
     let width = computePipline.threadExecutionWidth
@@ -120,16 +355,12 @@ extension MTLComputeCommandEncoder {
     let groupDepth = slices
     let groups = MTLSize.init(width: groupWidth, height: groupHeight, depth: groupDepth)
     
-//        print("groups: \(groups) ")
-//        print("threads per group: \(threadsPerGroup)")
-        
     setComputePipelineState(computePipline)
     
     dispatchThreadgroups(groups, threadsPerThreadgroup: threadsPerGroup)
   }
 }
 
-
 public extension MTLTexture {
   
   func stridableFloatArray<P>(stridable: Bool = true) -> [(index: Int, value: P)] {
@@ -180,60 +411,152 @@ public extension MTLTexture {
     return fArr
   }
   
+  func float32Array() -> [Float32] {
+    if pixelFormat == .rgba32Float {
+      let float32Array = floatArray { (f: Float32) -> Float32 in
+        return f
+      }
+      return float32Array
+    } else if pixelFormat == .rgba16Float {
+      
+      var float16Array = floatArray { (f: Float16) -> Float16 in
+        return f
+      }
+      return float16To32(input: &float16Array, count: float16Array.count)
+    } else {
+      fatalError()
+    }
+  }
+  
   func logDesc<T>(header: String = "", stridable: Bool = true) -> T? {
     print(header)
     print("texture: \(self)")
-        let res: [(index: Int, value: T)] = stridableFloatArray(stridable: stridable)
-        print(res)
-  
-//        if textureType == .type2DArray {
-//            for i in 0..<arrayLength{
-//                var str: String = "slice: \(i): \n"
-//                let bytes = UnsafeMutableRawPointer.allocate(byteCount: width * height * 4 * MemoryLayout<T>.size, alignment: MemoryLayout<T>.alignment)
-//                let bytesPerRow = width * depth * 4 * MemoryLayout<T>.size
-//                let bytesPerImage = width * height * depth * 4 * MemoryLayout<T>.size
-//                let region = MTLRegion.init(origin: MTLOrigin.init(x: 0, y: 0, z: 0), size: MTLSize.init(width: width, height: height, depth: depth))
-//                getBytes(bytes, bytesPerRow: bytesPerRow, bytesPerImage: bytesPerImage, from: region, mipmapLevel: 0, slice: i)
-//                let p = bytes.assumingMemoryBound(to: T.self)
-//                str += "2d array count : \(width * height * depth * 4) \n"
-//                if stridable && width * height * depth * 4 > 100 {
-//                    for j in stride(from: 0, to: width * height * depth * 4 , by: width * height * depth * 4 / 100){
-//                        str += " index \(j): \(p[j])"
-//                    }
-//                } else {
-//                    for j in 0..<width * height * depth * 4 {
-//                        str += " index \(j): \(p[j])"
-//                    }
-//                }
-//
-//                bytes.deallocate()
-//                print(str)
-//            }
-//        } else if textureType == .type2D {
-//            var str: String = "texture 2D: "
-//            let bytes = UnsafeMutableRawPointer.allocate(byteCount: width * height * 4 * MemoryLayout<T>.size, alignment: MemoryLayout<T>.alignment)
-//            let bytesPerRow = width * depth * 4 * MemoryLayout<T>.size
-//            let region = MTLRegion.init(origin: MTLOrigin.init(x: 0, y: 0, z: 0), size: MTLSize.init(width: width, height: height, depth: depth))
-//            getBytes(bytes, bytesPerRow: bytesPerRow, from: region, mipmapLevel: 0)
-//            let p = bytes.assumingMemoryBound(to: T.self)
-//            str += "2d count : \(width * width * 4) \n"
-//
-//            if stridable {
-//                for j in stride(from: 0, to: width * height * 4, by: width * height * 4 / 100){
-//                    str += "index \(j): \(p[j]) "
-//                }
-//            } else {
-//                for j in 0..<width * height * 4 {
-//                    str += "index \(j): \(p[j]) "
-//                }
-//            }
-//
-//            print(str)
-//            bytes.deallocate()
-//        }
+    //        let res: [(index: Int, value: T)] = stridableFloatArray(stridable: stridable)
+    //        print(res)
+    
+    if textureType == .type2DArray {
+      for i in 0..<arrayLength{
+        var str: String = "slice: \(i): \n"
+        let bytes = UnsafeMutableRawPointer.allocate(byteCount: width * height * 4 * MemoryLayout<T>.size, alignment: MemoryLayout<T>.alignment)
+        let bytesPerRow = width * depth * 4 * MemoryLayout<T>.size
+        let bytesPerImage = width * height * depth * 4 * MemoryLayout<T>.size
+        let region = MTLRegion.init(origin: MTLOrigin.init(x: 0, y: 0, z: 0), size: MTLSize.init(width: width, height: height, depth: depth))
+        getBytes(bytes, bytesPerRow: bytesPerRow, bytesPerImage: bytesPerImage, from: region, mipmapLevel: 0, slice: i)
+        let p = bytes.assumingMemoryBound(to: T.self)
+        str += "2d array count : \(width * height * depth * 4) \n"
+        if stridable && width * height * depth * 4 > 20 {
+          for j in stride(from: 0, to: width * height * depth * 4 , by: width * height * depth * 4 / 20){
+            str += " index \(j): \(p[j])"
+          }
+        } else {
+          for j in 0..<width * height * depth * 4 {
+            str += " index \(j): \(p[j])"
+          }
+        }
+        
+        bytes.deallocate()
+        print(str)
+      }
+    } else if textureType == .type2D {
+      var str: String = "texture 2D: "
+      let bytes = UnsafeMutableRawPointer.allocate(byteCount: width * height * 4 * MemoryLayout<T>.size, alignment: MemoryLayout<T>.alignment)
+      let bytesPerRow = width * depth * 4 * MemoryLayout<T>.size
+      let region = MTLRegion.init(origin: MTLOrigin.init(x: 0, y: 0, z: 0), size: MTLSize.init(width: width, height: height, depth: depth))
+      getBytes(bytes, bytesPerRow: bytesPerRow, from: region, mipmapLevel: 0)
+      let p = bytes.assumingMemoryBound(to: T.self)
+      str += "2d count : \(width * width * 4) \n"
+      
+      if stridable {
+        for j in stride(from: 0, to: width * height * 4, by: width * height * 4 / 20){
+          str += "index \(j): \(p[j]) "
+        }
+      } else {
+        for j in 0..<width * height * 4 {
+          str += "index \(j): \(p[j]) "
+        }
+      }
+      
+      print(str)
+      bytes.deallocate()
+    }
     return nil
     
   }
+  
+  // n c h w - dim
+  func toTensor(dim: (n: Int, c: Int, h: Int, w: Int)) -> [Float32] {
+    var textureArray: [Float32]
+    if pixelFormat == .rgba32Float {
+      textureArray = floatArray { (i : Float32) -> Float32 in
+        return i
+      }
+    } else if pixelFormat == .rgba16Float {
+      
+      var textureFloat16Array = floatArray { (i : Float16) -> Float16 in
+        return i
+      }
+      textureArray = float16To32(input: &textureFloat16Array, count: textureFloat16Array.count)
+    } else {
+      fatalError(" 目前还不支持其他类型 ")
+    }
+    
+    var output: [Float32] = []
+    for s in 0..<arrayLength {
+      for c in 0..<4{
+        for h in 0..<dim.h {
+          for w in 0..<dim.w {
+            if (s * 4 + c) < dim.c {
+              let textureValue = textureArray[dim.w * dim.h * 4 * s + h * dim.w * 4 + w * 4 + c]
+              output.append(textureValue)
+            }
+          }
+        }
+      }
+    }
+    return output
+  }
+  
+  func realNHWC(dim: (n: Int, h: Int, w: Int, c: Int)) -> [Float32] {
+//    print("origin dim: \(dim)")
+//    print("texture: ")
+//    print(self)
+    
+    var textureArray: [Float32]
+    if pixelFormat == .rgba32Float {
+      textureArray = floatArray { (i : Float32) -> Float32 in
+        return i
+      }
+    } else if pixelFormat == .rgba16Float {
+      var textureFloat16Array = floatArray { (i : Float16) -> Float16 in
+        return i
+      }
+      textureArray = float16To32(input: &textureFloat16Array, count: textureFloat16Array.count)
+    } else {
+      fatalError(" 目前还不支持其他类型 ")
+    }
+    
+    var output: [Float32] = []
+    let numOfASlice = dim.h * dim.w * 4
+    for h in 0..<dim.h {
+      for w in 0..<dim.w {
+        for sliceIndex in 0..<arrayLength {
+          if sliceIndex * 4 + 4 > dim.c {
+            for i in 0..<(4 - ((sliceIndex * 4 + 4) - dim.c)) {
+              let value = textureArray[sliceIndex * numOfASlice + h * dim.w * 4 + w * 4 + i]
+              output.append(value)
+            }
+          } else {
+            for i in 0..<4 {
+              let value = textureArray[sliceIndex * numOfASlice + h * dim.w * 4 + w * 4 + i]
+              output.append(value)
+            }
+          }
+        }
+      }
+    }
+    return output
+  }
+  
 }
 
 
@@ -270,11 +593,13 @@ public extension MTLBuffer {
     return texture
   }
   
-    
-
+  func array<T>() -> [T] {
+    var array: [T] = []
+    let pointer = contents().bindMemory(to: T.self, capacity: length)
+    for i in 0..<(length / MemoryLayout<T>.size) {
+      array.append(pointer[i])
+    }
+    return array;
+  }
 }
 
-
-
-
-

metal/paddle-mobile/paddle-mobile/Common/PaddleMobileUnitTest.swift

-//
-//  TestConvAddBatchNormRelu.swift
-//  paddle-mobile-demo
-//
-//  Created by liuRuiLong on 2018/7/25.
-//  Copyright © 2018年 orange. All rights reserved.
-//
+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
 
 import Metal
 import Foundation
@@ -17,6 +23,204 @@ public class PaddleMobileUnitTest {
         queue = inQueue
     }
     
+    private func indentPrintTensor(tensor: [Float32], dim: [Int], ix: [Int], indentLevel: Int) {
+        let indent = Array.init(repeating: " ", count: indentLevel).joined(separator: "")
+        var tx = ix
+        if dim.count == indentLevel + 1 {
+            var log: String = indent + "["
+            for i in 0..<dim[indentLevel] {
+                tx = ix
+                tx[indentLevel] = i
+                for x in 1..<dim.count {
+                    for y in 0..<x {
+                        tx[y] *= dim[x]
+                    }
+                }
+                let c = tx.reduce(0) { $0 + $1 }
+                if i > 0 {
+                    log += ", "
+                }
+                log += tensor[c].description
+            }
+            log += "]"
+            if (indentLevel > 0) && (ix[indentLevel - 1] < dim[indentLevel - 1] - 1) {
+                log += ","
+            }
+            print(log)
+        } else {
+            print(indent + "[")
+            for i in 0..<dim[indentLevel] {
+                tx[indentLevel] = i
+                indentPrintTensor(tensor: tensor, dim: dim, ix: tx, indentLevel: indentLevel + 1)
+            }
+            if (indentLevel > 0) && (ix[indentLevel - 1] < dim[indentLevel - 1] - 1) {
+                print(indent + "],")
+            } else {
+                print(indent + "]")
+            }
+        }
+    }
+    
+    private func tensorPrint(tensor: [Float32], dim: [Int]) {
+        var detectPos = -1
+        var odim = 1
+        var ndim = dim
+        for i in 0..<dim.count {
+            if dim[i] == -1 {
+                if detectPos == -1 {
+                    detectPos = i
+                } else {
+                    detectPos = -2
+                }
+            } else if dim[i] <= 0 {
+                detectPos = -3
+            } else {
+                odim *= dim[i]
+            }
+        }
+        assert(detectPos >= -1)
+        if (detectPos == -1) {
+            assert(tensor.count == odim)
+        } else {
+            assert(tensor.count % odim == 0)
+            ndim[detectPos] = tensor.count / odim
+        }
+        indentPrintTensor(tensor: tensor, dim: ndim, ix: dim.map { $0 * 0 }, indentLevel: 0)
+    }
+    
+    public func testConcat() {
+//        let buffer = queue.makeCommandBuffer() ?! "buffer is nil"
+//        var it: [[Float32]] = []
+//        for _ in 0..<7 {
+//            it.append((0..<12).map { Float32($0) })
+//        }
+//        let input = it.map { device.tensor2texture(value: $0, dim: [3, 4]) }
+//        let output = device.tensor2texture(value: [Float32](), dim: [3, 28])
+//
+//        let param = ConcatTestParam.init(
+//            input: input,
+//            output: output,
+//            dims: [[3, 4], [3, 4], [3, 4], [3, 4], [3, 4], [3, 4], [3, 4]],
+//            axis: 1,
+//            odim: [3, 28]
+//        )
+//        let concatKernel = ConcatKernel<Float32>.init(device: device, testParam: param)
+//        concatKernel.test(cmdBuffer: buffer, param: param)
+//        buffer.addCompletedHandler { (buffer) in
+//            for i in 0..<it.count {
+//                let _: Float32? = input[i].logDesc()
+//                self.tensorPrint(tensor: it[i], dim: [3, 4])
+//            }
+//            let _: Float32? = output.logDesc()
+//            let tx: [Float32] = self.device.texture2tensor(texture: output, dim: [3, 28])
+//            self.tensorPrint(tensor: tx, dim: [3, 28])
+//        }
+//
+//        buffer.commit()
+    }
+    
+    public func testReshape() {
+//        let buffer = queue.makeCommandBuffer() ?! "buffer is nil"
+//        let input: [Float32] = (0..<24).map { Float32($0) }
+//        let inTexture = device.tensor2texture(value: input, dim: [2, 3, 4])
+//        let outTexture = device.tensor2texture(value: [Float32](), dim: [4, 6])
+//        let mp = ReshapeMetalParam.init(
+//            idim: (1, 2, 3, 4),
+//            itrans: (0, 1, 2, 3),
+//            odim: (1, 1, 4, 6),
+//            otrans: (0, 1, 2, 3)
+//        )
+//        let param = ReshapeTestParam.init(
+//            inputTexture: inTexture,
+//            outputTexture: outTexture,
+//            param: mp
+//        )
+//        let reshapeKernel = ReshapeKernel<Float32>.init(device: device, testParam: param)
+//        reshapeKernel.test(commandBuffer: buffer, testParam: param)
+//        buffer.addCompletedHandler { (buffer) in
+//            let _: Float32? = inTexture.logDesc()
+//            let _: Float32? = outTexture.logDesc()
+//            self.tensorPrint(tensor: input, dim: [2, 3, 4])
+//            let tx: [Float32] = self.device.texture2tensor(texture: outTexture, dim: [4, 6])
+//            self.tensorPrint(tensor: tx, dim: [4, 6])
+//        }
+        
+//        let input: [Float32] = (0..<24).map { Float32($0) }
+//        let inTexture = device.tensor2texture(value: input, dim: [2, 3, 4])
+//        let outTexture = device.tensor2texture(value: [Float32](), dim: [24])
+//        let mp = ReshapeMetalParam.init(
+//            idim: (1, 2, 3, 4),
+//            itrans: (0, 1, 2, 3),
+//            odim: (1, 1, 1, 24),
+//            otrans: (0, 1, 2, 3)
+//        )
+//        let param = ReshapeTestParam.init(
+//            inputTexture: inTexture,
+//            outputTexture: outTexture,
+//            param: mp
+//        )
+//        let reshapeKernel = ReshapeKernel<Float32>.init(device: device, testParam: param)
+//        reshapeKernel.test(commandBuffer: buffer, testParam: param)
+//        buffer.addCompletedHandler { (buffer) in
+//            let _: Float32? = inTexture.logDesc()
+//            let _: Float32? = outTexture.logDesc()
+//            self.tensorPrint(tensor: input, dim: [2, 3, 4])
+//            let tx: [Float32] = self.device.texture2tensor(texture: outTexture, dim: [24])
+//            self.tensorPrint(tensor: tx, dim: [24])
+//        }
+//
+//        
+//        buffer.commit()
+    }
+    
+    public func testTranspose() {
+
+        let buffer = queue.makeCommandBuffer() ?! "buffer is nil"
+//        var input: [Float32] = []
+//        for i in 0..<72 {
+//            input.append(Float32(i))
+//        }
+////        let inputTexture = device.makeFloatTexture(value: input, textureWidth: 3, textureHeight: 2, arrayLength: 3)
+//        let inputTexture = device.tensor2texture(value: input, dim: [4, 3, 2, 3]);
+//        // group 1
+//        let outputTexture = device.tensor2texture(value: [Float32](), dim: [3, 3, 2, 4])
+//        let param = TransposeTestParam.init(inputTexture: inputTexture, outputTexture: outputTexture, iC: 3, oC: 4, axis: [3, 1, 2, 0])
+////        let param = TransposeTestParam.init(inputTexture: inputTexture, outputTexture: outputTexture, iC: 4, oC: 2, axis: [3, 0, 2, 1])
+////        // group 2
+////        let outputTexture = device.makeFloatTexture(value: [Float32](), textureWidth: 3, textureHeight: 3, arrayLength: 6)
+////        let param = TransposeTestParam.init(inputTexture: inputTexture, outputTexture: outputTexture, iC: 4, oC: 4, axis: [3, 0, 2, 1])
+////
+//        let transposeKernel = TransposeKernel<Float32>.init(device: device, testParam: param)
+//
+//        transposeKernel.test(commandBuffer: buffer, param: param)
+//
+//        buffer.addCompletedHandler { (buffer) in
+//            let _: Float32? = inputTexture.logDesc(header: "input texture", stridable: false)
+//            let _: Float32? = outputTexture.logDesc(header: "output texture", stridable: false)
+//            self.tensorPrint(tensor: input, dim: [4, 3, 2, 3])
+//            let tx: [Float32] = self.device.texture2tensor(texture: outputTexture, dim: [3, 3, 2, 4])
+//            self.tensorPrint(tensor: tx, dim: [3, 3, 2, 4])
+//        }
+//
+//        let input: [Float32] = (0..<24).map { Float32($0) }
+//        let inputTexture = device.tensor2texture(value: input, dim: [2, 3, 4])
+//        let outputTexture = device.tensor2texture(value: [Float](), dim: [3, 4, 2])
+//        let param = TransposeTestParam.init(inputTexture: inputTexture, outputTexture: outputTexture, iC: 4, oC: 2, axis: [0, 2, 3, 1])
+//        let transposeKernel = TransposeKernel<Float32>.init(device: device, testParam: param)
+//
+//        transposeKernel.test(commandBuffer: buffer, param: param)
+//
+//        buffer.addCompletedHandler { (buffer) in
+//            let _: Float32? = inputTexture.logDesc(header: "input texture", stridable: false)
+//            let _: Float32? = outputTexture.logDesc(header: "output texture", stridable: false)
+//            self.tensorPrint(tensor: input, dim: [2, 3, 4])
+//            let tx: [Float32] = self.device.texture2tensor(texture: outputTexture, dim: [3, 4, 2])
+//            self.tensorPrint(tensor: tx, dim: [3, 4, 2])
+//        }
+//        
+        buffer.commit()
+    }
+    
     public func testConvAddBnRelu() {
         let buffer = queue.makeCommandBuffer() ?! " buffer is nil "
         
@@ -116,7 +320,7 @@ public class PaddleMobileUnitTest {
         let offsetX = filterSize.width/2 - paddings.0
         let offsetY = filterSize.height/2 - paddings.1
         
-        let metalParam = MetalConvParam.init(offsetX: Int16(offsetX), offsetY: Int16(offsetY), offsetZ: 0, strideX: UInt16(stride.0), strideY: UInt16(stride.1), paddedZ: UInt16(paddings.0))
+        let metalParam = MetalConvParam.init(offsetX: Int16(offsetX), offsetY: Int16(offsetY), offsetZ: 0, strideX: UInt16(stride.0), strideY: UInt16(stride.1), dilationX: UInt16(1), dilationY: UInt16(1))
         
         let param = ConvAddBatchNormReluTestParam.init(inInputTexture: inputeTexture, inOutputTexture: outputTexture, inMetalParam: metalParam, inFilterBuffer: filterBuffer, inBiaseBuffer: biaseBuffer, inNewScaleBuffer: newScalueBuffer, inNewBiaseBuffer: newBiaseBuffer, inFilterSize: filterSize)
         
@@ -132,16 +336,6 @@ public class PaddleMobileUnitTest {
         }
         
         buffer.commit()
-        
-        
-//        let inputTexture = device.makeFloatTexture(value: <#T##[P]#>, textureWidth: <#T##Int#>, textureHeight: <#T##Int#>, arrayLength: <#T##Int#>)
-        
-        
-//        let param = ConvAddBatchNormReluTestParam.init(inInputTexture: <#T##MTLTexture#>, inOutputTexture: <#T##MTLTexture#>, inMetalParam: <#T##MetalConvParam#>, inFilterBuffer: <#T##MTLBuffer#>, inBiaseBuffer: <#T##MTLBuffer#>, inNewScaleBuffer: <#T##MTLBuffer#>, inNewBiaseBuffer: <#T##MTLBuffer#>, inFilterSize: <#T##(width: Int, height: Int, channel: Int)#>)
-        
-//        ConvAddBatchNormReluKernel.init(device: <#T##MTLDevice#>, testParam: <#T##ConvAddBatchNormReluTestParam#>)
-        
-        
     }
 }
 

metal/paddle-mobile/paddle-mobile/Common/Tools.swift

-//
-//  Tools.swift
-//  paddle-mobile
-//
-//  Created by liuRuiLong on 2018/7/26.
-//  Copyright © 2018年 orange. All rights reserved.
-//
+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
 
 import Foundation
 

metal/paddle-mobile/paddle-mobile/Common/Types.swift

@@ -13,6 +13,7 @@
  limitations under the License. */
 
 import Foundation
+import Accelerate
 
 public protocol SummableMultipliable: Equatable {
   static func +(lhs: Self, rhs: Self) -> Self
@@ -52,9 +53,6 @@ extension Float16: PrecisionType {
   public init(inFloat: Float32) {
     self = Int16(inFloat)
   }
-    
-    
-    
 }
 
 extension Float32: PrecisionType {
@@ -81,12 +79,162 @@ extension Float32: PrecisionType {
   }
 }
 
-public enum DataLayout {
-    case NCHW
-    case NHWC
+public func float32ToFloat16(input: UnsafeMutablePointer<Float32>, output: UnsafeMutableRawPointer, count: Int) {
+  var float32Buffer = vImage_Buffer(data: input,  height: 1, width: UInt(count), rowBytes: count * 4)
+  var float16buffer = vImage_Buffer(data: output, height: 1, width: UInt(count), rowBytes: count * 2)
+  guard vImageConvert_PlanarFtoPlanar16F(&float32Buffer, &float16buffer, 0) == kvImageNoError else {
+    fatalError(" float 32 to float 16 error ! ")
+  }
+}
+
+public func float16To32(input: UnsafeMutablePointer<Float16>, count: Int) -> [Float32] {
+  var output = Array<Float>.init(repeating: 0.0, count: count)
+  float16to32(input: input, output: &output, count: count)
+  return output
+}
+
+public func float16to32(input: UnsafeMutablePointer<Float16>, output: UnsafeMutablePointer<Float32>, count: Int) {
+  var bufferFloat16 = vImage_Buffer(data: input,  height: 1, width: UInt(count), rowBytes: count * 2)
+  var bufferFloat32 = vImage_Buffer(data: output, height: 1, width: UInt(count), rowBytes: count * 4)
+  if vImageConvert_Planar16FtoPlanarF(&bufferFloat16, &bufferFloat32, 0) != kvImageNoError {
+    fatalError(" convert float16 to float32 error")
+  }
+}
+
+// N - 0   C - 1   H - 2   W - 3
+struct DataLayout {
+  
+  static func NCHW(dim: Dim = Dim.init(inDim: [0, 0, 0, 0])) -> DataLayout {
+    return DataLayout.init([(.N, dim[0]), (.C, dim[1]), (.H, dim[2]), (.W, dim[3])])
+  }
+  
+  static func NHWC(dim: Dim = Dim.init(inDim: [0, 0, 0, 0])) -> DataLayout {
+    return DataLayout.init([(.N, dim[0]), (.H, dim[1]), (.W, dim[2]), (.C, dim[3])])
+  }
+  
+  func count() -> Int {
+    return layoutWithDim.count
+  }
+  
+  var N: Int? {
+    get {
+      for layoutDim in layoutWithDim {
+        if layoutDim.0 == .N {
+          return layoutDim.1
+        }
+      }
+      return nil
+    }
+    set {
+      var newN = (Layout.N, newValue)
+      if let index = layoutWithDim.index(where: { (layout: Layout, dim: Int) -> Bool in
+        return layout == .N
+      }) {
+        fatalError()
+      }
+    }
+  }
+  var C: Int? {
+    get {
+      for layoutDim in layoutWithDim {
+        if layoutDim.0 == .C {
+          return layoutDim.1
+        }
+      }
+      return nil
+    }
+    set {
+      var newN = (Layout.C, newValue)
+      if let index = layoutWithDim.index(where: { (layout: Layout, dim: Int) -> Bool in
+        return layout == .N
+      }) {
+        fatalError()
+      }
+    }
+  }
+  var H: Int? {
+    get {
+      for layoutDim in layoutWithDim {
+        if layoutDim.0 == .H {
+          return layoutDim.1
+        }
+      }
+      return nil
+    }
+    set {
+      var newN = (Layout.H, newValue)
+      if let index = layoutWithDim.index(where: { (layout: Layout, dim: Int) -> Bool in
+        return layout == .H
+      }) {
+        fatalError()
+      }
+    }
+  }
+  var W: Int? {
+    get {
+      for layoutDim in layoutWithDim {
+        if layoutDim.0 == .W {
+          return layoutDim.1
+        }
+      }
+      return nil
+    }
+    set {
+      var newN = (Layout.W, newValue)
+      if let index = layoutWithDim.index(where: { (layout: Layout, dim: Int) -> Bool in
+        return layout == .W
+      }) {
+        fatalError()
+      }
+    }
+  }
+  
+  
+  init(_ inLayout: [(Layout, Int)]) {
+    layoutWithDim = inLayout
+  }
+  
+  func layout() -> [Layout] {
+    return layoutWithDim.map({ (layout: Layout, dim: Int) -> Layout in
+      return layout
+    })
+  }
+  
+  var layoutWithDim: [(Layout, Int)] = [(.N, 0), (.C, 0), (.H, 0), (.W, 0)]
+  
+  func convertTo(inLayout: [Layout]) {
+    
+  }
+  
+  enum Layout: Int{
+    case N = 0
+    case C = 1
+    case H = 2
+    case W = 3
+    static func defaultLayout() -> [Layout] {
+      return [N, C, H, W]
+    }
+  }
+}
+
+extension DataLayout: Equatable {
+  public static func == (lhs: DataLayout, rhs: DataLayout) -> Bool {
+    if lhs.layoutWithDim.count == rhs.layoutWithDim.count {
+      var result = true
+      for i in 0..<lhs.layoutWithDim.count {
+        result = (lhs.layoutWithDim[i].0 == rhs.layoutWithDim[i].0)
+        if !result {
+          break
+        }
+      }
+      return result
+    } else {
+      return false
+    }
+  }
 }
 
-protocol Variant: CustomStringConvertible, CustomDebugStringConvertible {
+public protocol Variant: CustomStringConvertible, CustomDebugStringConvertible {
 }
 
 extension Tensor: Variant {
@@ -95,7 +243,7 @@ extension Tensor: Variant {
 extension Texture: Variant {
 }
 
-extension ResultHolder: Variant {
+extension GPUResultHolder: Variant {
 }
 
 extension InputTexture: Variant {
@@ -104,3 +252,43 @@ extension InputTexture: Variant {
 extension MTLTexture where Self: Variant {
   
 }
+
+class FetchHolder: Variant {
+  var resultBuffer: MTLBuffer?
+  var dim: [Int]
+  var capacity: Int
+  
+  init(inCapacity: Int, inDim: [Int]) {
+    capacity = inCapacity
+    dim = inDim
+  }
+  
+  func initBuffer(device: MTLDevice) {
+    resultBuffer = device.makeBuffer(length: capacity * 4, options: [])
+  }
+  
+  var result: UnsafeMutablePointer<Float32> {
+    guard let inResultBuffer = resultBuffer else {
+      fatalError()
+    }
+    return inResultBuffer.contents().bindMemory(to: Float32.self, capacity: capacity)
+  }
+  
+}
+
+extension FetchHolder: CustomStringConvertible, CustomDebugStringConvertible {
+  var description: String {
+    fatalError()
+//    return "\(result)"
+  }
+  
+  var debugDescription: String {
+    fatalError()
+//    return "\(result)"
+  }
+  
+  
+}
+
+
+

metal/paddle-mobile/paddle-mobile/Executor.swift

-/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
- 
- Licensed under the Apache License, Version 2.0 (the "License");
- you may not use this file except in compliance with the License.
- You may obtain a copy of the License at
- 
- http://www.apache.org/licenses/LICENSE-2.0
- 
- Unless required by applicable law or agreed to in writing, software
- distributed under the License is distributed on an "AS IS" BASIS,
- WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- See the License for the specific language governing permissions and
- limitations under the License. */
-
-import Foundation
-
-public class ResultHolder<P: PrecisionType> {
-    public let dim: [Int]
-    public let resultArr: [P]
-    public let elapsedTime: Double
-    public init(inDim: [Int], inResult: [P], inElapsedTime: Double) {
-        dim = inDim
-        resultArr = inResult
-        elapsedTime = inElapsedTime
-    }
-}
-
-extension ResultHolder: CustomDebugStringConvertible, CustomStringConvertible {
-    public var debugDescription: String {
-        var str = ""
-        str += "Dim: \(dim) \n value:[ "
-        if resultArr.count < 20 {
-            for d in resultArr {
-                str += " \(d) "
-            }
-        } else {
-            for d in stride(from: 0, to: resultArr.count, by: resultArr.count/20) {
-                str += " \(resultArr[d]) "
-            }
-        }
-        str += " ]"
-        return str
-    }
-    
-    public var description: String {
-        return debugDescription
-    }
-}
-
-public class Executor<P: PrecisionType> {
-    var ops: [Runable & InferShaperable] = []
-    let program: Program
-    let device: MTLDevice
-    let queue: MTLCommandQueue
-    public init(inDevice:MTLDevice, inQueue: MTLCommandQueue, inProgram: Program) throws {
-        program = inProgram
-        device = inDevice
-        queue = inQueue
-        for block in inProgram.programDesc.blocks {
-            //block.ops.count
-            for i in 0..<block.ops.count {
-                let op = block.ops[i]
-                do {
-                    let op = try OpCreator<P>.shared.creat(device: inDevice, opDesc: op, scope: inProgram.scope)
-                    op.inferShape()
-                    ops.append(op)
-                } catch let error {
-                    throw error
-                }
-            }
-            
-//            for op in block.ops {
-//                do {
-//                    let op = try OpCreator<P>.shared.creat(device: inDevice, opDesc: op, scope: inProgram.scope)
-//                    op.inferShape()
-//                    ops.append(op)
-//                } catch let error {
-//                    throw error
-//                }
-//            }
-        }
-    }
-    
-    public func predict(input: MTLTexture, expect: [Int], completionHandle: @escaping (ResultHolder<P>) -> Void, preProcessKernle: CusomKernel? = nil) throws {
-        guard let buffer = queue.makeCommandBuffer() else {
-            throw PaddleMobileError.predictError(message: "CommandBuffer is nil")
-        }
-        let resInput: MTLTexture
-        if let inPre = preProcessKernle {
-            do {
-                try inPre.compute(inputTexuture: input, commandBuffer: buffer)
-                resInput = inPre.outputTexture
-            } catch let error {
-                throw error
-            }
-        } else {
-            resInput = input
-        }
-        
-        let beforeDate = Date.init()
-        let inputTexture = InputTexture.init(inMTLTexture: resInput, inExpectDim: Dim.init(inDim: expect))
-        program.scope.setInput(input: inputTexture)
- 
-        for op in ops {
-            do {
-                try op.run(device: device, buffer: buffer)
-            } catch let error {
-                throw error
-            }
-        }
-        
-        buffer.addCompletedHandler { (commandbuffer) in
-//            let inputArr = resInput.floatArray(res: { (p:P) -> P in
-//                return p
-//            })
-//            print(inputArr)
-            
-//            let stridableInput: [(index: Int, value: Float)] = input.stridableFloatArray()
-//            print(stridableInput)
-            
-//            let _: Flo? = input.logDesc(header: "input: ", stridable: true)
-//            for op in self.ops {
-//                op.delogOutput()
-//            }
-//            return
-            
-//            self.ops[2].delogOutput()
-            
-            
-            let afterDate = Date.init()
-            
-            guard let outputVar = self.program.scope.output() else {
-                fatalError("output nil")
-            }
-
-            guard let output = outputVar as? Texture<P> else {
-                fatalError("output var type error")
-            }
-            let resultHodlder = ResultHolder<P>.init(inDim: output.dim.dims, inResult: output.metalTexture.floatArray(res: { (p:P) -> P in
-                return p
-            }), inElapsedTime: afterDate.timeIntervalSince(beforeDate))
-            completionHandle(resultHodlder)
-        }
-        buffer.commit()
-    }
-    
-    public func clear() {
-        program.scope.clear()
-    }
-    
-}
-
-//public let paddle_executor: Executor = Executor.init()

metal/paddle-mobile/paddle-mobile/Genet.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+public class Genet: Net {
+  @objc public override init(device: MTLDevice) {
+    super.init(device: device)
+    means = [128.0, 128.0, 128.0]
+    scale = 0.017
+    except = 0
+    modelPath = Bundle.main.path(forResource: "genet_model", ofType: nil) ?! "model null"
+    paramPath = Bundle.main.path(forResource: "genet_params", ofType: nil) ?! "para null"
+    modelDir = ""
+    preprocessKernel = GenetPreProccess.init(device: device)
+    dim = (n: 1, h: 128, w: 128, c: 3)
+  }
+  
+  @objc override public init(device: MTLDevice,paramPointer: UnsafeMutableRawPointer, paramSize:Int, modePointer: UnsafeMutableRawPointer, modelSize: Int) {
+    super.init(device:device,paramPointer:paramPointer,paramSize:paramSize,modePointer:modePointer,modelSize:modelSize)
+    means = [128.0, 128.0, 128.0]
+    scale = 0.017
+    except = 0
+    modelPath = ""
+    paramPath = ""
+    modelDir = ""
+    preprocessKernel = GenetPreProccess.init(device: device)
+    dim = (n: 1, h: 128, w: 128, c: 3)
+  }
+
+  class GenetPreProccess: CusomKernel {
+    init(device: MTLDevice) {
+      let s = CusomKernel.Shape.init(inWidth: 128, inHeight: 128, inChannel: 3)
+      super.init(device: device, inFunctionName: "genet_preprocess", outputDim: s, usePaddleMobileLib: false)
+    }
+  }
+  
+  override  public func resultStr(res: ResultHolder) -> String {
+//    fatalError()
+    return " \(res.result![0]) ... "
+  }
+  
+}

metal/paddle-mobile/paddle-mobile/Loader.swift

-/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
- 
- Licensed under the Apache License, Version 2.0 (the "License");
- you may not use this file except in compliance with the License.
- You may obtain a copy of the License at
- 
- http://www.apache.org/licenses/LICENSE-2.0
- 
- Unless required by applicable law or agreed to in writing, software
- distributed under the License is distributed on an "AS IS" BASIS,
- WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- See the License for the specific language governing permissions and
- limitations under the License. */
-
-import Foundation
-import SwiftProtobuf
-
-public class Loader<P: PrecisionType> {
-    class ParaLoader {
-        let file: UnsafeMutablePointer<FILE>
-        let fileSize: Int
-        var nowIndex: Int
-        init(paramPath: String) throws {
-            guard let tmpFile = fopen(paramPath, "rb") else {
-                throw PaddleMobileError.loaderError(message: "open param file error" + paramPath)
-            }
-            file = tmpFile
-            fseek(file, 0, SEEK_END)
-            fileSize = ftell(file)
-            guard fileSize > 0 else {
-                throw PaddleMobileError.loaderError(message: "param file size is too small")
-            }
-            rewind(file)
-            nowIndex = 0
-        }
-        
-        func read(tensor: Tensor<P>) throws {
-            guard nowIndex <= fileSize else {
-                throw PaddleMobileError.loaderError(message: "out of the file range")
-            }
-            
-            func pointerReader<T>(type: T.Type) -> T {
-                let ptr = UnsafeMutablePointer<T>.allocate(capacity: MemoryLayout<T>.size)
-                fread(ptr, 1, MemoryLayout<T>.size, file)
-                nowIndex += MemoryLayout<T>.size
-                let pointee = ptr.pointee
-                ptr.deinitialize(count: MemoryLayout<UInt32>.size)
-                ptr.deallocate()
-                return pointee
-            }
-            
-            let _ = pointerReader(type: UInt32.self)
-            let lodLevel = pointerReader(type: UInt64.self)
-            for _ in 0..<lodLevel {
-                let size = pointerReader(type: UInt64.self)
-                for _ in 0..<Int(size/UInt64(MemoryLayout<size_t>.size)){
-                    _ = pointerReader(type: size_t.self)
-                }
-            }
-            
-            let _ = pointerReader(type: UInt32.self)
-            
-            let tensorDescSize = pointerReader(type: Int32.self)
-            
-            fseek(file, Int(tensorDescSize), SEEK_CUR)
-            nowIndex += Int(tensorDescSize)
-            
-            /*
-             这里没有根据 Data Type 去判断, 而是从外部泛型直接指定了精度
-             */
-            
-            //现在模型传入模型为  Float 类型, 这块应该根据模型来
-//            let tmpCapacity = MemoryLayout<Float>.size * tensor.numel()
-//            let tmpPointer = UnsafeMutablePointer<Float>.allocate(capacity: tmpCapacity);
-            let bytesRead = fread(tensor.data.pointer, 1, tensor.data.size, file)
-            
-            guard bytesRead == tensor.data.size else {
-                throw PaddleMobileError.loaderError(message: "param read size error")
-            }
-            
-            // TODO: use script to convert
-//            let bytesRead = fread(tmpPointer, 1, tmpCapacity, file)
-//            for i in 0..<tensor.numel() {
-//                tensor.data[i] = P.init(inFloat: tmpPointer[i])
-//            }
-//            tmpPointer.deinitialize(count: tmpCapacity)
-//            tmpPointer.deallocate()
-            
-            nowIndex += bytesRead
-        }
-        
-        deinit {
-            fclose(file)
-        }
-    }
-    public init(){}
-    public func load(device: MTLDevice, modelPath: String, paraPath: String) throws -> Program{
-        guard let modelData = try? Data.init(contentsOf: URL.init(fileURLWithPath: modelPath)) else {
-            throw PaddleMobileError.loaderError(message: "load " + modelPath + " failed !")
-        }
-        
-        do {
-            let protoProgram = try PaddleMobile_Framework_Proto_ProgramDesc.init(
-                serializedData: modelData)
-            
-            let originProgramDesc = ProgramDesc.init(protoProgram: protoProgram)
-            let programDesc = ProgramOptimize<P>.init().optimize(originProgramDesc: originProgramDesc)
-            print(programDesc)
-
-            guard let paraLoader = try? ParaLoader.init(paramPath: paraPath) else {
-                throw PaddleMobileError.loaderError(message: "load para error")
-            }
-            
-            guard programDesc.blocks.count > 0 else {
-                throw PaddleMobileError.loaderError(message: "count of blocks must greater than 0")
-            }
-            
-            // to get feed key and fetch key
-            let block = programDesc.blocks[0]
-            guard let firstOp = block.ops.first, let lastOp = block.ops.last else {
-                throw PaddleMobileError.loaderError(message: "at least two operator")
-            }
-            guard firstOp.type == gFeedType, lastOp.type == gFetchType else {
-                throw PaddleMobileError.loaderError(message: "the first op is not feed or the last op is not fetch")
-            }
-            
-            guard let inputKey = opInfos[gFeedType]?.inputs.first, let outKey = opInfos[gFetchType]?.outputs.first else {
-                throw PaddleMobileError.loaderError(message: "the feed input key or fetch output key not found")
-            }
-            guard let feedKey = firstOp.inputs[inputKey]?.first, let fetchKey = lastOp.outputs[outKey]?.first else {
-                throw PaddleMobileError.loaderError(message: "feed key or fetch key not found")
-            }
-            
-            let scope = Scope.init(inFeedKey: feedKey, inFetchKey: fetchKey)
-            
-            // to load memory
-            for block in programDesc.blocks {
-                for varDesc in block.vars {
-                    if (varDesc.type == .LodTensor) {
-                        guard let tensorDesc = varDesc.tensorDesc else {
-                            throw PaddleMobileError.loaderError(message: "get tensor desc failed")
-                        }
-                        
-//                        guard (try? tensorDesc.dataType.dataTypeSize()) == MemoryLayout<P>.size else {
-//                            throw PaddleMobileError.memoryError(message: "PrecisionType not support")
-//                        }
-                        
-                        if (varDesc.persistable
-                            && varDesc.type != .FeedMiniBatch
-                            && varDesc.type != .FetchList) {
-                            let dimArr = tensorDesc.dims
-                            
-                            guard dimArr.count > 0 else {
-                                throw PaddleMobileError.loaderError(message: "tensor desc dim size error")
-                            }
-                            
-                            let dim = Dim.init(inDim: dimArr)
-                            let tensor = Tensor<P>.init(inDim: dim, inLayout: tensorDesc.dataLayout)
-                            do {
-                                try paraLoader.read(tensor: tensor)
-                            } catch let error {
-                                throw error
-                            }
-                            tensor.convert(to: .NHWC)
-//                            tensor.initBuffer(device: device)
-                            scope[varDesc.name] = tensor
-                        } else {
-                            let dim = Dim.init(inDim: tensorDesc.NHWCDim)
-                            scope[varDesc.name] = Texture<P>.init(device: device, inDim: dim)
-                        }
-                    } else {
-                        if varDesc.name == fetchKey {
-                            scope[varDesc.name] = ResultHolder<P>.init(inDim: [], inResult: [], inElapsedTime: 0.0)
-                        } else if varDesc.name == feedKey {
-                        }
-                    }
-                }
-            }
-            
-            let program = Program.init(inProgramDesc: programDesc, inParamPath: paraPath, inScope: scope)
-            
-            return program
-        } catch _ {
-            throw PaddleMobileError.loaderError(message: "protobuf decoder error")
-        }
-    }
-}

metal/paddle-mobile/paddle-mobile/MobileNet.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class MobileNet: Net{
+  
+  class MobilenetPreProccess: CusomKernel {
+    init(device: MTLDevice) {
+      let s = CusomKernel.Shape.init(inWidth: 224, inHeight: 224, inChannel: 3)
+      super.init(device: device, inFunctionName: "mobilenet_preprocess", outputDim: s, usePaddleMobileLib: false)
+    }
+  }
+  
+  class PreWords {
+    var contents: [String] = []
+    init(fileName: String, type: String = "txt", inBundle: Bundle = Bundle.main) {
+      if let filePath = inBundle.path(forResource: fileName, ofType: type) {
+        let string = try! String.init(contentsOfFile: filePath)
+        contents = string.components(separatedBy: CharacterSet.newlines).filter{$0.count > 10}.map{
+          String($0[$0.index($0.startIndex, offsetBy: 10)...])
+        }
+      }else{
+        fatalError("no file call \(fileName)")
+      }
+    }
+    subscript(index: Int) -> String {
+      return contents[index]
+    }
+  }
+  
+  let labels = PreWords.init(fileName: "synset")
+  
+  override public func resultStr(res: ResultHolder) -> String {
+    guard let resPointer = res.result else {
+      fatalError()
+    }
+    var s: [String] = []
+    (0..<res.capacity).map { resPointer[$0] }.top(r: 5).enumerated().forEach{
+      s.append(String(format: "%d: %@ (%3.2f%%)", $0 + 1, labels[$1.0], $1.1 * 100))
+    }
+    return s.joined(separator: "\n")
+  }
+  
+
+  
+  override init(device: MTLDevice) {
+    super.init(device: device)
+    means = [123.68, 116.78, 103.94]
+    scale = 0.017
+    except = 0
+    modelPath = Bundle.main.path(forResource: "model", ofType: nil) ?! "model null"
+    paramPath = Bundle.main.path(forResource: "params", ofType: nil) ?! "para null"
+    modelDir = ""
+    preprocessKernel = MobilenetPreProccess.init(device: device)
+    dim = (n: 1, h: 224, w: 224, c: 3)
+  }
+}
+

metal/paddle-mobile/paddle-mobile/MobileNetSSD.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+public class MobileNet_ssd_hand: Net{
+  @objc public override init(device: MTLDevice) {
+    super.init(device: device)
+    means = [123.68, 116.78, 103.94]
+    scale = 0.017
+    except = 2
+    modelPath = Bundle.main.path(forResource: "ssd_hand_model", ofType: nil) ?! "model null"
+    paramPath = Bundle.main.path(forResource: "ssd_hand_params", ofType: nil) ?! "para null"
+    modelDir = ""
+    preprocessKernel = MobilenetssdPreProccess.init(device: device)
+    dim = (n: 1, h: 300, w: 300, c: 3)
+  }
+  
+  @objc override public init(device: MTLDevice,paramPointer: UnsafeMutableRawPointer, paramSize:Int, modePointer: UnsafeMutableRawPointer, modelSize: Int) {
+    super.init(device:device,paramPointer:paramPointer,paramSize:paramSize,modePointer:modePointer,modelSize:modelSize)
+    means = [123.68, 116.78, 103.94]
+    scale = 0.017
+    except = 2
+    modelPath = ""
+    paramPath = ""
+    modelDir = ""
+    preprocessKernel = MobilenetssdPreProccess.init(device: device)
+    dim = (n: 1, h: 300, w: 300, c: 3)
+  }
+  
+  class MobilenetssdPreProccess: CusomKernel {
+    init(device: MTLDevice) {
+      let s = CusomKernel.Shape.init(inWidth: 300, inHeight: 300, inChannel: 3)
+      super.init(device: device, inFunctionName: "mobilenet_ssd_preprocess", outputDim: s, usePaddleMobileLib: false)
+    }
+  }
+  
+  override public func resultStr(res: ResultHolder) -> String {
+    return " \(res)"
+  }
+  
+  override func fetchResult(paddleMobileRes: GPUResultHolder) -> ResultHolder {
+
+//    guard let interRes = paddleMobileRes.intermediateResults else {
+//      fatalError(" need have inter result ")
+//    }
+//
+//    guard let scores = interRes["Scores"], scores.count > 0, let score = scores[0] as?  Texture<Float32> else {
+//      fatalError(" need score ")
+//    }
+//
+//    guard let bboxs = interRes["BBoxes"], bboxs.count > 0, let bbox = bboxs[0] as? Texture<Float32> else {
+//      fatalError()
+//    }
+//
+//    var scoreFormatArr: [Float32] = score.metalTexture.realNHWC(dim: (n: score.padToFourDim[0], h: score.padToFourDim[1], w: score.padToFourDim[2], c: score.padToFourDim[3]))
+////    print("score: ")
+////    print(scoreFormatArr.strideArray())
+////
+//    var bboxArr = bbox.metalTexture.float32Array()
+////    print("bbox: ")
+////    print(bboxArr.strideArray())
+//
+//    let nmsCompute = NMSCompute.init()
+//    nmsCompute.scoreThredshold = 0.01
+//    nmsCompute.nmsTopK = 400
+//    nmsCompute.keepTopK = 200
+//    nmsCompute.nmsEta = 1.0
+//    nmsCompute.nmsThreshold = 0.45
+//    nmsCompute.background_label = 0;
+//
+//    nmsCompute.scoreDim = [NSNumber.init(value: score.tensorDim[0]), NSNumber.init(value: score.tensorDim[1]), NSNumber.init(value: score.tensorDim[2])]
+//
+//    nmsCompute.bboxDim = [NSNumber.init(value: bbox.tensorDim[0]), NSNumber.init(value: bbox.tensorDim[1]), NSNumber.init(value: bbox.tensorDim[2])]
+//    guard let result = nmsCompute.compute(withScore: &scoreFormatArr, andBBoxs: &bboxArr) else {
+//      fatalError( " result error " )
+//    }
+//
+//    let output: [Float32] = result.map { $0.floatValue }
+//
+//
+//    return output
+    fatalError()
+  }
+  
+
+  
+ 
+}

metal/paddle-mobile/paddle-mobile/MobilenetSSD_AR.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+public class MobileNet_ssd_AR: Net{
+  @objc public override init(device: MTLDevice) {
+    super.init(device: device)
+    means = [103.94, 116.78, 123.68]
+    scale = 1
+    except = 2
+    modelPath = Bundle.main.path(forResource: "ar_model", ofType: nil) ?! "model null"
+    paramPath = Bundle.main.path(forResource: "ar_params", ofType: nil) ?! "para null"
+    modelDir = ""
+    preprocessKernel = MobilenetssdPreProccess.init(device: device)
+    dim = (n: 1, h: 160, w: 160, c: 3)
+  }
+  
+  @objc override public init(device: MTLDevice,paramPointer: UnsafeMutableRawPointer, paramSize:Int, modePointer: UnsafeMutableRawPointer, modelSize: Int) {
+    super.init(device:device,paramPointer:paramPointer,paramSize:paramSize,modePointer:modePointer,modelSize:modelSize)
+    means = [103.94, 116.78, 123.68]
+    scale = 1
+    except = 2
+    modelPath = ""
+    paramPath = ""
+    modelDir = ""
+    preprocessKernel = MobilenetssdPreProccess.init(device: device)
+    dim = (n: 1, h: 160, w: 160, c: 3)
+  }
+  
+  class MobilenetssdPreProccess: CusomKernel {
+    init(device: MTLDevice) {
+      let s = CusomKernel.Shape.init(inWidth: 160, inHeight: 160, inChannel: 3)
+      super.init(device: device, inFunctionName: "mobilent_ar_preprocess", outputDim: s, usePaddleMobileLib: false)
+    }
+  }
+  
+  override public func resultStr(res: ResultHolder) -> String {
+    return " \(res.result![0])"
+  }
+  
+  override func fetchResult(paddleMobileRes: GPUResultHolder) -> ResultHolder {
+    guard let interRes = paddleMobileRes.intermediateResults else {
+      fatalError(" need have inter result ")
+    }
+    
+    guard let scores = interRes["Scores"], scores.count > 0, let score = scores[0] as?  FetchHolder else {
+      fatalError(" need score ")
+    }
+    
+    guard let bboxs = interRes["BBoxes"], bboxs.count > 0, let bbox = bboxs[0] as? FetchHolder else {
+      fatalError()
+    }
+    
+//    let startDate = Date.init()
+    
+//    print("scoreFormatArr: ")
+//print((0..<score.capacity).map{ score.result[$0] }.strideArray())
+//
+//    print("bbox arr: ")
+//
+//    print((0..<bbox.capacity).map{ bbox.result[$0] }.strideArray())
+    
+    let nmsCompute = NMSCompute.init()
+    nmsCompute.scoreThredshold = 0.25
+    nmsCompute.nmsTopK = 100
+    nmsCompute.keepTopK = 100
+    nmsCompute.nmsEta = 1.0
+    nmsCompute.nmsThreshold = 0.449999988
+    nmsCompute.background_label = 0;
+    nmsCompute.scoreDim = [NSNumber.init(value: score.dim[0]), NSNumber.init(value: score.dim[1]), NSNumber.init(value: score.dim[2])]
+    nmsCompute.bboxDim = [NSNumber.init(value: bbox.dim[0]), NSNumber.init(value: bbox.dim[1]), NSNumber.init(value: bbox.dim[2])]
+    guard let result = nmsCompute.compute(withScore: score.result, andBBoxs: bbox.result) else {
+      fatalError( " result error " )
+    }
+    let resultHolder = ResultHolder.init(inResult: result.output, inCapacity: Int(result.outputSize))
+//    for i in 0..<Int(result.outputSize) {
+//
+//      print("i \(i) : \(result.output[i])")
+//    }
+//    print(Date.init().timeIntervalSince(startDate))
+
+//    print(resultHolder.result![0])
+    return resultHolder
+  }
+  
+  override func updateProgram(program: Program) {
+    for i in [56, 66, 76, 86, 93, 99] {
+      let opDesc = program.programDesc.blocks[0].ops[i]
+      let output = opDesc.outputs["Out"]!.first!
+      let v = program.scope[output]!
+      let originTexture = v as! Texture<Float32>
+      originTexture.tensorDim = Dim.init(inDim: [originTexture.tensorDim[1] / 7, originTexture.tensorDim[0] * 7])
+      
+      originTexture.dim = Dim.init(inDim: [1, 1, originTexture.dim[3] / 7, originTexture.dim[2] * 7])
+      
+      originTexture.padToFourDim = Dim.init(inDim: [1, 1, originTexture.padToFourDim[3] / 7, originTexture.padToFourDim[2] * 7])
+      
+      program.scope[output] = originTexture
+      
+      if i == 99 {
+        opDesc.attrs["axis"] = 0
+      } else {
+        opDesc.attrs["shape"] = originTexture.tensorDim.dims.map { Int32($0) }
+      }
+    }
+    
+    for i in [58, 59, 88, 89, 95, 96, 68, 69, 78, 79] {
+      let opDesc = program.programDesc.blocks[0].ops[i]
+      let output = opDesc.outputs["Out"]!.first!
+      let v = program.scope[output]!
+      
+      
+      
+      let originTexture = v as! Texture<Float32>
+      originTexture.tensorDim = Dim.init(inDim: [originTexture.tensorDim[1], originTexture.tensorDim[2]])
+      opDesc.attrs["shape"] = originTexture.tensorDim.dims.map { Int32($0) }
+    }
+    
+    for i in [60, 101, 90, 97, 70, 80] {
+      let opDesc = program.programDesc.blocks[0].ops[i]
+      let output = opDesc.outputs["Out"]!.first!
+      let v = program.scope[output]!
+      let originTexture = v as! Texture<Float32>
+      originTexture.tensorDim = Dim.init(inDim: [originTexture.tensorDim[1], originTexture.tensorDim[2]])
+      opDesc.attrs["axis"] = (opDesc.attrs["axis"]! as! Int) - 1
+    }
+    
+    for i in [102] {
+      let opDesc = program.programDesc.blocks[0].ops[i]
+      for output in opDesc.outputs["Out"]! {
+        let v = program.scope[output]!
+        let originTexture = v as! Texture<Float32>
+        originTexture.tensorDim = Dim.init(inDim: [originTexture.tensorDim[1], originTexture.tensorDim[2]])
+      }
+      opDesc.attrs["axis"] = (opDesc.attrs["axis"]! as! Int) - 1
+      print(" split axis \(opDesc.attrs["axis"])")
+    }
+    // 99
+  }
+  
+}

metal/paddle-mobile/paddle-mobile/Net.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+
+import Foundation
+
+public class ResultHolder: NSObject {
+  @objc public let result: UnsafeMutablePointer<Float32>?
+  @objc public let capacity: Int
+
+  init(inResult: UnsafeMutablePointer<Float32>?, inCapacity: Int) {
+    result = inResult
+    capacity = inCapacity
+  }
+  
+  @objc public func releasePointer() {
+    result?.deinitialize(count: capacity)
+    result?.deallocate()
+  }
+}
+
+public class Net: NSObject {
+  var except: Int = 0
+  var means: [Float] = []
+  var scale: Float = 0.0
+  var dim: (n: Int, h: Int, w: Int, c: Int) = (n: 0, h: 0, w: 0, c: 0)
+  var preprocessKernel: CusomKernel? = nil
+  var paramPointer: UnsafeMutableRawPointer? = nil
+  var paramSize: Int = 0
+  var modelPointer: UnsafeMutableRawPointer? = nil
+  var modelSize: Int = 0
+  var modelPath: String = ""
+  var paramPath: String = ""
+  var modelDir: String = ""
+  @objc public init(device: MTLDevice,paramPointer: UnsafeMutableRawPointer, paramSize:Int, modePointer: UnsafeMutableRawPointer, modelSize: Int) {
+      self.paramPointer = paramPointer
+      self.paramSize = paramSize
+      self.modelPointer = modePointer
+      self.modelSize = modelSize
+      super.init()
+  }
+
+  
+  public func resultStr(res: ResultHolder) -> String {
+    fatalError()
+  }
+  
+  func fetchResult(paddleMobileRes: GPUResultHolder) -> ResultHolder {
+    return ResultHolder.init(inResult: paddleMobileRes.resultPointer, inCapacity: paddleMobileRes.capacity)
+  }
+  
+  @objc public init(device: MTLDevice) {
+    super.init()
+  }
+  
+  func updateProgram(program: Program) {
+
+  }
+}

metal/paddle-mobile/paddle-mobile/Operators/Base/OpCreator.swift

@@ -43,14 +43,31 @@ class OpCreator<P: PrecisionType> {
         [gConvType                  :     ConvOp<P>.creat,
          gBatchNormType             :     BatchNormOp<P>.creat,
          gReluType                  :     ReluOp<P>.creat,
-         gElementwiseAdd            :     ElementwiseAddOp<P>.creat,
+         gElementwiseAddType        :     ElementwiseAddOp<P>.creat,
          gFeedType                  :     FeedOp<P>.creat,
          gFetchType                 :     FetchOp<P>.creat,
          gConvAddBatchNormReluType  :     ConvAddBatchNormReluOp<P>.creat,
          gPooType                   :     PoolOp<P>.creat,
          gSoftmaxType               :     SoftmaxOp<P>.creat,
          gReshapeType               :     ReshapeOp<P>.creat,
-         gConvAddType               :     ConvAddOp<P>.creat]
+         gConvAddType               :     ConvAddOp<P>.creat,
+         gDepthConvType             :     DepthConvOp<P>.creat,
+         gConcatType                :     ConcatOp<P>.creat,
+         gBoxcoderType              :     BoxcoderOp<P>.creat,
+         gConvBnReluType            :     ConvBNReluOp<P>.creat,
+         gDwConvBnReluType          :     DwConvBNReluOp<P>.creat,
+         gMulticlassNMSType         :     MulticlassNMSOp<P>.creat,
+         gTransposeType             :     TransposeOp<P>.creat,
+         gPriorBoxType              :     PriorBoxOp<P>.creat,
+         gPreluType                 :     PreluOp<P>.creat,
+         gConv2dTransposeType       :     ConvTransposeOp<P>.creat,
+         gBilinearInterpType        :     BilinearInterpOp<P>.creat,
+         gSplit                     :     SplitOp<P>.creat,
+         gShape                     :     ShapeOp<P>.creat,
+         gFlatten                   :     FlattenOp<P>.creat,
+         gConvAddPreluType          :     ConvAddPreluOp<P>.creat,
+         gConvAddAddPreluType       :     ConvAddAddPreluOp<P>.creat,
+         gElementwiseAddPreluType:   ElementwiseAddPreluOp<P>.creat]
   
     private init(){}
 }

metal/paddle-mobile/paddle-mobile/Operators/Base/OpParam.swift

@@ -39,8 +39,19 @@ protocol OpParam {
   static func output<VarType: Variant>(outputs: [String : [String]], from: Scope) throws -> VarType
   static func outputY<VarType: Variant>(outputs: [String : [String]], from: Scope) throws -> VarType
   static func inputY<VarType: Variant>(inputs: [String : [String]], from: Scope) throws -> VarType
+  
+  static func inputImage<VarType: Variant>(inputs: [String : [String]], from: Scope) throws -> VarType
+  
+  static func outputBoxes<VarType: Variant>(outputs: [String : [String]], from: Scope) throws -> VarType
+  
   static func outputOut<VarType: Variant>(outputs: [String : [String]], from: Scope) throws -> VarType
+  
+  static func outputVariances<VarType: Variant>(outputs: [String : [String]], from: Scope) throws -> VarType
+  
   static func getAttr<T>(key: String, attrs: [String : Attr]) throws -> T
+  
+  static func paramInputAlpha<VarType: Variant>(inputs: [String : [String]], from: Scope) throws -> VarType
+  
 }
 
 extension OpParam {
@@ -52,22 +63,63 @@ extension OpParam {
     guard let mapKeys = map[key], mapKeys.count > 0 else {
       throw PaddleMobileError.paramError(message: key + " not found in \(map) or maped values is empty")
     }
-        guard let variant = from[mapKeys[0]], let v = variant as? VarType else {
+    guard let variant = from[mapKeys[0]] else {
       throw PaddleMobileError.paramError(message: mapKeys[0] + " not found in scope")
     }
+    
+    guard let v = variant as? VarType else {
+      throw PaddleMobileError.paramError(message: " type error")
+
+    }
     return v
   }
   
+  static func outputVariances<VarType: Variant>(outputs: [String : [String]], from: Scope) throws -> VarType {
+    do {
+      let tensorVariances: VarType = try getFirstTensor(key: "Variances", map: outputs, from: from)
+      return tensorVariances
+    } catch let error {
+      throw error
+    }
+  }
+  
+  static func paramInputAlpha<VarType: Variant>(inputs: [String : [String]], from: Scope) throws -> VarType {
+    do {
+      let alphaTensor: VarType = try getFirstTensor(key: "Alpha", map: inputs, from: from)
+      return alphaTensor
+    } catch let error {
+      throw error
+    }
+  }
+  
+  
+  static func inputImage<VarType: Variant>(inputs: [String : [String]], from: Scope) throws -> VarType {
+    do {
+      let tensorImage: VarType = try getFirstTensor(key: "Image", map: inputs, from: from)
+      return tensorImage
+    } catch let error {
+      throw error
+    }
+  }
+  
   static func inputX<VarType: Variant>(inputs: [String : [String]], from: Scope) throws -> VarType {
     do {
       let tensorX: VarType = try getFirstTensor(key: "X", map: inputs, from: from)
-            
       return tensorX
     } catch let error {
       throw error
     }
   }
   
+  static func outputBoxes<VarType: Variant>(outputs: [String : [String]], from: Scope) throws -> VarType {
+    do {
+      let tensorBox: VarType = try getFirstTensor(key: "Boxes", map: outputs, from: from)
+      return tensorBox
+    } catch let error {
+      throw error
+    }
+  }
+  
   static func input<VarType: Variant>(inputs: [String : [String]], from: Scope) throws -> VarType {
     do {
       let tensorInput: VarType = try getFirstTensor(key: "Input", map: inputs, from: from)

metal/paddle-mobile/paddle-mobile/Operators/Base/Operator.swift

@@ -19,12 +19,20 @@ protocol Fusion {
   static func fusionNode() -> Node
   static func change() -> [String : [(from: String, to: String)]]
   static func fusionType() -> String
+  static func needCheck() -> [(Int, String)]
+}
+extension Fusion {
+  static func needCheck() -> [(Int, String)] {
+    return []
+  }
 }
 
 protocol Runable {
   func run(device: MTLDevice, buffer: MTLCommandBuffer) throws
   func runImpl(device: MTLDevice,buffer: MTLCommandBuffer) throws
   func delogOutput()
+  func inputVariant() -> [String : [Variant]]
+  func computeMiddleResult(device: MTLDevice, buffer: MTLCommandBuffer)
 }
 
 extension Runable where Self: OperatorProtocol{
@@ -34,10 +42,19 @@ extension Runable where Self: OperatorProtocol{
     } catch let error {
       throw error
     }
-//        print(type + ": " + para.outputDesc())
+  }
+  
+  func inputVariant() -> [String : [Variant]] {
+//    return [:]
+    fatalError(" op \(type) need implement inputVariant")
+  }
+  
+  func computeMiddleResult(device: MTLDevice, buffer: MTLCommandBuffer) {
+    fatalError(" need implement ")
   }
   
   func delogOutput() {
+    
     print(type + ": has no implementation" )
   }
 }
@@ -97,6 +114,7 @@ class Operator <KernelType:  Computable , ParameterType>: OperatorProtocol where
   let scope: Scope
   var kernel: KerType
   required init(device: MTLDevice, opDesc: OpDesc, inScope: Scope) throws {
+//    print("create op: \(opDesc.type)")
     type = opDesc.type
     scope = inScope
     inputs = opDesc.inputs
@@ -118,22 +136,57 @@ let gFeedType                   = "feed"
 let gConvType                   = "conv2d"
 let gBatchNormType              = "batch_norm"
 let gReluType                   = "relu"
-let gElementwiseAdd             = "elementwise_add"
+let gElementwiseAddType         = "elementwise_add"
 let gConvAddBatchNormReluType   = "conv_add_batchnorm_relu"
 let gPooType                    = "pool2d"
 let gSoftmaxType                = "softmax"
 let gReshapeType                = "reshape"
 let gConvAddType                = "conv_add"
+let gDepthConvType              = "depthwise_conv2d"
+let gPriorBoxType               = "prior_box"
+let gTransposeType              = "transpose"
+let gConcatType                 = "concat"
+let gBoxcoderType               = "box_coder"
+let gMulticlassNMSType          = "multiclass_nms"
+let gConvBnReluType             = "conv_bn_relu"
+let gDwConvBnReluType           = "depth_conv_bn_relu"
+let gPreluType                  = "prelu"
+let gConv2dTransposeType        = "conv2d_transpose"
+let gBilinearInterpType         = "bilinear_interp"
+let gSplit                      = "split"
+let gShape                      = "shape"
+let gFlatten                    = "flatten"
+let gConvAddPreluType           = "conv_add_prelu"
+let gConvAddAddPreluType        = "conv_add_add_prelu"
+let gElementwiseAddPreluType = "elementwise_add_prelu"
 
 
 let opInfos = [gConvType                    : (inputs: ["Input"], outputs: ["Output"]),
                gBatchNormType               : (inputs: ["X"], outputs: ["Y"]),
                gReluType                    : (inputs: ["X"], outputs: ["Out"]),
-               gElementwiseAdd              : (inputs: ["X"], outputs: ["Out"]),
+               gElementwiseAddType          : (inputs: ["X"], outputs: ["Out"]),
                gFeedType                    : (inputs: ["X"], outputs: ["Out"]),
                gFetchType                   : (inputs: ["X"], outputs: ["Out"]),
                gConvAddBatchNormReluType    : (inputs: ["Input"], outputs: ["Out"]),
                gPooType                     : (inputs: ["X"], outputs: ["Out"]),
                gSoftmaxType                 : (inputs: ["X"], outputs: ["Out"]),
                gReshapeType                 : (inputs: ["X"], outputs: ["Out"]),
-               gConvAddType                 : (inputs: ["Input"], outputs: ["Out"])]
+               gConvAddType                 : (inputs: ["Input"], outputs: ["Out"]),
+               gDepthConvType               : (inputs: ["Input"], outputs: ["Output"]),
+               gConcatType                  : (inputs: ["X"], outputs: ["Out"]),
+               gBoxcoderType                : (inputs: ["PriorBox", "PriorBoxVar", "TargetBox"], outputs: ["OutputBox"]),
+               gTransposeType               : (inputs: ["X"], outputs: ["Out"]),
+               gConvBnReluType              : (inputs: ["Input"], outputs: ["Out"]),
+               gDwConvBnReluType            : (inputs: ["Input"], outputs: ["Out"]),
+               gMulticlassNMSType           : (inputs: ["BBoxes", "Scores"], outputs: ["Out"]),
+               gPriorBoxType                : (inputs: ["Input", "Image"], outputs: ["Boxes", "Variances"]),
+               gPreluType                   : (inputs: ["X"], outputs: ["Out"]),
+               gConv2dTransposeType         : (inputs: ["Input"], outputs: ["Output"]),
+               gBilinearInterpType          : (inputs: ["X"], outputs: ["Out"]),
+               gSplit                       : (inputs: ["X"], outputs: ["Out"]),
+               gShape                       : (inputs: ["Input"], outputs: ["Out"]),
+               gFlatten                     : (inputs: ["X"], outputs: ["Out"]),
+               gConvAddPreluType            : (inputs: ["Input"], outputs: ["Out"]),
+               gConvAddAddPreluType         : (inputs: ["Input"], outputs: ["Out"]),
+               gElementwiseAddPreluType  :  (inputs: ["X"], outputs: ["Out"])
+              ]

metal/paddle-mobile/paddle-mobile/Operators/BatchNormOp.swift

-///* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
-// 
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-// 
-// http://www.apache.org/licenses/LICENSE-2.0
-// 
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License. */
+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
 
 import Foundation
 
@@ -19,34 +19,36 @@ class BatchNormParam<P: PrecisionType>: OpParam {
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       input = try BatchNormParam.inputX(inputs: opDesc.inputs, from: inScope)
+      if input.transpose != [0, 2, 3, 1] {
+        fatalError("batch norm only accepts NHWC")
+      }
       output = try BatchNormParam.outputY(outputs: opDesc.outputs, from: inScope)
-            inputBias = try BatchNormParam.inputBiase(inputs: opDesc.paraInputs, from: inScope)
-            inputMean = try BatchNormParam.inputMean(inputs: opDesc.paraInputs, from: inScope)
-            inputScale = try BatchNormParam.inputScale(inputs: opDesc.paraInputs, from: inScope)
-            inputVariance = try BatchNormParam.inputVariance(inputs: opDesc.paraInputs, from: inScope)
+      bias = try BatchNormParam.getFirstTensor(key: "Bias", map: opDesc.paraInputs, from: inScope)
+      mean = try BatchNormParam.getFirstTensor(key: "Mean", map: opDesc.paraInputs, from: inScope)
+      scale = try BatchNormParam.getFirstTensor(key: "Scale", map: opDesc.paraInputs, from: inScope)
+      variance = try BatchNormParam.getFirstTensor(key: "Variance", map: opDesc.paraInputs, from: inScope)
       epsilon = try BatchNormParam.getAttr(key: "epsilon", attrs: opDesc.attrs)
       momentum = try BatchNormParam.getAttr(key: "momentum", attrs: opDesc.attrs)
-            is_test = try BatchNormParam.getAttr(key: "is_test", attrs: opDesc.attrs)
     } catch let error {
       throw error
     }
   }
   let input: Texture<P>
   var output: Texture<P>
-    let inputBias: Tensor<ParamPrecisionType>
-    let inputMean: Tensor<ParamPrecisionType>
-    let inputScale: Tensor<ParamPrecisionType>
-    let inputVariance: Tensor<ParamPrecisionType>
+  let bias: Tensor<P>
+  let mean: Tensor<P>
+  let scale: Tensor<P>
+  let variance: Tensor<P>
   let epsilon: Float
   let momentum: Float
-    let is_test: Bool
 }
 
 class BatchNormOp<P: PrecisionType>: Operator<BatchNormKernel<P>, BatchNormParam<P>>, Runable, Creator, InferShaperable{
+  typealias OpType = BatchNormOp<P>
+
   func inferShape() {
     para.output.dim = para.input.dim
   }
-    typealias OpType = BatchNormOp<P>
   func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
     do {
       try kernel.compute(commandBuffer: buffer, param: para)
@@ -54,9 +56,11 @@ class BatchNormOp<P: PrecisionType>: Operator<BatchNormKernel<P>, BatchNormParam
       throw error
     }
   }
-}
-
-
-
-
   
+  func delogOutput() {
+    print(" \(type) output: ")
+    let device = para.output.metalTexture!.device
+    let outputArray: [Float32] = device.texture2tensor(texture: para.output.metalTexture, dim: para.output.tensorDim.dims, transpose: para.output.transpose)
+    print(outputArray.strideArray())
+  }
+}

metal/paddle-mobile/paddle-mobile/Operators/BilinearInterpOp.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class BilinearInterpParam<P: PrecisionType>: OpParam {
+  typealias ParamPrecisionType = P
+  required init(opDesc: OpDesc, inScope: Scope) throws {
+    do {
+      input = try BilinearInterpParam.inputX(inputs: opDesc.inputs, from: inScope)
+      output = try BilinearInterpParam.outputOut(outputs: opDesc.outputs, from: inScope)
+      out_h = try BilinearInterpParam.getAttr(key: "out_h", attrs: opDesc.attrs)
+      out_w = try BilinearInterpParam.getAttr(key: "out_w", attrs: opDesc.attrs)
+    } catch let error {
+      throw error
+    }
+    if (input.transpose != [0, 2, 3, 1]) || (input.tensorDim.cout() != 4) {
+      fatalError()
+    }
+  }
+  let input: Texture<P>
+  var output: Texture<P>
+  let out_h: Int
+  let out_w: Int
+}
+
+class BilinearInterpOp<P: PrecisionType>: Operator<BilinearInterpKernel<P>, BilinearInterpParam<P>>, Runable, Creator, InferShaperable{
+  
+  typealias OpType = BilinearInterpOp<P>
+
+  func inferShape() {
+    //        para.output.dim = para.input.dim
+  }
+  
+  func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
+    do {
+      try kernel.compute(commandBuffer: buffer, param: para)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  func delogOutput() {
+    print(" \(type) output: ")
+    let device = para.output.metalTexture!.device
+    let outputArray: [Float32] = device.texture2tensor(texture: para.output.metalTexture, dim: para.output.tensorDim.dims, transpose: para.output.transpose)
+//    print(outputArray)
+    print(outputArray.strideArray())
+  }
+  
+}
+
+
+
+
+
+

metal/paddle-mobile/paddle-mobile/Operators/BoxcoderOp.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class BoxcoderParam<P: PrecisionType>: OpParam {
+  typealias ParamPrecisionType = P
+  required init(opDesc: OpDesc, inScope: Scope) throws {
+    do {
+      priorBox = try BoxcoderParam.getFirstTensor(key: "PriorBox", map: opDesc.inputs, from: inScope)
+      priorBoxVar = try BoxcoderParam.getFirstTensor(key: "PriorBoxVar", map: opDesc.inputs, from: inScope)
+      targetBox = try BoxcoderParam.getFirstTensor(key: "TargetBox", map: opDesc.inputs, from: inScope)
+      output = try BoxcoderParam.getFirstTensor(key: "OutputBox", map: opDesc.outputs, from: inScope)
+      codeType = try BoxcoderParam.getAttr(key: "code_type", attrs: opDesc.attrs)
+      boxNormalized = try BoxcoderParam.getAttr(key: "box_normalized", attrs: opDesc.attrs)
+    } catch let error {
+      throw error
+    }
+    assert(priorBox.tensorDim.cout() == 2)
+    assert(priorBoxVar.tensorDim.cout() == 2)
+    assert(targetBox.tensorDim.cout() == 3)
+    assert(output.tensorDim.cout() == 3)
+    assert(priorBox.transpose == [0, 1, 2, 3])
+    assert(priorBoxVar.transpose == [0, 1, 2, 3])
+    assert(targetBox.transpose == [0, 1, 2, 3])
+    assert(codeType == "decode_center_size") // encode_center_size is not implemented
+    assert((targetBox.tensorDim.cout() == 3) && (targetBox.tensorDim[0] == 1)) // N must be 1 (only handle batch size = 1)
+  }
+  let priorBox: Texture<P>
+  let priorBoxVar: Texture<P>
+  let targetBox: Texture<P>
+  var output: Texture<P>
+  let codeType: String
+  let boxNormalized: Bool
+}
+
+class BoxcoderOp<P: PrecisionType>: Operator<BoxcoderKernel<P>, BoxcoderParam<P>>, Runable, Creator, InferShaperable{
+  
+  typealias OpType = BoxcoderOp<P>
+
+  func inferShape() {
+    //        para.output.dim = para.input.dim
+  }
+  
+  func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
+    do {
+      try kernel.compute(commandBuffer: buffer, param: para)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  func delogOutput() {
+    print(" \(type) output: ")
+    let device = para.output.metalTexture!.device
+    let pbv : [Float32] = device.texture2tensor(texture: para.priorBoxVar.metalTexture!, dim: para.priorBoxVar.tensorDim.dims, transpose: para.priorBoxVar.transpose)
+    let pb : [Float32] = device.texture2tensor(texture: para.priorBox.metalTexture!, dim: para.priorBox.tensorDim.dims, transpose: para.priorBox.transpose)
+    let tb : [Float32] = device.texture2tensor(texture: para.targetBox.metalTexture!, dim: para.targetBox.tensorDim.dims, transpose: para.targetBox.transpose)
+    let out : [Float32] = device.texture2tensor(texture: para.output.metalTexture!, dim: para.output.tensorDim.dims, transpose: para.output.transpose)
+    print(" prior box var ")
+    print(pbv.strideArray())
+    print(" target box ")
+    print(tb.strideArray())
+    print(" prior box ")
+    print(pb.strideArray())
+    print(" output ")
+    print(out.strideArray())
+  }
+  
+}
+
+
+
+
+
+

metal/paddle-mobile/paddle-mobile/Operators/CNNMPSConvOp.swift

+//
+//  CNNConvAddBatchNormReluOp.swift
+//  paddle-mobile
+
+import Foundation
+
+class CNNMPSConvTestParam: TestParam {
+    var outputTexture: MTLTexture?
+    var metalParam: MetalConvParam
+    let filterPointer: UnsafeMutableRawPointer
+    let biasePointer: UnsafeMutablePointer<Float>
+    let filterSize: (width: Int, height: Int, channel: Int)
+    init(inMetalParam: MetalConvParam, inFilter: [Float], inBiase: [Float], inFilterSize: (width: Int, height: Int, channel: Int)) {
+        metalParam = inMetalParam
+        filterPointer = UnsafeMutableRawPointer.init(mutating: inFilter)
+        biasePointer = UnsafeMutablePointer.init(mutating: inBiase)
+        filterSize = inFilterSize
+    }
+}
+
+@available(iOS 10.0, *)
+class CNNMPSConvOp<P: PrecisionType>: Operator<CNNConvKernel<P>, CNNConvParam<P>>, Runable, Creator, InferShaperable, Fusion {
+    
+    typealias OpType = CNNMPSConvOp<P>
+
+    required init(device: MTLDevice, opDesc: OpDesc, inScope: Scope) throws {
+        fatalError()
+    }
+    
+    func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
+        do {
+            try kernel.compute(commandBuffer: buffer, param: para)
+        } catch let error {
+            throw error
+        }
+    }
+    
+    func delogOutput() {
+    }
+    
+    static func fusionNode() -> Node {
+        let beginNode = Node.init(inType: gConvType)
+        _ = beginNode-->Node.init(inType: gElementwiseAdd);
+        return beginNode
+    }
+    
+    static func change() -> [String : [(from: String, to: String)]] {
+        return [:]
+    }
+    
+    static func fusionType() -> String {
+        return gMPSCNNConvType
+    }
+    func inferShape() {
+        let inDims = para.input.dim
+        let filterDim = para.filter.dim
+        let strides = para.stride
+        let paddings = para.paddings
+        let dilations = para.dilations
+        
+        var outDim = [inDims[0]]
+        for i in 0..<strides.count {
+            let dilation: Int = Int(dilations[i])
+            let filterSize: Int = filterDim[i + 1]
+            let inputSize: Int = inDims[i + 1]
+            let padding: Int = Int(paddings[i])
+            let stride: Int = Int(strides[i])
+            let dKernel = dilation * (filterSize - 1) + 1
+            let outputSize = (inputSize + 2 * padding - dKernel) / stride + 1
+            outDim.append(outputSize)
+        }
+        outDim.append(filterDim[0])
+        para.output.dim = Dim.init(inDim: outDim)
+    }
+}

metal/paddle-mobile/paddle-mobile/Operators/ConcatOp.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class ConcatParam<P: PrecisionType>: OpParam {
+  typealias ParamPrecisionType = P
+  required init(opDesc: OpDesc, inScope: Scope) throws {
+    do {
+      guard let xlist = opDesc.inputs["X"] else {
+        fatalError()
+      }
+      for x in xlist {
+        guard let variant = inScope[x], let v = variant as? Texture<P> else {
+          fatalError()
+        }
+        if transpose.count == 0 {
+          transpose = v.transpose
+        }
+        if v.transpose != transpose {
+          fatalError()
+        }
+       
+        input.append(v)
+      }
+      axis = try ConcatParam.getAttr(key: "axis", attrs: opDesc.attrs)
+      output = try ConcatParam.outputOut(outputs: opDesc.outputs, from: inScope)
+    } catch let error {
+      throw error
+    }
+  }
+  var input: [Texture<P>] = []
+  var output: Texture<P>
+  var transpose: [Int] = []
+  let axis: Int
+}
+
+class ConcatOp<P: PrecisionType>: Operator<ConcatKernel<P>, ConcatParam<P>>, Runable, Creator, InferShaperable{
+  
+  typealias OpType = ConcatOp<P>
+
+  func inferShape() {
+    //        let dim = para.input.reduce([0, 0]) {[$0[0] + $1.dim[0], $1.dim[1]]}
+    //        para.output.dim = Dim.init(inDim: dim)
+  }
+  
+  func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
+    do {
+      try kernel.compute(commandBuffer: buffer, param: para)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  func delogOutput() {
+    print(" \(type) output: ")
+    
+    let device = para.output.metalTexture!.device
+    let outputArray: [Float32] = device.texture2tensor(texture: para.output.metalTexture, dim: para.output.tensorDim.dims, transpose: para.output.transpose)
+    print(outputArray.strideArray())
+  }
+  
+}
+
+
+

metal/paddle-mobile/paddle-mobile/Operators/ConvAddAddPreluOp.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class ConvAddAddPreluParam<P: PrecisionType>: OpParam {
+  typealias ParamPrecisionType = P
+  required init(opDesc: OpDesc, inScope: Scope) throws {
+    do {
+      filter = try ConvAddAddPreluParam.inputFilter(paraInputs: opDesc.paraInputs, from: inScope)
+      input = try ConvAddAddPreluParam.input(inputs: opDesc.inputs, from: inScope)
+      output = try ConvAddAddPreluParam.outputOut(outputs: opDesc.outputs, from: inScope)
+      stride = try ConvAddAddPreluParam.getAttr(key: "strides", attrs: opDesc.attrs)
+      paddings = try ConvAddAddPreluParam.getAttr(key: "paddings", attrs: opDesc.attrs)
+      dilations = try ConvAddAddPreluParam.getAttr(key: "dilations", attrs: opDesc.attrs)
+      groups = try ConvAddAddPreluParam.getAttr(key: "groups", attrs: opDesc.attrs)
+      alpha = try ConvAddAddPreluParam.paramInputAlpha(inputs: opDesc.paraInputs, from: inScope)
+      mode = try ConvAddAddPreluParam.getAttr(key: "mode", attrs: opDesc.attrs)
+      y = try ConvAddAddPreluParam.inputY(inputs: opDesc.paraInputs, from: inScope)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  let input: Texture<P>
+  let y: Tensor<ParamPrecisionType>
+  let filter: Tensor<ParamPrecisionType>
+  let mode: String
+  let alpha: Tensor<P>
+  var output: Texture<P>
+  let stride: [Int32]
+  let paddings: [Int32]
+  let dilations: [Int32]
+  let groups: Int
+}
+
+class ConvAddAddPreluOp<P: PrecisionType>: Operator<ConvAddAddPreluKernel<P>, ConvAddAddPreluParam<P>>, Runable, Creator, InferShaperable, Fusion{
+  typealias OpType = ConvAddAddPreluOp<P>
+  
+  static func fusionNode() -> Node {
+    let beginNode = Node.init(inType: gConvType)
+    _ = beginNode
+      --> Node.init(inType: gElementwiseAddType) --> Node.init(inType: gElementwiseAddType) --> Node.init(inType: gPreluType)
+    return beginNode
+  }
+  
+  static func change() -> [String : [(from: String, to: String)]] {
+    return [:]
+  }
+  
+  static func fusionType() -> String {
+    return gConvAddAddPreluType
+  }
+  
+  static func needCheck() -> [(Int, String)] {
+    return [(2, "Y"), (2, "X")]
+  }
+  
+  
+  
+  func inferShape() {
+    let inDims = para.input.dim
+    let filterDim = para.filter.dim
+    let strides = para.stride
+    let paddings = para.paddings
+    let dilations = para.dilations
+    
+    var outDim = [inDims[0]]
+    for i in 0..<strides.count {
+      let dilation: Int = Int(dilations[i])
+      let filterSize: Int = filterDim[i + 1]
+      let inputSize: Int = inDims[i + 1]
+      let padding: Int = Int(paddings[i])
+      let stride: Int = Int(strides[i])
+      let dKernel = dilation * (filterSize - 1) + 1
+      let outputSize = (inputSize + 2 * padding - dKernel) / stride + 1
+      outDim.append(outputSize)
+    }
+    outDim.append(filterDim[0])
+    para.output.dim = Dim.init(inDim: outDim)
+  }
+  
+  func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
+    do {
+      try kernel.compute(commandBuffer: buffer, param: para)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  
+  func delogOutput() {
+    print(" \(type) output: ")
+    print(para.output.metalTexture.toTensor(dim: (n: para.output.tensorDim[0], c: para.output.tensorDim[1], h: para.output.tensorDim[2], w: para.output.tensorDim[3])).strideArray())
+  }
+  
+}

metal/paddle-mobile/paddle-mobile/Operators/ConvAddBatchNormReluOp.swift

@@ -14,10 +14,12 @@
 
 import Foundation
 
+
 class ConvAddBatchNormReluParam<P: PrecisionType>: OpParam {
   typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
+      
       filter = try ConvAddBatchNormReluParam.inputFilter(paraInputs: opDesc.paraInputs, from: inScope)
       input = try ConvAddBatchNormReluParam.input(inputs: opDesc.inputs, from: inScope)
       output = try ConvAddBatchNormReluParam.outputOut(outputs: opDesc.outputs, from: inScope)
@@ -29,6 +31,7 @@ class ConvAddBatchNormReluParam<P: PrecisionType>: OpParam {
       groups = try ConvAddBatchNormReluParam.getAttr(key: "groups", attrs: opDesc.attrs)
       variance = try ConvAddBatchNormReluParam.inputVariance(inputs: opDesc.paraInputs, from: inScope)
       bias = try ConvAddBatchNormReluParam.inputBiase(inputs: opDesc.paraInputs, from: inScope)
+      
       scale = try ConvAddBatchNormReluParam.inputScale(inputs: opDesc.paraInputs, from: inScope)
       mean = try ConvAddBatchNormReluParam.inputMean(inputs: opDesc.paraInputs, from: inScope)
       y = try ConvAddBatchNormReluParam.inputY(inputs: opDesc.paraInputs, from: inScope)
@@ -57,6 +60,7 @@ class ConvAddBatchNormReluParam<P: PrecisionType>: OpParam {
 }
 
 class ConvAddBatchNormReluOp<P: PrecisionType>: Operator<ConvAddBatchNormReluKernel<P>, ConvAddBatchNormReluParam<P>>, Runable, Creator, InferShaperable, Fusion{
+  
   typealias OpType = ConvAddBatchNormReluOp<P>
   
   func inferShape() {
@@ -92,7 +96,7 @@ class ConvAddBatchNormReluOp<P: PrecisionType>: Operator<ConvAddBatchNormReluKer
   static func fusionNode() -> Node {
     let beginNode = Node.init(inType: gConvType)
     _ = beginNode
-            --> Node.init(inType: gElementwiseAdd)
+      --> Node.init(inType: gElementwiseAddType)
       --> Node.init(inType: gBatchNormType)
       --> Node.init(inType: gReluType)
     return beginNode
@@ -107,26 +111,20 @@ class ConvAddBatchNormReluOp<P: PrecisionType>: Operator<ConvAddBatchNormReluKer
   }
   
   func delogOutput() {
+    print(" conv add batchnorm relu output ")
+    print(para.output.toTensor().strideArray())
+    //        let _: P? = para.input.metalTexture.logDesc(header: "conv add batchnorm relu input: ", stridable: false)
+    //        para.filter.logDataPointer(header: "filter data pointer: ")
+    //        print("filter: \(para.filter)")
     
-//        let _: P? = para.input.metalTexture.logDesc(header: "conv add batchnorm relu input: ", stridable: false)
-//        para.filter.logDataPointer(header: "filter data pointer: ")
-//        print("filter: \(para.filter)")
+    //        print("biase: \(para.y)")
+    //        print("padding: \(para.paddings)")
+    //        print("stride: \(para.stride)")
     
-//        print("biase: \(para.y)")
-//        print("padding: \(para.paddings)")
-//        print("stride: \(para.stride)")
-        
-//        let _: P? = para.y.buffer?.logDesc(header: " biase: ", stridable: false)
-//        let _: P? = para.newBiase?.logDesc(header: "new biase: ", stridable: false)
-//        let _: P? = para.newScale?.logDesc(header: "new scale: ", stridable: false)
-        
-        let output = para.output.metalTexture.floatArray { (p: P) -> P in
-            return p
-        }
-//
-        writeToLibrary(fileName: "output_112x112x32_2", array: output)
-        print(" write done")
+    //        let _: P? = para.y.buffer?.logDesc(header: " biase: ", stridable: false)
+    //        let _: P? = para.newBiase?.logDesc(header: "new biase: ", stridable: false)
+    //        let _: P? = para.newScale?.logDesc(header: "new scale: ", stridable: false)
     
-//        let _: P? = para.output.metalTexture.logDesc(header: "conv add batchnorm relu output: ", stridable: false)
+    //        let _: P? = para.output.metalTexture.logDesc(header: "conv add batchnorm relu output: ", stridable: false)
   }
 }

metal/paddle-mobile/paddle-mobile/Operators/ConvAddOp.swift

@@ -25,6 +25,7 @@ class ConvAddParam<P: PrecisionType>: OpParam {
       paddings = try ConvAddParam.getAttr(key: "paddings", attrs: opDesc.attrs)
       dilations = try ConvAddParam.getAttr(key: "dilations", attrs: opDesc.attrs)
       groups = try ConvAddParam.getAttr(key: "groups", attrs: opDesc.attrs)
+      
       y = try ConvAddParam.inputY(inputs: opDesc.paraInputs, from: inScope)
     } catch let error {
       throw error
@@ -43,10 +44,12 @@ class ConvAddParam<P: PrecisionType>: OpParam {
 }
 
 class ConvAddOp<P: PrecisionType>: Operator<ConvAddKernel<P>, ConvAddParam<P>>, Runable, Creator, InferShaperable, Fusion{
+  typealias OpType = ConvAddOp<P>
+
   static func fusionNode() -> Node {
     let beginNode = Node.init(inType: gConvType)
     _ = beginNode
-            --> Node.init(inType: gElementwiseAdd)
+      --> Node.init(inType: gElementwiseAddType)
     return beginNode
   }
   
