Merge branch 'develop' of https://github.com/PaddlePaddle/paddle-mobile into develop

# Conflicts:
#	src/operators/op_param.h
#	tools/op.cmake

doc/design_doc.md

+# paddle-mobile 设计文档
+
+
+#### 以下是 paddle-mobile 代码的执行流程图:
+
+![执行流程图](./images/flow_chart.png "执行流程图")
+
+
+
+#### 主要分为: Loader 模块、 Program 模块、 Executor 模块、 op 模块、 kernel 模块、scope variable Tensor 模块
+
+#### 下面展开说一下各个模块的作用以及设计思路
+
+### 一. Loader
+先来看一下模型, 模型分为两种结构:
+ 一种为参数文件是散开的, 如下图, 红框为模型结构的 protobuf 文件, 其余为参数文件
+
+![模型描述](./images/model_desc.png "模型描述")
+
+另一种为参数文件结合在一起的, 如下图, 红框内为模型结构描述的 protobuf 文件, 另一个文件为结合在一起的参数文件
+
+![模型描述combined](./images/model_desc_combined.png "模型描述combined")
+
+loader 模块的作用是将模型结构信息 load 进内存, 将红框内的 protobuf 文件 load 进内存, 并对模型结构进行优化(如将几个细粒度的 op 融合成 粗粒度的 op, 如将 conv、 add、 batchnorm、 relu 融合为 conv\_add\_batchnorm\_relu).
+方便进行算法优化.
+
+__那么为什么融合在一起能够做算法优化 ?__
+
+如果未融合的 conv add batchnorm relu 运算是这样的
+
+```
+[n]
+[conv_res] = conv([n])
+
+for &res in conv_res {
+	res = add_biase(res)
+}
+
+for &res in conv_res {
+	res = batchnorm(res)
+}
+
+for &res in conv_res {
+	res = relu(res)
+}
+
+```
+融合后的 conv\_add\_batchnorm\_relu 运算是这样的:
+
+```
+[n]
+[conv_res] = conv([n])
+
+for &res in conv_res {
+	res = relu(batchnorm(add_biase(res)))
+}
+
+```
+由于 conv 可以转换为两个大矩阵相乘, 更进一步可以分为若干个一行一列的小矩阵相乘, 那最终的运算是这样的:
+
+```
+[n]
+for &res in [res] {
+	res = relu(batchnorm(add_biase(A * B)))
+}
+
+其中 A 和 B 为 1 * k 和 k * 1 矩阵
+
+```
+
+
+
+### 二. Program
+
+program 为 loader 模块的结果, 包含了优化前的模型结构对象, 以及优化后的模型结构对象, 此模块基本对应着 paddle 模型的结构, 关于paddle 模型的一些概念的定义, 详细设计可以参考 [program.md](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md), 以下是一个简单的概况: 
+
+* programDesc 中包含着若干个(googlenet mobilenet yolo squeezenet resnet 常见的模型只有一个)可以嵌套的 block, blocks中的第一个block中的某个 op 可能会执行 blocks 中后边 block 中的一系列 op 运算(只有多个block才会有此概念)
+* block 包含着 ops 和 vars
+* ops 为一系列 op 的描述, 描述着每个 op 的类型, 输入输出, 所需参数
+* vars 里包含的为所有 op 运算所需的参数描述
+
+### 三. Executor
+
+executor 主要是用于 op 运算的上层调度操作, 主要有两个操作,  executor 实例化 和 暴露给上层的 predict 方法
+
+* executor 实例化过程中, 主要进行了这几个操作 
+	1. 根据 loader 产出的 program 初始化 operator 对象 
+	2. 分配所有需要用到的内存, 包括每个op 的输入输出, 权重参数, 目前模型的权重参数文件的内存格式为 NCHW, op 的输入输出中间矩阵参数也是 NCHW 格式
+	3. 调用每个 op 的 init 方法, init 方法是每个 op 实现者进行参数预处理的地方, 有助于减少 predict 的耗时
+
+* predict, 主要用于拿到外部的输入, 顺序调用 op 的 run 方法进行运算, 并返回最终的结果.
+
+
+### 四. op
+关于 op 模块代码的详细设计可以参考 [operator部分代码设计](https://github.com/PaddlePaddle/paddle-mobile/issues/300), operator主要包含一个kernel用于运算、一个 param 用于存储属性, operator 主要有三个操作, Init、RunImp、InferShape
+
+* Init: Init 函数主要用于参数预处理, 如对 batchNorm 参数进行预处理, 可以将 batchNorm 运算转化为 a * x + b 形式的运算, 这个函数也会调用, kernel 的 Init 函数对 kernel 进行初始化
+* RunImp: RunImp 函数会调用自己的kernel 的 compute 方法进行运算
+* InferShape: InferShape 函数会根据输入和参数得出输出的形状, 这个函数会在 executor 实例化时, 内存初始化前调用
+
+每个 operator 都需要进行注册才可以被使用, 以 conv 为例, 需在 conv_op.cpp 底部这样写: 
+
+```c++
+// 三个平台都注册了 conv op
+namespace ops = paddle_mobile::operators;
+#ifdef PADDLE_MOBILE_CPU
+USE_OP_CPU(conv2d);
+REGISTER_OPERATOR_CPU(conv2d, ops::ConvOp);
+#endif
+
+#ifdef PADDLE_MOBILE_MALI_GPU
+USE_OP_MALI_GPU(conv2d);
+REGISTER_OPERATOR_MALI_GPU(conv2d, ops::ConvOp);
+#endif
+
+#ifdef PADDLE_MOBILE_FPGA
+USE_OP_FPGA(conv2d);
+REGISTER_OPERATOR_FPGA(conv2d, ops::ConvOp);
+#endif
+
+```
+
+__一个关于包大小的优化__:
+
+每个 operator 都由一个宏控制编译, 如 conv_op.h(除了 conv_op.h ,  conv_op.cpp、conv_kernle.h、conv_kernle.cpp 也都需要加此宏控制)
+
+```c++
+
+#ifdef CONV_OP    //这个宏控制着 conv_op 是否被编译, 除了 conv_op.h ,  conv_op.cpp、conv_kernle.h conv_kernle.cpp 也都需要加此宏控制
+
+#pragma once
+
+#include <string>
+#include "framework/operator.h"
+#include "operators/kernel/conv_kernel.h"
+
+namespace paddle_mobile {
+namespace operators {
+using std::string;
+template <typename DeviceType, typename T>
+class ConvOp
+	//impl  
+};
+
+}  // namespace operators
+}  // namespace paddle_mobile
+
+#endif
+
+```
+这样做的目的是为了根据不同类型的网络编译特定的op, 在 cmake 中已经配置好不同网络编译的宏, 如果你要进行编译支持 yolo 的模型, 仅需执行:
+
+```sh
+cd toools
+sh build.sh android yolo
+
+```
+这样只会编译 yolo 所包含的四种 op, 极大的减小了包体积和编译时间
+
+### 五. kernel
+kernel 为 op 的底层运算实现, 主要有两个函数, Init 和 Compute, 分别用来初始化、预处理 和 运算操作, 值得提出的是, kernel 会根据泛型特化到不同的平台, 如图所示:
+
+![设备特化](./images/devices.png "设备特化")
+
+不同平台的 kernel 实现, 为同一个 kernel 类不同泛型的特化实现, 目前有三个平台, arm、mali、fpga, 图中的 central-arm-func\ 目录为 op kernel 的 arm 实现, 它承担了 arm\ 目录下 kernel 的底层实现, 同时 arm 处理器作为中央处理器, central-arm-func\ 也可以作为其他协处理器的底层实现, 如: fpga 的某一个 op kernel 还没有 fpga 协处理器的实现, 就可以直接调用使用这里的 arm 实现.
+
+__如果你有兴趣新增一个协处理器实现, 就可以在次添加一个 kernel 目录, 提供协处理器实现, 如果某个 kernel 你没有实现完, 你也可以直接使用 arm 实现__
+
+### 六. scope variable Tensor
+* scope 用来存储管理所需用到的所有 variable(用来存储不同类型的对象, 主要是矩阵Tensor, 也就是说 scpoe 管理着 op 运算过程中所有参数矩阵, 输入输出矩阵), 可以将 scope 理解为一个 map, 这里在 map 上封了一层 scope 的概念是为了方便内存管理
+* variable 可以用来存储不同类型的对象, paddle-mobile 里主要用它来存储矩阵 Tensor
+* tensor 代表着矩阵, 通过泛型可以用来存储不同类型的矩阵, 但需要注意的是, 存入和取出时的类型必须保持一致, 如果类型不一致,  使用 inline const T \*data() const 获取指针会不能通过类型检查, 通过  inline T \*mutable_data() 获取指针会重新分配内存, 以下是关于 Tensor 的一些小概念:
+	1. DDim: 用来存储矩阵的维度信息.
+	2. Slice(): 这个函数用来获取 N 维 (NCHW中的 N) 上切片
+	3. 当实例化未分配内存时, 调用 inline T *mutable_data() 会分配内存
+
+
+
+
+
+

