Merge branch 'dev-latest' of https://github.com/hjchen2/paddle-mobile into dev-latest

.gitignore

@@ -93,3 +93,6 @@ metal/paddle-mobile/paddle-mobile/CPU/libpaddle-mobile.a
 *.xcuserdatad/
 */xcuserdata/
 /venv/
+
+metal/paddle-mobile-demo/paddle-mobile-demo/images
+metal/paddle-mobile-demo/paddle-mobile-demo/models

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

@@ -17,7 +17,7 @@ import MetalKit
 import Foundation
 import paddle_mobile
 
-class MetalHelper {
+public class MetalHelper {
   let device: MTLDevice
   let queue: MTLCommandQueue
   let textureLoader: MTKTextureLoader

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

@@ -23,23 +23,23 @@ 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)]
+                                               .yolo : Runner.init(inNet: YoloNet.init(device: MetalHelper.shared.device), commandQueue: MetalHelper.shared.queue, inPlatform: platform),
+                                               .mobilenet_combined : Runner.init(inNet: MobileNetCombined.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)]
+let netSupport: [SupportModel : Net] = [.yolo : YoloNet.init(device: MetalHelper.shared.device), .mobilenet_combined : MobileNetCombined.init(device: MetalHelper.shared.device)]
 
 enum SupportModel: String{
   //  case mobilenet = "mobilenet"
 //  case mobilenet_ssd    = "mobilenetssd"
-  case genet            = "genet"
-  case mobilenet_ssd_ar = "mobilenetssd_ar"
+  case yolo            = "yolo"
+  case mobilenet_combined = "mobilenet_combined"
   
   static func supportedModels() -> [SupportModel] {
     // .mobilenet,
     // .mobilenet_ssd,
-    return [.genet, .mobilenet_ssd_ar]
+    return [.yolo, .mobilenet_combined]
   }
 }
 
@@ -98,7 +98,7 @@ class ViewController: UIViewController {
   }
   
   @IBAction func predictAct(_ sender: Any) {
-    let max = 50
+    let max = 1
     switch platform {
     case .GPU:
       guard let inTexture = toPredictTexture else {
@@ -106,11 +106,11 @@ class ViewController: UIViewController {
         return
       }
       
-      for _ in 0..<10{
-        runner.predict(texture: inTexture) { (success, resultHolder)  in
-          resultHolder?.releasePointer()
-        }
-      }
+//      for _ in 0..<1{
+//        runner.predict(texture: inTexture) { (success, resultHolder)  in
+//          resultHolder?.releasePointer()
+//        }
+//      }
       