@@ -58,9 +61,8 @@ class ConvAddOp<P: PrecisionType>: Operator<ConvAddKernel<P>, ConvAddParam<P>>,
     return gConvAddType
   }
   
-    typealias OpType = ConvAddOp<P>
-    
   func inferShape() {
+    
     let inDims = para.input.dim
     let filterDim = para.filter.dim
     let strides = para.stride
@@ -90,4 +92,25 @@ class ConvAddOp<P: PrecisionType>: Operator<ConvAddKernel<P>, ConvAddParam<P>>,
     }
   }
   
+  func delogOutput() {
+//    print("op \(type): ")
+//    print(" padding: ")
+//    print(para.paddings)
+//    print("stride: ")
+//    print(para.stride)
+//    print("dilations: ")
+//    print(para.dilations)
+//    print(" para input dim: ")
+//    print(para.input.dim)
+//    print(" para filter dim: ")
+//    print(para.filter.dim)
+//    print(" para output dim: ")
+//    print(para.output.dim)
+//    print(" biase: ")
+//    let biase: [Float32] = para.y.buffer.array()
+//    print(biase)
+    
+    print(" \(type) output: ")
+    print(para.output.metalTexture.toTensor(dim: (n: para.output.tensorDim[0], c: para.output.tensorDim[1], h: para.output.tensorDim[2], w: para.output.tensorDim[3])).strideArray())
+  }
 }

metal/paddle-mobile/paddle-mobile/Operators/ConvAddPreluOp.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class ConvAddPreluParam<P: PrecisionType>: OpParam {
+  typealias ParamPrecisionType = P
+  required init(opDesc: OpDesc, inScope: Scope) throws {
+    do {
+      filter = try ConvAddPreluParam.inputFilter(paraInputs: opDesc.paraInputs, from: inScope)
+      input = try ConvAddPreluParam.input(inputs: opDesc.inputs, from: inScope)
+      output = try ConvAddPreluParam.outputOut(outputs: opDesc.outputs, from: inScope)
+      stride = try ConvAddPreluParam.getAttr(key: "strides", attrs: opDesc.attrs)
+      paddings = try ConvAddPreluParam.getAttr(key: "paddings", attrs: opDesc.attrs)
+      dilations = try ConvAddPreluParam.getAttr(key: "dilations", attrs: opDesc.attrs)
+      groups = try ConvAddPreluParam.getAttr(key: "groups", attrs: opDesc.attrs)
+      alpha = try ConvAddPreluParam.paramInputAlpha(inputs: opDesc.paraInputs, from: inScope)
+      mode = try ConvAddPreluParam.getAttr(key: "mode", attrs: opDesc.attrs)
+      y = try ConvAddPreluParam.inputY(inputs: opDesc.paraInputs, from: inScope)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  let input: Texture<P>
+  let y: Tensor<ParamPrecisionType>
+  let filter: Tensor<ParamPrecisionType>
+  let mode: String
+  let alpha: Tensor<P>
+  var output: Texture<P>
+  let stride: [Int32]
+  let paddings: [Int32]
+  let dilations: [Int32]
+  let groups: Int
+}
+
+class ConvAddPreluOp<P: PrecisionType>: Operator<ConvAddPreluKernel<P>, ConvAddPreluParam<P>>, Runable, Creator, InferShaperable, Fusion{
+  typealias OpType = ConvAddPreluOp<P>
+  
+  static func fusionNode() -> Node {
+    let beginNode = Node.init(inType: gConvType)
+    _ = beginNode
+      --> Node.init(inType: gElementwiseAddType) --> Node.init(inType: gPreluType)
+    return beginNode
+  }
+  
+  static func change() -> [String : [(from: String, to: String)]] {
+    return [:]
+  }
+  
+  static func fusionType() -> String {
+    return gConvAddPreluType
+  }
+  
+  func inferShape() {
+    let inDims = para.input.dim
+    let filterDim = para.filter.dim
+    let strides = para.stride
+    let paddings = para.paddings
+    let dilations = para.dilations
+    
+    var outDim = [inDims[0]]
+    for i in 0..<strides.count {
+      let dilation: Int = Int(dilations[i])
+      let filterSize: Int = filterDim[i + 1]
+      let inputSize: Int = inDims[i + 1]
+      let padding: Int = Int(paddings[i])
+      let stride: Int = Int(strides[i])
+      let dKernel = dilation * (filterSize - 1) + 1
+      let outputSize = (inputSize + 2 * padding - dKernel) / stride + 1
+      outDim.append(outputSize)
+    }
+    outDim.append(filterDim[0])
+    para.output.dim = Dim.init(inDim: outDim)
+  }
+  
+  func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
+    do {
+      try kernel.compute(commandBuffer: buffer, param: para)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  func delogOutput() {
+    print(" \(type) output: ")
+    print(para.output.metalTexture.toTensor(dim: (n: para.output.tensorDim[0], c: para.output.tensorDim[1], h: para.output.tensorDim[2], w: para.output.tensorDim[3])).strideArray())
+  }
+  
+}

metal/paddle-mobile/paddle-mobile/Operators/ConvBNReluOp.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class ConvBNReluParam<P: PrecisionType>: OpParam {
+  typealias ParamPrecisionType = P
+  required init(opDesc: OpDesc, inScope: Scope) throws {
+    do {
+      filter = try ConvBNReluParam.inputFilter(paraInputs: opDesc.paraInputs, from: inScope)
+      input = try ConvBNReluParam.input(inputs: opDesc.inputs, from: inScope)
+      output = try ConvBNReluParam.outputOut(outputs: opDesc.outputs, from: inScope)
+      stride = try ConvBNReluParam.getAttr(key: "strides", attrs: opDesc.attrs)
+      paddings = try ConvBNReluParam.getAttr(key: "paddings", attrs: opDesc.attrs)
+      dilations = try ConvBNReluParam.getAttr(key: "dilations", attrs: opDesc.attrs)
+      epsilon = try ConvBNReluParam.getAttr(key: "epsilon", attrs: opDesc.attrs)
+      
+      groups = try ConvBNReluParam.getAttr(key: "groups", attrs: opDesc.attrs)
+      variance = try ConvBNReluParam.inputVariance(inputs: opDesc.paraInputs, from: inScope)
+      bias = try ConvBNReluParam.inputBiase(inputs: opDesc.paraInputs, from: inScope)
+      scale = try ConvBNReluParam.inputScale(inputs: opDesc.paraInputs, from: inScope)
+      mean = try ConvBNReluParam.inputMean(inputs: opDesc.paraInputs, from: inScope)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  let input: Texture<P>
+  
+  let variance: Tensor<ParamPrecisionType>
+  let bias: Tensor<ParamPrecisionType>
+  let mean: Tensor<ParamPrecisionType>
+  let scale: Tensor<ParamPrecisionType>
+  let filter: Tensor<ParamPrecisionType>
+  let epsilon: Float32
+  var newScale: MTLBuffer?
+  var newBiase: MTLBuffer?
+  
+  var output: Texture<P>
+  let stride: [Int32]
+  let paddings: [Int32]
+  let dilations: [Int32]
+  let groups: Int
+}
+
+class ConvBNReluOp<P: PrecisionType>: Operator<ConvBNReluKernel<P>, ConvBNReluParam<P>>, Runable, Creator, InferShaperable, Fusion{
+  typealias OpType = ConvBNReluOp<P>
+  
+  func inputs() -> [Variant] {
+    return [para.input, para.variance, para.bias, para.mean, para.scale, para.filter]
+  }
+  
+  
+  func inferShape() {
+    let inDims = para.input.dim
+    let filterDim = para.filter.dim
+    let strides = para.stride
+    let paddings = para.paddings
+    let dilations = para.dilations
+    
+    var outDim = [inDims[0]]
+    for i in 0..<strides.count {
+      let dilation: Int = Int(dilations[i])
+      let filterSize: Int = filterDim[i + 1]
+      let inputSize: Int = inDims[i + 1]
+      let padding: Int = Int(paddings[i])
+      let stride: Int = Int(strides[i])
+      let dKernel = dilation * (filterSize - 1) + 1
+      let outputSize = (inputSize + 2 * padding - dKernel) / stride + 1
+      outDim.append(outputSize)
+    }
+    outDim.append(filterDim[0])
+    para.output.dim = Dim.init(inDim: outDim)
+  }
+  
+  func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
+    do {
+      try kernel.compute(commandBuffer: buffer, param: para)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  static func fusionNode() -> Node {
+    let beginNode = Node.init(inType: gConvType)
+    _ = beginNode
+      --> Node.init(inType: gBatchNormType)
+      --> Node.init(inType: gReluType)
+    return beginNode
+  }
+  
+  static func change() -> [String : [(from: String, to: String)]] {
+    return [:]
+  }
+  
+  static func fusionType() -> String {
+    return gConvBnReluType
+  }
+  
+  func delogOutput() {
+    print(" \(type) output: ")
+    print(para.output.metalTexture.toTensor(dim: (n: para.output.padToFourDim[0], c: para.output.padToFourDim[1], h: para.output.padToFourDim[2], w: para.output.padToFourDim[3])).strideArray())
+  }
+  
+}

metal/paddle-mobile/paddle-mobile/Operators/ConvOp.swift

@@ -41,14 +41,8 @@ class ConvParam<P: PrecisionType>: OpParam {
 }
 
 class ConvOp<P: PrecisionType>: Operator<ConvKernel<P>, ConvParam<P>>, Runable, Creator, InferShaperable {
-    required init(device: MTLDevice, opDesc: OpDesc, inScope: Scope) throws {
-        do {
-            try super.init(device: device, opDesc: opDesc, inScope: inScope)
-        } catch let error {
-            throw error
-        }
+  typealias OpType = ConvOp<P>
 
-    }
   func inferShape() {
     let inDims = para.input.dim
     let filterDim = para.filter.dim
@@ -71,7 +65,6 @@ class ConvOp<P: PrecisionType>: Operator<ConvKernel<P>, ConvParam<P>>, Runable,
     para.output.dim = Dim.init(inDim: outDim)
   }
   
-    typealias OpType = ConvOp<P>
   func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
     do {
       try kernel.compute(commandBuffer: buffer, param: para)
@@ -82,7 +75,7 @@ class ConvOp<P: PrecisionType>: Operator<ConvKernel<P>, ConvParam<P>>, Runable,
   
   func delogOutput() {
     print("conv output : ")
-        print(para.output.metalTexture)
-//        let _: Float16? = para.output.metalTexture.logDesc()
+    print(para.output.toTensor().strideArray())
+    //        let _: Float16? = para.output.metalTexture.logDesc()
   }
 }

metal/paddle-mobile/paddle-mobile/Operators/ConvTransposeOp.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class ConvTransposeParam<P: PrecisionType>: ConvParam<P> {
+  typealias ParamPrecisionType = P
+  required init(opDesc: OpDesc, inScope: Scope) throws {
+    do {
+      try super.init(opDesc: opDesc, inScope: inScope)
+    } catch let error {
+      throw error
+    }
+  }
+}
+
+class ConvTransposeOp<P: PrecisionType>: Operator<ConvTransposeKernel<P>, ConvTransposeParam<P>>, Runable, Creator, InferShaperable{
+  
+  typealias OpType = ConvTransposeOp<P>
+  
+  func inferShape() {
+    // para.output.dim = para.input.dim
+  }
+  
+  func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
+    do {
+      try kernel.compute(commandBuffer: buffer, param: para)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  func delogOutput() {
+  
+    print(" \(type) output: ")
+    let padToFourDim = para.output.padToFourDim
+    if para.output.transpose == [0, 1, 2, 3] {
+      let outputArray: [Float32] = para.output.metalTexture.realNHWC(dim: (n: padToFourDim[0], h: padToFourDim[1], w: padToFourDim[2], c: padToFourDim[3]))
+      print(outputArray.strideArray())
+    } else if para.output.transpose == [0, 2, 3, 1] {
+      let output = para.output.metalTexture.toTensor(dim: (n: para.output.tensorDim[0], c: para.output.tensorDim[1], h: para.output.tensorDim[2], w: para.output.tensorDim[3]))
+      print(output.strideArray())
+    } else {
+      print(" not implement")
+    }
+  }
+}

metal/paddle-mobile/paddle-mobile/Operators/DepthwiseConvOp.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class DepthConvOp<P: PrecisionType>: Operator<ConvKernel<P>, ConvParam<P>>, Runable, Creator, InferShaperable {
+
+  typealias OpType = DepthConvOp<P>
+
+  required init(device: MTLDevice, opDesc: OpDesc, inScope: Scope) throws {
+    do {
+      try super.init(device: device, opDesc: opDesc, inScope: inScope)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  func inferShape() {
+    let inDims = para.input.dim
+    let filterDim = para.filter.dim
+    let strides = para.stride
+    let paddings = para.paddings
+    let dilations = para.dilations
+    
+    var outDim = [inDims[0]]
+    for i in 0..<strides.count {
+      let dilation: Int = Int(dilations[i])
+      let filterSize: Int = filterDim[i + 1]
+      let inputSize: Int = inDims[i + 1]
+      let padding: Int = Int(paddings[i])
+      let stride: Int = Int(strides[i])
+      let dKernel = dilation * (filterSize - 1) + 1
+      let outputSize = (inputSize + 2 * padding - dKernel) / stride + 1
+      outDim.append(outputSize)
+    }
+    outDim.append(filterDim[0])
+    para.output.dim = Dim.init(inDim: outDim)
+  }
+  
+  func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
+    do {
+      try kernel.compute(commandBuffer: buffer, param: para)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  func delogOutput() {
+    print(" \(type) output: ")
+    print(para.output.metalTexture.toTensor(dim: (n: para.output.padToFourDim[0], c: para.output.padToFourDim[1], h: para.output.padToFourDim[2], w: para.output.padToFourDim[3])).strideArray())
+  }
+}

metal/paddle-mobile/paddle-mobile/Operators/DwConvBNReluOp.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class DwConvBNReluOp<P: PrecisionType>: Operator<ConvBNReluKernel<P>, ConvBNReluParam<P>>, Runable, Creator, InferShaperable, Fusion{
+  typealias OpType = ConvBNReluOp<P>
+  
+  func inferShape() {
+    let inDims = para.input.dim
+    let filterDim = para.filter.dim
+    let strides = para.stride
+    let paddings = para.paddings
+    let dilations = para.dilations
+    
+    var outDim = [inDims[0]]
+    for i in 0..<strides.count {
+      let dilation: Int = Int(dilations[i])
+      let filterSize: Int = filterDim[i + 1]
+      let inputSize: Int = inDims[i + 1]
+      let padding: Int = Int(paddings[i])
+      let stride: Int = Int(strides[i])
+      let dKernel = dilation * (filterSize - 1) + 1
+      let outputSize = (inputSize + 2 * padding - dKernel) / stride + 1
+      outDim.append(outputSize)
+    }
+    outDim.append(filterDim[0])
+    para.output.dim = Dim.init(inDim: outDim)
+  }
+  
+  func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
+    do {
+      try kernel.compute(commandBuffer: buffer, param: para)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  static func fusionNode() -> Node {
+    let beginNode = Node.init(inType: gDepthConvType)
+    _ = beginNode
+      --> Node.init(inType: gBatchNormType)
+      --> Node.init(inType: gReluType)
+    return beginNode
+  }
+  
+  static func change() -> [String : [(from: String, to: String)]] {
+    return [:]
+  }
+  
+  static func fusionType() -> String {
+    return gDwConvBnReluType
+  }
+  
+  func delogOutput() {
+    print(" \(type) output: ")
+    print(para.output.metalTexture.toTensor(dim: (n: para.output.padToFourDim[0], c: para.output.padToFourDim[1], h: para.output.padToFourDim[2], w: para.output.padToFourDim[3])).strideArray())
+  }
+}

metal/paddle-mobile/paddle-mobile/Operators/ElementwiseAddOp.swift

@@ -18,29 +18,76 @@ class ElementwiseAddParam<P: PrecisionType>: OpParam {
   typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
-            input = try ElementwiseAddParam.inputX(inputs: opDesc.inputs, from: inScope)
-            inputY = try ElementwiseAddParam.inputY(inputs: opDesc.paraInputs, from: inScope)
-            
+      inputX = try ElementwiseAddParam.inputX(inputs: opDesc.inputs, from: inScope)
       output = try ElementwiseAddParam.outputOut(outputs: opDesc.outputs, from: inScope)
       axis = try ElementwiseAddParam.getAttr(key: "axis", attrs: opDesc.attrs)
     } catch let error {
       throw error
     }
+    do {
+      inputY = try ElementwiseAddParam.inputY(inputs: opDesc.paraInputs, from: inScope)
+    } catch _ {
+      let tensorY: Tensor<P> = try ElementwiseAddParam.inputY(inputs: opDesc.paraInputs, from: inScope)
+      let device = inputX.metalTexture!.device
+      inputY = Texture.init(device: device, inDim: tensorY.dim)
+      let value: [P] = Array(UnsafeBufferPointer(start: tensorY.data.pointer, count: tensorY.dim.numel()))
+      inputY.metalTexture = device.tensor2texture(value: value, dim: tensorY.dim.dims, transpose: [0, 1, 2, 3], inComputePrecision: computePrecision)
+    }
+    
+//    required init(device: MTLDevice, param: ElementwiseAddParam<P>) {
+//      param.output.initTexture(device: device, inTranspose: param.inputX.transpose, computePrecision: computePrecision)
+//      if computePrecision == .Float32 {
+//        super.init(device: device, inFunctionName: "elementwise_add")
+//      } else if computePrecision == .Float16 {
+//        super.init(device: device, inFunctionName: "elementwise_add_half")
+//      } else {
+//        fatalError()
+//      }
+//    }
+    
+    var offset = axis
+    if axis == -1 {
+      offset = inputX.tensorDim.cout() - inputY.tensorDim.cout()
+    }
+    for i in 0..<(inputY.tensorDim.cout()) {
+      assert(inputX.tensorDim[offset + i] == inputY.tensorDim[i])
     }
-    let input: Texture<P>
-    let inputY: Tensor<P>
+  }
+  
+  var inputX: Texture<P>
+  var inputY: Texture<P>
   var output: Texture<P>
-    let axis: Int
+  var axis: Int
 }
 
 class ElementwiseAddOp<P: PrecisionType>: Operator<ElementwiseAddKernel<P>, ElementwiseAddParam<P>>, Runable, Creator, InferShaperable{
+  typealias OpType = ElementwiseAddOp<P>
   
   func inferShape() {
-        para.output.dim = para.input.dim
+//    para.output.dim = para.input.dim
   }
   
-    typealias OpType = ElementwiseAddOp<P>
   func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
+    do {
+      try kernel.compute(commandBuffer: buffer, param: para)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  func delogOutput() {
+    print(" \(type) output: ")
+    print(para.output)
+    
+    let padToFourDim = para.output.padToFourDim
+    if para.output.transpose == [0, 1, 2, 3] {
+      let outputArray: [Float32] = para.output.metalTexture.realNHWC(dim: (n: padToFourDim[0], h: padToFourDim[1], w: padToFourDim[2], c: padToFourDim[3]))
+      print(outputArray.strideArray())
+    } else if para.output.transpose == [0, 2, 3, 1] {
+      print(para.output.metalTexture.toTensor(dim: (n: para.output.tensorDim[0], c: para.output.tensorDim[1], h: para.output.tensorDim[2], w: para.output.tensorDim[3])).strideArray())
+    } else {
+      print(" not implement")
+    }
   }
 }
 

metal/paddle-mobile/paddle-mobile/Operators/ElementwiseAddPreluOp.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class ElementwiseAddPreluParam<P: PrecisionType>: OpParam {
+  typealias ParamPrecisionType = P
+  required init(opDesc: OpDesc, inScope: Scope) throws {
+    do {
+      alpha = try ElementwiseAddPreluParam.paramInputAlpha(inputs: opDesc.paraInputs, from: inScope)
+      mode = try ElementwiseAddPreluParam.getAttr(key: "mode", attrs: opDesc.attrs)
+      inputX = try ElementwiseAddPreluParam.inputX(inputs: opDesc.inputs, from: inScope)
+      output = try ElementwiseAddPreluParam.outputOut(outputs: opDesc.outputs, from: inScope)
+      axis = try ElementwiseAddPreluParam.getAttr(key: "axis", attrs: opDesc.attrs)
+    } catch let error {
+      throw error
+    }
+    do {
+      inputY = try ElementwiseAddPreluParam.inputY(inputs: opDesc.paraInputs, from: inScope)
+    } catch _ {
+      let tensorY: Tensor<P> = try ElementwiseAddPreluParam.inputY(inputs: opDesc.paraInputs, from: inScope)
+      let device = inputX.metalTexture!.device
+      inputY = Texture.init(device: device, inDim: tensorY.dim)
+      let value: [P] = Array(UnsafeBufferPointer(start: tensorY.data.pointer, count: tensorY.dim.numel()))
+      inputY.metalTexture = device.tensor2texture(value: value, dim: tensorY.dim.dims, transpose: [0, 1, 2, 3], inComputePrecision: computePrecision)
+    }
+    
+    //    required init(device: MTLDevice, param: ElementwiseAddParam<P>) {
+    //      param.output.initTexture(device: device, inTranspose: param.inputX.transpose, computePrecision: computePrecision)
+    //      if computePrecision == .Float32 {
+    //        super.init(device: device, inFunctionName: "elementwise_add")
+    //      } else if computePrecision == .Float16 {
+    //        super.init(device: device, inFunctionName: "elementwise_add_half")
+    //      } else {
+    //        fatalError()
+    //      }
+    //    }
+    
+    var offset = axis
+    if axis == -1 {
+      offset = inputX.tensorDim.cout() - inputY.tensorDim.cout()
+    }
+    for i in 0..<(inputY.tensorDim.cout()) {
+      assert(inputX.tensorDim[offset + i] == inputY.tensorDim[i])
+    }
+  }
+  
+  let mode: String
+  let alpha: Tensor<P>
+  var inputX: Texture<P>
+  var inputY: Texture<P>
+  var output: Texture<P>
+  var axis: Int
+}
+
+class ElementwiseAddPreluOp<P: PrecisionType>: Operator<ElementwiseAddPreluKernel<P>, ElementwiseAddPreluParam<P>>, Runable, Creator, InferShaperable, Fusion{
+  static func fusionNode() -> Node {
+    let beginNode = Node.init(inType: gElementwiseAddType)
+    _ = beginNode
+      --> Node.init(inType: gPreluType)
+    return beginNode
+  }
+  
+  static func change() -> [String : [(from: String, to: String)]] {
+    return [:]
+  }
+  
+  static func fusionType() -> String {
+    return gElementwiseAddPreluType
+  }
+  
+  typealias OpType = ElementwiseAddPreluOp<P>
+  
+  func inferShape() {
+    //    para.output.dim = para.input.dim
+  }
+  
+  func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
+    do {
+      try kernel.compute(commandBuffer: buffer, param: para)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  
+  
+  func delogOutput() {
+    print(" \(type) output: ")
+    print(para.output)
+    
+    let padToFourDim = para.output.padToFourDim
+    if para.output.transpose == [0, 1, 2, 3] {
+      let outputArray: [Float32] = para.output.metalTexture.realNHWC(dim: (n: padToFourDim[0], h: padToFourDim[1], w: padToFourDim[2], c: padToFourDim[3]))
+      print(outputArray.strideArray())
+    } else if para.output.transpose == [0, 2, 3, 1] {
+      print(para.output.metalTexture.toTensor(dim: (n: para.output.tensorDim[0], c: para.output.tensorDim[1], h: para.output.tensorDim[2], w: para.output.tensorDim[3])).strideArray())
+    } else {
+      print(" not implement")
+    }
+  }
+}
+
+
+
+
+
+

metal/paddle-mobile/paddle-mobile/Operators/FeedOp.swift

@@ -50,19 +50,18 @@ class FeedOp<P: PrecisionType>: Operator<Texture2DTo2DArrayKernel<P>, FeedParam<
       throw error
     }
     
-//        let resizeKernel = ResizeKernel<P>.init(device: device)
-//        let resizeParam = ResizeParam.init(input: para.input.mtlTexture, output: para.output.metalTexture, expectDim: para.input.expectDim)
-//        do {
-//            try resizeKernel.compute(commandBuffer: buffer, param: resizeParam)
-//        } catch let error {
-//            throw error
-//        }
+    //        let resizeKernel = ResizeKernel<P>.init(device: device)
+    //        let resizeParam = ResizeParam.init(input: para.input.mtlTexture, output: para.output.metalTexture, expectDim: para.input.expectDim)
+    //        do {
+    //            try resizeKernel.compute(commandBuffer: buffer, param: resizeParam)
+    //        } catch let error {
+    //            throw error
+    //        }
   }
   
   func delogOutput() {
-//        para.input.mtlTexture.logDesc()
-//        let _: P? = para.input.mtlTexture.logDesc(header: "feed input: ", stridable: true)
-//        let _: P? = para.output.metalTexture.logDesc(header: "feed output: ", stridable: false)
+    print(" \(type) output: ")
+    print(para.output.metalTexture.toTensor(dim: (n: para.output.padToFourDim[0], c: para.output.padToFourDim[1], h: para.output.padToFourDim[2], w: para.output.padToFourDim[3])).strideArray())
   }
 }
 

metal/paddle-mobile/paddle-mobile/Operators/FetchOp.swift

@@ -15,14 +15,15 @@
 import Foundation
 
 class FetchParam<P: PrecisionType>: OpParam{
-    var output: Texture<P>
+  var output: FetchHolder
   let input: Texture<P>
   let scope: Scope
   required init(opDesc: OpDesc, inScope: Scope) throws {
     scope = inScope
     do {
       input = try FetchParam.inputX(inputs: opDesc.inputs, from: inScope)
-            output = input
+      output = FetchHolder.init(inCapacity: input.numel(), inDim: input.tensorDim.dims)
+      scope.setOutput(output: output)
     } catch let error {
       throw error
     }
@@ -34,21 +35,53 @@ class FetchParam<P: PrecisionType>: OpParam{
 class FetchKernel<P: PrecisionType>: Kernel, Computable {
   
   func compute(commandBuffer: MTLCommandBuffer, param: FetchParam<P>) throws {
+    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+      throw PaddleMobileError.predictError(message: " encode is nil")
+    }
+    encoder.setTexture(param.input.metalTexture, index: 0)
+    encoder.setBuffer(param.output.resultBuffer!, offset: 0, index: 0)
+    encoder.dispatch(computePipline: pipline, outTexture: param.input.metalTexture)
+    encoder.endEncoding()
   }
   
   required init(device: MTLDevice, param: FetchParam<P>) {
-        super.init(device: device, inFunctionName: "texture2d_to_2d_array")
+    param.output.initBuffer(device: device)
+    if computePrecision == .Float16 {
+      if param.input.transpose == [0, 2, 3, 1] {
+        super.init(device: device, inFunctionName: "fetch_half")
+      } else {
+//        fatalError(" not support ")
+        super.init(device: device, inFunctionName: "fetch_placeholder_half")
+        print(" not support ")
+      }
+    } else if computePrecision == .Float32 {
+      if param.input.transpose == [0, 2, 3, 1] {
+        super.init(device: device, inFunctionName: "fetch")
+      } else {
+        print(" not support ")
+        super.init(device: device, inFunctionName: "fetch_placeholder")
+//        fatalError(" not support ")        
+      }
+    } else {
+      fatalError(" not support ")
+    }
   }
 }
 
-class FetchOp<P: PrecisionType>: Operator< FetchKernel<P>, FetchParam<P>>, Runable, Creator, InferShaperable{
+class FetchOp<P: PrecisionType>: Operator< FetchKernel<P>, FetchParam<P>>, Runable, Creator, InferShaperable {
+  
+  typealias OpType = FetchOp<P>
+
   func inferShape() {
     print(para.input.dim)
   }
   
-    typealias OpType = FetchOp<P>
   func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
-        scope.setOutput(output: para.output)
+    do {
+      try kernel.compute(commandBuffer: buffer, param: para)
+    } catch let error {
+      throw error
+    }
   }
 }
 

metal/paddle-mobile/paddle-mobile/Operators/FlattenOp.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class FlattenParam<P: PrecisionType>: OpParam {
+  typealias ParamPrecisionType = P
+  required init(opDesc: OpDesc, inScope: Scope) throws {
+    do {
+      input = try FlattenParam.inputX(inputs: opDesc.inputs, from: inScope)
+      output = try FlattenParam.outputOut(outputs: opDesc.outputs, from: inScope)
+      axis = try FlattenParam.getAttr(key: "axis", attrs: opDesc.attrs)
+    } catch let error {
+      throw error
+    }
+  }
+  let input: Texture<P>
+  var output: Texture<P>
+  let axis: Int
+}
+
+
+class FlattenOp<P: PrecisionType>: Operator<FlattenKernel<P>, FlattenParam<P>>, Runable, Creator, InferShaperable{
+  
+  typealias OpType = FlattenOp<P>
+
+  func inferShape() {
+    //        para.output.dim = para.input.dim
+  }
+  
+  func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
+    do {
+      try kernel.compute(commandBuffer: buffer, param: para)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  func delogOutput() {
+    print(" \(type) output: ")
+    let device = para.output.metalTexture!.device
+    let outputArray: [Float32] = device.texture2tensor(texture: para.output.metalTexture, dim: para.output.tensorDim.dims, transpose: para.output.transpose)
+    print(outputArray.strideArray())
+  }
+  
+}
+
+
+
+
+
+

metal/paddle-mobile/paddle-mobile/Operators/Kernels/Base/Kernel.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Metal
+import Foundation
+
+public protocol TestParam {
+}
+
+public protocol Testable {
+  associatedtype TestParamType: TestParam
+  func test(commandBuffer: MTLCommandBuffer, param: TestParamType)
+  init(device: MTLDevice, testParam: TestParamType)
+}
+
+
+protocol Computable {
+  associatedtype ParamType: OpParam
+  func compute(commandBuffer: MTLCommandBuffer, param: ParamType) throws
+  init(device: MTLDevice, param: ParamType)
+}
+
+protocol KernelProtocol {
+  var pipline: MTLComputePipelineState { get set }
+  var functionName: String { get set }
+  
+}
+
+open class Kernel {
+  let pipline: MTLComputePipelineState
+  let functionName: String
+  public init(device: MTLDevice, inFunctionName: String, usePaddleMobileLib: Bool = true) {
+    pipline = device.pipeLine(funcName: inFunctionName, inPaddleMobileLib: usePaddleMobileLib)
+    functionName = inFunctionName
+  }
+}
+
+open class CusomKernel: Kernel {
+  public struct Shape {
+    public let width: Int
+    public let height: Int
+    public let channel: Int
+    public init(inWidth: Int, inHeight: Int, inChannel: Int){
+      width = inWidth
+      height = inHeight
+      channel = inChannel
+    }
+  }
+  public let outputTexture: MTLTexture
+  public init(device: MTLDevice, inFunctionName: String, outputDim: Shape, usePaddleMobileLib: Bool = false) {
+    let textureDesc = MTLTextureDescriptor.init()
+    textureDesc.textureType = .type2D
+    textureDesc.width = outputDim.width
+    textureDesc.height = outputDim.height
+    textureDesc.depth = (outputDim.channel + 3) / 4
+    
+    if computePrecision == .Float16 {
+      textureDesc.pixelFormat = .rgba16Float
+    } else if computePrecision == .Float32 {
+      textureDesc.pixelFormat = .rgba32Float
+    } else {
+      fatalError()
+    }
+    
+    textureDesc.usage = [.shaderRead, .shaderWrite]
+    textureDesc.storageMode = .shared
+    outputTexture = device.makeTexture(descriptor: textureDesc) ?! " make texture error "
+    
+    super.init(device: device, inFunctionName: inFunctionName, usePaddleMobileLib: usePaddleMobileLib)
+  }
+  
+  public func compute(inputTexuture: MTLTexture, commandBuffer: MTLCommandBuffer) throws {
+    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+      throw PaddleMobileError.predictError(message: " encode is nil")
+    }
+    encoder.setTexture(inputTexuture, index: 0)
+    encoder.setTexture(outputTexture, index: 1)
+    encoder.dispatch(computePipline: pipline, outTexture: outputTexture)
+    encoder.endEncoding()
+  }
+  
+}
+

metal/paddle-mobile/paddle-mobile/Operators/Kernels/BatchNormKernel.swift

@@ -15,40 +15,27 @@
 import Foundation
 
 class BatchNormKernel<P: PrecisionType>: Kernel, Computable {
-    var newScale: MTLBuffer
-    var newBias: MTLBuffer
-    
   required init(device: MTLDevice, param: BatchNormParam<P>) {
-        guard let newScale = device.makeBuffer(length: param.inputScale.buffer.length) else {
-            fatalError()
+    let count = param.variance.dim.numel()
+    let varianceP = param.variance.data.pointer
+    let meanP = param.mean.data.pointer
+    let scaleP = param.scale.data.pointer
+    let biasP = param.bias.data.pointer
+    for i in 0..<count {
+      let invStd = P(1 / (Float32(varianceP[i]) + param.epsilon).squareRoot())
+      biasP[i] = biasP[i] - meanP[i] * invStd * scaleP[i]
+      scaleP[i] = invStd * scaleP[i]
     }
-        guard let newBias = device.makeBuffer(length: param.inputBias.buffer.length) else {
-            fatalError()
-        }
-        self.newScale = newScale
-        self.newBias = newBias
 
+    param.bias.initBuffer(device: device, precision: computePrecision)
+    param.scale.initBuffer(device: device, precision: computePrecision)
+    param.output.initTexture(device: device, inTranspose: param.input.transpose, computePrecision: computePrecision)
+    if computePrecision == .Float32 {
       super.init(device: device, inFunctionName: "batchnorm")
-        
-        let varianceBuffer : MTLBuffer = param.inputVariance.buffer
-        
-        var invStd: [Float32] = Array(repeating: 0, count: varianceBuffer.length)
-        let varianceContents = varianceBuffer.contents().assumingMemoryBound(to: P.self)
-        for i in 0..<(varianceBuffer.length / MemoryLayout<P>.stride) {
-            invStd[i] = 1 / (Float32(varianceContents[i]) + param.epsilon).squareRoot()
-        }
-        
-        let newScaleContents = newScale.contents().assumingMemoryBound(to: P.self)
-        let newBiasContents = newBias.contents().assumingMemoryBound(to: P.self)
-        let scale : MTLBuffer = param.inputScale.buffer
-        let scaleContents = scale.contents().assumingMemoryBound(to: P.self)
-        let bias : MTLBuffer = param.inputBias.buffer
-        let biasContents = bias.contents().assumingMemoryBound(to: P.self)
-        let meanContents = param.inputMean.buffer.contents().assumingMemoryBound(to: P.self)
-        
-        for i in 0..<(newScale.length / MemoryLayout<P>.stride) {
-            newScaleContents[i] = P(invStd[i] * Float32(scaleContents[i]))
-            newBiasContents[i] = P(Float32(biasContents[i]) - Float32(meanContents[i]) * invStd[i] * Float32(scaleContents[i]))
+    } else if computePrecision == .Float16 {
+      super.init(device: device, inFunctionName: "batchnorm_half")
+    } else {
+      fatalError()
     }
   }
   
@@ -56,11 +43,10 @@ class BatchNormKernel<P: PrecisionType>: Kernel, Computable {
     guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
       throw PaddleMobileError.predictError(message: " encoder is nil")
     }
-        print("BatchNorm compute")
     encoder.setTexture(param.input.metalTexture, index: 0)
     encoder.setTexture(param.output.metalTexture, index: 1)
-        encoder.setBuffer(newScale, offset: 0, index: 0)
-        encoder.setBuffer(newBias, offset: 0, index: 1)
+    encoder.setBuffer(param.scale.buffer, offset: 0, index: 0)
+    encoder.setBuffer(param.bias.buffer, offset: 0, index: 1)
     encoder.dispatch(computePipline: pipline, outTexture: param.output.metalTexture)
     encoder.endEncoding()
   }

metal/paddle-mobile/paddle-mobile/Operators/Kernels/BatchNormReluKernel.swift

+//
+//  BatchNormRelu.swift
+//  paddle-mobile
+//
+//  Created by zhangxinjun on 2018/8/23.
+//  Copyright © 2018年 orange. All rights reserved.
+//
+
+import Foundation
+
+
+class BatchNormReluParam<P: PrecisionType>: BatchNormParam<P> {
+    
+}
+
+class BatchNormReluKernel<P: PrecisionType>: Kernel, Computable{
+    
+    
+    typealias ParamType = BatchNormReluParam<P>
+    var newScale: MTLBuffer
+    var newBias: MTLBuffer
+    
+    required init(device: MTLDevice, testParam: BatchNormReluTestParam) {
+        
+        newScale = testParam.newScaleBuffer
+        newBias = testParam.newBiaseBuffer
+        
+        super.init(device: device, inFunctionName: "batch_norm_relu_3x3")
+    }
+    
+    required init(device: MTLDevice, param: BatchNormReluParam<P>) {
+        guard let newScale = device.makeBuffer(length: param.inputScale.buffer.length) else {
+            fatalError()
+        }
+        guard let newBias = device.makeBuffer(length: param.inputBias.buffer.length) else {
+            fatalError()
+        }
+        self.newScale = newScale
+        self.newBias = newBias
+        
+        super.init(device: device, inFunctionName: "batch_norm_relu_3x3")
+        
+        
+        let varianceBuffer : MTLBuffer = param.inputVariance.buffer
+        
+        var invStd: [Float32] = Array(repeating: 0, count: varianceBuffer.length)
+        let varianceContents = varianceBuffer.contents().assumingMemoryBound(to: P.self)
+        for i in 0..<(varianceBuffer.length / MemoryLayout<P>.stride) {
+            invStd[i] = 1 / (Float32(varianceContents[i]) + param.epsilon).squareRoot()
+        }
+        
+        let newScaleContents = newScale.contents().assumingMemoryBound(to: P.self)
+        let newBiasContents = newBias.contents().assumingMemoryBound(to: P.self)
+        let scale : MTLBuffer = param.inputScale.buffer
+        let scaleContents = scale.contents().assumingMemoryBound(to: P.self)
+        let bias : MTLBuffer = param.inputBias.buffer
+        let biasContents = bias.contents().assumingMemoryBound(to: P.self)
+        let meanContents = param.inputMean.buffer.contents().assumingMemoryBound(to: P.self)
+        
+        for i in 0..<(newScale.length / MemoryLayout<P>.stride) {
+            newScaleContents[i] = P(invStd[i] * Float32(scaleContents[i]))
+            newBiasContents[i] = P(Float32(biasContents[i]) - Float32(meanContents[i]) * invStd[i] * Float32(scaleContents[i]))
+        }
+    }
+    
+    func compute(commandBuffer: MTLCommandBuffer, param: BatchNormReluParam<P>) throws {
+        guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+            fatalError()
+        }
+        encoder.setTexture(param.input as? MTLTexture, index: 0)
+        encoder.setTexture(param.output as? MTLTexture, index: 1)
+        encoder.setBuffer(newScale, offset: 0, index: 1)
+        encoder.setBuffer(newBias, offset: 0, index: 1)
+        encoder.dispatch(computePipline: pipline, outTexture: param.output as! MTLTexture)
+        encoder.endEncoding()
+    }
+    
+    func testCompute(commandBuffer: MTLCommandBuffer, testParam: BatchNormReluTestParam) throws {
+        guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+            fatalError()
+        }
+        encoder.setTexture(testParam.inputTexture, index: 0)
+        encoder.setTexture(testParam.outputTexture, index: 1)
+        encoder.setBuffer(newScale, offset: 0, index: 0)
+        encoder.setBuffer(newBias, offset: 0, index: 1)
+        encoder.dispatch(computePipline: pipline, outTexture: testParam.outputTexture)
+        encoder.endEncoding()
+    }
+    
+    
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/BilinearInterpKernel.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+struct BilinearInterpMetalParam {
+  var ratio_h: Float32
+  var ratio_w: Float32
+}
+
+class BilinearInterpKernel<P: PrecisionType>: Kernel, Computable{
+  func compute(commandBuffer: MTLCommandBuffer, param: BilinearInterpParam<P>) throws {
+    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+      throw PaddleMobileError.predictError(message: " encode is nil")
+    }
+    
+    encoder.setTexture(param.input.metalTexture, index: 0)
+    encoder.setTexture(param.output.metalTexture, index: 1)
+    var ratio_h: Float32 = 0
+    var ratio_w: Float32 = 0
+    if param.output.tensorDim.dims[2] > 1 {
+      ratio_h = Float32(param.input.tensorDim.dims[2]-1) / Float32(param.output.tensorDim.dims[2]-1)
+    }
+    if param.output.tensorDim.dims[3] > 1 {
+      ratio_w = Float32(param.input.tensorDim.dims[3]-1) / Float32(param.output.tensorDim.dims[3]-1)
+    }
+    var p = BilinearInterpMetalParam.init(ratio_h: ratio_h, ratio_w: ratio_w)
+    encoder.setBytes(&p, length: MemoryLayout<BilinearInterpMetalParam>.size, index: 0)
+    encoder.dispatch(computePipline: pipline, outTexture: param.output.metalTexture)
+    encoder.endEncoding()
+  }
+  
+  required init(device: MTLDevice, param: BilinearInterpParam<P>) {
+    param.output.initTexture(device: device, inTranspose: param.input.transpose, computePrecision: computePrecision)
+    if computePrecision == .Float32 {
+      super.init(device: device, inFunctionName: "bilinear_interp_float")
+    } else if computePrecision == .Float16 {
+      super.init(device: device, inFunctionName: "bilinear_interp_half")
+    } else {
+      fatalError()
+    }
+  }
+  
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/BoxcoderKernel.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+struct BoxcoderMetalParam {
+}
+
+class BoxcoderKernel<P: PrecisionType>: Kernel, Computable{
+  func compute(commandBuffer: MTLCommandBuffer, param: BoxcoderParam<P>) throws {
+    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+      throw PaddleMobileError.predictError(message: " encode is nil")
+    }
+    encoder.setTexture(param.priorBox.metalTexture, index: 0)
+    encoder.setTexture(param.priorBoxVar.metalTexture, index: 1)
+    encoder.setTexture(param.targetBox.metalTexture, index: 2)
+    encoder.setTexture(param.output.metalTexture, index: 3)
+    var bmp = BoxcoderMetalParam.init()
+    encoder.setBytes(&bmp, length: MemoryLayout<BoxcoderMetalParam>.size, index: 0)
+    encoder.dispatch(computePipline: pipline, outTexture: param.output.metalTexture)
+    encoder.endEncoding()
+  }
+  
+  required init(device: MTLDevice, param: BoxcoderParam<P>) {
+    param.output.initTexture(device: device, inTranspose: [0, 3, 1, 2], computePrecision: computePrecision)
+    if computePrecision == .Float32 {
+      super.init(device: device, inFunctionName: "boxcoder_float")
+    } else if computePrecision == .Float16 {
+      super.init(device: device, inFunctionName: "boxcoder_half")
+    } else {
+      fatalError()
+    }
+  }
+  
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/CNNConvKernel.swift

+//
+//  CNNConvKernel.swift
+//  paddle-mobile
+//
+
+import Foundation
+import Metal
+import Accelerate
+import MetalPerformanceShaders
+
+@available(iOS 10.0, *)
+class WeightsDataSource: NSObject, MPSCNNConvolutionDataSource  {
+    
+    let desc: MPSCNNConvolutionDescriptor
+    let weight:UnsafeMutableRawPointer
+    let bias:UnsafeMutablePointer<Float>
+    
+    
+    
+    init(inDesc: MPSCNNConvolutionDescriptor, inWeight: UnsafeMutableRawPointer, inBias: UnsafeMutablePointer<Float>) {
+        desc = inDesc
+        weight = inWeight
+        bias = inBias
+    }
+    
+    
+    func dataType() -> MPSDataType {
+        return .float32
+    }
+    
+    func descriptor() -> MPSCNNConvolutionDescriptor {
+        return desc
+    }
+    
+    func weights() -> UnsafeMutableRawPointer {
+        return self.weight
+    }
+    
+    func biasTerms() -> UnsafeMutablePointer<Float>? {
+        return self.bias
+    }
+    
+    func load() -> Bool {
+        return true
+    }
+    
+    func purge() {
+    }
+    
+    func label() -> String? {
+        return "Conv"
+    }
+    
+    
+}
+
+@available(iOS 10.0, *)
+class CNNConvParam<P: PrecisionType>: OpParam{
+    
+    typealias ParamPrecisionType = P
+    required init(opDesc: OpDesc, inScope: Scope) throws {
+        do {
+            filter = try CNNConvParam.inputFilter(paraInputs: opDesc.paraInputs, from: inScope)
+            input = try CNNConvParam.input(inputs: opDesc.inputs, from: inScope)
+            output = try CNNConvParam.outputOut(outputs: opDesc.outputs, from: inScope)
+            stride = try CNNConvParam.getAttr(key: "strides", attrs: opDesc.attrs)
+            paddings = try CNNConvParam.getAttr(key: "paddings", attrs: opDesc.attrs)
+            // 暂时不用关心
+            dilations = try CNNConvParam.getAttr(key: "dilations", attrs: opDesc.attrs)
+            // 暂时不用关心
+            groups = try CNNConvParam.getAttr(key: "groups", attrs: opDesc.attrs)
+            
+            variance = try CNNConvParam.inputVariance(inputs: opDesc.paraInputs, from: inScope)
+            // bias
+            y = try CNNConvParam.inputY(inputs: opDesc.paraInputs, from: inScope)
+        } catch let error {
+            throw error
+        }
+    }
+    
+    var input: Texture<P>
+    let variance: Tensor<ParamPrecisionType>
+    let y: Tensor<ParamPrecisionType>
+    let filter: Tensor<ParamPrecisionType>
+    var output: Texture<P>
+    let stride: [Int32]
+    let paddings: [Int32]
+    let dilations: [Int32]
+    let groups: Int
+}
+
+@available(iOS 10.0, *)
+class CNNConvKernel<P: PrecisionType>: Kernel, Computable {
+    
+    typealias ParamType = CNNConvParam<P>
+    
+    var mpsImageCreator: MpsImageCreator<P>?
+    var activation:MPSCNNNeuron?
+    var conv:MPSCNNConvolution?
+    var weightDataSource:WeightsDataSource?
+    var param: CNNConvParam<P>?
+    var device: MTLDevice?
+    
+    
+    required init(device:MTLDevice, testParam:CNNMPSConvTestParam) {
+        self.device = device
+        
+        let desc = MPSCNNConvolutionDescriptor(kernelWidth: testParam.filterSize.width, kernelHeight: testParam.filterSize.height, inputFeatureChannels: testParam.filterSize.channel, outputFeatureChannels: testParam.filterSize.channel, neuronFilter: activation)
+        
+        desc.strideInPixelsX = Int(testParam.metalParam.offsetX)
+        desc.strideInPixelsY = Int(testParam.metalParam.offsetY)
+        
+        
+        weightDataSource = WeightsDataSource(inDesc: desc, inWeight:testParam.filterPointer, inBias:testParam.biasePointer)
+        
+        if #available(iOS 11.0, *) {
+            conv = MPSCNNConvolution(device: self.device!, weights: weightDataSource!)
+        } else {
+            // Fallback on earlier versions
+        }
+        
+        super.init(device: device, inFunctionName: "")
+    }
+
+    required init(device:MTLDevice, param:CNNConvParam<P>) {
+        
+        self.device = device
+
+        let inChannels: Int
+        let outChannels: Int
+        
+        if param.y.dim.cout() == 4 {
+            inChannels = (param.y.dim[3])
+            outChannels = inChannels
+        } else {
+            inChannels = 0
+            outChannels = inChannels
+        }
+        
+        let desc = MPSCNNConvolutionDescriptor(kernelWidth: param.filter.width, kernelHeight: param.filter.height, inputFeatureChannels: inChannels, outputFeatureChannels: outChannels, neuronFilter: activation)
+        
+        desc.strideInPixelsX = Int(param.stride[0])
+        desc.strideInPixelsY = Int(param.stride[1])
+        
+        
+        weightDataSource = WeightsDataSource(inDesc: desc, inWeight:param.filter.data.pointer as! UnsafeMutablePointer<Float>, inBias: param.y.data.pointer as! UnsafeMutablePointer<Float>)
+        
+        if #available(iOS 11.0, *) {
+            conv = MPSCNNConvolution(device: self.device!, weights: weightDataSource!)
+        } else {
+            // Fallback on earlier versions
+        }
+        
+        super.init(device: device, inFunctionName: "")
+    }
+
+    func compute(commandBuffer: MTLCommandBuffer, param: CNNConvParam<P>) throws {
+        let inputImage:MPSImage = (mpsImageCreator?.createMPSImage(device: device!))!
+        var outputImage = (mpsImageCreator?.createMPSImage(device: device!))!
+        
+        // 运算conv和add两个步骤，add用了bias偏差做为参数，被Metal API进行调用
+        conv?.encode(commandBuffer: commandBuffer, sourceImage: inputImage, destinationImage: outputImage)
+        
+        param.input = outputImage.texture as! Texture<P>
+    }
+    
+    func testCompute(commandBuffer: MTLCommandBuffer, testParam: CNNMPSConvTestParam) throws {
+        let inputImage:MPSImage = (mpsImageCreator?.createMPSImage(device: device!))!
+        var outputImage = (mpsImageCreator?.createMPSImage(device: device!))!
+        
+        // 运算conv和add两个步骤，add用了bias偏差做为参数，被Metal API进行调用
+        conv?.encode(commandBuffer: commandBuffer, sourceImage: inputImage, destinationImage: outputImage)
+        
+        testParam.outputTexture = outputImage.texture
+    }
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/Concat.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class ConcatKernel<P: PrecisionType>: Kernel, Computable{
+    func compute(commandBuffer: MTLCommandBuffer, param: ConcatParam<P>) throws {
+        guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+            throw PaddleMobileError.predictError(message: " encoder is nil")
+        }
+        encoder.setTexture(param.input.metalTexture, index: 0)
+        encoder.setTexture(param.output.metalTexture, index: 1)
+        encoder.dispatch(computePipline: pipline, outTexture: param.output.metalTexture)
+        encoder.endEncoding()
+    }
+    
+    required init(device: MTLDevice, param: ConcatParam<P>) {
+        super.init(device: device, inFunctionName: "concat")
+    }
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/ConcatKernel.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+struct ConcatTestParam: TestParam {
+  var input: [MTLTexture]
+  var output: MTLTexture
+  var dims: [[Int]]
+  var axis: Int
+  var odim: [Int]
+}
+
+struct ConcatMetalParam {
+  var odim: (Int32, Int32, Int32, Int32) = (1, 1, 1, 1)
+  var axis: Int32 = 0
+  var offset: Int32 = 0
+  var trans: (Int32, Int32, Int32, Int32) = (0, 1, 2, 3)
+  var vdim: (Int32, Int32, Int32, Int32, Int32, Int32) = (0, 0, 0, 0, 0, 0)
+}
+
+class ConcatKernel<P: PrecisionType>: Kernel, Computable{
+  var v = "normal"
+  var pm = ConcatMetalParam.init()
+  func compute(commandBuffer: MTLCommandBuffer, param: ConcatParam<P>) throws {
+    
+    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+      throw PaddleMobileError.predictError(message: " encode is nil")
+    }
+    let num = param.input.count
+    for i in 0..<num {
+      encoder.setTexture(param.input[i].metalTexture, index: i)
+    }
+    encoder.setTexture(param.output.metalTexture, index: num)
+    if v == "normal" {
+      encoder.setTexture(param.output.metalTexture, index: num + 1)
+    }
+    encoder.setBytes(&pm, length: MemoryLayout<ConcatMetalParam>.size, index: 0)
+    encoder.dispatch(computePipline: pipline, outTexture: param.output.metalTexture)
+    encoder.endEncoding()
+  }
+
+  required init(device: MTLDevice, param: ConcatParam<P>) {
+    param.output.initTexture(device: device, inTranspose: param.transpose, computePrecision: computePrecision)
+    let orank = param.output.tensorDim.cout()
+    let num = param.input.count
+    assert(num <= 6)
+    var axis = 4 - param.output.tensorDim.cout() + param.axis
+    for i in 0..<4 {
+      if param.transpose[i] == axis {
+        axis = i
+        break
+      }
+    }
+    pm.axis = Int32(axis)
+    pm.odim = (Int32(param.output.dim[0]), Int32(param.output.dim[1]), Int32(param.output.dim[2]), Int32(param.output.dim[3]))
+    pm.trans = (Int32(param.output.transpose[0]), Int32(param.output.transpose[1]), Int32(param.output.transpose[2]), Int32(param.output.transpose[3]))
+    var vdim: [Int] = [0, 0, 0, 0, 0, 0]
+    for i in 0..<num {
+      vdim[i] = param.input[i].dim[axis]
+    }
+    if orank == 4 {
+      if axis == 1 {
+        v = "y"
+      } else if axis == 2 {
+        v = "x"
+      } else {
+        if (param.output.dim[0] == 1) && axis == 3 {
+          var vz = true
+          for i in 0..<num {
+            if vdim[i] % 4 != 0 {
+              vz = false
+              break
+            }
+          }
+          if vz {
+            v = "z"
+            for i in 0..<num {
+              vdim[i] = vdim[i] / 4
+            }
+          }
+        }
+      }
+    } else if orank == 3 {
+      if axis == 2 {
+        v = "y"
+      } else if axis == 3 {
+        v = "x"
+      } else if axis == 1 {
+        var vz = true
+        for i in 0..<num {
+          if vdim[i] % 4 != 0 {
+            vz = false
+            break
+          }
+        }
+        if vz {
+          v = "z"
+          for i in 0..<num {
+            vdim[i] = vdim[i] / 4
+          }
+        }
+      }
+    } else {
+      if axis == 2 {
+        v = "y"
+      } else if axis == 3 {
+        var vx = true
+        for i in 0..<num {
+          if vdim[i] % 4 != 0 {
+            vx = false
+            break
+          }
+        }
+        if vx {
+          v = "x"
+          for i in 0..<num {
+            vdim[i] = vdim[i] / 4
+          }
+        }
+      }
+    }
+    pm.vdim = (Int32(vdim[0]), Int32(vdim[1]), Int32(vdim[2]), Int32(vdim[3]), Int32(vdim[4]), Int32(vdim[5]))
+    if computePrecision == .Float32 {
+      super.init(device: device, inFunctionName: "concat_\(orank)_\(num)_\(v)_float")
+    } else if computePrecision == .Float16 {
+      super.init(device: device, inFunctionName: "concat_\(orank)_\(num)_\(v)_half")
+    } else {
+      fatalError()
+    }
+  }
+  
+  required init(device: MTLDevice, testParam: ConcatTestParam) {
+    super.init(device: device, inFunctionName: "concat")
+  }
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/ConvAddAddPreluKernel.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class ConvAddAddPreluKernel<P: PrecisionType>: Kernel, Computable {
+  var metalParam: MetalConvParam!
+  required init(device: MTLDevice, param: ConvAddAddPreluParam<P>) {
+    param.output.initTexture(device: device, inTranspose: [0, 2, 3, 1], computePrecision: computePrecision)
+    param.filter.initBuffer(device: device, precision: computePrecision)
+    param.y.initBuffer(device: device, precision: computePrecision)
+    param.alpha.initBuffer(device: device, precision: computePrecision)
+    
+    if computePrecision == .Float16 {
+      if param.filter.width == 1 && param.filter.height == 1 {
+        if param.mode == "channel" {
+          super.init(device: device, inFunctionName: "conv_add_1x1_prelu_channel_half")
+        } else if param.mode == "element" {
+          super.init(device: device, inFunctionName: "conv_add_1x1_prelu_element_half")
+        } else {
+          super.init(device: device, inFunctionName: "conv_add_1x1_prelu_other_half")
+        }
+        
+      } else if param.filter.channel == 1 {
+        if param.mode == "channel" {
+          super.init(device: device, inFunctionName: "depthwise_conv_add_3x3_prelu_channel_half")
+        } else if param.mode == "element" {
+          super.init(device: device, inFunctionName: "depthwise_conv_add_3x3_prelu_element_half")
+        } else {
+          super.init(device: device, inFunctionName: "depthwise_conv_add_3x3_prelu_other_half")
+        }
+      } else if param.filter.width == 3 && param.filter.height == 3 {
+        if param.mode == "channel" {
+          super.init(device: device, inFunctionName: "conv_add_3x3_prelu_channel_half")
+        } else if param.mode == "element" {
+          super.init(device: device, inFunctionName: "conv_add_3x3_prelu_element_half")
+        } else {
+          super.init(device: device, inFunctionName: "conv_add_3x3_prelu_other_half")
+        }
+        
+      } else if param.filter.width == 1 && param.filter.height == 5 {
+        if param.mode == "channel" {
+          super.init(device: device, inFunctionName: "conv_add_5x1_prelu_channel_half")
+        } else if param.mode == "element" {
+          super.init(device: device, inFunctionName: "conv_add_5x1_prelu_element_half")
+        } else {
+          super.init(device: device, inFunctionName: "conv_add_5x1_prelu_other_half")
+        }
+      } else if param.filter.width == 5 && param.filter.height == 1 {
+        if param.mode == "channel" {
+          super.init(device: device, inFunctionName: "conv_add_1x5_prelu_channel_half")
+        } else if param.mode == "element" {
+          super.init(device: device, inFunctionName: "conv_add_1x5_prelu_element_half")
+        } else {
+          super.init(device: device, inFunctionName: "conv_add_1x5_prelu_other_half")
+        }
+      } else {
+        fatalError(" unsupport yet ")
+      }
+    } else if computePrecision == .Float32 {
+      if param.filter.width == 1 && param.filter.height == 1 {
+        if param.mode == "channel" {
+          super.init(device: device, inFunctionName: "conv_add_1x1_prelu_channel_float")
+        } else if param.mode == "element" {
+          super.init(device: device, inFunctionName: "conv_add_1x1_prelu_element_float")
+        } else {
+          super.init(device: device, inFunctionName: "conv_add_1x1_prelu_other_float")
+        }
+      } else if param.filter.channel == 1 {
+        if param.mode == "channel" {
+          super.init(device: device, inFunctionName: "depthwise_conv_add_3x3_prelu_channel_float")
+        } else if param.mode == "element" {
+          super.init(device: device, inFunctionName: "depthwise_conv_add_3x3_prelu_element_float")
+        } else {
+          super.init(device: device, inFunctionName: "depthwise_conv_add_3x3_prelu_other_float")
+        }
+      } else if param.filter.width == 3 && param.filter.height == 3 {
+        if param.mode == "channel" {
+          super.init(device: device, inFunctionName: "conv_add_3x3_prelu_channel_float")
+        } else if param.mode == "element" {
+          super.init(device: device, inFunctionName: "conv_add_3x3_prelu_element_float")
+        } else {
+          super.init(device: device, inFunctionName: "conv_add_3x3_prelu_other_float")
+        }
+        
+      } else if param.filter.width == 1 && param.filter.height == 5 {
+        if param.mode == "channel" {
+          super.init(device: device, inFunctionName: "conv_add_5x1_prelu_channel_float")
+        } else if param.mode == "element" {
+          super.init(device: device, inFunctionName: "conv_add_5x1_prelu_element_float")
+        } else {
+          super.init(device: device, inFunctionName: "conv_add_5x1_prelu_other_float")
+        }
+      } else if param.filter.width == 5 && param.filter.height == 1 {
+        if param.mode == "channel" {
+          super.init(device: device, inFunctionName: "conv_add_1x5_prelu_channel_float")
+        } else if param.mode == "element" {
+          super.init(device: device, inFunctionName: "conv_add_1x5_prelu_element_float")
+        } else {
+          super.init(device: device, inFunctionName: "conv_add_1x5_prelu_other_float")
+        }
+      } else {
+        fatalError(" unsupport yet ")
+      }
+    } else {
+      fatalError()
+    }
+    
+    let offsetY = (Int(param.dilations[1]) * (param.filter.height - 1) + 1)/2 - Int(param.paddings[1])
+    
+    let offsetX = (Int(param.dilations[0]) * (param.filter.width - 1) + 1)/2 - Int(param.paddings[0])
+    
+    //    print(" function: \(functionName)")
+    //    print("offset x: \(offsetX)")
+    //    print("offset y: \(offsetY)")
+    
+    let offsetZ = 0.0
+    let inMetalParam = MetalConvParam.init(offsetX: Int16(offsetX), offsetY: Int16(offsetY), offsetZ: Int16(offsetZ), strideX: UInt16(param.stride[0]), strideY: UInt16(param.stride[1]), dilationX: UInt16(param.dilations[0]), dilationY: UInt16(param.dilations[1]))
+    //    print("metal param: ")
+    //    print(inMetalParam)
+    
+    metalParam = inMetalParam
+  }
+  
+  func compute(commandBuffer: MTLCommandBuffer, param: ConvAddAddPreluParam<P>) throws {
+    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+      throw PaddleMobileError.predictError(message: " encode is nil")
+    }
+    
+    encoder.setTexture(param.input.metalTexture, index: 0)
+    encoder.setTexture(param.output.metalTexture, index: 1)
+    encoder.setBytes(&metalParam, length: MemoryLayout<MetalConvParam>.size, index: 0)
+    encoder.setBuffer(param.filter.buffer, offset: 0, index: 1)
+    encoder.setBuffer(param.y.buffer, offset: 0, index: 2)
+    encoder.setBuffer(param.alpha.buffer, offset: 0, index: 3)
+    encoder.dispatch(computePipline: pipline, outTexture: param.output.metalTexture)
+    encoder.endEncoding()
+  }
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/ConvAddBatchNormReluKernel.swift

@@ -49,22 +49,37 @@ class ConvAddBatchNormReluKernel<P: PrecisionType>: Kernel, Computable, Testable
   var metalParam: MetalConvParam!
   
   required init(device: MTLDevice, param: ConvAddBatchNormReluParam<P>) {
-        
+    param.output.initTexture(device: device, inTranspose: [0, 2, 3, 1], computePrecision: computePrecision)
+    param.filter.initBuffer(device: device, precision: computePrecision)
+    param.y.initBuffer(device: device, precision: computePrecision)
+    param.variance.initBuffer(device: device, precision: .Float32)
+    param.mean.initBuffer(device: device, precision: .Float32)
+    param.scale.initBuffer(device: device, precision: .Float32)
+    param.bias.initBuffer(device: device, precision: .Float32)
+    
+    if computePrecision == .Float32 {
       if param.filter.width == 1 && param.filter.height == 1 {
         super.init(device: device, inFunctionName: "conv_add_batch_norm_relu_1x1")
       } else if param.filter.channel == 1 {
         super.init(device: device, inFunctionName: "depthwise_conv_add_batch_norm_relu_3x3")
-        } else {
+      } else if param.filter.width == 3 && param.filter.height == 3 {
         super.init(device: device, inFunctionName: "conv_add_batch_norm_relu_3x3")
+      } else {
+        fatalError(" unsupport ")
+      }
+    } else if computePrecision == .Float16 {
+      if param.filter.width == 1 && param.filter.height == 1 {
+        super.init(device: device, inFunctionName: "conv_add_batch_norm_relu_1x1_half")
+      } else if param.filter.channel == 1 {
+        super.init(device: device, inFunctionName: "depthwise_conv_add_batch_norm_relu_3x3_half")
+      } else if param.filter.width == 3 && param.filter.height == 3 {
+        super.init(device: device, inFunctionName: "conv_add_batch_norm_relu_3x3_half")
+      } else {
+        fatalError(" unsupport ")
+      }
+    } else {
+      fatalError()
     }
-        
-        param.filter.initBuffer(device: device, precision: Tensor.BufferPrecision.Float32)
-        param.y.initBuffer(device: device, precision: Tensor.BufferPrecision.Float32)
-
-        param.variance.initBuffer(device: device)
-        param.mean.initBuffer(device: device)
-        param.scale.initBuffer(device: device)
-        param.bias.initBuffer(device: device)
     
     let offsetX = param.filter.width/2 - Int(param.paddings[0])
     let offsetY = param.filter.height/2 - Int(param.paddings[1])
@@ -73,7 +88,7 @@ class ConvAddBatchNormReluKernel<P: PrecisionType>: Kernel, Computable, Testable
     print("offset y: \(offsetY)")
     
     let offsetZ = 0.0
-        metalParam = MetalConvParam.init(offsetX: Int16(offsetX), offsetY: Int16(offsetY), offsetZ: Int16(offsetZ), strideX: UInt16(param.stride[0]), strideY: UInt16(param.stride[1]), paddedZ: UInt16(param.input.metalTexture.arrayLength * 4 - param.input.dim[3]))
+    metalParam = MetalConvParam.init(offsetX: Int16(offsetX), offsetY: Int16(offsetY), offsetZ: Int16(offsetZ), strideX: UInt16(param.stride[0]), strideY: UInt16(param.stride[1]), dilationX: UInt16(param.dilations[0]), dilationY: UInt16(param.dilations[1]))
     
     var invs: [P] = []
     let varianceContents = param.variance.buffer.contents().assumingMemoryBound(to: P.self)
@@ -93,8 +108,34 @@ class ConvAddBatchNormReluKernel<P: PrecisionType>: Kernel, Computable, Testable
       newScale[i] = invs[i] * scaleContents[i]
       newBiase[i] = biaseContents[i] - meanContents[i] * invs[i] * scaleContents[i]
     }
-        param.newBiase = device.makeBuffer(bytes: newBiase, length: param.bias.buffer.length)
-        param.newScale = device.makeBuffer(bytes: newScale, length: param.scale.buffer.length)
+    
+//    var newScaleFP16: UnsafeMutableRawPointer
+//
+//    float32ToFloat16(input: newScale as! UnsafeMutablePointer<Float32>, output: newScaleFP16, count: param.scale.buffer.length / MemoryLayout<P>.size)
+    
+    
+//    let newBiaseFloat16 = device.makeBuffer(length: <#T##Int#>, options: <#T##MTLResourceOptions#>)
+    
+    var newBiaseBuffer: MTLBuffer
+    var newScaleBuffer: MTLBuffer
+    
+    if computePrecision == .Float32 {
+      newBiaseBuffer = device.makeBuffer(bytes: newBiase, length: param.bias.buffer.length)!
+      newScaleBuffer = device.makeBuffer(bytes: newScale, length: param.scale.buffer.length)!
+    } else if computePrecision == .Float16 {
+      
+      newBiaseBuffer = device.makeBuffer(length: param.bias.buffer.length / 2)!
+      newScaleBuffer = device.makeBuffer(length: param.bias.buffer.length / 2)!
+      
+      float32ToFloat16(input: newBiase as! UnsafeMutablePointer<Float32>, output: newBiaseBuffer.contents(), count: param.bias.buffer.length / MemoryLayout<P>.size)
+      
+      float32ToFloat16(input: newScale as! UnsafeMutablePointer<Float32>, output: newScaleBuffer.contents(), count: param.scale.buffer.length / MemoryLayout<P>.size)
+    } else {
+      fatalError(" unsupport ")
+    }
+    
+    param.newBiase = newBiaseBuffer
+    param.newScale = newScaleBuffer
     
     newScale.deinitialize(count: param.scale.buffer.length)
     newScale.deallocate()

metal/paddle-mobile/paddle-mobile/Operators/Kernels/ConvAddKernel.swift

@@ -17,18 +17,58 @@ import Foundation
 class ConvAddKernel<P: PrecisionType>: Kernel, Computable {
   var metalParam: MetalConvParam!
   required init(device: MTLDevice, param: ConvAddParam<P>) {
+    param.output.initTexture(device: device, inTranspose: [0, 2, 3, 1], computePrecision: computePrecision)
+    param.filter.initBuffer(device: device, precision: computePrecision)
+    param.y.initBuffer(device: device, precision: computePrecision)
+    
+    if computePrecision == .Float16 {
+      if param.filter.width == 1 && param.filter.height == 1 {
+        super.init(device: device, inFunctionName: "conv_add_1x1_half")
+      } else if param.filter.channel == 1 {
+        super.init(device: device, inFunctionName: "depthwise_conv_add_3x3_half")
+      } else if param.filter.width == 3 && param.filter.height == 3 {
+        super.init(device: device, inFunctionName: "conv_add_3x3_half")
+      } else if param.filter.width == 1 && param.filter.height == 5 {
+        super.init(device: device, inFunctionName: "conv_add_5x1_half")
+      } else if param.filter.width == 5 && param.filter.height == 1 {
+        super.init(device: device, inFunctionName: "conv_add_1x5_half")
+      } else {
+        fatalError(" unsupport yet ")
+      }
+    } else if computePrecision == .Float32 {
+      if param.filter.width == 1 && param.filter.height == 1 {
         super.init(device: device, inFunctionName: "conv_add_1x1")
-        let offsetX = param.filter.width/2 - Int(param.paddings[0])
-        let offsetY = param.filter.height/2 - Int(param.paddings[1])
+      } else if param.filter.channel == 1 {
+        super.init(device: device, inFunctionName: "depthwise_conv_add_3x3")
+      } else if param.filter.width == 1 && param.filter.height == 5 {
+        super.init(device: device, inFunctionName: "conv_add_5x1")
+      } else if param.filter.width == 5 && param.filter.height == 1 {
+        super.init(device: device, inFunctionName: "conv_add_1x5")
+      } else if param.filter.width == 3 && param.filter.height == 3 {
+        super.init(device: device, inFunctionName: "conv_add_3x3")
+      } else {
+        fatalError(" unsupport yet ")
+      }
+    } else {
+      fatalError()
+    }
     
-        param.filter.initBuffer(device: device, precision: Tensor.BufferPrecision.Float32)
-        param.y.initBuffer(device: device, precision: Tensor.BufferPrecision.Float32)
 
-        print("offset x: \(offsetX)")
-        print("offset y: \(offsetY)")
+    
+    let offsetY = (Int(param.dilations[1]) * (param.filter.height - 1) + 1)/2 - Int(param.paddings[1])
+    
+    let offsetX = (Int(param.dilations[0]) * (param.filter.width - 1) + 1)/2 - Int(param.paddings[0])
+    
+//    print(" function: \(functionName)")
+//    print("offset x: \(offsetX)")
+//    print("offset y: \(offsetY)")
     
     let offsetZ = 0.0
-        metalParam = MetalConvParam.init(offsetX: Int16(offsetX), offsetY: Int16(offsetY), offsetZ: Int16(offsetZ), strideX: UInt16(param.stride[0]), strideY: UInt16(param.stride[1]), paddedZ: UInt16(param.input.metalTexture.arrayLength * 4 - param.input.dim[3]))
+    let inMetalParam = MetalConvParam.init(offsetX: Int16(offsetX), offsetY: Int16(offsetY), offsetZ: Int16(offsetZ), strideX: UInt16(param.stride[0]), strideY: UInt16(param.stride[1]), dilationX: UInt16(param.dilations[0]), dilationY: UInt16(param.dilations[1]))
+//    print("metal param: ")
+//    print(inMetalParam)
+    
+    metalParam = inMetalParam
   }
   
   func compute(commandBuffer: MTLCommandBuffer, param: ConvAddParam<P>) throws {

metal/paddle-mobile/paddle-mobile/Operators/Kernels/ConvAddPreluKernel.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class ConvAddPreluKernel<P: PrecisionType>: Kernel, Computable {
+  var metalParam: MetalConvParam!
+  required init(device: MTLDevice, param: ConvAddPreluParam<P>) {
+    param.output.initTexture(device: device, inTranspose: [0, 2, 3, 1], computePrecision: computePrecision)
+    param.filter.initBuffer(device: device, precision: computePrecision)
+    param.y.initBuffer(device: device, precision: computePrecision)
+    param.alpha.initBuffer(device: device, precision: computePrecision)
+
+    if computePrecision == .Float16 {
+      if param.filter.width == 1 && param.filter.height == 1 {
+        if param.mode == "channel" {
+          super.init(device: device, inFunctionName: "conv_add_1x1_prelu_channel_half")
+        } else if param.mode == "element" {
+          super.init(device: device, inFunctionName: "conv_add_1x1_prelu_element_half")
+        } else {
+          super.init(device: device, inFunctionName: "conv_add_1x1_prelu_other_half")
+        }
+        
+      } else if param.filter.channel == 1 {
+        if param.mode == "channel" {
+          super.init(device: device, inFunctionName: "depthwise_conv_add_3x3_prelu_channel_half")
+        } else if param.mode == "element" {
+          super.init(device: device, inFunctionName: "depthwise_conv_add_3x3_prelu_element_half")
+        } else {
+          super.init(device: device, inFunctionName: "depthwise_conv_add_3x3_prelu_other_half")
+        }
+      } else if param.filter.width == 3 && param.filter.height == 3 {
+        if param.mode == "channel" {
+          super.init(device: device, inFunctionName: "conv_add_3x3_prelu_channel_half")
+        } else if param.mode == "element" {
+          super.init(device: device, inFunctionName: "conv_add_3x3_prelu_element_half")
+        } else {
+          super.init(device: device, inFunctionName: "conv_add_3x3_prelu_other_half")
+        }
+        
+      } else if param.filter.width == 1 && param.filter.height == 5 {
+        if param.mode == "channel" {
+          super.init(device: device, inFunctionName: "conv_add_5x1_prelu_channel_half")
+        } else if param.mode == "element" {
+          super.init(device: device, inFunctionName: "conv_add_5x1_prelu_element_half")
+        } else {
+          super.init(device: device, inFunctionName: "conv_add_5x1_prelu_other_half")
+        }
+      } else if param.filter.width == 5 && param.filter.height == 1 {
+        if param.mode == "channel" {
+          super.init(device: device, inFunctionName: "conv_add_1x5_prelu_channel_half")
+        } else if param.mode == "element" {
+          super.init(device: device, inFunctionName: "conv_add_1x5_prelu_element_half")
+        } else {
+          super.init(device: device, inFunctionName: "conv_add_1x5_prelu_other_half")
+        }
+      } else {
+        fatalError(" unsupport yet ")
+      }
+    } else if computePrecision == .Float32 {
+      if param.filter.width == 1 && param.filter.height == 1 {
+        if param.mode == "channel" {
+          super.init(device: device, inFunctionName: "conv_add_1x1_prelu_channel_float")
+        } else if param.mode == "element" {
+          super.init(device: device, inFunctionName: "conv_add_1x1_prelu_element_float")
+        } else {
+          super.init(device: device, inFunctionName: "conv_add_1x1_prelu_other_float")
+        }
+      } else if param.filter.channel == 1 {
+        if param.mode == "channel" {
+          super.init(device: device, inFunctionName: "depthwise_conv_add_3x3_prelu_channel_float")
+        } else if param.mode == "element" {
+          super.init(device: device, inFunctionName: "depthwise_conv_add_3x3_prelu_element_float")
+        } else {
+          super.init(device: device, inFunctionName: "depthwise_conv_add_3x3_prelu_other_float")
+        }
+      } else if param.filter.width == 3 && param.filter.height == 3 {
+        if param.mode == "channel" {
+          super.init(device: device, inFunctionName: "conv_add_3x3_prelu_channel_float")
+        } else if param.mode == "element" {
+          super.init(device: device, inFunctionName: "conv_add_3x3_prelu_element_float")
+        } else {
+          super.init(device: device, inFunctionName: "conv_add_3x3_prelu_other_float")
+        }
+        
+      } else if param.filter.width == 1 && param.filter.height == 5 {
+        if param.mode == "channel" {
+          super.init(device: device, inFunctionName: "conv_add_5x1_prelu_channel_float")
+        } else if param.mode == "element" {
+          super.init(device: device, inFunctionName: "conv_add_5x1_prelu_element_float")
+        } else {
+          super.init(device: device, inFunctionName: "conv_add_5x1_prelu_other_float")
+        }
+      } else if param.filter.width == 5 && param.filter.height == 1 {
+        if param.mode == "channel" {
+          super.init(device: device, inFunctionName: "conv_add_1x5_prelu_channel_float")
+        } else if param.mode == "element" {
+          super.init(device: device, inFunctionName: "conv_add_1x5_prelu_element_float")
+        } else {
+          super.init(device: device, inFunctionName: "conv_add_1x5_prelu_other_float")
+        }
+      } else {
+        fatalError(" unsupport yet ")
+      }
+    } else {
+      fatalError()
+    }
+    
+    let offsetY = (Int(param.dilations[1]) * (param.filter.height - 1) + 1)/2 - Int(param.paddings[1])
+    
+    let offsetX = (Int(param.dilations[0]) * (param.filter.width - 1) + 1)/2 - Int(param.paddings[0])
+    
+    //    print(" function: \(functionName)")
+    //    print("offset x: \(offsetX)")
+    //    print("offset y: \(offsetY)")
+    
+    let offsetZ = 0.0
+    let inMetalParam = MetalConvParam.init(offsetX: Int16(offsetX), offsetY: Int16(offsetY), offsetZ: Int16(offsetZ), strideX: UInt16(param.stride[0]), strideY: UInt16(param.stride[1]), dilationX: UInt16(param.dilations[0]), dilationY: UInt16(param.dilations[1]))
+    //    print("metal param: ")
+    //    print(inMetalParam)
+    
+    metalParam = inMetalParam
+  }
+  
+  func compute(commandBuffer: MTLCommandBuffer, param: ConvAddPreluParam<P>) throws {
+    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+      throw PaddleMobileError.predictError(message: " encode is nil")
+    }
+    
+    encoder.setTexture(param.input.metalTexture, index: 0)
+    encoder.setTexture(param.output.metalTexture, index: 1)
+    encoder.setBytes(&metalParam, length: MemoryLayout<MetalConvParam>.size, index: 0)
+    encoder.setBuffer(param.filter.buffer, offset: 0, index: 1)
+    encoder.setBuffer(param.y.buffer, offset: 0, index: 2)
+    encoder.setBuffer(param.alpha.buffer, offset: 0, index: 3)
+    encoder.dispatch(computePipline: pipline, outTexture: param.output.metalTexture)
+    encoder.endEncoding()
+  }
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/ConvBNReluKernel.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+import MetalPerformanceShaders
+
+struct ConvBNReluTestParam: TestParam {
+  let inputTexture: MTLTexture
+  let outputTexture: MTLTexture
+  var metalParam: MetalConvParam
+  let filterBuffer: MTLBuffer
+  let biaseBuffer: MTLBuffer
+  let newScaleBuffer: MTLBuffer
+  let newBiaseBuffer: MTLBuffer
+  let filterSize: (width: Int, height: Int, channel: Int)
+  init(inInputTexture: MTLTexture, inOutputTexture: MTLTexture, inMetalParam: MetalConvParam, inFilterBuffer: MTLBuffer, inBiaseBuffer: MTLBuffer, inNewScaleBuffer: MTLBuffer, inNewBiaseBuffer: MTLBuffer, inFilterSize: (width: Int, height: Int, channel: Int)) {
+    
+    inputTexture = inInputTexture
+    outputTexture = inOutputTexture
+    metalParam = inMetalParam
+    filterBuffer = inFilterBuffer
+    biaseBuffer = inBiaseBuffer
+    newScaleBuffer = inNewScaleBuffer
+    newBiaseBuffer = inNewBiaseBuffer
+    filterSize = inFilterSize
+  }
+}
+
+class ConvBNReluKernel<P: PrecisionType>: Kernel, Computable, Testable {
+  required init(device: MTLDevice, testParam: ConvBNReluTestParam) {
+    if testParam.filterSize.width == 1 && testParam.filterSize.height == 1 {
+      super.init(device: device, inFunctionName: "conv_batch_norm_relu_1x1")
+    } else if testParam.filterSize.channel == 1 {
+      super.init(device: device, inFunctionName: "depthwise_conv_batch_norm_relu_3x3")
+    } else {
+      super.init(device: device, inFunctionName: "conv_batch_norm_relu_3x3")
+    }
+  }
+  
+  var metalParam: MetalConvParam!
+
+  required init(device: MTLDevice, param: ConvBNReluParam<P>) {
+    
+    param.output.initTexture(device: device, inTranspose: [0, 2, 3, 1], computePrecision: computePrecision)
+    param.filter.initBuffer(device: device, precision: computePrecision)
+    param.variance.initBuffer(device: device, precision: .Float32)
+    param.mean.initBuffer(device: device, precision: .Float32)
+    param.scale.initBuffer(device: device, precision: .Float32)
+    param.bias.initBuffer(device: device, precision: .Float32)
+    
+    if computePrecision == .Float32 {
+      if param.filter.width == 1 && param.filter.height == 1 {
+        super.init(device: device, inFunctionName: "conv_batch_norm_relu_1x1")
+      } else if param.filter.channel == 1 {
+        super.init(device: device, inFunctionName: "depthwise_conv_batch_norm_relu_3x3")
+      } else if param.filter.width == 3 && param.filter.height == 3 {
+        super.init(device: device, inFunctionName: "conv_batch_norm_relu_3x3")
+      } else {
+        fatalError(" unsupport ")
+      }
+    } else if computePrecision == .Float16 {
+      if param.filter.width == 1 && param.filter.height == 1 {
+        super.init(device: device, inFunctionName: "conv_batch_norm_relu_1x1_half")
+      } else if param.filter.channel == 1 {
+        super.init(device: device, inFunctionName: "depthwise_conv_batch_norm_relu_3x3_half")
+      } else if param.filter.width == 3 && param.filter.height == 3 {
+        super.init(device: device, inFunctionName: "conv_batch_norm_relu_3x3_half")
+      } else {
+        fatalError(" unsupport ")
+      }
+    } else {
+      fatalError()
+    }
+    
+   
+    
+    let offsetX = param.filter.width/2 - Int(param.paddings[0])
+    let offsetY = param.filter.height/2 - Int(param.paddings[1])
+    
+//    print(" param filter width: \(param.filter.width)")
+//    print(" param filter height: \(param.filter.height)")
+//
+//    print(" param paddings: \(param.paddings)")
+//
+//    print("ConvBNReluKernel offset x: \(offsetX)")
+//    print("ConvBNReluKernel offset y: \(offsetY)")
+    
+    let offsetZ = 0.0
+    
+    metalParam = MetalConvParam.init(offsetX: Int16(offsetX), offsetY: Int16(offsetY), offsetZ: Int16(offsetZ), strideX: UInt16(param.stride[0]), strideY: UInt16(param.stride[1]), dilationX: UInt16(param.dilations[0]), dilationY: UInt16(param.dilations[1]))
+    
+    var invs: [P] = []
+    let varianceContents = param.variance.buffer.contents().assumingMemoryBound(to: P.self)
+    
+    for i in 0..<param.variance.buffer.length/MemoryLayout<P>.stride {
+      let inv = 1.0/pow(Float32.init(varianceContents[i]) + param.epsilon, 0.5)
+      invs.append(P(inv))
+    }
+    
+    let newScale: UnsafeMutablePointer<P> = UnsafeMutablePointer<P>.allocate(capacity: param.scale.buffer.length)
+    let newBiase: UnsafeMutablePointer<P> = UnsafeMutablePointer<P>.allocate(capacity: param.bias.buffer.length)
+    
+    let scaleContents = param.scale.buffer.contents().assumingMemoryBound(to: P.self)
+    let biaseContents = param.bias.buffer.contents().assumingMemoryBound(to: P.self)
+    let meanContents = param.mean.buffer.contents().assumingMemoryBound(to: P.self)
+    for i in 0..<param.scale.buffer.length/MemoryLayout<P>.stride {
+      newScale[i] = invs[i] * scaleContents[i]
+      newBiase[i] = biaseContents[i] - meanContents[i] * invs[i] * scaleContents[i]
+    }
+    
+    var newBiaseBuffer: MTLBuffer
+    var newScaleBuffer: MTLBuffer
+    
+    if computePrecision == .Float32 {
+      newBiaseBuffer = device.makeBuffer(bytes: newBiase, length: param.bias.buffer.length)!
+      newScaleBuffer = device.makeBuffer(bytes: newScale, length: param.scale.buffer.length)!
+    } else if computePrecision == .Float16 {
+      
+      newBiaseBuffer = device.makeBuffer(length: param.bias.buffer.length / 2)!
+      newScaleBuffer = device.makeBuffer(length: param.bias.buffer.length / 2)!
+      
+      float32ToFloat16(input: newBiase as! UnsafeMutablePointer<Float32>, output: newBiaseBuffer.contents(), count: param.bias.buffer.length / MemoryLayout<P>.size)
+      
+      float32ToFloat16(input: newScale as! UnsafeMutablePointer<Float32>, output: newScaleBuffer.contents(), count: param.scale.buffer.length / MemoryLayout<P>.size)
+    } else {
+      fatalError(" unsupport ")
+    }
+    
+    param.newBiase = newBiaseBuffer
+    param.newScale = newScaleBuffer
+    
+    newScale.deinitialize(count: param.scale.buffer.length)
+    newScale.deallocate()
+    
+    newBiase.deinitialize(count: param.bias.buffer.length)
+    newBiase.deallocate()
+  }
+  
+  func compute(commandBuffer: MTLCommandBuffer, param: ConvBNReluParam<P>) throws {
+    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+      throw PaddleMobileError.predictError(message: " encode is nil")
+    }
+    
+    encoder.setTexture(param.input.metalTexture, index: 0)
+    encoder.setTexture(param.output.metalTexture, index: 1)
+    encoder.setBytes(&metalParam, length: MemoryLayout<MetalConvParam>.size, index: 0)
+    encoder.setBuffer(param.filter.buffer, offset: 0, index: 1)
+    encoder.setBuffer(param.newScale!, offset: 0, index: 2)
+    encoder.setBuffer(param.newBiase!, offset: 0, index: 3)
+    encoder.dispatch(computePipline: pipline, outTexture: param.output.metalTexture)
+    encoder.endEncoding()
+  }
+  
+  public func test(commandBuffer: MTLCommandBuffer, param: ConvBNReluTestParam) {
+    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+      fatalError()
+    }
+    
+    encoder.setTexture(param.inputTexture, index: 0)
+    encoder.setTexture(param.outputTexture, index: 1)
+    var inMetalParam = param.metalParam
+    encoder.setBytes(&inMetalParam, length: MemoryLayout<MetalConvParam>.size, index: 0)
+    encoder.setBuffer(param.filterBuffer, offset: 0, index: 1)
+    encoder.setBuffer(param.newScaleBuffer, offset: 0, index: 2)
+    encoder.setBuffer(param.newBiaseBuffer, offset: 0, index: 3)
+    encoder.dispatch(computePipline: pipline, outTexture: param.outputTexture)
+    encoder.endEncoding()
+  }
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/ConvKernel.metal