doc/development_doc.md

+# iOS开发文档
+
+## 编译
+
+### 一. 使用 build.sh 编译
+
+```sh 
+sh build.sh ios
+
+# 如果只想编译某个特定模型的 op, 则需执行以下命令
+sh build.sh ios googlenet
+
+# 在这个文件夹下, 你可以拿到生成的 .a 库
+cd ../build/release/ios/build
+
+```
+
+### 二. 使用 xcode 编译
+
+我们提供了 ios 开发更为熟悉的 xcode 编译环境:
+在 ios/ 目录下打开 PaddleMobile.xcworkspace 即可编译 PaddleMobile 或者 运行 Demo
+
+### 三. 集成
+
+#### 如使用 c++ 接口
+将 
+
+```
+libpaddle-mobile.a 
+io.h  
+program.h 
+types.h 
+lod_tensor.h 
+tensor.h
+```
+拖入工程, io.h 为接口文件, 可在 [github](https://github.com/PaddlePaddle/paddle-mobile/blob/develop/src/io/io.h)上查看接口注释
+
+#### 如使用 oc 接口
+将在xcode 编译生成的
+```
+libPaddleMobile.a 
+PaddleMobile.h
+```
+拖入工程, 接口如下:
+
+```
+/*
+	创建单例对象
+*/
++ (instancetype)sharedInstance;
+
+/*
+	load 模型, 开辟内存
+*/
+- (BOOL)load:(NSString *)modelPath andWeightsPath:(NSString *)weighsPath;
+
+/*
+	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 means:(NSArray<NSNumber *> *)means scale:(float)scale;
+
+/*
+	进行预测
+*/
+- (NSArray *)predict:(CGImageRef)image;
+
+/*
+	清理内存
+*/
+- (void)clear;
+
+```
+
+
+
+
+
+
+
+
+
+
+
+
+

ios/PaddleMobile.xcworkspace/contents.xcworkspacedata

+<?xml version="1.0" encoding="UTF-8"?>
+<Workspace
+   version = "1.0">
+   <FileRef
+      location = "group:PaddleMobileDemo/PaddleMobileDemo.xcodeproj">
+   </FileRef>
+   <FileRef
+      location = "group:PaddleMobile/PaddleMobile.xcodeproj">
+   </FileRef>
+</Workspace>

ios/PaddleMobile.xcworkspace/xcshareddata/IDEWorkspaceChecks.plist

+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+	<key>IDEDidComputeMac32BitWarning</key>
+	<true/>
+</dict>
+</plist>

ios/PaddleMobile.xcworkspace/xcuserdata/liuruilong.xcuserdatad/xcdebugger/Breakpoints_v2.xcbkptlist

+<?xml version="1.0" encoding="UTF-8"?>
+<Bucket
+   type = "0"
+   version = "2.0">
+</Bucket>

ios/PaddleMobile/PaddleMobile.xcodeproj/project.xcworkspace/contents.xcworkspacedata

+<?xml version="1.0" encoding="UTF-8"?>
+<Workspace
+   version = "1.0">
+   <FileRef
+      location = "self:PaddleMobile.xcodeproj">
+   </FileRef>
+</Workspace>

ios/PaddleMobile/PaddleMobile.xcodeproj/project.xcworkspace/xcshareddata/IDEWorkspaceChecks.plist

+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+	<key>IDEDidComputeMac32BitWarning</key>
+	<true/>
+</dict>
+</plist>

ios/PaddleMobile/PaddleMobile.xcodeproj/xcuserdata/liuruilong.xcuserdatad/xcschemes/xcschememanagement.plist

+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+	<key>SchemeUserState</key>
+	<dict>
+		<key>PaddleMobile.xcscheme</key>
+		<dict>
+			<key>orderHint</key>
+			<integer>1</integer>
+		</dict>
+	</dict>
+</dict>
+</plist>

ios/PaddleMobile/PaddleMobile/MacroDefine.h

+//
+//  MacroDefine.h
+//  PaddleMobile
+//
+//  Created by liuRuiLong on 2018/6/30.
+//  Copyright © 2018年 orange. All rights reserved.
+//
+
+#ifndef MacroDefine_h
+#define MacroDefine_h
+
+
+#endif /* MacroDefine_h */

ios/PaddleMobile/PaddleMobile/PaddleMobile.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 <CoreImage/CoreImage.h>
+#import <Foundation/Foundation.h>
+
+@interface PaddleMobile : NSObject
+
++ (instancetype)sharedInstance;
+- (BOOL)load:(NSString *)modelPath andWeightsPath:(NSString *)weighsPath;
+- (NSArray *)predict:(CGImageRef)image means:(NSArray<NSNumber *> *)means scale:(float)scale;
+- (NSArray *)predict:(CGImageRef)image;
+- (void)clear;
+
+@end

ios/PaddleMobile/PaddleMobile/PaddleMobile.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 "PaddleMobile.h"
+
+@implementation PaddleMobile
+
++ (instancetype)sharedInstance{
+  //TODO: imp
+  exit(0);
+}
+
+- (BOOL)load:(NSString *)modelPath andWeightsPath:(NSString *)weighsPath{
+  //TODO: imp
+  exit(0);
+}
+
+- (NSArray *)predict:(CGImageRef)image means:(NSArray<NSNumber *> *)means scale:(float)scale{
+  //TODO: imp
+  exit(0);
+}
+
+- (NSArray *)predict:(CGImageRef)image{
+  //TODO: imp
+  exit(0);
+}
+
+- (void)clear{
+  //TODO: imp
+  exit(0);
+}
+@end