       let startDate = Date.init()
       for i in 0..<max {

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

@@ -29,6 +29,11 @@
 		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 */; };
+		C28FDF8421B7858F0054EFAC /* MobileNetCombined.swift in Sources */ = {isa = PBXBuildFile; fileRef = C28FDF8221B7858F0054EFAC /* MobileNetCombined.swift */; };
+		C28FDF8521B7858F0054EFAC /* YoloNet.swift in Sources */ = {isa = PBXBuildFile; fileRef = C28FDF8321B7858F0054EFAC /* YoloNet.swift */; };
+		C28FE02F21BA68C00054EFAC /* Metal.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = C28FE02C21BA68C00054EFAC /* Metal.framework */; };
+		C28FE03021BA68C00054EFAC /* MetalPerformanceShaders.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = C28FE02D21BA68C00054EFAC /* MetalPerformanceShaders.framework */; };
+		C28FE03121BA68C00054EFAC /* MetalKit.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = C28FE02E21BA68C00054EFAC /* MetalKit.framework */; };
 		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 */; };
@@ -72,7 +77,6 @@
 		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 */; };
@@ -160,6 +164,11 @@
 		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>"; };
+		C28FDF8221B7858F0054EFAC /* MobileNetCombined.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = MobileNetCombined.swift; sourceTree = "<group>"; };
+		C28FDF8321B7858F0054EFAC /* YoloNet.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = YoloNet.swift; sourceTree = "<group>"; };
+		C28FE02C21BA68C00054EFAC /* Metal.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Metal.framework; path = System/Library/Frameworks/Metal.framework; sourceTree = SDKROOT; };
+		C28FE02D21BA68C00054EFAC /* MetalPerformanceShaders.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = MetalPerformanceShaders.framework; path = System/Library/Frameworks/MetalPerformanceShaders.framework; sourceTree = SDKROOT; };
+		C28FE02E21BA68C00054EFAC /* MetalKit.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = MetalKit.framework; path = System/Library/Frameworks/MetalKit.framework; sourceTree = SDKROOT; };
 		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>"; };
@@ -277,7 +286,9 @@
 			isa = PBXFrameworksBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
-				FC4FD97E2140F2C30073E130 /* libstdc++.tbd in Frameworks */,
+				C28FE02F21BA68C00054EFAC /* Metal.framework in Frameworks */,
+				C28FE03021BA68C00054EFAC /* MetalPerformanceShaders.framework in Frameworks */,
+				C28FE03121BA68C00054EFAC /* MetalKit.framework in Frameworks */,
 				D3831F70E7E0B565B9AC22DA /* Pods_paddle_mobile.framework in Frameworks */,
 				FC4FD97A2140E4980073E130 /* libpaddle-mobile.a in Frameworks */,
 			);
@@ -289,6 +300,9 @@
 		336CBE234BF5DE48658DE65F /* Frameworks */ = {
 			isa = PBXGroup;
 			children = (
+				C28FE02C21BA68C00054EFAC /* Metal.framework */,
+				C28FE02E21BA68C00054EFAC /* MetalKit.framework */,
+				C28FE02D21BA68C00054EFAC /* MetalPerformanceShaders.framework */,
 				FC4FD97D2140F2C30073E130 /* libstdc++.tbd */,
 				DD2E06330A1E7129C918DB46 /* Pods_paddle_mobile.framework */,
 			);
@@ -325,6 +339,8 @@
 		FC039B6C20E11C3C0081E9F8 /* paddle-mobile */ = {
 			isa = PBXGroup;
 			children = (
+				C28FDF8221B7858F0054EFAC /* MobileNetCombined.swift */,
+				C28FDF8321B7858F0054EFAC /* YoloNet.swift */,
 				FCE9D7B6214F869000B520C3 /* Net.swift */,
 				FC9A19E22148C31300CD9CBF /* MobilenetSSD_AR.swift */,
 				FC33B0EF2147659000714A93 /* MobileNet.swift */,
@@ -681,6 +697,7 @@
 				FC039BBA20E11CC20081E9F8 /* TensorDesc.swift in Sources */,
 				FC039BA020E11CB20081E9F8 /* Dim.swift in Sources */,
 				FC039BB820E11CC20081E9F8 /* framework.pb.swift in Sources */,
+				C28FDF8521B7858F0054EFAC /* YoloNet.swift in Sources */,
 				FC039B9920E11C9A0081E9F8 /* Types.swift in Sources */,
 				FC4CB74920F0B954007C0C6D /* ConvKernel.metal in Sources */,
 				FCA3A1632132A4AC00084FE5 /* ReshapeKernel.metal in Sources */,
@@ -725,6 +742,7 @@
 				FCE9D7B7214F869000B520C3 /* Net.swift in Sources */,
 				FC0E2DBE20EE460D009C1FAC /* BatchNormKernel.swift in Sources */,
 				FC039BAB20E11CBC0081E9F8 /* Operator.swift in Sources */,
+				C28FDF8421B7858F0054EFAC /* MobileNetCombined.swift in Sources */,
 				FCD04E6A20F319EC0007374F /* SoftmaxOp.swift in Sources */,
 				FC292C82214255BD00CF622F /* MobileNetSSD.swift in Sources */,
 				FCBCCC612122FBDF00D94F7E /* PriorBoxKernel.swift in Sources */,
@@ -885,9 +903,9 @@
 			buildSettings = {
 				CLANG_ENABLE_MODULES = YES;
 				CODE_SIGN_IDENTITY = "";
-				CODE_SIGN_STYLE = Automatic;
+				CODE_SIGN_STYLE = Manual;
 				DEFINES_MODULE = YES;
-				DEVELOPMENT_TEAM = A798K58VVL;
+				DEVELOPMENT_TEAM = "";
 				DYLIB_COMPATIBILITY_VERSION = 1;
 				DYLIB_CURRENT_VERSION = 1;
 				DYLIB_INSTALL_NAME_BASE = "@rpath";
@@ -904,10 +922,11 @@
 					"$(inherited)",
 					"$(PROJECT_DIR)/paddle-mobile/CPU",
 				);
-				MACH_O_TYPE = mh_dylib;
+				MACH_O_TYPE = staticlib;
 				MTL_LANGUAGE_REVISION = UseDeploymentTarget;
 				PRODUCT_BUNDLE_IDENTIFIER = "orange.paddle-mobile";
 				PRODUCT_NAME = "$(TARGET_NAME:c99extidentifier)";
+				PROVISIONING_PROFILE_SPECIFIER = "";
 				SKIP_INSTALL = YES;
 				SWIFT_OPTIMIZATION_LEVEL = "-Onone";
 				SWIFT_VERSION = 4.0;
@@ -921,9 +940,9 @@
 			buildSettings = {
 				CLANG_ENABLE_MODULES = YES;
 				CODE_SIGN_IDENTITY = "";
-				CODE_SIGN_STYLE = Automatic;
+				CODE_SIGN_STYLE = Manual;
 				DEFINES_MODULE = YES;
-				DEVELOPMENT_TEAM = A798K58VVL;
+				DEVELOPMENT_TEAM = "";
 				DYLIB_COMPATIBILITY_VERSION = 1;
 				DYLIB_CURRENT_VERSION = 1;
 				DYLIB_INSTALL_NAME_BASE = "@rpath";
@@ -940,10 +959,11 @@
 					"$(inherited)",
 					"$(PROJECT_DIR)/paddle-mobile/CPU",
 				);
-				MACH_O_TYPE = mh_dylib;
+				MACH_O_TYPE = staticlib;
 				MTL_LANGUAGE_REVISION = UseDeploymentTarget;
 				PRODUCT_BUNDLE_IDENTIFIER = "orange.paddle-mobile";
 				PRODUCT_NAME = "$(TARGET_NAME:c99extidentifier)";
+				PROVISIONING_PROFILE_SPECIFIER = "";
 				SKIP_INSTALL = YES;
 				SWIFT_VERSION = 4.0;
 				TARGETED_DEVICE_FAMILY = "1,2";