-/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
- 
- Licensed under the Apache License, Version 2.0 (the "License");
- you may not use this file except in compliance with the License.
- You may obtain a copy of the License at
- 
- http://www.apache.org/licenses/LICENSE-2.0
- 
- Unless required by applicable law or agreed to in writing, software
- distributed under the License is distributed on an "AS IS" BASIS,
- WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- See the License for the specific language governing permissions and
- limitations under the License. */
-
-#include <metal_stdlib>
-using namespace metal;
-
-struct MetalConvParam {
-    short offsetX;
-    short offsetY;
-    short offsetZ;
-    ushort strideX;
-    ushort strideY;
-};
-
-
-kernel void conv_add_batch_norm_relu_1x1_half(texture2d_array<half, access::sample> inTexture [[texture(0)]],
-                                         texture2d_array<half, access::write> outTexture [[texture(1)]],
-                                         constant MetalConvParam &param [[buffer(0)]],
-                                         const device half4 *weights [[buffer(1)]],
-                                         const device half4 *biase [[buffer(2)]],
-                                         const device float4 *new_scale [[buffer(3)]],
-                                         const device float4 *new_biase [[buffer(4)]],
-                                         uint3 gid [[thread_position_in_grid]]) {
-    
-    if (gid.x >= outTexture.get_width() ||
-        gid.y >= outTexture.get_height() ||
-        gid.z >= outTexture.get_array_size()) {
-        return;
-    }
-    
-    
-    ushort2 stride = ushort2(param.strideX, param.strideY);
-    ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
-    
-    constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
-    const uint kernelHXW = 1;
-    
-    uint input_arr_size = inTexture.get_array_size();
-    uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
-    
-    half4 output = half4(0.0);
-    
-    half4 input;
-    for (uint i = 0; i < input_arr_size; ++i) {
-        input = inTexture.sample(sample, float2(posInInput.x, posInInput.y), i);
-        half4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size  + i];
-        output.x += dot(input, weight_x);
-        
-        half4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size  + i];
-        output.y += dot(input, weight_y);
-        
-        half4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size  + i];
-        output.z += dot(input, weight_z);
-        
-        half4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + i];
-        output.w += dot(input, weight_w);
-    }
-    
-    output = half4(fmax((float4(output) + float4(biase[gid.z])) * new_scale[gid.z] + new_biase[gid.z], 0.0));
-    outTexture.write(output, gid.xy, gid.z);
-}
-
-kernel void conv_add_batch_norm_relu_3x3_half(texture2d_array<half, access::sample> inTexture [[texture(0)]],
-                                         texture2d_array<half, access::write> outTexture [[texture(1)]],
-                                         constant MetalConvParam &param [[buffer(0)]],
-                                         const device half4 *weights [[buffer(1)]],
-                                         const device half4 *biase [[buffer(2)]],
-                                         const device float4 *new_scale [[buffer(3)]],
-                                         const device float4 *new_biase [[buffer(4)]],
-                                         uint3 gid [[thread_position_in_grid]]) {
-    
-    if (gid.x >= outTexture.get_width() ||
-        gid.y >= outTexture.get_height() ||
-        gid.z >= outTexture.get_array_size()) {
-        return;
-    }
-    
-    ushort2 stride = ushort2(param.strideX, param.strideY);
-    const ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
-    
-    constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
-    const uint kernelHXW = 9;
-    uint input_arr_size = inTexture.get_array_size();
-    uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
-    
-    half4 output = half4(0.0);
-    
-    half4 input[9];
-    for (uint i = 0; i < input_arr_size; ++i) {
-        input[0] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y - 1), i);
-        input[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - 1), i);
-        input[2] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y - 1), i);
-        input[3] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y), i);
-        input[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), i);
-        input[5] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y), i);
-        input[6] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y + 1), i);
-        input[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + 1), i);
-        input[8] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y + 1), i);
-        for (int j = 0; j < 9; ++j) {
-            half4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size + j * input_arr_size + i];
-            output.x += dot(input[j], weight_x);
-            
-            half4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size + j * input_arr_size + i];
-            output.y += dot(input[j], weight_y);
-            
-            half4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size + j * input_arr_size + i];
-            output.z += dot(input[j], weight_z);
-            
-            half4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + j * input_arr_size + i];
-            output.w += dot(input[j], weight_w);
-        }
-    }
-    output = half4(fmax((float4(output) + float4(biase[gid.z])) * new_scale[gid.z] + new_biase[gid.z], 0.0));
-    outTexture.write(output, gid.xy, gid.z);
-}
-
-kernel void conv_add_1x1_half(texture2d_array<half, access::sample> inTexture [[texture(0)]],
-                         texture2d_array<half, access::write> outTexture [[texture(1)]],
-                         constant MetalConvParam &param [[buffer(0)]],
-                         const device half4 *weights [[buffer(1)]],
-                         const device half4 *biase [[buffer(2)]],
-                         uint3 gid [[thread_position_in_grid]]) {
-    
-    if (gid.x >= outTexture.get_width() ||
-        gid.y >= outTexture.get_height() ||
-        gid.z >= outTexture.get_array_size()) {
-        return;
-    }
-    
-    ushort2 stride = ushort2(param.strideX, param.strideY);
-    ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
-    
-    constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
-    const uint kernelHXW = 1;
-    
-    uint input_arr_size = inTexture.get_array_size();
-    uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
-    
-    half4 output = half4(0.0);
-    
-    half4 input;
-    for (uint i = 0; i < input_arr_size; ++i) {
-        input = inTexture.sample(sample, float2(posInInput.x, posInInput.y), i);
-        half4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size  + i];
-        output.x += dot(input, weight_x);
-        
-        half4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size  + i];
-        output.y += dot(input, weight_y);
-        
-        half4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size  + i];
-        output.z += dot(input, weight_z);
-        
-        half4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + i];
-        output.w += dot(input, weight_w);
-    }
-    output = output + biase[gid.z];
-    outTexture.write(output, gid.xy, gid.z);
-}
-
-kernel void depthwise_conv_add_batch_norm_relu_3x3_half(texture2d_array<half, access::sample> inTexture [[texture(0)]],
-                                                   texture2d_array<half, access::write> outTexture [[texture(1)]],
-                                                   constant MetalConvParam &param [[buffer(0)]],
-                                                   const device half *weights [[buffer(1)]],
-                                                   const device half4 *biase [[buffer(2)]],
-                                                   const device float4 *new_scale [[buffer(3)]],
-                                                   const device float4 *new_biase [[buffer(4)]],
-                                                   uint3 gid [[thread_position_in_grid]]) {
-    
-    if (gid.x >= outTexture.get_width() ||
-        gid.y >= outTexture.get_height() ||
-        gid.z >= outTexture.get_array_size()) {
-        return;
-    }
-    uint output_slice = gid.z;
-    ushort2 stride = ushort2(param.strideX, param.strideY);
-    ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
-    constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
-    const uint kernelHXW = 9;
-    uint weithTo = gid.z * kernelHXW * 4;
-    half4 output = half4(0.0);
-    half4 inputs[9];
-    inputs[0] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y - 1), output_slice);
-    inputs[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - 1), output_slice);
-    inputs[2] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y - 1), output_slice);
-    inputs[3] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y), output_slice);
-    inputs[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), output_slice);
-    inputs[5] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y), output_slice);
-    inputs[6] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y + 1), output_slice);
-    inputs[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + 1), output_slice);
-    inputs[8] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y + 1), output_slice);
-    for (int j = 0; j < 9; ++j) {
-        half4 input = inputs[j];
-        output.x += input.x * weights[weithTo + 0 * kernelHXW + j];
-        output.y += input.y * weights[weithTo + 1 * kernelHXW + j];
-        output.z += input.z * weights[weithTo + 2 * kernelHXW + j];
-        output.w += input.w * weights[weithTo + 3 * kernelHXW + j];
-    }
-    output = half4(fmax((float4(output) + float4(biase[gid.z])) * new_scale[gid.z] + new_biase[gid.z], 0.0));
-    outTexture.write(output, gid.xy, gid.z);
-}
-
-
-/*---------------------------------------------*/
-
-
-
-kernel void conv_add_batch_norm_relu_1x1(texture2d_array<float, access::sample> inTexture [[texture(0)]],
-                                         texture2d_array<float, access::write> outTexture [[texture(1)]],
-                                         constant MetalConvParam &param [[buffer(0)]],
-                                         const device float4 *weights [[buffer(1)]],
-                                         const device float4 *biase [[buffer(2)]],
-                                         const device float4 *new_scale [[buffer(3)]],
-                                         const device float4 *new_biase [[buffer(4)]],
-                                         uint3 gid [[thread_position_in_grid]]) {
-    
-    if (gid.x >= outTexture.get_width() ||
-        gid.y >= outTexture.get_height() ||
-        gid.z >= outTexture.get_array_size()) {
-        return;
-    }
-    
-    ushort2 stride = ushort2(param.strideX, param.strideY);
-    ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
-    
-    constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
-    const uint kernelHXW = 1;
-    
-    uint input_arr_size = inTexture.get_array_size();
-    uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
-    
-    float4 output = float4(0.0);
-    
-    float4 input;
-    for (uint i = 0; i < input_arr_size; ++i) {
-        input = inTexture.sample(sample, float2(posInInput.x, posInInput.y), i);
-        float4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size  + i];
-        output.x += dot(input, weight_x);
-        
-        float4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size  + i];
-        output.y += dot(input, weight_y);
-        
-        float4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size  + i];
-        output.z += dot(input, weight_z);
-        
-        float4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + i];
-        output.w += dot(input, weight_w);
-    }
-    output = fmax((output + biase[gid.z]) * new_scale[gid.z] + new_biase[gid.z], 0.0);
-    outTexture.write(output, gid.xy, gid.z);
-}
-
-kernel void conv_add_batch_norm_relu_3x3(texture2d_array<float, access::sample> inTexture [[texture(0)]],
-                                         texture2d_array<float, access::write> outTexture [[texture(1)]],
-                                         constant MetalConvParam &param [[buffer(0)]],
-                                         const device float4 *weights [[buffer(1)]],
-                                         const device float4 *biase [[buffer(2)]],
-                                         const device float4 *new_scale [[buffer(3)]],
-                                         const device float4 *new_biase [[buffer(4)]],
-                                         uint3 gid [[thread_position_in_grid]]) {
-    
-    if (gid.x >= outTexture.get_width() ||
-        gid.y >= outTexture.get_height() ||
-        gid.z >= outTexture.get_array_size()) {
-        return;
-    }
-    
-    ushort2 stride = ushort2(param.strideX, param.strideY);
-    const ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
-    
-    constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
-    const uint kernelHXW = 9;
-    uint input_arr_size = inTexture.get_array_size();
-    uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
-    
-    float4 output = float4(0.0);
-    
-    float4 input[9];
-    for (uint i = 0; i < input_arr_size; ++i) {
-        input[0] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y - 1), i);
-        input[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - 1), i);
-        input[2] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y - 1), i);
-        input[3] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y), i);
-        input[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), i);
-        input[5] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y), i);
-        input[6] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y + 1), i);
-        input[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + 1), i);
-        input[8] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y + 1), i);
-        for (int j = 0; j < 9; ++j) {
-            float4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size + j * input_arr_size + i];
-            output.x += dot(input[j], weight_x);
-            
-            float4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size + j * input_arr_size + i];
-            output.y += dot(input[j], weight_y);
-            
-            float4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size + j * input_arr_size + i];
-            output.z += dot(input[j], weight_z);
-            
-            float4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + j * input_arr_size + i];
-            output.w += dot(input[j], weight_w);
-        }
-    }
-    output = fmax((output + biase[gid.z]) * new_scale[gid.z] + new_biase[gid.z], 0.0);
-    outTexture.write(output, gid.xy, gid.z);
-}
-
-kernel void conv_add_1x1(texture2d_array<float, access::sample> inTexture [[texture(0)]],
-                                         texture2d_array<float, access::write> outTexture [[texture(1)]],
-                                         constant MetalConvParam &param [[buffer(0)]],
-                                         const device float4 *weights [[buffer(1)]],
-                                         const device float4 *biase [[buffer(2)]],
-                                         uint3 gid [[thread_position_in_grid]]) {
-    
-    if (gid.x >= outTexture.get_width() ||
-        gid.y >= outTexture.get_height() ||
-        gid.z >= outTexture.get_array_size()) {
-        return;
-    }
-    
-    ushort2 stride = ushort2(param.strideX, param.strideY);
-    ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
-    
-    constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
-    const uint kernelHXW = 1;
-    
-    uint input_arr_size = inTexture.get_array_size();
-    uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
-    
-    float4 output = float4(0.0);
-    
-    float4 input;
-    for (uint i = 0; i < input_arr_size; ++i) {
-        input = inTexture.sample(sample, float2(posInInput.x, posInInput.y), i);
-        float4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size  + i];
-        output.x += dot(input, weight_x);
-        
-        float4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size  + i];
-        output.y += dot(input, weight_y);
-        
-        float4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size  + i];
-        output.z += dot(input, weight_z);
-        
-        float4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + i];
-        output.w += dot(input, weight_w);
-    }
-    output = output + biase[gid.z];
-    outTexture.write(output, gid.xy, gid.z);
-}
-
-kernel void depthwise_conv_add_batch_norm_relu_3x3(texture2d_array<float, access::sample> inTexture [[texture(0)]],
-                                         texture2d_array<float, access::write> outTexture [[texture(1)]],
-                                         constant MetalConvParam &param [[buffer(0)]],
-                                         const device float *weights [[buffer(1)]],
-                                         const device float4 *biase [[buffer(2)]],
-                                         const device float4 *new_scale [[buffer(3)]],
-                                         const device float4 *new_biase [[buffer(4)]],
-                                         uint3 gid [[thread_position_in_grid]]) {
-    
-    if (gid.x >= outTexture.get_width() ||
-        gid.y >= outTexture.get_height() ||
-        gid.z >= outTexture.get_array_size()) {
-        return;
-    }
-    uint output_slice = gid.z;
-    ushort2 stride = ushort2(param.strideX, param.strideY);
-    ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
-    constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
-    const uint kernelHXW = 9;
-    uint weithTo = gid.z * kernelHXW * 4;
-    float4 output = float4(0.0);
-    float4 inputs[9];
-    inputs[0] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y - 1), output_slice);
-    inputs[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - 1), output_slice);
-    inputs[2] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y - 1), output_slice);
-    inputs[3] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y), output_slice);
-    inputs[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), output_slice);
-    inputs[5] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y), output_slice);
-    inputs[6] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y + 1), output_slice);
-    inputs[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + 1), output_slice);
-    inputs[8] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y + 1), output_slice);
-    for (int j = 0; j < 9; ++j) {
-        float4 input = inputs[j];
-        output.x += input.x * weights[weithTo + 0 * kernelHXW + j];
-        output.y += input.y * weights[weithTo + 1 * kernelHXW + j];
-        output.z += input.z * weights[weithTo + 2 * kernelHXW + j];
-        output.w += input.w * weights[weithTo + 3 * kernelHXW + j];
-    }
-    output = fmax((output + biase[gid.z]) * new_scale[gid.z] + new_biase[gid.z], 0.0);
-    outTexture.write(output, gid.xy, gid.z);
-}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/ConvKernel.swift

@@ -14,26 +14,35 @@
 
 import Foundation
 
-
 public struct MetalConvParam {
   let offsetX: Int16
   let offsetY: Int16
   let offsetZ: Int16
   let strideX: UInt16
   let strideY: UInt16
-    let paddedZ: UInt16
+  let dilationX: UInt16
+  let dilationY: UInt16
 }
 
 class ConvKernel<P: PrecisionType>: Kernel, Computable {
   var metalParam: MetalConvParam!
   required init(device: MTLDevice, param: ConvParam<P>) {
-        super.init(device: device, inFunctionName: "conv_add_1x1")
+    param.filter.initBuffer(device: device, precision: ComputePrecision.Float32)
+    if param.filter.width == 1 && param.filter.height == 1 {
+      super.init(device: device, inFunctionName: "conv_1x1")
+    } else if param.filter.channel == 1 {
+      super.init(device: device, inFunctionName: "depthwise_conv_3x3")
+    } else if param.filter.width == 3 && param.filter.height == 3 {
+      super.init(device: device, inFunctionName: "conv_3x3")
+    } else {
+      fatalError(" unsupport ")
+    }
+
     let offsetX = param.filter.dim[2]/2 - Int(param.paddings[0])
     let offsetY = param.filter.dim[1]/2 - Int(param.paddings[1])
     let offsetZ = 0.0
-        param.filter.initBuffer(device: device, precision: Tensor.BufferPrecision.Float32)
     
-        metalParam = MetalConvParam.init(offsetX: Int16(offsetX), offsetY: Int16(offsetY), offsetZ: Int16(offsetZ), strideX: UInt16(param.stride[0]), strideY: UInt16(param.stride[1]), paddedZ: UInt16(param.input.metalTexture.arrayLength * 4 - param.input.dim[3]))
+    metalParam = MetalConvParam.init(offsetX: Int16(offsetX), offsetY: Int16(offsetY), offsetZ: Int16(offsetZ), strideX: UInt16(param.stride[0]), strideY: UInt16(param.stride[1]), dilationX: UInt16(param.dilations[0]), dilationY: UInt16(param.dilations[1]))
   }
   
   func compute(commandBuffer: MTLCommandBuffer, param: ConvParam<P>) throws {
@@ -49,3 +58,5 @@ class ConvKernel<P: PrecisionType>: Kernel, Computable {
     encoder.endEncoding()
   }
 }
+
+

metal/paddle-mobile/paddle-mobile/Operators/Kernels/ConvTransposeKernel.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+struct MetalConvTransposeParam {
+  let kernelW: UInt16;
+  let kernelH: UInt16;
+  
+  let strideX: UInt16;
+  let strideY: UInt16;
+  
+  let paddingX: UInt16;
+  let paddingY: UInt16;
+  
+  let dilationX: UInt16;
+  let dilationY: UInt16;
+}
+
+class ConvTransposeKernel<P: PrecisionType>: Kernel, Computable{
+  var metalParam: MetalConvTransposeParam!
+  required init(device: MTLDevice, param: ConvTransposeParam<P>) {
+    param.output.initTexture(device: device, inTranspose: param.input.transpose, computePrecision: computePrecision)
+    param.filter.initBuffer(device: device, precision: computePrecision, convertToNHWC: false, withTranspose: true)
+    if computePrecision == .Float32 {
+      if param.stride == [2, 2] && param.stride == [2, 2] {
+        super.init(device: device, inFunctionName: "conv_transpose2x2_stride2")
+      } else {
+        fatalError(" -- conv transpose unsupported yet -- ")
+      }
+    } else if computePrecision == .Float16 {
+      if param.stride == [2, 2] && param.stride == [2, 2] {
+        super.init(device: device, inFunctionName: "conv_transpose2x2_stride2_half")
+      } else {
+        fatalError(" -- conv transpose unsupported yet -- ")
+      }
+    } else {
+      fatalError()
+    }
+    
+//    let filter: [Float32] = param.filter.buffer.array()
+//    print(" conv transpose filter")
+//    print(filter)
+    let kernelWidth = UInt16(param.filter.width)
+    let kernelHeight = UInt16(param.filter.height)
+    
+    let strideX = UInt16(param.stride[0])
+    let strideY = UInt16(param.stride[1])
+    let paddingX = UInt16(param.paddings[0])
+    let paddingY = UInt16(param.paddings[1])
+    let dilationX = UInt16(param.dilations[0])
+    let dilationY = UInt16(param.dilations[1])
+    
+    metalParam = MetalConvTransposeParam.init(kernelW: kernelWidth, kernelH: kernelHeight, strideX: strideX, strideY: strideY, paddingX: paddingX, paddingY: paddingY, dilationX: dilationX, dilationY: dilationY)
+
+  }
+  
+  func compute(commandBuffer: MTLCommandBuffer, param: ConvTransposeParam<P>) throws {
+    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+      throw PaddleMobileError.predictError(message: " encoder is nil")
+    }
+    
+    encoder.setTexture(param.input.metalTexture, index: 0)
+    encoder.setTexture(param.output.metalTexture, index: 1)
+    encoder.setBytes(&metalParam, length: MemoryLayout<MetalConvTransposeParam>.size, index: 0)
+    encoder.setBuffer(param.filter.buffer, offset: 0, index: 1)
+    encoder.dispatch(computePipline: pipline, outTexture: param.output.metalTexture)
+    encoder.endEncoding()
+  }
+}
+
+

metal/paddle-mobile/paddle-mobile/Operators/Kernels/ElementwiseAddKernel.swift

@@ -14,13 +14,60 @@
 
 import Foundation
 
+struct ElementwiseAddMetalParam {
+  var fast: Int32 = 0
+  var axis: Int32 = 0
+  var ylen: Int32 = 0
+  var xdim: (Int32, Int32, Int32, Int32) = (0, 0, 0, 0)
+  var xtrans: (Int32, Int32, Int32, Int32) = (0, 1, 2, 3)
+  var ydim: (Int32, Int32, Int32, Int32) = (0, 0, 0, 0)
+  var ytrans: (Int32, Int32, Int32, Int32) = (0, 1, 2, 3)
+}
 
 class ElementwiseAddKernel<P: PrecisionType>: Kernel, Computable {
+  var metalParam: ElementwiseAddMetalParam
   required init(device: MTLDevice, param: ElementwiseAddParam<P>) {
+    param.output.initTexture(device: device, inTranspose: param.inputX.transpose, computePrecision: computePrecision)
+    
+    metalParam = ElementwiseAddMetalParam.init()
+    
+    let xdim: [Int32] = (0..<4).map { Int32(param.inputX.dim[$0]) }
+    let ydim: [Int32] = (0..<4).map { Int32(param.inputY.dim[$0]) }
+    let xtrans: [Int32] = (0..<4).map { Int32(param.inputX.transpose[$0]) }
+    let ytrans: [Int32] = (0..<4).map { Int32(param.inputY.transpose[$0]) }
+    
+    metalParam.xdim = (xdim[0], xdim[1], xdim[2], xdim[3])
+    metalParam.ydim = (ydim[0], ydim[1], ydim[2], ydim[3])
+    metalParam.xtrans = (xtrans[0], xtrans[1], xtrans[2], xtrans[3])
+    metalParam.ytrans = (ytrans[0], ytrans[1], ytrans[2], ytrans[3])
+    if param.axis == -1 {
+      metalParam.axis = 4 - Int32(param.inputY.tensorDim.cout())
+    } else {
+      metalParam.axis = 4 - Int32(param.inputX.tensorDim.cout()) + Int32(param.axis)
+    }
+    metalParam.ylen = Int32(param.inputY.tensorDim.cout())
+    if (param.inputX.dim == param.inputY.dim) && (param.inputX.transpose == param.inputY.transpose) {
+      //      print("===> elementwise_add fast!!!")
+      metalParam.fast = 1
+    }
+    if computePrecision == .Float32 {
       super.init(device: device, inFunctionName: "elementwise_add")
+    } else if computePrecision == .Float16 {
+      super.init(device: device, inFunctionName: "elementwise_add_half")
+    } else {
+      fatalError()
+    }
   }
   
   func compute(commandBuffer: MTLCommandBuffer, param: ElementwiseAddParam<P>) throws {
-        
+    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+      throw PaddleMobileError.predictError(message: " encode is nil")
+    }
+    encoder.setTexture(param.inputX.metalTexture, index: 0)
+    encoder.setTexture(param.inputY.metalTexture, index: 1)
+    encoder.setTexture(param.output.metalTexture, index: 2)
+    encoder.setBytes(&metalParam, length: MemoryLayout<ElementwiseAddMetalParam>.size, index: 0)
+    encoder.dispatch(computePipline: pipline, outTexture: param.output.metalTexture)
+    encoder.endEncoding()
   }
 }

metal/paddle-mobile/paddle-mobile/Operators/Kernels/ElementwiseAddPreluKernel.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+
+class ElementwiseAddPreluKernel<P: PrecisionType>: Kernel, Computable {
+  var metalParam: ElementwiseAddMetalParam
+  required init(device: MTLDevice, param: ElementwiseAddPreluParam<P>) {
+    param.output.initTexture(device: device, inTranspose: param.inputX.transpose, computePrecision: computePrecision)
+    param.alpha.initBuffer(device: device, precision: computePrecision)
+   
+    metalParam = ElementwiseAddMetalParam.init()
+    
+    let xdim: [Int32] = (0..<4).map { Int32(param.inputX.dim[$0]) }
+    let ydim: [Int32] = (0..<4).map { Int32(param.inputY.dim[$0]) }
+    let xtrans: [Int32] = (0..<4).map { Int32(param.inputX.transpose[$0]) }
+    let ytrans: [Int32] = (0..<4).map { Int32(param.inputY.transpose[$0]) }
+    
+    metalParam.xdim = (xdim[0], xdim[1], xdim[2], xdim[3])
+    metalParam.ydim = (ydim[0], ydim[1], ydim[2], ydim[3])
+    metalParam.xtrans = (xtrans[0], xtrans[1], xtrans[2], xtrans[3])
+    metalParam.ytrans = (ytrans[0], ytrans[1], ytrans[2], ytrans[3])
+    if param.axis == -1 {
+      metalParam.axis = 4 - Int32(param.inputY.tensorDim.cout())
+    } else {
+      metalParam.axis = 4 - Int32(param.inputX.tensorDim.cout()) + Int32(param.axis)
+    }
+    metalParam.ylen = Int32(param.inputY.tensorDim.cout())
+    if (param.inputX.dim == param.inputY.dim) && (param.inputX.transpose == param.inputY.transpose) {
+      //      print("===> elementwise_add fast!!!")
+      metalParam.fast = 1
+    }
+    
+    if computePrecision == .Float32 {
+      if param.mode == "channel" {
+        super.init(device: device, inFunctionName: "elementwise_add_channel_float")
+      } else if param.mode == "element" {
+        super.init(device: device, inFunctionName: "elementwise_add_element_float")
+      } else {
+        super.init(device: device, inFunctionName: "elementwise_add_prelu_float")
+      }
+    } else if computePrecision == .Float16 {
+      if param.mode == "channel" {
+        super.init(device: device, inFunctionName: "elementwise_add_channel_half")
+      } else if param.mode == "element" {
+        super.init(device: device, inFunctionName: "elementwise_add_channel_half")
+      } else {
+        super.init(device: device, inFunctionName: "elementwise_add_channel_half")
+      }
+    } else {
+      fatalError()
+    }
+  }
+  
+  func compute(commandBuffer: MTLCommandBuffer, param: ElementwiseAddPreluParam<P>) throws {
+    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+      throw PaddleMobileError.predictError(message: " encode is nil")
+    }
+    encoder.setTexture(param.inputX.metalTexture, index: 0)
+    encoder.setTexture(param.inputY.metalTexture, index: 1)
+    encoder.setTexture(param.output.metalTexture, index: 2)
+    encoder.setBytes(&metalParam, length: MemoryLayout<ElementwiseAddMetalParam>.size, index: 0)
+    encoder.setBuffer(param.alpha.buffer, offset: 0, index: 1)
+    encoder.dispatch(computePipline: pipline, outTexture: param.output.metalTexture)
+    encoder.endEncoding()
+  }
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/FlattenKernel.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+struct FlattenMetalParam {
+  var idim: (Int32, Int32, Int32, Int32)
+  var itrans: (Int32, Int32, Int32, Int32)
+  var odim: (Int32, Int32, Int32, Int32)
+  var otrans: (Int32, Int32, Int32, Int32)
+}
+
+
+class FlattenKernel<P: PrecisionType>: Kernel, Computable{
+  
+  var metalParam: FlattenMetalParam
+  
+  required init(device: MTLDevice, param: FlattenParam<P>) {
+    param.output.initTexture(device: device, computePrecision: computePrecision)
+    var id: [Int32] = [1, 1, 1, 1]
+    for i in 0..<param.input.tensorDim.cout() {
+      id[4-param.input.tensorDim.cout()+i] = Int32(param.input.tensorDim[i])
+    }
+    let it: [Int32] = param.input.transpose.map { Int32($0) }
+    var od: [Int32] = [1, 1, 1, 1]
+    for i in 0..<param.output.tensorDim.cout() {
+      od[4-param.output.tensorDim.cout()+i] = Int32(param.output.tensorDim[i])
+    }
+    let ot: [Int32] = param.output.transpose.map { Int32($0) }
+    metalParam = FlattenMetalParam.init(
+      idim: (id[0], id[1], id[2], id[3]),
+      itrans: (it[0], it[1], it[2], it[3]),
+      odim: (od[0], od[1], od[2], od[3]),
+      otrans: (ot[0], ot[1], ot[2], ot[3])
+    )
+    let irank = param.input.tensorDim.cout()
+    let orank = param.output.tensorDim.cout()
+    assert(orank == 2)
+    if computePrecision == .Float32 {
+      super.init(device: device, inFunctionName: "reshape_\(irank)_2_float")
+    } else if computePrecision == .Float16 {
+      super.init(device: device, inFunctionName: "reshape_\(irank)_2_half")
+    } else {
+      fatalError()
+    }
+  }
+  
+  func compute(commandBuffer: MTLCommandBuffer, param: FlattenParam<P>) throws {
+    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+      throw PaddleMobileError.predictError(message: " encoder is nil")
+    }
+    
+    encoder.setTexture(param.input.metalTexture, index: 0)
+    encoder.setTexture(param.output.metalTexture, index: 1)
+
+    encoder.setBytes(&metalParam, length: MemoryLayout<ReshapeMetalParam>.size, index: 0)
+    encoder.dispatch(computePipline: pipline, outTexture: param.output.metalTexture)
+    encoder.endEncoding()
+  }
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/Kernel.swift

-/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
- 
- Licensed under the Apache License, Version 2.0 (the "License");
- you may not use this file except in compliance with the License.
- You may obtain a copy of the License at
- 
- http://www.apache.org/licenses/LICENSE-2.0
- 
- Unless required by applicable law or agreed to in writing, software
- distributed under the License is distributed on an "AS IS" BASIS,
- WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- See the License for the specific language governing permissions and
- limitations under the License. */
-
-import Metal
-import Foundation
-
-public protocol TestParam {
-}
-
-public protocol Testable {
-    associatedtype TestParamType: TestParam
-    func test(commandBuffer: MTLCommandBuffer, param: TestParamType)
-    init(device: MTLDevice, testParam: TestParamType)
-}
-
-
-protocol Computable {
-    associatedtype ParamType: OpParam
-    func compute(commandBuffer: MTLCommandBuffer, param: ParamType) throws
-    init(device: MTLDevice, param: ParamType)
-}
-
-protocol KernelProtocol {
-    var pipline: MTLComputePipelineState { get set }
-    var functionName: String { get set }
-   
-}
-
-open class Kernel {
-    let pipline: MTLComputePipelineState
-    let functionName: String
-    public init(device: MTLDevice, inFunctionName: String, usePaddleMobileLib: Bool = true) {
-        pipline = device.pipeLine(funcName: inFunctionName, inPaddleMobileLib: usePaddleMobileLib)
-        functionName = inFunctionName
-    }
-}
-
-open class CusomKernel: Kernel {
-    public struct Shape {
-        public let width: Int
-        public let height: Int
-        public let channel: Int
-        public init(inWidth: Int, inHeight: Int, inChannel: Int){
-            width = inWidth
-            height = inHeight
-            channel = inChannel
-        }
-    }
-    let outputTexture: MTLTexture
-    public init(device: MTLDevice, inFunctionName: String, outputDim: Shape, usePaddleMobileLib: Bool = false) {
-        let textureDesc = MTLTextureDescriptor.init()
-        textureDesc.textureType = .type2D
-        textureDesc.width = outputDim.width
-        textureDesc.height = outputDim.height
-        textureDesc.depth = (outputDim.channel + 3) / 4
-        textureDesc.pixelFormat = .rgba32Float
-        textureDesc.usage = [.shaderRead, .shaderWrite]
-        textureDesc.storageMode = .shared
-        outputTexture = device.makeTexture(descriptor: textureDesc) ?! " make texture error "
-
-        super.init(device: device, inFunctionName: inFunctionName, usePaddleMobileLib: usePaddleMobileLib)
-    }
-    
-    func compute(inputTexuture: MTLTexture, commandBuffer: MTLCommandBuffer) throws {
-        guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
-            throw PaddleMobileError.predictError(message: " encode is nil")
-        }
-        encoder.setTexture(inputTexuture, index: 0)
-        encoder.setTexture(outputTexture, index: 1)
-        encoder.dispatch(computePipline: pipline, outTexture: outputTexture)
-        encoder.endEncoding()
-    }
-    
-}
-

metal/paddle-mobile/paddle-mobile/Operators/Kernels/Kernels.metal

-/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
- 
- Licensed under the Apache License, Version 2.0 (the "License");
- you may not use this file except in compliance with the License.
- You may obtain a copy of the License at
- 
- http://www.apache.org/licenses/LICENSE-2.0
- 
- Unless required by applicable law or agreed to in writing, software
- distributed under the License is distributed on an "AS IS" BASIS,
- WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- See the License for the specific language governing permissions and
- limitations under the License. */
-
-#include <metal_stdlib>
-using namespace metal;
-
-struct OutputDim {
-    ushort width;
-    ushort height;
-    ushort strideX;
-    ushort strideY;
-};
-
-kernel void resize(texture2d<half, access::read> inTexture [[texture(0)]],
-                   texture2d_array<half, access::write> outTexture [[texture(1)]],
-                   constant OutputDim &params [[buffer(0)]],
-                   uint3 gid [[thread_position_in_grid]]) {
-    if (gid.x >= outTexture.get_width() ||
-        gid.y >= outTexture.get_height() ||
-        gid.z >= outTexture.get_array_size()) return;
-    
-    constexpr sampler s(coord::pixel, filter::nearest, address::clamp_to_zero);
-    const uint2 pos = gid.xy * uint2(params.strideX, params.strideY);
-    const half4 input = inTexture.read(pos);
-    outTexture.write(half4(input.x, input.y, input.z, input.w), gid.xy, gid.z);
-}
-
-kernel void relu(texture2d_array<half, access::sample> inTexture [[texture(0)]],
-                 texture2d_array<half, access::write> outTexture [[texture(1)]],
-                 uint3 gid [[thread_position_in_grid]]) {
-    if (gid.x >= outTexture.get_width() ||
-        gid.y >= outTexture.get_height() ||
-        gid.z >= outTexture.get_array_size()) return;
-    constexpr sampler s(coord::pixel, filter::nearest, address::clamp_to_zero);
-    const half4 input = inTexture.read(gid.xy, gid.z);
-    const float4 relu = fmax((float4)input, 0.0);
-    outTexture.write(half4(relu), gid.xy, gid.z);
-}
-
-kernel void elementwise_add(texture2d_array<half, access::read> inTexture [[texture(0)]],
-                            texture2d_array<half, access::write> outTexture [[texture(1)]],
-                            const device half4 *biasTerms [[buffer(0)]],
-                            uint3 gid [[thread_position_in_grid]]) {
-    if (gid.x >= outTexture.get_width() ||
-        gid.y >= outTexture.get_height() ||
-        gid.z >= outTexture.get_array_size()) return;
-    constexpr sampler s(coord::pixel, filter::nearest, address::clamp_to_zero);
-    const half4 input = inTexture.read(gid.xy, gid.z);
-    outTexture.write(input, gid.xy, gid.z);
-}
-
-kernel void batchnorm(texture2d_array<half, access::read> inTexture [[texture(0)]],
-                      texture2d_array<half, access::write> outTexture [[texture(1)]],
-                      const device half4 * newScale [[buffer(0)]],
-                      const device half4 * newBias [[buffer(1)]],
-                      uint3 gid [[thread_position_in_grid]]) {
-    if (gid.x >= outTexture.get_width() ||
-        gid.y >= outTexture.get_height() ||
-        gid.z >= outTexture.get_array_size()) return;
-    const half4 input = inTexture.read(gid.xy, gid.z);
-    half4 output = input * newScale[gid.z] + newBias[gid.z];
-    outTexture.write(output, gid.xy, gid.z);
-}
-
-//kernel void texture2d_to_2d_array(texture2d<half, access::read> inTexture [[texture(0)]],
-//                               texture2d_array<half, access::write> outTexture [[texture(1)]],
-//                               uint3 gid [[thread_position_in_grid]]) {
-//    if (gid.x >= inTexture.get_width() ||
-//        gid.y >= inTexture.get_height()){
-//        return;
-//    }
-//    const half4 input = inTexture.read(gid.xy);
-//    outTexture.write(input, gid.xy, 0);
-//}
-
-kernel void texture2d_to_2d_array(texture2d<float, access::read> inTexture [[texture(0)]],
-                                  texture2d_array<float, access::write> outTexture [[texture(1)]],
-                                  uint3 gid [[thread_position_in_grid]]) {
-    if (gid.x >= inTexture.get_width() ||
-        gid.y >= inTexture.get_height()){
-        return;
-    }
-    const float4 input = inTexture.read(gid.xy);
-    outTexture.write(input, gid.xy, 0);
-}
-
-
-kernel void texture2d_to_2d_array_half(texture2d<half, access::read> inTexture [[texture(0)]],
-                                  texture2d_array<half, access::write> outTexture [[texture(1)]],
-                                  uint3 gid [[thread_position_in_grid]]) {
-    if (gid.x >= inTexture.get_width() ||
-        gid.y >= inTexture.get_height()){
-        return;
-    }
-    const half4 input = inTexture.read(gid.xy);
-    outTexture.write(input, gid.xy, 0);
-}
-
-struct PoolParam {
-    int ksizeX;
-    int ksizeY;
-    int strideX;
-    int strideY;
-    int paddingX;
-    int paddingY;
-    int poolType;
-};
-
-kernel void pool(texture2d_array<float, access::read> inTexture [[texture(0)]],
-                 texture2d_array<float, access::write> outTexture [[texture(1)]],
-                 constant PoolParam &pm [[buffer(0)]],
-                 uint3 gid [[thread_position_in_grid]]) {
-    if (gid.x >= outTexture.get_width() ||
-        gid.y >= outTexture.get_height() ||
-        gid.z >= outTexture.get_array_size()) return;
-    int xmin = gid.x * pm.strideX - pm.paddingX;
-    int xmax = min(xmin + pm.ksizeX, int(inTexture.get_width()));
-    xmin = max(xmin, 0);
-    int ymin = gid.y * pm.strideX - pm.paddingX;
-    int ymax = min(ymin + pm.ksizeX, int(inTexture.get_height()));
-    ymin = max(ymin, 0);
-    
-    float4 r = 0;
-    if (pm.poolType == 0) {
-        r = inTexture.read(uint2(xmin, ymin), gid.z);
-        for (int x = xmin; x < xmax; x++) {
-            for (int y = ymin; y < ymax; y++) {
-                r = fmax(r, inTexture.read(uint2(x, y), gid.z));
-            }
-        }
-    } else if (pm.poolType == 1) {
-        for (int x = xmin; x < xmax; x++) {
-            for (int y = ymin; y < ymax; y++) {
-                r += inTexture.read(uint2(x, y), gid.z);
-            }
-        }
-        r /= pm.ksizeX * pm.ksizeY;
-    }
-    outTexture.write(r, gid.xy, gid.z);
-}
-
-
-kernel void pool_half(texture2d_array<half, access::read> inTexture [[texture(0)]],
-                 texture2d_array<half, access::write> outTexture [[texture(1)]],
-                 constant PoolParam &pm [[buffer(0)]],
-                 uint3 gid [[thread_position_in_grid]]) {
-    if (gid.x >= outTexture.get_width() ||
-        gid.y >= outTexture.get_height() ||
-        gid.z >= outTexture.get_array_size()) return;
-    int xmin = gid.x * pm.strideX - pm.paddingX;
-    int xmax = min(xmin + pm.ksizeX, int(inTexture.get_width()));
-    xmin = max(xmin, 0);
-    int ymin = gid.y * pm.strideX - pm.paddingX;
-    int ymax = min(ymin + pm.ksizeX, int(inTexture.get_height()));
-    ymin = max(ymin, 0);
-    
-    half4 r = 0;
-    if (pm.poolType == 0) {
-        r = inTexture.read(uint2(xmin, ymin), gid.z);
-        for (int x = xmin; x < xmax; x++) {
-            for (int y = ymin; y < ymax; y++) {
-                r = fmax(r, inTexture.read(uint2(x, y), gid.z));
-            }
-        }
-    } else if (pm.poolType == 1) {
-        for (int x = xmin; x < xmax; x++) {
-            for (int y = ymin; y < ymax; y++) {
-                r += inTexture.read(uint2(x, y), gid.z);
-            }
-        }
-        r /= pm.ksizeX * pm.ksizeY;
-    }
-    outTexture.write(r, gid.xy, gid.z);
-}
-
-kernel void reshape(texture2d_array<float, access::read> inTexture [[texture(0)]],
-                    texture2d_array<float, access::write> outTexture [[texture(1)]],
-                    uint3 gid [[thread_position_in_grid]]) {
-    if (gid.x >= outTexture.get_width() ||
-        gid.y >= outTexture.get_height() ||
-        gid.z >= outTexture.get_array_size()) return;
-    
-    float4 r = inTexture.read(uint2(0, 0), gid.z);
-    outTexture.write(r, gid.xy, gid.z);
-}
-
-kernel void reshape_half(texture2d_array<half, access::read> inTexture [[texture(0)]],
-                    texture2d_array<half, access::write> outTexture [[texture(1)]],
-                    uint3 gid [[thread_position_in_grid]]) {
-    if (gid.x >= outTexture.get_width() ||
-        gid.y >= outTexture.get_height() ||
-        gid.z >= outTexture.get_array_size()) return;
-    
-    half4 r = inTexture.read(uint2(0, 0), gid.z);
-    outTexture.write(r, gid.xy, gid.z);
-}
-
-kernel void softmax(texture2d_array<float, access::read> inTexture [[texture(0)]],
-                    texture2d_array<float, access::write> outTexture [[texture(1)]],
-                    uint3 gid [[thread_position_in_grid]]) {
-    if (gid.x >= outTexture.get_width() ||
-        gid.y >= outTexture.get_height() ||
-        gid.z >= outTexture.get_array_size()) return;
-    int zsize = inTexture.get_array_size();
-    float maxv = inTexture.read(uint2(0, 0), 0)[0];
-    for (int z = 0; z < zsize; z++) {
-        float4 r = inTexture.read(uint2(0, 0), z);
-        maxv = max(maxv, max(max(r[0], r[1]), max(r[2], r[3])));
-    }
-    float sum = 0;
-    for (int z = 0; z < zsize; z++) {
-        float4 r = inTexture.read(uint2(0, 0), z);
-        sum += exp(r[0] - maxv) + exp(r[1] - maxv) + exp(r[2] - maxv) + exp(r[3] - maxv);
-    }
-    float4 rr = inTexture.read(gid.xy, gid.z);
-    rr = exp(rr - maxv) / sum;
-    outTexture.write(rr, gid.xy, gid.z);
-}
-
-
-kernel void softmax_half(texture2d_array<half, access::read> inTexture [[texture(0)]],
-                    texture2d_array<half, access::write> outTexture [[texture(1)]],
-                    uint3 gid [[thread_position_in_grid]]) {
-    if (gid.x >= outTexture.get_width() ||
-        gid.y >= outTexture.get_height() ||
-        gid.z >= outTexture.get_array_size()) return;
-    int zsize = inTexture.get_array_size();
-    half maxv = inTexture.read(uint2(0, 0), 0)[0];
-    for (int z = 0; z < zsize; z++) {
-        half4 r = inTexture.read(uint2(0, 0), z);
-        maxv = max(maxv, max(max(r[0], r[1]), max(r[2], r[3])));
-    }
-    float sum = 0;
-    for (int z = 0; z < zsize; z++) {
-        half4 r = inTexture.read(uint2(0, 0), z);
-        sum += exp(r[0] - maxv) + exp(r[1] - maxv) + exp(r[2] - maxv) + exp(r[3] - maxv);
-    }
-    half4 rr = inTexture.read(gid.xy, gid.z);
-    rr = exp(rr - maxv) / sum;
-    outTexture.write(rr, gid.xy, gid.z);
-}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/MulticlassNMSKernel.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class MulticlassNMSKernel<P: PrecisionType>: Kernel, Computable{
+  let pipline1: MTLComputePipelineState
+
+  required init(device: MTLDevice, param: MulticlassNMSParam<P>) {
+    
+    param.middleOutput.initBuffer(device: device)
+    param.bboxOutput.initBuffer(device: device)
+    if computePrecision == .Float32 {
+      pipline1 = device.pipeLine(funcName: "nms_fetch_bbox", inPaddleMobileLib: true)
+      super.init(device: device, inFunctionName: "nms_fetch_result")
+    } else if computePrecision == .Float16 {
+      pipline1 = device.pipeLine(funcName: "nms_fetch_bbox_half", inPaddleMobileLib: true)
+      super.init(device: device, inFunctionName: "nms_fetch_result_half")
+    } else {
+      fatalError( " unsupport precision " )
+    }
+    
+  }
+  
+  func compute(commandBuffer: MTLCommandBuffer, param: MulticlassNMSParam<P>) throws {
+    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+      throw PaddleMobileError.predictError(message: " encode is nil")
+    }
+    
+    encoder.setTexture(param.scores.metalTexture, index: 0)
+    encoder.setBuffer(param.middleOutput.resultBuffer!, offset: 0, index: 0)
+    encoder.dispatch(computePipline: pipline, outTexture: param.scores.metalTexture)
+    encoder.endEncoding()
+    
+    guard let encoderBox = commandBuffer.makeComputeCommandEncoder() else {
+      throw PaddleMobileError.predictError(message: " encode is nil")
+    }
+    
+    encoderBox.setTexture(param.bboxes.metalTexture, index: 0)
+    encoderBox.setBuffer(param.bboxOutput.resultBuffer!, offset: 0, index: 0)
+    encoderBox.dispatch(computePipline: pipline1, outTexture: param.bboxes.metalTexture)
+    encoderBox.endEncoding()
+  }
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/PoolKernel.swift

@@ -25,12 +25,10 @@ struct PoolMetalParam {
 }
 
 class PoolKernel<P: PrecisionType>: Kernel, Computable{
-    func compute(commandBuffer: MTLCommandBuffer, param: PoolParam<P>) throws {
-        guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
-            throw PaddleMobileError.predictError(message: " encoder is nil")
-        }
-        encoder.setTexture(param.input.metalTexture, index: 0)
-        encoder.setTexture(param.output.metalTexture, index: 1)
+  var metalParam: PoolMetalParam
+  required init(device: MTLDevice, param: PoolParam<P>) {
+    param.output.initTexture(device: device, inTranspose: param.input.transpose, computePrecision: computePrecision)
+    
     var poolType: Int32
     switch param.poolType {
     case "max":
@@ -38,9 +36,9 @@ class PoolKernel<P: PrecisionType>: Kernel, Computable{
     case "avg":
       poolType = 1
     default:
-            throw PaddleMobileError.predictError(message: " unknown pooltype " + param.poolType)
+      fatalError()
     }
-        var pmp = PoolMetalParam.init(
+    metalParam = PoolMetalParam.init(
       ksizeX: param.ksize[0],
       ksizeY: param.ksize[1],
       strideX: param.stride[0],
@@ -49,12 +47,25 @@ class PoolKernel<P: PrecisionType>: Kernel, Computable{
       paddingY: param.padding[1],
       poolType: poolType
     )
-        encoder.setBytes(&pmp, length: MemoryLayout<PoolMetalParam>.size, index: 0)
-        encoder.dispatch(computePipline: pipline, outTexture: param.output.metalTexture)
-        encoder.endEncoding()
-    }
     
-    required init(device: MTLDevice, param: PoolParam<P>) {
+    if computePrecision == .Float32 {
       super.init(device: device, inFunctionName: "pool")
+    } else if computePrecision == .Float16 {
+      super.init(device: device, inFunctionName: "pool_half")
+    } else {
+      fatalError()
+    }
+  }
+  
+  func compute(commandBuffer: MTLCommandBuffer, param: PoolParam<P>) throws {
+    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+      throw PaddleMobileError.predictError(message: " encoder is nil")
+    }
+    encoder.setTexture(param.input.metalTexture, index: 0)
+    encoder.setTexture(param.output.metalTexture, index: 1)
+
+    encoder.setBytes(&metalParam, length: MemoryLayout<PoolMetalParam>.size, index: 0)
+    encoder.dispatch(computePipline: pipline, outTexture: param.output.metalTexture)
+    encoder.endEncoding()
   }
 }

metal/paddle-mobile/paddle-mobile/Operators/Kernels/PreluKernel.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class PreluKernel<P: PrecisionType>: Kernel, Computable{
+  required init(device: MTLDevice, param: PreluParam<P>) {
+    param.alpha.initBuffer(device: device, precision: computePrecision)
+    param.output.initTexture(device: device, inTranspose: param.input.transpose, computePrecision: computePrecision)
+    if computePrecision == .Float32 {
+      if param.mode == "channel" {
+        super.init(device: device, inFunctionName: "prelu_channel")
+      } else if param.mode == "element" {
+        super.init(device: device, inFunctionName: "prelu_element")
+      } else {
+        super.init(device: device, inFunctionName: "prelu_other")
+      }
+    } else if computePrecision == .Float16 {
+      if param.mode == "channel" {
+        super.init(device: device, inFunctionName: "prelu_channel_half")
+      } else if param.mode == "element" {
+        super.init(device: device, inFunctionName: "prelu_element_half")
+      } else {
+        super.init(device: device, inFunctionName: "prelu_other_half")
+      }
+    } else {
+      fatalError()
+    }
+  }
+  
+  func compute(commandBuffer: MTLCommandBuffer, param: PreluParam<P>) throws {
+    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+      throw PaddleMobileError.predictError(message: " encoder is nil")
+    }
+    
+    encoder.setTexture(param.input.metalTexture, index: 0)
+    encoder.setTexture(param.output.metalTexture, index: 1)
+    encoder.setBuffer(param.alpha.buffer, offset: 0, index: 0)
+    encoder.dispatch(computePipline: pipline, outTexture: param.output.metalTexture)
+    encoder.endEncoding()
+  }
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/PriorBoxKernel.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+struct PriorBoxMetalParam {
+  let offset: Float32
+  let stepWidth: Float32
+  let stepHeight: Float32
+  let minSize: Float32
+  let maxSize: Float32
+  let imageWidth: Float32
+  let imageHeight: Float32
+  let clip: Bool
+  let numPriors: uint
+  let aspecRatiosSize: uint
+  let minSizeSize: uint
+  let maxSizeSize: uint
+}
+
+class PriorBoxKernel<P: PrecisionType>: Kernel, Computable{
+  var metalParam: PriorBoxMetalParam!
+  
+  required init(device: MTLDevice, param: PriorBoxParam<P>) {
+    
+    let originDim = param.output.tensorDim;
+    
+    param.output.tensorDim = Dim.init(inDim: [1, originDim[0], originDim[1], originDim[2] * originDim[3]])
+    param.output.padToFourDim = Dim.init(inDim: [1, originDim[0], originDim[1], originDim[2] * originDim[3]])
+    
+    param.output.initTexture(device: device, inTranspose: [0, 1, 2, 3], computePrecision: computePrecision)
+    param.outputVariances.initTexture(device: device, inTranspose: [2, 0, 1, 3], computePrecision: computePrecision)
+    
+    
+    if computePrecision == .Float32 {
+      if param.min_max_aspect_ratios_order {
+        super.init(device: device, inFunctionName: "prior_box_MinMaxAspectRatiosOrder")
+      } else {
+        super.init(device: device, inFunctionName: "prior_box")
+      }
+      
+    } else if computePrecision == .Float16 {
+      if param.min_max_aspect_ratios_order {
+        super.init(device: device, inFunctionName: "prior_box_MinMaxAspectRatiosOrder_half")
+      } else {
+        super.init(device: device, inFunctionName: "prior_box_half")
+      }
+    } else {
+      fatalError()
+    }
+    
+    
+    guard param.minSizes.count == 1 else {
+      fatalError(" need implement ")
+    }
+    
+//    let n = 1
+//    let h = param.output.dim[1]
+//    let w = param.output.dim[2]
+//    let c = param.output.dim[3] * param.output.dim[0]
+//
+//    param.output.dim = Dim.init(inDim: [n, h, w, c])
+//    param.output.transpose = [0, 1, 2, 3]
+    
+    let imageWidth = Float32(param.inputImage.padToFourDim[3])
+    let imageHeight = Float32(param.inputImage.padToFourDim[2])
+    
+    let featureWidth = param.input.padToFourDim[3]
+    let featureHeight = param.input.padToFourDim[2]
+    
+    if param.stepW == 0 || param.stepH == 0 {
+      param.stepW = Float32(imageWidth) / Float32(featureWidth)
+      param.stepH = Float32(imageHeight) / Float32(featureHeight)
+    }
+    
+    var outputAspectRatior: [Float32] = []
+    outputAspectRatior.append(1.0)
+    
+    let epsilon = 1e-6
+    for ar in param.aspectRatios {
+      var alreadyExist = false
+      for outputAr in outputAspectRatior {
+        if fabs(Double(ar) - Double(outputAr)) < Double(epsilon) {
+          alreadyExist = true
+          break
+        }
+      }
+      
+      if !alreadyExist {
+        outputAspectRatior.append(ar)
+      }
+      if param.flip {
+        outputAspectRatior.append(1.0 / ar)
+      }
+    }
+    
+    if computePrecision == .Float16 {
+      let buffer = device.makeBuffer(length: outputAspectRatior.count * MemoryLayout<Float16>.size)
+      float32ToFloat16(input: &outputAspectRatior, output:(buffer?.contents())!, count: outputAspectRatior.count)
+      param.newAspectRatios = buffer
+
+    } else if computePrecision == .Float32 {
+      let buffer = device.makeBuffer(bytes: outputAspectRatior, length: outputAspectRatior.count * MemoryLayout<Float32>.size, options: [])
+      param.newAspectRatios = buffer
+    } else {
+      fatalError()
+    }
+    
+    let aspectRatiosSize = uint(outputAspectRatior.count)
+    
+    let maxSizeSize: uint = uint(param.maxSizes.count)
+    let minSizeSize: uint = uint(param.minSizes.count)
+    
+    let numPriors = aspectRatiosSize * minSizeSize + maxSizeSize
+    
+    let minSize = param.minSizes.last ?? 0.0
+    let maxSize = param.maxSizes.last ?? 0.0
+    
+    metalParam = PriorBoxMetalParam.init(offset: param.offset, stepWidth: param.stepW, stepHeight: param.stepH, minSize: minSize, maxSize: maxSize, imageWidth: imageWidth, imageHeight: imageHeight, clip: param.clip, numPriors: numPriors, aspecRatiosSize: aspectRatiosSize, minSizeSize: minSizeSize, maxSizeSize: maxSizeSize)
+    
+  }
+  
+  func compute(commandBuffer: MTLCommandBuffer, param: PriorBoxParam<P>) throws {
+    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+      throw PaddleMobileError.predictError(message: " encode is nil")
+    }
+    
+    encoder.setTexture(param.input.metalTexture, index: 0)
+    encoder.setTexture(param.output.metalTexture, index: 1)
+    encoder.setTexture(param.outputVariances.metalTexture, index: 2)
+    
+    encoder.setBuffer(param.newAspectRatios!, offset: 0, index: 0)
+    
+    encoder.setBytes(&metalParam, length: MemoryLayout<PriorBoxMetalParam>.size, index: 1)
+    
+    encoder.setBytes(param.variances, length: MemoryLayout<Float32>.size * param.variances.count, index: 2)
+    encoder.dispatch(computePipline: pipline, outTexture: param.output.metalTexture)
+    encoder.endEncoding()
+  }
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/ReluKernel.swift

@@ -26,6 +26,12 @@ class ReluKernel<P: PrecisionType>: Kernel, Computable{
   }
   
   required init(device: MTLDevice, param: ReluParam<P>) {
+    if computePrecision == .Float32 {
       super.init(device: device, inFunctionName: "relu")
+    } else if computePrecision == .Float16 {
+      super.init(device: device, inFunctionName: "relu_half")
+    } else {
+      fatalError()
+    }
   }
 }

metal/paddle-mobile/paddle-mobile/Operators/Kernels/ReshapeKernel.swift

@@ -14,8 +14,59 @@
 
 import Foundation
 
+struct ReshapeMetalParam {
+  var idim: (Int32, Int32, Int32, Int32)
+  var itrans: (Int32, Int32, Int32, Int32)
+  var odim: (Int32, Int32, Int32, Int32)
+  var otrans: (Int32, Int32, Int32, Int32)
+}
+
+struct ReshapeTestParam: TestParam {
+  let inputTexture: MTLTexture
+  let outputTexture: MTLTexture
+  let param: ReshapeMetalParam
+}
+
 class ReshapeKernel<P: PrecisionType>: Kernel, Computable{
+  
+  var metalParam: ReshapeMetalParam
+  
   required init(device: MTLDevice, param: ReshapeParam<P>) {
+    param.output.initTexture(device: device, computePrecision: computePrecision)
+    var id: [Int32] = [1, 1, 1, 1]
+    for i in 0..<param.input.tensorDim.cout() {
+      id[4-param.input.tensorDim.cout()+i] = Int32(param.input.tensorDim[i])
+    }
+    let it: [Int32] = param.input.transpose.map { Int32($0) }
+    var od: [Int32] = [1, 1, 1, 1]
+    for i in 0..<param.output.tensorDim.cout() {
+      od[4-param.output.tensorDim.cout()+i] = Int32(param.output.tensorDim[i])
+    }
+    let ot: [Int32] = param.output.transpose.map { Int32($0) }
+    metalParam = ReshapeMetalParam.init(
+      idim: (id[0], id[1], id[2], id[3]),
+      itrans: (it[0], it[1], it[2], it[3]),
+      odim: (od[0], od[1], od[2], od[3]),
+      otrans: (ot[0], ot[1], ot[2], ot[3])
+    )
+    let irank = param.input.tensorDim.cout()
+    let orank = param.output.tensorDim.cout()
+    if computePrecision == .Float32 {
+      super.init(device: device, inFunctionName: "reshape_\(irank)_\(orank)_float")
+    } else if computePrecision == .Float16 {
+      super.init(device: device, inFunctionName: "reshape_\(irank)_\(orank)_half")
+    } else {
+      fatalError()
+    }
+  }
+  
+  required init(device: MTLDevice, testParam: ReshapeTestParam) {
+    metalParam = ReshapeMetalParam.init(
+    idim: (0, 0, 0, 0),
+    itrans: (0, 0, 0, 0),
+    odim: (0, 0, 0, 0),
+    otrans: (0, 0, 0, 0)
+    )
     super.init(device: device, inFunctionName: "reshape")
   }
   
@@ -23,9 +74,24 @@ class ReshapeKernel<P: PrecisionType>: Kernel, Computable{
     guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
       throw PaddleMobileError.predictError(message: " encoder is nil")
     }
+
     encoder.setTexture(param.input.metalTexture, index: 0)
     encoder.setTexture(param.output.metalTexture, index: 1)
+
+    encoder.setBytes(&metalParam, length: MemoryLayout<ReshapeMetalParam>.size, index: 0)
     encoder.dispatch(computePipline: pipline, outTexture: param.output.metalTexture)
     encoder.endEncoding()
   }
+  
+//  func test(commandBuffer: MTLCommandBuffer, testParam: ReshapeTestParam) {
+//    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+//      fatalError()
+//    }
+//    encoder.setTexture(testParam.inputTexture, index: 0)
+//    encoder.setTexture(testParam.outputTexture, index: 1)
+//    var pm: ReshapeMetalParam = testParam.param
+//    encoder.setBytes(&pm, length: MemoryLayout<ReshapeMetalParam>.size, index: 0)
+//    encoder.dispatch(computePipline: pipline, outTexture: testParam.outputTexture)
+//    encoder.endEncoding()
+//  }
 }

metal/paddle-mobile/paddle-mobile/Operators/Kernels/ResizeBilinearKernel.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+struct ResizeBilinearMetalParam {
+  var ratio_h: Float32
+  var ratio_w: Float32
+}
+
+class ResizeBilinearKernel<P: PrecisionType>: Kernel, Computable{
+  func compute(commandBuffer: MTLCommandBuffer, param: ResizeBilinearParam<P>) throws {
+    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+      throw PaddleMobileError.predictError(message: " encode is nil")
+    }
+    
+    encoder.setTexture(param.input.metalTexture, index: 0)
+    encoder.setTexture(param.output.metalTexture, index: 1)
+    let ratio_h: Float32 = Float32(param.input.tensorDim.dims[2]) / Float32(param.output.tensorDim.dims[2])
+    let ratio_w: Float32 = Float32(param.input.tensorDim.dims[3]) / Float32(param.output.tensorDim.dims[3])
+    var p = ResizeBilinearMetalParam.init(ratio_h: ratio_h, ratio_w: ratio_w)
+    encoder.setBytes(&p, length: MemoryLayout<ConcatMetalParam>.size, index: 0)
+    encoder.dispatch(computePipline: pipline, outTexture: param.output.metalTexture)
+    encoder.endEncoding()
+  }
+  
+  required init(device: MTLDevice, param: ResizeBilinearParam<P>) {
+    param.output.initTexture(device: device, inTranspose: param.input.transpose, computePrecision: computePrecision)
+    if computePrecision == .Float32 {
+      super.init(device: device, inFunctionName: "resize_bilinear")
+    } else if computePrecision == .Float16 {
+      super.init(device: device, inFunctionName: "resize_bilinear_half")
+    } else {
+      fatalError()
+    }
+  }
+  
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/ResizeKernel.swift

-/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
- 
- Licensed under the Apache License, Version 2.0 (the "License");
- you may not use this file except in compliance with the License.
- You may obtain a copy of the License at
- 
- http://www.apache.org/licenses/LICENSE-2.0
- 
- Unless required by applicable law or agreed to in writing, software
- distributed under the License is distributed on an "AS IS" BASIS,
- WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- See the License for the specific language governing permissions and
- limitations under the License. */
-//
-//import Foundation
-//import MetalPerformanceShaders
-//
-//
-//struct ResizeParam: OpParam{
-//    typealias OutputType = <#type#>
-//    
-//    typealias ParamPrecisionType = <#type#>
-//    
-//    let input: MTLTexture
-//    let output: MTLTexture
-//    let expectDim: Dim
-//}
-//
-//struct OutputDim {
-//    let width: UInt16
-//    let height: UInt16
-//    let strideX: UInt16
-//    let strideY: UInt16
-//}
-//
-//class ResizeKernel<P: PrecisionType>: Kernel, Computable{
-//    var lanczos: MPSImageLanczosScale
-//    required init(device: MTLDevice, param: ResizeParam) {
-//        lanczos = MPSImageLanczosScale.init(device: device)
-//        super.init(device: device, inFunctionName: "resize")
-//    }
-//    func compute(commandBuffer: MTLCommandBuffer, param: ResizeParam) throws {
-////        guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
-////            throw PaddleMobileError.predictError(message: " encode is nil")
-////        }
-//        lanczos.encode(commandBuffer: commandBuffer, sourceTexture: param.input, destinationTexture: param.output)
-//        
-////        encoder.setTexture(param.input, index: 0)
-////        encoder.setTexture(param.output, index: 1)
-////        let strideX = param.input.width/param.expectDim[2]
-////        let strideY = param.input.height/param.expectDim[1]
-////        var outputDim = OutputDim.init(width: UInt16(param.expectDim[1]), height: UInt16(param.expectDim[2]), strideX: UInt16(strideX), strideY: UInt16(strideY))
-////        encoder.setBytes(&outputDim, length: MemoryLayout<OutputDim>.size, index: 0)
-////        encoder.dispatch(computePipline: pipline, outTexture: param.output)
-////        encoder.endEncoding()
-//    }
-//    
-//
-//    
-//    
-//}
-

metal/paddle-mobile/paddle-mobile/Operators/Kernels/ShapeKernel.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+struct ShapeMetalParam {
+}
+
+class ShapeKernel<P: PrecisionType>: Kernel, Computable{
+  func compute(commandBuffer: MTLCommandBuffer, param: ShapeParam<P>) throws {
+//    print("shape compute")
+//    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+//      throw PaddleMobileError.predictError(message: " encode is nil")
+//    }
+//    encoder.setTexture(param.output.metalTexture, index: 0)
+//    encoder.endEncoding()
+  }
+  
+  required init(device: MTLDevice, param: ShapeParam<P>) {
+    param.output.initTexture(device: device, computePrecision: computePrecision)
+    if computePrecision == .Float32 {
+      super.init(device: device, inFunctionName: "shape")
+    } else if computePrecision == .Float16 {
+      super.init(device: device, inFunctionName: "shape_half")
+    } else {
+      fatalError()
+    }
+  }
+  
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/SoftmaxKernel.swift

@@ -14,19 +14,38 @@
 
 import Foundation
 
+struct SoftmaxMetalParam {
+  let N: Int32
+  let K: Int32
+}
+
 class SoftmaxKernel<P: PrecisionType>: Kernel, Computable{
   
+  var metalParam: SoftmaxMetalParam
+  required init(device: MTLDevice, param: SoftmaxParam<P>) {
+    param.output.initTexture(device: device, computePrecision: computePrecision)
+    metalParam = SoftmaxMetalParam.init(
+      N: Int32(param.input.tensorDim[0]),
+      K: Int32(param.input.tensorDim[1])
+    )
+    if computePrecision == .Float32 {
+      super.init(device: device, inFunctionName: "softmax_float")
+    } else if computePrecision == .Float16 {
+      super.init(device: device, inFunctionName: "softmax_half")
+    } else {
+      fatalError()
+    }
+  }
+
   func compute(commandBuffer: MTLCommandBuffer, param: SoftmaxParam<P>) throws {
     guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
       throw PaddleMobileError.predictError(message: " encoder is nil")
     }
     encoder.setTexture(param.input.metalTexture, index: 0)
     encoder.setTexture(param.output.metalTexture, index: 1)
+    encoder.setBytes(&metalParam, length: MemoryLayout<SoftmaxMetalParam>.size, index: 0)
     encoder.dispatch(computePipline: pipline, outTexture: param.output.metalTexture)
     encoder.endEncoding()
   }
   
-    required init(device: MTLDevice, param: SoftmaxParam<P>) {
-        super.init(device: device, inFunctionName: "softmax")
-    }
 }

metal/paddle-mobile/paddle-mobile/Operators/Kernels/SplitKernel.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+struct SplitMetalParam {
+  var idim: (Int32, Int32, Int32, Int32) = (1, 1, 1, 1)
+  var axis: Int32 = 0
+  var offset: Int32 = 0
+  var trans: (Int32, Int32, Int32, Int32) = (0, 1, 2, 3)
+  var vdim: (Int32, Int32, Int32, Int32) = (0, 0, 0, 0)
+}
+
+class SplitKernel<P: PrecisionType>: Kernel, Computable{
+  var smp: SplitMetalParam
+  func compute(commandBuffer: MTLCommandBuffer, param: SplitParam<P>) throws {
+    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+      throw PaddleMobileError.predictError(message: " encode is nil")
+    }
+    encoder.setTexture(param.input.metalTexture, index: 0)
+    for i in 0..<param.outputList.count {
+      encoder.setTexture(param.outputList[i].metalTexture, index: i + 1)
+    }
+    encoder.setBytes(&smp, length: MemoryLayout<SplitMetalParam>.size, index: 0)
+    encoder.dispatch(computePipline: pipline, outTexture: param.input.metalTexture)
+    encoder.endEncoding()
+  }
+  
+  required init(device: MTLDevice, param: SplitParam<P>) {
+    //     param.output.initTexture(device: device, computePrecision: computePrecision)
+    let num = param.outputList.count
+    let rank = param.input.tensorDim.cout()
+    assert(num >= 2 && num <= 4)
+    for output in param.outputList {
+      output.initTexture(device: device, inTranspose: param.input.transpose, computePrecision: computePrecision)
+    }
+    smp = SplitMetalParam.init()
+    smp.idim = (Int32(param.input.dim[0]), Int32(param.input.dim[1]), Int32(param.input.dim[2]), Int32(param.input.dim[3]))
+    smp.axis = Int32(param.axis + param.input.dim.cout() - param.input.tensorDim.cout())
+    for i in 0..<4 {
+      if param.input.transpose[i] == smp.axis {
+        smp.axis = Int32(i)
+        break
+      }
+    }
+    smp.trans = (Int32(param.input.transpose[0]), Int32(param.input.transpose[1]), Int32(param.input.transpose[2]), Int32(param.input.transpose[3]))
+    var vdim: [Int32] = [0, 0, 0, 0]
+    for i in 0..<num {
+      vdim[i] = Int32(param.outputList[i].tensorDim[param.axis])
+    }
+    smp.vdim = (vdim[0], vdim[1], vdim[2], vdim[3])
+    var v = "normal"
+    if rank == 4 {
+      if smp.axis == 1 {
+        v = "y"
+      } else if smp.axis == 2 {
+        v = "x"
+      }
+    } else if rank == 3 {
+      if smp.axis == 2 {
+        v = "y"
+      } else if smp.axis == 3 {
+        v = "x"
+      }
+    } else if rank == 2 {
+      if smp.axis == 2 {
+        v = "y"
+      }
+    }
+    if v == "normal" {
+      fatalError("split unsupported")
+    }
+    if computePrecision == .Float32 {
+      super.init(device: device, inFunctionName: "split_\(rank)_\(num)_\(v)_float")
+    } else if computePrecision == .Float16 {
+      super.init(device: device, inFunctionName: "split_\(rank)_\(num)_\(v)_half")
+    } else {
+      fatalError()
+    }
+  }
+  
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/Texture2DTo2DArrayKernel.swift

@@ -32,6 +32,14 @@ class Texture2DTo2DArrayKernel<P: PrecisionType>: Kernel, Computable{
   }
   
   required init(device: MTLDevice, param: FeedParam<P>) {
+    param.output.initTexture(device: device, inTranspose: [0, 2, 3, 1], computePrecision: computePrecision)
+    if computePrecision == .Float16 {
+      super.init(device: device, inFunctionName: "texture2d_to_2d_array_half")
+    } else if computePrecision == .Float32 {
       super.init(device: device, inFunctionName: "texture2d_to_2d_array")
+    } else {
+      fatalError()
+    }
+    
   }
 }

metal/paddle-mobile/paddle-mobile/Operators/Kernels/TransposeKernel.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+struct TransposeMetalParam {
+  var iC: Int32 = 0
+  var oC: Int32 = 0
+  var axis: (Int32, Int32, Int32, Int32) = (0, 1, 2, 3)
+}
+
+class TransposeKernel<P: PrecisionType>: Kernel, Computable {
+  var metalParam: TransposeMetalParam = TransposeMetalParam.init()
+  required init(device: MTLDevice, param: TransposeParam<P>) {
+    param.output.initTexture(device: device, computePrecision: computePrecision)
+    let rank = param.input.tensorDim.cout()
+    var axis: [Int] = [0, 1, 2, 3]
+    for i in 0..<param.axis.count {
+      axis[4-rank+i] = 4 - rank + Int(param.axis[i])
+    }
+
+    var naxis: [Int] = [0, 0, 0, 0]
+    for i in 0..<4 {
+      for j in 0..<4 {
+        if param.input.transpose[j] == axis[i] {
+          naxis[i] = j
+          break
+        }
+      }
+    }
+    metalParam.iC = Int32(param.input.dim[param.input.transpose[3]])
+    metalParam.oC = Int32(param.output.dim[3])
+    metalParam.axis = (Int32(naxis[0]), Int32(naxis[1]), Int32(naxis[2]), Int32(naxis[3]))
+    var kernelFunc = "transpose_undefined"
+    if computePrecision == .Float16 {
+      if param.input.transpose == axis {
+        kernelFunc = "transpose_copy_half"
+      } else {
+        kernelFunc = "transpose_\(rank)_half"
+      }
+    } else if computePrecision == .Float32 {
+      if param.input.transpose == axis {
+        kernelFunc = "transpose_copy_float"
+      } else {
+        kernelFunc = "transpose_\(rank)_float"
+      }
+    } else {
+      fatalError()
+    }
+    print("===========>", kernelFunc)
+    print(metalParam)
+    super.init(device: device, inFunctionName: kernelFunc)
+  }
+  
+  func compute(commandBuffer: MTLCommandBuffer, param: TransposeParam<P>) throws {
+    guard let encoder = commandBuffer.makeComputeCommandEncoder() else {
+      throw PaddleMobileError.predictError(message: " encode is nil")
+    }
+  
+    encoder.setTexture(param.input.metalTexture, index: 0)
+    encoder.setTexture(param.output.metalTexture, index: 1)
+    encoder.setBytes(&metalParam, length: MemoryLayout<TransposeMetalParam>.size, index: 0)
+    encoder.dispatch(computePipline: pipline, outTexture: param.output.metalTexture)
+    encoder.endEncoding()
+  }
+  
+
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/BatchNormKernel.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+using namespace metal;
+
+kernel void batchnorm(texture2d_array<float, access::read> inTexture [[texture(0)]],
+                      texture2d_array<float, access::write> outTexture [[texture(1)]],
+                      const device float4 * nscale [[buffer(0)]],
+                      const device float4 * nbias [[buffer(1)]],
+                      uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) return;
+  const float4 input = inTexture.read(gid.xy, gid.z);
+  float4 output = input * nscale[gid.z] + nbias[gid.z];
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+kernel void batchnorm_half(texture2d_array<half, access::read> inTexture [[texture(0)]],
+                      texture2d_array<half, access::write> outTexture [[texture(1)]],
+                      const device half4 * newScale [[buffer(0)]],
+                      const device half4 * newBias [[buffer(1)]],
+                      uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) return;
+  const half4 input = inTexture.read(gid.xy, gid.z);
+  half4 output = input * newScale[gid.z] + newBias[gid.z];
+  outTexture.write(output, gid.xy, gid.z);
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/BatchNormRelu.metal

+//
+//  BatchNormRelu.metal
+//  paddle-mobile
+//
+
+#include <metal_stdlib>
+using namespace metal;
+
+struct MetalConvParam {
+    short offsetX;
+    short offsetY;
+    short offsetZ;
+    ushort strideX;
+    ushort strideY;
+};
+
+kernel void batch_norm_relu_3x3(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+                                         texture2d_array<float, access::write> outTexture [[texture(1)]],
+                                         const device float4 *new_scale [[buffer(0)]],
+                                         const device float4 *new_biase [[buffer(1)]],
+                                         uint3 gid [[thread_position_in_grid]]) {
+    
+    if (gid.x >= outTexture.get_width() ||
+        gid.y >= outTexture.get_height() ||
+        gid.z >= outTexture.get_array_size()) {
+        return;
+    }
+    
+    float4 input;
+    float4 output;
+    constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+    input = inTexture.sample(sample, gid.x, gid.y, gid.z);
+    output = fmax(input * new_scale[gid.z] + new_biase[gid.z], 0.0);
+    outTexture.write(output, gid.xy, gid.z);
+
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/BilinearInterp.inc.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#ifdef P
+
+#define CONCAT2(a, b) a ## b
+#define CONCAT2_(a, b) a ## _ ## b
+
+#define FUNC(f, p) CONCAT2_(f, p)
+#define VECTOR(p, n) CONCAT2(p, n)
+
+kernel void FUNC(bilinear_interp, P)(texture2d_array<P, access::read> input [[texture(0)]],
+                     texture2d_array<P, access::write> output [[texture(1)]],
+                     constant bilinear_interp_param & pm [[buffer(0)]],
+                     uint3 gid [[thread_position_in_grid]]) {
+  VECTOR(P, 4) r;
+  if ((input.get_width() == output.get_width()) && (input.get_height() == output.get_height())) {
+    r = input.read(gid.xy, gid.z);
+  } else {
+    P w = gid.x * pm.ratio_w;
+    P h = gid.y * pm.ratio_h;
+    uint w0 = w, h0 = h;
+    uint w1 = w0 + 1, h1 = h0 + 1;
+    P w1lambda = w - w0, h1lambda = h - h0;
+    P w2lambda = 1.0 - w1lambda, h2lambda = 1.0 - h1lambda;
+    if (w1 >= input.get_width()) w1 = w0;
+    if (h1 >= input.get_height()) h1 = h0;
+    VECTOR(P, 4) r0 = input.read(uint2(w0, h0), gid.z);
+    VECTOR(P, 4) r1 = input.read(uint2(w1, h0), gid.z);
+    VECTOR(P, 4) r2 = input.read(uint2(w0, h1), gid.z);
+    VECTOR(P, 4) r3 = input.read(uint2(w1, h1), gid.z);
+    r = h2lambda * (w2lambda * r0 + w1lambda * r1)
+      + h1lambda * (w2lambda * r2 + w1lambda * r3);
+  }
+  output.write(r, gid.xy, gid.z);
+}
+
+#endif

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/BilinearInterp.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+using namespace metal;
+
+struct bilinear_interp_param {
+  float ratio_h;
+  float ratio_w;
+};
+
+#define P float
+#include "BilinearInterp.inc.metal"
+#undef P
+
+#define P half
+#include "BilinearInterp.inc.metal"
+#undef P

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/BoxCoder.inc.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#ifdef P
+
+#define CONCAT2(a, b) a ## b
+#define CONCAT2_(a, b) a ## _ ## b
+
+#define FUNC(f, p) CONCAT2_(f, p)
+#define VECTOR(p, n) CONCAT2(p, n)
+kernel void FUNC(boxcoder, P)(texture2d_array<P, access::read> priorBox [[texture(0)]],
+                     texture2d_array<P, access::read> priorBoxVar [[texture(1)]],
+                     texture2d_array<P, access::read> targetBox [[texture(2)]],
+                     texture2d_array<P, access::write> output[[texture(3)]],
+                     uint3 gid [[thread_position_in_grid]]) {
+  VECTOR(P, 4) p = priorBox.read(uint2(0, gid.x), gid.z);
+  VECTOR(P, 4) pv = priorBoxVar.read(uint2(0, gid.x), gid.z);
+  VECTOR(P, 4) t;
+  t[0] = targetBox.read(uint2(0, gid.x), gid.z)[0];
+  t[1] = targetBox.read(uint2(1, gid.x), gid.z)[0];
+  t[2] = targetBox.read(uint2(2, gid.x), gid.z)[0];
+  t[3] = targetBox.read(uint2(3, gid.x), gid.z)[0];
+  
+  P px = (p.x + p.z) / 2;
+  P py = (p.y + p.w) / 2;
+  P pw = p.z - p.x;
+  P ph = p.w - p.y;
+  
+  P tx = pv.x * t.x * pw + px;
+  P ty = pv.y * t.y * ph + py;
+  P tw = exp(pv.z * t.z) * pw;
+  P th = exp(pv.w * t.w) * ph;
+  
+  VECTOR(P, 4) r;
+  r.x = tx - tw / 2;
+  r.y = ty - th / 2;
+  r.z = tx + tw / 2;
+  r.w = ty + th / 2;
+
+  output.write(r, gid.xy, gid.z);
+}
+
+#endif

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/BoxCoder.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+using namespace metal;
+
+#define P float
+#include "BoxCoder.inc.metal"
+#undef P
+#define P half
+#include "BoxCoder.inc.metal"
+#undef P

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/Common.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+using namespace metal;
+
+
+inline void xyzn2abcd_1(int xyzn[4], int abcd[4]) {
+  abcd[0] = abcd[1] = abcd[2] = 0;
+  abcd[3] = xyzn[0] * 4 + xyzn[3];
+}
+inline void xyzn2abcd_2(int xyzn[4], int abcd[4]) {
+  abcd[0] = abcd[1] = 0;
+  abcd[2] = xyzn[1];
+  abcd[3] = xyzn[0] * 4 + xyzn[3];
+}
+inline void xyzn2abcd_3(int xyzn[4], int abcd[4]) {
+  abcd[0] = 0;
+  abcd[3] = xyzn[0];
+  abcd[2] = xyzn[1];
+  abcd[1] = xyzn[2] * 4 + xyzn[3];
+}
+inline void xyzn2abcd_4(int C, int xyzn[4], int abcd[4]) {
+  abcd[2] = xyzn[0];
+  abcd[1] = xyzn[1];
+  uint t = xyzn[2] * 4 + xyzn[3];
+  abcd[0] = t / C;
+  abcd[3] = t % C;
+}
+
+inline void abcd2xyzn_1(int abcd[4], int xyzn[4]) {
+  xyzn[1] = xyzn[2] = 0;
+  xyzn[0] = abcd[3] / 4;
+  xyzn[1] = abcd[3] % 4;
+}
+inline void abcd2xyzn_2(int abcd[4], int xyzn[4]) {
+  xyzn[2] = 0;
+  xyzn[1] = abcd[2];
+  xyzn[0] = abcd[3] / 4;
+  xyzn[3] = abcd[3] % 4;
+}
+inline void abcd2xyzn_3(int abcd[4], int xyzn[4]) {
+  xyzn[0] = abcd[3];
+  xyzn[1] = abcd[2];
+  xyzn[2] = abcd[1] / 4;
+  xyzn[3] = abcd[1] % 4;
+}
+inline void abcd2xyzn_4(int C, int abcd[4], int xyzn[4]) {
+  xyzn[0] = abcd[2];
+  xyzn[1] = abcd[1];
+  uint t = abcd[0] * C + abcd[3];
+  xyzn[2] = t / 4;
+  xyzn[3] = t % 4;
+}
+
+inline void xyzn2abcd(int C, int xyzn[4], int abcd[4]) {
+  abcd[2] = xyzn[0];
+  abcd[1] = xyzn[1];
+  uint t = xyzn[2] * 4 + xyzn[3];
+  abcd[0] = t / C;
+  abcd[3] = t % C;
+}
+
+inline void abcd2xyzn(int C, int abcd[4], int xyzn[4]) {
+  xyzn[0] = abcd[2];
+  xyzn[1] = abcd[1];
+  uint t = abcd[0] * C + abcd[3];
+  xyzn[2] = t / 4;
+  xyzn[3] = t % 4;
+}
+
+inline int32_t abcd2index(int32_t dim[4], int32_t abcd[4]) {
+  int32_t r = abcd[0];
+  r = r * dim[1] + abcd[1];
+  r = r * dim[2] + abcd[2];
+  r = r * dim[3] + abcd[3];
+  return r;
+}
+
+inline void index2abcd(int32_t dim[4], int32_t ind, int32_t abcd[4]) {
+  abcd[3] = ind % dim[3]; ind /= dim[3];
+  abcd[2] = ind % dim[2]; ind /= dim[2];
+  abcd[1] = ind % dim[1]; ind /= dim[1];
+  abcd[0] = ind;
+}
+
+inline void trans(int32_t trans[4], int32_t ipos[4], int32_t opos[4]) {
+  for (int i = 0; i < 4; i++) {
+    opos[i] = ipos[trans[i]];
+  }
+}
+
+inline void invtrans(int32_t trans[4], int32_t ipos[4], int32_t opos[4]) {
+  for (int i = 0; i < 4; i++) {
+    opos[trans[i]] = ipos[i];
+  }
+}
+
+
+struct MetalConvParam {
+  short offsetX;
+  short offsetY;
+  short offsetZ;
+  ushort strideX;
+  ushort strideY;
+  ushort dilationX;
+  ushort dilationY;
+};
+

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/ConcatKernel.inc.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#ifdef P
+
+#define CONCAT2(a, b) a ## b
+#define CONCAT2_(a, b) a ## _ ## b
+#define CONCAT3_(a, b, c) a ## _ ## b ## _ ## c
+#define CONCAT4_(a, b, c, d) a ## _ ## b ## _ ## c ## _ ## d
+#define CONCAT5_(a, b, c, d, e) a ## _ ## b ## _ ## c ## _ ## d ## _ ## e
+
+#define FUNC(f, r, n, v, p) CONCAT5_(f, r, n, v, p)
+#define VECTOR(p, n) CONCAT2(p, n)
+#define FUNC_R(f, r) CONCAT2_(f, r)
+
+#if V == VX
+#define VV x
+#elif V == VY
+#define VV y
+#elif V == VZ
+#define VV z
+#else
+#define VV normal
+#endif
+
+#if V == VNORMAL
+//kernel void FUNC(concat, R, N, normal, P)(array<texture2d_array<P, access::read>, N> in [[texture(0)]],
+//                                     texture2d_array<P, access::read> out_x [[texture(N)]],
+//                                     texture2d_array<P, access::write> out [[texture(N+1)]],
+//                                     constant ConcatParam & pm [[buffer(0)]],
+//                                     uint3 gid [[thread_position_in_grid]]) {
+//}
+kernel void FUNC(concat, R, N, VV, P)(texture2d_array<P, access::read> in0 [[texture(0)]],
+                                          texture2d_array<P, access::read> in1 [[texture(1)]],
+#if N >= 3
+                                          texture2d_array<P, access::read> in2 [[texture(2)]],
+#endif
+#if N >= 4
+                                          texture2d_array<P, access::read> in3 [[texture(3)]],
+#endif
+#if N >= 5
+                                          texture2d_array<P, access::read> in4 [[texture(4)]],
+#endif
+#if N >= 6
+                                          texture2d_array<P, access::read> in5 [[texture(5)]],
+#endif
+                                          texture2d_array<P, access::read> inx [[texture(N)]],
+                                          texture2d_array<P, access::write> out [[texture(N+1)]],
+                                          constant ConcatParam & pm [[buffer(0)]],
+                                          uint3 gid [[thread_position_in_grid]]) {
+
+   ConcatParam cp = pm;
+   int xyzn[4] = {int(gid.x), int(gid.y), int(gid.z), 0}, abcd[4], oxyzn[4];
+   VECTOR(P, 4) r = inx.read(gid.xy, gid.z);
+   for (int i = 0; i < 4; i++) {
+     xyzn[3] = i;
+#if R == 4
+     xyzn2abcd_4(cp.odim[3], xyzn, abcd);
+#else
+     FUNC_R(xyzn2abcd, R)(xyzn, abcd);
+#endif
+     int k = abcd[cp.axis] - cp.offset;
+     if (k < 0) continue;
+     int j = 0;
+     for (; j < N; j++) {
+       if (k < cp.vdim[j]) {
+         break;
+       }
+       k -= cp.vdim[j];
+     }
+     if (j == N) {
+       continue;
+     }
+     int ta = cp.odim[cp.axis];
+     abcd[cp.axis] = k;
+     cp.odim[cp.axis] = cp.vdim[j];
+#if R == 4
+     abcd2xyzn_4(cp.odim[3], abcd, oxyzn);
+#else
+     FUNC_R(abcd2xyzn, R)(abcd, oxyzn);
+#endif
+     cp.odim[cp.axis] = ta;
+     switch (j) {
+       case 0: r[i] = in0.read(uint2(oxyzn[0], oxyzn[1]), oxyzn[2])[oxyzn[3]]; break;
+       case 1: r[i] = in1.read(uint2(oxyzn[0], oxyzn[1]), oxyzn[2])[oxyzn[3]]; break;
+#if N >= 3
+       case 2: r[i] = in2.read(uint2(oxyzn[0], oxyzn[1]), oxyzn[2])[oxyzn[3]]; break;
+#endif
+#if N >= 4
+       case 3: r[i] = in3.read(uint2(oxyzn[0], oxyzn[1]), oxyzn[2])[oxyzn[3]]; break;
+#endif
+#if N >= 5
+       case 4: r[i] = in4.read(uint2(oxyzn[0], oxyzn[1]), oxyzn[2])[oxyzn[3]]; break;
+#endif
+#if N >= 6
+       case 5: r[i] = in5.read(uint2(oxyzn[0], oxyzn[1]), oxyzn[2])[oxyzn[3]]; break;
+#endif
+     }
+   }
+   out.write(r, gid.xy, gid.z);
+}
+
+#endif // V == NORMAL
+
+
+
+#if V == VX
+kernel void FUNC(concat, R, N, VV, P)(texture2d_array<P, access::read> in0 [[texture(0)]],
+                                          texture2d_array<P, access::read> in1 [[texture(1)]],
+#if N >= 3
+                                          texture2d_array<P, access::read> in2 [[texture(2)]],
+#endif // N >= 3
+#if N >= 4
+                                          texture2d_array<P, access::read> in3 [[texture(3)]],
+#endif // N >= 4
+#if N >= 5
+                                          texture2d_array<P, access::read> in4 [[texture(4)]],
+#endif // N >= 5
+#if N >= 6
+                                          texture2d_array<P, access::read> in5 [[texture(5)]],
+#endif // N >= 6
+                                          texture2d_array<P, access::write> out [[texture(N)]],
+                                          constant ConcatParam & pm [[buffer(0)]],
+                                          uint3 gid [[thread_position_in_grid]]) {
+  int x = gid.x - pm.offset;
+  if (x < 0) return;
+  if (x < pm.vdim[0]) {
+    VECTOR(P, 4) r = in0.read(gid.xy, gid.z);
+    out.write(r, gid.xy, gid.z);
+    return;
+  }
+  x -= pm.vdim[0];
+  if (x < pm.vdim[1]) {
+    VECTOR(P, 4) r = in1.read(uint2(x, gid.y), gid.z);
+    out.write(r, gid.xy, gid.z);
+    return;
+  }
+#if N >= 3
+  x -= pm.vdim[1];
+  if (x < pm.vdim[2]) {
+    VECTOR(P, 4) r = in2.read(uint2(x, gid.y), gid.z);
+    out.write(r, gid.xy, gid.z);
+    return;
+  }
+#endif // N >= 3
+#if N >= 4
+  x -= pm.vdim[2];
+  if (x < pm.vdim[3]) {
+    VECTOR(P, 4) r = in3.read(uint2(x, gid.y), gid.z);
+    out.write(r, gid.xy, gid.z);
+    return;
+  }
+#endif // N >= 4
+#if N >= 5
+  x -= pm.vdim[3];
+  if (x < pm.vdim[4]) {
+    VECTOR(P, 4) r = in4.read(uint2(x, gid.y), gid.z);
+    out.write(r, gid.xy, gid.z);
+    return;
+  }
+#endif // N >= 5
+#if N >= 6
+  x -= pm.vdim[4];
+  if (x < pm.vdim[5]) {
+    VECTOR(P, 4) r = in5.read(uint2(x, gid.y), gid.z);
+    out.write(r, gid.xy, gid.z);
+    return;
+  }
+#endif // N >= 6
+}
+#endif // V == VX
+
+#if V == VY
+kernel void FUNC(concat, R, N, VV, P)(texture2d_array<P, access::read> in0 [[texture(0)]],
+                                      texture2d_array<P, access::read> in1 [[texture(1)]],
+#if N >= 3
+                                      texture2d_array<P, access::read> in2 [[texture(2)]],
+#endif // N >= 3
+#if N >= 4
+                                      texture2d_array<P, access::read> in3 [[texture(3)]],
+#endif // N >= 4
+#if N >= 5
+                                      texture2d_array<P, access::read> in4 [[texture(4)]],
+#endif // N >= 5
+#if N >= 6
+                                      texture2d_array<P, access::read> in5 [[texture(5)]],
+#endif // N >= 6
+                                      texture2d_array<P, access::write> out [[texture(N)]],
+                                      constant ConcatParam & pm [[buffer(0)]],
+                                      uint3 gid [[thread_position_in_grid]]) {
+  int y = gid.y - pm.offset;
+  if (y < 0) return;
+  if (y < pm.vdim[0]) {
+    VECTOR(P, 4)  r = in0.read(gid.xy, gid.z);
+    out.write(r, gid.xy, gid.z);
+    return;
+  }
+  y -= pm.vdim[0];
+  if (y < pm.vdim[1]) {
+    VECTOR(P, 4)  r = in1.read(uint2(gid.x, y), gid.z);
+    out.write(r, gid.xy, gid.z);
+    return;
+  }
+#if N >= 3
+  y -= pm.vdim[1];
+  if (y < pm.vdim[2]) {
+    VECTOR(P, 4)  r = in2.read(uint2(gid.x, y), gid.z);
+    out.write(r, gid.xy, gid.z);
+    return;
+  }
+#endif // N >= 3
+#if N >= 4
+  y -= pm.vdim[2];
+  if (y < pm.vdim[3]) {
+    VECTOR(P, 4)  r = in3.read(uint2(gid.x, y), gid.z);
+    out.write(r, gid.xy, gid.z);
+    return;
+  }
+#endif // N >= 4
+#if N >= 5
+  y -= pm.vdim[3];
+  if (y < pm.vdim[4]) {
+    VECTOR(P, 4)  r = in4.read(uint2(gid.x, y), gid.z);
+    out.write(r, gid.xy, gid.z);
+    return;
+  }
+#endif // N >= 5
+#if N >= 6
+  y -= pm.vdim[4];
+  if (y < pm.vdim[5]) {
+    VECTOR(P, 4)  r = in5.read(uint2(gid.x, y), gid.z);
+    out.write(r, gid.xy, gid.z);
+    return;
+  }
+#endif // N >= 6
+}
+#endif // V == VY
+
+#if V == VZ
+kernel void FUNC(concat, R, N, VV, P)(texture2d_array<P, access::read> in0 [[texture(0)]],
+                                      texture2d_array<P, access::read> in1 [[texture(1)]],
+#if N >= 3
+                                      texture2d_array<P, access::read> in2 [[texture(2)]],
+#endif // N >= 3
+#if N >= 4
+                                      texture2d_array<P, access::read> in3 [[texture(3)]],
+#endif // N >= 4
+#if N >= 5
+                                      texture2d_array<P, access::read> in4 [[texture(4)]],
+#endif // N >= 5
+#if N >= 6
+                                      texture2d_array<P, access::read> in5 [[texture(5)]],
+#endif // N >= 6
+                                      texture2d_array<P, access::write> out [[texture(N)]],
+                                      constant ConcatParam & pm [[buffer(0)]],
+                                      uint3 gid [[thread_position_in_grid]]) {
+  int z = gid.z - pm.offset;
+  if (z < 0) return;
+  if (z < pm.vdim[0]) {
+    VECTOR(P, 4) r = in0.read(gid.xy, gid.z);
+    out.write(r, gid.xy, gid.z);
+    return;
+  }
+  z -= pm.vdim[0];
+  if (z < pm.vdim[1]) {
+    VECTOR(P, 4)  r = in1.read(gid.xy, z);
+    out.write(r, gid.xy, gid.z);
+    return;
+  }
+#if N >= 3
+  z -= pm.vdim[1];
+  if (z < pm.vdim[2]) {
+    VECTOR(P, 4)  r = in2.read(gid.xy, z);
+    out.write(r, gid.xy, gid.z);
+    return;
+  }
+#endif // N >= 3
+#if N >= 4
+  z -= pm.vdim[2];
+  if (z < pm.vdim[3]) {
+    VECTOR(P, 4)  r = in3.read(gid.xy, z);
+    out.write(r, gid.xy, gid.z);
+    return;
+  }
+#endif // N >= 4
+#if N >= 5
+  z -= pm.vdim[3];
+  if (z < pm.vdim[4]) {
+    VECTOR(P, 4)  r = in4.read(gid.xy, z);
+    out.write(r, gid.xy, gid.z);
+    return;
+  }
+#endif // N >= 5
+#if N >= 6
+  z -= pm.vdim[4];
+  if (z < pm.vdim[5]) {
+    VECTOR(P, 4)  r = in5.read(gid.xy, z);
+    out.write(r, gid.xy, gid.z);
+    return;
+  }
+#endif // N >= 6
+}
+#endif // V == VZ
+
+
+#undef VV
+#endif // #ifdef P