ios/PaddleMobileDemo/PaddleMobileDemo.xcodeproj/project.pbxproj

+// !$*UTF8*$!
+{
+	archiveVersion = 1;
+	classes = {
+	};
+	objectVersion = 50;
+	objects = {
+
+/* Begin PBXBuildFile section */
+		FC086BC920E783AF00D85EF7 /* AppDelegate.m in Sources */ = {isa = PBXBuildFile; fileRef = FC086BC820E783AF00D85EF7 /* AppDelegate.m */; };
+		FC086BCC20E783AF00D85EF7 /* ViewController.m in Sources */ = {isa = PBXBuildFile; fileRef = FC086BCB20E783AF00D85EF7 /* ViewController.m */; };
+		FC086BCF20E783AF00D85EF7 /* Main.storyboard in Resources */ = {isa = PBXBuildFile; fileRef = FC086BCD20E783AF00D85EF7 /* Main.storyboard */; };
+		FC086BD120E783B100D85EF7 /* Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = FC086BD020E783B100D85EF7 /* Assets.xcassets */; };
+		FC086BD420E783B100D85EF7 /* LaunchScreen.storyboard in Resources */ = {isa = PBXBuildFile; fileRef = FC086BD220E783B100D85EF7 /* LaunchScreen.storyboard */; };
+		FC086BD720E783B100D85EF7 /* main.m in Sources */ = {isa = PBXBuildFile; fileRef = FC086BD620E783B100D85EF7 /* main.m */; };
+/* End PBXBuildFile section */
+
+/* Begin PBXFileReference section */
+		FC086BC420E783AF00D85EF7 /* PaddleMobileDemo.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = PaddleMobileDemo.app; sourceTree = BUILT_PRODUCTS_DIR; };
+		FC086BC720E783AF00D85EF7 /* AppDelegate.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = AppDelegate.h; sourceTree = "<group>"; };
+		FC086BC820E783AF00D85EF7 /* AppDelegate.m */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.objc; path = AppDelegate.m; sourceTree = "<group>"; };
+		FC086BCA20E783AF00D85EF7 /* ViewController.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = ViewController.h; sourceTree = "<group>"; };
+		FC086BCB20E783AF00D85EF7 /* ViewController.m */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.objc; path = ViewController.m; sourceTree = "<group>"; };
+		FC086BCE20E783AF00D85EF7 /* Base */ = {isa = PBXFileReference; lastKnownFileType = file.storyboard; name = Base; path = Base.lproj/Main.storyboard; sourceTree = "<group>"; };
+		FC086BD020E783B100D85EF7 /* Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = Assets.xcassets; sourceTree = "<group>"; };
+		FC086BD320E783B100D85EF7 /* Base */ = {isa = PBXFileReference; lastKnownFileType = file.storyboard; name = Base; path = Base.lproj/LaunchScreen.storyboard; sourceTree = "<group>"; };
+		FC086BD520E783B100D85EF7 /* Info.plist */ = {isa = PBXFileReference; lastKnownFileType = text.plist.xml; path = Info.plist; sourceTree = "<group>"; };
+		FC086BD620E783B100D85EF7 /* main.m */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.objc; path = main.m; sourceTree = "<group>"; };
+/* End PBXFileReference section */
+
+/* Begin PBXFrameworksBuildPhase section */
+		FC086BC120E783AF00D85EF7 /* Frameworks */ = {
+			isa = PBXFrameworksBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+/* End PBXFrameworksBuildPhase section */
+
+/* Begin PBXGroup section */
+		FC086BBB20E783AF00D85EF7 = {
+			isa = PBXGroup;
+			children = (
+				FC086BC620E783AF00D85EF7 /* PaddleMobileDemo */,
+				FC086BC520E783AF00D85EF7 /* Products */,
+			);
+			sourceTree = "<group>";
+		};
+		FC086BC520E783AF00D85EF7 /* Products */ = {
+			isa = PBXGroup;
+			children = (
+				FC086BC420E783AF00D85EF7 /* PaddleMobileDemo.app */,
+			);
+			name = Products;
+			sourceTree = "<group>";
+		};
+		FC086BC620E783AF00D85EF7 /* PaddleMobileDemo */ = {
+			isa = PBXGroup;
+			children = (
+				FC086BC720E783AF00D85EF7 /* AppDelegate.h */,
+				FC086BC820E783AF00D85EF7 /* AppDelegate.m */,
+				FC086BCA20E783AF00D85EF7 /* ViewController.h */,
+				FC086BCB20E783AF00D85EF7 /* ViewController.m */,
+				FC086BCD20E783AF00D85EF7 /* Main.storyboard */,
+				FC086BD020E783B100D85EF7 /* Assets.xcassets */,
+				FC086BD220E783B100D85EF7 /* LaunchScreen.storyboard */,
+				FC086BD520E783B100D85EF7 /* Info.plist */,
+				FC086BD620E783B100D85EF7 /* main.m */,
+			);
+			path = PaddleMobileDemo;
+			sourceTree = "<group>";
+		};
+/* End PBXGroup section */
+
+/* Begin PBXNativeTarget section */
+		FC086BC320E783AF00D85EF7 /* PaddleMobileDemo */ = {
+			isa = PBXNativeTarget;
+			buildConfigurationList = FC086BDA20E783B100D85EF7 /* Build configuration list for PBXNativeTarget "PaddleMobileDemo" */;
+			buildPhases = (
+				FC086BC020E783AF00D85EF7 /* Sources */,
+				FC086BC120E783AF00D85EF7 /* Frameworks */,
+				FC086BC220E783AF00D85EF7 /* Resources */,
+			);
+			buildRules = (
+			);
+			dependencies = (
+			);
+			name = PaddleMobileDemo;
+			productName = PaddleMobileDemo;
+			productReference = FC086BC420E783AF00D85EF7 /* PaddleMobileDemo.app */;
+			productType = "com.apple.product-type.application";
+		};
+/* End PBXNativeTarget section */
+
+/* Begin PBXProject section */
+		FC086BBC20E783AF00D85EF7 /* Project object */ = {
+			isa = PBXProject;
+			attributes = {
+				LastUpgradeCheck = 0930;
+				ORGANIZATIONNAME = orange;
+				TargetAttributes = {
+					FC086BC320E783AF00D85EF7 = {
+						CreatedOnToolsVersion = 9.3.1;
+					};
+				};
+			};
+			buildConfigurationList = FC086BBF20E783AF00D85EF7 /* Build configuration list for PBXProject "PaddleMobileDemo" */;
+			compatibilityVersion = "Xcode 9.3";
+			developmentRegion = en;
+			hasScannedForEncodings = 0;
+			knownRegions = (
+				en,
+				Base,
+			);
+			mainGroup = FC086BBB20E783AF00D85EF7;
+			productRefGroup = FC086BC520E783AF00D85EF7 /* Products */;
+			projectDirPath = "";
+			projectRoot = "";
+			targets = (
+				FC086BC320E783AF00D85EF7 /* PaddleMobileDemo */,
+			);
+		};
+/* End PBXProject section */
+
+/* Begin PBXResourcesBuildPhase section */
+		FC086BC220E783AF00D85EF7 /* Resources */ = {
+			isa = PBXResourcesBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+				FC086BD420E783B100D85EF7 /* LaunchScreen.storyboard in Resources */,
+				FC086BD120E783B100D85EF7 /* Assets.xcassets in Resources */,
+				FC086BCF20E783AF00D85EF7 /* Main.storyboard in Resources */,
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+/* End PBXResourcesBuildPhase section */
+
+/* Begin PBXSourcesBuildPhase section */
+		FC086BC020E783AF00D85EF7 /* Sources */ = {
+			isa = PBXSourcesBuildPhase;
+			buildActionMask = 2147483647;
+			files = (
+				FC086BCC20E783AF00D85EF7 /* ViewController.m in Sources */,
+				FC086BD720E783B100D85EF7 /* main.m in Sources */,
+				FC086BC920E783AF00D85EF7 /* AppDelegate.m in Sources */,
+			);
+			runOnlyForDeploymentPostprocessing = 0;
+		};
+/* End PBXSourcesBuildPhase section */
+
+/* Begin PBXVariantGroup section */
+		FC086BCD20E783AF00D85EF7 /* Main.storyboard */ = {
+			isa = PBXVariantGroup;
+			children = (
+				FC086BCE20E783AF00D85EF7 /* Base */,