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

@@ -13,6 +13,8 @@
  limitations under the License. */
 
 import Foundation
+import MetalKit
+import CoreMedia
 
 fileprivate var defaultMetalLibrary: MTLLibrary?
 fileprivate var paddleMobileMetalLibrary: MTLLibrary?

metal/paddle-mobile/paddle-mobile/MobileNetCombined.swift

+//
+//  MobileNetCombined.swift
+//  paddle-mobile
+//
+//  Created by Xiao,Haichun on 2018/12/5.
+//  Copyright © 2018 orange. All rights reserved.
+//
+
+import Foundation
+
+public class MobileNetCombined: Net {
+    @objc public override init(device: MTLDevice) {
+        super.init(device: device)
+        means = [0, 0, 0]
+        scale = 1
+        except = 0
+        modelPath = Bundle.main.path(forResource: "combined_mobilenet_model", ofType: nil) ?! "model null"
+        paramPath = Bundle.main.path(forResource: "combined_mobilenet_params", ofType: nil) ?! "para null"
+        modelDir = ""
+        //preprocessKernel = GenetPreProccess.init(device: device)
+        dim = (n: 1, h: 416, w: 416, 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 = [0, 0, 0]
+        scale = 1
+        except = 0
+        modelPath = ""
+        paramPath = ""
+        modelDir = ""
+        //preprocessKernel = GenetPreProccess.init(device: device)
+        dim = (n: 1, h: 416, w: 416, 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/Net.swift

@@ -14,6 +14,7 @@
 
 
 import Foundation
+import Metal
 
 public class ResultHolder: NSObject {
   @objc public let result: UnsafeMutablePointer<Float32>?

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

@@ -67,7 +67,8 @@ class OpCreator<P: PrecisionType> {
          gFlatten                   :     FlattenOp<P>.creat,
          gConvAddPreluType          :     ConvAddPreluOp<P>.creat,
          gConvAddAddPreluType       :     ConvAddAddPreluOp<P>.creat,
-         gElementwiseAddPreluType:   ElementwiseAddPreluOp<P>.creat]
+         gElementwiseAddPreluType   :     ElementwiseAddPreluOp<P>.creat,
+         gFusionConvAddType         :     ConvAddOp<P>.creat]
   
     private init(){}
 }

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

@@ -26,7 +26,7 @@ protocol OpParam {
   var output: OutputType { get set }
   func outputDesc() -> String
   