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/ConcatKernel.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+#include "Common.metal"
+
+using namespace metal;
+
+struct ConcatParam {
+  int32_t odim[4];
+  int32_t axis;
+  int32_t offset;
+  int32_t trans[4];
+  int32_t vdim[6];
+};
+
+#define VNORMAL 1
+#define VX 2
+#define VY 3
+#define VZ 4
+
+// >> fast mode
+// only support concat_{2,3,4}_{2,3,4,5,6}_y_{float,half}
+// only support concat_{3,4}_{2,3,4,5,6}_x_{float,half}
+// only support concat_{1,2,3,4}_{2,3,4,5,6}_z_{float,half}
+// >> normal mode (loop mode)
+// ssd-ar: (R=4, N=3, V=z), (R=3, N=2, V=y), (R=2, N=5, V=x), (R=3, N=5, V=x)
+// ssd: (R=2, N=6, V=y), (R=3, N=6, V=y)
+// genet: (R=4, N=2, V=normal)
+
+// ssd-ar: (R=3, N=5, V=x)
+#define V VX
+  #define R 3
+    #define N 5
+      #define P float
+        #include "ConcatKernel.inc.metal"
+      #undef P
+      #define P half
+        #include "ConcatKernel.inc.metal"
+      #undef P
+    #undef N
+  #undef R
+#undef V
+
+// ssd-ar: (R=2, N=5, V=x)
+#define V VX
+  #define R 2
+    #define N 5
+      #define P float
+        #include "ConcatKernel.inc.metal"
+      #undef P
+      #define P half
+        #include "ConcatKernel.inc.metal"
+      #undef P
+    #undef N
+  #undef R
+#undef V
+
+
+// ssd-ar: (R=3, N=2, V=y)
+#define V VY
+  #define R 3
+    #define N 2
+      #define P float
+        #include "ConcatKernel.inc.metal"
+      #undef P
+      #define P half
+        #include "ConcatKernel.inc.metal"
+      #undef P
+    #undef N
+  #undef R
+#undef V
+
+// ssd-ar: (R=4, N=3, V=z)
+#define V VZ
+  #define R 4
+    #define N 3
+      #define P float
+        #include "ConcatKernel.inc.metal"
+      #undef P
+      #define P half
+        #include "ConcatKernel.inc.metal"
+      #undef P
+    #undef N
+  #undef R
+#undef V
+
+
+// ssd: (R=2, N=6, V=y)
+#define V VY
+  #define R 2
+    #define N 6
+      #define P float
+        #include "ConcatKernel.inc.metal"
+      #undef P
+      #define P half
+        #include "ConcatKernel.inc.metal"
+      #undef P
+    #undef N
+  #undef R
+#undef V
+
+// ssd: (R=3, N=6, V=y)
+#define V VY
+  #define R 3
+    #define N 6
+      #define P float
+        #include "ConcatKernel.inc.metal"
+      #undef P
+      #define P half
+        #include "ConcatKernel.inc.metal"
+      #undef P
+    #undef N
+  #undef R
+#undef V
+
+#define V VNORMAL
+  #define R 4
+    #define N 2
+      #define P float
+        #include "ConcatKernel.inc.metal"
+      #undef P
+      #define P half
+        #include "ConcatKernel.inc.metal"
+      #undef P
+    #undef N
+  #undef R
+#undef V
+
+
+#define V VY
+  #define R 2
+    #define N 2
+      #define P float
+        #include "ConcatKernel.inc.metal"
+      #undef P
+      #define P half
+        #include "ConcatKernel.inc.metal"
+      #undef P
+    #undef N
+  #undef R
+#undef V
+
+
+#define V VY
+  #define R 2
+    #define N 5
+      #define P float
+        #include "ConcatKernel.inc.metal"
+      #undef P
+      #define P half
+        #include "ConcatKernel.inc.metal"
+      #undef P
+    #undef N
+  #undef R
+#undef V
+
+
+
+

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/ConvAddBNReluKernel.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+#include "Common.metal"
+using namespace metal;
+
+
+kernel void conv_add_batch_norm_relu_1x1_half(
+            texture2d_array<half, access::sample> inTexture [[texture(0)]],
+            texture2d_array<half, access::write> outTexture [[texture(1)]],
+            constant MetalConvParam &param [[buffer(0)]],
+            const device half4 *weights [[buffer(1)]],
+            const device half4 *biase [[buffer(2)]],
+            const device half4 *new_scale [[buffer(3)]],
+            const device half4 *new_biase [[buffer(4)]],
+            uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 1;
+  
+  uint input_arr_size = inTexture.get_array_size();
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  float4 output = float4(0.0);
+  
+  half4 input;
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input = inTexture.sample(sample, float2(posInInput.x, posInInput.y), i);
+    half4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size  + i];
+    output.x += dot(input, weight_x);
+    
+    half4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size  + i];
+    output.y += dot(input, weight_y);
+    
+    half4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size  + i];
+    output.z += dot(input, weight_z);
+    
+    half4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + i];
+    output.w += dot(input, weight_w);
+  }
+  output = fmax((output + float4(biase[gid.z])) * float4(new_scale[gid.z]) + float4(new_biase[gid.z]), 0.0);
+  outTexture.write(half4(output), gid.xy, gid.z);
+}
+
+kernel void conv_add_batch_norm_relu_3x3_half(
+            texture2d_array<half, access::sample> inTexture [[texture(0)]],
+            texture2d_array<half, access::write> outTexture [[texture(1)]],
+            constant MetalConvParam &param [[buffer(0)]],
+            const device half4 *weights [[buffer(1)]],
+            const device half4 *biase [[buffer(2)]],
+            const device half4 *new_scale [[buffer(3)]],
+            const device half4 *new_biase [[buffer(4)]],
+            uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  const ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 9;
+  uint input_arr_size = inTexture.get_array_size();
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  float4 output = float4(0.0);
+  
+  half4 input[9];
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input[0] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y - 1), i);
+    input[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - 1), i);
+    input[2] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y - 1), i);
+    input[3] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y), i);
+    input[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), i);
+    input[5] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y), i);
+    input[6] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y + 1), i);
+    input[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + 1), i);
+    input[8] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y + 1), i);
+    for (int j = 0; j < 9; ++j) {
+      half4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.x += dot(input[j], weight_x);
+      
+      half4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.y += dot(input[j], weight_y);
+      
+      half4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.z += dot(input[j], weight_z);
+      
+      half4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.w += dot(input[j], weight_w);
+    }
+  }
+  output = fmax((output + float4(biase[gid.z])) * float4(new_scale[gid.z]) + float4(new_biase[gid.z]), 0.0);
+  outTexture.write(half4(output), gid.xy, gid.z);
+}
+
+kernel void depthwise_conv_add_batch_norm_relu_3x3_half(
+            texture2d_array<half, access::sample> inTexture [[texture(0)]],
+            texture2d_array<half, access::write> outTexture [[texture(1)]],
+            constant MetalConvParam &param [[buffer(0)]],
+            const device half *weights [[buffer(1)]],
+            const device half4 *biase [[buffer(2)]],
+            const device half4 *new_scale [[buffer(3)]],
+            const device half4 *new_biase [[buffer(4)]],
+            uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  uint output_slice = gid.z;
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 9;
+  uint weithTo = gid.z * kernelHXW * 4;
+  float4 output = float4(0.0);
+  half4 inputs[9];
+  inputs[0] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y - 1), output_slice);
+  inputs[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - 1), output_slice);
+  inputs[2] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y - 1), output_slice);
+  inputs[3] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y), output_slice);
+  inputs[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), output_slice);
+  inputs[5] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y), output_slice);
+  inputs[6] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y + 1), output_slice);
+  inputs[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + 1), output_slice);
+  inputs[8] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y + 1), output_slice);
+  for (int j = 0; j < 9; ++j) {
+    half4 input = inputs[j];
+    output.x += input.x * weights[weithTo + 0 * kernelHXW + j];
+    output.y += input.y * weights[weithTo + 1 * kernelHXW + j];
+    output.z += input.z * weights[weithTo + 2 * kernelHXW + j];
+    output.w += input.w * weights[weithTo + 3 * kernelHXW + j];
+  }
+  output = fmax((output + float4(biase[gid.z])) * float4(new_scale[gid.z]) + float4(new_biase[gid.z]), 0.0);
+  outTexture.write(half4(output), gid.xy, gid.z);
+}
+
+
+
+/*---------------------------------------------*/
+
+
+
+kernel void conv_add_batch_norm_relu_1x1(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+                                         texture2d_array<float, access::write> outTexture [[texture(1)]],
+                                         constant MetalConvParam &param [[buffer(0)]],
+                                         const device float4 *weights [[buffer(1)]],
+                                         const device float4 *biase [[buffer(2)]],
+                                         const device float4 *new_scale [[buffer(3)]],
+                                         const device float4 *new_biase [[buffer(4)]],
+                                         uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 1;
+  
+  uint input_arr_size = inTexture.get_array_size();
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  float4 output = float4(0.0);
+  
+  float4 input;
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input = inTexture.sample(sample, float2(posInInput.x, posInInput.y), i);
+    float4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size  + i];
+    output.x += dot(input, weight_x);
+    
+    float4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size  + i];
+    output.y += dot(input, weight_y);
+    
+    float4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size  + i];
+    output.z += dot(input, weight_z);
+    
+    float4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + i];
+    output.w += dot(input, weight_w);
+  }
+  output = fmax((output + biase[gid.z]) * new_scale[gid.z] + new_biase[gid.z], 0.0);
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+kernel void conv_add_batch_norm_relu_3x3(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+                                         texture2d_array<float, access::write> outTexture [[texture(1)]],
+                                         constant MetalConvParam &param [[buffer(0)]],
+                                         const device float4 *weights [[buffer(1)]],
+                                         const device float4 *biase [[buffer(2)]],
+                                         const device float4 *new_scale [[buffer(3)]],
+                                         const device float4 *new_biase [[buffer(4)]],
+                                         uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  const ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 9;
+  uint input_arr_size = inTexture.get_array_size();
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  float4 output = float4(0.0);
+  
+  float4 input[9];
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input[0] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y - 1), i);
+    input[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - 1), i);
+    input[2] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y - 1), i);
+    input[3] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y), i);
+    input[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), i);
+    input[5] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y), i);
+    input[6] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y + 1), i);
+    input[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + 1), i);
+    input[8] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y + 1), i);
+    for (int j = 0; j < 9; ++j) {
+      float4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.x += dot(input[j], weight_x);
+      
+      float4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.y += dot(input[j], weight_y);
+      
+      float4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.z += dot(input[j], weight_z);
+      
+      float4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.w += dot(input[j], weight_w);
+    }
+  }
+  output = fmax((output + biase[gid.z]) * new_scale[gid.z] + new_biase[gid.z], 0.0);
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+kernel void depthwise_conv_add_batch_norm_relu_3x3(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+                                                   texture2d_array<float, access::write> outTexture [[texture(1)]],
+                                                   constant MetalConvParam &param [[buffer(0)]],
+                                                   const device float *weights [[buffer(1)]],
+                                                   const device float4 *biase [[buffer(2)]],
+                                                   const device float4 *new_scale [[buffer(3)]],
+                                                   const device float4 *new_biase [[buffer(4)]],
+                                                   uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  uint output_slice = gid.z;
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 9;
+  uint weithTo = gid.z * kernelHXW * 4;
+  float4 output = float4(0.0);
+  float4 inputs[9];
+  inputs[0] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y - 1), output_slice);
+  inputs[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - 1), output_slice);
+  inputs[2] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y - 1), output_slice);
+  inputs[3] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y), output_slice);
+  inputs[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), output_slice);
+  inputs[5] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y), output_slice);
+  inputs[6] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y + 1), output_slice);
+  inputs[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + 1), output_slice);
+  inputs[8] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y + 1), output_slice);
+  for (int j = 0; j < 9; ++j) {
+    float4 input = inputs[j];
+    output.x += input.x * weights[weithTo + 0 * kernelHXW + j];
+    output.y += input.y * weights[weithTo + 1 * kernelHXW + j];
+    output.z += input.z * weights[weithTo + 2 * kernelHXW + j];
+    output.w += input.w * weights[weithTo + 3 * kernelHXW + j];
+  }
+  output = fmax((output + biase[gid.z]) * new_scale[gid.z] + new_biase[gid.z], 0.0);
+  outTexture.write(output, gid.xy, gid.z);
+}
+

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/ConvAddMetal.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+#include "Common.metal"
+
+using namespace metal;
+
+#pragma mark - convAdd
+kernel void conv_add_1x1(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+                         texture2d_array<float, access::write> outTexture [[texture(1)]],
+                         constant MetalConvParam &param [[buffer(0)]],
+                         const device float4 *weights [[buffer(1)]],
+                         const device float4 *biase [[buffer(2)]],
+                         uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 1;
+  
+  uint input_arr_size = inTexture.get_array_size();
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  float4 output = biase[gid.z];
+  
+  float4 input;
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input = inTexture.sample(sample, float2(posInInput.x, posInInput.y), i);
+    float4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size  + i];
+    output.x += dot(input, weight_x);
+    
+    float4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size  + i];
+    output.y += dot(input, weight_y);
+    
+    float4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size  + i];
+    output.z += dot(input, weight_z);
+    
+    float4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + i];
+    output.w += dot(input, weight_w);
+  }
+//  output = output + biase[gid.z];
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+kernel void conv_add_3x3(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+                         texture2d_array<float, access::write> outTexture [[texture(1)]],
+                         constant MetalConvParam &param [[buffer(0)]],
+                         const device float4 *weights [[buffer(1)]],
+                         const device float4 *biase [[buffer(2)]],
+                         uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  const ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  
+  const uint kernelHXW = 9;
+  
+  uint input_arr_size = inTexture.get_array_size();
+  
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  float4 output = biase[gid.z];
+  
+  ushort dilation_x = param.dilationX;
+  ushort dilation_y = param.dilationY;
+  
+  float4 input[9];
+  
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input[0] = inTexture.sample(sample, float2(posInInput.x - dilation_x, posInInput.y - dilation_y), i);
+    
+    input[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - dilation_y), i);
+    
+    input[2] = inTexture.sample(sample, float2(posInInput.x + dilation_x, posInInput.y - dilation_y), i);
+    
+    input[3] = inTexture.sample(sample, float2(posInInput.x - dilation_x,    posInInput.y), i);
+    
+    input[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), i);
+    
+    input[5] = inTexture.sample(sample, float2(posInInput.x + dilation_x,    posInInput.y), i);
+    
+    input[6] = inTexture.sample(sample, float2(posInInput.x - dilation_x,    posInInput.y + dilation_y), i);
+    
+    input[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + dilation_y), i);
+    
+    input[8] = inTexture.sample(sample, float2(posInInput.x + dilation_x,    posInInput.y + dilation_y), i);
+    
+    for (int j = 0; j < 9; ++j) {
+      float4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.x += dot(input[j], weight_x);
+      
+      float4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.y += dot(input[j], weight_y);
+      
+      float4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.z += dot(input[j], weight_z);
+      
+      float4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.w += dot(input[j], weight_w);
+    }
+  }
+//  output = output + biase[gid.z];
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+kernel void conv_add_5x1(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+                         texture2d_array<float, access::write> outTexture [[texture(1)]],
+                         constant MetalConvParam &param [[buffer(0)]],
+                         const device float4 *weights [[buffer(1)]],
+                         const device float4 *biase [[buffer(2)]],
+                         uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  const ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  
+  const uint kernelHXW = 5;
+  
+  uint input_arr_size = inTexture.get_array_size();
+  
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  float4 output = biase[gid.z];
+  
+  ushort dilation_y = param.dilationY;
+  float4 input[5];
+  
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input[0] = inTexture.sample(sample, float2(posInInput.x, posInInput.y - 2 * dilation_y), i);
+    
+    input[1] = inTexture.sample(sample, float2(posInInput.x, posInInput.y - dilation_y), i);
+    
+    input[2] = inTexture.sample(sample, float2(posInInput.x, posInInput.y), i);
+    
+    input[3] = inTexture.sample(sample, float2(posInInput.x, posInInput.y + dilation_y), i);
+    
+    input[4] = inTexture.sample(sample, float2(posInInput.x, posInInput.y + 2 * dilation_y), i);
+    
+    for (int j = 0; j < 5; ++j) {
+      float4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.x += dot(input[j], weight_x);
+      
+      float4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.y += dot(input[j], weight_y);
+      
+      float4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.z += dot(input[j], weight_z);
+      
+      float4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.w += dot(input[j], weight_w);
+    }
+  }
+//  output = output + biase[gid.z];
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+
+kernel void conv_add_1x5(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+                         texture2d_array<float, access::write> outTexture [[texture(1)]],
+                         constant MetalConvParam &param [[buffer(0)]],
+                         const device float4 *weights [[buffer(1)]],
+                         const device float4 *biase [[buffer(2)]],
+                         uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  const ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  
+  const uint kernelHXW = 5;
+  
+  uint input_arr_size = inTexture.get_array_size();
+  
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  float4 output = biase[gid.z];
+  
+  ushort dilation_x = param.dilationX;
+  float4 input[5];
+  
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input[0] = inTexture.sample(sample, float2(posInInput.x - 2 * dilation_x, posInInput.y), i);
+    
+    input[1] = inTexture.sample(sample, float2(posInInput.x - dilation_x, posInInput.y), i);
+    
+    input[2] = inTexture.sample(sample, float2(posInInput.x, posInInput.y), i);
+    
+    input[3] = inTexture.sample(sample, float2(posInInput.x + dilation_x, posInInput.y), i);
+    
+    input[4] = inTexture.sample(sample, float2(posInInput.x + 2 * dilation_x, posInInput.y), i);
+    
+    for (int j = 0; j < 5; ++j) {
+      float4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.x += dot(input[j], weight_x);
+      
+      float4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.y += dot(input[j], weight_y);
+      
+      float4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.z += dot(input[j], weight_z);
+      
+      float4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.w += dot(input[j], weight_w);
+    }
+  }
+//  output = output + biase[gid.z];
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+
+kernel void depthwise_conv_add_3x3(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+                                   texture2d_array<float, access::write> outTexture [[texture(1)]],
+                                   constant MetalConvParam &param [[buffer(0)]],
+                                   const device float *weights [[buffer(1)]],
+                                   const device float4 *biase [[buffer(2)]],
+                                   uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  uint output_slice = gid.z;
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 9;
+  uint weithTo = gid.z * kernelHXW * 4;
+  float4 output = biase[gid.z];
+  float4 inputs[9];
+  inputs[0] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y - 1), output_slice);
+  inputs[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - 1), output_slice);
+  inputs[2] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y - 1), output_slice);
+  inputs[3] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y), output_slice);
+  inputs[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), output_slice);
+  inputs[5] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y), output_slice);
+  inputs[6] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y + 1), output_slice);
+  inputs[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + 1), output_slice);
+  inputs[8] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y + 1), output_slice);
+  for (int j = 0; j < 9; ++j) {
+    float4 input = inputs[j];
+    output.x += input.x * weights[weithTo + 0 * kernelHXW + j];
+    output.y += input.y * weights[weithTo + 1 * kernelHXW + j];
+    output.z += input.z * weights[weithTo + 2 * kernelHXW + j];
+    output.w += input.w * weights[weithTo + 3 * kernelHXW + j];
+  }
+//  output = output + biase[gid.z];
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+
+#pragma mark - half
+
+kernel void conv_add_1x1_half(texture2d_array<half, access::sample> inTexture [[texture(0)]],
+                         texture2d_array<half, access::write> outTexture [[texture(1)]],
+                         constant MetalConvParam &param [[buffer(0)]],
+                         const device half4 *weights [[buffer(1)]],
+                         const device half4 *biase [[buffer(2)]],
+                         uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 1;
+  
+  uint input_arr_size = inTexture.get_array_size();
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  half4 output = biase[gid.z];
+  
+  half4 input;
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input = inTexture.sample(sample, float2(posInInput.x, posInInput.y), i);
+    half4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size  + i];
+    output.x += dot(input, weight_x);
+    
+    half4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size  + i];
+    output.y += dot(input, weight_y);
+    
+    half4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size  + i];
+    output.z += dot(input, weight_z);
+    
+    half4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + i];
+    output.w += dot(input, weight_w);
+  }
+//  output = output + float4(biase[gid.z]);
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+kernel void conv_add_3x3_half(texture2d_array<half, access::sample> inTexture [[texture(0)]],
+                         texture2d_array<half, access::write> outTexture [[texture(1)]],
+                         constant MetalConvParam &param [[buffer(0)]],
+                         const device half4 *weights [[buffer(1)]],
+                         const device half4 *biase [[buffer(2)]],
+                         uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  const ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 9;
+  uint input_arr_size = inTexture.get_array_size();
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  half4 output = biase[gid.z];
+  
+  ushort dilation_x = param.dilationX;
+  ushort dilation_y = param.dilationY;
+  
+  half4 input[9];
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input[0] = inTexture.sample(sample, float2(posInInput.x - dilation_x,    posInInput.y - dilation_y), i);
+    input[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - dilation_y), i);
+    input[2] = inTexture.sample(sample, float2(posInInput.x + dilation_x,    posInInput.y - dilation_y), i);
+    input[3] = inTexture.sample(sample, float2(posInInput.x - dilation_x,    posInInput.y), i);
+    input[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), i);
+    input[5] = inTexture.sample(sample, float2(posInInput.x + dilation_x,    posInInput.y), i);
+    input[6] = inTexture.sample(sample, float2(posInInput.x - dilation_x,    posInInput.y + dilation_y), i);
+    input[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + dilation_y), i);
+    input[8] = inTexture.sample(sample, float2(posInInput.x + dilation_x,    posInInput.y + dilation_y), i);
+    for (int j = 0; j < 9; ++j) {
+      half4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.x += dot(float4(input[j]), float4(weight_x));
+      
+      half4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.y += dot(float4(input[j]), float4(weight_y));
+      
+      half4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.z += dot(float4(input[j]), float4(weight_z));
+      
+      half4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.w += dot(float4(input[j]), float4(weight_w));
+    }
+  }
+//  output = output + float4(biase[gid.z]);
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+kernel void depthwise_conv_add_3x3_half(texture2d_array<half, access::sample> inTexture [[texture(0)]],
+                                   texture2d_array<half, access::write> outTexture [[texture(1)]],
+                                   constant MetalConvParam &param [[buffer(0)]],
+                                   const device half *weights [[buffer(1)]],
+                                   const device half4 *biase [[buffer(2)]],
+                                   uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  uint output_slice = gid.z;
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 9;
+  uint weithTo = gid.z * kernelHXW * 4;
+  half4 output = biase[gid.z];
+  half4 inputs[9];
+  inputs[0] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y - 1), output_slice);
+  inputs[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - 1), output_slice);
+  inputs[2] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y - 1), output_slice);
+  inputs[3] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y), output_slice);
+  inputs[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), output_slice);
+  inputs[5] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y), output_slice);
+  inputs[6] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y + 1), output_slice);
+  inputs[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + 1), output_slice);
+  inputs[8] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y + 1), output_slice);
+  for (int j = 0; j < 9; ++j) {
+    half4 input = inputs[j];
+    output.x += input.x * weights[weithTo + 0 * kernelHXW + j];
+    output.y += input.y * weights[weithTo + 1 * kernelHXW + j];
+    output.z += input.z * weights[weithTo + 2 * kernelHXW + j];
+    output.w += input.w * weights[weithTo + 3 * kernelHXW + j];
+  }
+//  output = output + float4(biase[gid.z]);
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+
+kernel void conv_add_5x1_half(texture2d_array<half, access::sample> inTexture [[texture(0)]],
+                         texture2d_array<half, access::write> outTexture [[texture(1)]],
+                         constant MetalConvParam &param [[buffer(0)]],
+                         const device half4 *weights [[buffer(1)]],
+                         const device half4 *biase [[buffer(2)]],
+                         uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  const ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  
+  const uint kernelHXW = 5;
+  
+  uint input_arr_size = inTexture.get_array_size();
+  
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  half4 output = biase[gid.z];
+  
+  ushort dilation_y = param.dilationY;
+  half4 input[5];
+  
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input[0] = inTexture.sample(sample, float2(posInInput.x, posInInput.y - 2 * dilation_y), i);
+    
+    input[1] = inTexture.sample(sample, float2(posInInput.x, posInInput.y - dilation_y), i);
+    
+    input[2] = inTexture.sample(sample, float2(posInInput.x, posInInput.y), i);
+    
+    input[3] = inTexture.sample(sample, float2(posInInput.x, posInInput.y + dilation_y), i);
+    
+    input[4] = inTexture.sample(sample, float2(posInInput.x, posInInput.y + 2 * dilation_y), i);
+    
+    for (int j = 0; j < 5; ++j) {
+      half4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.x += dot(input[j], weight_x);
+      
+      half4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.y += dot(input[j], weight_y);
+      
+      half4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.z += dot(input[j], weight_z);
+      
+      half4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.w += dot(input[j], weight_w);
+    }
+  }
+//  output = output + float4(biase[gid.z]);
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+
+kernel void conv_add_1x5_half(texture2d_array<half, access::sample> inTexture [[texture(0)]],
+                         texture2d_array<half, access::write> outTexture [[texture(1)]],
+                         constant MetalConvParam &param [[buffer(0)]],
+                         const device half4 *weights [[buffer(1)]],
+                         const device half4 *biase [[buffer(2)]],
+                         uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  const ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  
+  const uint kernelHXW = 5;
+  
+  uint input_arr_size = inTexture.get_array_size();
+  
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  half4 output = biase[gid.z];
+  
+  ushort dilation_x = param.dilationX;
+  half4 input[5];
+  
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input[0] = inTexture.sample(sample, float2(posInInput.x - 2 * dilation_x, posInInput.y), i);
+    
+    input[1] = inTexture.sample(sample, float2(posInInput.x - dilation_x, posInInput.y), i);
+    
+    input[2] = inTexture.sample(sample, float2(posInInput.x, posInInput.y), i);
+    
+    input[3] = inTexture.sample(sample, float2(posInInput.x + dilation_x, posInInput.y), i);
+    
+    input[4] = inTexture.sample(sample, float2(posInInput.x + 2 * dilation_x, posInInput.y), i);
+    
+    for (int j = 0; j < 5; ++j) {
+      half4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.x += dot(input[j], weight_x);
+      
+      half4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.y += dot(input[j], weight_y);
+      
+      half4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.z += dot(input[j], weight_z);
+      
+      half4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.w += dot(input[j], weight_w);
+    }
+  }
+//  output = output + float4(biase[gid.z]);
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+
+kernel void test_conv_add_3x3(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+                              texture2d_array<float, access::write> outTexture [[texture(1)]],
+                              constant MetalConvParam &param [[buffer(0)]],
+                              const device float4 *weights [[buffer(1)]],
+                              const device float4 *biase [[buffer(2)]],
+                              uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  if (gid.x > 0 || gid.y > 0 || gid.z > 0) { return; }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  const ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  
+  const uint kernelHXW = 9;
+  
+  uint input_arr_size = inTexture.get_array_size();
+  
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  float4 output = float4(0.0);
+  
+  ushort dilation_x = param.dilationX;
+  ushort dilation_y = param.dilationY;
+  
+  float4 input[9];
+  
+  for (uint i = 0; i < input_arr_size; ++i) {
+    
+    input[0] = inTexture.sample(sample, float2(posInInput.x - dilation_x, posInInput.y - dilation_y), i);
+    
+    input[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - dilation_y), i);
+    
+    input[2] = inTexture.sample(sample, float2(posInInput.x + dilation_x, posInInput.y - dilation_y), i);
+    
+    input[3] = inTexture.sample(sample, float2(posInInput.x - dilation_x,    posInInput.y), i);
+    
+    input[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), i);
+    
+    input[5] = inTexture.sample(sample, float2(posInInput.x + dilation_x,    posInInput.y), i);
+    
+    input[6] = inTexture.sample(sample, float2(posInInput.x - dilation_x,    posInInput.y + dilation_y), i);
+    
+    input[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + dilation_y), i);
+    
+    input[8] = inTexture.sample(sample, float2(posInInput.x + dilation_x,    posInInput.y + dilation_y), i);
+    
+    for (int j = 0; j < 9; ++j) {
+      float4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.x += dot(input[j], weight_x);
+      
+      float4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.y += dot(input[j], weight_y);
+      
+      float4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.z += dot(input[j], weight_z);
+      
+      float4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.w += dot(input[j], weight_w);
+    }
+  }
+  //  output = output + biase[gid.z];
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+
+

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/ConvAddPrelu.inc.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#ifdef P
+
+#include "Macro.metal"
+
+
+#pragma mark - convAdd
+kernel void FUNC3_(conv_add_1x1, PRELU_TYPE, P)(texture2d_array<P, access::sample> inTexture [[texture(0)]],
+                         texture2d_array<P, access::write> outTexture [[texture(1)]],
+                         constant MetalConvParam &param [[buffer(0)]],
+                         const device VECTOR(P, 4) *weights [[buffer(1)]],
+                         const device VECTOR(P, 4) *biase [[buffer(2)]],
+#ifdef PRELU_CHANNEL
+                         const device VECTOR(P, 4) *alpha [[buffer(3)]],
+#endif
+#ifdef PRELU_ELEMENT
+                         const device VECTOR(P, 4) *alpha [[buffer(3)]],
+#endif
+#ifdef PRELU_OTHER
+                         const device P *alpha [[buffer(3)]],
+#endif
+                         uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 1;
+  
+  uint input_arr_size = inTexture.get_array_size();
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  VECTOR(P, 4) output = biase[gid.z];
+  
+  VECTOR(P, 4) input;
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input = inTexture.sample(sample,float2(posInInput.x, posInInput.y), i);
+    VECTOR(P, 4) weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size  + i];
+    output.x += dot(input, weight_x);
+    
+    VECTOR(P, 4) weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size  + i];
+    output.y += dot(input, weight_y);
+    
+    VECTOR(P, 4) weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size  + i];
+    output.z += dot(input, weight_z);
+    
+    VECTOR(P, 4) weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + i];
+    output.w += dot(input, weight_w);
+  }
+  
+//  output = output + float4(biase[gid.z]);
+  
+#ifdef PRELU_CHANNEL
+  VECTOR(P, 4) alpha_value = alpha[gid.z];
+  output.x = output.x > 0 ? output.x : (alpha_value.x * output.x);
+  output.y = output.y > 0 ? output.y : (alpha_value.y * output.y);
+  output.z = output.z > 0 ? output.z : (alpha_value.z * output.z);
+  output.w = output.w > 0 ? output.w : (alpha_value.w * output.w);
+#endif
+#ifdef PRELU_ELEMENT
+  int alpha_to = (gid.y * outTexture.get_width() + gid.x) * outTexture.get_array_size();
+  VECTOR(P, 4) alpha_value = alpha[alpha_to + gid.z];
+  output.x = output.x > 0 ? output.x : (alpha_value.x * output.x);
+  output.y = output.y > 0 ? output.y : (alpha_value.y * output.y);
+  output.z = output.z > 0 ? output.z : (alpha_value.z * output.z);
+  output.w = output.w > 0 ? output.w : (alpha_value.w * output.w);
+#endif
+#ifdef PRELU_OTHER
+  P alpha_value = alpha[0];
+  output.x = output.x > 0 ? output.x : (alpha_value * output.x);
+  output.y = output.y > 0 ? output.y : (alpha_value * output.y);
+  output.z = output.z > 0 ? output.z : (alpha_value * output.z);
+  output.w = output.w > 0 ? output.w : (alpha_value * output.w);
+#endif
+  outTexture.write(VECTOR(P, 4)(output), gid.xy, gid.z);
+}
+
+kernel void FUNC3_(conv_add_3x3, PRELU_TYPE, P)(texture2d_array<P, access::sample> inTexture [[texture(0)]],
+    texture2d_array<P, access::write> outTexture [[texture(1)]],
+    constant MetalConvParam &param [[buffer(0)]],
+    const device VECTOR(P, 4) *weights [[buffer(1)]],
+    const device VECTOR(P, 4) *biase [[buffer(2)]],
+#ifdef PRELU_CHANNEL
+     const device VECTOR(P, 4) *alpha [[buffer(3)]],
+#endif
+#ifdef PRELU_ELEMENT
+     const device VECTOR(P, 4) *alpha [[buffer(3)]],
+#endif
+#ifdef PRELU_OTHER
+     const device P *alpha [[buffer(3)]],
+#endif
+     uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  const ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+
+  const uint kernelHXW = 9;
+
+  uint input_arr_size = inTexture.get_array_size();
+
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+
+  VECTOR(P, 4) output = biase[gid.z];
+
+  ushort dilation_x = param.dilationX;
+  ushort dilation_y = param.dilationY;
+
+  VECTOR(P, 4) input[9];
+
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input[0] = inTexture.sample(sample, float2(posInInput.x - dilation_x, posInInput.y - dilation_y), i);
+
+    input[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - dilation_y), i);
+
+    input[2] = inTexture.sample(sample, float2(posInInput.x + dilation_x, posInInput.y - dilation_y), i);
+
+    input[3] = inTexture.sample(sample, float2(posInInput.x - dilation_x,    posInInput.y), i);
+
+    input[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), i);
+
+    input[5] = inTexture.sample(sample, float2(posInInput.x + dilation_x,    posInInput.y), i);
+
+    input[6] = inTexture.sample(sample, float2(posInInput.x - dilation_x,    posInInput.y + dilation_y), i);
+
+    input[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + dilation_y), i);
+
+    input[8] = inTexture.sample(sample, float2(posInInput.x + dilation_x,    posInInput.y + dilation_y), i);
+
+    for (int j = 0; j < 9; ++j) {
+      VECTOR(P, 4) weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.x += dot(input[j], weight_x);
+
+      VECTOR(P, 4) weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.y += dot(input[j], weight_y);
+
+      VECTOR(P, 4) weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.z += dot(input[j], weight_z);
+
+      VECTOR(P, 4) weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.w += dot(input[j], weight_w);
+    }
+  }
+//  output = output + float4(biase[gid.z]);
+  
+#ifdef PRELU_CHANNEL
+  VECTOR(P, 4) alpha_value = alpha[gid.z];
+  output.x = output.x > 0 ? output.x : (alpha_value.x * output.x);
+  output.y = output.y > 0 ? output.y : (alpha_value.y * output.y);
+  output.z = output.z > 0 ? output.z : (alpha_value.z * output.z);
+  output.w = output.w > 0 ? output.w : (alpha_value.w * output.w);
+#endif
+#ifdef PRELU_ELEMENT
+  int alpha_to = (gid.y * outTexture.get_width() + gid.x) * outTexture.get_array_size();
+  VECTOR(P, 4) alpha_value = alpha[alpha_to + gid.z];
+  output.x = output.x > 0 ? output.x : (alpha_value.x * output.x);
+  output.y = output.y > 0 ? output.y : (alpha_value.y * output.y);
+  output.z = output.z > 0 ? output.z : (alpha_value.z * output.z);
+  output.w = output.w > 0 ? output.w : (alpha_value.w * output.w);
+#endif
+#ifdef PRELU_OTHER
+  P alpha_value = alpha[0];
+  output.x = output.x > 0 ? output.x : (alpha_value * output.x);
+  output.y = output.y > 0 ? output.y : (alpha_value * output.y);
+  output.z = output.z > 0 ? output.z : (alpha_value * output.z);
+  output.w = output.w > 0 ? output.w : (alpha_value * output.w);
+#endif
+  outTexture.write(VECTOR(P, 4)(output), gid.xy, gid.z);
+}
+
+kernel void FUNC3_(conv_add_5x1, PRELU_TYPE, P)(texture2d_array<P, access::sample> inTexture [[texture(0)]],
+                         texture2d_array<P, access::write> outTexture [[texture(1)]],
+                         constant MetalConvParam &param [[buffer(0)]],
+                         const device VECTOR(P, 4) *weights [[buffer(1)]],
+                         const device VECTOR(P, 4) *biase [[buffer(2)]],
+#ifdef PRELU_CHANNEL
+                        const device VECTOR(P, 4) *alpha [[buffer(3)]],
+#endif
+#ifdef PRELU_ELEMENT
+                        const device VECTOR(P, 4) *alpha [[buffer(3)]],
+#endif
+#ifdef PRELU_OTHER
+                        const device P *alpha [[buffer(3)]],
+#endif
+                         uint3 gid [[thread_position_in_grid]]) {
+
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  const ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+
+  const uint kernelHXW = 5;
+
+  uint input_arr_size = inTexture.get_array_size();
+
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+
+  VECTOR(P, 4) output = biase[gid.z];;
+
+  ushort dilation_y = param.dilationY;
+  VECTOR(P, 4) input[5];
+
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input[0] = inTexture.sample(sample, float2(posInInput.x, posInInput.y - 2 * dilation_y), i);
+
+    input[1] = inTexture.sample(sample, float2(posInInput.x, posInInput.y - dilation_y), i);
+
+    input[2] = inTexture.sample(sample, float2(posInInput.x, posInInput.y), i);
+
+    input[3] = inTexture.sample(sample, float2(posInInput.x, posInInput.y + dilation_y), i);
+
+    input[4] = inTexture.sample(sample, float2(posInInput.x, posInInput.y + 2 * dilation_y), i);
+
+    for (int j = 0; j < 5; ++j) {
+      VECTOR(P, 4) weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.x += dot(input[j], weight_x);
+
+      VECTOR(P, 4) weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.y += dot(input[j], weight_y);
+
+      VECTOR(P, 4) weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.z += dot(input[j], weight_z);
+
+      VECTOR(P, 4) weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.w += dot(input[j], weight_w);
+    }
+  }
+  
+#ifdef PRELU_CHANNEL
+  VECTOR(P, 4) alpha_value = alpha[gid.z];
+  output.x = output.x > 0 ? output.x : (alpha_value.x * output.x);
+  output.y = output.y > 0 ? output.y : (alpha_value.y * output.y);
+  output.z = output.z > 0 ? output.z : (alpha_value.z * output.z);
+  output.w = output.w > 0 ? output.w : (alpha_value.w * output.w);
+#endif
+#ifdef PRELU_ELEMENT
+  int alpha_to = (gid.y * outTexture.get_width() + gid.x) * outTexture.get_array_size();
+  VECTOR(P, 4) alpha_value = alpha[alpha_to + gid.z];
+  output.x = output.x > 0 ? output.x : (alpha_value.x * output.x);
+  output.y = output.y > 0 ? output.y : (alpha_value.y * output.y);
+  output.z = output.z > 0 ? output.z : (alpha_value.z * output.z);
+  output.w = output.w > 0 ? output.w : (alpha_value.w * output.w);
+#endif
+#ifdef PRELU_OTHER
+  P alpha_value = alpha[0];
+  output.x = output.x > 0 ? output.x : (alpha_value * output.x);
+  output.y = output.y > 0 ? output.y : (alpha_value * output.y);
+  output.z = output.z > 0 ? output.z : (alpha_value * output.z);
+  output.w = output.w > 0 ? output.w : (alpha_value * output.w);
+#endif
+  outTexture.write(VECTOR(P, 4)(output), gid.xy, gid.z);
+}
+
+
+kernel void FUNC3_(conv_add_1x5, PRELU_TYPE, P)(texture2d_array<P, access::sample> inTexture [[texture(0)]],
+                         texture2d_array<P, access::write> outTexture [[texture(1)]],
+                         constant MetalConvParam &param [[buffer(0)]],
+                         const device VECTOR(P, 4) *weights [[buffer(1)]],
+                         const device VECTOR(P, 4) *biase [[buffer(2)]],
+#ifdef PRELU_CHANNEL
+                         const device VECTOR(P, 4) *alpha [[buffer(3)]],
+#endif
+#ifdef PRELU_ELEMENT
+                         const device VECTOR(P, 4) *alpha [[buffer(3)]],
+#endif
+#ifdef PRELU_OTHER
+                         const device P *alpha [[buffer(3)]],
+#endif
+                         uint3 gid [[thread_position_in_grid]]) {
+
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  const ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+
+  const uint kernelHXW = 5;
+
+  uint input_arr_size = inTexture.get_array_size();
+
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+
+  VECTOR(P, 4) output = biase[gid.z];
+
+  ushort dilation_x = param.dilationX;
+  VECTOR(P, 4) input[5];
+
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input[0] = inTexture.sample(sample, float2(posInInput.x - 2 * dilation_x, posInInput.y), i);
+
+    input[1] = inTexture.sample(sample, float2(posInInput.x - dilation_x, posInInput.y), i);
+
+    input[2] = inTexture.sample(sample, float2(posInInput.x, posInInput.y), i);
+
+    input[3] = inTexture.sample(sample, float2(posInInput.x + dilation_x, posInInput.y), i);
+
+    input[4] = inTexture.sample(sample, float2(posInInput.x + 2 * dilation_x, posInInput.y), i);
+
+    for (int j = 0; j < 5; ++j) {
+      VECTOR(P, 4) weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.x += dot(input[j], weight_x);
+
+      VECTOR(P, 4) weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.y += dot(input[j], weight_y);
+
+      VECTOR(P, 4) weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.z += dot(input[j], weight_z);
+
+      VECTOR(P, 4) weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.w += dot(input[j], weight_w);
+    }
+  }
+  
+#ifdef PRELU_CHANNEL
+  VECTOR(P, 4) alpha_value = alpha[gid.z];
+  output.x = output.x > 0 ? output.x : (alpha_value.x * output.x);
+  output.y = output.y > 0 ? output.y : (alpha_value.y * output.y);
+  output.z = output.z > 0 ? output.z : (alpha_value.z * output.z);
+  output.w = output.w > 0 ? output.w : (alpha_value.w * output.w);
+#endif
+#ifdef PRELU_ELEMENT
+  int alpha_to = (gid.y * outTexture.get_width() + gid.x) * outTexture.get_array_size();
+  VECTOR(P, 4) alpha_value = alpha[alpha_to + gid.z];
+  output.x = output.x > 0 ? output.x : (alpha_value.x * output.x);
+  output.y = output.y > 0 ? output.y : (alpha_value.y * output.y);
+  output.z = output.z > 0 ? output.z : (alpha_value.z * output.z);
+  output.w = output.w > 0 ? output.w : (alpha_value.w * output.w);
+#endif
+#ifdef PRELU_OTHER
+  P alpha_value = alpha[0];
+  output.x = output.x > 0 ? output.x : (alpha_value * output.x);
+  output.y = output.y > 0 ? output.y : (alpha_value * output.y);
+  output.z = output.z > 0 ? output.z : (alpha_value * output.z);
+  output.w = output.w > 0 ? output.w : (alpha_value * output.w);
+#endif
+  outTexture.write(VECTOR(P, 4)(output), gid.xy, gid.z);
+}
+
+kernel void FUNC3_(depthwise_conv_add_3x3, PRELU_TYPE, P)(texture2d_array<P, access::sample> inTexture [[texture(0)]],
+    texture2d_array<P, access::write> outTexture [[texture(1)]],
+    constant MetalConvParam &param [[buffer(0)]],
+    const device P *weights [[buffer(1)]],
+    const device VECTOR(P, 4) *biase [[buffer(2)]],
+#ifdef PRELU_CHANNEL
+    const device VECTOR(P, 4) *alpha [[buffer(3)]],
+#endif
+#ifdef PRELU_ELEMENT
+    const device VECTOR(P, 4) *alpha [[buffer(3)]],
+#endif
+#ifdef PRELU_OTHER
+    const device P *alpha [[buffer(3)]],
+#endif
+    uint3 gid [[thread_position_in_grid]]) {
+
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  uint output_slice = gid.z;
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 9;
+  uint weithTo = gid.z * kernelHXW * 4;
+  VECTOR(P, 4) output = biase[gid.z];
+  VECTOR(P, 4) inputs[9];
+  inputs[0] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y - 1), output_slice);
+  inputs[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - 1), output_slice);
+  inputs[2] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y - 1), output_slice);
+  inputs[3] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y), output_slice);
+  inputs[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), output_slice);
+  inputs[5] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y), output_slice);
+  inputs[6] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y + 1), output_slice);
+  inputs[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + 1), output_slice);
+  inputs[8] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y + 1), output_slice);
+  for (int j = 0; j < 9; ++j) {
+    VECTOR(P, 4) input = inputs[j];
+    output.x += input.x * weights[weithTo + 0 * kernelHXW + j];
+    output.y += input.y * weights[weithTo + 1 * kernelHXW + j];
+    output.z += input.z * weights[weithTo + 2 * kernelHXW + j];
+    output.w += input.w * weights[weithTo + 3 * kernelHXW + j];
+  }
+  
+#ifdef PRELU_CHANNEL
+  VECTOR(P, 4) alpha_value = alpha[gid.z];
+  output.x = output.x > 0 ? output.x : (alpha_value.x * output.x);
+  output.y = output.y > 0 ? output.y : (alpha_value.y * output.y);
+  output.z = output.z > 0 ? output.z : (alpha_value.z * output.z);
+  output.w = output.w > 0 ? output.w : (alpha_value.w * output.w);
+#endif
+#ifdef PRELU_ELEMENT
+  int alpha_to = (gid.y * outTexture.get_width() + gid.x) * outTexture.get_array_size();
+  VECTOR(P, 4) alpha_value = alpha[alpha_to + gid.z];
+  output.x = output.x > 0 ? output.x : (alpha_value.x * output.x);
+  output.y = output.y > 0 ? output.y : (alpha_value.y * output.y);
+  output.z = output.z > 0 ? output.z : (alpha_value.z * output.z);
+  output.w = output.w > 0 ? output.w : (alpha_value.w * output.w);
+#endif
+#ifdef PRELU_OTHER
+  P alpha_value = alpha[0];
+  output.x = output.x > 0 ? output.x : (alpha_value * output.x);
+  output.y = output.y > 0 ? output.y : (alpha_value * output.y);
+  output.z = output.z > 0 ? output.z : (alpha_value * output.z);
+  output.w = output.w > 0 ? output.w : (alpha_value * output.w);
+#endif
+  outTexture.write(VECTOR(P, 4)(output), gid.xy, gid.z);
+}
+
+#endif
+

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/ConvAddPreluKernel.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+#include "Common.metal"
+using namespace metal;
+
+#define P float
+
+  #define PRELU_CHANNEL prelu_channel
+  #define PRELU_TYPE prelu_channel
+    #include "ConvAddPrelu.inc.metal"
+  #undef  PRELU_TYPE
+  #undef  PRELU_CHANNEL
+
+  #define PRELU_ELEMENT prelu_element
+  #define PRELU_TYPE prelu_element
+    #include "ConvAddPrelu.inc.metal"
+  #undef  PRELU_TYPE
+  #undef  PRELU_ELEMENT
+
+  #define PRELU_OTHER   prelu_other
+  #define PRELU_TYPE prelu_other
+    #include "ConvAddPrelu.inc.metal"
+  #undef  PRELU_TYPE
+  #undef  PRELU_OTHER
+
+#undef P
+
+#define P half
+
+  #define PRELU_CHANNEL prelu_channel
+  #define PRELU_TYPE prelu_channel
+    #include "ConvAddPrelu.inc.metal"
+  #undef  PRELU_TYPE
+  #undef  PRELU_CHANNEL
+
+  #define PRELU_ELEMENT prelu_element
+  #define PRELU_TYPE prelu_element
+    #include "ConvAddPrelu.inc.metal"
+  #undef  PRELU_TYPE
+  #undef  PRELU_ELEMENT
+
+  #define PRELU_OTHER   prelu_other
+  #define PRELU_TYPE prelu_other
+    #include "ConvAddPrelu.inc.metal"
+  #undef  PRELU_TYPE
+  #undef  PRELU_OTHER
+
+#undef P
+
+
+
+

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/ConvBNReluKernel.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+#include "Common.metal"
+
+using namespace metal;
+
+#pragma mark - conv bn relu
+kernel void conv_batch_norm_relu_1x1(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+                                     texture2d_array<float, access::write> outTexture [[texture(1)]],
+                                     constant MetalConvParam &param [[buffer(0)]],
+                                     const device float4 *weights [[buffer(1)]],
+                                     const device float4 *new_scale [[buffer(2)]],
+                                     const device float4 *new_biase [[buffer(3)]],
+                                     uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 1;
+  
+  uint input_arr_size = inTexture.get_array_size();
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  float4 output = float4(0.0);
+  
+  float4 input;
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input = inTexture.sample(sample, float2(posInInput.x, posInInput.y), i);
+    float4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size  + i];
+    output.x += dot(input, weight_x);
+    
+    float4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size  + i];
+    output.y += dot(input, weight_y);
+    
+    float4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size  + i];
+    output.z += dot(input, weight_z);
+    
+    float4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + i];
+    output.w += dot(input, weight_w);
+  }
+  output = fmax(output * new_scale[gid.z] + new_biase[gid.z], 0.0);
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+kernel void conv_batch_norm_relu_3x3(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+                                     texture2d_array<float, access::write> outTexture [[texture(1)]],
+                                     constant MetalConvParam &param [[buffer(0)]],
+                                     const device float4 *weights [[buffer(1)]],
+                                     const device float4 *new_scale [[buffer(2)]],
+                                     const device float4 *new_biase [[buffer(3)]],
+                                     uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  const ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 9;
+  uint input_arr_size = inTexture.get_array_size();
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  float4 output = float4(0.0);
+  
+  float4 input[9];
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input[0] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y - 1), i);
+    input[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - 1), i);
+    input[2] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y - 1), i);
+    input[3] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y), i);
+    input[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), i);
+    input[5] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y), i);
+    input[6] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y + 1), i);
+    input[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + 1), i);
+    input[8] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y + 1), i);
+    for (int j = 0; j < 9; ++j) {
+      float4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.x += dot(input[j], weight_x);
+      
+      float4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.y += dot(input[j], weight_y);
+      
+      float4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.z += dot(input[j], weight_z);
+      
+      float4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.w += dot(input[j], weight_w);
+    }
+  }
+  output = fmax(output * new_scale[gid.z] + new_biase[gid.z], 0.0);
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+kernel void depthwise_conv_batch_norm_relu_3x3(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+                                               texture2d_array<float, access::write> outTexture [[texture(1)]],
+                                               constant MetalConvParam &param [[buffer(0)]],
+                                               const device float *weights [[buffer(1)]],
+                                               const device float4 *new_scale [[buffer(2)]],
+                                               const device float4 *new_biase [[buffer(3)]],
+                                               uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  uint output_slice = gid.z;
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 9;
+  uint weithTo = gid.z * kernelHXW * 4;
+  float4 output = float4(0.0);
+  float4 inputs[9];
+  inputs[0] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y - 1), output_slice);
+  inputs[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - 1), output_slice);
+  inputs[2] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y - 1), output_slice);
+  inputs[3] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y), output_slice);
+  inputs[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), output_slice);
+  inputs[5] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y), output_slice);
+  inputs[6] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y + 1), output_slice);
+  inputs[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + 1), output_slice);
+  inputs[8] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y + 1), output_slice);
+  for (int j = 0; j < 9; ++j) {
+    float4 input = inputs[j];
+    output.x += input.x * weights[weithTo + 0 * kernelHXW + j];
+    output.y += input.y * weights[weithTo + 1 * kernelHXW + j];
+    output.z += input.z * weights[weithTo + 2 * kernelHXW + j];
+    output.w += input.w * weights[weithTo + 3 * kernelHXW + j];
+  }
+  output = fmax(output * new_scale[gid.z] + new_biase[gid.z], 0.0);
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+#pragma mark - half
+kernel void conv_batch_norm_relu_1x1_half(texture2d_array<half, access::sample> inTexture [[texture(0)]],
+                                     texture2d_array<half, access::write> outTexture [[texture(1)]],
+                                     constant MetalConvParam &param [[buffer(0)]],
+                                     const device half4 *weights [[buffer(1)]],
+                                     const device half4 *new_scale [[buffer(2)]],
+                                     const device half4 *new_biase [[buffer(3)]],
+                                     uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 1;
+  
+  uint input_arr_size = inTexture.get_array_size();
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  float4 output = float4(0.0);
+  
+  half4 input;
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input = inTexture.sample(sample, float2(posInInput.x, posInInput.y), i);
+    half4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size  + i];
+    output.x += dot(float4(input), float4(weight_x));
+    
+    half4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size  + i];
+    output.y += dot(float4(input), float4(weight_y));
+    
+    half4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size  + i];
+    output.z += dot(float4(input), float4(weight_z));
+    
+    half4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + i];
+    output.w += dot(float4(input), float4(weight_w));
+  }
+  output = fmax(output * float4(new_scale[gid.z]) + float4(new_biase[gid.z]), 0.0);
+  outTexture.write(half4(output), gid.xy, gid.z);
+}
+
+kernel void conv_batch_norm_relu_3x3_half(texture2d_array<half, access::sample> inTexture [[texture(0)]],
+                                     texture2d_array<half, access::write> outTexture [[texture(1)]],
+                                     constant MetalConvParam &param [[buffer(0)]],
+                                     const device half4 *weights [[buffer(1)]],
+                                     const device half4 *new_scale [[buffer(2)]],
+                                     const device half4 *new_biase [[buffer(3)]],
+                                     uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  const ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 9;
+  uint input_arr_size = inTexture.get_array_size();
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  float4 output = float4(0.0);
+  
+  half4 input[9];
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input[0] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y - 1), i);
+    input[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - 1), i);
+    input[2] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y - 1), i);
+    input[3] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y), i);
+    input[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), i);
+    input[5] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y), i);
+    input[6] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y + 1), i);
+    input[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + 1), i);
+    input[8] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y + 1), i);
+    for (int j = 0; j < 9; ++j) {
+      half4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.x += dot(float4(input[j]), float4(weight_x));
+      
+      half4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.y += dot(float4(input[j]), float4(weight_y));
+      
+      half4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.z += dot(float4(input[j]), float4(weight_z));
+      
+      half4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.w += dot(float4(input[j]), float4(weight_w));
+    }
+  }
+  output = fmax(output * float4(new_scale[gid.z]) + float4(new_biase[gid.z]), 0.0);
+  outTexture.write(half4(output), gid.xy, gid.z);
+}
+
+kernel void depthwise_conv_batch_norm_relu_3x3_half(texture2d_array<half, access::sample> inTexture [[texture(0)]],
+                                               texture2d_array<half, access::write> outTexture [[texture(1)]],
+                                               constant MetalConvParam &param [[buffer(0)]],
+                                               const device half *weights [[buffer(1)]],
+                                               const device half4 *new_scale [[buffer(2)]],
+                                               const device half4 *new_biase [[buffer(3)]],
+                                               uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  uint output_slice = gid.z;
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 9;
+  uint weithTo = gid.z * kernelHXW * 4;
+  float4 output = float4(0.0);
+  half4 inputs[9];
+  inputs[0] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y - 1), output_slice);
+  inputs[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - 1), output_slice);
+  inputs[2] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y - 1), output_slice);
+  inputs[3] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y), output_slice);
+  inputs[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), output_slice);
+  inputs[5] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y), output_slice);
+  inputs[6] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y + 1), output_slice);
+  inputs[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + 1), output_slice);
+  inputs[8] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y + 1), output_slice);
+  for (int j = 0; j < 9; ++j) {
+    half4 input = inputs[j];
+    output.x += input.x * weights[weithTo + 0 * kernelHXW + j];
+    output.y += input.y * weights[weithTo + 1 * kernelHXW + j];
+    output.z += input.z * weights[weithTo + 2 * kernelHXW + j];
+    output.w += input.w * weights[weithTo + 3 * kernelHXW + j];
+  }
+  output = fmax(output * float4(new_scale[gid.z]) + float4(new_biase[gid.z]), 0.0);
+  outTexture.write(half4(output), gid.xy, gid.z);
+}
+

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/ConvKernel.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+#include "Common.metal"
+using namespace metal;
+
+// conv
+#pragma mark -- conv
+kernel void conv_3x3(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+                     texture2d_array<float, access::write> outTexture [[texture(1)]],
+                     constant MetalConvParam &param [[buffer(0)]],
+                     const device float4 *weights [[buffer(1)]],
+                     uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  const ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 9;
+  uint input_arr_size = inTexture.get_array_size();
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  float4 output = float4(0.0);
+  
+  float4 input[9];
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input[0] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y - 1), i);
+    input[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - 1), i);
+    input[2] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y - 1), i);
+    input[3] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y), i);
+    input[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), i);
+    input[5] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y), i);
+    input[6] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y + 1), i);
+    input[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + 1), i);
+    input[8] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y + 1), i);
+    for (int j = 0; j < 9; ++j) {
+      float4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.x += dot(input[j], weight_x);
+      
+      float4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.y += dot(input[j], weight_y);
+      
+      float4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.z += dot(input[j], weight_z);
+      
+      float4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.w += dot(input[j], weight_w);
+    }
+  }
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+kernel void depthwise_conv_3x3(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+                               texture2d_array<float, access::write> outTexture [[texture(1)]],
+                               constant MetalConvParam &param [[buffer(0)]],
+                               const device float *weights [[buffer(1)]],
+                               uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  uint output_slice = gid.z;
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 9;
+  uint weithTo = gid.z * kernelHXW * 4;
+  float4 output = float4(0.0);
+  float4 inputs[9];
+  inputs[0] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y - 1), output_slice);
+  inputs[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - 1), output_slice);
+  inputs[2] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y - 1), output_slice);
+  inputs[3] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y), output_slice);
+  inputs[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), output_slice);
+  inputs[5] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y), output_slice);
+  inputs[6] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y + 1), output_slice);
+  inputs[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + 1), output_slice);
+  inputs[8] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y + 1), output_slice);
+  for (int j = 0; j < 9; ++j) {
+    float4 input = inputs[j];
+    output.x += input.x * weights[weithTo + 0 * kernelHXW + j];
+    output.y += input.y * weights[weithTo + 1 * kernelHXW + j];
+    output.z += input.z * weights[weithTo + 2 * kernelHXW + j];
+    output.w += input.w * weights[weithTo + 3 * kernelHXW + j];
+  }
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+kernel void conv_1x1(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+                     texture2d_array<float, access::write> outTexture [[texture(1)]],
+                     constant MetalConvParam &param [[buffer(0)]],
+                     const device float4 *weights [[buffer(1)]],
+                     uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 1;
+  
+  uint input_arr_size = inTexture.get_array_size();
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  float4 output = float4(0.0);
+  
+  float4 input;
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input = inTexture.sample(sample, float2(posInInput.x, posInInput.y), i);
+    float4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size  + i];
+    output.x += dot(input, weight_x);
+    
+    float4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size  + i];
+    output.y += dot(input, weight_y);
+    
+    float4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size  + i];
+    output.z += dot(input, weight_z);
+    
+    float4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + i];
+    output.w += dot(input, weight_w);
+  }
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+
+kernel void conv_3x3_half(texture2d_array<half, access::sample> inTexture [[texture(0)]],
+                     texture2d_array<half, access::write> outTexture [[texture(1)]],
+                     constant MetalConvParam &param [[buffer(0)]],
+                     const device half4 *weights [[buffer(1)]],
+                     uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  const ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 9;
+  uint input_arr_size = inTexture.get_array_size();
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  float4 output = float4(0.0);
+  
+  half4 input[9];
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input[0] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y - 1), i);
+    input[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - 1), i);
+    input[2] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y - 1), i);
+    input[3] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y), i);
+    input[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), i);
+    input[5] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y), i);
+    input[6] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y + 1), i);
+    input[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + 1), i);
+    input[8] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y + 1), i);
+    for (int j = 0; j < 9; ++j) {
+      half4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.x += dot(float4(input[j]), float4(weight_x));
+      
+      half4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.y += dot(float4(input[j]), float4(weight_y));
+      
+      half4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.z += dot(float4(input[j]), float4(weight_z));
+      
+      half4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + j * input_arr_size + i];
+      output.w += dot(float4(input[j]), float4(weight_w));
+    }
+  }
+  outTexture.write(half4(output), gid.xy, gid.z);
+}
+
+kernel void depthwise_conv_3x3_half(texture2d_array<half, access::sample> inTexture [[texture(0)]],
+                               texture2d_array<half, access::write> outTexture [[texture(1)]],
+                               constant MetalConvParam &param [[buffer(0)]],
+                               const device half *weights [[buffer(1)]],
+                               uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  uint output_slice = gid.z;
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 9;
+  uint weithTo = gid.z * kernelHXW * 4;
+  float4 output = float4(0.0);
+  half4 inputs[9];
+  inputs[0] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y - 1), output_slice);
+  inputs[1] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y - 1), output_slice);
+  inputs[2] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y - 1), output_slice);
+  inputs[3] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y), output_slice);
+  inputs[4] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y), output_slice);
+  inputs[5] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y), output_slice);
+  inputs[6] = inTexture.sample(sample, float2(posInInput.x - 1,    posInInput.y + 1), output_slice);
+  inputs[7] = inTexture.sample(sample, float2(posInInput.x,        posInInput.y + 1), output_slice);
+  inputs[8] = inTexture.sample(sample, float2(posInInput.x + 1,    posInInput.y + 1), output_slice);
+  for (int j = 0; j < 9; ++j) {
+    half4 input = inputs[j];
+    output.x += float(input.x) * float(weights[weithTo + 0 * kernelHXW + j]);
+    output.y += float(input.y) * float(weights[weithTo + 1 * kernelHXW + j]);
+    output.z += float(input.z) * float(weights[weithTo + 2 * kernelHXW + j]);
+    output.w += float(input.w) * float(weights[weithTo + 3 * kernelHXW + j]);
+  }
+  outTexture.write(half4(output), gid.xy, gid.z);
+}
+
+kernel void conv_1x1_half(texture2d_array<half, access::sample> inTexture [[texture(0)]],
+                     texture2d_array<half, access::write> outTexture [[texture(1)]],
+                     constant MetalConvParam &param [[buffer(0)]],
+                     const device half4 *weights [[buffer(1)]],
+                     uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  ushort2 stride = ushort2(param.strideX, param.strideY);
+  ushort2 posInInput = ushort2(gid.xy) * stride + ushort2(param.offsetX, param.offsetY);
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint kernelHXW = 1;
+  
+  uint input_arr_size = inTexture.get_array_size();
+  uint weithTo = gid.z * kernelHXW * input_arr_size * 4;
+  
+  float4 output = float4(0.0);
+  
+  half4 input;
+  for (uint i = 0; i < input_arr_size; ++i) {
+    input = inTexture.sample(sample, float2(posInInput.x, posInInput.y), i);
+    half4 weight_x = weights[weithTo + 0 * kernelHXW * input_arr_size  + i];
+    output.x += dot(float4(input), float4(weight_x));
+    
+    half4 weight_y = weights[weithTo + 1 * kernelHXW * input_arr_size  + i];
+    output.y += dot(float4(input), float4(weight_y));
+    
+    half4 weight_z = weights[weithTo + 2 * kernelHXW * input_arr_size  + i];
+    output.z += dot(float4(input), float4(weight_z));
+    
+    half4 weight_w = weights[weithTo + 3 * kernelHXW * input_arr_size + i];
+    output.w += dot(float4(input), float4(weight_w));
+  }
+  outTexture.write(half4(output), gid.xy, gid.z);
+}
+
+

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/ConvTransposeKernel.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+using namespace metal;
+
+struct MetalConvTransposeParam{
+  ushort kernelW;
+  ushort kernelH;
+  
+  ushort strideX;
+  ushort strideY;
+  
+  ushort paddingX;
+  ushort paddingY;
+  
+  ushort dilationX;
+  ushort dilationY;
+};
+
+kernel void conv_transpose2x2_stride2(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+                                      texture2d_array<float, access::write> outTexture [[texture(1)]],
+                                      constant MetalConvTransposeParam &param [[buffer(0)]],
+                                      const device float4 *weights [[buffer(1)]],
+                                      uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  int input_array_size = inTexture.get_array_size();
+  int kernel_index_x = gid.x % 2;
+  int kernel_index_y = gid.y % 2;
+  int kernel_index = kernel_index_y * 2 + kernel_index_x;
+  int kernel_to = gid.z * input_array_size * 4 * 4 + (kernel_index * input_array_size);
+  int input_x = gid.x / 2;
+  int input_y = gid.y / 2;
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  float4 output = float4(0.0);
+  for (int i = 0; i < input_array_size; ++i) {
+    
+    float4 input = inTexture.sample(sample, float2(input_x, input_y), i);
+    
+    float4 kernel_slice0 = weights[kernel_to + input_array_size * 4 * 0 + i];
+    float4 kernel_slice1 = weights[kernel_to + input_array_size * 4 * 1 + i];
+    float4 kernel_slice2 = weights[kernel_to + input_array_size * 4 * 2 + i];
+    float4 kernel_slice3 = weights[kernel_to + input_array_size * 4 * 3 + i];
+    
+    output.x += dot(input, kernel_slice0);
+    
+    output.y += dot(input, kernel_slice1);
+    
+    output.z += dot(input, kernel_slice2);
+    
+    output.w += dot(input, kernel_slice3);
+  }
+  
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+kernel void conv_transpose2x2_stride2_half(texture2d_array<half, access::sample> inTexture [[texture(0)]],
+                                      texture2d_array<half, access::write> outTexture [[texture(1)]],
+                                      constant MetalConvTransposeParam &param [[buffer(0)]],
+                                      const device half4 *weights [[buffer(1)]],
+                                      uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  int input_array_size = inTexture.get_array_size();
+  int kernel_index_x = gid.x % 2;
+  int kernel_index_y = gid.y % 2;
+  int kernel_index = kernel_index_y * 2 + kernel_index_x;
+  int kernel_to = gid.z * input_array_size * 4 * 4 + (kernel_index * input_array_size);
+  int input_x = gid.x / 2;
+  int input_y = gid.y / 2;
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  float4 output = float4(0.0);
+  for (int i = 0; i < input_array_size; ++i) {
+    
+    half4 input = inTexture.sample(sample, float2(input_x, input_y), i);
+    
+    half4 kernel_slice0 = weights[kernel_to + input_array_size * 4 * 0 + i];
+    half4 kernel_slice1 = weights[kernel_to + input_array_size * 4 * 1 + i];
+    half4 kernel_slice2 = weights[kernel_to + input_array_size * 4 * 2 + i];
+    half4 kernel_slice3 = weights[kernel_to + input_array_size * 4 * 3 + i];
+    
+    output.x += dot(float4(input), float4(kernel_slice0));
+    
+    output.y += dot(float4(input), float4(kernel_slice1));
+    
+    output.z += dot(float4(input), float4(kernel_slice2));
+    
+    output.w += dot(float4(input), float4(kernel_slice3));
+  }
+  
+  outTexture.write(half4(output), gid.xy, gid.z);
+}
+
+//kernel void conv_transpose(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+//                           texture2d_array<float, access::write> outTexture [[texture(1)]],
+//                           constant MetalConvTransposeParam &param [[buffer(0)]],
+//                           const device float4 *weights [[buffer(1)]],
+//                           uint3 gid [[thread_position_in_grid]]){
+//  if (gid.x >= outTexture.get_width() ||
+//      gid.y >= outTexture.get_height() ||
+//      gid.z >= outTexture.get_array_size()) {
+//    return;
+//  }
+//
+//  int input_array_size = inTexture.get_array_size();
+//
+//  uint kernel_one_output_slice = input_array_size * param.kernelW * param.kernelH;
+//
+//  uint kernel_stride_z = gid.z * 4 * (kernel_one_output_slice);
+//
+//  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+//
+//  float4 output;
+//
+//  for (int w = 0; w < param.kernelW; ++w) {
+//    int top = gid.x - w * param.dilationX + param.paddingX;
+//    int input_x = top / param.strideX;
+//    if (top < 0 || input_x >= int(inTexture.get_width())) {
+//      continue;
+//    }
+//
+//    for (int h = 0; h < param.kernelH; ++h) {
+//      int top_y = gid.y - h * param.dilationY + param.paddingY;
+//      int input_y = top_y / param.strideY;
+//      if (top_y < 0 || input_y >= int(inTexture.get_height())) {
+//        continue;
+//      }
+//
+//      uint kernel_index = (w * param.kernelH + h) * inTexture.get_array_size();
+//
+//      for (int slice = 0; slice < input_array_size; ++slice) {
+//
+//        float4 input;
+//        float4 kernel_slice = weights[kernel_stride_z + 0 * kernel_one_output_slice + kernel_index + slice];
+//        float4 kernel_slice1 = weights[kernel_stride_z + 1 * kernel_one_output_slice + kernel_index + slice];
+//
+//        float4 kernel_slice2 = weights[kernel_stride_z + 2 * kernel_one_output_slice + kernel_index + slice];
+//
+//        float4 kernel_slice3 = weights[kernel_stride_z + 3 * kernel_one_output_slice + kernel_index + slice];
+//
+//        input = inTexture.sample(sample, float2(input_x,    input_y), slice);
+//        output.x += dot(input, kernel_slice);
+//        output.y += dot(input, kernel_slice1);
+//        output.z += dot(input, kernel_slice2);
+//        output.w += dot(input, kernel_slice3);
+//      }
+//    }
+//  }
+//
+//  outTexture.write(output, gid.xy, gid.z);
+//}
+//

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/Elementwise.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+#include "Common.metal"
+
+using namespace metal;
+
+struct ElementwiseAddParam {
+  int32_t fast;
+  int32_t axis;
+  int32_t ylen;
+  int32_t xdim[4];
+  int32_t xtrans[4];
+  int32_t ydim[4];
+  int32_t ytrans[4];
+};
+
+kernel void elementwise_add(texture2d_array<float, access::read> inputX [[texture(0)]],
+                            texture2d_array<float, access::read> inputY [[texture(1)]],
+                            texture2d_array<float, access::write> outTexture [[texture(2)]],
+                            constant ElementwiseAddParam &pm [[buffer(0)]],
+                            uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) return;
+  float4 rx, ry;
+
+  if (pm.fast == 1) {
+    rx = inputX.read(gid.xy, gid.z);
+    ry = inputY.read(gid.xy, gid.z);
+  } else {
+    rx = inputX.read(gid.xy, gid.z);
+    int32_t x_xyzn[4] = {int32_t(gid.x), int32_t(gid.y), int32_t(gid.z), 0}, x_abcd[4], t_abcd[4];
+    int32_t y_abcd[4] = {0, 0, 0, 0}, y_xyzn[4];
+    int32_t xtrans[4] = {pm.xtrans[0], pm.xtrans[1], pm.xtrans[2], pm.xtrans[3]};
+    int32_t ytrans[4] = {pm.ytrans[0], pm.ytrans[1], pm.ytrans[2], pm.ytrans[3]};
+    int32_t yshift = 4 - pm.ylen - pm.axis;
+    for (int n = 0; n < 4; n++) {
+      x_xyzn[3] = n;
+      xyzn2abcd(pm.xdim[3], x_xyzn, x_abcd);
+      invtrans(xtrans, x_abcd, t_abcd);
+      for (int k = pm.axis; k < (pm.axis + pm.ylen); k++) {
+        y_abcd[yshift+k] = t_abcd[k];
+      }
+      trans(ytrans, y_abcd, t_abcd);
+      abcd2xyzn(pm.ydim[3], t_abcd, y_xyzn);
+      ry[n] = inputY.read(uint2(y_xyzn[0], y_xyzn[1]), y_xyzn[2])[y_xyzn[3]];
+    }
+  }
+  float4 r = rx + ry;
+  outTexture.write(r, gid.xy, gid.z);
+}
+
+kernel void elementwise_add_half(texture2d_array<half, access::read> inputX [[texture(0)]],
+                            texture2d_array<half, access::read> inputY [[texture(1)]],
+                            texture2d_array<half, access::write> outTexture [[texture(2)]],
+                            constant ElementwiseAddParam &pm [[buffer(0)]],
+                            uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) return;
+  half4 rx, ry;
+
+  if (pm.fast == 1) {
+    rx = inputX.read(gid.xy, gid.z);
+    ry = inputY.read(gid.xy, gid.z);
+  } else {
+    rx = inputX.read(gid.xy, gid.z);
+    int32_t x_xyzn[4] = {int32_t(gid.x), int32_t(gid.y), int32_t(gid.z), 0}, x_abcd[4], t_abcd[4];
+    int32_t y_abcd[4] = {0, 0, 0, 0}, y_xyzn[4];
+    int32_t xtrans[4] = {pm.xtrans[0], pm.xtrans[1], pm.xtrans[2], pm.xtrans[3]};
+    int32_t ytrans[4] = {pm.ytrans[0], pm.ytrans[1], pm.ytrans[2], pm.ytrans[3]};
+    int32_t yshift = 4 - pm.ylen - pm.axis;
+    for (int n = 0; n < 4; n++) {
+      x_xyzn[3] = n;
+      xyzn2abcd(pm.xdim[3], x_xyzn, x_abcd);
+      invtrans(xtrans, x_abcd, t_abcd);
+      for (int k = pm.axis; k < (pm.axis + pm.ylen); k++) {
+        y_abcd[yshift+k] = t_abcd[k];
+      }
+      trans(ytrans, y_abcd, t_abcd);
+      abcd2xyzn(pm.ydim[3], t_abcd, y_xyzn);
+      ry[n] = inputY.read(uint2(y_xyzn[0], y_xyzn[1]), y_xyzn[2])[y_xyzn[3]];
+    }
+  }
+  half4 r = rx + ry;
+  outTexture.write(r, gid.xy, gid.z);
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/ElementwiseAddPreluKernel.inc.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#ifdef P
+
+#include <metal_stdlib>
+#include "Macro.metal"
+
+using namespace metal;
+
+kernel void FUNC3_(elementwise_add, PRELU_TYPE, P)(texture2d_array<P, access::read> inputX [[texture(0)]],
+                                 texture2d_array<P, access::read> inputY [[texture(1)]],
+                                 texture2d_array<P, access::write> outTexture [[texture(2)]],
+                                 constant ElementwiseAddParam &pm [[buffer(0)]],
+#ifdef PRELU_CHANNEL
+                                 const device VECTOR(P, 4) *alpha [[buffer(1)]],
+#endif
+#ifdef PRELU_ELEMENT
+                                 const device VECTOR(P, 4) *alpha [[buffer(1)]],
+#endif
+#ifdef PRELU_OTHER
+                                 const device P *alpha [[buffer(1)]],
+#endif
+                                 uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) return;
+  VECTOR(P, 4) rx, ry;
+  
+  if (pm.fast == 1) {
+    rx = inputX.read(gid.xy, gid.z);
+    ry = inputY.read(gid.xy, gid.z);
+    } else {
+      rx = inputX.read(gid.xy, gid.z);
+      int32_t x_xyzn[4] = {int32_t(gid.x), int32_t(gid.y), int32_t(gid.z), 0}, x_abcd[4], t_abcd[4];
+      int32_t y_abcd[4] = {0, 0, 0, 0}, y_xyzn[4];
+      int32_t xtrans[4] = {pm.xtrans[0], pm.xtrans[1], pm.xtrans[2], pm.xtrans[3]};
+      int32_t ytrans[4] = {pm.ytrans[0], pm.ytrans[1], pm.ytrans[2], pm.ytrans[3]};
+      int32_t yshift = 4 - pm.ylen - pm.axis;
+      for (int n = 0; n < 4; n++) {
+        x_xyzn[3] = n;
+        xyzn2abcd(pm.xdim[3], x_xyzn, x_abcd);
+        invtrans(xtrans, x_abcd, t_abcd);
+        for (int k = pm.axis; k < (pm.axis + pm.ylen); k++) {
+          y_abcd[yshift+k] = t_abcd[k];
+        }
+        trans(ytrans, y_abcd, t_abcd);
+        abcd2xyzn(pm.ydim[3], t_abcd, y_xyzn);
+        ry[n] = inputY.read(uint2(y_xyzn[0], y_xyzn[1]), y_xyzn[2])[y_xyzn[3]];
+      }
+  }
+  VECTOR(P, 4) output = rx + ry;
+  
+#ifdef PRELU_CHANNEL
+  VECTOR(P, 4) alpha_value = alpha[gid.z];
+  output.x = output.x > 0 ? output.x : (alpha_value.x * output.x);
+  output.y = output.y > 0 ? output.y : (alpha_value.y * output.y);
+  output.z = output.z > 0 ? output.z : (alpha_value.z * output.z);
+  output.w = output.w > 0 ? output.w : (alpha_value.w * output.w);
+#endif
+#ifdef PRELU_ELEMENT
+  int alpha_to = (gid.y * outTexture.get_width() + gid.x) * outTexture.get_array_size();
+  VECTOR(P, 4) alpha_value = alpha[alpha_to + gid.z];
+  output.x = output.x > 0 ? output.x : (alpha_value.x * output.x);
+  output.y = output.y > 0 ? output.y : (alpha_value.y * output.y);
+  output.z = output.z > 0 ? output.z : (alpha_value.z * output.z);
+  output.w = output.w > 0 ? output.w : (alpha_value.w * output.w);
+#endif
+#ifdef PRELU_OTHER
+  P alpha_value = alpha[0];
+  output.x = output.x > 0 ? output.x : (alpha_value * output.x);
+  output.y = output.y > 0 ? output.y : (alpha_value * output.y);
+  output.z = output.z > 0 ? output.z : (alpha_value * output.z);
+  output.w = output.w > 0 ? output.w : (alpha_value * output.w);
+#endif
+  
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+#endif

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/ElementwiseAddPreluKernel.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+#include "Common.metal"
+using namespace metal;
+
+struct ElementwiseAddParam {
+  int32_t fast;
+  int32_t axis;
+  int32_t ylen;
+  int32_t xdim[4];
+  int32_t xtrans[4];
+  int32_t ydim[4];
+  int32_t ytrans[4];
+};
+
+#define P float
+
+#define PRELU_CHANNEL prelu_channel
+#define PRELU_TYPE channel
+#include "ElementwiseAddPreluKernel.inc.metal"
+#undef  PRELU_TYPE
+#undef  PRELU_CHANNEL
+
+#define PRELU_ELEMENT element
+#define PRELU_TYPE prelu_element
+#include "ElementwiseAddPreluKernel.inc.metal"
+#undef  PRELU_TYPE
+#undef  PRELU_ELEMENT
+
+#define PRELU_OTHER   other
+#define PRELU_TYPE prelu_other
+#include "ElementwiseAddPreluKernel.inc.metal"
+#undef  PRELU_TYPE
+#undef  PRELU_OTHER
+
+#undef P
+
+#define P half
+
+#define PRELU_CHANNEL channel
+#define PRELU_TYPE channel
+#include "ElementwiseAddPreluKernel.inc.metal"
+#undef  PRELU_TYPE
+#undef  PRELU_CHANNEL
+
+#define PRELU_ELEMENT element
+#define PRELU_TYPE prelu_element
+#include "ElementwiseAddPreluKernel.inc.metal"
+#undef  PRELU_TYPE
+#undef  PRELU_ELEMENT
+
+#define PRELU_OTHER   other
+#define PRELU_TYPE prelu_other
+#include "ElementwiseAddPreluKernel.inc.metal"
+#undef  PRELU_TYPE
+#undef  PRELU_OTHER
+
+#undef P
+
+
+
+

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/FetchKernel.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+using namespace metal;
+
+kernel void fetch(texture2d_array<float, access::read> inTexture [[texture(0)]],
+                       device float *output [[buffer(0)]],
+                      uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= inTexture.get_width() ||
+      gid.y >= inTexture.get_height() ||
+      gid.z >= inTexture.get_array_size()) {
+    return;
+  }
+  
+  int input_width = inTexture.get_width();
+  int input_height = inTexture.get_height();
+  const float4 input = inTexture.read(gid.xy, gid.z);
+  int output_to = 4 * input_width * input_height;
+  output[gid.z * output_to + 0 * input_width * input_height + gid.y * input_width + gid.x] = input.x;
+  output[gid.z * output_to + 1 * input_width * input_height + gid.y * input_width + gid.x] = input.y;
+//  output[gid.z * output_to + 2 * input_width * input_height + gid.y * input_width + gid.x] = input.z;
+//  output[gid.z * output_to + 3 * input_width * input_height + gid.y * input_width + gid.x] = input.w;
+}
+
+
+kernel void fetch_half(texture2d_array<half, access::read> inTexture [[texture(0)]],
+                  device float * output [[buffer(0)]],
+                  uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= inTexture.get_width() ||
+      gid.y >= inTexture.get_height() ||
+      gid.z >= inTexture.get_array_size()) {
+    return;
+  }
+  
+  int input_width = inTexture.get_width();
+  int input_height = inTexture.get_height();
+  const half4 input = inTexture.read(gid.xy, gid.z);
+  int output_to = 4 * input_width * input_height;
+  output[gid.z * output_to + 0 * input_width * input_height + gid.y * input_width + gid.x] = input.x;
+  output[gid.z * output_to + 1 * input_width * input_height + gid.y * input_width + gid.x] = input.y;
+//  output[gid.z * output_to + 2 * input_width * input_height + gid.y * input_width + gid.x] = input.z;
+//  output[gid.z * output_to + 3 * input_width * input_height + gid.y * input_width + gid.x] = input.w;
+  
+}
+
+kernel void fetch_placeholder(texture2d_array<float, access::read> inTexture [[texture(0)]],
+                              device float *output [[buffer(0)]],
+                              uint3 gid [[thread_position_in_grid]]) {
+  
+}
+
+kernel void fetch_placeholder_half(texture2d_array<half, access::read> inTexture [[texture(0)]],
+                              device float *output [[buffer(0)]],
+                              uint3 gid [[thread_position_in_grid]]) {
+}
+
+

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/Kernels.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+#include "Common.metal"
+using namespace metal;
+
+// 占位函数, 啥也没干
+kernel void place_holder(texture2d<half, access::read> inTexture [[texture(0)]],
+                         texture2d_array<half, access::write> outTexture [[texture(1)]],
+                         uint3 gid [[thread_position_in_grid]]) {
+}
+
+struct OutputDim {
+  ushort width;
+  ushort height;
+  ushort strideX;
+  ushort strideY;
+};
+
+kernel void resize(texture2d<half, access::read> inTexture [[texture(0)]],
+                   texture2d_array<half, access::write> outTexture [[texture(1)]],
+                   constant OutputDim &params [[buffer(0)]],
+                   uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) return;
+  
+  constexpr sampler s(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const uint2 pos = gid.xy * uint2(params.strideX, params.strideY);
+  const half4 input = inTexture.read(pos);
+  outTexture.write(half4(input.x, input.y, input.z, input.w), gid.xy, gid.z);
+}
+
+
+kernel void texture2d_to_2d_array(texture2d<float, access::read> inTexture [[texture(0)]],
+                                  texture2d_array<float, access::write> outTexture [[texture(1)]],
+                                  uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= inTexture.get_width() ||
+      gid.y >= inTexture.get_height()){
+    return;
+  }
+  const float4 input = inTexture.read(gid.xy);
+  outTexture.write(input, gid.xy, 0);
+}
+
+kernel void texture2d_to_2d_array_half(texture2d<half, access::read> inTexture [[texture(0)]],
+                                      texture2d_array<half, access::write> outTexture [[texture(1)]],
+                                      uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= inTexture.get_width() ||
+      gid.y >= inTexture.get_height()){
+    return;
+  }
+  const half4 input = inTexture.read(gid.xy);
+  outTexture.write(input, gid.xy, 0);
+}
+
+

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/Macro.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+using namespace metal;
+
+
+#define CONCAT2(a, b) a ## b
+#define CONCAT2_(a, b) a ## _ ## b
+#define CONCAT3_(a, b, c) a ## _ ## b ## _ ## c
+#define CONCAT4_(a, b, c, d) a ## _ ## b ## _ ## c ## _ ## d
+#define CONCAT5_(a, b, c, d, e) a ## _ ## b ## _ ## c ## _ ## d ## _ ## e
+
+#define FUNC(f, r, n, v, p) CONCAT5_(f, r, n, v, p)
+#define VECTOR(p, n) CONCAT2(p, n)
+
+#define FUNC3_(a, b, c) CONCAT3_(a, b, c)
+

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/NMSFetchResultKernel.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+using namespace metal;
+
+kernel void nms_fetch_result(texture2d_array<float, access::read> inTexture [[texture(0)]],
+    device float *output [[buffer(0)]],
+    uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= inTexture.get_width() ||
+      gid.y >= inTexture.get_height() ||
+      gid.z >= inTexture.get_array_size()) {
+    return;
+  }
+  
+  int input_width = inTexture.get_width();
+  const float4 input = inTexture.read(gid.xy, gid.z);
+  output[gid.y * input_width + gid.x] = input.x;
+  
+}
+
+
+kernel void nms_fetch_result_half(texture2d_array<half, access::read> inTexture [[texture(0)]],
+                             device float *output [[buffer(0)]],
+                             uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= inTexture.get_width() ||
+      gid.y >= inTexture.get_height() ||
+      gid.z >= inTexture.get_array_size()) {
+    return;
+  }
+  
+  int input_width = inTexture.get_width();
+  const half4 input = inTexture.read(gid.xy, gid.z);
+  output[gid.y * input_width + gid.x] = input.x;
+}
+
+kernel void nms_fetch_bbox(texture2d_array<float, access::read> inTexture [[texture(0)]],
+    device float4 *output [[buffer(0)]],
+    uint3 gid [[thread_position_in_grid]]) {
+  
+  if (gid.x >= inTexture.get_width() ||
+      gid.y >= inTexture.get_height() ||
+      gid.z >= inTexture.get_array_size()) {
+    return;
+  }
+  
+  int input_width = inTexture.get_width();
+//  int input_height = inTexture.get_height();
+  const float4 input = inTexture.read(gid.xy, gid.z);
+  output[gid.y * input_width + gid.x] = input;
+}
+
+kernel void nms_fetch_bbox_half(texture2d_array<half, access::read> inTexture [[texture(0)]],
+                           device float4 *output [[buffer(0)]],
+                           uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= inTexture.get_width() ||
+      gid.y >= inTexture.get_height() ||
+      gid.z >= inTexture.get_array_size()) {
+    return;
+  }
+  
+  int input_width = inTexture.get_width();
+//  int input_height = inTexture.get_height();
+  const half4 input = inTexture.read(gid.xy, gid.z);
+  output[gid.y * input_width + gid.x] = float4(input);
+}
+