+			);
+			name = Main.storyboard;
+			sourceTree = "<group>";
+		};
+		FC086BD220E783B100D85EF7 /* LaunchScreen.storyboard */ = {
+			isa = PBXVariantGroup;
+			children = (
+				FC086BD320E783B100D85EF7 /* Base */,
+			);
+			name = LaunchScreen.storyboard;
+			sourceTree = "<group>";
+		};
+/* End PBXVariantGroup section */
+
+/* Begin XCBuildConfiguration section */
+		FC086BD820E783B100D85EF7 /* Debug */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				ALWAYS_SEARCH_USER_PATHS = NO;
+				CLANG_ANALYZER_NONNULL = YES;
+				CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE;
+				CLANG_CXX_LANGUAGE_STANDARD = "gnu++14";
+				CLANG_CXX_LIBRARY = "libc++";
+				CLANG_ENABLE_MODULES = YES;
+				CLANG_ENABLE_OBJC_ARC = YES;
+				CLANG_ENABLE_OBJC_WEAK = YES;
+				CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES;
+				CLANG_WARN_BOOL_CONVERSION = YES;
+				CLANG_WARN_COMMA = YES;
+				CLANG_WARN_CONSTANT_CONVERSION = YES;
+				CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES;
+				CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR;
+				CLANG_WARN_DOCUMENTATION_COMMENTS = YES;
+				CLANG_WARN_EMPTY_BODY = YES;
+				CLANG_WARN_ENUM_CONVERSION = YES;
+				CLANG_WARN_INFINITE_RECURSION = YES;
+				CLANG_WARN_INT_CONVERSION = YES;
+				CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES;
+				CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES;
+				CLANG_WARN_OBJC_LITERAL_CONVERSION = YES;
+				CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR;
+				CLANG_WARN_RANGE_LOOP_ANALYSIS = YES;
+				CLANG_WARN_STRICT_PROTOTYPES = YES;
+				CLANG_WARN_SUSPICIOUS_MOVE = YES;
+				CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE;
+				CLANG_WARN_UNREACHABLE_CODE = YES;
+				CLANG_WARN__DUPLICATE_METHOD_MATCH = YES;
+				CODE_SIGN_IDENTITY = "iPhone Developer";
+				COPY_PHASE_STRIP = NO;
+				DEBUG_INFORMATION_FORMAT = dwarf;
+				ENABLE_STRICT_OBJC_MSGSEND = YES;
+				ENABLE_TESTABILITY = YES;
+				GCC_C_LANGUAGE_STANDARD = gnu11;
+				GCC_DYNAMIC_NO_PIC = NO;
+				GCC_NO_COMMON_BLOCKS = YES;
+				GCC_OPTIMIZATION_LEVEL = 0;
+				GCC_PREPROCESSOR_DEFINITIONS = (
+					"DEBUG=1",
+					"$(inherited)",
+				);
+				GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
+				GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR;
+				GCC_WARN_UNDECLARED_SELECTOR = YES;
+				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
+				GCC_WARN_UNUSED_FUNCTION = YES;
+				GCC_WARN_UNUSED_VARIABLE = YES;
+				IPHONEOS_DEPLOYMENT_TARGET = 11.3;
+				MTL_ENABLE_DEBUG_INFO = YES;
+				ONLY_ACTIVE_ARCH = YES;
+				SDKROOT = iphoneos;
+			};
+			name = Debug;
+		};
+		FC086BD920E783B100D85EF7 /* Release */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				ALWAYS_SEARCH_USER_PATHS = NO;
+				CLANG_ANALYZER_NONNULL = YES;
+				CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE;
+				CLANG_CXX_LANGUAGE_STANDARD = "gnu++14";
+				CLANG_CXX_LIBRARY = "libc++";
+				CLANG_ENABLE_MODULES = YES;
+				CLANG_ENABLE_OBJC_ARC = YES;
+				CLANG_ENABLE_OBJC_WEAK = YES;
+				CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES;
+				CLANG_WARN_BOOL_CONVERSION = YES;
+				CLANG_WARN_COMMA = YES;
+				CLANG_WARN_CONSTANT_CONVERSION = YES;
+				CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES;
+				CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR;
+				CLANG_WARN_DOCUMENTATION_COMMENTS = YES;
+				CLANG_WARN_EMPTY_BODY = YES;
+				CLANG_WARN_ENUM_CONVERSION = YES;
+				CLANG_WARN_INFINITE_RECURSION = YES;
+				CLANG_WARN_INT_CONVERSION = YES;
+				CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES;
+				CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES;
+				CLANG_WARN_OBJC_LITERAL_CONVERSION = YES;
+				CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR;
+				CLANG_WARN_RANGE_LOOP_ANALYSIS = YES;
+				CLANG_WARN_STRICT_PROTOTYPES = YES;
+				CLANG_WARN_SUSPICIOUS_MOVE = YES;
+				CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE;
+				CLANG_WARN_UNREACHABLE_CODE = YES;
+				CLANG_WARN__DUPLICATE_METHOD_MATCH = YES;
+				CODE_SIGN_IDENTITY = "iPhone Developer";
+				COPY_PHASE_STRIP = NO;
+				DEBUG_INFORMATION_FORMAT = "dwarf-with-dsym";
+				ENABLE_NS_ASSERTIONS = NO;
+				ENABLE_STRICT_OBJC_MSGSEND = YES;
+				GCC_C_LANGUAGE_STANDARD = gnu11;
+				GCC_NO_COMMON_BLOCKS = YES;
+				GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
+				GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR;
+				GCC_WARN_UNDECLARED_SELECTOR = YES;
+				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
+				GCC_WARN_UNUSED_FUNCTION = YES;
+				GCC_WARN_UNUSED_VARIABLE = YES;
+				IPHONEOS_DEPLOYMENT_TARGET = 11.3;
+				MTL_ENABLE_DEBUG_INFO = NO;
+				SDKROOT = iphoneos;
+				VALIDATE_PRODUCT = YES;
+			};
+			name = Release;
+		};
+		FC086BDB20E783B100D85EF7 /* Debug */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
+				CODE_SIGN_STYLE = Automatic;
+				DEVELOPMENT_TEAM = Z5M2UUN5YV;
+				INFOPLIST_FILE = PaddleMobileDemo/Info.plist;
+				LD_RUNPATH_SEARCH_PATHS = (
+					"$(inherited)",
+					"@executable_path/Frameworks",
+				);
+				PRODUCT_BUNDLE_IDENTIFIER = orange.PaddleMobileDemo;
+				PRODUCT_NAME = "$(TARGET_NAME)";
+				TARGETED_DEVICE_FAMILY = "1,2";
+			};
+			name = Debug;
+		};
+		FC086BDC20E783B100D85EF7 /* Release */ = {
+			isa = XCBuildConfiguration;
+			buildSettings = {
+				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
+				CODE_SIGN_STYLE = Automatic;
+				DEVELOPMENT_TEAM = Z5M2UUN5YV;
+				INFOPLIST_FILE = PaddleMobileDemo/Info.plist;
+				LD_RUNPATH_SEARCH_PATHS = (
+					"$(inherited)",
+					"@executable_path/Frameworks",
+				);
+				PRODUCT_BUNDLE_IDENTIFIER = orange.PaddleMobileDemo;
+				PRODUCT_NAME = "$(TARGET_NAME)";
+				TARGETED_DEVICE_FAMILY = "1,2";
+			};
+			name = Release;
+		};
+/* End XCBuildConfiguration section */
+
+/* Begin XCConfigurationList section */
+		FC086BBF20E783AF00D85EF7 /* Build configuration list for PBXProject "PaddleMobileDemo" */ = {
+			isa = XCConfigurationList;
+			buildConfigurations = (
+				FC086BD820E783B100D85EF7 /* Debug */,
+				FC086BD920E783B100D85EF7 /* Release */,
+			);
+			defaultConfigurationIsVisible = 0;
+			defaultConfigurationName = Release;
+		};
+		FC086BDA20E783B100D85EF7 /* Build configuration list for PBXNativeTarget "PaddleMobileDemo" */ = {
+			isa = XCConfigurationList;
+			buildConfigurations = (
+				FC086BDB20E783B100D85EF7 /* Debug */,
+				FC086BDC20E783B100D85EF7 /* Release */,
+			);
+			defaultConfigurationIsVisible = 0;
+			defaultConfigurationName = Release;
+		};
+/* End XCConfigurationList section */
+	};
+	rootObject = FC086BBC20E783AF00D85EF7 /* Project object */;
+}