-  associatedtype ParamPrecisionType: PrecisionType
+  //associatedtype ParamPrecisionType: PrecisionType
   init(opDesc: OpDesc, inScope: Scope) throws
   static func getFirstTensor<VarType: Variant>(key: String, map: [String : [String]], from: Scope) throws -> VarType
   static func inputX<VarType: Variant>(inputs: [String : [String]], from: Scope) throws -> VarType

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

@@ -159,6 +159,7 @@ let gFlatten                    = "flatten"
 let gConvAddPreluType           = "conv_add_prelu"
 let gConvAddAddPreluType        = "conv_add_add_prelu"
 let gElementwiseAddPreluType    = "elementwise_add_prelu"
+let gFusionConvAddType          = "fusion_conv_add"
 
 
 let opInfos = [gConvType                    : (inputs: ["Input"], outputs: ["Output"]),
@@ -188,5 +189,6 @@ let opInfos = [gConvType                    : (inputs: ["Input"], outputs: ["Out
                gFlatten                     : (inputs: ["X"], outputs: ["Out"]),
                gConvAddPreluType            : (inputs: ["Input"], outputs: ["Out"]),
                gConvAddAddPreluType         : (inputs: ["Input"], outputs: ["Out"]),
-               gElementwiseAddPreluType  :  (inputs: ["X"], outputs: ["Out"])
+               gElementwiseAddPreluType     : (inputs: ["X"], outputs: ["Out"]),
+               gFusionConvAddType           : (inputs: ["Input"], outputs: ["Out"])
               ]

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

@@ -13,9 +13,10 @@
  limitations under the License. */
 
 import Foundation
+import Metal
 
 class BatchNormParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       input = try BatchNormParam.inputX(inputs: opDesc.inputs, from: inScope)

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

@@ -13,9 +13,10 @@
  limitations under the License. */
 
 import Foundation
+import Metal
 
 class BilinearInterpParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       input = try BilinearInterpParam.inputX(inputs: opDesc.inputs, from: inScope)

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

@@ -15,7 +15,7 @@
 import Foundation
 
 class BoxcoderParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       priorBox = try BoxcoderParam.getFirstTensor(key: "PriorBox", map: opDesc.inputs, from: inScope)

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

@@ -15,7 +15,7 @@
 import Foundation
 
 class ConcatParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       guard let xlist = opDesc.inputs["X"] else {

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

@@ -13,9 +13,10 @@
  limitations under the License. */
 
 import Foundation
+import Metal
 
 class ConvAddAddPreluParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       filter = try ConvAddAddPreluParam.inputFilter(paraInputs: opDesc.paraInputs, from: inScope)
@@ -34,8 +35,8 @@ class ConvAddAddPreluParam<P: PrecisionType>: OpParam {
   }
   
   let input: Texture<P>
-  let y: Tensor<ParamPrecisionType>
-  let filter: Tensor<ParamPrecisionType>
+  let y: Tensor<P>
+  let filter: Tensor<P>
   let mode: String
   let alpha: Tensor<P>
   var output: Texture<P>

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

@@ -16,7 +16,7 @@ import Foundation
 
 
 class ConvAddBatchNormReluParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       
@@ -42,12 +42,12 @@ class ConvAddBatchNormReluParam<P: PrecisionType>: OpParam {
   
   let input: Texture<P>
   
-  let variance: Tensor<ParamPrecisionType>
-  let bias: Tensor<ParamPrecisionType>
-  let mean: Tensor<ParamPrecisionType>
-  let scale: Tensor<ParamPrecisionType>
-  let y: Tensor<ParamPrecisionType>
-  let filter: Tensor<ParamPrecisionType>
+  let variance: Tensor<P>
+  let bias: Tensor<P>
+  let mean: Tensor<P>
+  let scale: Tensor<P>
+  let y: Tensor<P>
+  let filter: Tensor<P>
   let epsilon: Float32
   var newScale: MTLBuffer?
   var newBiase: MTLBuffer?