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/PoolKernel.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+#include "Common.metal"
+using namespace metal;
+
+struct PoolParam {
+  int ksizeX;
+  int ksizeY;
+  int strideX;
+  int strideY;
+  int paddingX;
+  int paddingY;
+  int poolType;
+};
+
+kernel void pool(texture2d_array<float, access::read> inTexture [[texture(0)]],
+                 texture2d_array<float, access::write> outTexture [[texture(1)]],
+                 constant PoolParam &pm [[buffer(0)]],
+                 uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) return;
+  int xmin = gid.x * pm.strideX - pm.paddingX;
+  int xmax = min(xmin + pm.ksizeX, int(inTexture.get_width()));
+  xmin = max(xmin, 0);
+  int ymin = gid.y * pm.strideX - pm.paddingX;
+  int ymax = min(ymin + pm.ksizeX, int(inTexture.get_height()));
+  ymin = max(ymin, 0);
+  
+  float4 r = 0;
+  if (pm.poolType == 0) {
+    r = inTexture.read(uint2(xmin, ymin), gid.z);
+    for (int x = xmin; x < xmax; x++) {
+      for (int y = ymin; y < ymax; y++) {
+        r = fmax(r, inTexture.read(uint2(x, y), gid.z));
+      }
+    }
+  } else if (pm.poolType == 1) {
+    for (int x = xmin; x < xmax; x++) {
+      for (int y = ymin; y < ymax; y++) {
+        r += inTexture.read(uint2(x, y), gid.z);
+      }
+    }
+    r /= pm.ksizeX * pm.ksizeY;
+  }
+  outTexture.write(r, gid.xy, gid.z);
+}
+
+kernel void pool_half(texture2d_array<half, access::read> inTexture [[texture(0)]],
+                      texture2d_array<half, access::write> outTexture [[texture(1)]],
+                      constant PoolParam &pm [[buffer(0)]],
+                      uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) return;
+  int xmin = gid.x * pm.strideX - pm.paddingX;
+  int xmax = min(xmin + pm.ksizeX, int(inTexture.get_width()));
+  xmin = max(xmin, 0);
+  int ymin = gid.y * pm.strideX - pm.paddingX;
+  int ymax = min(ymin + pm.ksizeX, int(inTexture.get_height()));
+  ymin = max(ymin, 0);
+  
+  half4 r = 0;
+  if (pm.poolType == 0) {
+    r = inTexture.read(uint2(xmin, ymin), gid.z);
+    for (int x = xmin; x < xmax; x++) {
+      for (int y = ymin; y < ymax; y++) {
+        r = fmax(r, inTexture.read(uint2(x, y), gid.z));
+      }
+    }
+  } else if (pm.poolType == 1) {
+    for (int x = xmin; x < xmax; x++) {
+      for (int y = ymin; y < ymax; y++) {
+        r += inTexture.read(uint2(x, y), gid.z);
+      }
+    }
+    r /= pm.ksizeX * pm.ksizeY;
+  }
+  outTexture.write(r, gid.xy, gid.z);
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/PreluKernel.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+using namespace metal;
+
+kernel void prelu_channel(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+                           texture2d_array<float, access::write> outTexture [[texture(1)]],
+                           const device float4 *alpha [[buffer(0)]],
+                           uint3 gid [[thread_position_in_grid]]){
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  float4 input = inTexture.sample(sample, float2(gid.x, gid.y), gid.z);
+  float4 alpha_value = alpha[gid.z];
+  float4 output;
+  output.x = input.x > 0 ? input.x : (alpha_value.x * input.x);
+  output.y = input.y > 0 ? input.y : (alpha_value.y * input.y);
+  output.z = input.z > 0 ? input.z : (alpha_value.z * input.z);
+  output.w = input.w > 0 ? input.w : (alpha_value.w * input.w);
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+kernel void prelu_element(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+                          texture2d_array<float, access::write> outTexture [[texture(1)]],
+                          const device float4 *alpha [[buffer(0)]],
+                          uint3 gid [[thread_position_in_grid]]){
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  float4 input = inTexture.sample(sample, float2(gid.x, gid.y), gid.z);
+
+  int alpha_to = (gid.y * inTexture.get_width() + gid.x) * inTexture.get_array_size();
+  float4 alpha_value = alpha[alpha_to + gid.z];
+
+  float4 output;
+  output.x = input.x > 0 ? input.x : (alpha_value.x * input.x);
+  output.y = input.y > 0 ? input.y : (alpha_value.y * input.y);
+  output.z = input.z > 0 ? input.z : (alpha_value.z * input.z);
+  output.w = input.w > 0 ? input.w : (alpha_value.w * input.w);
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+kernel void prelu_other(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+                          texture2d_array<float, access::write> outTexture [[texture(1)]],
+                          const device float *alpha [[buffer(0)]],
+                          uint3 gid [[thread_position_in_grid]]){
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  float4 input = inTexture.sample(sample, float2(gid.x, gid.y), gid.z);
+  float alpha_value = alpha[0];
+  float4 output;
+  output.x = input.x > 0 ? input.x : (alpha_value * input.x);
+  output.y = input.y > 0 ? input.y : (alpha_value * input.y);
+  output.z = input.z > 0 ? input.z : (alpha_value * input.z);
+  output.w = input.w > 0 ? input.w : (alpha_value * input.w);
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+
+kernel void prelu_channel_half(texture2d_array<half, access::sample> inTexture [[texture(0)]],
+                          texture2d_array<half, access::write> outTexture [[texture(1)]],
+                          const device half4 *alpha [[buffer(0)]],
+                          uint3 gid [[thread_position_in_grid]]){
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  half4 input = inTexture.sample(sample, float2(gid.x, gid.y), gid.z);
+  half4 alpha_value = alpha[gid.z];
+  half4 output;
+  output.x = input.x > 0 ? input.x : (alpha_value.x * input.x);
+  output.y = input.y > 0 ? input.y : (alpha_value.y * input.y);
+  output.z = input.z > 0 ? input.z : (alpha_value.z * input.z);
+  output.w = input.w > 0 ? input.w : (alpha_value.w * input.w);
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+kernel void prelu_element_half(texture2d_array<half, access::sample> inTexture [[texture(0)]],
+                          texture2d_array<half, access::write> outTexture [[texture(1)]],
+                          const device half4 *alpha [[buffer(0)]],
+                          uint3 gid [[thread_position_in_grid]]){
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  half4 input = inTexture.sample(sample, float2(gid.x, gid.y), gid.z);
+  
+  int alpha_to = (gid.y * inTexture.get_width() + gid.x) * inTexture.get_array_size();
+  half4 alpha_value = alpha[alpha_to + gid.z];
+  
+  half4 output;
+  output.x = input.x > 0 ? input.x : (alpha_value.x * input.x);
+  output.y = input.y > 0 ? input.y : (alpha_value.y * input.y);
+  output.z = input.z > 0 ? input.z : (alpha_value.z * input.z);
+  output.w = input.w > 0 ? input.w : (alpha_value.w * input.w);
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+kernel void prelu_other_half(texture2d_array<half, access::sample> inTexture [[texture(0)]],
+                        texture2d_array<half, access::write> outTexture [[texture(1)]],
+                        const device half *alpha [[buffer(0)]],
+                        uint3 gid [[thread_position_in_grid]]){
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) {
+    return;
+  }
+  
+  constexpr sampler sample(coord::pixel, filter::nearest, address::clamp_to_zero);
+  half4 input = inTexture.sample(sample, float2(gid.x, gid.y), gid.z);
+  half alpha_value = alpha[0];
+  half4 output;
+  output.x = input.x > 0 ? input.x : (alpha_value * input.x);
+  output.y = input.y > 0 ? input.y : (alpha_value * input.y);
+  output.z = input.z > 0 ? input.z : (alpha_value * input.z);
+  output.w = input.w > 0 ? input.w : (alpha_value * input.w);
+  outTexture.write(output, gid.xy, gid.z);
+}
+
+

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/PriorBoxKernel.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+using namespace metal;
+
+struct PriorBoxMetalParam {
+  float offset;
+  float stepWidth;
+  float stepHeight;
+  float minSize;
+  float maxSize;
+  float imageWidth;
+  float imageHeight;
+  
+  bool clip;
+  
+  uint numPriors;
+  uint aspecRatiosSize;
+  uint minSizeSize;
+  uint maxSizeSize;
+};
+
+kernel void prior_box(texture2d_array<float, access::read> inTexture [[texture(0)]],
+                      texture2d_array<float, access::write> outBoxTexture [[texture(1)]],
+                      texture2d_array<float, access::write> varianceTexture [[texture(2)]],
+                      const device float *aspect_ratios [[buffer(0)]],
+                      constant PriorBoxMetalParam &param [[buffer(1)]],
+                      const device float4 *variances [[buffer(2)]],
+                      uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= outBoxTexture.get_width() ||
+      gid.y >= outBoxTexture.get_height() ||
+      gid.z >= outBoxTexture.get_array_size()) return;
+  
+  float center_x = (gid.x + param.offset) * param.stepWidth;
+  float center_y = (gid.y + param.offset) * param.stepHeight;
+  
+  float box_width, box_height;
+  
+  if (gid.z < param.aspecRatiosSize) {
+    float ar = aspect_ratios[gid.z];
+    box_width = param.minSize * sqrt(ar) / 2;
+    box_height = param.minSize / sqrt(ar) / 2;
+    float4 box;
+    box.x = (center_x - box_width) / param.imageWidth;
+    box.y = (center_y - box_height) / param.imageHeight;
+    box.z = (center_x + box_width) / param.imageWidth;
+    box.w = (center_y + box_height) / param.imageHeight;
+    
+    float4 res;
+    if (param.clip) {
+      res = fmin(fmax(box, 0.0), 1.0);
+    } else {
+      res = box;
+    }
+    
+    outBoxTexture.write(res, gid.xy, gid.z);
+  } else if (gid.z >= param.aspecRatiosSize) {
+    if (param.maxSizeSize > 0) {
+      box_width = box_height = sqrt(param.minSize * param.maxSize) / 2;
+      float4 max_box;
+      max_box.x = (center_x - box_width) / param.imageWidth;
+      max_box.y = (center_y - box_height) / param.imageHeight;
+      max_box.z = (center_x + box_width) / param.imageWidth;
+      max_box.w = (center_y + box_height) / param.imageHeight;
+
+      float4 res;
+      if (param.clip) {
+        res = min(max(max_box, 0.0), 1.0);
+      } else {
+        res = max_box;
+      }
+      outBoxTexture.write(max_box, gid.xy, gid.z);
+    }
+  }
+  
+  float4 variance = variances[0];
+  if (gid.z < param.numPriors) {
+    float4 variances_output;
+    variances_output.x = variance.x;
+    variances_output.y = variance.y;
+    variances_output.z = variance.z;
+    variances_output.w = variance.w;
+    varianceTexture.write(variances_output, gid.xy, gid.z);
+  }
+}
+
+
+kernel void prior_box_half(texture2d_array<half, access::read> inTexture [[texture(0)]],
+                      texture2d_array<half, access::write> outBoxTexture [[texture(1)]],
+                      texture2d_array<half, access::write> varianceTexture [[texture(2)]],
+                      const device half *aspect_ratios [[buffer(0)]],
+                      constant PriorBoxMetalParam &param [[buffer(1)]],
+                      const device float4 *variances [[buffer(2)]],
+                      uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= outBoxTexture.get_width() ||
+      gid.y >= outBoxTexture.get_height() ||
+      gid.z >= outBoxTexture.get_array_size()) return;
+  
+  float center_x = (gid.x + param.offset) * param.stepWidth;
+  float center_y = (gid.y + param.offset) * param.stepHeight;
+  
+  float box_width, box_height;
+  
+  if (gid.z < param.aspecRatiosSize) {
+    half ar = aspect_ratios[gid.z];
+    box_width = param.minSize * sqrt(ar) / 2;
+    box_height = param.minSize / sqrt(ar) / 2;
+    float4 box;
+    box.x = (center_x - box_width) / param.imageWidth;
+    box.y = (center_y - box_height) / param.imageHeight;
+    box.z = (center_x + box_width) / param.imageWidth;
+    box.w = (center_y + box_height) / param.imageHeight;
+    
+    float4 res;
+    if (param.clip) {
+      res = fmin(fmax(box, 0.0), 1.0);
+    } else {
+      res = box;
+    }
+    
+    outBoxTexture.write(half4(res), gid.xy, gid.z);
+  } else if (gid.z >= param.aspecRatiosSize) {
+    if (param.maxSizeSize > 0) {
+      box_width = box_height = sqrt(param.minSize * param.maxSize) / 2;
+      float4 max_box;
+      max_box.x = (center_x - box_width) / param.imageWidth;
+      max_box.y = (center_y - box_height) / param.imageHeight;
+      max_box.z = (center_x + box_width) / param.imageWidth;
+      max_box.w = (center_y + box_height) / param.imageHeight;
+      
+      float4 res;
+      if (param.clip) {
+        res = min(max(max_box, 0.0), 1.0);
+      } else {
+        res = max_box;
+      }
+      outBoxTexture.write(half4(max_box), gid.xy, gid.z);
+    }
+  }
+  
+  float4 variance = variances[0];
+  if (gid.z < param.numPriors) {
+    float4 variances_output;
+    variances_output.x = variance.x;
+    variances_output.y = variance.y;
+    variances_output.z = variance.z;
+    variances_output.w = variance.w;
+    varianceTexture.write(half4(variances_output), gid.xy, gid.z);
+  }
+}
+
+
+
+kernel void prior_box_MinMaxAspectRatiosOrder(texture2d_array<float, access::read> inTexture [[texture(0)]],
+                      texture2d_array<float, access::write> outBoxTexture [[texture(1)]],
+                      texture2d_array<float, access::write> varianceTexture [[texture(2)]],
+                      const device float *aspect_ratios [[buffer(0)]],
+                      constant PriorBoxMetalParam &param [[buffer(1)]],
+                      const device float4 *variances [[buffer(2)]],
+                      uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= outBoxTexture.get_width() ||
+      gid.y >= outBoxTexture.get_height() ||
+      gid.z >= outBoxTexture.get_array_size()) return;
+  
+  float center_x = (gid.x + param.offset) * param.stepWidth;
+  float center_y = (gid.y + param.offset) * param.stepHeight;
+  
+  float box_width, box_height;
+  
+  
+  
+  if (gid.z == 0) {
+    box_width = box_height = param.minSize / 2;
+    
+    float4 box;
+    box.x = (center_x - box_width) / param.imageWidth;
+    box.y = (center_y - box_height) / param.imageHeight;
+    box.z = (center_x + box_width) / param.imageWidth;
+    box.w = (center_y + box_height) / param.imageHeight;
+    
+    float4 res;
+    if (param.clip) {
+      res = fmin(fmax(box, 0.0), 1.0);
+    } else {
+      res = box;
+    }
+
+    outBoxTexture.write(res, gid.xy, gid.z);
+  }
+  
+  if (gid.z == 1 && param.maxSizeSize > 0) {
+    
+    box_width = box_height = sqrt(param.minSize * param.maxSize) / 2;
+    float4 max_box;
+    max_box.x = (center_x - box_width) / param.imageWidth;
+    max_box.y = (center_y - box_height) / param.imageHeight;
+    max_box.z = (center_x + box_width) / param.imageWidth;
+    max_box.w = (center_y + box_height) / param.imageHeight;
+    
+    float4 res;
+    if (param.clip) {
+      res = min(max(max_box, 0.0), 1.0);
+    } else {
+      res = max_box;
+    }
+    outBoxTexture.write(res, gid.xy, gid.z);
+  }
+  
+  int aspect_to = 0;
+  if (param.maxSizeSize > 0) {
+    aspect_to = gid.z - 2;
+  } else {
+    aspect_to = gid.z - 1;
+  }
+  
+
+  
+  
+  if (aspect_to >= 0 && aspect_to < int(param.aspecRatiosSize)) {
+    
+    int skip = 0;
+    for (int i = 0; i < aspect_to + 1; ++i) {
+      if (fabs(aspect_ratios[i] - 1.) < 1e-6) {
+        skip += 1;
+      }
+    }
+    aspect_to += skip;
+    
+    float ar = aspect_ratios[aspect_to];
+    
+    box_width = param.minSize * sqrt(ar) / 2;
+    box_height = param.minSize / sqrt(ar) / 2;
+    float4 box;
+    box.x = (center_x - box_width) / param.imageWidth;
+    box.y = (center_y - box_height) / param.imageHeight;
+    box.z = (center_x + box_width) / param.imageWidth;
+    box.w = (center_y + box_height) / param.imageHeight;
+    
+    float4 res;
+    if (param.clip) {
+      res = fmin(fmax(box, 0.0), 1.0);
+    } else {
+      res = box;
+    }
+    
+    outBoxTexture.write(res, gid.xy, gid.z);
+  }
+  
+  float4 variance = variances[0];
+  if (gid.z < param.numPriors) {
+    float4 variances_output;
+    variances_output.x = variance.x;
+    variances_output.y = variance.y;
+    variances_output.z = variance.z;
+    variances_output.w = variance.w;
+    varianceTexture.write(variances_output, gid.xy, gid.z);
+  }
+}
+
+
+kernel void prior_box_MinMaxAspectRatiosOrder_half(texture2d_array<half, access::read> inTexture [[texture(0)]],
+                           texture2d_array<half, access::write> outBoxTexture [[texture(1)]],
+                           texture2d_array<half, access::write> varianceTexture [[texture(2)]],
+                           const device half *aspect_ratios [[buffer(0)]],
+                           constant PriorBoxMetalParam &param [[buffer(1)]],
+                           const device float4 *variances [[buffer(2)]],
+                           uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= outBoxTexture.get_width() ||
+      gid.y >= outBoxTexture.get_height() ||
+      gid.z >= outBoxTexture.get_array_size()) return;
+  
+  float center_x = (gid.x + param.offset) * param.stepWidth;
+  float center_y = (gid.y + param.offset) * param.stepHeight;
+  
+  float box_width, box_height;
+  
+  
+  
+  if (gid.z == 0) {
+    box_width = box_height = param.minSize / 2;
+    
+    float4 box;
+    box.x = (center_x - box_width) / param.imageWidth;
+    box.y = (center_y - box_height) / param.imageHeight;
+    box.z = (center_x + box_width) / param.imageWidth;
+    box.w = (center_y + box_height) / param.imageHeight;
+    
+    float4 res;
+    if (param.clip) {
+      res = fmin(fmax(box, 0.0), 1.0);
+    } else {
+      res = box;
+    }
+    
+    outBoxTexture.write(half4(res), gid.xy, gid.z);
+  }
+  
+  if (gid.z == 1 && param.maxSizeSize > 0) {
+    
+    box_width = box_height = sqrt(param.minSize * param.maxSize) / 2;
+    float4 max_box;
+    max_box.x = (center_x - box_width) / param.imageWidth;
+    max_box.y = (center_y - box_height) / param.imageHeight;
+    max_box.z = (center_x + box_width) / param.imageWidth;
+    max_box.w = (center_y + box_height) / param.imageHeight;
+    
+    float4 res;
+    if (param.clip) {
+      res = min(max(max_box, 0.0), 1.0);
+    } else {
+      res = max_box;
+    }
+    outBoxTexture.write(half4(res), gid.xy, gid.z);
+  }
+  
+  int aspect_to = 0;
+  if (param.maxSizeSize > 0) {
+    aspect_to = gid.z - 2;
+  } else {
+    aspect_to = gid.z - 1;
+  }
+  
+  if (aspect_to > 0 && aspect_to < int(param.aspecRatiosSize) && fabs(aspect_ratios[aspect_to] - 1.) > 1e-6) {
+    float ar = aspect_ratios[aspect_to];
+    
+    box_width = param.minSize * sqrt(ar) / 2;
+    box_height = param.minSize / sqrt(ar) / 2;
+    float4 box;
+    box.x = (center_x - box_width) / param.imageWidth;
+    box.y = (center_y - box_height) / param.imageHeight;
+    box.z = (center_x + box_width) / param.imageWidth;
+    box.w = (center_y + box_height) / param.imageHeight;
+    
+    float4 res;
+    if (param.clip) {
+      res = fmin(fmax(box, 0.0), 1.0);
+    } else {
+      res = box;
+    }
+    
+    outBoxTexture.write(half4(res), gid.xy, gid.z);
+  }
+  
+  float4 variance = variances[0];
+  if (gid.z < param.numPriors) {
+    float4 variances_output;
+    variances_output.x = variance.x;
+    variances_output.y = variance.y;
+    variances_output.z = variance.z;
+    variances_output.w = variance.w;
+    varianceTexture.write(half4(variances_output), gid.xy, gid.z);
+  }
+}
+
+

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/ReluKernel.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+using namespace metal;
+
+
+kernel void relu_half(texture2d_array<half, access::sample> inTexture [[texture(0)]],
+                 texture2d_array<half, access::write> outTexture [[texture(1)]],
+                 uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) return;
+  constexpr sampler s(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const half4 input = inTexture.read(gid.xy, gid.z);
+  const float4 relu = fmax((float4)input, 0.0);
+  outTexture.write(half4(relu), gid.xy, gid.z);
+}
+
+kernel void relu(texture2d_array<float, access::sample> inTexture [[texture(0)]],
+                      texture2d_array<float, access::write> outTexture [[texture(1)]],
+                      uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) return;
+  constexpr sampler s(coord::pixel, filter::nearest, address::clamp_to_zero);
+  const float4 input = inTexture.read(gid.xy, gid.z);
+  const float4 relu = fmax((float4)input, 0.0);
+  outTexture.write(float4(relu), gid.xy, gid.z);
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/ReshapeKernel.inc.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#ifdef P
+
+#define CONCAT2(a, b) a ## b
+#define CONCAT2_(a, b) a ## _ ## b
+#define CONCAT3_(a, b, c) a ## _ ## b ## _ ## c
+#define CONCAT4_(a, b, c, d) a ## _ ## b ## _ ## c ## _ ## d
+
+#define FUNC(f, r1, r2, p) CONCAT4_(f, r1, r2, p)
+#define VECTOR(p, n) CONCAT2(p, n)
+#define FUNC_R(f, r) CONCAT2_(f, r)
+
+kernel void FUNC(reshape, RIN, ROUT, P)(texture2d_array<P, access::read> inTexture [[texture(0)]],
+                    texture2d_array<P, access::write> outTexture [[texture(1)]],
+                    constant ReshapeParam &rp [[buffer(0)]],
+                    uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) return;
+
+  int oxyzn[4] = {int(gid.x), int(gid.y), int(gid.z), 0}, oabcd[4], ixyzn[4], iabcd[4];
+  ReshapeParam lrp = rp;
+  int oC = lrp.odim[lrp.otrans[3]];
+  int iC = lrp.idim[lrp.itrans[3]];
+  int count = lrp.odim[0] * lrp.odim[1] * lrp.odim[2] * lrp.odim[3];
+  VECTOR(P, 4) r;
+  for (int n = 0; n < 4; n++) {
+    oxyzn[3] = n;
+#if ROUT == 4
+    xyzn2abcd_4(oC, oxyzn, oabcd);
+#else
+    FUNC_R(xyzn2abcd, ROUT)(oxyzn, oabcd);
+#endif
+    int tabcd[4];
+    invtrans(lrp.otrans, oabcd, tabcd);
+    int index = abcd2index(lrp.odim, tabcd);
+    if (index < count) {
+      index2abcd(lrp.idim, index, tabcd);
+      trans(lrp.itrans, tabcd, iabcd);
+#if RIN == 4
+      abcd2xyzn_4(iC, iabcd, ixyzn);
+#else
+      FUNC_R(abcd2xyzn, RIN)(iabcd, ixyzn);
+#endif
+      r[n] = inTexture.read(uint2(ixyzn[0], ixyzn[1]), ixyzn[2])[ixyzn[3]];
+    } else {
+      r[n] = 0;
+    }
+  }
+  outTexture.write(r, gid.xy, gid.z);
+}
+
+#endif

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/ReshapeKernel.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONRITIONS OF ANY KINR, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+#include "Common.metal"
+
+using namespace metal;
+
+struct ReshapeParam {
+  int32_t idim[4];
+  int32_t itrans[4];
+  int32_t odim[4];
+  int32_t otrans[4];
+};
+
+#define P float
+#define RIN 4
+#define ROUT 4
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 3
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 2
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 1
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#undef RIN
+
+#define RIN 3
+#define ROUT 4
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 3
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 2
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 1
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#undef RIN
+
+#define RIN 2
+#define ROUT 4
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 3
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 2
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 1
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#undef RIN
+
+#define RIN 1
+#define ROUT 4
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 3
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 2
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 1
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#undef RIN
+
+#undef P
+
+#define P half
+#define RIN 4
+#define ROUT 4
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 3
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 2
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 1
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#undef RIN
+
+#define RIN 3
+#define ROUT 4
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 3
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 2
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 1
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#undef RIN
+
+#define RIN 2
+#define ROUT 4
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 3
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 2
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 1
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#undef RIN
+
+#define RIN 1
+#define ROUT 4
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 3
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 2
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#define ROUT 1
+#include "ReshapeKernel.inc.metal"
+#undef ROUT
+#undef RIN
+#undef P

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/ResizeBilinear.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+using namespace metal;
+
+struct resize_bilinear_param {
+//  int32_t out_h;
+//  int32_t out_w;
+  float ratio_h;
+  float ratio_w;
+};
+
+kernel void resize_bilinear(texture2d_array<float, access::read> input [[texture(0)]],
+                     texture2d_array<float, access::write> output [[texture(2)]],
+                     constant resize_bilinear_param & pm [[buffer(0)]],
+                     uint3 gid [[thread_position_in_grid]]) {
+  float4 r;
+  if ((input.get_width() == output.get_width()) && (input.get_height() == output.get_height())) {
+    r = input.read(gid.xy, gid.z);
+  } else {
+    float w = gid.x * pm.ratio_w;
+    float h = gid.y * pm.ratio_h;
+    uint w0 = w, h0 = h;
+    uint w1 = w0 + 1, h1 = h0 + 1;
+    float w1lambda = w - w0, h1lambda = h - h0;
+    float w2lambda = 1.0 - w1lambda, h2lambda = 1.0 - h1lambda;
+    if (w1 >= input.get_width()) w1 = w0;
+    if (h1 >= input.get_height()) h1 = h0;
+    float4 r0 = input.read(uint2(w0, h0), gid.z);
+    float4 r1 = input.read(uint2(w1, h0), gid.z);
+    float4 r2 = input.read(uint2(w0, h1), gid.z);
+    float4 r3 = input.read(uint2(w1, h1), gid.z);
+    r = h2lambda * (w2lambda * r0 + w1lambda * r1) + h1lambda * (w2lambda * r2 + w1lambda * r3);
+  }
+  output.write(r, gid.xy, gid.z);
+}
+
+kernel void resize_bilinear_half(texture2d_array<half, access::read> input [[texture(0)]],
+                            texture2d_array<half, access::write> output [[texture(2)]],
+                            constant resize_bilinear_param & pm [[buffer(0)]],
+                            uint3 gid [[thread_position_in_grid]]) {
+  
+  half4 r;
+  if ((input.get_width() == output.get_width()) && (input.get_height() == output.get_height())) {
+    r = input.read(gid.xy, gid.z);
+  } else {
+    half w = gid.x * pm.ratio_w;
+    half h = gid.y * pm.ratio_h;
+    uint w0 = w, h0 = h;
+    uint w1 = w0 + 1, h1 = h0 + 1;
+    half w1lambda = w - w0, h1lambda = h - h0;
+    half w2lambda = 1.0 - w1lambda, h2lambda = 1.0 - h1lambda;
+    if (w1 >= input.get_width()) w1 = w0;
+    if (h1 >= input.get_height()) h1 = h0;
+    half4 r0 = input.read(uint2(w0, h0), gid.z);
+    half4 r1 = input.read(uint2(w1, h0), gid.z);
+    half4 r2 = input.read(uint2(w0, h1), gid.z);
+    half4 r3 = input.read(uint2(w1, h1), gid.z);
+    r = h2lambda * (w2lambda * r0 + w1lambda * r1) + h1lambda * (w2lambda * r2 + w1lambda * r3);
+  }
+  output.write(r, gid.xy, gid.z);
+  output.write(r, gid.xy, gid.z);
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/Shape.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+using namespace metal;
+
+kernel void shape() {
+}
+kernel void shape_half() {
+}

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/Softmax.inc.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#ifdef P
+
+#define CONCAT2(a, b) a ## b
+#define CONCAT2_(a, b) a ## _ ## b
+
+#define FUNC(f, p) CONCAT2_(f, p)
+#define VECTOR(p, n) CONCAT2(p, n)
+
+kernel void FUNC(softmax, P)(texture2d_array<P, access::read> inTexture [[texture(0)]],
+                    texture2d_array<P, access::write> outTexture [[texture(1)]],
+                    constant SoftmaxParam &sp [[buffer(0)]],
+                    uint3 gid [[thread_position_in_grid]]) {
+  if (gid.x >= outTexture.get_width() ||
+      gid.y >= outTexture.get_height() ||
+      gid.z >= outTexture.get_array_size()) return;
+//  int zsize = inTexture.get_array_size();
+  P maxv = inTexture.read(uint2(0, gid.y), 0)[0];
+  int group = sp.K / 4;
+  int remain = sp.K % 4;
+  for (int x = 0; x < group; x++) {
+    VECTOR(P, 4) r = inTexture.read(uint2(x, gid.y), 0);
+    maxv = max(maxv, max(r[0], max(r[1], max(r[2], r[3]))));
+  }
+  if (remain > 0) {
+    VECTOR(P, 4) r = inTexture.read(uint2(group, gid.y), 0);
+    for (int i = 0; i < remain; i++) {
+      maxv = max(maxv, r[i]);
+    }
+  }
+  VECTOR(P, 4) rsum = {0, 0, 0, 0};
+  for (int x = 0; x < group; x++) {
+    VECTOR(P, 4) r = inTexture.read(uint2(x, gid.y), 0);
+    rsum += exp(r - maxv);
+  }
+  P sum = rsum[0] + rsum[1] + rsum[2] + rsum[3];
+  if (remain > 0) {
+    VECTOR(P, 4) r = inTexture.read(uint2(group, gid.y), 0);
+    for (int i = 0; i < remain; i++) {
+      sum += exp(r[i] - maxv);
+    }
+  }
+  VECTOR(P, 4) rr = inTexture.read(gid.xy, gid.z);
+  rr = exp(rr - maxv) / sum;
+  outTexture.write(rr, gid.xy, gid.z);
+}
+
+#endif

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/Softmax.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+using namespace metal;
+
+struct SoftmaxParam {
+  int N;
+  int K;
+};
+
+#define P float
+#include "Softmax.inc.metal"
+#undef P
+
+#define P half
+#include "Softmax.inc.metal"
+#undef P

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/Split.inc.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#ifdef P
+
+#define CONCAT2(a, b) a ## b
+#define CONCAT2_(a, b) a ## _ ## b
+#define CONCAT3_(a, b, c) a ## _ ## b ## _ ## c
+#define CONCAT4_(a, b, c, d) a ## _ ## b ## _ ## c ## _ ## d
+#define CONCAT5_(a, b, c, d, e) a ## _ ## b ## _ ## c ## _ ## d ## _ ## e
+
+#define FUNC(f, r, n, v, p) CONCAT5_(f, r, n, v, p)
+#define VECTOR(p, n) CONCAT2(p, n)
+#define FUNC_R(f, r) CONCAT2_(f, r)
+
+#if V == VX
+#define VV x
+#elif V == VY
+#define VV y
+#elif V == VZ
+#define VV z
+#else
+#define VV normal
+#endif
+
+#if V == VY
+kernel void FUNC(split, R, N, VV, P)(texture2d_array<P, access::read> input [[texture(0)]],
+                                 texture2d_array<P, access::write> out1 [[texture(1)]],
+                                 texture2d_array<P, access::write> out2 [[texture(2)]],
+#if N >= 3
+                                 texture2d_array<P, access::write> out3 [[texture(3)]],
+#endif // N >= 3
+#if N >= 4
+                                 texture2d_array<P, access::write> out4 [[texture(4)]],
+#endif // N >= 4
+                                 constant SplitParam &sp [[buffer(0)]],
+                                 uint3 gid [[thread_position_in_grid]]) {
+
+  VECTOR(P, 4) r = input.read(gid.xy, gid.z);
+  int y = gid.y - sp.offset;
+  if (y < sp.vdim[0]) {
+    out1.write(r, gid.xy, gid.z);
+    return;
+  }
+  y -= sp.vdim[0];
+  if (y < sp.vdim[1]) {
+    out2.write(r, uint2(gid.x, y), gid.z);
+    return;
+  }
+#if N >= 3
+  y -= sp.vdim[1];
+  if (y < sp.vdim[2]) {
+    out3.write(r, uint2(gid.x, y), gid.z);
+    return;
+  }
+#endif // N >= 3
+#if N >= 4
+  y -= sp.vdim[2];
+  if (y < sp.vdim[3]) {
+    out4.write(r, uint2(gid.x, y), gid.z);
+    return;
+  }
+#endif // N >= 4
+}
+#endif // V == VY
+
+
+#if V == VX
+kernel void FUNC(split, R, N, VV, P)(texture2d_array<P, access::read> input [[texture(0)]],
+                                     texture2d_array<P, access::write> out1 [[texture(1)]],
+                                     texture2d_array<P, access::write> out2 [[texture(2)]],
+#if N >= 3
+                                     texture2d_array<P, access::write> out3 [[texture(3)]],
+#endif // N >= 3
+#if N >= 4
+                                     texture2d_array<P, access::write> out4 [[texture(4)]],
+#endif // N >= 4
+                                     constant SplitParam &sp [[buffer(0)]],
+                                     uint3 gid [[thread_position_in_grid]]) {
+  VECTOR(P, 4) r = input.read(gid.xy, gid.z);
+  int x = gid.x;
+  if (x < sp.vdim[0]) {
+    out1.write(r, gid.xy, gid.z);
+    return;
+  }
+  x -= sp.vdim[0];
+  if (x < sp.vdim[1]) {
+    out2.write(r, uint2(x, gid.y), gid.z);
+    return;
+  }
+#if N >= 3
+  x -= sp.vdim[1];
+  if (x < sp.vdim[2]) {
+    out3.write(r, uint2(x, gid.y), gid.z);
+    return;
+  }
+#endif // N >= 3
+#if N >= 4
+  x -= sp.vdim[2];
+  if (x < sp.vdim[3]) {
+    out4.write(r, uint2(x, gid.y), gid.z);
+    return;
+  }
+#endif // N >= 4
+}
+#endif // V == VX
+
+
+
+#undef VV
+#endif

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/Split.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+#include "Common.metal"
+
+using namespace metal;
+
+struct SplitParam {
+  int32_t idim[4];
+  int32_t axis;
+  int32_t offset;
+  int32_t trans[4];
+  int32_t vdim[4];
+};
+
+#define VNORMAL 1
+#define VX 2
+#define VY 3
+#define VZ 4
+
+// only support split_{2, 3, 4}_{2, 3, 4}_y_{float, half}
+// only support split_{3, 4}_{2, 3, 4}_x_{float, half}
+
+
+//// ssd-ar: (R=3, N=2, V=y)
+#define V VY
+  #define R 3
+    #define N 2
+      #define P float
+        #include "Split.inc.metal"
+      #undef P
+      #define P half
+        #include "Split.inc.metal"
+      #undef P
+    #undef N
+  #undef R
+#undef V
+
+
+//// ssd-ar: (R=2, N=2, V=y)
+#define V VY
+  #define R 2
+    #define N 2
+      #define P float
+        #include "Split.inc.metal"
+      #undef P
+      #define P half
+        #include "Split.inc.metal"
+      #undef P
+    #undef N
+  #undef R
+#undef V

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/TransposeKernel.inc.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#ifdef P
+
+#define CONCAT2(a, b) a ## b
+#define CONCAT2_(a, b) a ## _ ## b
+#define CONCAT3_(a, b, c) a ## _ ## b ## _ ## c
+
+#define FUNC(f, r, p) CONCAT3_(f, r, p)
+#define VECTOR(p, n) CONCAT2(p, n)
+
+kernel void FUNC(transpose, R, P)(texture2d_array<P, access::read> inTexture [[texture(0)]],
+                      texture2d_array<P, access::write> outTexture [[texture(1)]],
+                      constant TransposeParam &pm [[buffer(0)]],
+                      uint3 gid [[thread_position_in_grid]]) {
+  VECTOR(P, 4) r;
+  int oxyzn[4] = {int(gid.x), int(gid.y), int(gid.z), 0};
+  int iabcd[4], oabcd[4], ixyzn[4];
+  for (int n = 0; n < 4; n++) {
+    oxyzn[3] = n;
+#if R == 4
+    xyzn2abcd_4(pm.oC, oxyzn, iabcd);
+#endif // R == 4
+#if R == 3
+    xyzn2abcd_3(oxyzn, oabcd);
+#endif // R == 3
+#if R == 2
+    xyzn2abcd_2(oxyzn, oabcd);
+#endif // R == 2
+    iabcd[pm.axis[0]] = oabcd[0];
+    iabcd[pm.axis[1]] = oabcd[1];
+    iabcd[pm.axis[2]] = oabcd[2];
+    iabcd[pm.axis[3]] = oabcd[3];
+#if R == 4
+    abcd2xyzn_4(pm.iC, iabcd, ixyzn);
+#endif // R == 4
+#if R == 3
+    abcd2xyzn_3(iabcd, ixyzn);
+#endif // R == 3
+#if R == 2
+    abcd2xyzn_2(iabcd, ixyzn);
+#endif // R == 2
+    r[n] = inTexture.read(uint2(ixyzn[0], ixyzn[1]), ixyzn[2])[ixyzn[3]];
+  }
+  outTexture.write(r, gid.xy, gid.z);
+}
+
+#endif

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/TransposeKernel.metal

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#include <metal_stdlib>
+#include "Common.metal"
+using namespace metal;
+
+struct TransposeParam {
+  int iC;
+  int oC;
+  int axis[4];
+};
+
+kernel void transpose_copy_float(texture2d_array<float, access::read> inTexture [[texture(0)]],
+                           texture2d_array<float, access::write> outTexture [[texture(1)]],
+                           constant TransposeParam &pm [[buffer(0)]],
+                           uint3 gid [[thread_position_in_grid]]) {
+  outTexture.write(inTexture.read(gid.xy, gid.z), gid.xy, gid.z);
+}
+kernel void transpose_copy_half(texture2d_array<half, access::read> inTexture [[texture(0)]],
+                           texture2d_array<half, access::write> outTexture [[texture(1)]],
+                           constant TransposeParam &pm [[buffer(0)]],
+                           uint3 gid [[thread_position_in_grid]]) {
+  outTexture.write(inTexture.read(gid.xy, gid.z), gid.xy, gid.z);
+}
+
+#define R 4
+  #define P float
+    #include "TransposeKernel.inc.metal"
+  #undef P
+  #define P half
+    #include "TransposeKernel.inc.metal"
+  #undef P
+#undef R
+
+#define R 3
+  #define P float
+    #include "TransposeKernel.inc.metal"
+  #undef P
+  #define P half
+    #include "TransposeKernel.inc.metal"
+  #undef P
+#undef R
+
+#define R 2
+  #define P float
+    #include "TransposeKernel.inc.metal"
+  #undef P
+  #define P half
+    #include "TransposeKernel.inc.metal"
+  #undef P
+#undef R

metal/paddle-mobile/paddle-mobile/Operators/MulticlassNMSOp.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class MulticlassNMSParam<P: PrecisionType>: OpParam {
+  typealias ParamPrecisionType = P
+  required init(opDesc: OpDesc, inScope: Scope) throws {
+    do {
+      scores = try MulticlassNMSParam.getFirstTensor(key: "Scores", map: opDesc.inputs, from: inScope)
+      bboxes = try MulticlassNMSParam.getFirstTensor(key: "BBoxes", map: opDesc.inputs, from: inScope)
+      output = try MulticlassNMSParam.outputOut(outputs: opDesc.outputs, from: inScope)
+      
+      middleOutput = FetchHolder.init(inCapacity: scores.tensorDim.numel(), inDim: scores.tensorDim.dims)
+      
+      bboxOutput = FetchHolder.init(inCapacity: bboxes.tensorDim.numel(), inDim: bboxes.tensorDim.dims)
+    } catch let error {
+      throw error
+    }
+  }
+  var bboxOutput: FetchHolder
+  var middleOutput: FetchHolder
+  let scores: Texture<P>
+  let bboxes: Texture<P>
+  var output: Texture<P>
+}
+
+class MulticlassNMSOp<P: PrecisionType>: Operator<MulticlassNMSKernel<P>, MulticlassNMSParam<P>>, Runable, Creator, InferShaperable{
+
+  func inputVariant() -> [String : [Variant]] {
+    return ["Scores" : [para.middleOutput], "BBoxes" : [para.bboxOutput]]
+  }
+  
+  func computeMiddleResult(device: MTLDevice, buffer: MTLCommandBuffer) {
+    do {
+      try kernel.compute(commandBuffer: buffer, param: para)
+    } catch let _ {
+      fatalError()
+    }
+  }
+  
+  func inferShape() {
+    // para.output.dim = para.input.dim
+  }
+  
+  typealias OpType =  MulticlassNMSOp<P>
+  func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
+
+  }
+  
+  func delogOutput() {
+    print(" nms - output: ")
+    print(para.bboxes.metalTexture.float32Array().strideArray())
+  }
+}
+
+
+

metal/paddle-mobile/paddle-mobile/Operators/PoolOp.swift

@@ -26,10 +26,11 @@ class PoolParam<P: PrecisionType>: OpParam {
       padding = try PoolParam.getAttr(key: "paddings", attrs: opDesc.attrs)
       ceilMode = try PoolParam.getAttr(key: "ceil_mode", attrs: opDesc.attrs)
       globalPooling = try PoolParam.getAttr(key: "global_pooling", attrs: opDesc.attrs)
+      assert(input.transpose == [0, 2, 3, 1])
     } catch let error {
       throw error
     }
-//        let buffer = input.metalTexture.buffer.contents().assumingMemoryBound(to: P.self)
+    //        let buffer = input.metalTexture.buffer.contents().assumingMemoryBound(to: P.self)
   }
   let input: Texture<P>
   var output: Texture<P>
@@ -43,11 +44,12 @@ class PoolParam<P: PrecisionType>: OpParam {
 
 class PoolOp<P: PrecisionType>: Operator<PoolKernel<P>, PoolParam<P>>, Runable, Creator, InferShaperable{
   
+  typealias OpType = PoolOp<P>
+
   func inferShape() {
     // para.output.dim = para.input.dim
   }
   
-    typealias OpType = PoolOp<P>
   func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
     do {
       try kernel.compute(commandBuffer: buffer, param: para)
@@ -57,12 +59,16 @@ class PoolOp<P: PrecisionType>: Operator<PoolKernel<P>, PoolParam<P>>, Runable,
   }
   
   func delogOutput() {
-        print("pool2d delog")
-        let _: P? = para.input.metalTexture.logDesc(header: "pool2d input: ", stridable: true)
-        print(para.ksize)
-        print(para.stride)
-        print(para.padding)
-        print(para.poolType)
-        let _: P? = para.output.metalTexture.logDesc(header: "pool2d output: ", stridable: true)
+    print(" \(type) output: ")
+    print(para.output.metalTexture.toTensor(dim: (n: para.output.tensorDim[0], c: para.output.tensorDim[1], h: para.output.tensorDim[2], w: para.output.tensorDim[3])).strideArray())
+
+    
+//    print("pool2d delog")
+//    let _: P? = para.input.metalTexture.logDesc(header: "pool2d input: ", stridable: true)
+//    print(para.ksize)
+//    print(para.stride)
+//    print(para.padding)
+//    print(para.poolType)
+//    let _: P? = para.output.metalTexture.logDesc(header: "pool2d output: ", stridable: true)
   }
 }

metal/paddle-mobile/paddle-mobile/Operators/PreluOp.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class PreluParam<P: PrecisionType>: OpParam {
+  typealias ParamPrecisionType = P
+  required init(opDesc: OpDesc, inScope: Scope) throws {
+    do {
+      input = try PreluParam.inputX(inputs: opDesc.inputs, from: inScope)
+      output = try PreluParam.outputOut(outputs: opDesc.outputs, from: inScope)
+      alpha = try PreluParam.paramInputAlpha(inputs: opDesc.paraInputs, from: inScope)
+      mode = try PreluParam.getAttr(key: "mode", attrs: opDesc.attrs)
+    } catch let error {
+      throw error
+    }
+  }
+  let mode: String
+  let alpha: Tensor<P>
+  let input: Texture<P>
+  var output: Texture<P>
+}
+
+class PreluOp<P: PrecisionType>: Operator<PreluKernel<P>, PreluParam<P>>, Runable, Creator, InferShaperable{
+  
+  typealias OpType = PreluOp<P>
+
+  func inferShape() {
+    // para.output.dim = para.input.dim
+  }
+  
+  func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
+    do {
+      try kernel.compute(commandBuffer: buffer, param: para)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  func delogOutput() {
+    print(" \(type) input: ")
+    print(para.input.metalTexture.toTensor(dim: (n: para.input.padToFourDim[0], c: para.input.padToFourDim[1], h: para.input.padToFourDim[2], w: para.input.padToFourDim[3])).strideArray())
+    
+    print(" \(type) Alpha: ")
+    let _: Float32? = para.alpha.buffer.logDesc(header: " alpha: ", stridable: false)
+    
+    print(" \(type) output: ")
+    print(para.output.metalTexture.toTensor(dim: (n: para.output.padToFourDim[0], c: para.output.padToFourDim[1], h: para.output.padToFourDim[2], w: para.output.padToFourDim[3])).strideArray())
+  }
+  
+//    print("softmax delog")
+//    let _: P? = para.input.metalTexture.logDesc(header: "softmax input: ", stridable: false)
+//    let _: P? = para.output.metalTexture.logDesc(header: "softmax output: ", stridable: false)
+}

metal/paddle-mobile/paddle-mobile/Operators/PriorBoxOp.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class PriorBoxParam<P: PrecisionType>: OpParam {
+  typealias ParamPrecisionType = P
+  required init(opDesc: OpDesc, inScope: Scope) throws {
+    do {
+      min_max_aspect_ratios_order = try PriorBoxParam.getAttr(key: "min_max_aspect_ratios_order", attrs: opDesc.attrs)
+    } catch _ {
+    }
+    
+    do {
+      input = try PriorBoxParam.input(inputs: opDesc.inputs, from: inScope)
+      output = try PriorBoxParam.outputBoxes(outputs: opDesc.outputs, from: inScope)
+      inputImage = try PriorBoxParam.inputImage(inputs: opDesc.inputs, from: inScope)
+      outputVariances = try PriorBoxParam.outputVariances(outputs: opDesc.outputs, from: inScope)
+      minSizes = try PriorBoxParam.getAttr(key: "min_sizes", attrs: opDesc.attrs)
+      maxSizes = try PriorBoxParam.getAttr(key: "max_sizes", attrs: opDesc.attrs)
+      aspectRatios = try PriorBoxParam.getAttr(key: "aspect_ratios", attrs: opDesc.attrs)
+      variances = try PriorBoxParam.getAttr(key: "variances", attrs: opDesc.attrs)
+      flip = try PriorBoxParam.getAttr(key: "flip", attrs: opDesc.attrs)
+      clip = try PriorBoxParam.getAttr(key: "clip", attrs: opDesc.attrs)
+      stepW = try PriorBoxParam.getAttr(key: "step_w", attrs: opDesc.attrs)
+      stepH = try PriorBoxParam.getAttr(key: "step_h", attrs: opDesc.attrs)
+      offset = try PriorBoxParam.getAttr(key: "offset", attrs: opDesc.attrs)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  var min_max_aspect_ratios_order: Bool = false
+  let minSizes: [Float32]
+  let maxSizes: [Float32]
+  let aspectRatios: [Float32]
+  var newAspectRatios: MTLBuffer?
+  let variances: [Float32]
+  let flip: Bool
+  let clip: Bool
+  var stepW: Float32
+  var stepH: Float32
+  let offset: Float32
+  
+  let input: Texture<P>
+  let inputImage: Texture<P>
+  var output: Texture<P>
+  let outputVariances: Texture<P>
+}
+
+class PriorBoxOp<P: PrecisionType>: Operator<PriorBoxKernel<P>, PriorBoxParam<P>>, Runable, Creator, InferShaperable{
+  
+  typealias OpType = PriorBoxOp<P>
+
+  func inferShape() {
+  }
+  
+  func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
+    do {
+      try kernel.compute(commandBuffer: buffer, param: para)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  func delogOutput() {
+
+    print(" \(type) output: ")
+    // output
+//    let outputArray = para.output.metalTexture.float32Array()
+//    print(outputArray.strideArray())
+//    let device = para.input.metalTexture!.device
+//    let boxes:[Float32] = device.texture2tensor(texture: para.output.metalTexture!, dim: para.output.tensorDim.dims, transpose: [2,0,1,3])
+//    let variances:[Float32] = device.texture2tensor(texture: para.outputVariances.metalTexture!, dim: para.outputVariances.tensorDim.dims, transpose: [2,0,1,3])
+//    print("boxes: ")
+//    print(boxes.strideArray())
+//    print("variances: ")
+//    print(variances.strideArray())
+    // output
+    print(" \(type) output: ")
+    
+    let box = para.output.metalTexture.realNHWC(dim: (para.output.dim[0], para.output.dim[1], para.output.dim[2], para.output.dim[3]))
+    print(" dim: \(para.output.dim)")
+    print(box.strideArray())
+//    print((0..<box.count).map { (index: $0, value: box[$0])})
+//    print(para.output.realNHWC().strideArray())
+    
+//    let padToFourDim = para.output.padToFourDim
+//    if para.output.transpose == [0, 1, 2, 3] {
+//      let outputArray: [Float32] = para.output.metalTexture.realNHWC(dim: (n: padToFourDim[0], h: padToFourDim[1], w: padToFourDim[2], c: padToFourDim[3]), texturePrecision: computePrecision)
+//      print(outputArray.strideArray())
+//    } else if para.output.transpose == [0, 2, 3, 1] {
+//      print(para.output.metalTexture.toTensor(dim: (n: padToFourDim[0], c: padToFourDim[1], h: padToFourDim[2], w: padToFourDim[3]), texturePrecision: computePrecision).strideArray())
+//    } else {
+//      print(" not implement")
+//    }
+    
+//    writeToLibrary(fileName: "box_out", array: outputArray)
+    
+    // output variance
+//    let outputVarianceArray = para.outputVariances.metalTexture.floatArray { (o: Float32) -> Float32 in
+//      return o
+//    }
+//
+//    print(" output variance: \(outputVarianceArray)")
+    
+//    writeToLibrary(fileName: "variance_out", array: outputVarianceArray)
+    
+  }
+}
+
+
+

metal/paddle-mobile/paddle-mobile/Operators/ReluOp.swift

-///* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
-// 
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-// 
-// http://www.apache.org/licenses/LICENSE-2.0
-// 
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License. */
+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
 
 import Foundation
 
@@ -30,11 +31,12 @@ class ReluParam<P: PrecisionType>: OpParam {
 
 class ReluOp<P: PrecisionType>: Operator<ReluKernel<P>, ReluParam<P>>, Runable, Creator, InferShaperable{
   
+  typealias OpType = ReluOp<P>
+  
   func inferShape() {
     para.output.dim = para.input.dim
   }
   
-    typealias OpType = ReluOp<P>
   func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
     do {
       try kernel.compute(commandBuffer: buffer, param: para)
@@ -42,6 +44,15 @@ class ReluOp<P: PrecisionType>: Operator<ReluKernel<P>, ReluParam<P>>, Runable,
       throw error
     }
   }
+  
+  func delogOutput() {
+    print(" \(type) output: ")
+    print(para.output.metalTexture.toTensor(dim: (n: para.output.tensorDim[0], c: para.output.tensorDim[1], h: para.output.tensorDim[2], w: para.output.tensorDim[3])).strideArray())
+    let device = para.output.metalTexture!.device
+    let outputArray: [Float32] = device.texture2tensor(texture: para.output.metalTexture, dim: para.output.tensorDim.dims, transpose: para.output.transpose)
+    print(outputArray.strideArray())
+  }
+  
 }
 
 

metal/paddle-mobile/paddle-mobile/Operators/ReshapeOp.swift

@@ -20,21 +20,46 @@ class ReshapeParam<P: PrecisionType>: OpParam {
     do {
       input = try ReshapeParam.inputX(inputs: opDesc.inputs, from: inScope)
       output = try ReshapeParam.outputOut(outputs: opDesc.outputs, from: inScope)
+      shape = try ReshapeParam.getAttr(key: "shape", attrs: opDesc.attrs)
+        
+      var s: [Int] = shape.map { Int($0) }
+      
+      var di = -1
+      var ml = 1
+      for i in 0..<s.count {
+        if s[i] == -1 {
+          di = i
+          continue
+        }
+        ml *= s[i]
+      }
+      if di >= 0 {
+        s[di] = input.dim.numel() / ml
+      }
+      output.tensorDim = Dim.init(inDim: s)
+      var dim: [Int] = [1, 1, 1, 1]
+      for i in 0..<s.count {
+        dim[4-s.count+i] = s[i]
+      }
+      output.padToFourDim = Dim.init(inDim: dim)
+      output.dim = output.padToFourDim
     } catch let error {
       throw error
     }
   }
   let input: Texture<P>
+  let shape: [Int32]
   var output: Texture<P>
 }
 
 class ReshapeOp<P: PrecisionType>: Operator<ReshapeKernel<P>, ReshapeParam<P>>, Runable, Creator, InferShaperable{
   
+  typealias OpType = ReshapeOp<P>
+
   func inferShape() {
     // para.output.dim = para.input.dim
   }
   
-    typealias OpType = ReshapeOp<P>
   func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
     do {
       try kernel.compute(commandBuffer: buffer, param: para)
@@ -44,7 +69,9 @@ class ReshapeOp<P: PrecisionType>: Operator<ReshapeKernel<P>, ReshapeParam<P>>,
   }
   func delogOutput() {
     print("reshape delog")
-        let _: P? = para.input.metalTexture.logDesc(header: "reshape input: ", stridable: false)
-        let _: P? = para.output.metalTexture.logDesc(header: "reshape output: ", stridable: false)
+    let device = para.output.metalTexture!.device
+    let outputArray: [Float32] = device.texture2tensor(texture: para.output.metalTexture, dim: para.output.tensorDim.dims, transpose: para.output.transpose)
+    print(outputArray.strideArray())
+//    print(outputArray)
   }
 }

metal/paddle-mobile/paddle-mobile/Operators/ResizeBilinearOp.swift

+///* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License. */
+
+import Foundation
+
+class ResizeBilinearParam<P: PrecisionType>: OpParam {
+  typealias ParamPrecisionType = P
+  required init(opDesc: OpDesc, inScope: Scope) throws {
+    do {
+      input = try ResizeBilinearParam.inputX(inputs: opDesc.inputs, from: inScope)
+//      if (input.transpose != [0, 2, 3, 1]) || (input.tensorDim.cout() != 4) {
+//        fatalError()
+//      }
+      output = try ResizeBilinearParam.outputOut(outputs: opDesc.outputs, from: inScope)
+      out_h = try ResizeBilinearParam.getAttr(key: "out_h", attrs: opDesc.attrs)
+      out_w = try ResizeBilinearParam.getAttr(key: "out_w", attrs: opDesc.attrs)
+    } catch let error {
+      throw error
+    }
+  }
+  let input: Texture<P>
+  var output: Texture<P>
+  let out_h: Int32
+  let out_w: Int32
+}
+
+class ResizeBilinearOp<P: PrecisionType>: Operator<ResizeBilinearKernel<P>, ResizeBilinearParam<P>>, Runable, Creator, InferShaperable{
+  
+  typealias OpType = ResizeBilinearOp<P>
+
+  func inferShape() {
+    //        para.output.dim = para.input.dim
+  }
+  
+  func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
+    do {
+      try kernel.compute(commandBuffer: buffer, param: para)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  func delogOutput() {
+    print(" \(type) output: ")
+  }
+  
+}
+
+
+
+
+
+

metal/paddle-mobile/paddle-mobile/Operators/ShapeOp.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class ShapeParam<P: PrecisionType>: OpParam {
+  typealias ParamPrecisionType = P
+  required init(opDesc: OpDesc, inScope: Scope) throws {
+    do {
+      input = try ShapeParam.input(inputs: opDesc.inputs, from: inScope)
+      output = try ShapeParam.outputOut(outputs: opDesc.outputs, from: inScope)
+    } catch let error {
+      throw error
+    }
+  }
+  var output: Texture<P>
+  let input: Texture<P>
+}
+
+class ShapeOp<P: PrecisionType>: Operator<ShapeKernel<P>, ShapeParam<P>>, Runable, Creator, InferShaperable{
+  
+  typealias OpType = ShapeOp<P>
+
+  func inferShape() {
+    //        para.output.dim = para.input.dim
+  }
+  
+  func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
+    do {
+      try kernel.compute(commandBuffer: buffer, param: para)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  func delogOutput() {
+    print(" \(type) output: ")
+  }
+  
+}
+
+
+
+
+
+

metal/paddle-mobile/paddle-mobile/Operators/SoftmaxOp.swift

@@ -20,6 +20,13 @@ class SoftmaxParam<P: PrecisionType>: OpParam {
     do {
       input = try SoftmaxParam.inputX(inputs: opDesc.inputs, from: inScope)
       output = try SoftmaxParam.outputOut(outputs: opDesc.outputs, from: inScope)
+      
+      assert(input.tensorDim.dims.count == 2)
+      assert(input.transpose == [0, 1, 2, 3])
+      
+      output.dim = input.dim
+      output.tensorDim = input.tensorDim
+      output.padToFourDim = input.padToFourDim
     } catch let error {
       throw error
     }
@@ -29,12 +36,12 @@ class SoftmaxParam<P: PrecisionType>: OpParam {
 }
 
 class SoftmaxOp<P: PrecisionType>: Operator<SoftmaxKernel<P>, SoftmaxParam<P>>, Runable, Creator, InferShaperable{
+  typealias OpType = SoftmaxOp<P>
 
   func inferShape() {
     // para.output.dim = para.input.dim
   }
   
-    typealias OpType = SoftmaxOp<P>
   func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
     do {
       try kernel.compute(commandBuffer: buffer, param: para)
@@ -42,9 +49,14 @@ class SoftmaxOp<P: PrecisionType>: Operator<SoftmaxKernel<P>, SoftmaxParam<P>>,
       throw error
     }
   }
+  
   func delogOutput() {
     print("softmax delog")
-        let _: P? = para.input.metalTexture.logDesc(header: "softmax input: ", stridable: false)
-        let _: P? = para.output.metalTexture.logDesc(header: "softmax output: ", stridable: false)
+    print(para.input)
+    
+    print(para.output)
+    let padToFourDim = para.output.padToFourDim
+    let outputArray: [Float32] = para.output.metalTexture.realNHWC(dim: (n: padToFourDim[0], h: padToFourDim[1], w: padToFourDim[2], c: padToFourDim[3]))
+    print(outputArray.strideArray())
   }
 }

metal/paddle-mobile/paddle-mobile/Operators/SplitOp.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class SplitParam<P: PrecisionType>: OpParam {
+  typealias ParamPrecisionType = P
+  required init(opDesc: OpDesc, inScope: Scope) throws {
+    do {
+      input = try SplitParam.inputX(inputs: opDesc.inputs, from: inScope)
+      output = Texture<P>.init(device: input.metalTexture!.device, inDim: input.dim)
+      axis = try SplitParam.getAttr(key: "axis", attrs: opDesc.attrs)
+      sections = try SplitParam.getAttr(key: "sections", attrs: opDesc.attrs)
+      if axis < 0 {
+        axis = input.tensorDim.cout() + axis
+      }
+      guard let outlist = opDesc.outputs["Out"] else {
+        fatalError()
+      }
+      for out in outlist {
+        guard let variant = inScope[out], let v = variant as? Texture<P> else {
+          fatalError()
+        }
+        outputList.append(v)
+        sections.append(Int32(v.tensorDim.dims[axis]))
+      }
+    } catch let error {
+      throw error
+    }
+  }
+  
+  var axis: Int
+  let input: Texture<P>
+  var output: Texture<P>
+  var outputList: [Texture<P>] = []
+  var sections: [Int32] = []
+}
+
+class SplitOp<P: PrecisionType>: Operator<SplitKernel<P>, SplitParam<P>>, Runable, Creator, InferShaperable{
+  
+  typealias OpType = SplitOp<P>
+
+  func inferShape() {
+    //        para.output.dim = para.input.dim
+  }
+  
+  func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
+    do {
+      try kernel.compute(commandBuffer: buffer, param: para)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  func delogOutput() {
+    print(" \(type) output: ")
+    let device = para.input.metalTexture!.device
+    for out in para.outputList {
+      let arr: [Float32] = device.texture2tensor(texture: out.metalTexture, dim: out.tensorDim.dims, transpose: out.transpose)
+      print(arr.strideArray())
+    }
+  }
+  
+}
+
+
+
+
+
+

metal/paddle-mobile/paddle-mobile/Operators/TransposeOp.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+class TransposeParam<P: PrecisionType>: OpParam {
+  typealias ParamPrecisionType = P
+  required init(opDesc: OpDesc, inScope: Scope) throws {
+    do {
+      input = try TransposeParam.inputX(inputs: opDesc.inputs, from: inScope)
+      output = try TransposeParam.outputOut(outputs: opDesc.outputs, from: inScope)
+      axis = try TransposeParam.getAttr(key: "axis", attrs: opDesc.attrs)
+    } catch let error {
+      throw error
+    }
+  }
+  let input: Texture<P>
+  var output: Texture<P>
+  let axis: [Int32]
+}
+
+class TransposeOp<P: PrecisionType>: Operator<TransposeKernel<P>, TransposeParam<P>>, Runable, Creator, InferShaperable{
+  
+  typealias OpType = TransposeOp<P>
+
+  func inferShape() {
+    //para.output.dim = para.input.dim
+  }
+  
+  func runImpl(device: MTLDevice, buffer: MTLCommandBuffer) throws {
+    do {
+      try kernel.compute(commandBuffer: buffer, param: para)
+    } catch let error {
+      throw error
+    }
+  }
+  
+  func delogOutput() {
+    print(" \(type) output: ")
+    let device = para.output.metalTexture!.device
+    let outputArray: [Float32] = device.texture2tensor(texture: para.output.metalTexture, dim: para.output.tensorDim.dims, transpose: para.output.transpose)
+    print(outputArray.strideArray())
+  }
+}
+
+
+

metal/paddle-mobile/paddle-mobile/PaddleMobileGPU.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#import <Metal/Metal.h>
+#import <Foundation/Foundation.h>
+
+typedef enum : NSUInteger {
+  MobileNetType,
+  MobileNetSSDType,
+  GenetType,
+} NetType;
+
+@interface PaddleMobileGPUResult: NSObject
+
+@property (assign, nonatomic) float *output;
+
+@property (assign, nonatomic) int outputSize;
+
+-(void)releaseOutput;
+
+@end
+
+@interface ModelConfig: NSObject
+
+/*
+ * 预处理需要用到的值 (三个)
+ */
+@property (strong, nonatomic) NSArray<NSNumber *> *means;
+/*
+ * 预处理需要用到的 scale 值
+ */
+@property (assign, nonatomic) float scale;
+
+/*
+ * 输出维度信息  [n c h w]
+ */
+@property (strong, nonatomic) NSArray<NSNumber *> *dims;
+
+
+/*
+ * 模型参数内存地址
+ */
+@property (assign, nonatomic) void *paramPointer;
+
+/*
+ * 模型参数占用内存大小 (kb)
+ */
+@property (assign, nonatomic) int paramSize;
+
+/*
+ * 模型内存地址
+ */
+@property (assign, nonatomic) void *modelPointer;
+
+/*
+ * 模型占用内存大小 (kb)
+ */
+@property (assign, nonatomic) int modelSize;
+
+@end
+
+@interface PaddleMobileGPU: NSObject
+
+/*
+ * 初始化
+ */
+-(instancetype)initWithCommandQueue:(id<MTLCommandQueue>)queue net:(NetType)netType modelConfig:(ModelConfig *)config;
+
+/*
+ * paramPointer 模型参数内存地址
+ * paramSize    模型参数占用内存大小 (kb)
+ * modelPointer 模型内存地址
+ * modelSize    模型占用内存大小 (kb)
+ */
+-(BOOL)load;
+
+/*
+ * texture:     需要进行预测的图像转换的 texture
+ * completion:  预测完成回调
+ */
+-(void)predict:(id<MTLTexture>)texture withCompletion:(void (^)(BOOL, NSArray<NSNumber *> *))completion;
+
+/*
+ * texture:     需要进行预测的图像转换的 texture
+ * completion:  预测完成回调
+ */
+-(void)predict:(id<MTLTexture>)texture withResultCompletion:(void (^)(BOOL, PaddleMobileGPUResult *))completion;
+
+/*
+ * 清理内存
+ */
+-(void)clear;
+
+@end
+
+

metal/paddle-mobile/paddle-mobile/PaddleMobileGPU.m

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#import <Foundation/Foundation.h>
+#import "PaddleMobileGPU.h"
+#import "paddle_mobile.h"
+#import <paddle_mobile/paddle_mobile-Swift.h>
+
+@implementation ModelConfig
+@end
+
+@interface PaddleMobileGPUResult ()
+
+@property (strong, nonatomic) ResultHolder *resultHolder;
+
+- (void)setOutputResult:(ResultHolder *)resultHolder;
+
+@end
+
+@implementation PaddleMobileGPUResult
+- (void)setOutputResult:(ResultHolder *)resultHolder {
+  self.resultHolder = resultHolder;
+  self.output = resultHolder.result;
+  self.outputSize = resultHolder.capacity;
+}
+
+-(void)releaseOutput {
+  [self.resultHolder releasePointer];
+}
+@end
+
+@interface PaddleMobileGPU ()
+{
+  Runner *runner;
+}
+@end
+
+@implementation PaddleMobileGPU
+
+-(instancetype)initWithCommandQueue:(id<MTLCommandQueue>)queue net:(NetType)netType modelConfig:(ModelConfig *)config {
+  self = [super init];
+  if (self) {
+    Net *net = nil;
+    if (netType == GenetType) {
+      net = [[Genet alloc] initWithDevice:queue.device paramPointer:config.paramPointer paramSize:config.paramSize modePointer:config.modelPointer modelSize:config.modelSize];
+    } else if (netType == MobileNetSSDType) {
+      net = [[MobileNet_ssd_AR alloc] initWithDevice:queue.device paramPointer:config.paramPointer paramSize:config.paramSize modePointer:config.modelPointer modelSize:config.modelSize];
+    } else if (netType == MobileNetType) {
+      
+    }
+    runner = [[Runner alloc] initInNet:net commandQueue:queue inPlatform:PlatformGPU];
+  }
+  return self;
+}
+
+-(BOOL)load {
+  return [runner load];
+}
+
+-(void)predict:(id<MTLTexture>)texture withCompletion:(void (^)(BOOL, NSArray<NSNumber *> *))completion {
+  [runner predictWithTexture:texture completion:^(BOOL success, ResultHolder * _Nullable result) {
+    NSMutableArray<NSNumber *> *resultArray = [NSMutableArray arrayWithCapacity:result.capacity];
+    for (int i = 0; i < result.capacity; ++i) {
+      [resultArray addObject:[NSNumber numberWithFloat:result.result[i]]];
+    }
+    completion(success, resultArray);
+    [result releasePointer];
+    
+  }];
+}
+
+-(void)predict:(id<MTLTexture>)texture withResultCompletion:(void (^)(BOOL, PaddleMobileGPUResult *))completion {
+  [runner predictWithTexture:texture completion:^(BOOL success, ResultHolder * _Nullable result) {
+    PaddleMobileGPUResult *gpuResult = [[PaddleMobileGPUResult alloc] init];
+    [gpuResult setOutputResult:result];
+    completion(success, gpuResult);
+  }];
+}
+
+-(void)clear {
+  [runner clear];
+}
+
+@end

metal/paddle-mobile/paddle-mobile/Program/BlockDesc.swift

@@ -14,7 +14,7 @@
 
 import Foundation
 
-struct BlockDesc {
+class BlockDesc {
     let index: Int
     let parentIndex: Int
     let vars: [VarDesc]
@@ -48,7 +48,9 @@ extension BlockDesc: CustomStringConvertible, CustomDebugStringConvertible {
     var description: String {
         var str = ""
         
-        for op in ops {
+        for i in 0..<ops.count {
+          str += " op \(i): "
+          let op = ops[i]
           str += op.description
         }
         

metal/paddle-mobile/paddle-mobile/Program/OpDesc.swift

@@ -14,7 +14,7 @@
 
 import Foundation
 
-struct OpDesc {
+class OpDesc {
   let inputs: [String : [String]]
   var paraInputs: [String : [String]]
   var outputs: [String : [String]]

metal/paddle-mobile/paddle-mobile/Program/Program.swift

@@ -14,7 +14,7 @@
 
 import Foundation
 
-public struct Program {
+public class Program {
     let paramPath: String
     let programDesc: ProgramDesc
     let scope: Scope
@@ -23,4 +23,9 @@ public struct Program {
         paramPath = inParamPath
         scope = inScope
     }
+    init(inProgramDesc: ProgramDesc, inScope: Scope) {
+        programDesc = inProgramDesc
+        scope = inScope
+        paramPath = ""
+    }
 }

metal/paddle-mobile/paddle-mobile/Program/ProgramDesc.swift

@@ -14,7 +14,7 @@
 
 import Foundation
 
-public struct ProgramDesc {
+public class ProgramDesc {
     var blocks: [BlockDesc] = []
     init(protoProgram: PaddleMobile_Framework_Proto_ProgramDesc) {
         for block in protoProgram.blocks {

metal/paddle-mobile/paddle-mobile/Program/ProgramOptimize.swift

@@ -35,6 +35,22 @@ class Node {
     type = inType
   }
   
+  subscript(index: Int) -> [Node] {
+    var nodes: [Node] = []
+    getNodesWithLocation(index: index, nowIndex: 0, nodes: &nodes)
+    return nodes
+  }
+  
+  func getNodesWithLocation(index: Int, nowIndex: Int, nodes: inout [Node]) {
+    if index == nowIndex {
+      nodes.append(self)
+    }
+    
+    for output in outputs {
+      output.getNodesWithLocation(index: index, nowIndex: nowIndex + 1, nodes: &nodes)
+    }
+  }
+  
   static func -->(lNode: Node, rNode: Node) -> Node {
     lNode.outputs.append(rNode)
     rNode.inputs.append(lNode)
@@ -53,6 +69,7 @@ class Node {
   
   func to(depth: UInt) -> Node {
     let beginNode = Node.init(inType: type)
+    beginNode.opDesc = opDesc
     to(depth: depth - 1, withNode: beginNode)
     return beginNode
   }
@@ -77,7 +94,7 @@ class Node {
     
     for attr in inOpdesc.attrs {
       beginNode.opDesc?.attrs[attr.key] = attr.value
-//            print(beginNode.opDesc?.attrs)
+      //            print(beginNode.opDesc?.attrs)
     }
     
     for paraInput in inOpdesc.paraInputs {
@@ -114,11 +131,33 @@ class Node {
     
     for output in outputs {
       let node = Node.init(inType: output.type)
+      node.opDesc = output.opDesc
       withNode.outputs.append(node)
       output.to(depth: depth - 1, withNode: node)
     }
   }
   
+  func relationship() -> [String : Node]{
+    var map: [String : Node] = [:]
+    relationship(map: &map)
+    return map
+  }
+  
+  private func relationship(map: inout [String : Node]) {
+    guard let inOpDesc = opDesc else {
+      return
+    }
+    
+    for output in inOpDesc.outputs {
+      for outputKey in output.value {
+        map[outputKey] = self
+      }
+    }
+    
+    for output in outputs {
+      output.relationship(map: &map)
+    }
+  }
   
 }
 
@@ -143,7 +182,16 @@ extension Node: Equatable {
 }
 
 class ProgramOptimize<P: PrecisionType> {
-    let fusionOps: [Fusion.Type] = [ConvAddBatchNormReluOp<P>.self, ConvAddOp<P>.self]
+  // register fusion
+  let fusionOps: [Fusion.Type] = [ConvAddBatchNormReluOp<P>.self,
+//                                  ConvAddAddPreluOp<P>.self,
+                                  ConvAddPreluOp<P>.self,
+                                  ConvAddOp<P>.self,
+                                  ConvBNReluOp<P>.self,
+                                  DwConvBNReluOp<P>.self,
+                                  ElementwiseAddPreluOp<P>.self
+  ]
+  
   func optimize(originProgramDesc: ProgramDesc) -> ProgramDesc {
     
     guard originProgramDesc.blocks.count == 1 else {
@@ -152,7 +200,7 @@ class ProgramOptimize<P: PrecisionType> {
     
     var mapForNodeChain: [String : Node] = [:]
     var nodes: [Node] = []
-        var typeMapNodes: [String : [Node]] = [:]
+    var typeMapNodes: [String : [(node: Node, output: [String : Node])]] = [:]
     let block = originProgramDesc.blocks[0]
     for opDesc in block.ops {
       guard let opInputKeys = opInfos[opDesc.type]?.inputs, let outputKeys = opInfos[opDesc.type]?.outputs else {
@@ -181,10 +229,10 @@ class ProgramOptimize<P: PrecisionType> {
       nodes.append(node)
       
       if var inNodes = typeMapNodes[opDesc.type] {
-                    inNodes.append(node)
+        inNodes.append((node, mapForNodeChain))
         typeMapNodes[opDesc.type] = inNodes
       } else {
-                    typeMapNodes[opDesc.type] = [node]
+        typeMapNodes[opDesc.type] = [(node, mapForNodeChain)]
       }
     }
     
@@ -193,10 +241,43 @@ class ProgramOptimize<P: PrecisionType> {
       let depth = fusionNode.depth()
       if let toMatchNodes = typeMapNodes[fusionNode.type] {
         for node in toMatchNodes {
-                        let toNode = node.to(depth: depth)
+          
+          let toNode = node.node.to(depth: depth)
           if toNode == fusionNode {   // match
+            var canFolder = true
+            let relationshipMap = toNode.relationship()
+            
+            for toCheck in fusion.needCheck() {
+              //              let nodes = toCheck
+              let checkNodes = toNode[toCheck.0]
+              
+              for checkNode in checkNodes {
+                let inputToChecks = checkNode.opDesc?.inputs[toCheck.1] ?? []
+                for inputToCheck in inputToChecks {
+                  if node.output[inputToCheck] == nil {
+                    if relationshipMap[inputToCheck] == nil {
+                      canFolder = false
+                    }
+                  }
+                }
+                
+                let paramInputToChecks = checkNode.opDesc?.paraInputs[toCheck.1] ?? []
+                for paramInputToCheck in paramInputToChecks {
+                  if node.output[paramInputToCheck] == nil {
+                    if relationshipMap[paramInputToCheck] == nil {
+                      canFolder = false
+                    }
+                  }
+                }
+              }
+            }
+            
+            if !canFolder {
+              continue
+            }
+            
             var removeNodes: [Node] = []
-                            node.folderWith(fusion: fusion, removedNodes: &removeNodes)
+            node.node.folderWith(fusion: fusion, removedNodes: &removeNodes)
             for removeNode in removeNodes {
               nodes.remove(element: removeNode)
             }

metal/paddle-mobile/paddle-mobile/Program/TensorDesc.swift

@@ -14,18 +14,18 @@
 
 import Foundation
 
-struct TensorDesc {
+class TensorDesc {
     let dims: [Int]
     let dataType: VarTypeType
-    let dataLayout: DataLayout = .NCHW
+    let dataLayout: DataLayout = DataLayout.NCHW()
     var NCHWDim: [Int] {
         get {
             if dims.count != 4 {
                 return dims
             }
-            if dataLayout == .NCHW {
+            if dataLayout == DataLayout.NCHW() {
                 return dims
-            } else if dataLayout == .NHWC{
+            } else if dataLayout == DataLayout.NHWC() {
                 var resultDims = dims
                 resultDims.swapAt(1, 3)
                 return resultDims
@@ -40,9 +40,9 @@ struct TensorDesc {
             if dims.count != 4 {
                 return dims
             }
-            if dataLayout == .NHWC {
+            if dataLayout == DataLayout.NHWC() {
                 return dims
-            } else if dataLayout == .NCHW{
+            } else if dataLayout == DataLayout.NCHW() {
                 var resultDims = dims
                 resultDims.swapAt(1, 3)
                 return resultDims
@@ -53,7 +53,7 @@ struct TensorDesc {
     }
     
     init(protoTensorDesc: PaddleMobile_Framework_Proto_VarType.TensorDesc) {
-        dims = protoTensorDesc.dims.map{ Int($0) > 0 ? Int($0) : 1 }
+        dims = protoTensorDesc.dims.map{ Int($0) > 0 ? Int($0) : abs(Int($0)) }
         dataType = VarTypeType.init(rawValue: protoTensorDesc.dataType.rawValue) ?? .ErrorType
     }
     

metal/paddle-mobile/paddle-mobile/Program/VarDesc.swift

@@ -56,7 +56,7 @@ enum VarTypeType: Int {
     }
 }
 
-struct VarDesc {
+class VarDesc {
     let name: String
     let persistable: Bool
     let type: VarTypeType

metal/paddle-mobile/paddle-mobile/framework/Dim.swift

@@ -31,15 +31,14 @@ public struct Dim {
         return dims.reduce(1) { $0 * $1 }
     }
     
-    static func ==(left: Dim, right: Dim) -> Bool {
+    public static func ==(left: Dim, right: Dim) -> Bool {
         return left.dims == right.dims;
     }
     
-    subscript(index: Int) -> Int {
+    public subscript(index: Int) -> Int {
         return dims[index];
     }
     
-    
     private(set) var dims: [Int]
     private init(){
         fatalError()

metal/paddle-mobile/paddle-mobile/framework/Executor.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+
+
+let testTo = 81
+
+var isTest = false
+
+let computePrecision: ComputePrecision = .Float16
+
+public class GPUResultHolder {
+  public let dim: [Int]
+  public let capacity: Int
+  public var resultPointer: UnsafeMutablePointer<Float32>?
+  public var intermediateResults: [String : [Variant]]?
+  public let elapsedTime: Double
+  public init(inDim: [Int], inPointer: UnsafeMutablePointer<Float32>?, inCapacity: Int, inElapsedTime: Double, inIntermediateResults: [String : [Variant]]? = nil) {
+    dim = inDim
+    capacity = inCapacity
+    
+    if let inInPointer = inPointer {
+      resultPointer = UnsafeMutablePointer<Float32>.allocate(capacity: inCapacity)
+      resultPointer?.initialize(from: inInPointer, count: inCapacity)
+    }
+    
+    elapsedTime = inElapsedTime
+    intermediateResults = inIntermediateResults
+  }
+  
+}
+
+extension GPUResultHolder: CustomDebugStringConvertible, CustomStringConvertible {
+  public var debugDescription: String {
+//    var str = ""
+//    str += "Dim: \(dim) \n value:[ "
+//    if resultArr.count < 20 {
+//      for d in resultArr {
+//        str += " \(d) "
+//      }
+//    } else {
+//      for d in stride(from: 0, to: resultArr.count, by: resultArr.count/20) {
+//        str += " \(resultArr[d]) "
+//      }
+//    }
+//    str += " ]"
+//    return str
+    fatalError()
+  }
+  
+  public var description: String {
+    return debugDescription
+  }
+}
+
+public class Executor<P: PrecisionType> {
+  var ops: [Runable & InferShaperable] = []
+  let program: Program
+  let device: MTLDevice
+  let inflightSemaphore: DispatchSemaphore
+  let queue: MTLCommandQueue
+  public init(inDevice:MTLDevice, inQueue: MTLCommandQueue, inProgram: Program) throws {
+    self.inflightSemaphore = DispatchSemaphore(value: 3)
+    program = inProgram
+    device = inDevice
+    queue = inQueue
+//    print("before for ")
+//print(program.scope.vars["fea_pyramid1_mbox_conf_flat.Flatten.output.1.tmp_0"])
+    
+    
+    for block in inProgram.programDesc.blocks {
+      //block.ops.count
+      for i in 0..<block.ops.count {
+        let opDesc = block.ops[i]
+        do {
+//          print("in for i \(i): ")
+//      print(program.scope.vars["fea_pyramid1_mbox_conf_flat.Flatten.output.1.tmp_0"])
+//
+//          if i == 56 {
+//          print(program.scope.vars["fea_pyramid1_mbox_conf_flat.Flatten.output.1.tmp_0"])
+//
+//          }
+          
+          let op = try OpCreator<P>.shared.creat(device: inDevice, opDesc: opDesc, scope: inProgram.scope)
+          ops.append(op)
+        } catch let error {
+          throw error
+        }
+      }
+    }
+  }
+  
+  public func predict(input: MTLTexture, dim: [Int], completionHandle: @escaping (GPUResultHolder) -> Void, preProcessKernle: CusomKernel? = nil, except: Int = 0) throws {
+    guard let buffer = queue.makeCommandBuffer() else {
+      throw PaddleMobileError.predictError(message: "CommandBuffer is nil")
+    }
+    inflightSemaphore.wait()
+    
+    let resInput: MTLTexture
+    if let inPre = preProcessKernle {
+      do {
+        try inPre.compute(inputTexuture: input, commandBuffer: buffer)
+        resInput = inPre.outputTexture
+      } catch let error {
+        throw error
+      }
+    } else {
+      resInput = input
+    }
+    
+    let beforeDate = Date.init()
+    let inputTexture = InputTexture.init(inMTLTexture: resInput, inExpectDim: Dim.init(inDim: dim))
+    program.scope.setInput(input: inputTexture)
+    //(ops.count - except)
+    for i in 0..<(ops.count - except) {
+      let op = ops[i]
+      do {
+        try op.run(device: device, buffer: buffer)
+      } catch let error {
+        throw error
+      }
+    }
+    
+    var outputTextures: [String : [Variant]]?
+    if except > 0 {
+      ops[ops.count - except].computeMiddleResult(device: device, buffer: buffer)
+      outputTextures = ops[ops.count - except].inputVariant()
+    }
+    
+    buffer.addCompletedHandler { [weak self] (commandbuffer) in
+//      let inputArr = resInput.toTensor(dim: (n: dim[0], c: dim[3], h: dim[1], w: dim[2]))
+//      print(inputArr.strideArray())
+//
+////      print(dim)
+//      writeToLibrary(fileName: "test_image_ssd_ar", array: inputArr)
+//      print(" write done ")
+
+//      print("write to library done")
+//      return
+//                  print(inputArr)
+//
+//                  let stridableInput: [(index: Int, value: Float)] = input.stridableFloatArray()
+//                  print(stridableInput)
+//
+//                  let _: Flo? = input.logDesc(header: "input: ", stridable: true)
+//      for i in 0..<self!.ops.count {
+//        let op = self!.ops[i]
+//        print(" 第 \(i) 个 op: ")
+//        op.delogOutput()
+//      }
+      
+//      return;
+//      self!.ops[testTo - 2].delogOutput()
+//      self!.ops[testTo - 1].delogOutput()
+//      self!.ops[5].delogOutput()
+
+//      return
+      
+      guard let SSelf = self else {
+//        return
+        fatalError()
+      }
+      
+      let afterDate = Date.init()
+      var resultHolder: GPUResultHolder
+      if except > 0 {
+        resultHolder = GPUResultHolder.init(inDim: [], inPointer: nil, inCapacity: 0, inElapsedTime: afterDate.timeIntervalSince(beforeDate), inIntermediateResults: outputTextures)
+      } else {
+        let outputVar: Variant = SSelf.program.scope.output()!
+        let output: FetchHolder = outputVar as! FetchHolder
+//        let beforeToTensorDate = Date.init()
+
+        resultHolder = GPUResultHolder.init(inDim: output.dim, inPointer: output.result, inCapacity: output.capacity, inElapsedTime: afterDate.timeIntervalSince(beforeDate))
+        
+//        let timeToTensor = Date.init().timeIntervalSince(beforeToTensorDate)
+//        print(timeToTensor)
+      }
+
+      completionHandle(resultHolder)
+      SSelf.inflightSemaphore.signal()
+    }
+    buffer.commit()
+  }
+  
+  public func clear() {
+    program.scope.clear()
+  }
+  
+}

metal/paddle-mobile/paddle-mobile/framework/Loader.swift

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ 
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+ 
+ http://www.apache.org/licenses/LICENSE-2.0
+ 
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+import Foundation
+import SwiftProtobuf
+
+public class Loader<P: PrecisionType> {
+  class ParaLoader {
+    let file: UnsafeMutablePointer<FILE>
+    let fileSize: Int
+    var nowIndex: Int
+    init(paramPath: String) throws {
+      guard let tmpFile = fopen(paramPath, "rb") else {
+        throw PaddleMobileError.loaderError(message: "open param file error" + paramPath)
+      }
+      file = tmpFile
+      fseek(file, 0, SEEK_END)
+      fileSize = ftell(file)
+      guard fileSize > 0 else {
+        throw PaddleMobileError.loaderError(message: "param file size is too small")
+      }
+      rewind(file)
+      nowIndex = 0
+    }
+    
+    func read(tensor: Tensor<P>) throws {
+      guard nowIndex <= fileSize else {
+        throw PaddleMobileError.loaderError(message: "out of the file range")
+      }
+      
+      func pointerReader<T>(type: T.Type) -> T {
+        let ptr = UnsafeMutablePointer<T>.allocate(capacity: MemoryLayout<T>.size)
+        fread(ptr, 1, MemoryLayout<T>.size, file)
+        nowIndex += MemoryLayout<T>.size
+        let pointee = ptr.pointee
+        ptr.deinitialize(count: MemoryLayout<UInt32>.size)
+        ptr.deallocate()
+        return pointee
+      }
+      
+      let _ = pointerReader(type: UInt32.self)
+      let lodLevel = pointerReader(type: UInt64.self)
+      for _ in 0..<lodLevel {
+        let size = pointerReader(type: UInt64.self)
+        for _ in 0..<Int(size/UInt64(MemoryLayout<size_t>.size)){
+          _ = pointerReader(type: size_t.self)
+        }
+      }
+      
+      let _ = pointerReader(type: UInt32.self)
+      
+      let tensorDescSize = pointerReader(type: Int32.self)
+      
+      fseek(file, Int(tensorDescSize), SEEK_CUR)
+      nowIndex += Int(tensorDescSize)
+      
+      /*
+       这里没有根据 Data Type 去判断, 而是从外部泛型直接指定了精度
+       */
+      
+      //现在模型传入模型为  Float 类型, 这块应该根据模型来
+      //            let tmpCapacity = MemoryLayout<Float>.size * tensor.numel()
+      //            let tmpPointer = UnsafeMutablePointer<Float>.allocate(capacity: tmpCapacity);
+      let bytesRead = fread(tensor.data.pointer, 1, tensor.data.size, file)
+      
+      guard bytesRead == tensor.data.size else {
+        throw PaddleMobileError.loaderError(message: "param read size error")
+      }
+      
+      // TODO: use script to convert
+      //            let bytesRead = fread(tmpPointer, 1, tmpCapacity, file)
+      //            for i in 0..<tensor.numel() {
+      //                tensor.data[i] = P.init(inFloat: tmpPointer[i])
+      //            }
+      //            tmpPointer.deinitialize(count: tmpCapacity)
+      //            tmpPointer.deallocate()
+      
+      nowIndex += bytesRead
+    }
+    
+    deinit {
+      fclose(file)
+    }
+  }
+  class ParaLoaderWithPointer {
+    var paramPointer: UnsafeMutableRawPointer
+      let paramSize: Int
+      var nowIndex: Int
+      init(pPointer: UnsafeMutableRawPointer,pSize:Int) throws {
+          paramPointer = UnsafeMutableRawPointer.init(pPointer)
+          paramSize = pSize
+          nowIndex = 0
+      }
+    
+      func read(tensor: Tensor<P>) throws {
+        guard nowIndex <= paramSize else {
+          throw PaddleMobileError.loaderError(message: "out of the file range")
+        }
+        var readerIndex: Int = 0
+        func pointerReader<T>(type: T.Type) -> T {
+          let ptr = UnsafeMutablePointer<T>.allocate(capacity: MemoryLayout<T>.size)
+          memcpy(ptr, paramPointer.advanced(by: Int(readerIndex)), MemoryLayout<T>.size)
+          nowIndex += MemoryLayout<T>.size
+          readerIndex += MemoryLayout<T>.size
+          let pointee = ptr.pointee
+          ptr.deinitialize(count: MemoryLayout<UInt32>.size)
+          ptr.deallocate()
+          
+          return pointee
+        }
+        let _ = pointerReader(type: UInt32.self)
+        let lodLevel = pointerReader(type: UInt64.self)
+        for _ in 0..<lodLevel {
+          let size = pointerReader(type: UInt64.self)
+          for _ in 0..<Int(size/UInt64(MemoryLayout<size_t>.size)){
+            _ = pointerReader(type: size_t.self)
+          }
+        }
+        
+        let _ = pointerReader(type: UInt32.self)
+        let tensorDescSize = pointerReader(type: Int32.self)
+        
+        paramPointer = paramPointer.advanced(by: Int(readerIndex))
+        paramPointer = paramPointer.advanced(by: Int(tensorDescSize))
+        nowIndex += Int(tensorDescSize)
+        
+        let _ = memcpy(tensor.data.pointer, paramPointer, tensor.data.size)
+        paramPointer = paramPointer.advanced(by: Int(tensor.data.size))
+        nowIndex += tensor.data.size
+    }
+    deinit {
+    }
+  }
+  public init(){}
+  func loadModelandParam(_ device:MTLDevice,_ modelData:Data, _ paraLoaderPointer:ParaLoaderWithPointer?, _ paraLoader:ParaLoader?) throws -> Program {
+    do {
+      let protoProgram = try PaddleMobile_Framework_Proto_ProgramDesc.init(
+        serializedData: modelData)
+      
+      let originProgramDesc = ProgramDesc.init(protoProgram: protoProgram)
+      let programDesc = ProgramOptimize<P>.init().optimize(originProgramDesc: originProgramDesc)
+      print(programDesc)
+      
+      guard programDesc.blocks.count > 0 else {
+        throw PaddleMobileError.loaderError(message: "count of blocks must greater than 0")
+      }
+      
+      // to get feed key and fetch key
+      let block = programDesc.blocks[0]
+      guard let firstOp = block.ops.first, let lastOp = block.ops.last else {
+        throw PaddleMobileError.loaderError(message: "at least two operator")
+      }
+      
+      guard firstOp.type == gFeedType, lastOp.type == gFetchType else {
+        throw PaddleMobileError.loaderError(message: "the first op is not feed or the last op is not fetch")
+      }
+      
+      guard let inputKey = opInfos[gFeedType]?.inputs.first, let outKey = opInfos[gFetchType]?.outputs.first else {
+        throw PaddleMobileError.loaderError(message: "the feed input key or fetch output key not found")
+      }
+      guard let feedKey = firstOp.inputs[inputKey]?.first, let fetchKey = lastOp.outputs[outKey]?.first else {
+        throw PaddleMobileError.loaderError(message: "feed key or fetch key not found")
+      }
+      
+      let scope = Scope.init(inFeedKey: feedKey, inFetchKey: fetchKey)
+      
+      // to load memory
+      for block in programDesc.blocks {
+        for varDesc in block.vars {
+          if (varDesc.type == .LodTensor) {
+            guard let tensorDesc = varDesc.tensorDesc else {
+              throw PaddleMobileError.loaderError(message: "get tensor desc failed")
+            }
+            
+            if (varDesc.persistable
+              && varDesc.type != .FeedMiniBatch
+              && varDesc.type != .FetchList) {
+              let dimArr = tensorDesc.dims
+              
+              guard dimArr.count > 0 else {
+                throw PaddleMobileError.loaderError(message: "tensor desc dim size error")
+              }
+              
+              let dim = Dim.init(inDim: dimArr)
+              let tensor = Tensor<P>.init(inDim: dim, inLayout: tensorDesc.dataLayout)
+              do {
+                if paraLoaderPointer != nil {
+                  try paraLoaderPointer!.read(tensor: tensor)
+                }
+                
+                if paraLoader != nil {
+                  try paraLoader!.read(tensor: tensor)
+                }
+              } catch let error {
+                throw error
+              }
+              //              tensor.convert(to: DataLayout.NHWC())
+              //                            tensor.initBuffer(device: device)
+              scope[varDesc.name] = tensor
+            } else {
+              let dim = Dim.init(inDim: tensorDesc.dims)
+              scope[varDesc.name] = Texture<P>.init(device: device, inDim: dim)
+            }
+          } else {
+            if varDesc.name == fetchKey {
+//              scope[varDesc.name] = ResultHolder.init(inDim: [], inResult: [], inCapacity: <#Int#>, inElapsedTime: 0.0)
+            } else if varDesc.name == feedKey {
+            }
+          }
+        }
+      }
+      
+      let program = Program.init(inProgramDesc: programDesc, inScope: scope)
+      
+      return program
+    } catch _ {
+      throw PaddleMobileError.loaderError(message: "protobuf decoder error")
+    }
+  }
+  public func load(device:MTLDevice, paramPointer: UnsafeMutableRawPointer, paramSize:Int, modePointer: UnsafeMutableRawPointer, modelSize: Int) throws -> Program {
+    let modelData = Data.init(bytes:modePointer, count:modelSize)
+    guard let paraLoader = try? ParaLoaderWithPointer.init(pPointer: paramPointer,pSize: paramSize) else {
+      throw PaddleMobileError.loaderError(message: "load para error")
+    }
+    do {
+      let program = try loadModelandParam(device,modelData,paraLoader,nil)
+      return program
+    } catch let error {
+      throw error
+    }
+  }
+    
+  public func load(device: MTLDevice, modelPath: String, paraPath: String) throws -> Program{
+    guard let modelData = try? Data.init(contentsOf: URL.init(fileURLWithPath: modelPath)) else {
+      throw PaddleMobileError.loaderError(message: "load " + modelPath + " failed !")
+    }
+    guard let paraLoader = try? ParaLoader.init(paramPath: paraPath) else {
+      throw PaddleMobileError.loaderError(message: "load para error")
+    }
+    
+    do {
+      let program = try loadModelandParam(device,modelData,nil,paraLoader)
+      return program
+    } catch let error {
+      throw error
+    }
+  }
+}

metal/paddle-mobile/paddle-mobile/framework/Tensor.swift

@@ -12,7 +12,6 @@
  See the License for the specific language governing permissions and
  limitations under the License. */
 