ios/PaddleMobileDemo/PaddleMobileDemo.xcodeproj/project.xcworkspace/contents.xcworkspacedata

+<?xml version="1.0" encoding="UTF-8"?>
+<Workspace
+   version = "1.0">
+   <FileRef
+      location = "self:PaddleMobileDemo.xcodeproj">
+   </FileRef>
+</Workspace>

ios/PaddleMobileDemo/PaddleMobileDemo.xcodeproj/project.xcworkspace/xcshareddata/IDEWorkspaceChecks.plist

+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+	<key>IDEDidComputeMac32BitWarning</key>
+	<true/>
+</dict>
+</plist>

ios/PaddleMobileDemo/PaddleMobileDemo.xcodeproj/xcuserdata/liuruilong.xcuserdatad/xcschemes/xcschememanagement.plist

+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+	<key>SchemeUserState</key>
+	<dict>
+		<key>PaddleMobileDemo.xcscheme</key>
+		<dict>
+			<key>orderHint</key>
+			<integer>0</integer>
+		</dict>
+	</dict>
+</dict>
+</plist>

ios/PaddleMobileDemo/PaddleMobileDemo/AppDelegate.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 <UIKit/UIKit.h>
+
+@interface AppDelegate : UIResponder <UIApplicationDelegate>
+
+@property (strong, nonatomic) UIWindow *window;
+
+
+@end
+

ios/PaddleMobileDemo/PaddleMobileDemo/AppDelegate.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 "AppDelegate.h"
+
+@interface AppDelegate ()
+
+@end
+
+@implementation AppDelegate
+
+
+- (BOOL)application:(UIApplication *)application didFinishLaunchingWithOptions:(NSDictionary *)launchOptions {
+    // Override point for customization after application launch.
+    return YES;
+}
+
+
+- (void)applicationWillResignActive:(UIApplication *)application {
+    // Sent when the application is about to move from active to inactive state. This can occur for certain types of temporary interruptions (such as an incoming phone call or SMS message) or when the user quits the application and it begins the transition to the background state.
+    // Use this method to pause ongoing tasks, disable timers, and invalidate graphics rendering callbacks. Games should use this method to pause the game.
+}
+
+
+- (void)applicationDidEnterBackground:(UIApplication *)application {
+    // Use this method to release shared resources, save user data, invalidate timers, and store enough application state information to restore your application to its current state in case it is terminated later.
+    // If your application supports background execution, this method is called instead of applicationWillTerminate: when the user quits.
+}
+
+
+- (void)applicationWillEnterForeground:(UIApplication *)application {
+    // Called as part of the transition from the background to the active state; here you can undo many of the changes made on entering the background.
+}
+
+
+- (void)applicationDidBecomeActive:(UIApplication *)application {
+    // Restart any tasks that were paused (or not yet started) while the application was inactive. If the application was previously in the background, optionally refresh the user interface.
+}
+
+
+- (void)applicationWillTerminate:(UIApplication *)application {
+    // Called when the application is about to terminate. Save data if appropriate. See also applicationDidEnterBackground:.
+}
+
+
+@end

ios/PaddleMobileDemo/PaddleMobileDemo/Assets.xcassets/AppIcon.appiconset/Contents.json

+{
+  "images" : [
+    {
+      "idiom" : "iphone",
+      "size" : "20x20",
+      "scale" : "2x"
+    },
+    {
+      "idiom" : "iphone",
+      "size" : "20x20",
+      "scale" : "3x"
+    },
+    {
+      "idiom" : "iphone",
+      "size" : "29x29",
+      "scale" : "2x"
+    },
+    {
+      "idiom" : "iphone",
+      "size" : "29x29",
+      "scale" : "3x"
+    },
+    {
+      "idiom" : "iphone",
+      "size" : "40x40",
+      "scale" : "2x"
+    },
+    {
+      "idiom" : "iphone",
+      "size" : "40x40",
+      "scale" : "3x"
+    },
+    {
+      "idiom" : "iphone",
+      "size" : "60x60",
+      "scale" : "2x"
+    },
+    {
+      "idiom" : "iphone",
+      "size" : "60x60",
+      "scale" : "3x"
+    },
+    {
+      "idiom" : "ipad",
+      "size" : "20x20",
+      "scale" : "1x"
+    },
+    {
+      "idiom" : "ipad",
+      "size" : "20x20",
+      "scale" : "2x"
+    },
+    {
+      "idiom" : "ipad",
+      "size" : "29x29",
+      "scale" : "1x"
+    },
+    {
+      "idiom" : "ipad",
+      "size" : "29x29",
+      "scale" : "2x"
+    },
+    {
+      "idiom" : "ipad",
+      "size" : "40x40",
+      "scale" : "1x"
+    },
+    {
+      "idiom" : "ipad",
+      "size" : "40x40",
+      "scale" : "2x"
+    },
+    {
+      "idiom" : "ipad",
+      "size" : "76x76",
+      "scale" : "1x"
+    },
+    {
+      "idiom" : "ipad",
+      "size" : "76x76",
+      "scale" : "2x"
+    },
+    {
+      "idiom" : "ipad",
+      "size" : "83.5x83.5",
+      "scale" : "2x"
+    },
+    {
+      "idiom" : "ios-marketing",
+      "size" : "1024x1024",
+      "scale" : "1x"
+    }
+  ],
+  "info" : {
+    "version" : 1,
+    "author" : "xcode"
+  }
+}
\ No newline at end of file

ios/PaddleMobileDemo/PaddleMobileDemo/Assets.xcassets/Contents.json

+{
+  "info" : {
+    "version" : 1,
+    "author" : "xcode"
+  }
+}
\ No newline at end of file

ios/PaddleMobileDemo/PaddleMobileDemo/Base.lproj/LaunchScreen.storyboard

+<?xml version="1.0" encoding="UTF-8" standalone="no"?>
+<document type="com.apple.InterfaceBuilder3.CocoaTouch.Storyboard.XIB" version="3.0" toolsVersion="13122.16" systemVersion="17A277" targetRuntime="iOS.CocoaTouch" propertyAccessControl="none" useAutolayout="YES" launchScreen="YES" useTraitCollections="YES" useSafeAreas="YES" colorMatched="YES" initialViewController="01J-lp-oVM">
+    <dependencies>
+        <plugIn identifier="com.apple.InterfaceBuilder.IBCocoaTouchPlugin" version="13104.12"/>
+        <capability name="Safe area layout guides" minToolsVersion="9.0"/>
+        <capability name="documents saved in the Xcode 8 format" minToolsVersion="8.0"/>
+    </dependencies>
+    <scenes>
+        <!--View Controller-->
+        <scene sceneID="EHf-IW-A2E">
+            <objects>
+                <viewController id="01J-lp-oVM" sceneMemberID="viewController">
+                    <view key="view" contentMode="scaleToFill" id="Ze5-6b-2t3">
+                        <rect key="frame" x="0.0" y="0.0" width="375" height="667"/>
+                        <autoresizingMask key="autoresizingMask" widthSizable="YES" heightSizable="YES"/>
+                        <color key="backgroundColor" red="1" green="1" blue="1" alpha="1" colorSpace="custom" customColorSpace="sRGB"/>
+                        <viewLayoutGuide key="safeArea" id="6Tk-OE-BBY"/>
+                    </view>
+                </viewController>
+                <placeholder placeholderIdentifier="IBFirstResponder" id="iYj-Kq-Ea1" userLabel="First Responder" sceneMemberID="firstResponder"/>
+            </objects>
+            <point key="canvasLocation" x="53" y="375"/>
+        </scene>
+    </scenes>
+</document>