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

@@ -15,7 +15,7 @@
 import Foundation
 
 class ConvAddParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       filter = try ConvAddParam.inputFilter(paraInputs: opDesc.paraInputs, from: inScope)
@@ -33,8 +33,8 @@ class ConvAddParam<P: PrecisionType>: OpParam {
   }
   
   let input: Texture<P>
-  let y: Tensor<ParamPrecisionType>
-  let filter: Tensor<ParamPrecisionType>
+  let y: Tensor<P>
+  let filter: Tensor<P>
   
   var output: Texture<P>
   let stride: [Int32]

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

@@ -15,7 +15,7 @@
 import Foundation
 
 class ConvAddPreluParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       filter = try ConvAddPreluParam.inputFilter(paraInputs: opDesc.paraInputs, from: inScope)
@@ -34,8 +34,8 @@ class ConvAddPreluParam<P: PrecisionType>: OpParam {
   }
   
   let input: Texture<P>
-  let y: Tensor<ParamPrecisionType>
-  let filter: Tensor<ParamPrecisionType>
+  let y: Tensor<P>
+  let filter: Tensor<P>
   let mode: String
   let alpha: Tensor<P>
   var output: Texture<P>

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

@@ -15,7 +15,7 @@
 import Foundation
 
 class ConvBNReluParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       filter = try ConvBNReluParam.inputFilter(paraInputs: opDesc.paraInputs, from: inScope)
@@ -38,11 +38,11 @@ class ConvBNReluParam<P: PrecisionType>: OpParam {
   
   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 variance: Tensor<P>
+  let bias: Tensor<P>
+  let mean: Tensor<P>
+  let scale: Tensor<P>
+  let filter: Tensor<P>
   let epsilon: Float32
   var newScale: MTLBuffer?
   var newBiase: MTLBuffer?

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

@@ -15,7 +15,7 @@
 import Foundation
 
 class ConvParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       filter = try ConvParam.inputFilter(paraInputs: opDesc.paraInputs, from: inScope)
@@ -32,7 +32,7 @@ class ConvParam<P: PrecisionType>: OpParam {
   }
   
   let input: Texture<P>
-  let filter: Tensor<ParamPrecisionType>
+  let filter: Tensor<P>
   var output: Texture<P>
   let stride: [Int32]
   let paddings: [Int32]

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

@@ -15,7 +15,7 @@
 import Foundation
 
 class ConvTransposeParam<P: PrecisionType>: ConvParam<P> {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
     required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       try super.init(opDesc: opDesc, inScope: inScope)

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

@@ -13,9 +13,10 @@
 // limitations under the License. */
 
 import Foundation
+import Metal
 
 class ElementwiseAddParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       inputX = try ElementwiseAddParam.inputX(inputs: opDesc.inputs, from: inScope)

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

@@ -13,9 +13,10 @@
  limitations under the License. */
 
 import Foundation
+import Metal
 
 class ElementwiseAddPreluParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       alpha = try ElementwiseAddPreluParam.paramInputAlpha(inputs: opDesc.paraInputs, from: inScope)

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

@@ -13,6 +13,8 @@
  limitations under the License. */
 
 import Foundation
+import MetalKit
+import CoreMedia
 
 class FeedParam<P: PrecisionType>: OpParam{
   var output: Texture<P>
@@ -30,7 +32,7 @@ class FeedParam<P: PrecisionType>: OpParam{
     }
   }
   
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
 }
 
 class FeedOp<P: PrecisionType>: Operator<Texture2DTo2DArrayKernel<P>, FeedParam<P>>, Runable, Creator, InferShaperable {

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

@@ -13,6 +13,7 @@
  limitations under the License. */
 
 import Foundation
+import Metal
 
 class FetchParam<P: PrecisionType>: OpParam{
   var output: FetchHolder
@@ -29,7 +30,7 @@ class FetchParam<P: PrecisionType>: OpParam{
     }
   }
   
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
 }
 
 class FetchKernel<P: PrecisionType>: Kernel, Computable {

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

@@ -15,7 +15,7 @@
 import Foundation
 
 class FlattenParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       input = try FlattenParam.inputX(inputs: opDesc.inputs, from: inScope)

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

@@ -13,6 +13,7 @@
  limitations under the License. */
 
 import Foundation
+import Metal
 
 struct ConcatTestParam: TestParam {
   var input: [MTLTexture]

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

@@ -13,6 +13,7 @@
  limitations under the License. */
 
 import Foundation
+import Metal
 
 struct ConvAddBatchNormReluTestParam: TestParam {
   let inputTexture: MTLTexture

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

@@ -12,6 +12,7 @@
  See the License for the specific language governing permissions and
  limitations under the License. */
 