-import Accelerate
 import Foundation
 
 protocol Tensorial: CustomStringConvertible, CustomDebugStringConvertible{
@@ -27,10 +26,11 @@ extension Tensorial {
   }
 }
 
-class Tensor<P: PrecisionType>: Tensorial {
-    enum BufferPrecision {
+public enum ComputePrecision {
   case Float32, Float16
-    }
+}
+
+class Tensor<P: PrecisionType>: Tensorial {
   
   var data: Data
   var dim: Dim
@@ -57,11 +57,11 @@ class Tensor<P: PrecisionType>: Tensorial {
       pointer.deallocate()
     }
     deinit {
-//            release()
+      //            release()
     }
   }
   
-    required init(inDim: Dim, inLayout: DataLayout = .NCHW) {
+  required init(inDim: Dim, inLayout: DataLayout = DataLayout.NCHW()) {
     dim = inDim
     let size = inDim.numel() * MemoryLayout<P>.size
     let pointer = UnsafeMutablePointer<P>.allocate(capacity: size)
@@ -78,13 +78,13 @@ class Tensor<P: PrecisionType>: Tensorial {
       return
     }
     
-        guard layout == .NCHW && to == .NHWC else {
+    guard layout == DataLayout.NCHW() && to == DataLayout.NHWC() else {
       // other not support
       return
     }
     let newPointer = UnsafeMutablePointer<P>.allocate(capacity: data.size)
     
-        if layout == .NCHW {
+    if layout == DataLayout.NCHW() {
       NCHW2NHWC(newPtr: newPointer)
     }
     
@@ -93,20 +93,34 @@ class Tensor<P: PrecisionType>: Tensorial {
     layout = to
   }
   
-    func float32ToFloat16(input: UnsafeMutablePointer<Float32>, output: UnsafeMutableRawPointer, count: Int) {
-        var float32Buffer = vImage_Buffer(data: input,  height: 1, width: UInt(count), rowBytes: count * 4)
-        var float16buffer = vImage_Buffer(data: output, height: 1, width: UInt(count), rowBytes: count * 2)
-        guard vImageConvert_PlanarFtoPlanar16F(&float32Buffer, &float16buffer, 0) == kvImageNoError else {
-            fatalError(" float 32 to float 16 error ! ")
+
+  
+  func initBuffer(device: MTLDevice, precision: ComputePrecision = .Float16, convertToNHWC: Bool = true, withTranspose: Bool = false) {
+    if convertToNHWC {
+//      print(layout)
+      convert(to: DataLayout.NHWC())
     }
+    
+    if withTranspose {
+      let transposePointer = UnsafeMutablePointer<P>.allocate(capacity: numel())
+      let n = dim[0]
+      let hwc = numel()/n
+      for j in 0..<hwc {
+        for i in 0..<n {
+          //data[i * hwc + j]
+          transposePointer[j * n + i] = data[i * hwc + j]
+        }
+      }
+
+      dim.swapeDimAt(index1: 0, index2: 3)
+      data.release()
+      data.pointer = transposePointer
     }
     
-    func initBuffer(device: MTLDevice, precision: BufferPrecision = .Float32) {
     guard let floatPointer = data.pointer as? UnsafeMutablePointer<Float32> else {
       fatalError(" not support yet ")
     }
     
-        
     let precisionSize: Int
     switch precision {
     case .Float32:
@@ -116,7 +130,7 @@ class Tensor<P: PrecisionType>: Tensorial {
     }
     
     if dim.cout() == 4 {
-            if layout == .NHWC {
+      if layout == DataLayout.NHWC() {
         let C = dim[3]
         let cSlices = (C + 3) / 4
         let paddedC = cSlices * 4
@@ -146,6 +160,8 @@ class Tensor<P: PrecisionType>: Tensorial {
             for j in 0..<paddedC {
               if j < C {
                 dstPtr[j] = tmpPointer[j]
+              } else {
+                dstPtr[j] = 0
               }
             }
             tmpPointer += C
@@ -159,17 +175,59 @@ class Tensor<P: PrecisionType>: Tensorial {
             float32ToFloat16(input: convertedPointer, output: buffer.contents(), count: count)
           }
           
+          convertedPointer.deinitialize(count: count)
+          convertedPointer.deallocate()
+        }
+      } else {
+        let C = dim[3]
+        let cSlices = (C + 3) / 4
+        let paddedC = cSlices * 4
+        let count = paddedC * dim[0] * dim[1] * dim[2]
+        if C == paddedC {
+          buffer = device.makeBuffer(length: count * precisionSize)
+          switch precision {
+          case .Float32:
+            buffer?.contents().copyMemory(from: data.pointer, byteCount: count * MemoryLayout<P>.stride)
+          case .Float16:
+            float32ToFloat16(input: floatPointer, output: buffer.contents(), count: count)
+          }
+        } else if C == 1 {
+          fatalError(" not support ")
+        } else {
+          buffer = device.makeBuffer(length: count * precisionSize)
+          let convertedPointer = UnsafeMutablePointer<Float32>.allocate(capacity: count)
+          var tmpPointer = floatPointer
+          var dstPtr = convertedPointer
+          for _ in 0..<dim[0] * dim[1] * dim[2] {
+            for j in 0..<paddedC {
+              if j < C {
+                dstPtr[j] = tmpPointer[j]
+              } else {
+                dstPtr[j] = 0
+              }
+            }
+            tmpPointer += C
+            dstPtr += paddedC
+          }
+          
+          switch precision {
+          case .Float32:
+            buffer?.contents().copyMemory(from: convertedPointer, byteCount: count * MemoryLayout<P>.stride)
+          case .Float16:
+            float32ToFloat16(input: convertedPointer, output: buffer.contents(), count: count)
+          }
           convertedPointer.deinitialize(count: count)
           convertedPointer.deallocate()
         }
       }
     } else if dim.cout() == 1 {
-            buffer = device.makeBuffer(length: numel() * precisionSize)
+      let num = ((numel() + 3) / 4) * 4
+      buffer = device.makeBuffer(length: num * precisionSize)
       switch precision {
       case .Float32:
-                buffer?.contents().copyMemory(from: data.pointer, byteCount: numel() * MemoryLayout<P>.stride)
+        buffer?.contents().copyMemory(from: data.pointer, byteCount: num * MemoryLayout<P>.stride)
       case .Float16:
-                float32ToFloat16(input: floatPointer, output: buffer.contents(), count: numel())
+        float32ToFloat16(input: floatPointer, output: buffer.contents(), count: num)
       }
     } else {
       fatalError(" not support !")
@@ -232,7 +290,6 @@ class Tensor<P: PrecisionType>: Tensorial {
   }
 }
 
-
 extension Tensor {
   
   var debugDescription: String {

metal/paddle-mobile/paddle-mobile/framework/Texture.swift

@@ -22,7 +22,6 @@ class InputTexture {
     mtlTexture = inMTLTexture
     expectDim = inExpectDim
   }
-    
 }
 
 extension InputTexture {
@@ -39,46 +38,100 @@ extension InputTexture {
   }
 }
 
+
+/*
+ 4 维 tensor 存储 texture，要考虑 transpose
+ transpose 之后的维度是 [a, b, c, d]，对应的texture_2darray
+ .width = c
+ .height = b
+ .len = a * d + 3 / 4
+ 
+低于 4 维的 tensor，transpose 必须为 [0, 1, 2, 3] 既不考虑 transpose
+ 
+// TODO transpose 对于低维 tensor 的扩展原则。。。
+// [a, b] -> [1, 1, a, b] transpose 必须为 [0, 1, x, x]
+// [a] -> [1, 1, 1, a] transpose 必须为 [0, 1, 2, 3]
+// [a, b, c] -> [1, a, b, c] tranpose 必须为 [0, x, x, x]
+
+3 维 tensor [a, b, c] 对应的 texture_2darray,
+.width = c
+.height = b
+.len = a + 3 / 4
+ 
+ 2 维 tensor [a, b] 对应的 texture_2darray
+ .width = b + 3 / 4
+ .height = a
+ .len = 1
+ 
+ 1 维 tensor [a] 对应的 texture_2darray
+ .width = a + 3 / 4
+ .height = 1
+ .len = 1
+ */
+
+
 public class Texture<P: PrecisionType>: Tensorial {
   var dim: Dim
-    let textureDesc: MTLTextureDescriptor
-    var metalTexture: MTLTexture
+  public var tensorDim: Dim
+  public var padToFourDim: Dim
+  private var textureDesc: MTLTextureDescriptor!
+  public var metalTexture: MTLTexture!
+  var transpose: [Int] = [0, 1, 2, 3]
+  
+  func toTensor() -> [Float32] {
+    guard  padToFourDim.cout() == 4 else {
+      fatalError("- not support -")
+    }
+    return metalTexture.toTensor(dim: (n: dim[0], c: dim[3], h: dim[1], w: dim[2]))
+  }
+  
+  func realNHWC() -> [Float32] {
+    guard padToFourDim.cout() == 4 else {
+      fatalError(" - not support - ")
+    }
+    return metalTexture.realNHWC(dim: (n: padToFourDim[0], h: padToFourDim[1], w: padToFourDim[2], c: padToFourDim[3]))
+  }
+  
+  func initTexture(device: MTLDevice, inTranspose: [Int] = [0, 1, 2, 3], computePrecision: ComputePrecision = .Float16) {
+    transpose = inTranspose
+    for i in 0..<(4 - tensorDim.cout()) {
+      if i != inTranspose[i] {
+        fatalError()
+      }
+    }
+    let newDim = transpose.map { padToFourDim[$0] }
+    
+    let newLayout = transpose.map { layout.layoutWithDim[$0] }
+    
+    layout = DataLayout.init(newLayout)
+    dim = Dim.init(inDim: newDim)
     
-    init(device: MTLDevice, inDim: Dim, inLayout: DataLayout = .NHWC) {
-        dim = inDim
-        layout = inLayout
     let tmpTextureDes = MTLTextureDescriptor.init()
-        if inDim.cout() == 1 {
-            tmpTextureDes.width = inDim[0]
-            tmpTextureDes.textureType = .type1D
-        } else if inDim.cout() == 4 {
-            tmpTextureDes.height = inDim[1]
-            tmpTextureDes.width = inDim[2]
-//            print("n : \(inDim[0])")
-//            print(inDim[3] * inDim[0])
-            tmpTextureDes.depth = 1
-            tmpTextureDes.arrayLength = (inDim[3] * inDim[0] + 3)/4
     tmpTextureDes.textureType = .type2DArray
-        } else if inDim.cout() == 2 {
-            tmpTextureDes.height = 1
-            tmpTextureDes.width = 1
     tmpTextureDes.depth = 1
-            tmpTextureDes.arrayLength = (inDim[0] * inDim[1] + 3)/4
-            tmpTextureDes.textureType = .type2DArray
-        } else {
-            fatalError(" not suuprt ")
+    
+    switch tensorDim.cout() {
+    case 4:
+      tmpTextureDes.width = newDim[2]
+      tmpTextureDes.height = newDim[1]
+      tmpTextureDes.arrayLength = ((newDim[0]) * (newDim[3]) + 3) / 4
+    case 3:
+      tmpTextureDes.width = newDim[3]
+      tmpTextureDes.height = newDim[2]
+      tmpTextureDes.arrayLength = (newDim[1] + 3) / 4
+    case 2, 1:
+      tmpTextureDes.width = (newDim[3] + 3) / 4
+      tmpTextureDes.height = newDim[2]
+      tmpTextureDes.arrayLength = 1
+    default:
+      fatalError("unreachable")
     }
    
-        if MemoryLayout<P>.size == 1 {
-            tmpTextureDes.pixelFormat = .rgba8Unorm
-        } else if MemoryLayout<P>.size == 2 {
+    if computePrecision == .Float16 {
       tmpTextureDes.pixelFormat = .rgba16Float
-        } else if MemoryLayout<P>.size == 4 {
-//            tmpTextureDes.pixelFormat = .r32Float
+    } else if computePrecision == .Float32 {
       tmpTextureDes.pixelFormat = .rgba32Float
-
     }
-//        tmpTextureDes.pixelFormat = .rgba16Float
     
     tmpTextureDes.usage = [.shaderRead, .shaderWrite]
     tmpTextureDes.storageMode = .shared
@@ -86,42 +139,25 @@ public class Texture<P: PrecisionType>: Tensorial {
     metalTexture = device.makeTexture(descriptor: tmpTextureDes) ?! " texture nil "
   }
   
-//    required public init(inDim: Dim, inLayout: DataLayout = .NHWC, inTexture: MTLTexture) {
-//        dim = inDim
-//        layout = inLayout
-//        metalTexture = inTexture
-//        let tmpTextureDes = MTLTextureDescriptor.init()
-//        
-//        if inDim.cout() == 1 {
-//            tmpTextureDes.width = inDim[0]
-//            tmpTextureDes.textureType = .type1D
-//        } else if inDim.cout() == 2 {
-//            tmpTextureDes.height = inDim[0]
-//            tmpTextureDes.width = inDim[1]
-//            tmpTextureDes.textureType = .type2D
-//        } else if inDim.cout() == 3 {
-//            fatalError(" not support texture dim 3")
-//        } else if inDim.cout() == 4 {
-//            tmpTextureDes.height = inDim[1]
-//            tmpTextureDes.width = inDim[2]
-//            tmpTextureDes.depth = inDim[3] * inDim[1]
-//            tmpTextureDes.textureType = .type2DArray
-//        }
-//        
-//        tmpTextureDes.pixelFormat = .r32Float
-//        tmpTextureDes.storageMode = .shared
-//        textureDesc = tmpTextureDes
-//        let device = MTLCreateSystemDefaultDevice()
-//        metalTexture = device!.makeTexture(descriptor: tmpTextureDes)!
-//    }
-    
-//    init() {
-//        dim = Dim.init(inDim: [])
-//        layout = .NCHW
-//        let device = MTLCreateSystemDefaultDevice()
-//        textureDesc = MTLTextureDescriptor.init()
-//        metalTexture = device!.makeTexture(descriptor: textureDesc)!
-//    }
+  init(device: MTLDevice, inDim: Dim) {
+    var fourDim: Dim
+    if inDim.cout() == 4 {
+      fourDim = inDim
+    } else if inDim.cout() < 4 {
+      var fourDimNum: [Int] = []
+      for _ in 0..<(4 - inDim.cout()) {
+        fourDimNum.append(1)
+      }
+      fourDimNum.append(contentsOf: inDim.dims)
+      fourDim = Dim.init(inDim: fourDimNum)
+    } else {
+      fatalError(" not support ")
+    }
+    tensorDim = inDim
+    dim = fourDim
+    padToFourDim = fourDim
+    layout = DataLayout.init([(.N, fourDim[0]), (.C, fourDim[1]), (.H, fourDim[2]), (.W, fourDim[3])])
+  }
   
   private(set) var layout: DataLayout
 }

metal/paddle-mobile/paddle-mobile/paddle_mobile.h

@@ -14,12 +14,15 @@
 
 #pragma once
 
+#import "PaddleMobileCPU.h"
+#import "CPUCompute.h"
+#import "PaddleMobileGPU.h"
 #import <UIKit/UIKit.h>
 
 //! Project version number for paddle_mobile.
-FOUNDATION_EXPORT double paddle_mobileVersionNumber;
+//FOUNDATION_EXPORT double paddle_mobileVersionNumber;
 
 //! Project version string for paddle_mobile.
-FOUNDATION_EXPORT const unsigned char paddle_mobileVersionString[];
+//FOUNDATION_EXPORT const unsigned char paddle_mobileVersionString[];
 
 

python/tools/imagetools/imagetools.py

+# coding=utf-8
+import cv2
+from array import array
+
+
+def resize_take_rgbs(path, shape_h_w):
+    print '--------------resize_take_rgbs-----------------begin'
+    image = cv2.imread(path)
+    # print image.shape
+    cv2.imshow("before", image)
+
+    print_rgb(image[0, 0])
+    # image len may be for .just check it
+    # image.resize(shape_h_w)
+
+    image = cv2.resize(image, (shape_h_w[0], shape_h_w[1]))
+
+    cv2.imshow("after", image)
+    print image.shape
+    height = shape_h_w[0]
+    width = shape_h_w[1]
+
+    rs_ = []
+    gs_ = []
+    bs_ = []
+    for h in range(0, height):
+        for w in range(0, width):
+            bs_.append(image[h, w, 0])
+            gs_.append(image[h, w, 1])
+            rs_.append(image[h, w, 2])
+
+    # print image[2, 2, 0]/255.
+    print len(bs_)
+    print len(gs_)
+    print len(rs_)
+    print '--------------resize_take_rgbs-----------------end'
+    return bs_, gs_, rs_
+
+
+def print_rgb((b, g, r)):
+    print "像素 - R:%d,G:%d,B:%d" % (r, g, b)  # 显示像素值
+    #
+    # image[0, 0] = (100, 150, 200)  # 更改位置(0,0)处的像素
+    #
+    # (b, g, r) = image[0, 0]  # 再次读取(0,0)像素
+    # print "位置(0,0)处的像素 - 红:%d,绿:%d,蓝:%d" % (r, g, b)  # 显示更改后的像素值
+    #
+    # corner = image[0:100, 0:100]  # 读取像素块
+    # cv2.imshow("Corner", corner)  # 显示读取的像素块
+    #
+    # image[0:100, 0:100] = (0, 255, 0);  # 更改读取的像素块
+    #
+    # cv2.imshow("Updated", image)  # 显示图像
+    #
+    # cv2.waitKey(0)  # 程序暂停
+
+
+def save_to_file(to_file_name, array):
+    to_file = open(to_file_name, "wb")
+    array.tofile(to_file)
+    to_file.close()

python/tools/imagetools/img2nchw.py

+# coding=utf-8
+import cv2
+from array import array
+import imagetools as tools
+from enum import Enum
+
+
+class ChannelType(Enum):
+    RGB = 0,
+    BGR = 1
+
+
+def combine_bgrs_nchw(bgrs, means_b_g_r, scale, channel_type=ChannelType.BGR):
+    print '--------------combine_bgrs_nchw-----------------begin'
+    print "scale: %f" % scale
+    print means_b_g_r
+    # print len(bgrs)
+    bs = bgrs[0]
+    gs = bgrs[1]
+    rs = bgrs[2]
+
+    assert len(bs) == len(gs) == len(rs)
+    print len(bs)
+    bgrs_float_array = array('f')
+
+    if channel_type == ChannelType.BGR:
+        print 'bgr'
+        for i in range(0, len(bs)):
+            bgrs_float_array.append((bs[i] - means_b_g_r[0]) * scale)  # b
+        for i in range(0, len(gs)):
+            bgrs_float_array.append((gs[i] - means_b_g_r[1]) * scale)  # g
+        for i in range(0, len(rs)):
+            bgrs_float_array.append((rs[i] - means_b_g_r[2]) * scale)  # r
+    elif channel_type == ChannelType.RGB:
+        print 'rgb'
+
+        for i in range(0, len(rs)):
+            bgrs_float_array.append((rs[i] - means_b_g_r[2]) * scale)  # r
+        for i in range(0, len(gs)):
+            bgrs_float_array.append((gs[i] - means_b_g_r[1]) * scale)  # g
+        for i in range(0, len(bs)):
+            bgrs_float_array.append((bs[i] - means_b_g_r[0]) * scale)  # b
+
+    print len(bgrs_float_array)
+
+    print '------------------'
+    print bgrs_float_array[0]
+    print bgrs_float_array[416 * 416 * 2 + 416 * 2 + 2]
+
+    # for i in range(0, 9):
+    #     print'bs %d' % i
+    #     print bs[i] / 255.
+
+    print bs[416 * 2 + 2] / 255.
+    print '--------------combine_bgrs_nchw-----------------end'
+
+    return bgrs_float_array
+
+
+# bgrs = tools.resize_take_rgbs('banana.jpeg', (224, 224, 3))
+# array = combine_bgrs_nchw(bgrs, (103.94, 116.78, 123.68), 0.017, array,ChannelType.BGR)
+# tools.save_to_file('banana_1_3_224_224_nchw_float')
+
+# cv2.waitKey(0)
+
+
+bgrs = tools.resize_take_rgbs('datas/newyolo.jpg', (416, 416, 3))
+array = combine_bgrs_nchw(bgrs, (0, 0, 0), 1. / 255, ChannelType.RGB)
+tools.save_to_file('datas/desktop_1_3_416_416_nchw_float', array)

python/tools/imagetools/img2nhwc.py

+# coding=utf-8
+import cv2
+from array import array
+import imagetools as tools
+
+
+def combine_bgrs_nhwc(bgrs, means_b_g_r, scale):
+    print "scale: %f" % scale
+    print means_b_g_r
+    # print len(bgrs)
+    bs = bgrs[0]
+    gs = bgrs[1]
+    rs = bgrs[2]
+    assert len(bs) == len(gs) == len(rs)
+    # print len(bs)
+    bgrs_float_array = array('f')
+    for i in range(0, len(bs)):
+        bgrs_float_array.append((rs[i] - means_b_g_r[2]) * scale)  # r
+        bgrs_float_array.append((gs[i] - means_b_g_r[1]) * scale)  # g
+        bgrs_float_array.append((bs[i] - means_b_g_r[0]) * scale)  # b
+
+    print len(bgrs_float_array)
+
+    print '------------------'
+    print bgrs_float_array[0]
+    print bgrs_float_array[999]
+    return bgrs_float_array
+
+
+bgrs = tools.resize_take_rgbs('newyolo_1.jpg', (416, 416, 3))
+array = combine_bgrs_nhwc(bgrs, (0, 0, 0), 1.0 / 255)
+tools.save_to_file('desktop_1_3_416_416_nhwc_float', array)
+
+cv2.waitKey(0)

python/tools/imagetools/numpy2binary.py

+# coding=utf-8
+
+# 这个脚本是可以将numpy合并到二进制
+import cv2
+import numpy as np
+import imagetools as tools
+from array import array
+
+#
+# image = cv2.imread(path)
+# print image.shape
+#
+# print_rgb(image[0, 0])
+# # image len may be for .just check it
+# image.resize(shape_h_w)
+
+
+data = np.fromfile('datas/img.res')
+print data.size
+print data[0]
+
+data.reshape(1, 3, 416, 416)
+out_array = array('f')
+print'--------------------'
+print data.size
+print data[0]
+
+print '如果是nhwc --------'
+# rgb rgb rgb rgb rgb
+print data[416 * 3 * 2 + 3 * 2 + 2]
+# print data[2]
+
+print '如果是nchw --------'
+# rgb rgb rgb rgb rgb
+print data[416 * 416 * 2 + 416 * 2 + 2]
+# print data[2]
+
+# 明明是nchw
+
+for i in range(0, data.size):
+    out_array.append(data[i])
+
+print len(out_array)
+
+print out_array[416 * 416 * 2 + 416 * 2 + 2]
+
+tools.save_to_file('datas/in_put_1_3_416_416_2', out_array)

python/tools/mdl2fluid/float2halffloat.py

+# encoding:utf-8
+import math
+import re
+
+
+def Real2HalfFloat(data):
+    MINNUM = -65536
+    MAXNUM = 65535
+    FloatVal = 0
+    if data:
+        if data < MINNUM:
+            data = MINNUM
+        if data > MAXNUM:
+            data = MAXNUM
+
+        sign = 0
+        if data < 0:
+            sign = 1
+            data = -data
+
+        exp = math.floor((math.log2(data)))
+        expout = exp + 16
+
+        Mantial = round(data / pow(2, exp - 10)) - 1024
+
+        if expout <= 0:
+            FloatVal = 0
+        else:
+            FloatVal = sign * 32768 + expout * 1024 + Mantial
+    return FloatVal
+
+
+def ReadCfloatData(sourcefile):
+    input = []
+    with open(sourcfile, 'r') as f:
+        for line in f.readlines():
+            line = line.strip()
+            line = re.sub('\s+', ' ', line)  # 两个数字间多个空格
+            input.append(line.split(' '))
+    destfile = sourcefile.replace('.dat', '')
+    destfile = destfile.replace('.txt', '')
+    destfile += 'Out.dat'
+    with open(destfile, 'w') as fw:
+        for i in range(len(input)):
+            if len(input[i]) == 2:
+                real = Real2HalfFloat(float(input[i][0]))
+                imag = Real2HalfFloat(float(input[i][1]))
+                result = real * 65536 + imag
+                if imag and not real:
+                    fw.write('0x0000' + "%X" % result + '\n')
+                elif not imag and not real:
+                    fw.write('0x00000000' + '\n')
+                else:
+                    fw.write('0x' + "%X" % result + '\n')
+            elif len(input[i]) == 1:
+                result = Real2HalfFloat(float(input[i][0]))
+                if result:
+                    fw.write('0x' + "%X" % result + '\n')
+                else:
+                    fw.write('0x0000' + '\n')
+
+
+if __name__ == '__main__':
+    print('Tips: Input number 0 if you want to exit!\n')
+    while True:
+        sourcfile = input("input source file:\n")
+        if sourcfile is '0':
+            break
+        ReadCfloatData(sourcfile)
+        print('Transfer Success!')

python/tools/mdl2fluid/framework.proto

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+syntax = "proto2";
+option optimize_for = LITE_RUNTIME;
+package paddle_mobile.framework.proto;
+
+enum AttrType {
+  INT = 0;
+  FLOAT = 1;
+  STRING = 2;
+  INTS = 3;
+  FLOATS = 4;
+  STRINGS = 5;
+  BOOLEAN = 6;
+  BOOLEANS = 7;
+  BLOCK = 8;
+  LONG = 9;
+}
+
+// OpDesc describes an instance of a C++ framework::OperatorBase
+// derived class type.
+message OpDesc {
+
+  message Attr {
+    required string name = 1;
+    required AttrType type = 2;
+    optional int32 i = 3;
+    optional float f = 4;
+    optional string s = 5;
+    repeated int32 ints = 6;
+    repeated float floats = 7;
+    repeated string strings = 8;
+    optional bool b = 10;
+    repeated bool bools = 11;
+    optional int32 block_idx = 12;
+    optional int64 l = 13;
+  };
+
+  message Var {
+    required string parameter = 1;
+    repeated string arguments = 2;
+  };
+
+  required string type = 3;
+  repeated Var inputs = 1;
+  repeated Var outputs = 2;
+  repeated Attr attrs = 4;
+  optional bool is_target = 5 [ default = false ];
+};
+
+// OpProto describes a C++ framework::OperatorBase derived class.
+message OpProto {
+
+  // VarProto describes the C++ type framework::Variable.
+  message Var {
+    required string name = 1;
+    required string comment = 2;
+
+    optional bool duplicable = 3 [ default = false ];
+    optional bool intermediate = 4 [ default = false ];
+    optional bool dispensable = 5 [ default = false ];
+  }
+
+  // AttrProto describes the C++ type Attribute.
+  message Attr {
+    required string name = 1;
+    required AttrType type = 2;
+    required string comment = 3;
+    // If that attribute is generated, it means the Paddle third
+    // language binding has responsibility to fill that
+    // attribute. End-User should not set that attribute.
+    optional bool generated = 4 [ default = false ];
+  }
+
+  required string type = 1;
+  repeated Var inputs = 2;
+  repeated Var outputs = 3;
+  repeated Attr attrs = 4;
+  required string comment = 5;
+}
+
+message VarType {
+  enum Type {
+    // Pod Types
+    BOOL = 0;
+    INT16 = 1;
+    INT32 = 2;
+    INT64 = 3;
+    FP16 = 4;
+    FP32 = 5;
+    FP64 = 6;
+
+    // Other types that may need additional descriptions
+    LOD_TENSOR = 7;
+    SELECTED_ROWS = 8;
+    FEED_MINIBATCH = 9;
+    FETCH_LIST = 10;
+    STEP_SCOPES = 11;
+    LOD_RANK_TABLE = 12;
+    LOD_TENSOR_ARRAY = 13;
+    PLACE_LIST = 14;
+    READER = 15;
+    CHANNEL = 16;
+    // Any runtime decided variable type is raw
+    // raw variables should manage their own allocations
+    // in operators like nccl_op
+    RAW = 17;
+    TUPLE = 18;
+  }
+
+  required Type type = 1;
+
+  message TensorDesc {
+    // Should only be PODType. Is enforced in C++
+    required Type data_type = 1;
+    repeated int64 dims = 2; // [UNK, 640, 480] is saved as [-1, 640, 480]
+  }
+  optional TensorDesc selected_rows = 2;
+
+  message LoDTensorDesc {
+    required TensorDesc tensor = 1;
+    optional int32 lod_level = 2 [ default = 0 ];
+  }
+  optional LoDTensorDesc lod_tensor = 3;
+
+  message LoDTensorArrayDesc {
+    required TensorDesc tensor = 1;
+    optional int32 lod_level = 2 [ default = 0 ];
+  }
+  optional LoDTensorArrayDesc tensor_array = 4;
+
+  message ReaderDesc { repeated LoDTensorDesc lod_tensor = 1; }
+  optional ReaderDesc reader = 5;
+
+  message ChannelDesc {
+    required Type data_type = 1;
+    required int64 capacity = 2;
+  }
+  optional ChannelDesc channel = 6;
+
+  message Tuple { repeated Type element_type = 1; }
+  optional Tuple tuple = 7;
+}
+
+message VarDesc {
+  required string name = 1;
+  required VarType type = 2;
+  optional bool persistable = 3 [ default = false ];
+}
+
+message BlockDesc {
+  required int32 idx = 1;
+  required int32 parent_idx = 2;
+  repeated VarDesc vars = 3;
+  repeated OpDesc ops = 4;
+  optional int32 forward_block_idx = 5 [ default = -1 ];
+}
+
+// Please refer to
+// https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/program.md
+// for more details.
+// TODO(panyx0718): A model can have multiple programs. Need a
+// way to distinguish them. Maybe ID or name?
+message ProgramDesc { repeated BlockDesc blocks = 1; }

python/tools/mdl2fluid/framework_pb2.py

+# Generated by the protocol buffer compiler.  DO NOT EDIT!
+# source: framework.proto
+
+import sys
+_b=sys.version_info[0]<3 and (lambda x:x) or (lambda x:x.encode('latin1'))
+from google.protobuf.internal import enum_type_wrapper
+from google.protobuf import descriptor as _descriptor
+from google.protobuf import message as _message
+from google.protobuf import reflection as _reflection
+from google.protobuf import symbol_database as _symbol_database
+from google.protobuf import descriptor_pb2
+# @@protoc_insertion_point(imports)
+
+_sym_db = _symbol_database.Default()
+
+
+
+
+DESCRIPTOR = _descriptor.FileDescriptor(
+  name='framework.proto',
+  package='paddle_mobile.framework.proto',
+  syntax='proto2',
+  serialized_pb=_b('\n\x0f\x66ramework.proto\x12\x1dpaddle_mobile.framework.proto\"\xe5\x03\n\x06OpDesc\x12\x0c\n\x04type\x18\x03 \x02(\t\x12\x39\n\x06inputs\x18\x01 \x03(\x0b\x32).paddle_mobile.framework.proto.OpDesc.Var\x12:\n\x07outputs\x18\x02 \x03(\x0b\x32).paddle_mobile.framework.proto.OpDesc.Var\x12\x39\n\x05\x61ttrs\x18\x04 \x03(\x0b\x32*.paddle_mobile.framework.proto.OpDesc.Attr\x12\x18\n\tis_target\x18\x05 \x01(\x08:\x05\x66\x61lse\x1a\xd3\x01\n\x04\x41ttr\x12\x0c\n\x04name\x18\x01 \x02(\t\x12\x35\n\x04type\x18\x02 \x02(\x0e\x32\'.paddle_mobile.framework.proto.AttrType\x12\t\n\x01i\x18\x03 \x01(\x05\x12\t\n\x01\x66\x18\x04 \x01(\x02\x12\t\n\x01s\x18\x05 \x01(\t\x12\x0c\n\x04ints\x18\x06 \x03(\x05\x12\x0e\n\x06\x66loats\x18\x07 \x03(\x02\x12\x0f\n\x07strings\x18\x08 \x03(\t\x12\t\n\x01\x62\x18\n \x01(\x08\x12\r\n\x05\x62ools\x18\x0b \x03(\x08\x12\x11\n\tblock_idx\x18\x0c \x01(\x05\x12\t\n\x01l\x18\r \x01(\x03\x1a+\n\x03Var\x12\x11\n\tparameter\x18\x01 \x02(\t\x12\x11\n\targuments\x18\x02 \x03(\t\"\xcf\x03\n\x07OpProto\x12\x0c\n\x04type\x18\x01 \x02(\t\x12:\n\x06inputs\x18\x02 \x03(\x0b\x32*.paddle_mobile.framework.proto.OpProto.Var\x12;\n\x07outputs\x18\x03 \x03(\x0b\x32*.paddle_mobile.framework.proto.OpProto.Var\x12:\n\x05\x61ttrs\x18\x04 \x03(\x0b\x32+.paddle_mobile.framework.proto.OpProto.Attr\x12\x0f\n\x07\x63omment\x18\x05 \x02(\t\x1ax\n\x03Var\x12\x0c\n\x04name\x18\x01 \x02(\t\x12\x0f\n\x07\x63omment\x18\x02 \x02(\t\x12\x19\n\nduplicable\x18\x03 \x01(\x08:\x05\x66\x61lse\x12\x1b\n\x0cintermediate\x18\x04 \x01(\x08:\x05\x66\x61lse\x12\x1a\n\x0b\x64ispensable\x18\x05 \x01(\x08:\x05\x66\x61lse\x1av\n\x04\x41ttr\x12\x0c\n\x04name\x18\x01 \x02(\t\x12\x35\n\x04type\x18\x02 \x02(\x0e\x32\'.paddle_mobile.framework.proto.AttrType\x12\x0f\n\x07\x63omment\x18\x03 \x02(\t\x12\x18\n\tgenerated\x18\x04 \x01(\x08:\x05\x66\x61lse\"\xb9\n\n\x07VarType\x12\x39\n\x04type\x18\x01 \x02(\x0e\x32+.paddle_mobile.framework.proto.VarType.Type\x12H\n\rselected_rows\x18\x02 \x01(\x0b\x32\x31.paddle_mobile.framework.proto.VarType.TensorDesc\x12H\n\nlod_tensor\x18\x03 \x01(\x0b\x32\x34.paddle_mobile.framework.proto.VarType.LoDTensorDesc\x12O\n\x0ctensor_array\x18\x04 \x01(\x0b\x32\x39.paddle_mobile.framework.proto.VarType.LoDTensorArrayDesc\x12\x41\n\x06reader\x18\x05 \x01(\x0b\x32\x31.paddle_mobile.framework.proto.VarType.ReaderDesc\x12\x43\n\x07\x63hannel\x18\x06 \x01(\x0b\x32\x32.paddle_mobile.framework.proto.VarType.ChannelDesc\x12;\n\x05tuple\x18\x07 \x01(\x0b\x32,.paddle_mobile.framework.proto.VarType.Tuple\x1aZ\n\nTensorDesc\x12>\n\tdata_type\x18\x01 \x02(\x0e\x32+.paddle_mobile.framework.proto.VarType.Type\x12\x0c\n\x04\x64ims\x18\x02 \x03(\x03\x1ah\n\rLoDTensorDesc\x12\x41\n\x06tensor\x18\x01 \x02(\x0b\x32\x31.paddle_mobile.framework.proto.VarType.TensorDesc\x12\x14\n\tlod_level\x18\x02 \x01(\x05:\x01\x30\x1am\n\x12LoDTensorArrayDesc\x12\x41\n\x06tensor\x18\x01 \x02(\x0b\x32\x31.paddle_mobile.framework.proto.VarType.TensorDesc\x12\x14\n\tlod_level\x18\x02 \x01(\x05:\x01\x30\x1aV\n\nReaderDesc\x12H\n\nlod_tensor\x18\x01 \x03(\x0b\x32\x34.paddle_mobile.framework.proto.VarType.LoDTensorDesc\x1a_\n\x0b\x43hannelDesc\x12>\n\tdata_type\x18\x01 \x02(\x0e\x32+.paddle_mobile.framework.proto.VarType.Type\x12\x10\n\x08\x63\x61pacity\x18\x02 \x02(\x03\x1aJ\n\x05Tuple\x12\x41\n\x0c\x65lement_type\x18\x01 \x03(\x0e\x32+.paddle_mobile.framework.proto.VarType.Type\"\x8e\x02\n\x04Type\x12\x08\n\x04\x42OOL\x10\x00\x12\t\n\x05INT16\x10\x01\x12\t\n\x05INT32\x10\x02\x12\t\n\x05INT64\x10\x03\x12\x08\n\x04\x46P16\x10\x04\x12\x08\n\x04\x46P32\x10\x05\x12\x08\n\x04\x46P64\x10\x06\x12\x0e\n\nLOD_TENSOR\x10\x07\x12\x11\n\rSELECTED_ROWS\x10\x08\x12\x12\n\x0e\x46\x45\x45\x44_MINIBATCH\x10\t\x12\x0e\n\nFETCH_LIST\x10\n\x12\x0f\n\x0bSTEP_SCOPES\x10\x0b\x12\x12\n\x0eLOD_RANK_TABLE\x10\x0c\x12\x14\n\x10LOD_TENSOR_ARRAY\x10\r\x12\x0e\n\nPLACE_LIST\x10\x0e\x12\n\n\x06READER\x10\x0f\x12\x0b\n\x07\x43HANNEL\x10\x10\x12\x07\n\x03RAW\x10\x11\x12\t\n\x05TUPLE\x10\x12\"i\n\x07VarDesc\x12\x0c\n\x04name\x18\x01 \x02(\t\x12\x34\n\x04type\x18\x02 \x02(\x0b\x32&.paddle_mobile.framework.proto.VarType\x12\x1a\n\x0bpersistable\x18\x03 \x01(\x08:\x05\x66\x61lse\"\xb5\x01\n\tBlockDesc\x12\x0b\n\x03idx\x18\x01 \x02(\x05\x12\x12\n\nparent_idx\x18\x02 \x02(\x05\x12\x34\n\x04vars\x18\x03 \x03(\x0b\x32&.paddle_mobile.framework.proto.VarDesc\x12\x32\n\x03ops\x18\x04 \x03(\x0b\x32%.paddle_mobile.framework.proto.OpDesc\x12\x1d\n\x11\x66orward_block_idx\x18\x05 \x01(\x05:\x02-1\"G\n\x0bProgramDesc\x12\x38\n\x06\x62locks\x18\x01 \x03(\x0b\x32(.paddle_mobile.framework.proto.BlockDesc*}\n\x08\x41ttrType\x12\x07\n\x03INT\x10\x00\x12\t\n\x05\x46LOAT\x10\x01\x12\n\n\x06STRING\x10\x02\x12\x08\n\x04INTS\x10\x03\x12\n\n\x06\x46LOATS\x10\x04\x12\x0b\n\x07STRINGS\x10\x05\x12\x0b\n\x07\x42OOLEAN\x10\x06\x12\x0c\n\x08\x42OOLEANS\x10\x07\x12\t\n\x05\x42LOCK\x10\x08\x12\x08\n\x04LONG\x10\tB\x02H\x03')
+)
+_sym_db.RegisterFileDescriptor(DESCRIPTOR)
+
+_ATTRTYPE = _descriptor.EnumDescriptor(
+  name='AttrType',
+  full_name='paddle_mobile.framework.proto.AttrType',
+  filename=None,
+  file=DESCRIPTOR,
+  values=[
+    _descriptor.EnumValueDescriptor(
+      name='INT', index=0, number=0,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='FLOAT', index=1, number=1,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='STRING', index=2, number=2,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='INTS', index=3, number=3,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='FLOATS', index=4, number=4,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='STRINGS', index=5, number=5,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='BOOLEAN', index=6, number=6,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='BOOLEANS', index=7, number=7,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='BLOCK', index=8, number=8,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='LONG', index=9, number=9,
+      options=None,
+      type=None),
+  ],
+  containing_type=None,
+  options=None,
+  serialized_start=2708,
+  serialized_end=2833,
+)
+_sym_db.RegisterEnumDescriptor(_ATTRTYPE)
+
+AttrType = enum_type_wrapper.EnumTypeWrapper(_ATTRTYPE)
+INT = 0
+FLOAT = 1
+STRING = 2
+INTS = 3
+FLOATS = 4
+STRINGS = 5
+BOOLEAN = 6
+BOOLEANS = 7
+BLOCK = 8
+LONG = 9
+
+
+_VARTYPE_TYPE = _descriptor.EnumDescriptor(
+  name='Type',
+  full_name='paddle_mobile.framework.proto.VarType.Type',
+  filename=None,
+  file=DESCRIPTOR,
+  values=[
+    _descriptor.EnumValueDescriptor(
+      name='BOOL', index=0, number=0,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='INT16', index=1, number=1,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='INT32', index=2, number=2,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='INT64', index=3, number=3,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='FP16', index=4, number=4,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='FP32', index=5, number=5,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='FP64', index=6, number=6,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='LOD_TENSOR', index=7, number=7,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='SELECTED_ROWS', index=8, number=8,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='FEED_MINIBATCH', index=9, number=9,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='FETCH_LIST', index=10, number=10,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='STEP_SCOPES', index=11, number=11,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='LOD_RANK_TABLE', index=12, number=12,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='LOD_TENSOR_ARRAY', index=13, number=13,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='PLACE_LIST', index=14, number=14,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='READER', index=15, number=15,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='CHANNEL', index=16, number=16,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='RAW', index=17, number=17,
+      options=None,
+      type=None),
+    _descriptor.EnumValueDescriptor(
+      name='TUPLE', index=18, number=18,
+      options=None,
+      type=None),
+  ],
+  containing_type=None,
+  options=None,
+  serialized_start=2072,
+  serialized_end=2342,
+)
+_sym_db.RegisterEnumDescriptor(_VARTYPE_TYPE)
+
+
+_OPDESC_ATTR = _descriptor.Descriptor(
+  name='Attr',
+  full_name='paddle_mobile.framework.proto.OpDesc.Attr',
+  filename=None,
+  file=DESCRIPTOR,
+  containing_type=None,
+  fields=[
+    _descriptor.FieldDescriptor(
+      name='name', full_name='paddle_mobile.framework.proto.OpDesc.Attr.name', index=0,
+      number=1, type=9, cpp_type=9, label=2,
+      has_default_value=False, default_value=_b("").decode('utf-8'),
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='type', full_name='paddle_mobile.framework.proto.OpDesc.Attr.type', index=1,
+      number=2, type=14, cpp_type=8, label=2,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='i', full_name='paddle_mobile.framework.proto.OpDesc.Attr.i', index=2,
+      number=3, type=5, cpp_type=1, label=1,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='f', full_name='paddle_mobile.framework.proto.OpDesc.Attr.f', index=3,
+      number=4, type=2, cpp_type=6, label=1,
+      has_default_value=False, default_value=float(0),
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='s', full_name='paddle_mobile.framework.proto.OpDesc.Attr.s', index=4,
+      number=5, type=9, cpp_type=9, label=1,
+      has_default_value=False, default_value=_b("").decode('utf-8'),
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='ints', full_name='paddle_mobile.framework.proto.OpDesc.Attr.ints', index=5,
+      number=6, type=5, cpp_type=1, label=3,
+      has_default_value=False, default_value=[],
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='floats', full_name='paddle_mobile.framework.proto.OpDesc.Attr.floats', index=6,
+      number=7, type=2, cpp_type=6, label=3,
+      has_default_value=False, default_value=[],
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='strings', full_name='paddle_mobile.framework.proto.OpDesc.Attr.strings', index=7,
+      number=8, type=9, cpp_type=9, label=3,
+      has_default_value=False, default_value=[],
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='b', full_name='paddle_mobile.framework.proto.OpDesc.Attr.b', index=8,
+      number=10, type=8, cpp_type=7, label=1,
+      has_default_value=False, default_value=False,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='bools', full_name='paddle_mobile.framework.proto.OpDesc.Attr.bools', index=9,
+      number=11, type=8, cpp_type=7, label=3,
+      has_default_value=False, default_value=[],
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='block_idx', full_name='paddle_mobile.framework.proto.OpDesc.Attr.block_idx', index=10,
+      number=12, type=5, cpp_type=1, label=1,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='l', full_name='paddle_mobile.framework.proto.OpDesc.Attr.l', index=11,
+      number=13, type=3, cpp_type=2, label=1,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+  ],
+  extensions=[
+  ],
+  nested_types=[],
+  enum_types=[
+  ],
+  options=None,
+  is_extendable=False,
+  syntax='proto2',
+  extension_ranges=[],
+  oneofs=[
+  ],
+  serialized_start=280,
+  serialized_end=491,
+)
+
+_OPDESC_VAR = _descriptor.Descriptor(
+  name='Var',
+  full_name='paddle_mobile.framework.proto.OpDesc.Var',
+  filename=None,
+  file=DESCRIPTOR,
+  containing_type=None,
+  fields=[
+    _descriptor.FieldDescriptor(
+      name='parameter', full_name='paddle_mobile.framework.proto.OpDesc.Var.parameter', index=0,
+      number=1, type=9, cpp_type=9, label=2,
+      has_default_value=False, default_value=_b("").decode('utf-8'),
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='arguments', full_name='paddle_mobile.framework.proto.OpDesc.Var.arguments', index=1,
+      number=2, type=9, cpp_type=9, label=3,
+      has_default_value=False, default_value=[],
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+  ],
+  extensions=[
+  ],
+  nested_types=[],
+  enum_types=[
+  ],
+  options=None,
+  is_extendable=False,
+  syntax='proto2',
+  extension_ranges=[],
+  oneofs=[
+  ],
+  serialized_start=493,
+  serialized_end=536,
+)
+
+_OPDESC = _descriptor.Descriptor(
+  name='OpDesc',
+  full_name='paddle_mobile.framework.proto.OpDesc',
+  filename=None,
+  file=DESCRIPTOR,
+  containing_type=None,
+  fields=[
+    _descriptor.FieldDescriptor(
+      name='type', full_name='paddle_mobile.framework.proto.OpDesc.type', index=0,
+      number=3, type=9, cpp_type=9, label=2,
+      has_default_value=False, default_value=_b("").decode('utf-8'),
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='inputs', full_name='paddle_mobile.framework.proto.OpDesc.inputs', index=1,
+      number=1, type=11, cpp_type=10, label=3,
+      has_default_value=False, default_value=[],
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='outputs', full_name='paddle_mobile.framework.proto.OpDesc.outputs', index=2,
+      number=2, type=11, cpp_type=10, label=3,
+      has_default_value=False, default_value=[],
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='attrs', full_name='paddle_mobile.framework.proto.OpDesc.attrs', index=3,
+      number=4, type=11, cpp_type=10, label=3,
+      has_default_value=False, default_value=[],
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='is_target', full_name='paddle_mobile.framework.proto.OpDesc.is_target', index=4,
+      number=5, type=8, cpp_type=7, label=1,
+      has_default_value=True, default_value=False,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+  ],
+  extensions=[
+  ],
+  nested_types=[_OPDESC_ATTR, _OPDESC_VAR, ],
+  enum_types=[
+  ],
+  options=None,
+  is_extendable=False,
+  syntax='proto2',
+  extension_ranges=[],
+  oneofs=[
+  ],
+  serialized_start=51,
+  serialized_end=536,
+)
+
+
+_OPPROTO_VAR = _descriptor.Descriptor(
+  name='Var',
+  full_name='paddle_mobile.framework.proto.OpProto.Var',
+  filename=None,
+  file=DESCRIPTOR,
+  containing_type=None,
+  fields=[
+    _descriptor.FieldDescriptor(
+      name='name', full_name='paddle_mobile.framework.proto.OpProto.Var.name', index=0,
+      number=1, type=9, cpp_type=9, label=2,
+      has_default_value=False, default_value=_b("").decode('utf-8'),
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='comment', full_name='paddle_mobile.framework.proto.OpProto.Var.comment', index=1,
+      number=2, type=9, cpp_type=9, label=2,
+      has_default_value=False, default_value=_b("").decode('utf-8'),
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='duplicable', full_name='paddle_mobile.framework.proto.OpProto.Var.duplicable', index=2,
+      number=3, type=8, cpp_type=7, label=1,
+      has_default_value=True, default_value=False,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='intermediate', full_name='paddle_mobile.framework.proto.OpProto.Var.intermediate', index=3,
+      number=4, type=8, cpp_type=7, label=1,
+      has_default_value=True, default_value=False,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='dispensable', full_name='paddle_mobile.framework.proto.OpProto.Var.dispensable', index=4,
+      number=5, type=8, cpp_type=7, label=1,
+      has_default_value=True, default_value=False,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+  ],
+  extensions=[
+  ],
+  nested_types=[],
+  enum_types=[
+  ],
+  options=None,
+  is_extendable=False,
+  syntax='proto2',
+  extension_ranges=[],
+  oneofs=[
+  ],
+  serialized_start=762,
+  serialized_end=882,
+)
+
+_OPPROTO_ATTR = _descriptor.Descriptor(
+  name='Attr',
+  full_name='paddle_mobile.framework.proto.OpProto.Attr',
+  filename=None,
+  file=DESCRIPTOR,
+  containing_type=None,
+  fields=[
+    _descriptor.FieldDescriptor(
+      name='name', full_name='paddle_mobile.framework.proto.OpProto.Attr.name', index=0,
+      number=1, type=9, cpp_type=9, label=2,
+      has_default_value=False, default_value=_b("").decode('utf-8'),
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='type', full_name='paddle_mobile.framework.proto.OpProto.Attr.type', index=1,
+      number=2, type=14, cpp_type=8, label=2,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='comment', full_name='paddle_mobile.framework.proto.OpProto.Attr.comment', index=2,
+      number=3, type=9, cpp_type=9, label=2,
+      has_default_value=False, default_value=_b("").decode('utf-8'),
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='generated', full_name='paddle_mobile.framework.proto.OpProto.Attr.generated', index=3,
+      number=4, type=8, cpp_type=7, label=1,
+      has_default_value=True, default_value=False,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+  ],
+  extensions=[
+  ],
+  nested_types=[],
+  enum_types=[
+  ],
+  options=None,
+  is_extendable=False,
+  syntax='proto2',
+  extension_ranges=[],
+  oneofs=[
+  ],
+  serialized_start=884,
+  serialized_end=1002,
+)
+
+_OPPROTO = _descriptor.Descriptor(
+  name='OpProto',
+  full_name='paddle_mobile.framework.proto.OpProto',
+  filename=None,
+  file=DESCRIPTOR,
+  containing_type=None,
+  fields=[
+    _descriptor.FieldDescriptor(
+      name='type', full_name='paddle_mobile.framework.proto.OpProto.type', index=0,
+      number=1, type=9, cpp_type=9, label=2,
+      has_default_value=False, default_value=_b("").decode('utf-8'),
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='inputs', full_name='paddle_mobile.framework.proto.OpProto.inputs', index=1,
+      number=2, type=11, cpp_type=10, label=3,
+      has_default_value=False, default_value=[],
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='outputs', full_name='paddle_mobile.framework.proto.OpProto.outputs', index=2,
+      number=3, type=11, cpp_type=10, label=3,
+      has_default_value=False, default_value=[],
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='attrs', full_name='paddle_mobile.framework.proto.OpProto.attrs', index=3,
+      number=4, type=11, cpp_type=10, label=3,
+      has_default_value=False, default_value=[],
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='comment', full_name='paddle_mobile.framework.proto.OpProto.comment', index=4,
+      number=5, type=9, cpp_type=9, label=2,
+      has_default_value=False, default_value=_b("").decode('utf-8'),
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+  ],
+  extensions=[
+  ],
+  nested_types=[_OPPROTO_VAR, _OPPROTO_ATTR, ],
+  enum_types=[
+  ],
+  options=None,
+  is_extendable=False,
+  syntax='proto2',
+  extension_ranges=[],
+  oneofs=[
+  ],
+  serialized_start=539,
+  serialized_end=1002,
+)
+
+
+_VARTYPE_TENSORDESC = _descriptor.Descriptor(
+  name='TensorDesc',
+  full_name='paddle_mobile.framework.proto.VarType.TensorDesc',
+  filename=None,
+  file=DESCRIPTOR,
+  containing_type=None,
+  fields=[
+    _descriptor.FieldDescriptor(
+      name='data_type', full_name='paddle_mobile.framework.proto.VarType.TensorDesc.data_type', index=0,
+      number=1, type=14, cpp_type=8, label=2,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='dims', full_name='paddle_mobile.framework.proto.VarType.TensorDesc.dims', index=1,
+      number=2, type=3, cpp_type=2, label=3,
+      has_default_value=False, default_value=[],
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+  ],
+  extensions=[
+  ],
+  nested_types=[],
+  enum_types=[
+  ],
+  options=None,
+  is_extendable=False,
+  syntax='proto2',
+  extension_ranges=[],
+  oneofs=[
+  ],
+  serialized_start=1501,
+  serialized_end=1591,
+)
+
+_VARTYPE_LODTENSORDESC = _descriptor.Descriptor(
+  name='LoDTensorDesc',
+  full_name='paddle_mobile.framework.proto.VarType.LoDTensorDesc',
+  filename=None,
+  file=DESCRIPTOR,
+  containing_type=None,
+  fields=[
+    _descriptor.FieldDescriptor(
+      name='tensor', full_name='paddle_mobile.framework.proto.VarType.LoDTensorDesc.tensor', index=0,
+      number=1, type=11, cpp_type=10, label=2,
+      has_default_value=False, default_value=None,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='lod_level', full_name='paddle_mobile.framework.proto.VarType.LoDTensorDesc.lod_level', index=1,
+      number=2, type=5, cpp_type=1, label=1,
+      has_default_value=True, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+  ],
+  extensions=[
+  ],
+  nested_types=[],
+  enum_types=[
+  ],
+  options=None,
+  is_extendable=False,
+  syntax='proto2',
+  extension_ranges=[],
+  oneofs=[
+  ],
+  serialized_start=1593,
+  serialized_end=1697,
+)
+
+_VARTYPE_LODTENSORARRAYDESC = _descriptor.Descriptor(
+  name='LoDTensorArrayDesc',
+  full_name='paddle_mobile.framework.proto.VarType.LoDTensorArrayDesc',
+  filename=None,
+  file=DESCRIPTOR,
+  containing_type=None,
+  fields=[
+    _descriptor.FieldDescriptor(
+      name='tensor', full_name='paddle_mobile.framework.proto.VarType.LoDTensorArrayDesc.tensor', index=0,
+      number=1, type=11, cpp_type=10, label=2,
+      has_default_value=False, default_value=None,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='lod_level', full_name='paddle_mobile.framework.proto.VarType.LoDTensorArrayDesc.lod_level', index=1,
+      number=2, type=5, cpp_type=1, label=1,
+      has_default_value=True, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+  ],
+  extensions=[
+  ],
+  nested_types=[],
+  enum_types=[
+  ],
+  options=None,
+  is_extendable=False,
+  syntax='proto2',
+  extension_ranges=[],
+  oneofs=[
+  ],
+  serialized_start=1699,
+  serialized_end=1808,
+)
+
+_VARTYPE_READERDESC = _descriptor.Descriptor(
+  name='ReaderDesc',
+  full_name='paddle_mobile.framework.proto.VarType.ReaderDesc',
+  filename=None,
+  file=DESCRIPTOR,
+  containing_type=None,
+  fields=[
+    _descriptor.FieldDescriptor(
+      name='lod_tensor', full_name='paddle_mobile.framework.proto.VarType.ReaderDesc.lod_tensor', index=0,
+      number=1, type=11, cpp_type=10, label=3,
+      has_default_value=False, default_value=[],
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+  ],
+  extensions=[
+  ],
+  nested_types=[],
+  enum_types=[
+  ],
+  options=None,
+  is_extendable=False,
+  syntax='proto2',
+  extension_ranges=[],
+  oneofs=[
+  ],
+  serialized_start=1810,
+  serialized_end=1896,
+)
+
+_VARTYPE_CHANNELDESC = _descriptor.Descriptor(
+  name='ChannelDesc',
+  full_name='paddle_mobile.framework.proto.VarType.ChannelDesc',
+  filename=None,
+  file=DESCRIPTOR,
+  containing_type=None,
+  fields=[
+    _descriptor.FieldDescriptor(
+      name='data_type', full_name='paddle_mobile.framework.proto.VarType.ChannelDesc.data_type', index=0,
+      number=1, type=14, cpp_type=8, label=2,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='capacity', full_name='paddle_mobile.framework.proto.VarType.ChannelDesc.capacity', index=1,
+      number=2, type=3, cpp_type=2, label=2,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+  ],
+  extensions=[
+  ],
+  nested_types=[],
+  enum_types=[
+  ],
+  options=None,
+  is_extendable=False,
+  syntax='proto2',
+  extension_ranges=[],
+  oneofs=[
+  ],
+  serialized_start=1898,
+  serialized_end=1993,
+)
+
+_VARTYPE_TUPLE = _descriptor.Descriptor(
+  name='Tuple',
+  full_name='paddle_mobile.framework.proto.VarType.Tuple',
+  filename=None,
+  file=DESCRIPTOR,
+  containing_type=None,
+  fields=[
+    _descriptor.FieldDescriptor(
+      name='element_type', full_name='paddle_mobile.framework.proto.VarType.Tuple.element_type', index=0,
+      number=1, type=14, cpp_type=8, label=3,
+      has_default_value=False, default_value=[],
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+  ],
+  extensions=[
+  ],
+  nested_types=[],
+  enum_types=[
+  ],
+  options=None,
+  is_extendable=False,
+  syntax='proto2',
+  extension_ranges=[],
+  oneofs=[
+  ],
+  serialized_start=1995,
+  serialized_end=2069,
+)
+
+_VARTYPE = _descriptor.Descriptor(
+  name='VarType',
+  full_name='paddle_mobile.framework.proto.VarType',
+  filename=None,
+  file=DESCRIPTOR,
+  containing_type=None,
+  fields=[
+    _descriptor.FieldDescriptor(
+      name='type', full_name='paddle_mobile.framework.proto.VarType.type', index=0,
+      number=1, type=14, cpp_type=8, label=2,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='selected_rows', full_name='paddle_mobile.framework.proto.VarType.selected_rows', index=1,
+      number=2, type=11, cpp_type=10, label=1,
+      has_default_value=False, default_value=None,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='lod_tensor', full_name='paddle_mobile.framework.proto.VarType.lod_tensor', index=2,
+      number=3, type=11, cpp_type=10, label=1,
+      has_default_value=False, default_value=None,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='tensor_array', full_name='paddle_mobile.framework.proto.VarType.tensor_array', index=3,
+      number=4, type=11, cpp_type=10, label=1,
+      has_default_value=False, default_value=None,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='reader', full_name='paddle_mobile.framework.proto.VarType.reader', index=4,
+      number=5, type=11, cpp_type=10, label=1,
+      has_default_value=False, default_value=None,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='channel', full_name='paddle_mobile.framework.proto.VarType.channel', index=5,
+      number=6, type=11, cpp_type=10, label=1,
+      has_default_value=False, default_value=None,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='tuple', full_name='paddle_mobile.framework.proto.VarType.tuple', index=6,
+      number=7, type=11, cpp_type=10, label=1,
+      has_default_value=False, default_value=None,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+  ],
+  extensions=[
+  ],
+  nested_types=[_VARTYPE_TENSORDESC, _VARTYPE_LODTENSORDESC, _VARTYPE_LODTENSORARRAYDESC, _VARTYPE_READERDESC, _VARTYPE_CHANNELDESC, _VARTYPE_TUPLE, ],
+  enum_types=[
+    _VARTYPE_TYPE,
+  ],
+  options=None,
+  is_extendable=False,
+  syntax='proto2',
+  extension_ranges=[],
+  oneofs=[
+  ],
+  serialized_start=1005,
+  serialized_end=2342,
+)
+
+
+_VARDESC = _descriptor.Descriptor(
+  name='VarDesc',
+  full_name='paddle_mobile.framework.proto.VarDesc',
+  filename=None,
+  file=DESCRIPTOR,
+  containing_type=None,
+  fields=[
+    _descriptor.FieldDescriptor(
+      name='name', full_name='paddle_mobile.framework.proto.VarDesc.name', index=0,
+      number=1, type=9, cpp_type=9, label=2,
+      has_default_value=False, default_value=_b("").decode('utf-8'),
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='type', full_name='paddle_mobile.framework.proto.VarDesc.type', index=1,
+      number=2, type=11, cpp_type=10, label=2,
+      has_default_value=False, default_value=None,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='persistable', full_name='paddle_mobile.framework.proto.VarDesc.persistable', index=2,
+      number=3, type=8, cpp_type=7, label=1,
+      has_default_value=True, default_value=False,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+  ],
+  extensions=[
+  ],
+  nested_types=[],
+  enum_types=[
+  ],
+  options=None,
+  is_extendable=False,
+  syntax='proto2',
+  extension_ranges=[],
+  oneofs=[
+  ],
+  serialized_start=2344,
+  serialized_end=2449,
+)
+
+
+_BLOCKDESC = _descriptor.Descriptor(
+  name='BlockDesc',
+  full_name='paddle_mobile.framework.proto.BlockDesc',
+  filename=None,
+  file=DESCRIPTOR,
+  containing_type=None,
+  fields=[
+    _descriptor.FieldDescriptor(
+      name='idx', full_name='paddle_mobile.framework.proto.BlockDesc.idx', index=0,
+      number=1, type=5, cpp_type=1, label=2,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='parent_idx', full_name='paddle_mobile.framework.proto.BlockDesc.parent_idx', index=1,
+      number=2, type=5, cpp_type=1, label=2,
+      has_default_value=False, default_value=0,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='vars', full_name='paddle_mobile.framework.proto.BlockDesc.vars', index=2,
+      number=3, type=11, cpp_type=10, label=3,
+      has_default_value=False, default_value=[],
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='ops', full_name='paddle_mobile.framework.proto.BlockDesc.ops', index=3,
+      number=4, type=11, cpp_type=10, label=3,
+      has_default_value=False, default_value=[],
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+    _descriptor.FieldDescriptor(
+      name='forward_block_idx', full_name='paddle_mobile.framework.proto.BlockDesc.forward_block_idx', index=4,
+      number=5, type=5, cpp_type=1, label=1,
+      has_default_value=True, default_value=-1,
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+  ],
+  extensions=[
+  ],
+  nested_types=[],
+  enum_types=[
+  ],
+  options=None,
+  is_extendable=False,
+  syntax='proto2',
+  extension_ranges=[],
+  oneofs=[
+  ],
+  serialized_start=2452,
+  serialized_end=2633,
+)
+
+
+_PROGRAMDESC = _descriptor.Descriptor(
+  name='ProgramDesc',
+  full_name='paddle_mobile.framework.proto.ProgramDesc',
+  filename=None,
+  file=DESCRIPTOR,
+  containing_type=None,
+  fields=[
+    _descriptor.FieldDescriptor(
+      name='blocks', full_name='paddle_mobile.framework.proto.ProgramDesc.blocks', index=0,
+      number=1, type=11, cpp_type=10, label=3,
+      has_default_value=False, default_value=[],
+      message_type=None, enum_type=None, containing_type=None,
+      is_extension=False, extension_scope=None,
+      options=None),
+  ],
+  extensions=[
+  ],
+  nested_types=[],
+  enum_types=[
+  ],
+  options=None,
+  is_extendable=False,
+  syntax='proto2',
+  extension_ranges=[],
+  oneofs=[
+  ],
+  serialized_start=2635,
+  serialized_end=2706,
+)
+
+_OPDESC_ATTR.fields_by_name['type'].enum_type = _ATTRTYPE
+_OPDESC_ATTR.containing_type = _OPDESC
+_OPDESC_VAR.containing_type = _OPDESC
+_OPDESC.fields_by_name['inputs'].message_type = _OPDESC_VAR
+_OPDESC.fields_by_name['outputs'].message_type = _OPDESC_VAR
+_OPDESC.fields_by_name['attrs'].message_type = _OPDESC_ATTR
+_OPPROTO_VAR.containing_type = _OPPROTO
+_OPPROTO_ATTR.fields_by_name['type'].enum_type = _ATTRTYPE
+_OPPROTO_ATTR.containing_type = _OPPROTO
+_OPPROTO.fields_by_name['inputs'].message_type = _OPPROTO_VAR
+_OPPROTO.fields_by_name['outputs'].message_type = _OPPROTO_VAR
+_OPPROTO.fields_by_name['attrs'].message_type = _OPPROTO_ATTR
+_VARTYPE_TENSORDESC.fields_by_name['data_type'].enum_type = _VARTYPE_TYPE
+_VARTYPE_TENSORDESC.containing_type = _VARTYPE
+_VARTYPE_LODTENSORDESC.fields_by_name['tensor'].message_type = _VARTYPE_TENSORDESC
+_VARTYPE_LODTENSORDESC.containing_type = _VARTYPE
+_VARTYPE_LODTENSORARRAYDESC.fields_by_name['tensor'].message_type = _VARTYPE_TENSORDESC
+_VARTYPE_LODTENSORARRAYDESC.containing_type = _VARTYPE
+_VARTYPE_READERDESC.fields_by_name['lod_tensor'].message_type = _VARTYPE_LODTENSORDESC
+_VARTYPE_READERDESC.containing_type = _VARTYPE
+_VARTYPE_CHANNELDESC.fields_by_name['data_type'].enum_type = _VARTYPE_TYPE
+_VARTYPE_CHANNELDESC.containing_type = _VARTYPE
+_VARTYPE_TUPLE.fields_by_name['element_type'].enum_type = _VARTYPE_TYPE
+_VARTYPE_TUPLE.containing_type = _VARTYPE
+_VARTYPE.fields_by_name['type'].enum_type = _VARTYPE_TYPE
+_VARTYPE.fields_by_name['selected_rows'].message_type = _VARTYPE_TENSORDESC
+_VARTYPE.fields_by_name['lod_tensor'].message_type = _VARTYPE_LODTENSORDESC
+_VARTYPE.fields_by_name['tensor_array'].message_type = _VARTYPE_LODTENSORARRAYDESC
+_VARTYPE.fields_by_name['reader'].message_type = _VARTYPE_READERDESC
+_VARTYPE.fields_by_name['channel'].message_type = _VARTYPE_CHANNELDESC
+_VARTYPE.fields_by_name['tuple'].message_type = _VARTYPE_TUPLE
+_VARTYPE_TYPE.containing_type = _VARTYPE
+_VARDESC.fields_by_name['type'].message_type = _VARTYPE
+_BLOCKDESC.fields_by_name['vars'].message_type = _VARDESC
+_BLOCKDESC.fields_by_name['ops'].message_type = _OPDESC
+_PROGRAMDESC.fields_by_name['blocks'].message_type = _BLOCKDESC
+DESCRIPTOR.message_types_by_name['OpDesc'] = _OPDESC
+DESCRIPTOR.message_types_by_name['OpProto'] = _OPPROTO
+DESCRIPTOR.message_types_by_name['VarType'] = _VARTYPE
+DESCRIPTOR.message_types_by_name['VarDesc'] = _VARDESC
+DESCRIPTOR.message_types_by_name['BlockDesc'] = _BLOCKDESC
+DESCRIPTOR.message_types_by_name['ProgramDesc'] = _PROGRAMDESC
+DESCRIPTOR.enum_types_by_name['AttrType'] = _ATTRTYPE
+
+OpDesc = _reflection.GeneratedProtocolMessageType('OpDesc', (_message.Message,), dict(
+
+  Attr = _reflection.GeneratedProtocolMessageType('Attr', (_message.Message,), dict(
+    DESCRIPTOR = _OPDESC_ATTR,
+    __module__ = 'framework_pb2'
+    # @@protoc_insertion_point(class_scope:paddle_mobile.framework.proto.OpDesc.Attr)
+    ))
+  ,
+
+  Var = _reflection.GeneratedProtocolMessageType('Var', (_message.Message,), dict(
+    DESCRIPTOR = _OPDESC_VAR,
+    __module__ = 'framework_pb2'
+    # @@protoc_insertion_point(class_scope:paddle_mobile.framework.proto.OpDesc.Var)
+    ))
+  ,
+  DESCRIPTOR = _OPDESC,
+  __module__ = 'framework_pb2'
+  # @@protoc_insertion_point(class_scope:paddle_mobile.framework.proto.OpDesc)
+  ))
+_sym_db.RegisterMessage(OpDesc)
+_sym_db.RegisterMessage(OpDesc.Attr)
+_sym_db.RegisterMessage(OpDesc.Var)
+
+OpProto = _reflection.GeneratedProtocolMessageType('OpProto', (_message.Message,), dict(
+
+  Var = _reflection.GeneratedProtocolMessageType('Var', (_message.Message,), dict(
+    DESCRIPTOR = _OPPROTO_VAR,
+    __module__ = 'framework_pb2'
+    # @@protoc_insertion_point(class_scope:paddle_mobile.framework.proto.OpProto.Var)
+    ))
+  ,
+
+  Attr = _reflection.GeneratedProtocolMessageType('Attr', (_message.Message,), dict(
+    DESCRIPTOR = _OPPROTO_ATTR,
+    __module__ = 'framework_pb2'
+    # @@protoc_insertion_point(class_scope:paddle_mobile.framework.proto.OpProto.Attr)
+    ))
+  ,
+  DESCRIPTOR = _OPPROTO,
+  __module__ = 'framework_pb2'
+  # @@protoc_insertion_point(class_scope:paddle_mobile.framework.proto.OpProto)
+  ))
+_sym_db.RegisterMessage(OpProto)
+_sym_db.RegisterMessage(OpProto.Var)
+_sym_db.RegisterMessage(OpProto.Attr)
+
+VarType = _reflection.GeneratedProtocolMessageType('VarType', (_message.Message,), dict(
+
+  TensorDesc = _reflection.GeneratedProtocolMessageType('TensorDesc', (_message.Message,), dict(
+    DESCRIPTOR = _VARTYPE_TENSORDESC,
+    __module__ = 'framework_pb2'
+    # @@protoc_insertion_point(class_scope:paddle_mobile.framework.proto.VarType.TensorDesc)
+    ))
+  ,
+
+  LoDTensorDesc = _reflection.GeneratedProtocolMessageType('LoDTensorDesc', (_message.Message,), dict(
+    DESCRIPTOR = _VARTYPE_LODTENSORDESC,
+    __module__ = 'framework_pb2'
+    # @@protoc_insertion_point(class_scope:paddle_mobile.framework.proto.VarType.LoDTensorDesc)
+    ))
+  ,
+
+  LoDTensorArrayDesc = _reflection.GeneratedProtocolMessageType('LoDTensorArrayDesc', (_message.Message,), dict(
+    DESCRIPTOR = _VARTYPE_LODTENSORARRAYDESC,
+    __module__ = 'framework_pb2'
+    # @@protoc_insertion_point(class_scope:paddle_mobile.framework.proto.VarType.LoDTensorArrayDesc)
+    ))
+  ,
+
+  ReaderDesc = _reflection.GeneratedProtocolMessageType('ReaderDesc', (_message.Message,), dict(
+    DESCRIPTOR = _VARTYPE_READERDESC,
+    __module__ = 'framework_pb2'
+    # @@protoc_insertion_point(class_scope:paddle_mobile.framework.proto.VarType.ReaderDesc)
+    ))
+  ,
+
+  ChannelDesc = _reflection.GeneratedProtocolMessageType('ChannelDesc', (_message.Message,), dict(
+    DESCRIPTOR = _VARTYPE_CHANNELDESC,
+    __module__ = 'framework_pb2'
+    # @@protoc_insertion_point(class_scope:paddle_mobile.framework.proto.VarType.ChannelDesc)
+    ))
+  ,
+
+  Tuple = _reflection.GeneratedProtocolMessageType('Tuple', (_message.Message,), dict(
+    DESCRIPTOR = _VARTYPE_TUPLE,
+    __module__ = 'framework_pb2'
+    # @@protoc_insertion_point(class_scope:paddle_mobile.framework.proto.VarType.Tuple)
+    ))
+  ,
+  DESCRIPTOR = _VARTYPE,
+  __module__ = 'framework_pb2'
+  # @@protoc_insertion_point(class_scope:paddle_mobile.framework.proto.VarType)
+  ))
+_sym_db.RegisterMessage(VarType)
+_sym_db.RegisterMessage(VarType.TensorDesc)
+_sym_db.RegisterMessage(VarType.LoDTensorDesc)
+_sym_db.RegisterMessage(VarType.LoDTensorArrayDesc)
+_sym_db.RegisterMessage(VarType.ReaderDesc)
+_sym_db.RegisterMessage(VarType.ChannelDesc)
+_sym_db.RegisterMessage(VarType.Tuple)
+
+VarDesc = _reflection.GeneratedProtocolMessageType('VarDesc', (_message.Message,), dict(
+  DESCRIPTOR = _VARDESC,
+  __module__ = 'framework_pb2'
+  # @@protoc_insertion_point(class_scope:paddle_mobile.framework.proto.VarDesc)
+  ))
+_sym_db.RegisterMessage(VarDesc)
+
+BlockDesc = _reflection.GeneratedProtocolMessageType('BlockDesc', (_message.Message,), dict(
+  DESCRIPTOR = _BLOCKDESC,
+  __module__ = 'framework_pb2'
+  # @@protoc_insertion_point(class_scope:paddle_mobile.framework.proto.BlockDesc)
+  ))
+_sym_db.RegisterMessage(BlockDesc)
+
+ProgramDesc = _reflection.GeneratedProtocolMessageType('ProgramDesc', (_message.Message,), dict(
+  DESCRIPTOR = _PROGRAMDESC,
+  __module__ = 'framework_pb2'
+  # @@protoc_insertion_point(class_scope:paddle_mobile.framework.proto.ProgramDesc)
+  ))
+_sym_db.RegisterMessage(ProgramDesc)
+
+
+DESCRIPTOR.has_options = True
+DESCRIPTOR._options = _descriptor._ParseOptions(descriptor_pb2.FileOptions(), _b('H\003'))
+# @@protoc_insertion_point(module_scope)

python/tools/mdl2fluid/loader.py

+import datetime
+import json
+import os
+
+import google.protobuf as pbg
+import framework_pb2 as framework_pb2
+
+
+def loadmdl(json_path):
+    print('mdl json path : ' + json_path)
+    with open(json_path, 'r') as f:
+        json_dick = json.load(f)
+        # print(json_dick)
+        layers = (json_dick['layer'])
+        for layer in layers:
+            print(layer)
+
+

python/tools/mdl2fluid/mdl2fluid.py

+import json
+import os
+
+import framework_pb2 as framework_pb2
+import op_types as types
+from swicher import Swichter
+import shutil
+
+
+def load_mdl(mdl_json_path):
+    # print('mdl json path : ' + mdl_json_path)
+    with open(mdl_json_path, 'r') as f:
+        return json.load(f)
+
+
+class Converter:
+    'convert mdlmodel to fluidmodel'
+
+    def __init__(self, mdl_json_path):
+        self.mdl_json_path = mdl_json_path
+        print mdl_json_path
+        self.mdl_json = load_mdl(self.mdl_json_path)
+        self.program_desc = framework_pb2.ProgramDesc()
+        self.weight_list_ = []
+        self.deepwise_weight_list_ = []
+        # print(json_dick)
+        # layers = (json_dick['layer'])
+        # for layer in layers:
+        #     print(layer)
+
+    def convert(self):
+        print 'convert begin.....'
+        # add block_desc
+        block_desc = self.program_desc.blocks.add()
+        block_desc.idx = 0
+        block_desc.parent_idx = -1
+        self.package_ops(block_desc)
+        self.package_vars(block_desc)
+        print 'blocks: '
+        print self.program_desc.blocks
+        print 'convert end.....'
+        desc_serialize_to_string = self.program_desc.SerializeToString()
+        shutil.rmtree('newyolo/')
+        shutil.copytree('multiobjects/float32s_nchw_with_head', 'newyolo/')
+
+        f = open("newyolo/__model__", "wb")
+        f.write(desc_serialize_to_string)
+        f.close()
+
+    def package_ops(self, block_desc):
+
+        self.add_op_feed(block_desc)
+
+        # add ops with layer
+        if 'layer' in self.mdl_json:
+
+            layers_ = self.mdl_json['layer']
+            for layer in layers_:
+                desc_ops_add = block_desc.ops.add()
+
+                # print layer
+                # for i in layer:
+                #     print i
+                if 'name' in layer:
+                    l_name = layer['name']
+                if 'type' in layer:
+                    self.package_ops_type(desc_ops_add, layer)
+
+                if 'weight' in layer:
+                    self.package_ops_weight2inputs(desc_ops_add, layer)
+
+                if 'output' in layer:
+                    self.package_ops_outputs(desc_ops_add, layer)
+
+                if 'input' in layer:
+                    self.package_ops_inputs(desc_ops_add, layer)
+
+                self.package_ops_attrs(desc_ops_add, layer)
+
+        self.add_op_fetch(block_desc)
+
+    def add_op_feed(self, block_desc):
+        desc_ops_add = block_desc.ops.add()
+        inputs_add = desc_ops_add.inputs.add()
+        inputs_add.parameter = 'X'
+        inputs_add.arguments.append('feed')
+        desc_ops_add.type = 'feed'
+        outputs_add = desc_ops_add.outputs.add()
+        outputs_add.parameter = 'Out'
+        outputs_add.arguments.append('data')
+        attrs_add = desc_ops_add.attrs.add()
+        attrs_add.name = 'col'
+        # boolean
+        attrs_add.type = 0
+        attrs_add.i = 0
+
+    def add_op_fetch(self, block_desc):
+        desc_ops_add = block_desc.ops.add()
+        inputs_add = desc_ops_add.inputs.add()
+        inputs_add.parameter = 'X'
+        inputs_add.arguments.append('conv_pred_87')
+        desc_ops_add.type = 'fetch'
+        outputs_add = desc_ops_add.outputs.add()
+        outputs_add.parameter = 'Out'
+        outputs_add.arguments.append('fetch')
+        attrs_add = desc_ops_add.attrs.add()
+        attrs_add.name = 'col'
+        # boolean
+        attrs_add.type = 0
+        attrs_add.i = 0
+
+    @staticmethod
+    def package_ops_attrs(desc_ops_add, layer):
+        # print l_params
+        # print desc_ops_add.type
+        if desc_ops_add.type == types.op_fluid_fusion_conv_add:
+            Converter.pack_fusion_conv_add_attr(desc_ops_add, layer)
+        elif desc_ops_add.type == types.op_fluid_relu:
+            # fusion_conv_add : attrs
+            attrs_add = desc_ops_add.attrs.add()
+            attrs_add.name = 'use_mkldnn'
+            # boolean
+            attrs_add.type = 6
+            attrs_add.b = 0
+
+    @staticmethod
+    def pack_fusion_conv_add_attr(desc_ops_add, layer):
+
+        # fusion_conv_add : attrs
+        attrs_add = desc_ops_add.attrs.add()
+        attrs_add.name = 'workspace_size_MB'
+        # 0-->INT
+        attrs_add.type = 0
+        attrs_add.i = 4096
+
+        attrs_add = desc_ops_add.attrs.add()
+        attrs_add.name = 'data_format'
+        # 2-->STRING
+        attrs_add.type = 2
+        attrs_add.s = 'AnyLayout'
+
+        attrs_add = desc_ops_add.attrs.add()
+        attrs_add.name = 'use_mkldnn'
+        # boolean
+        attrs_add.type = 6
+        attrs_add.b = 0
+
+        attrs_add = desc_ops_add.attrs.add()
+        attrs_add.name = 'use_cudnn'
+        # boolean
+        attrs_add.type = 6
+        attrs_add.b = 1
+
+        attrs_add = desc_ops_add.attrs.add()
+        attrs_add.name = 'dilations'
+        # ints
+        attrs_add.type = 3
+        attrs_add.ints.append(1)
+        attrs_add.ints.append(1)
+
+        attrs_add = desc_ops_add.attrs.add()
+        attrs_add.name = 'axis'
+        # int
+        attrs_add.type = 0
+        attrs_add.i = 1
+
+        if 'param' in layer:
+            l_params = layer['param']
+
+            attrs_add = desc_ops_add.attrs.add()
+            attrs_add.name = 'paddings'
+            # ints
+            attrs_add.type = 3
+            attrs_add.ints.append(l_params[types.fusion_conv_add_attrs_dict.get('paddings')])
+            attrs_add.ints.append(l_params[types.fusion_conv_add_attrs_dict.get('paddings')])
+
+            attrs_add = desc_ops_add.attrs.add()
+            attrs_add.name = 'strides'
+            # ints
+            attrs_add.type = 3
+            attrs_add.ints.append(l_params[types.fusion_conv_add_attrs_dict.get('strides')])
+            attrs_add.ints.append(l_params[types.fusion_conv_add_attrs_dict.get('strides')])
+
+            attrs_add = desc_ops_add.attrs.add()
+            attrs_add.name = 'groups'
+            # int
+            attrs_add.type = 0
+            attrs_add.i = l_params[types.fusion_conv_add_attrs_dict.get('groups')]
+            # attrs_add.i = 1
+
+        #
+        # op_attrs_tupl = types.op_io_dict.get(desc_ops_add.type) \
+        #     .get(types.mdl_attrs_key)
+        #
+        #
+        #
+        #
+        # # group stride padding
+        # print '----------------------'
+        # for i, val in enumerate(op_attrs_tupl):
+        #     attrs_add = desc_ops_add.attrs.add()
+        #     attr_name = op_attrs_tupl[i]
+        #     print attr_name
+        #     attrs_add.name = attr_name
+        #     attrs_add.type = types.fluid_attrs_type_dict.get(attr_name)
+        #     attrs_add.
+        #     print l_params[types.fusion_conv_add_attrs_dict.get(attr_name)]
+
+        # for p in l_params:
+        #     attrs_add = desc_ops_add.attrs.add()
+
+    @staticmethod
+    def package_ops_inputs(desc_ops_add, layer):
+        l_inputs = layer['input']
+        for i in l_inputs:
+            inputs_add = desc_ops_add.inputs.add()
+            # print i
+            inputs_add.parameter = types.op_io_dict.get(desc_ops_add.type).get(types.mdl_inputs_key)
+            inputs_add.arguments.append(i)
+
+    @staticmethod
+    def package_ops_outputs(desc_ops_add, layer):
+        l_outputs = layer['output']
+        for o in l_outputs:
+            # print o
+            outputs_add = desc_ops_add.outputs.add()
+            outputs_add.parameter = types.op_io_dict.get(desc_ops_add.type).get(types.mdl_outputs_key)
+            outputs_add.arguments.append(o)
+
+    def package_ops_weight2inputs(self, desc_ops_add, layer):
+        l_weights = layer['weight']
+        for w in l_weights:
+            self.weight_list_.append(w)
+
+        if layer['type'] == 'DepthwiseConvolutionLayer':
+            # print l_weights[0]
+            self.deepwise_weight_list_.append(l_weights[0])
+
+        op_weight_tup = types.op_io_dict.get(desc_ops_add.type).get(types.mdl_weight_key)
+        # print len(op_weight_tup)
+        for i, val in enumerate(op_weight_tup):
+            # print i
+            # print val
+            inputs_add = desc_ops_add.inputs.add()
+            inputs_add.parameter = op_weight_tup[i]
+            inputs_add.arguments.append(l_weights[i])
+
+        # for w in l_weights:
+        #     inputs_add = desc_ops_add.inputs.add()
+        #     # print w
+        #     inputs_add.parameter = op_weight_tup[0]
+        #     inputs_add.arguments.append(w)
+
+    @staticmethod
+    def package_ops_type(desc_ops_add, layer):
+        l_type = layer['type']
+        # print l_type
+        # print mdl2fluid_op_layer_dict.get(l_type)
+        desc_ops_add.type = types.mdl2fluid_op_layer_dict.get(l_type)
+
+    def package_vars(self, block_desc):
+        vars_add = block_desc.vars.add()
+        vars_add.name = 'feed'
+        vars_add.type.type = 9  # 9 is FEED_MINIBATCH
+        vars_add.persistable = 1
+        # fetch
+        vars_add = block_desc.vars.add()
+        vars_add.name = 'fetch'
+        vars_add.type.type = 10  # 10 is fetch list
+        vars_add.persistable = 1
+
+        json_matrix_ = self.mdl_json['matrix']
+        # print json_matrix_
+        for j in json_matrix_:
+            vars_add = block_desc.vars.add()
+            vars_add.name = j
+            vars_add.type.type = 7  # 7 is lodtensor
+            # print j
+            tensor = vars_add.type.lod_tensor.tensor
+            tensor.data_type = 5  # 5 is FP32
+
+            # print json_matrix_
+
+            dims_of_matrix = json_matrix_.get(j)
+            # dims_size = len(dims_of_matrix)
+            # print dims_size
+
+            # if dims_size == 4:
+            #     tensor.dims.append(dims_of_matrix[0])  # N
+            #     tensor.dims.append(dims_of_matrix[3])  # C
+            #     tensor.dims.append(dims_of_matrix[1])  # H
+            #     tensor.dims.append(dims_of_matrix[2])  # W
+            # else:
+
+            # issues in mdl model filter swich n and c
+            if j in self.deepwise_weight_list_ and len(dims_of_matrix) == 4:
+                print j
+                tensor.dims.append(dims_of_matrix[1])
+                tensor.dims.append(dims_of_matrix[0])
+                tensor.dims.append(dims_of_matrix[2])
+                tensor.dims.append(dims_of_matrix[3])
+                print tensor.dims
+            else:
+                for dims in dims_of_matrix:
+                    # print dims
+                    tensor.dims.append(dims)
+
+            if j in self.weight_list_:
+                vars_add.persistable = 1
+                dims_size = len(dims_of_matrix)
+                # print dims_size
+                if dims_size == 4:
+                    # convert weight from nhwc to nchw
+                    Swichter().nhwc2nchw_one_slice_add_head(
+                        '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nhwc/' + j + '.bin',
+                        '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nchw_with_head/' + j,
+                        '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nchw/' + j + '.tmp',
+                        dims_of_matrix[0],
+                        dims_of_matrix[1],
+                        dims_of_matrix[2],
+                        dims_of_matrix[3]
+                    )
+                else:
+                    Swichter().copy_add_head(
+                        '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nhwc/' + j + '.bin',
+                        '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nchw_with_head/' + j,
+                        '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nchw/' + j + '.tmp'
+                    )
+            else:
+                vars_add.persistable = 0
+
+
+mdl_path = "/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/YOLO_Universal.json"
+converter = Converter(mdl_path)
+converter.convert()

python/tools/mdl2fluid/model_combine.py

+# coding=utf-8
+import os
+
+path = "yolo_v2_tofile_source/"  # 文件夹目录
+to_file_path = "yolo_v2_tofile_combined/params"
+files = os.listdir(path)  # 得到文件夹下的所有文件名称
+files.sort(cmp=None, key=str.lower)
+to_file = open(to_file_path, "wb")
+
+for file in files:  # 遍历文件夹
+    if not os.path.isdir(file):  # 判断是否是文件夹，不是文件夹才打开
+        f = open(path + "/" + file)  # 打开文件
+        name = f.name
+        print 'name:  ' + name
+        from_file = open(name, "rb")
+        to_file.write(from_file.read())
+        from_file.close()
+
+to_file.close()

python/tools/mdl2fluid/model_reader.py

+import os
+
+import framework_pb2 as framework_pb2
+
+
+def read_model(model_path):
+    print('read_model.')
+    path_8 = unicode(model_path, 'utf8')
+
+    try:
+        with open(path_8, "rb") as f_model:
+            print get_file_size(model_path)
+            desc = framework_pb2.ProgramDesc()
+            desc.ParseFromString(f_model.read())
+            print desc
+            # print desc.blocks
+
+    except IOError:
+        print ": File not found.  Creating a new file."
+
+
+def get_file_size(file_path):
+    file_path = unicode(file_path, 'utf8')
+    fsize = os.path.getsize(file_path)
+    fsize = fsize / float(1024 * 1024)
+    return round(fsize, 2)
+
+
+path = "newyolo/__model__"
+read_model(path)

python/tools/mdl2fluid/op_types.py

+# coding=utf-8
+
+# mdl layers
+layer_mdl_conv = 'ConvolutionLayer'
+layer_mdl_deepwise_conv = 'DepthwiseConvolutionLayer'
+layer_mdl_relu = 'ReluLayer'
+layer_mdl_pointwise_add = 'PointwiseConvolutionLayer'
+
+# fluid ops
+op_fluid_fusion_conv_add = 'fusion_conv_add'
+op_fluid_relu = 'relu'
+
+# dict mdk layer ---  fluid op
+mdl2fluid_op_layer_dict = {
+    layer_mdl_conv: op_fluid_fusion_conv_add,
+    layer_mdl_deepwise_conv: op_fluid_fusion_conv_add,
+    layer_mdl_relu: op_fluid_relu,
+    layer_mdl_pointwise_add: op_fluid_fusion_conv_add
+}
+
+mdl_outputs_key = "outputs"
+mdl_inputs_key = "inputs"
+mdl_weight_key = "weights"
+mdl_attrs_key = "params"
+
+# dict of mdl-input _out param  to fluid input out attrs
+fusion_conv_add_dict = {
+    mdl_inputs_key: 'Input',
+    mdl_outputs_key: 'Out',
+    mdl_weight_key: ('Filter', 'Y'),
+    mdl_attrs_key: (
+        # 'workspace_size_MB', 'use_mkldnn', 'use_cudnn', 'data_format','dilations',
+        # dilations =  [1,1]
+        'groups', 'paddings', 'strides'
+        # 'axis'
+    )
+}
+
+relu_dict = {
+    mdl_inputs_key: 'X',
+    mdl_outputs_key: 'Out',
+    mdl_weight_key: ()
+
+}
+# mdl layers  ---  fluid ops
+op_io_dict = {
+    'fusion_conv_add': fusion_conv_add_dict,
+    'relu': relu_dict
+}
+
+# fluid attr key  ---  mdl params key
+fusion_conv_add_attrs_dict = {
+    'paddings': 'pad',
+    'strides': 'stride',
+    'groups': 'group'
+}
+# fluid attr key  ---  mdl params key
+fluid_attrs_type_dict = {
+    'paddings': 0,
+    'strides': 6,
+    'groups': 6
+}
+
+# '': "bias_term",    是不是要add   目前 yolo的模型都是 bias_term = 1
+
+
+# attrs {
+#       name: "axis"
+#       type: INT
+#       i: 1
+#     }
+
+
+# attrs_name = {
+#     'name': "workspace_size_MB",
+#     'type': 'INT',
+#     'i': '4096'
+# }
+# attrs
+# {
+#     name: "data_format"
+#     type: STRING
+#     s: "AnyLayout"
+# }
+# attrs
+# {
+#     name: "use_mkldnn"
+#     type: BOOLEAN
+#     b: false
+# }
+# attrs
+# {
+#     name: "use_cudnn"
+#     type: BOOLEAN
+#     b: true
+# }
+# attrs
+# {
+#     name: "dilations"
+#     type: INTS
+#     ints: 1
+#     ints: 1
+# }
+# attrs
+# {
+#     name: "groups"
+#     type: INT
+#     i: 1
+# }
+# attrs
+# {
+#     name: "paddings"
+#     type: INTS
+#     ints: 0
+#     ints: 0
+# }
+# attrs
+# {
+#     name: "strides"
+#     type: INTS
+#     ints: 1
+#     ints: 1
+# }

python/tools/mdl2fluid/swicher.py

+from array import array
+
+
+class Swichter:
+    def __init__(self):
+        pass
+
+    def nhwc2nchw_one_slice(self, from_file_name, to_file_name, batch, channel, height, width):
+        from_file = open(from_file_name, "rb")
+        to_file = open(to_file_name, "wb")
+
+        float_array = array("f")
+        float_array.fromfile(from_file, width * height * batch * channel)
+        float_write_array = array("f")
+
+        for b in range(batch):
+            for c in range(channel):
+                for h in range(height):
+                    for w in range(width):
+                        float_value = float_array[b * channel * width * height
+                                                  + channel * (h * width + w) + c]
+
+                        float_write_array.append(float_value)
+
+        float_write_array.tofile(to_file)
+        from_file.close()
+        to_file.close()
+
+    def copy(self, from_file_name, to_file_name):
+        from_file = open(from_file_name, "rb")
+        to_file = open(to_file_name, "wb")
+
+        to_file.write(from_file.read())
+        from_file.close()
+        to_file.close()
+
+    def nhwc2nchw_one_slice_add_head(self, from_file_name, to_file_name, tmp_file_name, batch, channel, height, width):
+        from_file = open(from_file_name, "rb")
+        tmp_file = open(tmp_file_name, "wb+")
+        float_array = array("f")
+        float_array.fromfile(from_file, width * height * batch * channel)
+        float_write_array = array("f")
+
+        for b in range(batch):
+            for c in range(channel):
+                for h in range(height):
+                    for w in range(width):
+                        float_value = float_array[b * channel * width * height
+                                                  + channel * (h * width + w) + c]
+
+                        float_write_array.append(float_value)
+
+        float_write_array.tofile(tmp_file)
+        tmp_file.close()
+        from_file.close()
+
+        tmp_file = open(tmp_file_name, "rb")
+        to_file = open(to_file_name, "wb")
+
+        tmp = tmp_file.read()
+        head = self.read_head('/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/yolo/conv1_biases')
+        to_file.write(head)
+        to_file.write(tmp)
+        tmp_file.close()
+        to_file.close()
+
+    def read_head(self, head_file):
+        from_file = open(head_file, "rb")
+        read = from_file.read(24)
+        # print read
+        from_file.close()
+        # print read
+        return read
+
+    def copy_add_head(self, from_file_name, to_file_name, tmp_file_name):
+        from_file = open(from_file_name, "rb")
+        to_file = open(to_file_name, "wb")
+        # tmp_file = open(tmp_file_name, "wb")
+
+        head = self.read_head('/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/yolo/conv1_biases')
+        to_file.write(head)
+        to_file.write(from_file.read())
+        from_file.close()
+        to_file.close()
+        pass
+
+    def copy_padding_add_head(self, from_file_name, to_file_name, tmp_file_name, padding):
+        print'padding  = %d' % padding
+        from_file = open(from_file_name, "rb")
+        # print len(from_file.read())
+        from_file.seek(padding, 0)
+
+        read = from_file.read()
+        print len(read)
+
+        to_file = open(to_file_name, "wb")
+        # tmp_file = open(tmp_file_name, "wb")
+
+        head = self.read_head('/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/yolo/conv1_biases')
+        to_file.write(head)
+        to_file.write(read)
+        from_file.close()
+        to_file.close()
+        pass
+
+# Swichter().nhwc2nchw_one_slice_add_head(
+#     '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nhwc/conv1_0.bin',
+#     '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nchw_with_head/conv1_0',
+#     '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nchw/.tmp',
+#     32,
+#     3, 3, 3)
+
+# Swichter().read_head('/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/yolo/conv1_biases')
+
+# Swichter().copy_add_head('datas/model.0.0.weight', 'datas/conv1_0', '')

src/fpga/api.cpp

@@ -12,28 +12,30 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "api.h"
+#include "fpga/api.h"
 #include <fcntl.h>
-#include <stdio.h>
-#include <stdlib.h>
 #include <sys/ioctl.h>
+#include <sys/mman.h>
 #include <algorithm>
-#include <cstring>
-#include "bias_scale.h"
-#include "filter.h"
-#include "image.h"
-
+#include <map>
+#include "fpga/bias_scale.h"
+#include "fpga/filter.h"
+#include "fpga/image.h"
 #define FPGA_TEST_MODE
+#define PADDLE_MOBILE_OS_LINUX
 
 namespace paddle_mobile {
 namespace fpga {
 
 static int fd = -1;
 static const char *device_path = "/dev/fpgadrv0";
+static std::map<void *, size_t> memory_map;
 
 static inline int do_ioctl(int req, const void *arg) {
 #ifdef PADDLE_MOBILE_OS_LINUX
-  return ioctl(req, (unsigned int64_t)arg);
+  int result = ioctl(fd, req, (uint64_t)arg);
+  PADDLE_MOBILE_ENFORCE(result == 0, "ioctl didn't return correctly");
+  return result;
 #else
   return -1;
 #endif
@@ -48,50 +50,94 @@ int open_device() {
 