ios/PaddleMobileDemo/PaddleMobileDemo/Base.lproj/Main.storyboard

+<?xml version="1.0" encoding="UTF-8"?>
+<document type="com.apple.InterfaceBuilder3.CocoaTouch.Storyboard.XIB" version="3.0" toolsVersion="13122.16" systemVersion="17A277" targetRuntime="iOS.CocoaTouch" propertyAccessControl="none" useAutolayout="YES" useTraitCollections="YES" useSafeAreas="YES" colorMatched="YES" initialViewController="BYZ-38-t0r">
+    <dependencies>
+        <plugIn identifier="com.apple.InterfaceBuilder.IBCocoaTouchPlugin" version="13104.12"/>
+        <capability name="Safe area layout guides" minToolsVersion="9.0"/>
+        <capability name="documents saved in the Xcode 8 format" minToolsVersion="8.0"/>
+    </dependencies>
+    <scenes>
+        <!--View Controller-->
+        <scene sceneID="tne-QT-ifu">
+            <objects>
+                <viewController id="BYZ-38-t0r" customClass="ViewController" customModuleProvider="" sceneMemberID="viewController">
+                    <view key="view" contentMode="scaleToFill" id="8bC-Xf-vdC">
+                        <rect key="frame" x="0.0" y="0.0" width="375" height="667"/>
+                        <autoresizingMask key="autoresizingMask" widthSizable="YES" heightSizable="YES"/>
+                        <color key="backgroundColor" red="1" green="1" blue="1" alpha="1" colorSpace="custom" customColorSpace="sRGB"/>
+                        <viewLayoutGuide key="safeArea" id="6Tk-OE-BBY"/>
+                    </view>
+                </viewController>
+                <placeholder placeholderIdentifier="IBFirstResponder" id="dkx-z0-nzr" sceneMemberID="firstResponder"/>
+            </objects>
+        </scene>
+    </scenes>
+</document>

ios/PaddleMobileDemo/PaddleMobileDemo/Info.plist

+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+	<key>CFBundleDevelopmentRegion</key>
+	<string>$(DEVELOPMENT_LANGUAGE)</string>
+	<key>CFBundleExecutable</key>
+	<string>$(EXECUTABLE_NAME)</string>
+	<key>CFBundleIdentifier</key>
+	<string>$(PRODUCT_BUNDLE_IDENTIFIER)</string>
+	<key>CFBundleInfoDictionaryVersion</key>
+	<string>6.0</string>
+	<key>CFBundleName</key>
+	<string>$(PRODUCT_NAME)</string>
+	<key>CFBundlePackageType</key>
+	<string>APPL</string>
+	<key>CFBundleShortVersionString</key>
+	<string>1.0</string>
+	<key>CFBundleVersion</key>
+	<string>1</string>
+	<key>LSRequiresIPhoneOS</key>
+	<true/>
+	<key>UILaunchStoryboardName</key>
+	<string>LaunchScreen</string>
+	<key>UIMainStoryboardFile</key>
+	<string>Main</string>
+	<key>UIRequiredDeviceCapabilities</key>
+	<array>
+		<string>armv7</string>
+	</array>
+	<key>UISupportedInterfaceOrientations</key>
+	<array>
+		<string>UIInterfaceOrientationPortrait</string>
+		<string>UIInterfaceOrientationLandscapeLeft</string>
+		<string>UIInterfaceOrientationLandscapeRight</string>
+	</array>
+	<key>UISupportedInterfaceOrientations~ipad</key>
+	<array>
+		<string>UIInterfaceOrientationPortrait</string>
+		<string>UIInterfaceOrientationPortraitUpsideDown</string>
+		<string>UIInterfaceOrientationLandscapeLeft</string>
+		<string>UIInterfaceOrientationLandscapeRight</string>
+	</array>
+</dict>
+</plist>

ios/PaddleMobileDemo/PaddleMobileDemo/ViewController.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 <UIKit/UIKit.h>
+
+@interface ViewController : UIViewController
+
+
+@end
+

ios/PaddleMobileDemo/PaddleMobileDemo/ViewController.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 "ViewController.h"
+
+@interface ViewController ()
+
+@end
+
+@implementation ViewController
+
+- (void)viewDidLoad {
+    [super viewDidLoad];
+}
+
+
+- (void)didReceiveMemoryWarning {
+    [super didReceiveMemoryWarning];
+}
+
+
+@end

ios/PaddleMobileDemo/PaddleMobileDemo/main.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 <UIKit/UIKit.h>
+#import "AppDelegate.h"
+
+int main(int argc, char * argv[]) {
+    @autoreleasepool {
+        return UIApplicationMain(argc, argv, nil, NSStringFromClass([AppDelegate class]));
+    }
+}

src/common/log.h

@@ -174,7 +174,10 @@ struct ToLog;
 struct Print {
   friend struct ToLog;
   template <typename T>
-  Print &operator<<(T const &value) {}
+  Print &operator<<(T const &value) {
+    Print p = Print();
+    return p;
+  }
 
  private:
 };

src/common/types.cpp

@@ -23,8 +23,9 @@ const std::string G_OP_TYPE_BOX_CODER = "box_coder";
 const std::string G_OP_TYPE_CONCAT = "concat";
 const std::string G_OP_TYPE_ELEMENTWISE_ADD = "elementwise_add";
 const std::string G_OP_TYPE_FUSION_CONV_ADD_RELU = "fusion_conv_add_relu";
-const std::string G_OP_TYPE_FC = "fc";
-const std::string G_OP_TYPE_CONV_ADD = "conv_add";
+const std::string G_OP_TYPE_FUSION_CONV_ADD_BN_RELU = "fusion_conv_add_bn_relu";
+const std::string G_OP_TYPE_FC = "fusion_fc";
+const std::string G_OP_TYPE_FUSION_CONV_ADD = "fusion_conv_add";
 const std::string G_OP_TYPE_LRN = "lrn";
 const std::string G_OP_TYPE_MUL = "mul";
 const std::string G_OP_TYPE_MULTICLASS_NMS = "multiclass_nms";
@@ -46,7 +47,7 @@ std::unordered_map<
     std::string, std::pair<std::vector<std::string>, std::vector<std::string>>>
     op_input_output_key = {
         {G_OP_TYPE_CONV, {{"Input"}, {"Output"}}},
-        {G_OP_TYPE_CONV_ADD, {{"Input"}, {"Out"}}},
+        {G_OP_TYPE_FUSION_CONV_ADD, {{"Input"}, {"Out"}}},
         {G_OP_TYPE_RELU, {{"X"}, {"Out"}}},
         {G_OP_TYPE_SOFTMAX, {{"X"}, {"Out"}}},
         {G_OP_TYPE_MUL, {{"X"}, {"Out"}}},
@@ -61,6 +62,7 @@ std::unordered_map<
         {G_OP_TYPE_TRANSPOSE, {{"X"}, {"Out"}}},
         {G_OP_TYPE_BOX_CODER,
          {{"PriorBox", "PriorBoxVar", "TargetBox"}, {"OutputBox"}}},
+        {G_OP_TYPE_FUSION_CONV_ADD_BN_RELU, {{"Input"}, {"Out"}}},
         {G_OP_TYPE_PRIOR_BOX, {{"Image", "Input"}, {"Boxes", "Variances"}}},
         {G_OP_TYPE_MULTICLASS_NMS, {{"BBoxes", "Scores"}, {"Out"}}},
         {G_OP_TYPE_FC, {{"X", "Y", "Z"}, {"Out"}}},