+import Metal
 import Foundation
 
 struct ShapeMetalParam {

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

@@ -13,6 +13,8 @@
  limitations under the License. */
 
 import Foundation
+import MetalKit
+import CoreMedia
 
 struct Texture2DTo2DArrayParam {
   let input: MTLTexture

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

@@ -15,7 +15,7 @@
 import Foundation
 
 class MulticlassNMSParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       scores = try MulticlassNMSParam.getFirstTensor(key: "Scores", map: opDesc.inputs, from: inScope)

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

@@ -15,7 +15,7 @@
 import Foundation
 
 class PoolParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       input = try PoolParam.inputX(inputs: opDesc.inputs, from: inScope)

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

@@ -15,7 +15,7 @@
 import Foundation
 
 class PreluParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       input = try PreluParam.inputX(inputs: opDesc.inputs, from: inScope)

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

@@ -15,7 +15,7 @@
 import Foundation
 
 class PriorBoxParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //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)

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

@@ -16,7 +16,7 @@
 import Foundation
 
 class ReluParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       input = try ReluParam.inputX(inputs: opDesc.inputs, from: inScope)

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

@@ -13,9 +13,10 @@
  limitations under the License. */
 
 import Foundation
+import Metal
 
 class ReshapeParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       input = try ReshapeParam.inputX(inputs: opDesc.inputs, from: inScope)

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

@@ -15,7 +15,7 @@
 import Foundation
 
 class ShapeParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+ // typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       input = try ShapeParam.input(inputs: opDesc.inputs, from: inScope)

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

@@ -13,16 +13,17 @@
  limitations under the License. */
 
 import Foundation
+import Metal
 
 class SoftmaxParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     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])
+      //assert(input.tensorDim.dims.count == 2)
+      //assert(input.transpose == [0, 1, 2, 3])
       
       output.dim = input.dim
       output.tensorDim = input.tensorDim

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

@@ -15,7 +15,7 @@
 import Foundation
 
 class SplitParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       input = try SplitParam.inputX(inputs: opDesc.inputs, from: inScope)

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

@@ -13,9 +13,10 @@
  limitations under the License. */
 
 import Foundation
+import Metal
 
 class TransposeParam<P: PrecisionType>: OpParam {
-  typealias ParamPrecisionType = P
+  //typealias ParamPrecisionType = P
   required init(opDesc: OpDesc, inScope: Scope) throws {
     do {
       input = try TransposeParam.inputX(inputs: opDesc.inputs, from: inScope)

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

@@ -33,7 +33,7 @@ class ScaleKernel: CusomKernel {
   
 }
 
-public class Runner: NSObject {
+@objc public class Runner: NSObject {
   var program: Program?
   var executor: Executor<Float32>?
   var queue: MTLCommandQueue?

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

@@ -203,6 +203,7 @@ class ProgramOptimize<P: PrecisionType> {
     var typeMapNodes: [String : [(node: Node, output: [String : Node])]] = [:]
     let block = originProgramDesc.blocks[0]
     for opDesc in block.ops {
+        print(opDesc.type)
       guard let opInputKeys = opInfos[opDesc.type]?.inputs, let outputKeys = opInfos[opDesc.type]?.outputs else {
         fatalError()
       }

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

@@ -85,6 +85,7 @@ extension VarDesc: CustomStringConvertible, CustomDebugStringConvertible {
         if let inTensorDesc = tensorDesc {
             str += " dim size: \(inTensorDesc.dims.count) \n"
             str += "    dim: \(inTensorDesc.dims) \n"
+            str += "type:\(self.type) \n"
         } else {
             str += " no dim info"
         }

metal/paddle-mobile/paddle-mobile/YoloNet.swift

+//
+//  YoloNet.swift
+//  paddle-mobile
+//
+//  Created by Xiao,Haichun on 2018/12/5.
+//  Copyright © 2018 orange. All rights reserved.
+//
+
+import Foundation
+import Metal
+
+public class YoloNet: Net {
+    @objc public override init(device: MTLDevice) {
+        super.init(device: device)
+        means = [0, 0, 0]
+        scale = 1
+        except = 0
+        modelPath = Bundle.main.path(forResource: "yolo_model", ofType: nil) ?! "model null"
+        paramPath = Bundle.main.path(forResource: "yolo_params", ofType: nil) ?! "para null"
+        modelDir = ""
+        //preprocessKernel = GenetPreProccess.init(device: device)
+        dim = (n: 1, h: 224, w: 224, 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 = [0, 0, 0]
+        scale = 1
+        except = 0
+        modelPath = ""
+        paramPath = ""
+        modelDir = ""
+        //preprocessKernel = GenetPreProccess.init(device: device)
+        dim = (n: 1, h: 224, w: 224, 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/framework/Tensor.swift