 // memory management;
 void *fpga_malloc(size_t size) {
+  static uint64_t counter = 0;
+
 #ifdef PADDLE_MOBILE_OS_LINUX
-  return reinterpret_cast<void *>(
-      mmap64(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0));
+  auto ptr = mmap64(nullptr, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
 #else
-  return malloc(size);
+  auto ptr = malloc(size);
 #endif
+  counter += size;
+  memory_map.insert(std::make_pair(ptr, size));
+  DLOG << "Address: " << ptr << ", " << size << " bytes allocated. Total "
+       << counter << " bytes";
+  return ptr;
 }
 
 void fpga_free(void *ptr) {
+  static uint64_t counter = 0;
+  size_t size = 0;
+
+  auto iter = memory_map.find(ptr);  // std::map<void *, size_t>::iterator
+  if (iter != memory_map.end()) {
+    size = iter->second;
+    memory_map.erase(iter);
 #ifdef PADDLE_MOBILE_OS_LINUX
-  munmap(ptr, 0);
+    munmap(ptr, size);
 #else
     free(ptr);
 #endif
+    counter += size;
+    DLOG << "Address: " << ptr << ", " << size << " bytes freed. Total "
+         << counter << " bytes";
+  } else {
+    DLOG << "Invalid pointer";
+  }
 }
 
 void fpga_copy(void *dest, const void *src, size_t num) {
   memcpy(dest, src, num);
 }
 
-int ComputeFpgaConv(const struct WrapperConvArgs &args) {
-#ifdef FPGA_TEST_MODE
-/*DLOG << "   relu_enabled:" << args.relu_enabled
+int fpga_flush(void *address, size_t size) {
+  struct MemoryCacheArgs args = {nullptr};
+  args.address = address;
+  args.size = size;
+  return do_ioctl(IOCTL_MEMCACHE_FLUSH, &args);
+}
+
+int fpga_invalidate(void *address, size_t size) {
+  struct MemoryCacheArgs args = {nullptr};
+  args.address = address;
+  args.size = size;
+  return do_ioctl(IOCTL_MEMCACHE_INVAL, &args);
+}
+
+int ComputeBasicConv(const struct ConvArgs &args) {
+  DLOG << "======Compute Basic Conv======";
+  DLOG << "   relu_enabled:" << args.relu_enabled
        << "   sb_address:" << args.sb_address
        << "   filter_address:" << args.filter_address
        << "   filter_num:" << args.filter_num
        << "   group_num:" << args.group_num;
-DLOG << "   image_address:" << args.image.address
+  DLOG << "   image_address:" << args.image.address
        << "   image_scale_address:" << args.image.scale_address
        << "   image_channels:" << args.image.channels
        << "   image_height:" << args.image.height
        << "   image_width:" << args.image.width
        << "   pad_height:" << args.image.pad_height
        << "   pad_width:" << args.image.pad_width;
-DLOG << "   kernel_height:" << args.kernel.height
+  DLOG << "   kernel_height:" << args.kernel.height
        << "   kernel_width:" << args.kernel.width
        << "   stride_h:" << args.kernel.stride_h
        << "   stride_w:" << args.kernel.stride_w;
-DLOG << "   out_address:" << args.output.address
-     << "   out_scale_address:" << args.output.scale_address;*/
+  DLOG << "   out_address:" << args.output.address
+       << "   out_scale_address:" << args.output.scale_address;
+
+  return do_ioctl(IOCTL_CONFIG_CONV, &args);
+}
+
+int ComputeFpgaConv(const struct WrapperConvArgs &args) {
+#ifdef FPGA_TEST_MODE
+  DLOG << "=============ComputeFPGAConv===========";
+  DLOG << "   filter_num:" << args.filter_num
+       << "   group_num:" << args.group_num
+       << "   split_num:" << args.split_num;
 #endif
+
   int split_num = args.split_num;
   for (int i = 0; i < split_num; i++) {
-    do_ioctl(IOCTL_CONFIG_CONV, &args.conv_args[i]);
+    ComputeBasicConv(args.conv_args[i]);
   }
 
   if (split_num > 1) {
@@ -101,6 +147,7 @@ DLOG << "   out_address:" << args.output.address
 
 int ComputeFpgaPool(const struct PoolingArgs &args) {
 #ifdef FPGA_TEST_MODE
+  DLOG << "=============ComputeFpgaPool===========";
   DLOG << "   image_address:" << args.image.address
        << "   image_scale_address:" << args.image.scale_address
        << "   image_channels:" << args.image.channels
@@ -121,6 +168,7 @@ int ComputeFpgaPool(const struct PoolingArgs &args) {
 
 int ComputeFpgaEWAdd(const struct EWAddArgs &args) {
 #ifdef FPGA_TEST_MODE
+  DLOG << "=============ComputeFpgaEWAdd===========";
   DLOG << "   relu_enabled:" << args.relu_enabled << "   const0:" << args.const0
        << "   const1:" << args.const1;
   DLOG << "   image0_address:" << args.image0.address
@@ -145,8 +193,11 @@ int ComputeFpgaEWAdd(const struct EWAddArgs &args) {
 }
 int PerformBypass(const struct BypassArgs &args) {
 #ifdef FPGA_TEST_MODE
-  DLOG << "   layout_type:" << args.layout_type
-       << "   convert_type:" << args.convert_type;
+  DLOG << "=============ComputeFpgaBypass===========";
+  DLOG << "   input_type:" << args.input_data_type
+       << "   output_type:" << args.output_data_type
+       << "   input_layout_type:" << args.input_layout_type
+       << "   output_layout_type:" << args.output_layout_type;
   DLOG << "   image_address:" << args.image.address
        << "   image_scale_address:" << args.image.scale_address
        << "   image_channels:" << args.image.channels
@@ -162,29 +213,71 @@ int PerformBypass(const struct BypassArgs &args) {
 }
 
 int ComputeFPGAConcat(const struct ConcatArgs &args) {
+#ifdef FPGA_TEST_MODE
+  DLOG << "=============ComputeFpgaConcat===========";
+  DLOG << "   Image_num: " << args.image_num
+       << "   out_address:" << args.image_out
+       << "   out_scale_address:" << args.scale_out;
+  DLOG << "   image_height:" << args.height << "   image_width:" << args.width;
+  for (int i = 0; i < args.image_num; i++) {
+    DLOG << "   " << i << "th:        ";
+    DLOG << "   channel_num:" << args.channel_num[i]
+         << "   image_address:" << args.images_in[i]
+         << "   image_scale_address:" << args.scales_in[i];
+  }
+#endif
+
   image::concat_images(args.images_in, args.scales_in, args.image_out,
                        args.scale_out, args.image_num, args.channel_num,
                        args.height, args.width);
   return 0;
 }
 
+int get_align_image_cw(int cw) { return align_to_x(cw, IMAGE_ALIGNMENT); }
+
 void format_image(framework::Tensor *image_tensor) {
   auto dims = image_tensor->dims();
   auto channel = dims[1], height = dims[2], width = dims[3];
-  auto data_ptr = image_tensor->mutable_data<float>();
+  auto data_ptr = image_tensor->data<float>();
   size_t memory_size = channel * height * width * sizeof(float);
-  float *new_data = (float *)fpga_malloc(memory_size);
+  auto new_data = (float *)fpga_malloc(memory_size);  // NOLINT
   fpga_copy(new_data, data_ptr, memory_size);
   image::format_image(&new_data, channel, height, width);
   image_tensor->reset_data_ptr(new_data);
 }
 
-void format_ofm(framework::Tensor *ofm_tensor) {
+void format_fp16_ofm(framework::Tensor *ofm_tensor) {
   auto dims = ofm_tensor->dims();
+  size_t memory_size = 0;
+  if (dims.size() == 4) {
     auto channel = dims[1], height = dims[2], width = dims[3];
-  size_t memory_size =
+    memory_size =
         height * align_to_x(channel * width, IMAGE_ALIGNMENT) * sizeof(half);
-  ofm_tensor->reset_data_ptr(fpga_malloc(memory_size));
+  } else if (dims.size() == 2) {
+    memory_size = align_to_x(dims[1], IMAGE_ALIGNMENT) * sizeof(half);
+  } else {
+    DLOG << "Wrong ofm dimension";
+  }
+  auto p = fpga_malloc(memory_size);
+  memset(p, 0, memory_size);
+  ofm_tensor->reset_data_ptr(p);
+}
+
+void format_fp32_ofm(framework::Tensor *ofm_tensor) {
+  auto dims = ofm_tensor->dims();
+  size_t memory_size = 0;
+  if (dims.size() == 4) {
+    auto channel = dims[1], height = dims[2], width = dims[3];
+    memory_size =
+        height * align_to_x(channel * width, IMAGE_ALIGNMENT) * sizeof(float);
+  } else if (dims.size() == 2) {
+    memory_size = align_to_x(dims[1], IMAGE_ALIGNMENT) * sizeof(float);
+  } else {
+    DLOG << "Wrong ofm dimension";
+  }
+  auto p = fpga_malloc(memory_size);
+  memset(p, 0, memory_size);
+  ofm_tensor->reset_data_ptr(p);
 }
 
 float filter_find_max(framework::Tensor *filter_tensor) {
@@ -200,7 +293,7 @@ int get_plit_num(framework::Tensor *filter_tensor) {
   return filter::calc_split_num(num, div_capacity);
 }
 
-int get_element_num_per_div(framework::Tensor *filter_tensor, int group_num) {
+int get_filter_num_per_div(framework::Tensor *filter_tensor, int group_num) {
   auto dims = filter_tensor->dims();
   auto chw = dims[1] * dims[2] * dims[3];
   auto num = dims[0];
@@ -218,13 +311,13 @@ int get_aligned_filter_num(int num) {
 
 void format_filter(framework::Tensor *filter_tensor, float max_value,
                    int group_num) {
-  filter_tensor->scale[0] = float(max_value / 127.0);
-  filter_tensor->scale[1] = float(127.0 / max_value);
+  filter_tensor->scale[0] = float(max_value / 127.0);  // NOLINT
+  filter_tensor->scale[1] = float(127.0 / max_value);  // NOLINT
   auto dims = filter_tensor->dims();
   auto num = dims[0], channel = dims[1], height = dims[2], width = dims[3];
-  auto data_ptr = filter_tensor->mutable_data<float>();
+  auto data_ptr = filter_tensor->data<float>();
   size_t memory_size = num * channel * height * width * sizeof(float);
-  auto new_data = (float *)fpga_malloc(memory_size);
+  auto new_data = (float *)fpga_malloc(memory_size);  // NOLINT
   fpga_copy(new_data, data_ptr, memory_size);
   filter::format_filter(&new_data, num, channel, height, width, group_num,
                         max_value);
@@ -246,7 +339,7 @@ void format_concat_output(framework::Tensor *out, int height, int width,
 
   sum_cw = align_to_x(width * sum_channel, IMAGE_ALIGNMENT);
   auto data_ptr = fpga_malloc(height * sum_cw * sizeof(half));
-  auto ddim = framework::make_ddim({-1, sum_channel, height, width});
+  auto ddim = framework::make_ddim({1, sum_channel, height, width});
   out->Resize(ddim);
   out->reset_data_ptr(data_ptr);
 }
@@ -257,15 +350,16 @@ void fill_conv_arg(struct WrapperConvArgs *arg, framework::Tensor *input,
                    int padding_h, int padding_w, float *bs_ptr) {
   auto input_ptr = input->data<float>();
   auto filter_ptr = filter->data<float>();
-  auto out_ptr = out->mutable_data<float>();
+  auto out_ptr = out->data<float>();
 
   arg->group_num = (uint32_t)group_num;
-  arg->split_num = (uint32_t)fpga::get_plit_num(filter);
+  // Either group_num or split_num = 1;
+  arg->split_num = group_num == 1 ? (uint32_t)get_plit_num(filter) : 1;
   arg->filter_num = (uint32_t)filter->dims()[0];
   arg->output.address = out_ptr;
   arg->output.scale_address = out->scale;
-  arg->conv_args = (fpga::ConvArgs *)fpga::fpga_malloc(arg->split_num *
-                                                       sizeof(fpga::ConvArgs));
+  arg->conv_args =
+      (ConvArgs *)fpga_malloc(arg->split_num * sizeof(ConvArgs));  // NOLINT
 
   arg->concat_arg.image_num = arg->split_num;
   arg->concat_arg.image_out = out_ptr;
@@ -274,15 +368,17 @@ void fill_conv_arg(struct WrapperConvArgs *arg, framework::Tensor *input,
   arg->concat_arg.width = (uint32_t)filter->dims()[3];
 
   int n = arg->split_num;
-  arg->concat_arg.images_in = (half **)fpga::fpga_malloc(n * sizeof(int *));
-  arg->concat_arg.scales_in = (float **)fpga::fpga_malloc(n * sizeof(float *));
+  arg->concat_arg.images_in =
+      (half **)fpga_malloc(n * sizeof(int *));  // NOLINT
+  arg->concat_arg.scales_in =
+      (float **)fpga_malloc(n * sizeof(float *));  // NOLINT
   arg->concat_arg.channel_num =
-      (uint32_t *)fpga::fpga_malloc(n * sizeof(uint32_t));
+      (uint32_t *)fpga_malloc(n * sizeof(uint32_t));  // NOLINT
   arg->concat_arg.image_out = out_ptr;
 
-  const int channel = (int)out->dims()[1];
-  int element_num_per_div = fpga::get_element_num_per_div(filter, group_num);
-  int element_num = fpga::get_aligned_filter_element_num(
+  auto channel = (int)out->dims()[1];  // NOLINT
+  int filter_num_per_div = get_filter_num_per_div(filter, group_num);
+  int element_num = get_aligned_filter_element_num(
       filter->dims()[1] * filter->dims()[2] * filter->dims()[3]);
 
   for (int i = 0; i < n; i++) {
@@ -299,28 +395,30 @@ void fill_conv_arg(struct WrapperConvArgs *arg, framework::Tensor *input,
     arg->conv_args[i].image.scale_address = input->scale;
     arg->conv_args[i].image.pad_height = (uint32_t)padding_h;
     arg->conv_args[i].image.pad_width = (uint32_t)padding_w;
-    arg->conv_args[i].filter_address = &((int8_t *)filter_ptr)[i * element_num];
-    arg->conv_args[i].sb_address = &((int8_t *)bs_ptr)[i * element_num];
-    arg->conv_args[i].filter_num =
-        (uint32_t)(i == n - 1 ? fpga::get_aligned_filter_num(
-                                    channel - (n - 1) * element_num_per_div)
-                              : element_num_per_div);
+    arg->conv_args[i].filter_scale_address = filter->scale;
+    arg->conv_args[i].filter_address = &(
+        (int8_t *)filter_ptr)[i * element_num * filter_num_per_div];  // NOLINT
+    arg->conv_args[i].sb_address = &bs_ptr[i * filter_num_per_div * 2];
+    arg->conv_args[i].filter_num = (uint32_t)(
+        i == n - 1 ? channel - (n - 1) * filter_num_per_div  // NOLINT
+                   : filter_num_per_div);
 
     if (n > 1) {
       arg->conv_args[i].output.scale_address =
-          (float *)fpga::fpga_malloc(2 * sizeof(float));
-      arg->conv_args[i].output.address =
-          fpga::fpga_malloc(input->dims()[2] * input->dims()[3] *
-                            arg->conv_args[i].filter_num * sizeof(half));
-    }
-
-    else {
+          (float *)fpga_malloc(2 * sizeof(float));  // NOLINT
+      arg->conv_args[i].output.address = fpga_malloc(
+          input->dims()[2] *
+          align_to_x(input->dims()[3] * arg->conv_args[i].filter_num,
+                     IMAGE_ALIGNMENT) *
+          sizeof(half));
+    } else {
       arg->conv_args[i].output.scale_address = out->scale;
       arg->conv_args[i].output.address = out_ptr;
     }
 
-    arg->concat_arg.images_in[i] = (half *)arg->conv_args[i].output.address;
-    arg->concat_arg.scales_in[i] = (float *)arg->conv_args[i].sb_address;
+    arg->concat_arg.images_in[i] =
+        (half *)arg->conv_args[i].output.address;  // NOLINT
+    arg->concat_arg.scales_in[i] = arg->conv_args[i].output.scale_address;
     arg->concat_arg.channel_num[i] = arg->conv_args[i].filter_num;
   }
 }

src/fpga/api.h

@@ -20,28 +20,17 @@ limitations under the License. */
 #include <limits>
 #include "framework/tensor.h"
 
-// memory management;
-
 namespace paddle_mobile {
 namespace fpga {
 
-int open_device();
-int close_device();
-
-void* fpga_malloc(size_t size);
-void fpga_free(void* ptr);
-void fpga_copy(void* dst, const void* src, size_t num);
-
-enum DataConvertType {
-  DATA_NO_CONVERT = 0,
-  DATA_FP32_TO_FP16 = 1,
-  DATA_FP16_TO_FP32 = 2,
+enum DataType {
+  DATA_TYPE_FP32 = 1,
+  DATA_TYPE_FP16 = 0,
 };
 
-enum LayoutConvertType {
-  LAYOUT_NO_CONVERT = 0,
-  LAYOUT_CHW_TO_HWC = 1,
-  LAYOUT_HWC_TO_CHW = 2,
+enum LayoutType {
+  LAYOUT_CHW = 1,
+  LAYOUT_HWC = 0,
 };
 
 struct VersionArgs {
@@ -54,9 +43,6 @@ struct MemoryCopyArgs {
   size_t size;
 };
 
-/**
-Conv and Pooling kernel
-*/
 struct KernelArgs {
   uint32_t width;
   uint32_t height;
@@ -118,20 +104,21 @@ struct PoolingArgs {
   struct ImageOutputArgs output;
 };
 
-// elementwise add arguments
 struct EWAddArgs {
   bool relu_enabled;
 
-  float const0;  // output0 = const0 x input0 + const1 x input1;
-  float const1;
+  uint32_t const0;  // output0 = const0 x input0 + const1 x input1;
+  uint32_t const1;
   struct ImageInputArgs image0;
   struct ImageInputArgs image1;
   struct ImageOutputArgs output;
 };
 
 struct BypassArgs {
-  enum DataConvertType convert_type;
-  enum LayoutConvertType layout_type;
+  enum DataType input_data_type;
+  enum DataType output_data_type;
+  enum LayoutType input_layout_type;
+  enum LayoutType output_layout_type;
   struct ImageInputArgs image;
   struct ImageOutputArgs output;
 };
@@ -141,6 +128,16 @@ struct FpgaRegWriteArgs {
   uint64_t value;
 };
 
+struct FpgaRegReadArgs {
+  uint64_t address;
+  uint64_t value;
+};
+
+struct MemoryCacheArgs {
+  void* address;
+  size_t size;
+};
+
 #define IOCTL_FPGA_MAGIC 'FPGA'
 
 #define IOCTL_VERSION _IOW(IOCTL_FPGA_MAGIC, 01, struct VersionArgs)
@@ -148,6 +145,8 @@ struct FpgaRegWriteArgs {
 #define IOCTL_SEPARATOR_0 10
 
 #define IOCTL_MEM_COPY _IOW(IOCTL_FPGA_MAGIC, 11, struct MemoryCopyArgs)
+#define IOCTL_MEMCACHE_INVAL _IOW(IOCTL_FPGA_MAGIC, 12, struct MemoryCacheArgs)
+#define IOCTL_MEMCACHE_FLUSH _IOW(IOCTL_FPGA_MAGIC, 13, struct MemoryCacheArgs)
 
 #define IOCTL_SEPARATOR_1 20
 
@@ -184,6 +183,15 @@ enum FPGA_ERR_TYPE {
 
 //============================== API =============================
 
+int open_device();
+int close_device();
+
+void* fpga_malloc(size_t size);
+void fpga_free(void* ptr);
+void fpga_copy(void* dst, const void* src, size_t num);
+int fpga_flush(void* address, size_t size);
+int fpga_invalidate(void* address, size_t size);
+
 int PerformBypass(const struct BypassArgs& args);
 int ComputeFpgaConv(const struct WrapperConvArgs& args);
 int ComputeFpgaPool(const struct PoolingArgs& args);
@@ -192,11 +200,13 @@ int ComputeFPGAConcat(const struct ConcatArgs& args);
 
 static inline int align_to_x(int num, int x) { return (num + x - 1) / x * x; }
 
+int get_align_image_cw(int cw);
 void format_image(framework::Tensor* image_tensor);
-void format_ofm(framework::Tensor* ofm_tensor);  // only allocate memory
+void format_fp16_ofm(framework::Tensor* ofm_tensor);  // only allocate memory
+void format_fp32_ofm(framework::Tensor* ofm_tensor);
 
 float filter_find_max(framework::Tensor* filter_tensor);
-int get_element_num_per_div(framework::Tensor* filter_tensor, int group_num);
+int get_filter_num_per_div(framework::Tensor* filter_tensor, int group_num);
 int get_plit_num(framework::Tensor* filter_tensor);
 int get_aligned_filter_element_num(int chw);
 int get_aligned_filter_num(int num);

src/fpga/bias_scale.cpp

@@ -12,9 +12,9 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "bias_scale.h"
+#include "fpga/bias_scale.h"
 #include <memory.h>
-#include "api.h"
+#include "fpga/api.h"
 
 namespace paddle_mobile {
 namespace fpga {
@@ -29,7 +29,8 @@ void align_element(float **data_in, int num_per_div_before_alignment, int num) {
       align_to_x(num_per_div_before_alignment, BS_NUM_ALIGNMENT);
   int num_element =
       2 * div_num * num_per_div_after_alignment;  // including bias & scale
-  float *ptr_aligned = (float *)fpga_malloc(num_element * sizeof(float));
+  float *ptr_aligned =
+      (float *)fpga_malloc(num_element * sizeof(float));  // NOLINT
 
   memset(ptr_aligned, 0, num_element * sizeof(float));
 
@@ -59,7 +60,7 @@ void interleave(float **data_in, int num_after_alignment) {
 
   float *ptr_uninterleaved = *data_in;
   float *ptr_interleaved =
-      (float *)fpga_malloc(2 * num_after_alignment * sizeof(float));
+      (float *)fpga_malloc(2 * num_after_alignment * sizeof(float));  // NOLINT
   int num = num_after_alignment / 4;
   for (int i = 0; i < num; i++) {
     memcpy(ptr_interleaved + 8 * i, ptr_uninterleaved + 4 * i,
@@ -79,6 +80,7 @@ void format_bias_scale_array(float **bias_scale_array,
   int element_num_after_division =
       align_to_x(element_num_per_division, BS_NUM_ALIGNMENT);
   interleave(bias_scale_array, div_num * element_num_after_division);
+  fpga_flush(*bias_scale_array, 2 * element_num_after_division * sizeof(float));
 }
 
 }  // namespace bias_scale

src/fpga/filter.cpp

@@ -11,9 +11,11 @@ distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
-#include "filter.h"
+
+#include "fpga/filter.h"
 #include <memory.h>
-#include "api.h"
+#include <algorithm>
+#include "fpga/api.h"
 
 namespace paddle_mobile {
 namespace fpga {
@@ -55,7 +57,7 @@ void convert_to_hwc(char **data_in, int num, int channel, int height,
                     int width) {
   char *tmp = *data_in;
   int chw = channel * height * width;
-  char *data_tmp = (char *)fpga_malloc(chw * num * sizeof(char));
+  char *data_tmp = (char *)fpga_malloc(chw * num * sizeof(char));  // NOLINT
   for (int n = 0; n < num; n++) {
     int64_t amount_per_row = width * channel;
     for (int c = 0; c < channel; c++) {
@@ -88,11 +90,11 @@ void quantize(float **data_in, int data_size, float max) {
   float fix_range = 127;
   float scale = fix_range / max;
 
-  char *tmp_data = (char *)fpga_malloc(data_size * sizeof(char));
+  char *tmp_data = (char *)fpga_malloc(data_size * sizeof(char));  // NOLINT
   for (int i = 0; i < data_size; i++) {
-    tmp_data[i] = (char)((*data_in)[i] * scale);
+    tmp_data[i] = (char)((*data_in)[i] * scale);  // NOLINT
   }
-  *data_in = (float *)tmp_data;
+  *data_in = (float *)tmp_data;  // NOLINT
   fpga_free(tmp);
 }
 
@@ -101,9 +103,9 @@ void align_element(char **data_in, int num, int chw) {
   int j = 0;
   int align_chw = align_to_x(chw, FILTER_ELEMENT_ALIGNMENT);
   if (align_chw != chw) {
-    printf("align %d \n", align_chw);
     char *tmp = *data_in;
-    char *data_tmp = (char *)fpga_malloc(num * align_chw * sizeof(char));
+    char *data_tmp =
+        (char *)fpga_malloc(num * align_chw * sizeof(char));  // NOLINT
 
     memset(data_tmp, 0, num * align_chw);
     for (j = 0; j < num; j++) {
@@ -125,7 +127,7 @@ void align_num(char **data_in, int num_per_div_before_alignment, int num,
     int div_num =
         (num + num_per_div_before_alignment - 1) / num_per_div_before_alignment;
     int num_element = div_num * num_per_div_after_alignment * align_chw;
-    char *data_tmp = (char *)fpga_malloc(num_element * sizeof(char));
+    char *data_tmp = (char *)fpga_malloc(num_element * sizeof(char));  // NOLINT
 
     memset(data_tmp, 0, num_element * sizeof(char));
 
@@ -147,7 +149,8 @@ void reorder(char **data_in, int num_after_alignment, int chw) {
   int chw_align = align_to_x(chw, FILTER_ELEMENT_ALIGNMENT);
 
   char *data_tmp =
-      (char *)fpga_malloc(chw_align * num_after_alignment * sizeof(char));
+      (char *)fpga_malloc(chw_align * num_after_alignment *  // NOLINT
+                          sizeof(char));
   char *tmp = *data_in;
   for (index = 0; index < num_after_alignment; index++) {
     new_index = index / 32 * 32 + (index % 16 / 4 * 8) + (index % 16 % 4) +
@@ -164,10 +167,11 @@ void interleave(char **data_in, int num_after_alignment, int chw) {
   int j = 0;
   int k = 0;
   int interleave_per_num = 16;
-  ;
+
   int chw_align = align_to_x(chw, FILTER_ELEMENT_ALIGNMENT);
   char *data_tmp =
-      (char *)fpga_malloc(chw_align * num_after_alignment * sizeof(char));
+      (char *)fpga_malloc(chw_align * num_after_alignment *  // NOLINT
+                          sizeof(char));
   char *tmp = *data_in;
   int interleave_num = chw_align * 2 / interleave_per_num;
   for (i = 0; i < num_after_alignment; i += 2) {
@@ -200,13 +204,15 @@ void format_filter(float **data_in, int num, int channel, int height, int width,
 
   quantize(data_in, data_size, max);
 
-  char **quantize_data = (char **)data_in;
+  char **quantize_data = (char **)data_in;  // NOLINT
 
   convert_to_hwc(quantize_data, num, channel, height, width);
   align_element(quantize_data, num, chw);
   align_num(quantize_data, num_per_div_before_alignment, num, chw);
   reorder(quantize_data, num_after_alignment, chw);
   interleave(quantize_data, num_after_alignment, chw);
+  fpga_flush(*quantize_data, align_to_x(chw, FILTER_ELEMENT_ALIGNMENT) *
+                                 num_after_alignment * sizeof(char));
 }
 
 }  // namespace filter

src/fpga/image.cpp

@@ -12,9 +12,10 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "image.h"
+#include "fpga/image.h"
 #include <memory.h>
-#include "api.h"
+#include <algorithm>
+#include "fpga/api.h"
 
 namespace paddle_mobile {
 namespace fpga {
@@ -23,7 +24,7 @@ namespace image {
 void convert_to_hwc(float **data_in, int channel, int height, int width) {
   float *tmp = *data_in;
   float *data_tmp =
-      (float *)fpga_malloc(channel * height * width * sizeof(float));
+      (float *)fpga_malloc(channel * height * width * sizeof(float));  // NOLINT
   int64_t amount_per_row = width * channel;
   for (int c = 0; c < channel; c++) {
     for (int h = 0; h < height; h++) {
@@ -38,17 +39,18 @@ void convert_to_hwc(float **data_in, int channel, int height, int width) {
 }
 
 void align_element_conv(float **data_in, int height, int cw) {
-  int i = 0;
   int h = 0;
   int align_cw = align_to_x(cw, IMAGE_ALIGNMENT);
   if (align_cw != cw) {
     float *tmp = *data_in;
-    float *data_tmp = (float *)fpga_malloc(height * align_cw * sizeof(float));
+    float *data_tmp =
+        (float *)fpga_malloc(height * align_cw * sizeof(float));  // NOLINT
 
     memset(data_tmp, 0, height * align_cw * sizeof(float));
 
     for (h = 0; h < height; h++) {
-      memcpy((void *)(data_tmp + h * align_cw), (void *)(*data_in + h * cw),
+      memcpy((void *)(data_tmp + h * align_cw),  // NOLINT
+             (void *)(*data_in + h * cw),        // NOLINT
              cw * sizeof(float));
     }
 
@@ -60,6 +62,8 @@ void align_element_conv(float **data_in, int height, int cw) {
 void format_image(float **data_in, int channel, int height, int width) {
   convert_to_hwc(data_in, channel, height, width);
   align_element_conv(data_in, height, channel * width);
+  fpga_flush(*data_in, align_to_x(channel * width, IMAGE_ALIGNMENT) * height *
+                           sizeof(float));
 }
 
 void concat_images(int16_t **images_in, float **scales_in, void *image_out,
@@ -73,11 +77,17 @@ void concat_images(int16_t **images_in, float **scales_in, void *image_out,
   int align_each_in_area_cw = 0;
   int align_each_out_area_cw_differ = 0;
   int tmp_channel = 0;
-  *scale_out = 0;
+  scale_out[0] = 0.0;
+  scale_out[1] = 0.0;
   for (i = 0; i < image_num; i++) {
     each_out_line_channel += channel_num[i];
-    *scale_out = std::max(*scale_out, scales_in[i][0]);
+    scale_out[0] = std::max(*scale_out, scales_in[i][0]);
+    fpga_invalidate(images_in[i],
+                    height *
+                        align_to_x(channel_num[i] * width, IMAGE_ALIGNMENT) *
+                        sizeof(int16_t));
   }
+  scale_out[1] = 1 / scale_out[0];
   align_each_out_area_cw =
       align_to_x(each_out_line_channel * width, IMAGE_ALIGNMENT);
   align_each_out_area_cw_differ =
@@ -88,7 +98,7 @@ void concat_images(int16_t **images_in, float **scales_in, void *image_out,
       for (i = 0; i < image_num; i++) {
         align_each_in_area_cw =
             align_to_x(channel_num[i] * width, IMAGE_ALIGNMENT);
-        memcpy((int16_t *)image_out + tmp_channel +
+        memcpy((int16_t *)image_out + tmp_channel +  // NOLINT
                    k * align_each_out_area_cw_differ,
                images_in[i] + j * channel_num[i] + k * align_each_in_area_cw,
                channel_num[i] * sizeof(int16_t));
@@ -97,6 +107,8 @@ void concat_images(int16_t **images_in, float **scales_in, void *image_out,
       }
     }
   }
+
+  fpga_flush(image_out, height * align_each_out_area_cw * sizeof(int16_t));
 }
 
 }  // namespace image

src/framework/operator.cpp

@@ -62,13 +62,24 @@ void OperatorBase<Dtype>::Run() const {
   DLOG << "-------------" << type_ << "----------------------------";
   vector<string> input_keys = GetInputKeys();
   for (const auto key : input_keys) {
-    Tensor *input = GetVarValue<framework::LoDTensor>(key, inputs_, *scope_);
-    if (input) DLOG << type_ << " input- " << key << "=" << *input;
+    auto var_vec_in = inputs_.at(key);
+    for (int i = 0; i < var_vec_in.size(); ++i) {
+      auto vari = scope_->FindVar(var_vec_in[i]);
+      if (vari->IsInitialized()) {
+        Tensor *tensor = vari->template GetMutable<framework::LoDTensor>();
+        if (tensor) DLOG << type_ << " input- " << key << "=" << *tensor;
+      }
+    }
+  }
+  for (const auto key : GetOutKeys()) {
+    auto var_vec_out = outputs_.at(key);
+    for (int i = 0; i < var_vec_out.size(); ++i) {
+      auto vari = scope_->FindVar(var_vec_out[i]);
+      if (vari->IsInitialized()) {
+        Tensor *tensor = vari->template GetMutable<framework::LoDTensor>();
+        if (tensor) DLOG << type_ << " output- " << key << "=" << *tensor;
+      }
     }
-  vector<string> output_keys = GetOutKeys();
-  for (const auto key : output_keys) {
-    Tensor *out_ = GetVarValue<framework::LoDTensor>(key, outputs_, *scope_);
-    DLOG << type_ << " output- " << key << "=" << *out_;
   }
 #endif
 }

src/framework/tensor.h

@@ -289,12 +289,8 @@ class Tensor {
     virtual std::type_index type() const { return type_; }
 
     virtual void set_type(std::type_index type) { type_ = type; }
-#ifndef PADDLE_MOBILE_FPGA
-    /*! the pointer of memory block. */
+
     std::unique_ptr<uint8_t, memory::PODDeleter<uint8_t>> ptr_;
-#else
-    std::shared_ptr<uint8_t> ptr_;
-#endif
 
     /*! the size of memory block. */
     size_t size_;
@@ -323,10 +319,11 @@ class Tensor {
    * begins.
    */
   size_t offset_;
+
 #ifdef PADDLE_MOBILE_FPGA
- public:
+ public:  // NOLINT
   inline void reset_data_ptr(void *p) {
-    ((PlaceholderImpl *)(holder_.get()))->ptr_.reset((uint8_t *)p);
+    ((PlaceholderImpl *)(holder_.get()))->ptr_.reset((uint8_t *)p);  // NOLINT
   }
   float scale[2];  // scale[0]= MAX/127.0, scale[1]= 127.0/MAX
 #endif

src/io/loader.cpp

@@ -27,8 +27,8 @@ using framework::Variable;
  * @param scope
  */
 void InitMemoryFromProgram(
-    std::shared_ptr<framework::ProgramDesc> &originProgramDesc,
-    std::shared_ptr<framework::Scope> &scope) {
+    std::shared_ptr<framework::ProgramDesc> &originProgramDesc,  // NOLINT
+    std::shared_ptr<framework::Scope> &scope) {                  // NOLINT
   for (const auto &block : originProgramDesc.get()->Blocks()) {
     for (const auto &var_desc : block->Vars()) {
       auto var = scope.get()->Var(var_desc->Name());
@@ -45,7 +45,6 @@ void InitMemoryFromProgram(
           tensor->Resize(framework::make_ddim(dim));
         }
       } else {
-        // var_desc type is always lod tensor in any time?? (houjiang)
         // TODO(codeWorm): some.
       }
     }
@@ -62,7 +61,8 @@ void InitMemoryFromProgram(
  */
 template <typename Dtype, Precision P>
 void FusionAndPrintInfos(
-    bool &optimize, bool &can_add_split, framework::Program<Dtype, P> &program,
+    bool optimize, bool can_add_split,
+    framework::Program<Dtype, P> &program,  // NOLINT
     const std::shared_ptr<framework::ProgramDesc> &originProgramDesc) {
   if (optimize) {
     framework::ProgramOptimize program_optimize;
@@ -78,6 +78,7 @@ void FusionAndPrintInfos(
     originProgramDesc->Description("program: ");
   }
 }
+
 static size_t ReadBuffer(const char *file_name, uint8_t **out) {
   FILE *fp;
   fp = fopen(file_name, "rb");

src/io/paddle_mobile.cpp

@@ -21,13 +21,11 @@ void PaddleMobile<Dtype, P>::SetThreadNum(int num) {
 #ifdef _OPENMP
   omp_set_num_threads(num);
 #endif
-};
+}
 
 template <typename Dtype, Precision P>
-bool PaddleMobile<Dtype, P>::Load(const std::string &dirname,
-                                  bool optimize,
-                                  bool quantification,
-                                  bool loddable) {
+bool PaddleMobile<Dtype, P>::Load(const std::string &dirname, bool optimize,
+                                  bool quantification, bool loddable) {
   if (loader_.get() == nullptr) {
     loader_ = std::make_shared<Loader<Dtype, P>>();
   } else {
@@ -46,10 +44,8 @@ bool PaddleMobile<Dtype, P>::Load(const std::string &dirname,
 
 template <typename Dtype, Precision P>
 bool PaddleMobile<Dtype, P>::Load(const std::string &model_path,
-                                  const std::string &para_path,
-                                  bool optimize,
-                                  bool quantification,
-                                  bool loddable) {
+                                  const std::string &para_path, bool optimize,
+                                  bool quantification, bool loddable) {
   if (loader_.get() == nullptr) {
     loader_ = std::make_shared<Loader<Dtype, P>>();
   } else {
@@ -84,7 +80,8 @@ bool PaddleMobile<Dtype, P>::LoadCombinedMemory(
     executor_ = std::make_shared<Executor<Dtype, P>>(
         loader_->LoadCombinedMemory(model_len, model_buf, combined_params_len,
                                     combined_params_buf, optimise,
-                                    quantification), optimise);
+                                    quantification),
+        optimise);
   } else {
     LOG(kLOG_INFO) << "executor inited";
   }
@@ -122,6 +119,40 @@ PaddleMobile<Dtype, P>::~PaddleMobile() {
   loader_ = nullptr;
 }
 
+#ifdef PADDLE_MOBILE_FPGA
+
+template <typename Dtype, Precision P>
+void PaddleMobile<Dtype, P>::InjectVariable(const framework::Tensor &t,
+                                            string var_name) {
+  executor_->InjectVariable(t, var_name);
+}
+
+template <typename Dtype, Precision P>
+void PaddleMobile<Dtype, P>::FeedData(const framework::Tensor &t) {
+  executor_->FeedData(t);
+}
+
+template <typename Dtype, Precision P>
+std::shared_ptr<framework::Tensor> PaddleMobile<Dtype, P>::FetchResult(int id) {
+  return executor_->FetchResult(id);
+}
+
+template <typename Dtype, Precision P>
+void PaddleMobile<Dtype, P>::Predict_From_To(int start, int end) {
+  executor_->Predict_From_To(start, end);
+}
+
+template <typename Dtype, Precision P>
+void PaddleMobile<Dtype, P>::Predict_From(int start) {
+  executor_->Predict_From(start);
+}
+
+template <typename Dtype, Precision P>
+void PaddleMobile<Dtype, P>::Predict_To(int end) {
+  executor_->Predict_To(end);
+}
+#endif
+
 template class PaddleMobile<CPU, Precision::FP32>;
 template class PaddleMobile<FPGA, Precision::FP32>;
 template class PaddleMobile<GPU_MALI, Precision::FP32>;

src/io/paddle_mobile.h

@@ -34,15 +34,11 @@ class PaddleMobile {
 
  public:
   PaddleMobile() {}
-  bool Load(const std::string &dirname,
-            bool optimize = false,
-            bool quantification = false,
-            bool loddable = false);
+  bool Load(const std::string &dirname, bool optimize = false,
+            bool quantification = false, bool loddable = false);
 
-  bool Load(const std::string &model_path,
-            const std::string &para_path,
-            bool optimize = false,
-            bool quantification = false,
+  bool Load(const std::string &model_path, const std::string &para_path,
+            bool optimize = false, bool quantification = false,
             bool loddable = false);
 
   std::shared_ptr<framework::Tensor> Predict(const framework::Tensor &t);
@@ -61,6 +57,15 @@ class PaddleMobile {
 
   ~PaddleMobile();
 
+#ifdef PADDLE_MOBILE_FPGA
+  void InjectVariable(const framework::Tensor &t, string var_name);
+  void FeedData(const framework::Tensor &t);
+  std::shared_ptr<framework::Tensor> FetchResult(int id = -1);
+  void Predict_From_To(int start = 0, int end = -1);
+  void Predict_From(int start);
+  void Predict_To(int end);
+#endif
+
  private:
   std::shared_ptr<Loader<Dtype, P>> loader_;
   std::shared_ptr<Executor<Dtype, P>> executor_;

src/ios_io/PaddleMobileCPU.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License. */
+
+#pragma once
+
+#import <CoreImage/CoreImage.h>
+#import <Foundation/Foundation.h>
+
+@interface PaddleMobileCPUResult: NSObject
+
+@property (assign, nonatomic, readonly) float *output;
+
+@property (assign, nonatomic, readonly) int outputSize;
+
+-(void)releaseOutput;
+
+@end
+
+@interface PaddleMobileCPU : NSObject
+
+/*
+    创建对象
+*/
+- (instancetype)init;
+
+/*
+    load 模型, 开辟内存
+*/
+- (BOOL)load:(NSString *)modelPath andWeightsPath:(NSString *)weighsPath;
+
+/*
+  加载散开形式的模型, 需传入模型的目录
+*/
+- (BOOL)load:(NSString *)modelAndWeightPath;
+
+/*
+ * 从内存中加载模型
+ * */
+- (BOOL)LoadCombinedMemory:(size_t)modelLen
+               andModelBuf:(const uint8_t *)modelBuf
+         andModelParamsLen:(size_t)combinedParamsLen
+      andCombinedParamsBuf:(const uint8_t *)combinedParamsBuf;
+
+/*
+ *  对图像进行预处理, 需要外部开辟 output 内存, 外部释放 output 内存
+ * */
+-(void)preprocess:(CGImageRef)image
+           output:(float *)output
+            means:(NSArray<NSNumber *> *)means
+        scale:(float)scale
+        dim:(NSArray<NSNumber *> *)dim;
+
+/*
+ * 预测预处理后的数据, 返回结果使用结束需要调用其 realseOutput 函数进行释放
+ * */
+- (PaddleMobileCPUResult *)predictInput:(float *)input
+                                    dim:(NSArray<NSNumber *> *)dim;
+
+/*
+    进行预测, means 和 scale 为训练模型时的预处理参数, 如训练时没有做这些预处理则直接使用 predict
+*/
+- (NSArray *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim means:(NSArray<NSNumber *> *)means scale:(float)scale;
+
+/*
+    进行预测, 默认 means 为 0, scale 为 1.0
+*/
+- (NSArray *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim;
+
+/*
+    清理内存
+*/
+- (void)clear;
+
+@end

src/ios_io/PaddleMobile.mm → src/ios_io/PaddleMobileCPU.mm

@@ -12,24 +12,51 @@
  See the License for the specific language governing permissions and
  limitations under the License. */
 
-#import "PaddleMobile.h"
+#import "PaddleMobileCPU.h"
 
 #import "op_symbols.h"
+#include "framework/tensor.h"
 #import "io/paddle_mobile.h"
 
 #import <memory>
 #import <vector>
 
-@interface  PaddleMobile()
+
+@interface PaddleMobileCPUResult()
+
+-(void)toSetOutput:(float *)output;
+
+-(void)toSetOutputSize:(int)outputSize;
+
+@end
+
+@implementation PaddleMobileCPUResult
+
+-(void)releaseOutput {
+  delete [] _output;
+  _output = nil;
+  _outputSize = 0;
+}
+
+-(void)toSetOutput:(float *)output {
+  _output = output;
+}
+
+-(void)toSetOutputSize:(int)outputSize {
+  _outputSize = outputSize;
+}
+
+@end
+
+
+@interface  PaddleMobileCPU()
 {
   paddle_mobile::PaddleMobile<paddle_mobile::CPU, paddle_mobile::Precision::FP32> *pam_;
   BOOL loaded_;
-  std::vector<float> *predict_input_;
-
 }
 @end
 
-@implementation PaddleMobile
+@implementation PaddleMobileCPU
 
 static std::mutex shared_mutex;
 
@@ -66,6 +93,14 @@ static std::mutex shared_mutex;
   }
 }
 
+- (BOOL)LoadCombinedMemory:(size_t)modelLen
+               andModelBuf:(const uint8_t *)modelBuf
+         andModelParamsLen:(size_t)combinedParamsLen
+      andCombinedParamsBuf:(const uint8_t *)combinedParamsBuf {
+  pam_->SetThreadNum(2);
+  return loaded_ = pam_->LoadCombinedMemory(modelLen, modelBuf, combinedParamsLen, combinedParamsBuf);
+}
+
 - (BOOL)load:(NSString *)modelAndWeightPath{
   std::string model_path_str = std::string([modelAndWeightPath UTF8String]);
   if (loaded_ = pam_->Load(model_path_str)) {
@@ -75,6 +110,57 @@ static std::mutex shared_mutex;
   }
 }
 
+
+-(void)preprocess:(CGImageRef)image
+           output:(float *)output
+            means:(NSArray<NSNumber *> *)means
+        scale:(float)scale
+        dim:(NSArray<NSNumber *> *)dim {
+  std::lock_guard<std::mutex> lock(shared_mutex);
+
+  // dim to c++ vector, get numel
+  std::vector<int64_t > dim_vec;
+  int numel = 1;
+  for (int k = 0; k < dim.count; ++k) {
+    int d = dim[k].intValue;
+    numel *= d;
+    dim_vec.push_back(d);
+  }
+
+  const int sourceRowBytes = CGImageGetBytesPerRow(image);
+  const int imageWidth = CGImageGetWidth(image);
+  const int imageHeight = CGImageGetHeight(image);
+  const int imageChannels = 4;
+  CGDataProviderRef provider = CGImageGetDataProvider(image);
+  CFDataRef cfData = CGDataProviderCopyData(provider);
+  const UInt8 *input = CFDataGetBytePtr(cfData);
+
+  int wanted_input_width = dim_vec[3];
+  int wanted_input_height = dim_vec[2];
+  int wanted_input_channels = dim_vec[1];
+
+  for (int c = 0; c < wanted_input_channels; ++c) {
+    float *out_channel = output + c * wanted_input_height * wanted_input_width;
+    for (int y = 0; y < wanted_input_height; ++y) {
+      float *out_row = out_channel + y * wanted_input_width;
+      for (int x = 0; x < wanted_input_width; ++x) {
+        int in_row = (y * imageHeight) / wanted_input_height;
+        int in_col = (x * imageWidth) / wanted_input_width;
+        const UInt8 *in_pixel = input + (in_row * imageWidth * imageChannels) + (in_col * imageChannels);
+        float *out_pos = out_row + x;
+        if (c == 0) {
+          *out_pos = (in_pixel[c] - means[c].floatValue) * scale;
+        }else if (c == 1){
+          *out_pos = (in_pixel[c] - means[c].floatValue) * scale;
+        }else if (c == 2){
+          *out_pos = (in_pixel[c] - means[c].floatValue) * scale;
+        }
+      }
+    }
+  }
+
+}
+
 -(void)preprocess:(const UInt8 *)input output:(float *)output imageWidth:(int)imageWidth imageHeight:(int)imageHeight imageChannels:(int)imageChannels means:(NSArray<NSNumber *> *)means scale:(float)scale dim:(std::vector<int64_t>)dim{
   if (means == nil) {
     means = @[@0, @0, @0];
@@ -105,27 +191,54 @@ static std::mutex shared_mutex;
   }
 }
 
-- (NSArray *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim means:(NSArray<NSNumber *> *)means scale:(float)scale{
-//  printf(" hi i am here");
-  if (predict_input_) {
-//    printf(" fukc -- ");
-//    printf(" %d \n", predict_input_->size());
-    // dim to c++ vector, get numel
-    std::vector<int64_t > dim_vec = {1, 3, 300, 300};
-//    int numel = 1;
-//    for (int k = 0; k < dim.count; ++k) {
-//      int d = dim[k].intValue;
-//      numel *= d;
-//      dim_vec.push_back(d);
-//    }
-
+- (PaddleMobileCPUResult *)predictInput:(float *)input
+                      dim:(NSArray<NSNumber *> *)dim {
+  std::lock_guard<std::mutex> lock(shared_mutex);
+  if (!loaded_) {
+    printf("PaddleMobile doesn't be loaded yet");
+    return nil;
+  }
 
-    std::vector<float> cpp_result = pam_->Predict(*predict_input_, dim_vec);
+  if (dim.count != 4) {
+    printf("dim must have 4 elements");
     return nil;
   }
-//  printf(" predict one ");
 
-//  std::lock_guard<std::mutex> lock(shared_mutex);
+  // dim to c++ vector, get numel
+  std::vector<int64_t > dim_vec;
+  int numel = 1;
+  for (int k = 0; k < dim.count; ++k) {
+    int d = dim[k].intValue;
+    numel *= d;
+    dim_vec.push_back(d);
+  }
+
+  paddle_mobile::framework::Tensor input_tensor;
+
+  paddle_mobile::framework::DDim dims = paddle_mobile::framework::make_ddim(dim_vec);
+
+  float *input_ptr = input_tensor.mutable_data<float>(dims);
+
+  memcpy(input_ptr, input,
+         numel * sizeof(float));
+
+  std::shared_ptr<paddle_mobile::framework::Tensor> output = pam_->Predict(input_tensor);
+
+  float *output_pointer = new float[output->numel()];
+
+  memcpy(output_pointer, output->data<float>(),
+         output->numel() * sizeof(float));
+
+  PaddleMobileCPUResult *cpuResult = [[PaddleMobileCPUResult alloc] init];
+  [cpuResult toSetOutput: output_pointer];
+  [cpuResult toSetOutputSize: output->numel()];
+
+  return cpuResult;
+}
+
+- (NSArray *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim means:(NSArray<NSNumber *> *)means scale:(float)scale{
+//  printf(" predict one ");
+  std::lock_guard<std::mutex> lock(shared_mutex);
   if (!loaded_) {
     printf("PaddleMobile doesn't be loaded yet");
     return nil;
@@ -164,15 +277,13 @@ static std::mutex shared_mutex;
   }
 
   // input
-  std::vector<float> *predict_input = new std::vector<float>();
+  std::vector<float> predict_input;
   for (int j = 0; j < numel; ++j) {
-    predict_input->push_back(dataPointer[j]);
+    predict_input.push_back(dataPointer[j]);
   }
 
-  predict_input_ = predict_input;
-
   // predict
-  std::vector<float> cpp_result = pam_->Predict(*predict_input, dim_vec);
+  std::vector<float> cpp_result = pam_->Predict(predict_input, dim_vec);
 
   // result
   long count = 0;

src/ios_io/op_symbols.h

@@ -15,27 +15,48 @@
 #pragma once
 
 #include "operators/batchnorm_op.h"
+#include "operators/bilinear_interp_op.h"
 #include "operators/box_coder_op.h"
 #include "operators/concat_op.h"
 #include "operators/conv_op.h"
+#include "operators/conv_transpose_op.h"
+#include "operators/crf_op.h"
 #include "operators/depthwise_conv_op.h"
+#include "operators/dequantize_op.h"
 #include "operators/dropout_op.h"
 #include "operators/elementwise_add_op.h"
 #include "operators/feed_op.h"
 #include "operators/fetch_op.h"
-#include "operators/fusion_conv_add.h"
+#include "operators/flatten_op.h"
+#include "operators/fusion_conv_add_add_prelu_op.h"
+#include "operators/fusion_conv_add_bn_op.h"
 #include "operators/fusion_conv_add_bn_relu_op.h"
+#include "operators/fusion_conv_add_op.h"
+#include "operators/fusion_conv_add_prelu_op.h"
+#include "operators/fusion_conv_add_relu_op.h"
+#include "operators/fusion_conv_bn_add_relu_op.h"
 #include "operators/fusion_conv_bn_relu_op.h"
 #include "operators/fusion_dwconv_bn_relu_op.h"
+#include "operators/fusion_elementwise_add_relu_op.h"
 #include "operators/fusion_fc_op.h"
+#include "operators/fusion_fc_relu_op.h"
+#include "operators/gru_op.h"
 #include "operators/im2sequence_op.h"
+#include "operators/lookup_op.h"
 #include "operators/lrn_op.h"
 #include "operators/mul_op.h"
 #include "operators/multiclass_nms_op.h"
 #include "operators/pool_op.h"
+#include "operators/prelu_op.h"
 #include "operators/prior_box_op.h"
+#include "operators/quantize_op.h"
 #include "operators/relu_op.h"
 #include "operators/reshape_op.h"
+#include "operators/resize_op.h"
+#include "operators/scale_op.h"
+#include "operators/shape_op.h"
 #include "operators/sigmoid_op.h"
+#include "operators/slice_op.h"
 #include "operators/softmax_op.h"
+#include "operators/split_op.h"
 #include "operators/transpose_op.h"

src/operators/feed_op.h

@@ -24,7 +24,7 @@ namespace operators {
 template <typename DeviceType, typename T>
 class FeedOp : public framework::OperatorBase<DeviceType> {
  public:
-  FeedOp(const string &type, const VariableNameMap &inputs,
+  FeedOp(const std::string &type, const VariableNameMap &inputs,
          const VariableNameMap &outputs, const framework::AttributeMap attrs,
          std::shared_ptr<framework::Scope> scope)
       : framework::OperatorBase<DeviceType>(type, inputs, outputs, attrs,
@@ -38,25 +38,29 @@ class FeedOp : public framework::OperatorBase<DeviceType> {
   }
 
 #ifdef PADDLE_MOBILE_FPGA
+
   void Init() {
     Tensor *output = param_.Out();
-    fpga::format_ofm(output);
+    fpga::format_fp16_ofm(output);
   }
 
   void RunImpl() const {
-    auto input = reinterpret_cast<Tensor *>(param_.InputX());
+    auto input = (Tensor *)const_cast<LoDTensor *>(param_.InputX());  // NOLINT
     fpga::format_image(input);
     auto input_ptr = input->data<float>();
     Tensor *output = param_.Out();
-    auto output_ptr = output->mutable_data<half>();
-
-    fpga::BypassArgs args;
-    args.convert_type = fpga::DATA_FP32_TO_FP16;
-    args.layout_type = fpga::LAYOUT_NO_CONVERT;
-    args.image.address = input_ptr;
-    args.image.channels = input->dims()[1];
-    args.image.height = input->dims()[2];
-    args.image.width = input->dims()[3];
+    auto output_ptr = output->data<float>();
+
+    fpga::BypassArgs args = {fpga::DATA_TYPE_FP32};
+
+    args.input_data_type = fpga::DATA_TYPE_FP32;
+    args.output_data_type = fpga::DATA_TYPE_FP16;
+    args.input_layout_type = fpga::LAYOUT_CHW;
+    args.output_layout_type = fpga::LAYOUT_HWC;
+    args.image.address = (void *)input_ptr;  // NOLINT
+    args.image.channels = (uint32_t)input->dims()[1];
+    args.image.height = (uint32_t)input->dims()[2];
+    args.image.width = (uint32_t)input->dims()[3];
     args.image.pad_height = 0;
     args.image.pad_width = 0;
     args.output.address = output_ptr;

src/operators/kernel/central-arm-func/conv_add_arm_func.h

@@ -129,10 +129,13 @@ void ConvAddCompute(const FusionConvAddParam<CPU> &param) {
     //        param.Paddings(),
     //                               param.Filter(), param.Bias(),
     //                               param.Output(), false);
-
-    math::DepthwiseConv3x3s2p1v2(param.Input(), param.Filter(), param.Output(),
+    if (param.Paddings()[0] == 0) {
+      math::DepthwiseConv3x3s2p0(param.Input(), param.Filter(), param.Output(),
                                  *param.Bias(), true);
-
+    } else {
+      math::DepthwiseConv3x3s2p1v2(param.Input(), param.Filter(),
+                                   param.Output(), *param.Bias(), true);
+    }
   } else {
     ConvAddBasic(param);
   }

src/operators/kernel/central-arm-func/prior_box_arm_func.h

@@ -16,6 +16,7 @@ limitations under the License. */
 #pragma once
 
 #include <algorithm>
+#include <cmath>
 #include <vector>
 
 namespace paddle_mobile {
@@ -89,12 +90,8 @@ void PriorBoxCompute(const PriorBoxParam<CPU> &param) {
       int idx = 0;
       for (size_t s = 0; s < min_sizes.size(); ++s) {
         auto min_size = min_sizes[s];
-        // priors with different aspect ratios
-        for (float ar : aspect_ratios) {
-          box_width = min_size * sqrt(ar) / 2.;
-          box_height = min_size / sqrt(ar) / 2.;
-          /// box_width/2 , / img_width 为了得到feature map 相对于
-          /// 原图的归一化位置的比例。
+        if (param.MinMaxAspectRatiosOrder()) {
+          box_width = box_height = min_size / 2.;
           output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 0] =
               (center_x - box_width) / img_width;
           output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 1] =
@@ -104,24 +101,77 @@ void PriorBoxCompute(const PriorBoxParam<CPU> &param) {
           output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 3] =
               (center_y + box_height) / img_height;
           idx++;
+
+          if (max_sizes.size() > 0) {
+            auto max_size = max_sizes[s];
+            // square prior with size sqrt(minSize * maxSize)
+            box_width = box_height = sqrt(min_size * max_size) / 2.;
+            output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 +
+                                 0] = (center_x - box_width) / img_width;
+            output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 +
+                                 1] = (center_y - box_height) / img_height;
+            output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 +
+                                 2] = (center_x + box_width) / img_width;
+            output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 +
+                                 3] = (center_y + box_height) / img_height;
+            idx++;
+          }
+
+          // priors with different aspect ratios
+          for (float ar : aspect_ratios) {
+            if (fabs(ar - 1.) < 1e-6) {
+              continue;
+            }
+            box_width = min_size * sqrt(ar) / 2.;
+            box_height = min_size / sqrt(ar) / 2.;
+            /// box_width/2 , / img_width 为了得到feature map 相对于
+            /// 原图的归一化位置的比例。
+            output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 +
+                                 0] = (center_x - box_width) / img_width;
+            output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 +
+                                 1] = (center_y - box_height) / img_height;
+            output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 +
+                                 2] = (center_x + box_width) / img_width;
+            output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 +
+                                 3] = (center_y + box_height) / img_height;
+            idx++;
+          }
+
+        } else {
+          // priors with different aspect ratios
+          for (float ar : aspect_ratios) {
+            box_width = min_size * sqrt(ar) / 2.;
+            box_height = min_size / sqrt(ar) / 2.;
+            /// box_width/2 , / img_width 为了得到feature map 相对于
+            /// 原图的归一化位置的比例。
+            output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 +
+                                 0] = (center_x - box_width) / img_width;
+            output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 +
+                                 1] = (center_y - box_height) / img_height;
+            output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 +
+                                 2] = (center_x + box_width) / img_width;
+            output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 +
+                                 3] = (center_y + box_height) / img_height;
+            idx++;
           }
           if (!max_sizes.empty()) {
             auto max_size = max_sizes[s];
             // square prior with size sqrt(minSize * maxSize)
             box_width = box_height = sqrt(min_size * max_size) / 2.;
-          output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 0] =
-              (center_x - box_width) / img_width;
-          output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 1] =
-              (center_y - box_height) / img_height;
-          output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 2] =
-              (center_x + box_width) / img_width;
-          output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 3] =
-              (center_y + box_height) / img_height;
+            output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 +
+                                 0] = (center_x - box_width) / img_width;
+            output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 +
+                                 1] = (center_y - box_height) / img_height;
+            output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 +
+                                 2] = (center_x + box_width) / img_width;
+            output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 +
+                                 3] = (center_y + box_height) / img_height;
             idx++;
           }
         }
       }
     }
+  }
   if (clip) {
     math::Transform trans;
     ClipFunctor<float> clip_func;

src/operators/kernel/fpga/concat_kernel.cpp

@@ -24,10 +24,12 @@ bool ConcatKernel<FPGA, float>::Init(ConcatParam<FPGA> *param) {
   auto inputs = param->Inputs();
   auto out = param->Out();
   auto image_num = inputs.size();
-  auto images_in = (half **)fpga::fpga_malloc(image_num * sizeof(int *));
-  auto scales_in = (float **)fpga::fpga_malloc(image_num * sizeof(float *));
+  auto images_in =
+      (half **)fpga::fpga_malloc(image_num * sizeof(int *));  // NOLINT
+  auto scales_in =
+      (float **)fpga::fpga_malloc(image_num * sizeof(float *));  // NOLINT
   auto channel_num =
-      (uint32_t *)fpga::fpga_malloc(image_num * sizeof(uint32_t));
+      (uint32_t *)fpga::fpga_malloc(image_num * sizeof(uint32_t));  // NOLINT
 
   auto height = inputs[0]->dims()[2];
   auto width = inputs[0]->dims()[3];
@@ -36,22 +38,21 @@ bool ConcatKernel<FPGA, float>::Init(ConcatParam<FPGA> *param) {
     PADDLE_MOBILE_ENFORCE(
         input->dims()[2] == height && input->dims()[3] == width,
         "Image height & width should be unified");
-    images_in[i] = (half *)input->data<float>();
-    channel_num[i] = (uint32_t)inputs[i]->dims()[1];
+    images_in[i] = (half *)input->data<float>();      // NOLINT
+    channel_num[i] = (uint32_t)inputs[i]->dims()[1];  // NOLINT
     scales_in[i] = input->scale;
   }
-  fpga::format_concat_output(out, (int)height, (int)width, (int)image_num,
-                             channel_num);
+  fpga::format_concat_output(out, height, width, image_num, channel_num);
 
-  fpga::ConcatArgs concatArgs;
-  concatArgs.image_num = (uint32_t)image_num;
+  fpga::ConcatArgs concatArgs = {0};
+  concatArgs.image_num = image_num;
   concatArgs.images_in = images_in;
   concatArgs.scales_in = scales_in;
-  concatArgs.image_out = (half *)out->data<float>();
+  concatArgs.image_out = (half *)out->data<float>();  // NOLINT
   concatArgs.scale_out = out->scale;
   concatArgs.channel_num = channel_num;
-  concatArgs.height = (uint32_t)height;
-  concatArgs.width = (uint32_t)width;
+  concatArgs.height = height;
+  concatArgs.width = width;
   param->SetFpgaArgs(concatArgs);
   return true;
 }

src/operators/kernel/fpga/conv_add_bn_kernel.cpp

@@ -23,7 +23,7 @@ template <>
 bool ConvAddBNKernel<FPGA, float>::Init(FusionConvAddBNParam<FPGA> *param) {
   bool relu_enabled = false;
   auto input = const_cast<Tensor *>(param->Input());
-  auto input_ptr = input->data<float>();
+
   auto bias = param->Bias();
   auto bias_ptr = bias->data<float>();
   auto filter = const_cast<Tensor *>(param->Filter());
@@ -62,11 +62,11 @@ bool ConvAddBNKernel<FPGA, float>::Init(FusionConvAddBNParam<FPGA> *param) {
   fpga::format_filter(filter, max_value, param->Groups());
 
   int element_num_per_div =
-      fpga::get_element_num_per_div(filter, param->Groups());
+      fpga::get_filter_num_per_div(filter, param->Groups());
   fpga::format_bias_scale_array(&bs_ptr, element_num_per_div, channel);
-  fpga::format_ofm(out);
+  fpga::format_fp16_ofm(out);
 
-  fpga::WrapperConvArgs conv_arg;
+  fpga::WrapperConvArgs conv_arg = {0};
   fpga::fill_conv_arg(&conv_arg, input, out, filter, relu_enabled,
                       param->Groups(), param->Strides()[0], param->Strides()[1],
                       param->Paddings()[0], param->Paddings()[1], bs_ptr);
@@ -80,7 +80,6 @@ void ConvAddBNKernel<FPGA, float>::Compute(
     const FusionConvAddBNParam<FPGA> &param) const {
   fpga::ComputeFpgaConv(param.FpgaArgs());
 }
-template class ConvAddBNKernel<FPGA, float>;
 
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/fpga/conv_add_bn_relu_kernel.cpp

@@ -24,7 +24,6 @@ bool ConvAddBNReluKernel<FPGA, float>::Init(
     FusionConvAddBNReluParam<FPGA> *param) {
   bool relu_enabled = true;
   auto input = const_cast<Tensor *>(param->Input());
-  auto input_ptr = input->data<float>();
   const Tensor *bias = param->Bias();
   auto bias_ptr = bias->data<float>();
   auto filter = const_cast<Tensor *>(param->Filter());
@@ -39,7 +38,8 @@ bool ConvAddBNReluKernel<FPGA, float>::Init(
                         "Output channel should be equal to bias number");
 
   const int channel = out->dims()[1];
-  auto bs_ptr = (float *)fpga::fpga_malloc(2 * channel * sizeof(float));
+  auto bs_ptr =
+      (float *)fpga::fpga_malloc(2 * channel * sizeof(float));  // NOLINT
   auto new_scale = new Tensor();
   auto new_bias = new Tensor();
   auto new_scale_ptr = new_scale->mutable_data<float>({channel});
@@ -58,16 +58,14 @@ bool ConvAddBNReluKernel<FPGA, float>::Init(
 
   float max_value = fpga::filter_find_max(filter);
   fpga::format_filter(filter, max_value, param->Groups());
-  auto filter_ptr = filter->data<float>();
 
   int element_num_per_div =
-      fpga::get_element_num_per_div(filter, param->Groups());
+      fpga::get_filter_num_per_div(filter, param->Groups());
   fpga::format_bias_scale_array(&bs_ptr, element_num_per_div, channel);
 
-  fpga::format_ofm(out);
-  auto out_ptr = out->mutable_data<float>();
+  fpga::format_fp16_ofm(out);
 
-  fpga::WrapperConvArgs conv_arg;
+  fpga::WrapperConvArgs conv_arg = {0};
   fpga::fill_conv_arg(&conv_arg, input, out, filter, relu_enabled,
                       param->Groups(), param->Strides()[0], param->Strides()[1],
                       param->Paddings()[0], param->Paddings()[1], bs_ptr);
@@ -80,7 +78,6 @@ void ConvAddBNReluKernel<FPGA, float>::Compute(
     const FusionConvAddBNReluParam<FPGA> &param) const {
   fpga::ComputeFpgaConv(param.FpgaArgs());
 }
-template class ConvAddBNReluKernel<FPGA, float>;
 
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/fpga/conv_add_relu_kernel.cpp

@@ -23,7 +23,6 @@ template <>
 bool ConvAddReluKernel<FPGA, float>::Init(FusionConvAddReluParam<FPGA> *param) {
   bool relu_enabled = true;
   auto input = const_cast<Tensor *>(param->Input());
-  auto input_ptr = input->data<float>();
   const Tensor *bias = param->Bias();
   auto bias_ptr = bias->data<float>();
   auto filter = const_cast<Tensor *>(param->Filter());
@@ -32,7 +31,8 @@ bool ConvAddReluKernel<FPGA, float>::Init(FusionConvAddReluParam<FPGA> *param) {
   PADDLE_MOBILE_ENFORCE(out->dims()[1] == bias->dims()[0],
                         "Output channel should be equal to bias number");
   int channel = out->dims()[1];
-  auto bs_ptr = (float *)fpga::fpga_malloc(2 * channel * sizeof(float));
+  auto bs_ptr =
+      (float *)fpga::fpga_malloc(2 * channel * sizeof(float));  // NOLINT
   for (int i = 0; i < channel; i++) {
     bs_ptr[i + channel] = 1;
     bs_ptr[i] = bias_ptr[i];
@@ -40,16 +40,14 @@ bool ConvAddReluKernel<FPGA, float>::Init(FusionConvAddReluParam<FPGA> *param) {
 
   float max_value = fpga::filter_find_max(filter);
   fpga::format_filter(filter, max_value, param->Groups());
-  auto filter_ptr = filter->data<float>();
 
   int element_num_per_div =
-      fpga::get_element_num_per_div(filter, param->Groups());
+      fpga::get_filter_num_per_div(filter, param->Groups());
   fpga::format_bias_scale_array(&bs_ptr, element_num_per_div, channel);
 
-  fpga::format_ofm(out);
-  auto out_ptr = out->mutable_data<float>();
+  fpga::format_fp16_ofm(out);
 
-  fpga::WrapperConvArgs conv_arg;
+  fpga::WrapperConvArgs conv_arg = {0};
   fpga::fill_conv_arg(&conv_arg, input, out, filter, relu_enabled,
                       param->Groups(), param->Strides()[0], param->Strides()[1],
                       param->Paddings()[0], param->Paddings()[1], bs_ptr);
@@ -62,7 +60,6 @@ void ConvAddReluKernel<FPGA, float>::Compute(
     const FusionConvAddReluParam<FPGA> &param) const {
   fpga::ComputeFpgaConv(param.FpgaArgs());
 }
-template class ConvAddReluKernel<FPGA, float>;
 
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/fpga/conv_bn_kernel.cpp

@@ -15,7 +15,6 @@ limitations under the License. */
 #ifdef FUSION_CONVBN_OP
 
 #include "operators/kernel/conv_bn_kernel.h"
-#include "fpga/api.h"
 
 namespace paddle_mobile {
 namespace operators {
@@ -24,7 +23,6 @@ template <>
 bool ConvBNKernel<FPGA, float>::Init(FusionConvBNParam<FPGA> *param) {
   bool relu_enabled = false;
   auto input = const_cast<Tensor *>(param->Input());
-  auto input_ptr = input->data<float>();
   auto filter = const_cast<Tensor *>(param->Filter());
   auto out = param->Output();
   auto bn_mean_ptr = param->InputMean()->data<float>();
@@ -34,10 +32,9 @@ bool ConvBNKernel<FPGA, float>::Init(FusionConvBNParam<FPGA> *param) {
   const float epsilon = param->Epsilon();
   PADDLE_MOBILE_ENFORCE(out->dims()[1] == param->InputBias()->dims()[0],
                         "Output channel should be equal to bias number");
-
   const int channel = out->dims()[1];
   auto bs_ptr =
-      reinterpret_cast<float *>(fpga::fpga_malloc(2 * channel * sizeof(float)));
+      (float *)fpga::fpga_malloc(2 * channel * sizeof(float));  // // NOLINT
   auto new_scale = new Tensor();
   auto new_bias = new Tensor();
   auto new_scale_ptr = new_scale->mutable_data<float>({channel});
@@ -55,16 +52,14 @@ bool ConvBNKernel<FPGA, float>::Init(FusionConvBNParam<FPGA> *param) {
 
   float max_value = fpga::filter_find_max(filter);
   fpga::format_filter(filter, max_value, param->Groups());
-  auto filter_ptr = filter->data<float>();
 
   int element_num_per_div =
-      fpga::get_element_num_per_div(filter, param->Groups());
+      fpga::get_filter_num_per_div(filter, param->Groups());
   fpga::format_bias_scale_array(&bs_ptr, element_num_per_div, channel);
 
-  fpga::format_ofm(out);
-  auto out_ptr = out->mutable_data<float>();
+  fpga::format_fp16_ofm(out);
 
-  fpga::WrapperConvArgs conv_arg;
+  fpga::WrapperConvArgs conv_arg = {0};
   fpga::fill_conv_arg(&conv_arg, input, out, filter, relu_enabled,
                       param->Groups(), param->Strides()[0], param->Strides()[1],
                       param->Paddings()[0], param->Paddings()[1], bs_ptr);
@@ -77,7 +72,6 @@ void ConvBNKernel<FPGA, float>::Compute(
     const FusionConvBNParam<FPGA> &param) const {
   fpga::ComputeFpgaConv(param.FpgaArgs());
 }
-template class ConvBNKernel<FPGA, float>;
 
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/fpga/conv_bn_relu_kernel.cpp

@@ -23,7 +23,6 @@ template <>
 bool ConvBNReluKernel<FPGA, float>::Init(FusionConvBNReluParam<FPGA> *param) {
   bool relu_enabled = true;
   auto input = const_cast<Tensor *>(param->Input());
-  auto input_ptr = input->data<float>();
   auto filter = const_cast<Tensor *>(param->Filter());
   auto out = param->Output();
   auto bn_mean_ptr = param->InputMean()->data<float>();
@@ -34,7 +33,8 @@ bool ConvBNReluKernel<FPGA, float>::Init(FusionConvBNReluParam<FPGA> *param) {
   PADDLE_MOBILE_ENFORCE(out->dims()[1] == param->InputBias()->dims()[0],
                         "Output channel should be equal to bias number");
   const int channel = out->dims()[1];
-  auto bs_ptr = (float *)fpga::fpga_malloc(2 * channel * sizeof(float));
+  auto bs_ptr =
+      (float *)fpga::fpga_malloc(2 * channel * sizeof(float));  // NOLINT
   auto new_scale = new Tensor();
   auto new_bias = new Tensor();
   auto new_scale_ptr = new_scale->mutable_data<float>({channel});
@@ -52,29 +52,14 @@ bool ConvBNReluKernel<FPGA, float>::Init(FusionConvBNReluParam<FPGA> *param) {
 
   float max_value = fpga::filter_find_max(filter);
   fpga::format_filter(filter, max_value, param->Groups());
-  auto filter_ptr = filter->data<float>();
 
   int element_num_per_div =
-      fpga::get_element_num_per_div(filter, param->Groups());
+      fpga::get_filter_num_per_div(filter, param->Groups());
   fpga::format_bias_scale_array(&bs_ptr, element_num_per_div, channel);
 
-  fpga::format_ofm(out);
-  auto out_ptr = out->mutable_data<float>();
+  fpga::format_fp16_ofm(out);
 
-  fpga::WrapperConvArgs convArgs;
-  convArgs.group_num = (uint32_t)param->Groups();
-  convArgs.split_num = (uint32_t)fpga::get_plit_num(filter);
-  convArgs.filter_num = (uint32_t)filter->dims()[0];
-  convArgs.output.address = out_ptr;
-  convArgs.output.scale_address = out->scale;
-  convArgs.conv_args = (fpga::ConvArgs *)fpga::fpga_malloc(
-      convArgs.split_num * sizeof(fpga::ConvArgs));
-  param->SetFpgaArgs(convArgs);
-
-  int element_num = fpga::get_aligned_filter_element_num(
-      filter->dims()[1] * filter->dims()[2] * filter->dims()[3]);
-
-  fpga::WrapperConvArgs conv_arg;
+  fpga::WrapperConvArgs conv_arg = {0};
   fpga::fill_conv_arg(&conv_arg, input, out, filter, relu_enabled,
                       param->Groups(), param->Strides()[0], param->Strides()[1],
                       param->Paddings()[0], param->Paddings()[1], bs_ptr);
@@ -87,7 +72,6 @@ void ConvBNReluKernel<FPGA, float>::Compute(
     const FusionConvBNReluParam<FPGA> &param) const {
   fpga::ComputeFpgaConv(param.FpgaArgs());
 }
-template class ConvBNReluKernel<FPGA, float>;
 
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/fpga/dropout_kernel.cpp

@@ -27,13 +27,7 @@ bool DropoutKernel<FPGA, float>::Init(DropoutParam<FPGA> *param) {
 
 template <>
 void DropoutKernel<FPGA, float>::Compute(
-    const DropoutParam<FPGA> &param) const {
-  // auto *input_x = param.InputX();
-  // auto *out = param.Out();
-  // auto input_x_ptr = input_x->data<float>();
-  // auto out_ptr = out->mutable_data<float>();
-  // out_ptr = const_cast<float *>(input_x_ptr);
-}
+    const DropoutParam<FPGA> &param) const {}
 
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/fpga/elementwise_add_relu_kernel.cpp

@@ -27,10 +27,10 @@ bool ElementwiseAddReluKernel<FPGA, float>::Init(
   auto *out = param->Out();
   auto input_x_ptr = input_x->data<float>();
   auto input_y_ptr = input_y->data<float>();
-  fpga::format_ofm(out);
+  fpga::format_fp16_ofm(out);
   auto out_ptr = out->mutable_data<float>();
 
-  fpga::EWAddArgs ewaddArgs;
+  fpga::EWAddArgs ewaddArgs = {0};
   ewaddArgs.relu_enabled = relu_enabled;
   ewaddArgs.const0 = 1;
   ewaddArgs.const1 = 1;

src/operators/kernel/fpga/fc_relu_kernel.cpp

@@ -21,7 +21,6 @@ template <>
 bool FusionFcReluKernel<FPGA, float>::Init(FusionFcReluParam<FPGA> *param) {
   bool relu_enabled = true;
   auto input_x = const_cast<LoDTensor *>(param->InputX());
-  auto input_x_ptr = input_x->data<float>();
   auto filter = const_cast<Tensor *>(param->InputY());
   auto input_z = param->InputZ();
   auto input_z_ptr = input_z->data<float>();
@@ -29,7 +28,8 @@ bool FusionFcReluKernel<FPGA, float>::Init(FusionFcReluParam<FPGA> *param) {
   PADDLE_MOBILE_ENFORCE(input_x->dims()[1] == filter->dims()[0],
                         "Image channel should be equal to weight number");
   int channel = (uint32_t)out->dims()[1];
-  auto bs_ptr = (float *)fpga::fpga_malloc(2 * channel * sizeof(float));
+  auto bs_ptr =
+      (float *)fpga::fpga_malloc(2 * channel * sizeof(float));  // NOLINT
   for (int i = 0; i < channel; i++) {
     bs_ptr[i + channel] = 1;
     bs_ptr[i] = input_z_ptr[i];
@@ -47,14 +47,12 @@ bool FusionFcReluKernel<FPGA, float>::Init(FusionFcReluParam<FPGA> *param) {
   filter->Resize(framework::make_ddim({num, filter_channel, height, width}));
   float max_value = fpga::filter_find_max(filter);
   fpga::format_filter(filter, max_value, 1);
-  auto filter_ptr = filter->data<float>();
 
-  int element_num_per_div = fpga::get_element_num_per_div(filter, 1);
+  int element_num_per_div = fpga::get_filter_num_per_div(filter, 1);
   fpga::format_bias_scale_array(&bs_ptr, element_num_per_div, channel);
+  fpga::format_fp16_ofm(out);
 
-  auto out_ptr = out->mutable_data<float>();
-
-  fpga::WrapperConvArgs conv_arg;
+  fpga::WrapperConvArgs conv_arg = {0};
   fpga::fill_conv_arg(&conv_arg, input_x, out, filter, relu_enabled, 1, 1, 1, 0,
                       0, bs_ptr);
   param->SetFpgaArgs(conv_arg);
@@ -64,7 +62,7 @@ template <>
 void FusionFcReluKernel<FPGA, float>::Compute(
     const FusionFcReluParam<FPGA> &param) const {
   fpga::ComputeFpgaConv(param.FpgaArgs());
-};
+}
 
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/fpga/fusion_fc_kernel.cpp

@@ -22,7 +22,6 @@ template <>
 bool FusionFcKernel<FPGA, float>::Init(FusionFcParam<FPGA> *param) {
   bool relu_enabled = false;
   auto input_x = const_cast<LoDTensor *>(param->InputX());
-  auto input_x_ptr = input_x->data<float>();
   auto filter = const_cast<Tensor *>(param->InputY());
   const Tensor *input_z = param->InputZ();
   auto input_z_ptr = input_z->data<float>();
@@ -31,7 +30,8 @@ bool FusionFcKernel<FPGA, float>::Init(FusionFcParam<FPGA> *param) {
   PADDLE_MOBILE_ENFORCE(input_x->dims()[1] == filter->dims()[0],
                         "Image channel should be equal to weight number");
   int channel = (uint32_t)out->dims()[1];
-  auto bs_ptr = (float *)fpga::fpga_malloc(2 * channel * sizeof(float));
+  auto bs_ptr =
+      (float *)fpga::fpga_malloc(2 * channel * sizeof(float));  // NOLINT
   for (int i = 0; i < channel; i++) {
     bs_ptr[i + channel] = 1;
     bs_ptr[i] = input_z_ptr[i];
@@ -48,14 +48,12 @@ bool FusionFcKernel<FPGA, float>::Init(FusionFcParam<FPGA> *param) {
   filter->Resize(framework::make_ddim({num, filter_channel, height, width}));
   float max_value = fpga::filter_find_max(filter);
   fpga::format_filter(filter, max_value, 1);
-  auto filter_ptr = filter->data<float>();
 
-  int element_num_per_div = fpga::get_element_num_per_div(filter, 1);
+  int element_num_per_div = fpga::get_filter_num_per_div(filter, 1);
   fpga::format_bias_scale_array(&bs_ptr, element_num_per_div, channel);
+  fpga::format_fp16_ofm(out);
 
-  auto out_ptr = out->mutable_data<float>();
-
-  fpga::WrapperConvArgs conv_arg;
+  fpga::WrapperConvArgs conv_arg = {0};
   fpga::fill_conv_arg(&conv_arg, input_x, out, filter, relu_enabled, 1, 1, 1, 0,
                       0, bs_ptr);
   param->SetFpgaArgs(conv_arg);

src/operators/kernel/fpga/pool_kernel.cpp

@@ -24,13 +24,13 @@ bool PoolKernel<FPGA, float>::Init(PoolParam<FPGA> *param) {
   auto *input = const_cast<Tensor *>(param->Input());
   auto input_ptr = input->data<float>();
   Tensor *output = param->Output();
-  fpga::format_ofm(output);
+  fpga::format_fp16_ofm(output);
   auto output_ptr = output->mutable_data<float>();
   vector<int> ksize = param->Ksize();
   vector<int> strides = param->Strides();
   vector<int> paddings = param->Paddings();
 
-  fpga::PoolingArgs poolArgs;
+  fpga::PoolingArgs poolArgs = {0};
   poolArgs.image.address = input_ptr;
   poolArgs.image.channels = (uint32_t)input->dims()[1];
   poolArgs.image.height = (uint32_t)input->dims()[2];
@@ -39,7 +39,7 @@ bool PoolKernel<FPGA, float>::Init(PoolParam<FPGA> *param) {
   poolArgs.image.pad_width = (uint32_t)paddings[1];
   poolArgs.image.scale_address = input->scale;
   poolArgs.output.address = output_ptr;
-  poolArgs.output.scale_address = input->scale;
+  poolArgs.output.scale_address = output->scale;
   poolArgs.kernel.height = (uint32_t)ksize[0];
   poolArgs.kernel.width = (uint32_t)ksize[1];
   poolArgs.kernel.stride_h = (uint32_t)strides[0];
@@ -50,9 +50,7 @@ bool PoolKernel<FPGA, float>::Init(PoolParam<FPGA> *param) {
 
 template <>
 void PoolKernel<FPGA, float>::Compute(const PoolParam<FPGA> &param) const {
-#ifdef PADDLE_MOBILE_FPGA
   fpga::ComputeFpgaPool(param.FpgaArgs());
-#endif
 }
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/fpga/softmax_kernel.cpp

@@ -43,17 +43,24 @@ bool SoftmaxKernel<FPGA, float>::Init(SoftmaxParam<FPGA> *param) {
   args.output.scale_address = float_input->scale;
   param->SetFloatInput(float_input);
   param->SetFpgaArgs(args);
-
   return true;
 }
 
 template <>
 void SoftmaxKernel<FPGA, float>::Compute(
     const SoftmaxParam<FPGA> &param) const {
-  // SoftmaxCompute<float>(param);
+  Tensor *in_x = param.FloatInput();
+  Tensor *out = param.Out();
+
+  fpga::PerformBypass(param.FpgaArgs());
+  fpga::fpga_invalidate(
+      (void *)in_x->data<float>(),  // NOLINT
+      fpga::get_align_image_cw(in_x->dims()[1]) * sizeof(float));
+
+  math::SoftmaxFuntor<CPU, float>()(in_x, out);
+  fpga::fpga_flush(out->data<float>(), out->memory_size());
 }
 
-template class SoftmaxKernel<FPGA, float>;
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/math/depthwise_conv_3x3.cpp