src/common/types.h

@@ -79,7 +79,9 @@ extern const std::string G_OP_TYPE_CONCAT;
 extern const std::string G_OP_TYPE_ELEMENTWISE_ADD;
 extern const std::string G_OP_TYPE_FUSION_CONV_ADD_RELU;
 extern const std::string G_OP_TYPE_FC;
-extern const std::string G_OP_TYPE_CONV_ADD;
+extern const std::string G_OP_TYPE_FUSION_CONV_ADD;
+extern const std::string G_OP_TYPE_FUSION_CONV_ADD_BN_RELU;
+
 extern const std::string G_OP_TYPE_LRN;
 extern const std::string G_OP_TYPE_MUL;
 extern const std::string G_OP_TYPE_MULTICLASS_NMS;

src/common/variant.h

@@ -83,6 +83,7 @@ struct Variant {
       return *const_cast<T *>(reinterpret_cast<const T *>(&data));
     } else {
       PADDLE_MOBILE_THROW_EXCEPTION(" bad cast in variant ");
+      exit(0);
     }
   }
 

src/framework/attribute.h

@@ -129,6 +129,7 @@ class Attribute {
       return vistor(attr.variant_.Get<int64_t>());
     } else {
       PADDLE_MOBILE_THROW_EXCEPTION("type not support");
+      exit(0);
     }
   }
 

src/framework/data_layout.h

@@ -40,6 +40,7 @@ inline DataLayout StringToDataLayout(const std::string &str) {
     return DataLayout::kAnyLayout;
   } else {
     PADDLE_MOBILE_THROW_EXCEPTION("Unknown storage order string: %s", s.c_str())
+    exit(0);
   }
 }
 
@@ -52,6 +53,8 @@ inline std::string DataLayoutToString(const DataLayout &data_layout) {
     case DataLayout::kAnyLayout:
       return "ANY_LAYOUT";
     default:
+      PADDLE_MOBILE_THROW_EXCEPTION("Unknown storage order string ")
+      exit(0);
       break;
   }
 }

src/framework/ddim.h

@@ -58,7 +58,8 @@ struct DDim {
     } else if (d.var.TypeId() == typeid(Dim<9>).hash_code()) {
       return vistor(d.var.Get<Dim<9>>());
     } else {
-      DLOG << " dim not support";
+      PADDLE_MOBILE_ENFORCE(false, " dim not support");
+      exit(0);
     }
   }
 

src/framework/dim.h

@@ -129,6 +129,7 @@ int64_t &indexer(Dim<D> &dim, int idx) {
 template <>
 int64_t &indexer<0>(Dim<0> &dim, int idx) {
   PADDLE_MOBILE_THROW_EXCEPTION("Invalid index")
+  exit(0);
 }
 
 template <int D>
@@ -145,6 +146,7 @@ int64_t indexer(const Dim<D> &dim, int idx) {
 template <>
 int64_t indexer<0>(const Dim<0> &dim, int idx) {
   PADDLE_MOBILE_THROW_EXCEPTION("Invalid index")
+  exit(0);
 }
 
 }  // namespace

src/framework/operator.h

@@ -63,7 +63,7 @@ class OperatorBase {
   std::vector<string> GetOutKeys() const;
   virtual void RunImpl() const = 0;
 
-  virtual void Init() const = 0;
+  virtual void Init() = 0;
   /*
    * @b op 运算所需的输入, 如上一层的输出结果、卷积核
    * */
@@ -117,8 +117,8 @@ class OperatorWithKernel : public OperatorBase<Dtype> {
 
   virtual void InferShape() const = 0;
 
-  void Init() const {
-    PADDLE_MOBILE_ENFORCE(kernel_.Init(param_), "  %s kernel init failed",
+  void Init() {
+    PADDLE_MOBILE_ENFORCE(kernel_.Init(&param_), "  %s kernel init failed",
                           this->type_.c_str());
   }
 
@@ -146,7 +146,7 @@ class OpKernelBase {
   }
 #endif
   virtual void Compute(const P &para) const = 0;
-  virtual bool Init(const P &para) const { return true; };
+  virtual bool Init(P *para) { return true; };
   virtual ~OpKernelBase() = default;
 
  private:

src/framework/program/program-optimize/node.cpp

@@ -93,7 +93,8 @@ int Node::Depth(int begin) {
 
 Node &Node::Folder(
     int size, std::string type,
-    std::map<std::string, std::pair<std::string, std::string>> change,
+    std::map<std::string, std::vector<std::pair<std::string, std::string>>>
+        change,
     std::vector<std::shared_ptr<Node>> *removed_nodes) {
   std::shared_ptr<framework::OpDesc> op_desc =
       std::make_shared<framework::OpDesc>();
@@ -110,12 +111,15 @@ Node &Node::Folder(
 void Node::Folder(
     std::shared_ptr<framework::OpDesc> op_desc,
     std::vector<std::shared_ptr<Node>> *outputs, int index,
-    std::map<std::string, std::pair<std::string, std::string>> *change,
+    std::map<std::string, std::vector<std::pair<std::string, std::string>>>
+        *change,
     Node *begin_node, std::vector<std::shared_ptr<Node>> *removed_nodes) {
   if (change->find(this->type_) != change->end()) {
-    auto change_pair = (*change)[this->type_];
-    op_desc->GetInputs()[change_pair.second] =
-        this->op_desc_->GetInputs()[change_pair.first];
+    auto change_pairs = (*change)[this->type_];
+    for (const auto &change_pair : change_pairs) {
+      op_desc->GetInputs()[change_pair.second] =
+          this->op_desc_->GetInputs()[change_pair.first];
+    }
   }
 
   for (auto &attr_pair : this->op_desc_->attrs_) {

src/framework/program/program-optimize/node.h

@@ -17,6 +17,7 @@ limitations under the License. */
 #include <cinttypes>
 #include <map>
 #include <string>
+#include <utility>
 #include <vector>
 #include "common/log.h"
 #include "framework/program/op_desc.h"
@@ -43,7 +44,8 @@ class Node {
   int Depth(int begin = 0);
   Node &Folder(
       int size, std::string type,
-      std::map<std::string, std::pair<std::string, std::string>> change_map,
+      std::map<std::string, std::vector<std::pair<std::string, std::string>>>
+          change,
       std::vector<std::shared_ptr<Node>> *removed_nodes);
   std::vector<std::shared_ptr<framework::OpDesc>> OpDescs(int size);
   std::shared_ptr<framework::OpDesc> OpDescOfNode() { return op_desc_; }
@@ -56,7 +58,8 @@ class Node {
   void Folder(
       std::shared_ptr<framework::OpDesc> op_desc,
       std::vector<std::shared_ptr<Node>> *outputs, int index,
-      std::map<std::string, std::pair<std::string, std::string>> *change,
+      std::map<std::string, std::vector<std::pair<std::string, std::string>>>
+          *change,
       Node *begin_node, std::vector<std::shared_ptr<Node>> *removed_nodes);
   std::shared_ptr<framework::OpDesc> op_desc_;
 #ifdef PADDLE_MOBILE_DEBUG

src/operators/feed_op.h

@@ -32,7 +32,7 @@ class FeedOp : public framework::OperatorBase<DeviceType> {
         param_(inputs, outputs, attrs, *scope) {}
   void RunImpl() const { param_.Out()->ShareDataWith(*param_.InputX()); }
 