@@ -13,8 +13,10 @@
  limitations under the License. */
 
 import Foundation
+import MetalKit
+import CoreMedia
 
-protocol Tensorial: CustomStringConvertible, CustomDebugStringConvertible{
+protocol Tensorial: Variant {
   var dim: Dim { get set }
   func numel() -> Int
   var layout: DataLayout { get }
@@ -61,7 +63,7 @@ class Tensor<P: PrecisionType>: Tensorial {
     }
   }
   
-  required init(inDim: Dim, inLayout: DataLayout = DataLayout.NCHW()) {
+  init(inDim: Dim, inLayout: DataLayout = DataLayout.NCHW()) {
     dim = inDim
     let size = inDim.numel() * MemoryLayout<P>.size
     let pointer = UnsafeMutablePointer<P>.allocate(capacity: size)

src/fpga/V2/api.cpp

@@ -204,7 +204,8 @@ void fill_split_arg(struct SplitConvArgs *arg, framework::Tensor *input,
 
     arg->conv_arg[i].image.address = input_ptr;
     arg->conv_arg[i].image.scale_address = input->scale;
-    arg->conv_arg[i].image.channels = (uint32_t)input->dims()[1];
+    arg->conv_arg[i].image.channels =
+        (uint32_t)get_aligned_channel_num((int)(input->dims()[1]));  // NOLINT
     arg->conv_arg[i].image.height = (uint32_t)input->dims()[2];
     arg->conv_arg[i].image.width = (uint32_t)input->dims()[3];
     arg->conv_arg[i].image.pad_height = (uint32_t)padding_h;
@@ -216,7 +217,7 @@ void fill_split_arg(struct SplitConvArgs *arg, framework::Tensor *input,
     int num_after_alignment = filter::calc_aligned_num(
         arg->filter_num, (int)input->dims()[1]);  // NOLINT
     arg->conv_arg[i].free_space =
-        fpga_malloc(num_after_alignment * 2 * sizeof(half));
+        fpga_malloc(num_after_alignment * 2 * sizeof(float));  // half
   }
 }
 

src/fpga/V2/filter.cpp

@@ -16,7 +16,6 @@ limitations under the License. */
 #include <memory.h>
 #include <algorithm>
 #include "fpga/common/fpga_common.h"
-
 namespace paddle_mobile {
 namespace fpga {
 namespace filter {
@@ -88,12 +87,25 @@ void align_filter(float **data_in, int num, int channel, int height,
   *data_in = new_data;
   fpga_free(temp);
 }
-
+void convert_to_fp16(float **data_in, int data_size) {
+  float *tmp = *data_in;
+  // half_float::half *tmp_data = (half_float::half *)fpga_malloc(data_size *
+  // sizeof(half_float::half));
+  int16_t *tmp_data =
+      (int16_t *)fpga_malloc(data_size * sizeof(int16_t));  // NOLINT
+  for (int i = 0; i < data_size; i++) {
+    // tmp_data[i] = (half_float::half)((*data_in)[i]);
+    tmp_data[i] = fp32_2_fp16((*data_in)[i]);
+  }
+  *data_in = (float *)tmp_data;  // NOLINT
+  fpga_free(tmp);
+}
 void format_filter(float **data_in, int num, int channel, int height, int width,
                    int group_num, float max) {
   convert_to_hwc(data_in, num, channel, height, width);
   align_filter(data_in, num, channel, height, width);
   int pixel_num = calc_aligned_total_pixel_num(num, channel, height, width);
+  convert_to_fp16(data_in, pixel_num);
   fpga_flush(*data_in, pixel_num * sizeof(float));
 }
 