@@ -1465,7 +1465,7 @@ void DepthwiseConvAddBNRelu3x3s2p1v2(const Tensor *input, const Tensor *filter,
                                      Tensor *output, const Tensor *new_scale,
                                      const Tensor *new_bias, bool if_relu) {
 #if __ARM_NEON
-  //#ifdef _OPENMP
+  // #ifdef _OPENMP
   //  const float *newscale_data = new_scale->data<float>();
   //  const float *newbias_data = new_bias->data<float>();
   //
@@ -1645,7 +1645,7 @@ void DepthwiseConvAddBNRelu3x3s2p1v2(const Tensor *input, const Tensor *filter,
   //    }
   //  }
   //
-  //#else
+  // #else
 
   const float *input_data = input->data<float>();
   const float *filter_data = filter->data<float>();
@@ -1877,7 +1877,104 @@ void DepthwiseConvAddBNRelu3x3s2p1v2(const Tensor *input, const Tensor *filter,
     input_data += inhxw * c;
     output_data += outhxw * c;
   }
-//#endif
+// #endif
+#endif
+}
+
+void DepthwiseConv3x3s2p0(const Tensor *input, const Tensor *filter,
+                          Tensor *output, Tensor bias, bool if_bias) {
+#if __ARM_NEON
+
+  const int batch_size = static_cast<int>(input->dims()[0]);
+  const int input_channel = static_cast<int>(input->dims()[1]);
+
+  const int input_height = static_cast<int>(input->dims()[2]);
+  const int input_width = static_cast<int>(input->dims()[3]);
+  const int output_height = static_cast<int>(output->dims()[2]);
+  const int output_width = static_cast<int>(output->dims()[3]);
+  const int inhxw = input_height * input_width;
+  const int outhxw = output_height * output_width;
+
+  float32x4_t zero = vdupq_n_f32(0.0);
+  for (int b = 0; b < batch_size; b++) {
+#pragma omp parallel for
+    for (int c = 0; c < input_channel; c++) {
+      const float *filter_data = filter->data<float>() + c * 9;
+      const float *input_data = input->data<float>() + c * inhxw;
+      const float *bias_data = bias.data<float>() + c;
+      float *output_data = output->data<float>() + c * outhxw;
+      float w00 = filter_data[0];
+      float w01 = filter_data[1];
+      float w02 = filter_data[2];
+      float w10 = filter_data[3];
+      float w11 = filter_data[4];
+      float w12 = filter_data[5];
+      float w20 = filter_data[6];
+      float w21 = filter_data[7];
+      float w22 = filter_data[8];
+
+      float32x4_t biasv = vld1q_dup_f32(bias_data);
+
+      for (int i = 0; i < output_height; i += 1) {
+        for (int m = 0; m < output_width - 2; m += 3) {
+          float *output_ptr = output_data + i * output_width + m;
+          float32x4x2_t input_buff_top{}, input_buff_mid{}, input_buff_bottom{};
+          float32x4_t in0, in1, in2, in3, in4, in5, tmp0, tmp1, tmp2, tmp3,
+              tmp4, tmp5, out0;
+          input_buff_top =
+              vld2q_f32(input_data + (2 * i) * input_width + (2 * m));
+          input_buff_mid =
+              vld2q_f32(input_data + (2 * i + 1) * input_width + (2 * m));
+          input_buff_bottom =
+              vld2q_f32(input_data + (2 * i + 2) * input_width + (2 * m));
+
+          in0 = input_buff_top.val[0];
+          tmp0 = input_buff_top.val[1];
+          tmp1 = vextq_f32(in0, zero, 1);
+
+          in2 = input_buff_mid.val[0];
+          tmp2 = input_buff_mid.val[1];
+          tmp3 = vextq_f32(in2, zero, 1);
+
+          in4 = input_buff_bottom.val[0];
+          tmp4 = input_buff_bottom.val[1];
+          tmp5 = vextq_f32(in4, zero, 1);
+
+          out0 = vmulq_n_f32(in0, w00);
+          out0 = vmlaq_n_f32(out0, tmp0, w01);
+          out0 = vmlaq_n_f32(out0, tmp1, w02);
+          out0 = vmlaq_n_f32(out0, in2, w10);
+          out0 = vmlaq_n_f32(out0, tmp2, w11);
+          out0 = vmlaq_n_f32(out0, tmp3, w12);
+          out0 = vmlaq_n_f32(out0, in4, w20);
+          out0 = vmlaq_n_f32(out0, tmp4, w21);
+          out0 = vmlaq_n_f32(out0, tmp5, w22);
+          out0 = vaddq_f32(out0, biasv);
+
+          vst1q_lane_f32(output_ptr, out0, 0);
+          vst1q_lane_f32(output_ptr + 1, out0, 1);
+          vst1q_lane_f32(output_ptr + 2, out0, 2);
+        }
+        int m;
+        for (m = 0; m < output_width - 2; m += 3) {
+        }
+        for (int j = m; j < output_width; j++) {
+          output_data[i * output_width + j] =
+              input_data[(2 * i - 1) * input_width + 2 * j - 1] * w00 +
+              input_data[(2 * i - 1) * input_width + 2 * j] * w01 +
+              input_data[(2 * i - 1) * input_width + 2 * j + 1] * w02 +
+              input_data[(2 * i) * input_width + 2 * j - 1] * w10 +
+              input_data[(2 * i) * input_width + 2 * j] * w11 +
+              input_data[(2 * i) * input_width + 2 * j + 1] * w12 +
+              input_data[(2 * i + 1) * input_width + 2 * j - 1] * w20 +
+              input_data[(2 * i + 1) * input_width + 2 * j] * w21 +
+              input_data[(2 * i + 1) * input_width + 2 * j + 1] * w22;
+          output_data[i * output_width + j] += *bias_data;
+        }
+      }
+    }
+  }
+
 #endif
 }
 

src/operators/math/depthwise_conv_3x3.h

@@ -43,6 +43,9 @@ void DepthwiseConv3x3s2p1v2(const Tensor *input, const Tensor *filter,
 void DepthwiseConvAddBNRelu3x3s2p1v2(const Tensor *input, const Tensor *filter,
                                      Tensor *output, const Tensor *new_scale,
                                      const Tensor *new_bias, bool if_relu);
+
+void DepthwiseConv3x3s2p0(const Tensor *input, const Tensor *filter,
+                          Tensor *output, Tensor bias, bool if_bias);
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/math/im2col.cpp

@@ -74,7 +74,7 @@ class Im2ColFunctor<ColFormat::kCFO, CPU, T> {
     const int isize = im_height;
     bool pad1 = padding[0] > 0;
     bool pad2 =
-        (pad1 &&
+        (pad1 && padding[1] &&
          (((isize - 2 * padding[0] + filter_height) % stride[0] == 0) ? 1 : 0));
     int fill = isize % 2;
     if (stride[0] == 1 && filter_height == 3 && pad1 && pad2 &&

src/operators/math/math_func_neon.h

@@ -38,7 +38,6 @@ limitations under the License. */
  *
  *  (this is the zlib license)
  */
-
 #pragma once
 #include <arm_neon.h>
 

src/operators/math/math_function.cpp

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #include "operators/math/math_function.h"
 #include <cstring>
+#include <string>
 #include "operators/math/gemm.h"
 
 namespace paddle_mobile {
@@ -36,6 +37,27 @@ void matmul<float>(const framework::Tensor &matrix_a, bool trans_a,
   int N = dim_out[1];
   int K = (!trans_a) ? dim_a[1] : dim_a[0];
 
+  if (trans_a) {
+    int numel = matrix_a.numel();
+    int m = matrix_a.dims()[0];
+    int n = matrix_a.dims()[1];
+    float *tmp = (float *)(matrix_a.data<float>());  // NOLINT
+    float *a = static_cast<float *>(
+        paddle_mobile::memory::Alloc(sizeof(float) * numel));
+    int index = 0;
+    for (int j = 0; j < n; j++) {
+      for (int i = 0; i < m; i++) {
+        a[index++] = tmp[i * n + j];
+      }
+    }
+#ifdef _OPENMP
+    Sgemm_omp(M, N, K, alpha, a, K, matrix_b.data<float>(), N, beta,
+              matrix_out->data<float>(), N, relu, bias);
+#else
+    Sgemm(M, N, K, alpha, a, K, matrix_b.data<float>(), N, beta,
+          matrix_out->data<float>(), N, relu, bias);
+#endif
+  } else {
 #ifdef _OPENMP
     Sgemm_omp(M, N, K, alpha, matrix_a.data<float>(), K, matrix_b.data<float>(),
               N, beta, matrix_out->data<float>(), N, relu, bias);
@@ -43,6 +65,7 @@ void matmul<float>(const framework::Tensor &matrix_a, bool trans_a,
     Sgemm(M, N, K, alpha, matrix_a.data<float>(), K, matrix_b.data<float>(), N,
           beta, matrix_out->data<float>(), N, relu, bias);
 #endif
+  }
 }
 
 template <>
@@ -104,7 +127,7 @@ struct ClearTensor<CPU, T> {
   void operator()(framework::Tensor *tensor) {
     auto size = tensor->numel();
     auto *tensor_data = tensor->data<float>();
-    memset((void *)tensor_data, 0, sizeof(T) * size);
+    memset((void *)tensor_data, 0, sizeof(T) * size);  // NOLINT
   }
 };
 

src/operators/math/pool_3x3.cpp

@@ -31,251 +31,428 @@ using std::min;
 using std::vector;
 void Pool3x3Avgs1p1(const Tensor *input, Tensor *output) {
 #if __ARM_NEON
-  const int batch_size = input->dims()[0];
+  const int batch_size = static_cast<int>(input->dims()[0]);
+  const int input_channel = static_cast<int>(input->dims()[1]);
 
-  const int h_in = input->dims()[2];
+  const int input_height = static_cast<int>(input->dims()[2]);
+  const int input_width = static_cast<int>(input->dims()[3]);
+  const int output_height = static_cast<int>(output->dims()[2]);
+  const int output_width = static_cast<int>(output->dims()[3]);
 
-  const int w_in = input->dims()[3];
+  const int hxw = input_height * input_width;
 
-  const int output_channels = output->dims()[1];
-
-  const int h_out = output->dims()[2];
-  const int w_out = output->dims()[3];
-  const int outputdata_channel_stride = h_out * w_out;
-  const int inputdata_channel_stride = h_in * w_in;
-  const int input_batch_stride = output_channels * inputdata_channel_stride;
-  const int output_batch_stride = output_channels * outputdata_channel_stride;
-  float *out_data = output->data<float>();
-  const float *input_data = input->data<float>();
+  const int l = input_height;
 
   const float coef = 1.0 / 9.0;
-  for (int k = 0; k < batch_size; ++k) {
-#pragma omp parallel for
-    for (int c = 0; c < output_channels; ++c) {
-      const float *input_seg = input_data + c * inputdata_channel_stride;
-      float *output_seg = out_data + c * outputdata_channel_stride;
-      // four corner point
-      output_seg[0] = (input_seg[0] + input_seg[1] + input_seg[w_in] +
-                       input_seg[w_in + 1]) *
-                      coef;
-      output_seg[w_out - 1] =
-          (input_seg[w_in - 2] + input_seg[w_in - 1] + input_seg[w_in * 2 - 2] +
-           input_seg[2 * w_in - 1]) *
-          coef;
-      output_seg[(h_out - 1) * w_out] =
-          (input_seg[(h_in - 2) * w_in] + input_seg[(h_in - 2) * w_in + 1] +
-           input_seg[(h_in - 1) * w_in] + input_seg[(h_in - 1) * w_in + 1]) *
-          coef;
-      output_seg[h_out * w_out - 1] =
-          (input_seg[h_in * w_in - 1] + input_seg[h_in * w_in - 2] +
-           input_seg[(h_in - 1) * w_in - 1] +
-           input_seg[(h_in - 1) * w_in - 2]) *
-          coef;
-      // left side & right side
-      for (int i = 1; i < h_in - 1; ++i) {
-        output_seg[i * w_out] =
-            (input_seg[i * w_in - w_in] + input_seg[i * w_in - w_in + 1] +
-             input_seg[i * w_in] + input_seg[i * w_in + 1] +
-             input_seg[i * w_in + w_in] + input_seg[i * w_in + w_in + 1]) *
-            coef;
-        output_seg[i * w_out + w_out - 1] =
-            (input_seg[i * w_in - w_in + w_in - 2] +
-             input_seg[i * w_in - w_in + 1 + w_in - 2] +
-             input_seg[i * w_in + w_in - 2] +
-             input_seg[i * w_in + 1 + w_in - 2] +
-             input_seg[i * w_in + w_in + w_in - 2] +
-             input_seg[i * w_in + w_in + 1 + w_in - 2]) *
-            coef;
-      }
-      // top 1 row & bottom 1 row
-      const float *input_tmp = input_seg;
+  const float coef1 = 1.0 / 6.0;
+  const float coef2 = 1.0 / 4.0;
 
-      float32x4_t in0, in1, in2, in3, in4, in5, in6, in7, tmp0, tmp1, tmp2,
-          tmp3, tmp4, tmp5, sum, out0;
   float32x4_t v_coef = vdupq_n_f32(coef);
-      in0 = vld1q_f32(input_tmp);
-      in2 = vld1q_f32(input_tmp + w_in);
-      const float *input_tmp_end = input_tmp + (h_in - 2) * w_in;
-      in4 = vld1q_f32(input_tmp_end);
-      in6 = vld1q_f32(input_tmp_end + w_in);
-      int c_mid = w_out - 2;
-      auto output_ptr = output_seg + 1;
-      for (; c_mid > 3; c_mid -= 4) {
-        in1 = vld1q_f32(input_tmp + 4);
-        in3 = vld1q_f32(input_tmp + w_in + 4);
+  float32x4_t v_coef1 = vdupq_n_f32(coef1);
+
+  for (int b = 0; b < batch_size; b++) {
+#pragma omp parallel for
+    for (int c = 0; c < input_channel; c++) {
+      const float *input_data = input->data<float>() + c * hxw;
+      float *output_data = output->data<float>() + c * hxw;
+
+      for (int i = 1; i < output_height - 1; i++) {
+        float *output_ptr;
+        float32x4_t in0, in1, in2, in3, in4, in5, tmp0, tmp1, tmp2, tmp3, tmp4,
+            tmp5, out0;
+        for (int m = 1; m < output_width - 4; m += 4) {
+          output_ptr = output_data + i * output_width + m;
+          in0 = vld1q_f32(input_data + (i - 1) * input_width + m - 1);
+          in1 = vld1q_f32(input_data + (i - 1) * input_width + m + 3);
+          in2 = vld1q_f32(input_data + i * input_width + m - 1);
+          in3 = vld1q_f32(input_data + i * input_width + m + 3);
+          in4 = vld1q_f32(input_data + (i + 1) * input_width + m - 1);
+          in5 = vld1q_f32(input_data + (i + 1) * input_width + m + 3);
 
           tmp0 = vextq_f32(in0, in1, 1);
           tmp1 = vextq_f32(in0, in1, 2);
-
           tmp2 = vextq_f32(in2, in3, 1);
           tmp3 = vextq_f32(in2, in3, 2);
+          tmp4 = vextq_f32(in4, in5, 1);
+          tmp5 = vextq_f32(in4, in5, 2);
 
-        sum = vaddq_f32(in0, tmp0);
-        sum = vaddq_f32(sum, tmp1);
-        sum = vaddq_f32(sum, in2);
-        sum = vaddq_f32(sum, tmp2);
-        sum = vaddq_f32(sum, tmp3);
-
-        vst1q_f32(output_ptr, vmulq_f32(sum, v_coef));
-
-        in5 = vld1q_f32(input_tmp_end + 4);
-        in7 = vld1q_f32(input_tmp_end + w_in + 4);
-
-        tmp0 = vextq_f32(in4, in5, 1);
-        tmp1 = vextq_f32(in4, in5, 2);
-        tmp2 = vextq_f32(in6, in7, 1);
-        tmp3 = vextq_f32(in6, in7, 2);
-
-        sum = vaddq_f32(in0, tmp0);
-        sum = vaddq_f32(sum, tmp1);
-        sum = vaddq_f32(sum, in2);
-        sum = vaddq_f32(sum, tmp2);
-        sum = vaddq_f32(sum, tmp3);
-
-        vst1q_f32(output_ptr + (h_out - 1) * w_out, vmulq_f32(sum, v_coef));
-
-        // can optimize to each 8 stride.
-        input_tmp += 4;
-        input_tmp_end += 4;
-        output_ptr += 4;
-        in0 = in1;
-        in2 = in3;
-        in4 = in5;
-        in6 = in7;
-      }
-      // top right remain
-      float32x4_t pad0 = vdupq_n_f32(input_seg[w_in - 1]);
-      float32x4_t pad1 = vdupq_n_f32(input_seg[2 * w_in - 1]);
-
-      tmp0 = vextq_f32(in0, pad0, 1);
-      tmp1 = vextq_f32(in0, pad0, 2);
-      tmp2 = vextq_f32(in2, pad1, 2);
-      tmp3 = vextq_f32(in2, pad1, 2);
-
-      sum = vaddq_f32(in0, tmp0);
-      sum = vaddq_f32(sum, tmp1);
-      sum = vaddq_f32(sum, in2);
-      sum = vaddq_f32(sum, tmp2);
-      sum = vaddq_f32(sum, tmp3);
-      out0 = vmulq_f32(sum, v_coef);
+          out0 = in0;
+          out0 = vaddq_f32(out0, tmp0);
+          out0 = vaddq_f32(out0, tmp1);
+          out0 = vaddq_f32(out0, in2);
+          out0 = vaddq_f32(out0, tmp2);
+          out0 = vaddq_f32(out0, tmp3);
+          out0 = vaddq_f32(out0, in4);
+          out0 = vaddq_f32(out0, tmp4);
+          out0 = vaddq_f32(out0, tmp5);
+
+          vst1q_f32(output_ptr, vmulq_f32(out0, v_coef));
+        }
+        int m;
+        for (m = 1; (m + 3) < output_width - 1; m = m + 4) {
+        }
+
+        for (int j = m; j < output_width - 1; j++) {
+          output_data[i * output_width + j] =
+              input_data[(i - 1) * input_width + j - 1] +
+              input_data[(i - 1) * input_width + j] +
+              input_data[(i - 1) * input_width + j + 1] +
+              input_data[(i)*input_width + j - 1] +
+              input_data[(i)*input_width + j] +
+              input_data[(i)*input_width + j + 1] +
+              input_data[(i + 1) * input_width + j - 1] +
+              input_data[(i + 1) * input_width + j] +
+              input_data[(i + 1) * input_width + j + 1];
+          output_data[i * output_width + j] =
+              output_data[i * output_width + j] * coef;
+        }
+      }
+
+      output_data[0] =
+          input_data[0] + input_data[1] + input_data[l] + input_data[l + 1];
+      output_data[l - 1] = input_data[l - 2] + input_data[l - 1] +
+                           input_data[2 * l - 2] + input_data[2 * l - 1];
+      output_data[(l - 1) * l] =
+          input_data[(l - 2) * l] + input_data[(l - 2) * l + 1] +
+          input_data[(l - 1) * l] + input_data[(l - 1) * l + 1];
+      output_data[l * l - 1] = input_data[(l - 2) * (l + 1)] +
+                               input_data[(l - 2) * (l + 1) + 1] +
+                               input_data[l * l - 2] + input_data[l * l - 1];
+      output_data[0] = output_data[0] * coef2;
+      output_data[l - 1] = output_data[l - 1] * coef2;
+      output_data[(l - 1) * l] = output_data[(l - 1) * l] * coef2;
+      output_data[l * l - 1] = output_data[l * l - 1] * coef2;
+
+      for (int i = 1; i < l - 1; ++i) {
+        output_data[i * l] = input_data[i * l - l] + input_data[i * l - l + 1] +
+                             input_data[i * l] + input_data[i * l + 1] +
+                             input_data[i * l + l] + input_data[i * l + l + 1];
+
+        output_data[i * l + l - 1] =
+            input_data[i * l + l - 1 - l - 1] + input_data[i * l + l - 1 - l] +
+            input_data[i * l + l - 1 - 1] + input_data[i * l + l - 1] +
+            input_data[i * l + l - 1 + l - 1] + input_data[i * l + l - 1 + l];
+        output_data[i * l] = output_data[i * l] * coef1;
+        output_data[i * l + l - 1] = output_data[i * l + l - 1] * coef1;
+      }
+
+      int m;
+      for (m = 1; m < output_width - 4; m += 4) {
+        float *output_ptr = output_data + m;
+        float32x4_t in0, in1, in2, in3, tmp0, tmp1, tmp2, tmp3, out0;
+        in0 = vld1q_f32(input_data + m - 1);
+        in1 = vld1q_f32(input_data + m + 3);
+        in2 = vld1q_f32(input_data + input_width + m - 1);
+        in3 = vld1q_f32(input_data + input_width + m + 3);
+        tmp0 = vextq_f32(in0, in1, 1);
+        tmp1 = vextq_f32(in0, in1, 2);
+        tmp2 = vextq_f32(in2, in3, 1);
+        tmp3 = vextq_f32(in2, in3, 2);
+        out0 = in0;
+        out0 = vaddq_f32(out0, tmp0);
+        out0 = vaddq_f32(out0, tmp1);
+        out0 = vaddq_f32(out0, in2);
+        out0 = vaddq_f32(out0, tmp2);
+        out0 = vaddq_f32(out0, tmp3);
 
-      for (int i = 0; i < c_mid; ++i) {
-        if (i == 0) {
-          vst1q_lane_f32(output_ptr + i, out0, 0);
-        }
-        if (i == 1) {
-          vst1q_lane_f32(output_ptr + i, out0, 1);
-        }
-        if (i == 2) {
-          vst1q_lane_f32(output_ptr + i, out0, 2);
-        }
+        vst1q_f32(output_ptr, vmulq_f32(out0, v_coef1));
       }
 
-      // bottom_right remain
-      float32x4_t pad2 = vdupq_n_f32(input_seg[(h_in - 1) * w_in - 1]);
-      float32x4_t pad3 = vdupq_n_f32(input_seg[h_in * w_in - 1]);
-
-      tmp0 = vextq_f32(in4, pad2, 1);
-      tmp1 = vextq_f32(in4, pad2, 2);
-      tmp2 = vextq_f32(in6, pad3, 2);
-      tmp3 = vextq_f32(in6, pad3, 2);
-
-      sum = vaddq_f32(in4, tmp0);
-      sum = vaddq_f32(sum, tmp1);
-      sum = vaddq_f32(sum, in6);
-      sum = vaddq_f32(sum, tmp2);
-      sum = vaddq_f32(sum, tmp3);
-      out0 = vmulq_f32(sum, v_coef);
-
-      for (int i = 0; i < c_mid; ++i) {
-        if (i == 0) {
-          vst1q_lane_f32(output_ptr + (h_out - 1) * w_out + i, out0, 0);
+      for (m = 1; (m + 3) < output_width - 1; m += 4) {
       }
-        if (i == 1) {
-          vst1q_lane_f32(output_ptr + (h_out - 1) * w_out + i, out0, 1);
-        }
-        if (i == 2) {
-          vst1q_lane_f32(output_ptr + (h_out - 1) * w_out + i, out0, 2);
+      for (int j = m; j < output_width - 1; j++) {
+        output_data[j] = input_data[j - 1] + input_data[j] + input_data[j + 1] +
+                         input_data[input_width + j - 1] +
+                         input_data[input_width + j] +
+                         input_data[input_width + j + 1];
+        output_data[j] = output_data[j] * coef1;
       }
-      }
-      // mid
-      for (int j = 0; j < h_out - 2; ++j) {
-        output_ptr = output_seg + w_out * (j + 1) + 1;
-        input_tmp = input_seg + j * w_in;
 
-        in0 = vld1q_f32(input_tmp);
-        in2 = vld1q_f32(input_tmp + w_in);
-        in4 = vld1q_f32(input_tmp + 2 * w_in);
-        c_mid = w_out - 2;
-        for (; c_mid > 3; c_mid -= 4) {
-          in1 = vld1q_f32(input_tmp + 4);
-          in3 = vld1q_f32(input_tmp + w_in + 4);
-          in5 = vld1q_f32(input_tmp + 2 * w_in + 4);
+      for (m = 1; m < output_width - 4; m += 4) {
+        float *output_ptr =
+            output_data + (output_height - 1) * output_width + m;
 
+        float32x4_t in0, in1, in2, in3, tmp0, tmp1, tmp2, tmp3, out0;
+        in0 = vld1q_f32(input_data + (output_height - 2) * input_width + m - 1);
+        in1 = vld1q_f32(input_data + (output_height - 2) * input_width + m + 3);
+        in2 = vld1q_f32(input_data + (output_height - 1) * input_width + m - 1);
+        in3 = vld1q_f32(input_data + (output_height - 1) * input_width + m + 3);
         tmp0 = vextq_f32(in0, in1, 1);
         tmp1 = vextq_f32(in0, in1, 2);
         tmp2 = vextq_f32(in2, in3, 1);
         tmp3 = vextq_f32(in2, in3, 2);
-          tmp4 = vextq_f32(in4, in5, 1);
-          tmp5 = vextq_f32(in4, in5, 2);
-
-          sum = vaddq_f32(in0, tmp0);
-          sum = vaddq_f32(sum, tmp1);
-          sum = vaddq_f32(sum, in2);
-          sum = vaddq_f32(sum, tmp2);
-          sum = vaddq_f32(sum, tmp3);
-          sum = vaddq_f32(sum, in4);
-          sum = vaddq_f32(sum, tmp4);
-          sum = vaddq_f32(sum, tmp5);
-
-          out0 = vmulq_f32(sum, v_coef);
-          vst1q_f32(output_ptr, out0);
-          output_ptr += 4;
-          input_tmp += 4;
-          in0 = in1;
-          in2 = in3;
-          in4 = in5;
-        }
-        // mid remain
-        float32x4_t pad0 = vdupq_n_f32(input_seg[(j + 1) * w_in - 1]);
-        float32x4_t pad1 = vdupq_n_f32(input_seg[(j + 2) * w_in - 1]);
-        float32x4_t pad2 = vdupq_n_f32(input_seg[(j + 2) * w_in - 1]);
-
-        tmp0 = vextq_f32(in0, pad0, 1);
-        tmp1 = vextq_f32(in0, pad0, 2);
-        tmp2 = vextq_f32(in2, pad1, 1);
-        tmp3 = vextq_f32(in2, pad1, 2);
-        tmp4 = vextq_f32(in4, pad2, 1);
-        tmp5 = vextq_f32(in4, pad2, 2);
-
-        sum = vaddq_f32(in0, tmp0);
-        sum = vaddq_f32(sum, tmp1);
-        sum = vaddq_f32(sum, in2);
-        sum = vaddq_f32(sum, tmp2);
-        sum = vaddq_f32(sum, tmp3);
-        sum = vaddq_f32(sum, in4);
-        sum = vaddq_f32(sum, tmp4);
-        sum = vaddq_f32(sum, tmp5);
-        out0 = vmulq_f32(sum, v_coef);
-
-        for (int i = 0; i < c_mid; ++i) {
-          if (i == 0) {
-            vst1q_lane_f32(output_ptr + i, out0, 0);
-          }
-          if (i == 1) {
-            vst1q_lane_f32(output_ptr + i, out0, 1);
-          }
-          if (i == 2) {
-            vst1q_lane_f32(output_ptr + i, out0, 2);
-          }
-        }
-      }
-      //      input_data += inputdata_channel_stride;
-      //      out_data += outputdata_channel_stride;
-    }
-    input_data += input_batch_stride;
-    out_data += output_batch_stride;
-  }
+        out0 = in0;
+        out0 = vaddq_f32(out0, tmp0);
+        out0 = vaddq_f32(out0, tmp1);
+        out0 = vaddq_f32(out0, in2);
+        out0 = vaddq_f32(out0, tmp2);
+        out0 = vaddq_f32(out0, tmp3);
+
+        vst1q_f32(output_ptr, vmulq_f32(out0, v_coef1));
+      }
+      for (m = 1; (m + 3) < output_width - 1; m = m + 4) {
+      }
+      for (int j = m; j < output_width - 1; j++) {
+        output_data[(output_height - 1) * input_width + j] =
+            input_data[(output_height - 2) * input_width + j - 1] +
+            input_data[(output_height - 2) * input_width + j] +
+            input_data[(output_height - 2) * input_width + j + 1] +
+            input_data[(output_height - 1) * input_width + j - 1] +
+            input_data[(output_height - 1) * input_width + j] +
+            input_data[(output_height - 1) * input_width + j + 1];
+        output_data[(output_height - 1) * output_width + j] =
+            output_data[(output_height - 1) * output_width + j] * coef1;
+      }
+    }
+  }
+
+//  const int batch_size = input->dims()[0];
+//
+//  const int h_in = input->dims()[2];
+//
+//  const int w_in = input->dims()[3];
+//
+//  const int output_channels = output->dims()[1];
+//
+//  const int h_out = output->dims()[2];
+//  const int w_out = output->dims()[3];
+//  const int outputdata_channel_stride = h_out * w_out;
+//  const int inputdata_channel_stride = h_in * w_in;
+//  const int input_batch_stride = output_channels * inputdata_channel_stride;
+//  const int output_batch_stride = output_channels *
+//  outputdata_channel_stride; float *out_data = output->data<float>(); const
+//  float *input_data = input->data<float>();
+//
+//  const float coef = 1.0 / 9.0;
+//  for (int k = 0; k < batch_size; ++k) {
+// #pragma omp parallel for
+//    for (int c = 0; c < output_channels; ++c) {
+//      const float *input_seg = input_data + c * inputdata_channel_stride;
+//      float *output_seg = out_data + c * outputdata_channel_stride;
+//      // four corner point
+//      output_seg[0] = (input_seg[0] + input_seg[1] + input_seg[w_in] +
+//                       input_seg[w_in + 1]) *
+//                      coef;
+//      output_seg[w_out - 1] =
+//          (input_seg[w_in - 2] + input_seg[w_in - 1] + input_seg[w_in * 2 -
+//          2] +
+//           input_seg[2 * w_in - 1]) *
+//          coef;
+//      output_seg[(h_out - 1) * w_out] =
+//          (input_seg[(h_in - 2) * w_in] + input_seg[(h_in - 2) * w_in + 1] +
+//           input_seg[(h_in - 1) * w_in] + input_seg[(h_in - 1) * w_in + 1])
+//           *
+//          coef;
+//      output_seg[h_out * w_out - 1] =
+//          (input_seg[h_in * w_in - 1] + input_seg[h_in * w_in - 2] +
+//           input_seg[(h_in - 1) * w_in - 1] +
+//           input_seg[(h_in - 1) * w_in - 2]) *
+//          coef;
+//      // left side & right side
+//      for (int i = 1; i < h_in - 1; ++i) {
+//        output_seg[i * w_out] =
+//            (input_seg[i * w_in - w_in] + input_seg[i * w_in - w_in + 1] +
+//             input_seg[i * w_in] + input_seg[i * w_in + 1] +
+//             input_seg[i * w_in + w_in] + input_seg[i * w_in + w_in + 1]) *
+//            coef;
+//        output_seg[i * w_out + w_out - 1] =
+//            (input_seg[i * w_in - w_in + w_in - 2] +
+//             input_seg[i * w_in - w_in + 1 + w_in - 2] +
+//             input_seg[i * w_in + w_in - 2] +
+//             input_seg[i * w_in + 1 + w_in - 2] +
+//             input_seg[i * w_in + w_in + w_in - 2] +
+//             input_seg[i * w_in + w_in + 1 + w_in - 2]) *
+//            coef;
+//      }
+//      // top 1 row & bottom 1 row
+//      const float *input_tmp = input_seg;
+//
+//      float32x4_t in0, in1, in2, in3, in4, in5, in6, in7, tmp0, tmp1, tmp2,
+//          tmp3, tmp4, tmp5, sum, out0;
+//      float32x4_t v_coef = vdupq_n_f32(coef);
+//      in0 = vld1q_f32(input_tmp);
+//      in2 = vld1q_f32(input_tmp + w_in);
+//      const float *input_tmp_end = input_tmp + (h_in - 2) * w_in;
+//      in4 = vld1q_f32(input_tmp_end);
+//      in6 = vld1q_f32(input_tmp_end + w_in);
+//      int c_mid = w_out - 2;
+//      auto output_ptr = output_seg + 1;
+//      for (; c_mid > 3; c_mid -= 4) {
+//        in1 = vld1q_f32(input_tmp + 4);
+//        in3 = vld1q_f32(input_tmp + w_in + 4);
+//
+//        tmp0 = vextq_f32(in0, in1, 1);
+//        tmp1 = vextq_f32(in0, in1, 2);
+//
+//        tmp2 = vextq_f32(in2, in3, 1);
+//        tmp3 = vextq_f32(in2, in3, 2);
+//
+//        sum = vaddq_f32(in0, tmp0);
+//        sum = vaddq_f32(sum, tmp1);
+//        sum = vaddq_f32(sum, in2);
+//        sum = vaddq_f32(sum, tmp2);
+//        sum = vaddq_f32(sum, tmp3);
+//
+//        vst1q_f32(output_ptr, vmulq_f32(sum, v_coef));
+//
+//        in5 = vld1q_f32(input_tmp_end + 4);
+//        in7 = vld1q_f32(input_tmp_end + w_in + 4);
+//
+//        tmp0 = vextq_f32(in4, in5, 1);
+//        tmp1 = vextq_f32(in4, in5, 2);
+//        tmp2 = vextq_f32(in6, in7, 1);
+//        tmp3 = vextq_f32(in6, in7, 2);
+//
+//        sum = vaddq_f32(in0, tmp0);
+//        sum = vaddq_f32(sum, tmp1);
+//        sum = vaddq_f32(sum, in2);
+//        sum = vaddq_f32(sum, tmp2);
+//        sum = vaddq_f32(sum, tmp3);
+//
+//        vst1q_f32(output_ptr + (h_out - 1) * w_out, vmulq_f32(sum, v_coef));
+//
+//        // can optimize to each 8 stride.
+//        input_tmp += 4;
+//        input_tmp_end += 4;
+//        output_ptr += 4;
+//        in0 = in1;
+//        in2 = in3;
+//        in4 = in5;
+//        in6 = in7;
+//      }
+//      // top right remain
+//      float32x4_t pad0 = vdupq_n_f32(input_seg[w_in - 1]);
+//      float32x4_t pad1 = vdupq_n_f32(input_seg[2 * w_in - 1]);
+//
+//      tmp0 = vextq_f32(in0, pad0, 1);
+//      tmp1 = vextq_f32(in0, pad0, 2);
+//      tmp2 = vextq_f32(in2, pad1, 2);
+//      tmp3 = vextq_f32(in2, pad1, 2);
+//
+//      sum = vaddq_f32(in0, tmp0);
+//      sum = vaddq_f32(sum, tmp1);
+//      sum = vaddq_f32(sum, in2);
+//      sum = vaddq_f32(sum, tmp2);
+//      sum = vaddq_f32(sum, tmp3);
+//      out0 = vmulq_f32(sum, v_coef);
+//
+//      for (int i = 0; i < c_mid; ++i) {
+//        if (i == 0) {
+//          vst1q_lane_f32(output_ptr + i, out0, 0);
+//        }
+//        if (i == 1) {
+//          vst1q_lane_f32(output_ptr + i, out0, 1);
+//        }
+//        if (i == 2) {
+//          vst1q_lane_f32(output_ptr + i, out0, 2);
+//        }
+//      }
+//
+//      // bottom_right remain
+//      float32x4_t pad2 = vdupq_n_f32(input_seg[(h_in - 1) * w_in - 1]);
+//      float32x4_t pad3 = vdupq_n_f32(input_seg[h_in * w_in - 1]);
+//
+//      tmp0 = vextq_f32(in4, pad2, 1);
+//      tmp1 = vextq_f32(in4, pad2, 2);
+//      tmp2 = vextq_f32(in6, pad3, 2);
+//      tmp3 = vextq_f32(in6, pad3, 2);
+//
+//      sum = vaddq_f32(in4, tmp0);
+//      sum = vaddq_f32(sum, tmp1);
+//      sum = vaddq_f32(sum, in6);
+//      sum = vaddq_f32(sum, tmp2);
+//      sum = vaddq_f32(sum, tmp3);
+//      out0 = vmulq_f32(sum, v_coef);
+//
+//      for (int i = 0; i < c_mid; ++i) {
+//        if (i == 0) {
+//          vst1q_lane_f32(output_ptr + (h_out - 1) * w_out + i, out0, 0);
+//        }
+//        if (i == 1) {
+//          vst1q_lane_f32(output_ptr + (h_out - 1) * w_out + i, out0, 1);
+//        }
+//        if (i == 2) {
+//          vst1q_lane_f32(output_ptr + (h_out - 1) * w_out + i, out0, 2);
+//        }
+//      }
+//      // mid
+//      for (int j = 0; j < h_out - 2; ++j) {
+//        output_ptr = output_seg + w_out * (j + 1) + 1;
+//        input_tmp = input_seg + j * w_in;
+//
+//        in0 = vld1q_f32(input_tmp);
+//        in2 = vld1q_f32(input_tmp + w_in);
+//        in4 = vld1q_f32(input_tmp + 2 * w_in);
+//        c_mid = w_out - 2;
+//        for (; c_mid > 3; c_mid -= 4) {
+//          in1 = vld1q_f32(input_tmp + 4);
+//          in3 = vld1q_f32(input_tmp + w_in + 4);
+//          in5 = vld1q_f32(input_tmp + 2 * w_in + 4);
+//
+//          tmp0 = vextq_f32(in0, in1, 1);
+//          tmp1 = vextq_f32(in0, in1, 2);
+//          tmp2 = vextq_f32(in2, in3, 1);
+//          tmp3 = vextq_f32(in2, in3, 2);
+//          tmp4 = vextq_f32(in4, in5, 1);
+//          tmp5 = vextq_f32(in4, in5, 2);
+//
+//          sum = vaddq_f32(in0, tmp0);
+//          sum = vaddq_f32(sum, tmp1);
+//          sum = vaddq_f32(sum, in2);
+//          sum = vaddq_f32(sum, tmp2);
+//          sum = vaddq_f32(sum, tmp3);
+//          sum = vaddq_f32(sum, in4);
+//          sum = vaddq_f32(sum, tmp4);
+//          sum = vaddq_f32(sum, tmp5);
+//
+//          out0 = vmulq_f32(sum, v_coef);
+//          vst1q_f32(output_ptr, out0);
+//          output_ptr += 4;
+//          input_tmp += 4;
+//          in0 = in1;
+//          in2 = in3;
+//          in4 = in5;
+//        }
+//        // mid remain
+//        float32x4_t pad0 = vdupq_n_f32(input_seg[(j + 1) * w_in - 1]);
+//        float32x4_t pad1 = vdupq_n_f32(input_seg[(j + 2) * w_in - 1]);
+//        float32x4_t pad2 = vdupq_n_f32(input_seg[(j + 2) * w_in - 1]);
+//
+//        tmp0 = vextq_f32(in0, pad0, 1);
+//        tmp1 = vextq_f32(in0, pad0, 2);
+//        tmp2 = vextq_f32(in2, pad1, 1);
+//        tmp3 = vextq_f32(in2, pad1, 2);
+//        tmp4 = vextq_f32(in4, pad2, 1);
+//        tmp5 = vextq_f32(in4, pad2, 2);
+//
+//        sum = vaddq_f32(in0, tmp0);
+//        sum = vaddq_f32(sum, tmp1);
+//        sum = vaddq_f32(sum, in2);
+//        sum = vaddq_f32(sum, tmp2);
+//        sum = vaddq_f32(sum, tmp3);
+//        sum = vaddq_f32(sum, in4);
+//        sum = vaddq_f32(sum, tmp4);
+//        sum = vaddq_f32(sum, tmp5);
+//        out0 = vmulq_f32(sum, v_coef);
+//
+//        for (int i = 0; i < c_mid; ++i) {
+//          if (i == 0) {
+//            vst1q_lane_f32(output_ptr + i, out0, 0);
+//          }
+//          if (i == 1) {
+//            vst1q_lane_f32(output_ptr + i, out0, 1);
+//          }
+//          if (i == 2) {
+//            vst1q_lane_f32(output_ptr + i, out0, 2);
+//          }
+//        }
+//      }
+//      //      input_data += inputdata_channel_stride;
+//      //      out_data += outputdata_channel_stride;
+//    }
+//    input_data += input_batch_stride;
+//    out_data += output_batch_stride;
+//  }
 #endif
 }
 
@@ -662,6 +839,7 @@ void Pool3x3Avg(vector<int> strides, vector<int> paddings, const Tensor *input,
           wstart = max(wstart, 0);
           hend = min(hend, input_height);
           wend = min(wend, input_width);
+
           const float *pos1 = input_seg + hstart * input_width + wstart;
           const float *pos2 = input_seg + (hstart + 1) * input_width + wstart;
           const float *pos3 = input_seg + (hstart + 2) * input_width + wstart;
@@ -674,7 +852,8 @@ void Pool3x3Avg(vector<int> strides, vector<int> paddings, const Tensor *input,
                 sum += input_seg[h * input_width + w];
               }
             }
-            output_seg[ph * output_width + pw] = sum / 9.0;
+            output_seg[ph * output_width + pw] =
+                sum / ((hend - hstart) * (wend - wstart) * 1.0);
           } else {
 #if __aarch64__
 #else

src/operators/op_param.h

@@ -317,22 +317,23 @@ class OpParam {
   }
 };
 
-#ifdef CONV_OP
 template <typename Dtype>
-class ConvParam : OpParam {
+class ConvParam : public OpParam {
   typedef typename DtypeTensorTrait<Dtype>::gtype GType;
   typedef typename DtypeTensorTrait<Dtype>::rtype RType;
 
  public:
   ConvParam(const VariableNameMap &inputs, const VariableNameMap &outputs,
             const AttributeMap &attrs, const Scope &scope) {
-    filter_ = FilterFrom<GType>(inputs, scope);
-    input_ = InputFrom<GType>(inputs, scope);
-    output_ = OutputFrom<GType>(outputs, scope);
-    strides_ = GetAttr<vector<int>>("strides", attrs);
-    paddings_ = GetAttr<vector<int>>("paddings", attrs);
-    dilations_ = GetAttr<vector<int>>("dilations", attrs);
-    groups = GetAttr<int>("groups", attrs);
+    filter_ = OpParam::FilterFrom<GType>(inputs, scope);
+    input_ = OpParam::InputFrom<GType>(inputs, scope);
+    if (outputs.count("Output")) {
+      output_ = OpParam::OutputFrom<GType>(outputs, scope);
+    }
+    strides_ = OpParam::GetAttr<vector<int>>("strides", attrs);
+    paddings_ = OpParam::GetAttr<vector<int>>("paddings", attrs);
+    dilations_ = OpParam::GetAttr<vector<int>>("dilations", attrs);
+    groups = OpParam::GetAttr<int>("groups", attrs);
   }
 
   const RType *Input() const { return input_; }
@@ -360,7 +361,6 @@ class ConvParam : OpParam {
 };
 template <typename Dtype>
 Print &operator<<(Print &printer, const ConvParam<Dtype> &conv_param);
-#endif
 
 template <typename Dtype>
 class ElementwiseAddParam : OpParam {
@@ -652,6 +652,11 @@ class PriorBoxParam : public OpParam {
     max_sizes_ = GetAttr<vector<float>>("max_sizes", attrs);
     aspect_ratios_ = GetAttr<vector<float>>("aspect_ratios", attrs);
     variances_ = GetAttr<vector<float>>("variances", attrs);
+
+    if (HasAttr("min_max_aspect_ratios_order", attrs)) {
+      min_max_aspect_ratios_order_ =
+          GetAttr<bool>("min_max_aspect_ratios_order", attrs);
+    }
     flip_ = GetAttr<bool>("flip", attrs);
     clip_ = GetAttr<bool>("clip", attrs);
     step_w_ = GetAttr<float>("step_w", attrs);
@@ -684,6 +689,10 @@ class PriorBoxParam : public OpParam {
 
   const float &Offset() const { return offset_; }
 
+  const bool &MinMaxAspectRatiosOrder() const {
+    return min_max_aspect_ratios_order_;
+  }
+
  private:
   RType *input_;
   RType *input_image_;
@@ -698,6 +707,7 @@ class PriorBoxParam : public OpParam {
   float step_w_;
   float step_h_;
   float offset_;
+  bool min_max_aspect_ratios_order_;
 };
 #endif
 
@@ -761,7 +771,7 @@ class SoftmaxParam : public OpParam {
   fpga::BypassArgs fpga_bypass_args;
 
  public:
-  RType *FloatInput() {
+  RType *FloatInput() const {
     return float_input_x_ == nullptr ? input_x_ : float_input_x_.get();
   }
   void SetFloatInput(Tensor *input) { float_input_x_.reset(input); }
@@ -1260,52 +1270,29 @@ using FusionFcReluParam = FusionFcParam<DeviceType>;
 #endif
 
 template <typename Dtype>
-class FusionConvAddParam : public OpParam {
+class FusionConvAddParam : public ConvParam<Dtype> {
   typedef typename DtypeTensorTrait<Dtype>::gtype GType;
   typedef typename DtypeTensorTrait<Dtype>::rtype RType;
 
  public:
   FusionConvAddParam(const VariableNameMap &inputs,
                      const VariableNameMap &outputs, const AttributeMap &attrs,
-                     const Scope &scope) {
-    bias_ = InputYFrom<GType>(inputs, scope);
-    axis_ = GetAttr<int>("axis", attrs);
-    filter_ = FilterFrom<GType>(inputs, scope);
-    input_ = InputFrom<GType>(inputs, scope);
-    output_ = OutFrom<GType>(outputs, scope);
-    strides_ = GetAttr<vector<int>>("strides", attrs);
-    paddings_ = GetAttr<vector<int>>("paddings", attrs);
-    dilations_ = GetAttr<vector<int>>("dilations", attrs);
-    groups = GetAttr<int>("groups", attrs);
+                     const Scope &scope)
+      : ConvParam<Dtype>(inputs, outputs, attrs, scope) {
+    bias_ = OpParam::InputYFrom<GType>(inputs, scope);
+    axis_ = OpParam::GetAttr<int>("axis", attrs);
+    output_ = OpParam::OutFrom<GType>(outputs, scope);
   }
   RType *Bias() const { return bias_; }
 
   const int &Axis() const { return axis_; }
 
-  const RType *Input() const { return input_; }
-
-  const RType *Filter() const { return filter_; }
-
   RType *Output() const { return output_; }
 
-  const vector<int> &Strides() const { return strides_; }
-
-  const vector<int> &Paddings() const { return paddings_; }
-
-  const vector<int> &Dilations() const { return dilations_; }
-
-  const int &Groups() const { return groups; }
-
  protected:
   RType *bias_;
   int axis_;
-  RType *input_;
   RType *output_;
-  RType *filter_;
-  vector<int> strides_;
-  vector<int> paddings_;
-  vector<int> dilations_;
-  int groups;
 #ifdef PADDLE_MOBILE_FPGA
 
  private:
@@ -1332,58 +1319,33 @@ class FusionConvAddReluParam : public FusionConvAddParam<DeviceType> {
 #endif
 
 #ifdef FUSION_CONVADDPRELU_OP
-template <typename DeviceType>
-class FusionConvAddPReluParam : public OpParam {
-  typedef typename DtypeTensorTrait<DeviceType>::gtype GType;
-  typedef typename DtypeTensorTrait<DeviceType>::rtype RType;
+template <typename Dtype>
+class FusionConvAddPReluParam : public ConvParam<Dtype> {
+  typedef typename DtypeTensorTrait<Dtype>::gtype GType;
+  typedef typename DtypeTensorTrait<Dtype>::rtype RType;
 
  public:
   FusionConvAddPReluParam(const VariableNameMap &inputs,
                           const VariableNameMap &outputs,
-                          const AttributeMap &attrs, const Scope &scope) {
-    alpha_ = InputAlphaFrom<GType>(inputs, scope);
-    mode_ = GetAttr<std::string>("mode", attrs);
+                          const AttributeMap &attrs, const Scope &scope)
+      : ConvParam<Dtype>(inputs, outputs, attrs, scope) {
+    alpha_ = OpParam::InputAlphaFrom<GType>(inputs, scope);
+    mode_ = OpParam::GetAttr<std::string>("mode", attrs);
     framework::DDim dims = alpha_->dims();
-    bias_ = InputYFrom<GType>(inputs, scope);
-    axis_ = GetAttr<int>("axis", attrs);
-    filter_ = FilterFrom<GType>(inputs, scope);
-    input_ = InputFrom<GType>(inputs, scope);
-    output_ = OutFrom<GType>(outputs, scope);
-    strides_ = GetAttr<vector<int>>("strides", attrs);
-    paddings_ = GetAttr<vector<int>>("paddings", attrs);
-    dilations_ = GetAttr<vector<int>>("dilations", attrs);
-    groups = GetAttr<int>("groups", attrs);
+    bias_ = OpParam::InputYFrom<GType>(inputs, scope);
+    axis_ = OpParam::GetAttr<int>("axis", attrs);
+    output_ = OpParam::OutFrom<GType>(outputs, scope);
   }
   const RType *InputAlpha() const { return alpha_; }
   const std::string &Mode() const { return mode_; }
   RType *Bias() const { return bias_; }
-
   const int &Axis() const { return axis_; }
-
-  const RType *Input() const { return input_; }
-
-  const RType *Filter() const { return filter_; }
-
   RType *Output() const { return output_; }
 
-  const vector<int> &Strides() const { return strides_; }
-
-  const vector<int> &Paddings() const { return paddings_; }
-
-  const vector<int> &Dilations() const { return dilations_; }
-
-  const int &Groups() const { return groups; }
-
  protected:
   RType *bias_;
   int axis_;
-  RType *input_;
   RType *output_;
-  RType *filter_;
-  vector<int> strides_;
-  vector<int> paddings_;
-  vector<int> dilations_;
-  int groups;
   RType *alpha_;
   std::string mode_;
 #ifdef PADDLE_MOBILE_FPGA
@@ -1399,35 +1361,30 @@ class FusionConvAddPReluParam : public OpParam {
 #endif
 
 #ifdef FUSION_CONVADDADDPRELU_OP
-template <typename DeviceType>
-class FusionConvAddAddPReluParam : public OpParam {
-  typedef typename DtypeTensorTrait<DeviceType>::gtype GType;
-  typedef typename DtypeTensorTrait<DeviceType>::rtype RType;
+template <typename Dtype>
+class FusionConvAddAddPReluParam : public ConvParam<Dtype> {
+  typedef typename DtypeTensorTrait<Dtype>::gtype GType;
+  typedef typename DtypeTensorTrait<Dtype>::rtype RType;
 
  public:
   FusionConvAddAddPReluParam(const VariableNameMap &inputs,
                              const VariableNameMap &outputs,
-                             const AttributeMap &attrs, const Scope &scope) {
-    bias1_ = InputYFrom1<GType>(inputs, scope);
-    alpha_ = InputAlphaFrom<GType>(inputs, scope);
-    mode_ = GetAttr<std::string>("mode", attrs);
+                             const AttributeMap &attrs, const Scope &scope)
+      : ConvParam<Dtype>(inputs, outputs, attrs, scope) {
+    bias1_ = OpParam::InputYFrom1<GType>(inputs, scope);
+    alpha_ = OpParam::InputAlphaFrom<GType>(inputs, scope);
+    mode_ = OpParam::GetAttr<std::string>("mode", attrs);
     framework::DDim dims = alpha_->dims();
-    bias_ = InputYFrom<GType>(inputs, scope);
-    axis_ = GetAttr<int>("axis", attrs);
-    filter_ = FilterFrom<GType>(inputs, scope);
-    input_ = InputFrom<GType>(inputs, scope);
-    output_ = OutFrom<GType>(outputs, scope);
-    strides_ = GetAttr<vector<int>>("strides", attrs);
-    paddings_ = GetAttr<vector<int>>("paddings", attrs);
-    dilations_ = GetAttr<vector<int>>("dilations", attrs);
-    groups = GetAttr<int>("groups", attrs);
-    keyOutput_ = getkey("addOut", inputs, 0);
-    keyX1_ = getkey("addX", inputs, 1);
-    keyY1_ = getkey("Y", inputs, 1);
+    bias_ = OpParam::InputYFrom<GType>(inputs, scope);
+    output_ = OpParam::OutFrom<GType>(outputs, scope);
+    axis_ = OpParam::GetAttr<int>("axis", attrs);
+    keyOutput_ = OpParam::getkey("addOut", inputs, 0);
+    keyX1_ = OpParam::getkey("addX", inputs, 1);
+    keyY1_ = OpParam::getkey("Y", inputs, 1);
     if (keyX1_ == keyOutput_) {
-      bias1_ = InputYFrom1<GType>(inputs, scope);
+      bias1_ = OpParam::InputYFrom1<GType>(inputs, scope);
     } else if (keyY1_ == keyOutput_) {
-      bias1_ = InputXFrom1<GType>(inputs, scope);
+      bias1_ = OpParam::InputXFrom1<GType>(inputs, scope);
     }
   }
   const RType *InputAlpha() const { return alpha_; }
@@ -1437,31 +1394,12 @@ class FusionConvAddAddPReluParam : public OpParam {
   RType *Bias() const { return bias_; }
 
   const int &Axis() const { return axis_; }
-
-  const RType *Input() const { return input_; }
-
-  const RType *Filter() const { return filter_; }
-
   RType *Output() const { return output_; }
 
-  const vector<int> &Strides() const { return strides_; }
-
-  const vector<int> &Paddings() const { return paddings_; }
-
-  const vector<int> &Dilations() const { return dilations_; }
-
-  const int &Groups() const { return groups; }
-
  protected:
   RType *bias_;
   int axis_;
-  RType *input_;
   RType *output_;
-  RType *filter_;
-  vector<int> strides_;
-  vector<int> paddings_;
-  vector<int> dilations_;
-  int groups;
   RType *alpha_;
   std::string mode_;
   RType *bias1_;
@@ -1482,49 +1420,32 @@ class FusionConvAddAddPReluParam : public OpParam {
 
 #ifdef FUSION_CONVADDBNRELU_OP
 template <typename Dtype>
-class FusionConvAddBNReluParam : public OpParam {
+class FusionConvAddBNReluParam : public ConvParam<Dtype> {
   typedef typename DtypeTensorTrait<Dtype>::gtype GType;
   typedef typename DtypeTensorTrait<Dtype>::rtype RType;
 
  public:
   FusionConvAddBNReluParam(const VariableNameMap &inputs,
                            const VariableNameMap &outputs,
-                           const AttributeMap &attrs, const Scope &scope) {
-    bias_ = InputYFrom<GType>(inputs, scope);
-    axis_ = GetAttr<int>("axis", attrs);
-    filter_ = FilterFrom<GType>(inputs, scope);
-    input_ = InputFrom<GType>(inputs, scope);
-    output_ = OutFrom<GType>(outputs, scope);
-    strides_ = GetAttr<vector<int>>("strides", attrs);
-    paddings_ = GetAttr<vector<int>>("paddings", attrs);
-    dilations_ = GetAttr<vector<int>>("dilations", attrs);
-    groups = GetAttr<int>("groups", attrs);
-    input_bias_ = InputBiasFrom<GType>(inputs, scope);
-    input_mean_ = InputMeanFrom<GType>(inputs, scope);
-    input_scale_ = InputScaleFrom<GType>(inputs, scope);
-    input_variance_ = InputVarianceFrom<GType>(inputs, scope);
-    epsilon_ = GetAttr<float>("epsilon", attrs);
-    momentum_ = GetAttr<float>("momentum", attrs);
-    //    is_test_ = GetAttr<bool>("is_test", attrs);
+                           const AttributeMap &attrs, const Scope &scope)
+      : ConvParam<Dtype>(inputs, outputs, attrs, scope) {
+    bias_ = OpParam::InputYFrom<GType>(inputs, scope);
+    axis_ = OpParam::GetAttr<int>("axis", attrs);
+    output_ = OpParam::OutFrom<GType>(outputs, scope);
+    input_bias_ = OpParam::InputBiasFrom<GType>(inputs, scope);
+    input_mean_ = OpParam::InputMeanFrom<GType>(inputs, scope);
+    input_scale_ = OpParam::InputScaleFrom<GType>(inputs, scope);
+    input_variance_ = OpParam::InputVarianceFrom<GType>(inputs, scope);
+    epsilon_ = OpParam::GetAttr<float>("epsilon", attrs);
+    momentum_ = OpParam::GetAttr<float>("momentum", attrs);
+    //    is_test_ = OpParam::GetAttr<bool>("is_test", attrs);
   }
   RType *Bias() const { return bias_; }
 
   const int &Axis() const { return axis_; }
 
-  const RType *Input() const { return input_; }
-
-  const RType *Filter() const { return filter_; }
-
   RType *Output() const { return output_; }
 
-  const vector<int> &Strides() const { return strides_; }
-
-  const vector<int> &Paddings() const { return paddings_; }
-
-  const vector<int> &Dilations() const { return dilations_; }
-
-  const int &Groups() const { return groups; }
-
   const RType *InputBias() const { return input_bias_; }
 
   const RType *InputMean() const { return input_mean_; }
@@ -1550,13 +1471,7 @@ class FusionConvAddBNReluParam : public OpParam {
  protected:
   RType *bias_;
   int axis_;
-  RType *input_;
   RType *output_;
-  RType *filter_;
-  vector<int> strides_;
-  vector<int> paddings_;
-  vector<int> dilations_;
-  int groups;
   RType *input_bias_;
   RType *input_mean_;
   RType *input_scale_;
@@ -1580,57 +1495,40 @@ class FusionConvAddBNReluParam : public OpParam {
 
 #ifdef FUSION_CONVBNADDRELU_OP
 template <typename Dtype>
-class FusionConvBNAddReluParam : public OpParam {
+class FusionConvBNAddReluParam : public ConvParam<Dtype> {
   typedef typename DtypeTensorTrait<Dtype>::gtype GType;
   typedef typename DtypeTensorTrait<Dtype>::rtype RType;
 
  public:
   FusionConvBNAddReluParam(const VariableNameMap &inputs,
                            const VariableNameMap &outputs,
-                           const AttributeMap &attrs, const Scope &scope) {
-    bias_ = InputYFrom<GType>(inputs, scope);
-    axis_ = GetAttr<int>("axis", attrs);
-    filter_ = FilterFrom<GType>(inputs, scope);
-    input_ = InputFrom<GType>(inputs, scope);
-    output_ = OutFrom<GType>(outputs, scope);
-    strides_ = GetAttr<vector<int>>("strides", attrs);
-    paddings_ = GetAttr<vector<int>>("paddings", attrs);
-    dilations_ = GetAttr<vector<int>>("dilations", attrs);
-    groups = GetAttr<int>("groups", attrs);
-    input_bias_ = InputBiasFrom<GType>(inputs, scope);
-    input_mean_ = InputMeanFrom<GType>(inputs, scope);
-    input_scale_ = InputScaleFrom<GType>(inputs, scope);
-    input_variance_ = InputVarianceFrom<GType>(inputs, scope);
-    epsilon_ = GetAttr<float>("epsilon", attrs);
-    momentum_ = GetAttr<float>("momentum", attrs);
-    keyBNY_ = getkey("BNY", inputs, 0);
-    keyX_ = getkey("X", inputs, 0);
-    keyY_ = getkey("Y", inputs, 0);
+                           const AttributeMap &attrs, const Scope &scope)
+      : ConvParam<Dtype>(inputs, outputs, attrs, scope) {
+    bias_ = OpParam::InputYFrom<GType>(inputs, scope);
+    axis_ = OpParam::GetAttr<int>("axis", attrs);
+    output_ = OpParam::OutFrom<GType>(outputs, scope);
+    input_bias_ = OpParam::InputBiasFrom<GType>(inputs, scope);
+    input_mean_ = OpParam::InputMeanFrom<GType>(inputs, scope);
+    input_scale_ = OpParam::InputScaleFrom<GType>(inputs, scope);
+    input_variance_ = OpParam::InputVarianceFrom<GType>(inputs, scope);
+    epsilon_ = OpParam::GetAttr<float>("epsilon", attrs);
+    momentum_ = OpParam::GetAttr<float>("momentum", attrs);
+    keyBNY_ = OpParam::getkey("BNY", inputs, 0);
+    keyX_ = OpParam::getkey("X", inputs, 0);
+    keyY_ = OpParam::getkey("Y", inputs, 0);
     if (keyX_ == keyBNY_) {
-      bias_ = InputYFrom<GType>(inputs, scope);
+      bias_ = OpParam::InputYFrom<GType>(inputs, scope);
     } else if (keyY_ == keyBNY_) {
-      bias_ = InputXFrom<GType>(inputs, scope);
+      bias_ = OpParam::InputXFrom<GType>(inputs, scope);
     }
-    //    is_test_ = GetAttr<bool>("is_test", attrs);
+    //    is_test_ = OpParam::GetAttr<bool>("is_test", attrs);
   }
   RType *Bias() const { return bias_; }
 
   const int &Axis() const { return axis_; }
 
-  const RType *Input() const { return input_; }
-
-  const RType *Filter() const { return filter_; }
-
   RType *Output() const { return output_; }
 
-  const vector<int> &Strides() const { return strides_; }
-
-  const vector<int> &Paddings() const { return paddings_; }
-
-  const vector<int> &Dilations() const { return dilations_; }
-
-  const int &Groups() const { return groups; }
-
   const RType *InputBias() const { return input_bias_; }
 
   const RType *InputMean() const { return input_mean_; }
@@ -1656,13 +1554,7 @@ class FusionConvBNAddReluParam : public OpParam {
  protected:
   RType *bias_;
   int axis_;
-  RType *input_;
   RType *output_;
-  RType *filter_;
-  vector<int> strides_;
-  vector<int> paddings_;
-  vector<int> dilations_;
-  int groups;
   RType *input_bias_;
   RType *input_mean_;
   RType *input_scale_;
@@ -1689,44 +1581,26 @@ class FusionConvBNAddReluParam : public OpParam {
 
 #ifdef FUSION_CONVBN_OP
 template <typename Dtype>
-class FusionConvBNParam : public OpParam {
+class FusionConvBNParam : public ConvParam<Dtype> {
   typedef typename DtypeTensorTrait<Dtype>::gtype GType;
   typedef typename DtypeTensorTrait<Dtype>::rtype RType;
 
  public:
   FusionConvBNParam(const VariableNameMap &inputs,
                     const VariableNameMap &outputs, const AttributeMap &attrs,
-                    const Scope &scope) {
-    filter_ = FilterFrom<GType>(inputs, scope);
-    input_ = InputFrom<GType>(inputs, scope);
-    output_y_ = OutputYFrom<GType>(outputs, scope);
-    strides_ = GetAttr<vector<int>>("strides", attrs);
-    paddings_ = GetAttr<vector<int>>("paddings", attrs);
-    dilations_ = GetAttr<vector<int>>("dilations", attrs);
-    groups = GetAttr<int>("groups", attrs);
-    input_bias_ = InputBiasFrom<GType>(inputs, scope);
-    input_mean_ = InputMeanFrom<GType>(inputs, scope);
-    input_scale_ = InputScaleFrom<GType>(inputs, scope);
-    input_variance_ = InputVarianceFrom<GType>(inputs, scope);
-    epsilon_ = GetAttr<float>("epsilon", attrs);
-    momentum_ = GetAttr<float>("momentum", attrs);
-    //    is_test_ = GetAttr<bool>("is_test", attrs);
+                    const Scope &scope)
+      : ConvParam<Dtype>(inputs, outputs, attrs, scope) {
+    output_y_ = OpParam::OutputYFrom<GType>(outputs, scope);
+    input_bias_ = OpParam::InputBiasFrom<GType>(inputs, scope);
+    input_mean_ = OpParam::InputMeanFrom<GType>(inputs, scope);
+    input_scale_ = OpParam::InputScaleFrom<GType>(inputs, scope);
+    input_variance_ = OpParam::InputVarianceFrom<GType>(inputs, scope);
+    epsilon_ = OpParam::GetAttr<float>("epsilon", attrs);
+    momentum_ = OpParam::GetAttr<float>("momentum", attrs);
+    //    is_test_ = OpParam::GetAttr<bool>("is_test", attrs);
   }
-
-  const RType *Input() const { return input_; }
-
-  const RType *Filter() const { return filter_; }
-
   RType *Output() const { return output_y_; }
 
-  const vector<int> &Strides() const { return strides_; }
-
-  const vector<int> &Paddings() const { return paddings_; }
-
-  const vector<int> &Dilations() const { return dilations_; }
-
-  const int &Groups() const { return groups; }
-
   const RType *InputBias() const { return input_bias_; }
 
   const RType *InputMean() const { return input_mean_; }
@@ -1750,13 +1624,7 @@ class FusionConvBNParam : public OpParam {
   const RType *NewBias() const { return new_bias_; }
 
  protected:
-  RType *input_;
   RType *output_y_;
-  RType *filter_;
-  vector<int> strides_;
-  vector<int> paddings_;
-  vector<int> dilations_;
-  int groups;
   RType *input_bias_;
   RType *input_mean_;
   RType *input_scale_;
@@ -1780,49 +1648,32 @@ class FusionConvBNParam : public OpParam {
 
 #ifdef FUSION_CONVADDBN_OP
 template <typename Dtype>
-class FusionConvAddBNParam : public OpParam {
+class FusionConvAddBNParam : public ConvParam<Dtype> {
   typedef typename DtypeTensorTrait<Dtype>::gtype GType;
   typedef typename DtypeTensorTrait<Dtype>::rtype RType;
 
  public:
   FusionConvAddBNParam(const VariableNameMap &inputs,
                        const VariableNameMap &outputs,
-                       const AttributeMap &attrs, const Scope &scope) {
-    bias_ = InputYFrom<GType>(inputs, scope);
-    axis_ = GetAttr<int>("axis", attrs);
-    filter_ = FilterFrom<GType>(inputs, scope);
-    input_ = InputFrom<GType>(inputs, scope);
-    output_y_ = OutputYFrom<GType>(outputs, scope);
-    strides_ = GetAttr<vector<int>>("strides", attrs);
-    paddings_ = GetAttr<vector<int>>("paddings", attrs);
-    dilations_ = GetAttr<vector<int>>("dilations", attrs);
-    groups = GetAttr<int>("groups", attrs);
-    input_bias_ = InputBiasFrom<GType>(inputs, scope);
-    input_mean_ = InputMeanFrom<GType>(inputs, scope);
-    input_scale_ = InputScaleFrom<GType>(inputs, scope);
-    input_variance_ = InputVarianceFrom<GType>(inputs, scope);
-    epsilon_ = GetAttr<float>("epsilon", attrs);
-    momentum_ = GetAttr<float>("momentum", attrs);
-    //    is_test_ = GetAttr<bool>("is_test", attrs);
+                       const AttributeMap &attrs, const Scope &scope)
+      : ConvParam<Dtype>(inputs, outputs, attrs, scope) {
+    bias_ = OpParam::InputYFrom<GType>(inputs, scope);
+    axis_ = OpParam::GetAttr<int>("axis", attrs);
+    output_y_ = OpParam::OutputYFrom<GType>(outputs, scope);
+    input_bias_ = OpParam::InputBiasFrom<GType>(inputs, scope);
+    input_mean_ = OpParam::InputMeanFrom<GType>(inputs, scope);
+    input_scale_ = OpParam::InputScaleFrom<GType>(inputs, scope);
+    input_variance_ = OpParam::InputVarianceFrom<GType>(inputs, scope);
+    epsilon_ = OpParam::GetAttr<float>("epsilon", attrs);
+    momentum_ = OpParam::GetAttr<float>("momentum", attrs);
+    //    is_test_ = OpParam::GetAttr<bool>("is_test", attrs);
   }
   RType *Bias() const { return bias_; }
 
   const int &Axis() const { return axis_; }
 
-  const RType *Input() const { return input_; }
-
-  const RType *Filter() const { return filter_; }
-
   RType *Output() const { return output_y_; }
 
-  const vector<int> &Strides() const { return strides_; }
-
-  const vector<int> &Paddings() const { return paddings_; }
-
-  const vector<int> &Dilations() const { return dilations_; }
-
-  const int &Groups() const { return groups; }
-
   const RType *InputBias() const { return input_bias_; }
 
   const RType *InputMean() const { return input_mean_; }
@@ -1848,13 +1699,7 @@ class FusionConvAddBNParam : public OpParam {
  protected:
   RType *bias_;
   int axis_;
-  RType *input_;
   RType *output_y_;
-  RType *filter_;
-  vector<int> strides_;
-  vector<int> paddings_;
-  vector<int> dilations_;
-  int groups;
   RType *input_bias_;
   RType *input_mean_;
   RType *input_scale_;
@@ -1878,44 +1723,26 @@ class FusionConvAddBNParam : public OpParam {
 
 #ifdef FUSION_DWCONVBNRELU_OP
 template <typename Dtype>
-class FusionDWConvBNReluParam : public OpParam {
+class FusionDWConvBNReluParam : public ConvParam<Dtype> {
   typedef typename DtypeTensorTrait<Dtype>::gtype GType;
   typedef typename DtypeTensorTrait<Dtype>::rtype RType;
 
  public:
   FusionDWConvBNReluParam(const VariableNameMap &inputs,
                           const VariableNameMap &outputs,
-                          const AttributeMap &attrs, const Scope &scope) {
-    filter_ = FilterFrom<GType>(inputs, scope);
-    input_ = InputFrom<GType>(inputs, scope);
-    output_ = OutFrom<GType>(outputs, scope);
-    strides_ = GetAttr<vector<int>>("strides", attrs);
-    paddings_ = GetAttr<vector<int>>("paddings", attrs);
-    dilations_ = GetAttr<vector<int>>("dilations", attrs);
-    groups = GetAttr<int>("groups", attrs);
-    input_bias_ = InputBiasFrom<GType>(inputs, scope);
-    input_mean_ = InputMeanFrom<GType>(inputs, scope);
-    input_scale_ = InputScaleFrom<GType>(inputs, scope);
-    input_variance_ = InputVarianceFrom<GType>(inputs, scope);
-    epsilon_ = GetAttr<float>("epsilon", attrs);
-    momentum_ = GetAttr<float>("momentum", attrs);
-    //    is_test_ = GetAttr<bool>("is_test", attrs);
+                          const AttributeMap &attrs, const Scope &scope)
+      : ConvParam<Dtype>(inputs, outputs, attrs, scope) {
+    output_ = OpParam::OutFrom<GType>(outputs, scope);
+    input_bias_ = OpParam::InputBiasFrom<GType>(inputs, scope);
+    input_mean_ = OpParam::InputMeanFrom<GType>(inputs, scope);
+    input_scale_ = OpParam::InputScaleFrom<GType>(inputs, scope);
+    input_variance_ = OpParam::InputVarianceFrom<GType>(inputs, scope);
+    epsilon_ = OpParam::GetAttr<float>("epsilon", attrs);
+    momentum_ = OpParam::GetAttr<float>("momentum", attrs);
+    //    is_test_ = OpParam::GetAttr<bool>("is_test", attrs);
   }
-
-  const RType *Input() const { return input_; }
-
-  const RType *Filter() const { return filter_; }
-
   RType *Output() const { return output_; }
 
-  const vector<int> &Strides() const { return strides_; }
-
-  const vector<int> &Paddings() const { return paddings_; }
-
-  const vector<int> &Dilations() const { return dilations_; }
-
-  const int &Groups() const { return groups; }
-
   const RType *InputBias() const { return input_bias_; }
 
   const RType *InputMean() const { return input_mean_; }
@@ -1939,13 +1766,7 @@ class FusionDWConvBNReluParam : public OpParam {
   const RType *NewBias() const { return new_bias_; }
 
  protected:
-  RType *input_;
   RType *output_;
-  RType *filter_;
-  vector<int> strides_;
-  vector<int> paddings_;
-  vector<int> dilations_;
-  int groups;
   RType *input_bias_;
   RType *input_mean_;
   RType *input_scale_;
@@ -1961,45 +1782,26 @@ class FusionDWConvBNReluParam : public OpParam {
 
 #ifdef FUSION_CONVBNRELU_OP
 template <typename Dtype>
-class FusionConvBNReluParam : public OpParam {
+class FusionConvBNReluParam : public ConvParam<Dtype> {
   typedef typename DtypeTensorTrait<Dtype>::gtype GType;
   typedef typename DtypeTensorTrait<Dtype>::rtype RType;
 
  public:
   FusionConvBNReluParam(const VariableNameMap &inputs,
                         const VariableNameMap &outputs,
-                        const AttributeMap &attrs, const Scope &scope) {
-    filter_ = FilterFrom<GType>(inputs, scope);
-    input_ = InputFrom<GType>(inputs, scope);
-    output_ = OutFrom<GType>(outputs, scope);
-
-    strides_ = GetAttr<vector<int>>("strides", attrs);
-    paddings_ = GetAttr<vector<int>>("paddings", attrs);
-    dilations_ = GetAttr<vector<int>>("dilations", attrs);
-    groups = GetAttr<int>("groups", attrs);
-    input_bias_ = InputBiasFrom<GType>(inputs, scope);
-    input_mean_ = InputMeanFrom<GType>(inputs, scope);
-    input_scale_ = InputScaleFrom<GType>(inputs, scope);
-    input_variance_ = InputVarianceFrom<GType>(inputs, scope);
-    epsilon_ = GetAttr<float>("epsilon", attrs);
-    momentum_ = GetAttr<float>("momentum", attrs);
-    //    is_test_ = GetAttr<bool>("is_test", attrs);
+                        const AttributeMap &attrs, const Scope &scope)
+      : ConvParam<Dtype>(inputs, outputs, attrs, scope) {
+    output_ = OpParam::OutFrom<GType>(outputs, scope);
+    input_bias_ = OpParam::InputBiasFrom<GType>(inputs, scope);
+    input_mean_ = OpParam::InputMeanFrom<GType>(inputs, scope);
+    input_scale_ = OpParam::InputScaleFrom<GType>(inputs, scope);
+    input_variance_ = OpParam::InputVarianceFrom<GType>(inputs, scope);
+    epsilon_ = OpParam::GetAttr<float>("epsilon", attrs);
+    momentum_ = OpParam::GetAttr<float>("momentum", attrs);
+    //    is_test_ = OpParam::GetAttr<bool>("is_test", attrs);
   }
-
-  const RType *Input() const { return input_; }
-
-  const RType *Filter() const { return filter_; }
-
   RType *Output() const { return output_; }
 
-  const vector<int> &Strides() const { return strides_; }
-
-  const vector<int> &Paddings() const { return paddings_; }
-
-  const vector<int> &Dilations() const { return dilations_; }
-
-  const int &Groups() const { return groups; }
-
   const RType *InputBias() const { return input_bias_; }
 
   const RType *InputMean() const { return input_mean_; }
@@ -2023,13 +1825,7 @@ class FusionConvBNReluParam : public OpParam {
   const RType *NewBias() const { return new_bias_; }
 
  protected:
-  RType *input_;
   RType *output_;
-  RType *filter_;
-  vector<int> strides_;
-  vector<int> paddings_;
-  vector<int> dilations_;
-  int groups;
   RType *input_bias_;
   RType *input_mean_;
   RType *input_scale_;
@@ -2331,16 +2127,14 @@ class ShapeParam : public OpParam {
 };
 #endif
 
-template<typename Dtype>
+template <typename Dtype>
 class QuantizeParam : public OpParam {
   typedef typename DtypeTensorTrait<Dtype>::gtype GType;
   typedef typename DtypeTensorTrait<Dtype>::rtype RType;
 
  public:
-  QuantizeParam(const VariableNameMap &inputs,
-                const VariableNameMap &outputs,
-                const AttributeMap &attrs,
-                const Scope &scope) {
+  QuantizeParam(const VariableNameMap &inputs, const VariableNameMap &outputs,
+                const AttributeMap &attrs, const Scope &scope) {
     input_ = InputXFrom<GType>(inputs, scope);
     out_ = OutFrom<GType>(outputs, scope);
     if (HasAttr("is_static", attrs)) {
@@ -2375,16 +2169,14 @@ class QuantizeParam : public OpParam {
   RoundType round_type_ = ROUND_NEAREST_TO_EVEN;
 };
 
-template<typename Dtype>
+template <typename Dtype>
 class DequantizeParam : public OpParam {
   typedef typename DtypeTensorTrait<Dtype>::gtype GType;
   typedef typename DtypeTensorTrait<Dtype>::rtype RType;
 
  public:
-  DequantizeParam(const VariableNameMap &inputs,
-                const VariableNameMap &outputs,
-                const AttributeMap &attrs,
-                const Scope &scope) {
+  DequantizeParam(const VariableNameMap &inputs, const VariableNameMap &outputs,
+                  const AttributeMap &attrs, const Scope &scope) {
     input_ = InputXFrom<GType>(inputs, scope);
     out_ = OutFrom<GType>(outputs, scope);
     activation_scale_ = GetVarValue<GType>("Scale", inputs, scope);

src/protobuf-c/protobuf-c.h

@@ -202,34 +202,34 @@ size_t foo__bar__baz_bah__pack_to_buffer
 #include <stdint.h>
 
 #ifdef __cplusplus
-# define PROTOBUF_C__BEGIN_DECLS	extern "C" {
-# define PROTOBUF_C__END_DECLS		}
+#define PROTOBUF_C__BEGIN_DECLS extern "C" {
+#define PROTOBUF_C__END_DECLS }
 #else
-# define PROTOBUF_C__BEGIN_DECLS
-# define PROTOBUF_C__END_DECLS
+#define PROTOBUF_C__BEGIN_DECLS
+#define PROTOBUF_C__END_DECLS
 #endif
 
 PROTOBUF_C__BEGIN_DECLS
 
 #if defined(_WIN32) && defined(PROTOBUF_C_USE_SHARED_LIB)
-# ifdef PROTOBUF_C_EXPORT
-#  define PROTOBUF_C__API __declspec(dllexport)
-# else
-#  define PROTOBUF_C__API __declspec(dllimport)
-# endif
+#ifdef PROTOBUF_C_EXPORT
+#define PROTOBUF_C__API __declspec(dllexport)
 #else
-# define PROTOBUF_C__API
+#define PROTOBUF_C__API __declspec(dllimport)
+#endif
+#else
+#define PROTOBUF_C__API
 #endif
 
 #if !defined(PROTOBUF_C__NO_DEPRECATED) && \
     ((__GNUC__ > 3) || (__GNUC__ == 3 && __GNUC_MINOR__ >= 1))
-# define PROTOBUF_C__DEPRECATED __attribute__((__deprecated__))
+#define PROTOBUF_C__DEPRECATED __attribute__((__deprecated__))
 #else
-# define PROTOBUF_C__DEPRECATED
+#define PROTOBUF_C__DEPRECATED
 #endif
 
 #ifndef PROTOBUF_C__FORCE_ENUM_TO_BE_INT_SIZE
- #define PROTOBUF_C__FORCE_ENUM_TO_BE_INT_SIZE(enum_name) \
+#define PROTOBUF_C__FORCE_ENUM_TO_BE_INT_SIZE(enum_name) \
   , _##enum_name##_IS_INT_SIZE = INT_MAX
 #endif
 
@@ -441,9 +441,7 @@ protobuf_c_message_pack_to_buffer(&message, &tmp);
  */
 struct ProtobufCBuffer {
   /** Append function. Consumes the `len` bytes stored at `data`. */
-	void		(*append)(ProtobufCBuffer *buffer,
-				  size_t len,
-				  const uint8_t *data);
+  void (*append)(ProtobufCBuffer *buffer, size_t len, const uint8_t *data);
 };
 
 /**
@@ -733,10 +731,8 @@ struct ProtobufCService {
   /** Service descriptor. */
   const ProtobufCServiceDescriptor *descriptor;
   /** Function to invoke the service. */
-	void (*invoke)(ProtobufCService *service,
-		       unsigned method_index,
-		       const ProtobufCMessage *input,
-		       ProtobufCClosure closure,
+  void (*invoke)(ProtobufCService *service, unsigned method_index,
+                 const ProtobufCMessage *input, ProtobufCClosure closure,
                  void *closure_data);
   /** Function to destroy the service. */
   void (*destroy)(ProtobufCService *service);
@@ -772,8 +768,7 @@ struct ProtobufCServiceDescriptor {
  * \return A string containing the version number of protobuf-c.
  */
 PROTOBUF_C__API
-const char *
-protobuf_c_version(void);
+const char *protobuf_c_version(void);
 
 /**
  * Get the version of the protobuf-c library. Note that this is the version of
@@ -783,8 +778,7 @@ protobuf_c_version(void);
  *      protobuf-c, represented in base-10 as (MAJOR*1E6) + (MINOR*1E3) + PATCH.
  */
 PROTOBUF_C__API
-uint32_t
-protobuf_c_version_number(void);
+uint32_t protobuf_c_version_number(void);
 
 /**
  * The version of the protobuf-c headers, represented as a string using the same
@@ -818,10 +812,8 @@ protobuf_c_version_number(void);
  *      If not found or if the optimize_for = CODE_SIZE option was set.
  */
 PROTOBUF_C__API
-const ProtobufCEnumValue *
-protobuf_c_enum_descriptor_get_value_by_name(
-	const ProtobufCEnumDescriptor *desc,
-	const char *name);
+const ProtobufCEnumValue *protobuf_c_enum_descriptor_get_value_by_name(
+    const ProtobufCEnumDescriptor *desc, const char *name);
 
 /**
  * Look up a `ProtobufCEnumValue` from a `ProtobufCEnumDescriptor` by numeric
@@ -839,10 +831,8 @@ protobuf_c_enum_descriptor_get_value_by_name(
  *      If not found.
  */
 PROTOBUF_C__API
-const ProtobufCEnumValue *
-protobuf_c_enum_descriptor_get_value(
-	const ProtobufCEnumDescriptor *desc,
-	int value);
+const ProtobufCEnumValue *protobuf_c_enum_descriptor_get_value(
+    const ProtobufCEnumDescriptor *desc, int value);
 
 /**
  * Look up a `ProtobufCFieldDescriptor` from a `ProtobufCMessageDescriptor` by
@@ -858,10 +848,8 @@ protobuf_c_enum_descriptor_get_value(
  *      If not found or if the optimize_for = CODE_SIZE option was set.
  */
 PROTOBUF_C__API
-const ProtobufCFieldDescriptor *
-protobuf_c_message_descriptor_get_field_by_name(
-	const ProtobufCMessageDescriptor *desc,
-	const char *name);
+const ProtobufCFieldDescriptor *protobuf_c_message_descriptor_get_field_by_name(
+    const ProtobufCMessageDescriptor *desc, const char *name);
 
 /**
  * Look up a `ProtobufCFieldDescriptor` from a `ProtobufCMessageDescriptor` by
@@ -877,10 +865,8 @@ protobuf_c_message_descriptor_get_field_by_name(
  *      If not found.
  */
 PROTOBUF_C__API
-const ProtobufCFieldDescriptor *
-protobuf_c_message_descriptor_get_field(
-	const ProtobufCMessageDescriptor *desc,
-	unsigned value);
+const ProtobufCFieldDescriptor *protobuf_c_message_descriptor_get_field(
+    const ProtobufCMessageDescriptor *desc, unsigned value);
 
 /**
  * Determine the number of bytes required to store the serialised message.
@@ -891,8 +877,7 @@ protobuf_c_message_descriptor_get_field(
  *      Number of bytes.
  */
 PROTOBUF_C__API
-size_t
-protobuf_c_message_get_packed_size(const ProtobufCMessage *message);
+size_t protobuf_c_message_get_packed_size(const ProtobufCMessage *message);
 
 /**
  * Serialise a message from its in-memory representation.
@@ -911,8 +896,7 @@ protobuf_c_message_get_packed_size(const ProtobufCMessage *message);
  *      Number of bytes stored in `out`.
  */
 PROTOBUF_C__API
-size_t
-protobuf_c_message_pack(const ProtobufCMessage *message, uint8_t *out);
+size_t protobuf_c_message_pack(const ProtobufCMessage *message, uint8_t *out);
 
 /**
  * Serialise a message from its in-memory representation to a virtual buffer.
@@ -928,9 +912,7 @@ protobuf_c_message_pack(const ProtobufCMessage *message, uint8_t *out);
  *      Number of bytes passed to the virtual buffer.
  */
 PROTOBUF_C__API
-size_t
-protobuf_c_message_pack_to_buffer(
-	const ProtobufCMessage *message,
+size_t protobuf_c_message_pack_to_buffer(const ProtobufCMessage *message,
                                          ProtobufCBuffer *buffer);
 
 /**
@@ -951,12 +933,9 @@ protobuf_c_message_pack_to_buffer(
  *      If an error occurred during unpacking.
  */
 PROTOBUF_C__API
-ProtobufCMessage *
-protobuf_c_message_unpack(
-	const ProtobufCMessageDescriptor *descriptor,
-	ProtobufCAllocator *allocator,
-	size_t len,
-	const uint8_t *data);
+ProtobufCMessage *protobuf_c_message_unpack(
+    const ProtobufCMessageDescriptor *descriptor, ProtobufCAllocator *allocator,
+    size_t len, const uint8_t *data);
 
 /**
  * Free an unpacked message object.
@@ -971,9 +950,7 @@ protobuf_c_message_unpack(
  *      specify the default allocator.
  */
 PROTOBUF_C__API
-void
-protobuf_c_message_free_unpacked(
-	ProtobufCMessage *message,
+void protobuf_c_message_free_unpacked(ProtobufCMessage *message,
                                       ProtobufCAllocator *allocator);
 
 /**
@@ -988,11 +965,11 @@ protobuf_c_message_free_unpacked(
  *      Message is invalid.
  */
 PROTOBUF_C__API
-protobuf_c_boolean
-protobuf_c_message_check(const ProtobufCMessage *);
+protobuf_c_boolean protobuf_c_message_check(const ProtobufCMessage *);
 
 /** Message initialiser. */
-#define PROTOBUF_C_MESSAGE_INIT(descriptor) { descriptor, 0, NULL }
+#define PROTOBUF_C_MESSAGE_INIT(descriptor) \
+  { descriptor, 0, NULL }
 
 /**
  * Initialise a message object from a message descriptor.
@@ -1003,9 +980,7 @@ protobuf_c_message_check(const ProtobufCMessage *);
  *      Allocated block of memory of size `descriptor->sizeof_message`.
  */
 PROTOBUF_C__API
-void
-protobuf_c_message_init(
-	const ProtobufCMessageDescriptor *descriptor,
+void protobuf_c_message_init(const ProtobufCMessageDescriptor *descriptor,
                              void *message);
 
 /**
@@ -1015,8 +990,7 @@ protobuf_c_message_init(
  *      The service object to free.
  */
 PROTOBUF_C__API
-void
-protobuf_c_service_destroy(ProtobufCService *service);
+void protobuf_c_service_destroy(ProtobufCService *service);
 
 /**
  * Look up a `ProtobufCMethodDescriptor` by name.
@@ -1034,36 +1008,29 @@ protobuf_c_service_destroy(ProtobufCService *service);
 PROTOBUF_C__API
 const ProtobufCMethodDescriptor *
 protobuf_c_service_descriptor_get_method_by_name(
-	const ProtobufCServiceDescriptor *desc,
-	const char *name);
+    const ProtobufCServiceDescriptor *desc, const char *name);
 
 /**
  * Initialise a `ProtobufCBufferSimple` object.
  */
 #define PROTOBUF_C_BUFFER_SIMPLE_INIT(array_of_bytes)             \
-{                                                                       \
-	{ protobuf_c_buffer_simple_append },                            \
-	sizeof(array_of_bytes),                                         \
-	0,                                                              \
-	(array_of_bytes),                                               \
-	0,                                                              \
-	NULL                                                            \
-}
+  {                                                               \
+    {protobuf_c_buffer_simple_append}, sizeof(array_of_bytes), 0, \
+        (array_of_bytes), 0, NULL                                 \
+  }
 
 /**
  * Clear a `ProtobufCBufferSimple` object, freeing any allocated memory.
  */
 #define PROTOBUF_C_BUFFER_SIMPLE_CLEAR(simp_buf)                              \
-do {                                                                    \
+  do {                                                                        \
     if ((simp_buf)->must_free_data) {                                         \
       if ((simp_buf)->allocator != NULL)                                      \
-			(simp_buf)->allocator->free(                    \
-				(simp_buf)->allocator,                  \
-				(simp_buf)->data);			\
+        (simp_buf)->allocator->free((simp_buf)->allocator, (simp_buf)->data); \
       else                                                                    \
         free((simp_buf)->data);                                               \
     }                                                                         \
-} while (0)
+  } while (0)
 
 /**
  * The `append` method for `ProtobufCBufferSimple`.
@@ -1077,23 +1044,16 @@ do {                                                                    \
  *      Data to append.
  */
 PROTOBUF_C__API
-void
-protobuf_c_buffer_simple_append(
-	ProtobufCBuffer *buffer,
-	size_t len,
+void protobuf_c_buffer_simple_append(ProtobufCBuffer *buffer, size_t len,
                                      const unsigned char *data);
 
 PROTOBUF_C__API
-void
-protobuf_c_service_generated_init(
-	ProtobufCService *service,
-	const ProtobufCServiceDescriptor *descriptor,
+void protobuf_c_service_generated_init(
+    ProtobufCService *service, const ProtobufCServiceDescriptor *descriptor,
     ProtobufCServiceDestroy destroy);
 
 PROTOBUF_C__API
-void
-protobuf_c_service_invoke_internal(
-	ProtobufCService *service,
+void protobuf_c_service_invoke_internal(ProtobufCService *service,
                                         unsigned method_index,
                                         const ProtobufCMessage *input,
                                         ProtobufCClosure closure,

tools/build.sh

 #!/usr/bin/env bash
 NETS=""
-declare -a supportedNets=("googlenet" "mobilenet" "yolo" "squeezenet" "resnet" "mobilenetssd" "nlp")
+declare -a supportedNets=("googlenet" "mobilenet" "yolo" "squeezenet" "resnet" "mobilenetssd" "nlp" "mobilenetfssd" "genet")
 
 build_for_mac() {
     if [ ! `which brew` ]; then

tools/op.cmake

@@ -177,6 +177,7 @@ if (CON GREATER -1)
   set(FOUND_MATCH ON)
 endif()
 
+
 if(NOT FOUND_MATCH)
   message("--default--")
   set(BATCHNORM_OP ON)

tools/pre-commit.hooks/cpplint.hook

@@ -3,7 +3,7 @@
 TOTAL_ERRORS=0
 
 # The trick to remove deleted files: https://stackoverflow.com/a/2413151
-for file in $(git diff --cached --name-status | awk '$1 != "D" {print $2}' | grep -v ".pb.cpp" | grep -v ".pb.h" | grep -v "protobuf-c.*"); do
+for file in $(git diff --cached --name-status | awk '$1 != "D" {print $2}' | grep -v ".pb.cpp" | grep -v ".pb.h"); do
     cpplint $file;
     TOTAL_ERRORS=$(expr $TOTAL_ERRORS + $?);
 done