@@ -115,6 +127,7 @@ void format_fc_filter(float **data_in, int num, int channel, int height,
   convert_fc_filter(data_in, num, chw);
   align_filter(data_in, num, channel, height, width);
   int pixel_num = calc_aligned_total_pixel_num(num, channel, height, width);
+  convert_to_fp16(data_in, pixel_num);
   fpga_flush(*data_in, pixel_num * sizeof(float));
 }
 

src/fpga/common/fpga_common.cpp

@@ -113,6 +113,12 @@ int fpga_invalidate(void *address, size_t size) {
   return 0;
 #endif
 }
-
+uint64_t vaddr_to_paddr(void *address) {
+#ifdef PADDLE_MOBILE_ZU5
+  return driver::vaddr_to_paddr(address);
+#else
+  return 0;
+#endif
+}
 }  // namespace fpga
 }  // namespace paddle_mobile

src/operators/kernel/arm/quantize_kernel.cpp

@@ -197,21 +197,21 @@ void QuantizeKernel<CPU, float>::Compute(const QuantizeParam<CPU> &param) {
   }
   max_abs = std::max(max_abs, 1e-6f);
   param.online_scale_->mutable_data<float>()[0] = max_abs;
-  //  switch (param.round_type_) {
-  //    case ROUND_NEAREST_TO_EVEN:
-  //      Quantize<ROUND_NEAREST_TO_EVEN>(input, scale, output);
-  //      break;
-  //    case ROUND_NEAREST_TOWARDS_ZERO:
-  //      Quantize<ROUND_NEAREST_TOWARDS_ZERO>(input, scale, output);
-  //      break;
-  //    case ROUND_NEAREST_AWAY_ZERO:
-  //      Quantize<ROUND_NEAREST_AWAY_ZERO>(input, scale, output);
-  //      break;
-  //    default:
-  //      LOG(kLOG_ERROR) << "round type is not supported.";
-  //      break;
-  //  }
+  switch (param.round_type_) {
+    case ROUND_NEAREST_TO_EVEN:
+      Quantize<ROUND_NEAREST_TO_EVEN>(input, max_abs, param.offline_, output);
+      break;
+    case ROUND_NEAREST_TOWARDS_ZERO:
+      Quantize<ROUND_NEAREST_TOWARDS_ZERO>(input, max_abs, param.offline_,
+                                           output);
+      break;
+    case ROUND_NEAREST_AWAY_ZERO:
       Quantize<ROUND_NEAREST_AWAY_ZERO>(input, max_abs, param.offline_, output);
+      break;
+    default:
+      LOG(kLOG_ERROR) << "round type is not supported.";
+      break;
+  }
 }
 
 }  // namespace operators

src/operators/kernel/fpga/V2/feed_kernel.cpp

@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "operators/kernel/feed_kernel.h"
-
+#include "fpga/V2/filter.h"
 namespace paddle_mobile {
 namespace operators {
 
@@ -24,7 +24,6 @@ bool FeedKernel<FPGA, float>::Init(FeedParam<FPGA> *param) {
   fpga::format_fp16_ofm(output, aligned_channel);
   return true;
 }
-
 template <>
 void FeedKernel<FPGA, float>::Compute(const FeedParam<FPGA> &param) {
   auto input =
@@ -33,6 +32,9 @@ void FeedKernel<FPGA, float>::Compute(const FeedParam<FPGA> &param) {
   auto input_ptr = input->data<float>();
   Tensor *output = param.Out();
   auto output_ptr = output->data<float>();
+  auto channel = input->dims()[1];
+  uint32_t aligned_channels =
+      fpga::filter::calc_aligned_channel((int)channel);  // NOLINT
 
   fpga::BypassArgs args = {fpga::DATA_TYPE_FP32};
 
@@ -41,7 +43,7 @@ void FeedKernel<FPGA, float>::Compute(const FeedParam<FPGA> &param) {
   args.input_layout_type = fpga::LAYOUT_CHW;
   args.output_layout_type = fpga::LAYOUT_HWC;
   args.image.address = reinterpret_cast<void *>(input_ptr);
-  args.image.channels = (uint32_t)input->dims()[1];
+  args.image.channels = aligned_channels;
   args.image.height = (uint32_t)input->dims()[2];
   args.image.width = (uint32_t)input->dims()[3];
   args.image.pad_height = 0;

src/operators/kernel/fpga/V2/softmax_kernel.cpp

@@ -25,7 +25,7 @@ bool SoftmaxKernel<FPGA, float>::Init(SoftmaxParam<FPGA> *param) {
   auto input_ptr = input->data<float>();
   auto float_input = new Tensor;
   float_input->mutable_data<float>({1, input->dims()[1]});
-  fpga::format_fp32_ofm(float_input, 8);
+  fpga::format_fp32_ofm(float_input, 1024);
 
   fpga::BypassArgs args = {fpga::DATA_TYPE_FP16};
   args.input_layout_type = fpga::LAYOUT_HWC;