-  void Init() const {}
+  void Init() {}
 
   void InferShape() const {
     auto out_dims = param_.Out()->dims();

src/operators/fetch_op.h

@@ -33,7 +33,7 @@ class FetchOp : public framework::OperatorBase<DeviceType> {
         param_(inputs, outputs, attrs, *scope) {}
   void RunImpl() const { param_.Out()->ShareDataWith(*param_.InputX()); }
 
-  void Init() const {}
+  void Init() {}
 
   void InferShape() const {
     auto x_dims = param_.InputX()->dims();

src/operators/fusion_conv_add.cpp

@@ -50,8 +50,8 @@ template class FusionConvAddOp<CPU, float>;
 
 namespace ops = paddle_mobile::operators;
 #ifdef PADDLE_MOBILE_CPU
-USE_OP_CPU(conv_add);
-REGISTER_OPERATOR_CPU(conv_add, ops::FusionConvAddOp);
+USE_OP_CPU(fusion_conv_add);
+REGISTER_OPERATOR_CPU(fusion_conv_add, ops::FusionConvAddOp);
 #endif
 #ifdef PADDLE_MOBILE_MALI_GPU
 USE_OP_MALI_GPU(conv_add);

src/operators/fusion_conv_add.h

@@ -39,10 +39,10 @@ class FusionConvAddMatcher : public framework::FusionOpMatcher {
     vector<std::shared_ptr<framework::OpDesc>> origin_descs =
         node->OpDescs(node_.Depth());
     node->Folder(node_.Depth(), Type(),
-                 {{G_OP_TYPE_ELEMENTWISE_ADD, {"Y", "Y"}}}, removed_nodes);
+                 {{G_OP_TYPE_ELEMENTWISE_ADD, {{"Y", "Y"}}}}, removed_nodes);
   }
 
-  std::string Type() { return G_OP_TYPE_CONV_ADD; }
+  std::string Type() { return G_OP_TYPE_FUSION_CONV_ADD; }
 };
 
 template <typename DeviceType, typename T>
@@ -67,11 +67,13 @@ class FusionConvAddOp : public framework::OperatorWithKernel<
 };
 
 #ifdef PADDLE_MOBILE_CPU
+
 #ifndef CONV_ADD_REGISTER
 static framework::FusionOpRegistrar convadd_registrar(
     new FusionConvAddMatcher());
 #define CONV_ADD_REGISTER
 #endif
+
 #endif
 
 #ifdef PADDLE_MOBILE_MALI_GPU

src/operators/fusion_conv_add_bn_relu_op.cpp

+/* 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 FUSION_CONVADDBNRELU_OP
+
+#include "operators/fusion_conv_add_bn_relu_op.h"
+#include "operators/math/conv_func.h"
+
+namespace paddle_mobile {
+namespace operators {
+
+template <typename Dtype, typename T>
+void FusionConvAddBNReluOp<Dtype, T>::InferShape() const {
+  auto in_dims = this->param_.Input()->dims();
+  auto filter_dims = this->param_.Filter()->dims();
+  const std::vector<int> &strides = this->param_.Strides();
+  std::vector<int> paddings = this->param_.Paddings();
+  int groups = this->param_.Groups();
+  std::vector<int> dilations = this->param_.Dilations();
+
+  PADDLE_MOBILE_ENFORCE((in_dims.size() == filter_dims.size() &&
+                         dilations.size() == paddings.size() &&
+                         paddings.size() == strides.size()),
+                        "ConvParam is not suitable");
+
+  std::vector<int64_t> output_shape({in_dims[0], filter_dims[0]});
+  for (size_t i = 0; i < strides.size(); ++i) {
+    output_shape.push_back(
+        math::ConvOutputSize(in_dims[i + 2], filter_dims[i + 2], dilations[i],
+                             paddings[i], strides[i]));
+  }
+
+  framework::DDim ddim = framework::make_ddim(output_shape);
+  this->param_.Output()->Resize(ddim);
+}
+template class FusionConvAddBNReluOp<CPU, float>;
+}  // namespace operators
+}  // namespace paddle_mobile
+
+namespace ops = paddle_mobile::operators;
+#ifdef PADDLE_MOBILE_CPU
+USE_OP_CPU(fusion_conv_add_bn_relu);
+REGISTER_OPERATOR_CPU(fusion_conv_add_bn_relu, ops::FusionConvAddBNReluOp);
+#endif
+#ifdef PADDLE_MOBILE_MALI_GPU
+#endif
+#ifdef PADDLE_MOBILE_FPGA
+#endif
+
+#endif

src/operators/fusion_conv_add_bn_relu_op.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. */
+
+#ifdef FUSION_CONVADDBNRELU_OP
+
+#pragma once
+
+#include <string>
+#include <vector>
+#include "framework/operator.h"
+#include "framework/program/program-optimize/fusion_op_register.h"
+#include "op_param.h"
+#include "operators/kernel/conv_add_bn_relu_kernel.h"
+
+namespace paddle_mobile {
+namespace operators {
+using std::string;
+using std::vector;
+class FusionConvAddBNReluMatcher : public framework::FusionOpMatcher {
+ public:
+  FusionConvAddBNReluMatcher() {
+    node_ = framework::Node(G_OP_TYPE_CONV);
+    node_ > std::make_shared<framework::Node>(G_OP_TYPE_ELEMENTWISE_ADD) >
+        std::make_shared<framework::Node>(G_OP_TYPE_BATCHNORM) >
+        std::make_shared<framework::Node>(G_OP_TYPE_RELU);
+  }
+
+  void FolderNodes(
+      framework::Node *node,
+      std::vector<std::shared_ptr<framework::Node>> *removed_nodes) {
+    vector<std::shared_ptr<framework::OpDesc>> origin_descs =
+        node->OpDescs(node_.Depth());
+    node->Folder(node_.Depth(), Type(),
+                 {{G_OP_TYPE_ELEMENTWISE_ADD, {{"Y", "Y"}}},
+                  {G_OP_TYPE_BATCHNORM,
+                   {{"Scale", "Scale"},
+                    {"Mean", "Mean"},
+                    {"Bias", "Bias"},
+                    {"Variance", "Variance"}}}},
+                 removed_nodes);
+  }
+
+  std::string Type() { return G_OP_TYPE_FUSION_CONV_ADD_BN_RELU; }
+};
+
+template <typename DeviceType, typename T>
+class FusionConvAddBNReluOp
+    : public framework::OperatorWithKernel<
+          DeviceType, FusionConvAddBNReluParam,
+          operators::ConvAddBNReluKernel<DeviceType, T>> {
+ public:
+  FusionConvAddBNReluOp(const string &type, const VariableNameMap &inputs,
+                        const VariableNameMap &outputs,
+                        const framework::AttributeMap &attrs,
+                        std::shared_ptr<framework::Scope> scope)
+      : framework::OperatorWithKernel<
+            DeviceType, FusionConvAddBNReluParam,
+            operators::ConvAddBNReluKernel<DeviceType, T>>(
+            type, inputs, outputs, attrs, scope) {}
+
+  using framework::OperatorWithKernel<
+      DeviceType, FusionConvAddBNReluParam,
+      operators::ConvAddBNReluKernel<DeviceType, T>>::OperatorWithKernel;
+  void InferShape() const override;
+
+ protected:
+};
+
+#ifdef PADDLE_MOBILE_CPU
+
+//#ifndef FUSION_CONV_ADD_BN_RELU_REGISTER
+// static framework::FusionOpRegistrar fusion_conv_add_bn_relu_registrar(
+//    new FusionConvAddBNReluMatcher());
+//#define FUSION_CONV_ADD_BN_RELU_REGISTER
+//#endif
+
+#endif
+
+#ifdef PADDLE_MOBILE_MALI_GPU
+
+#ifndef FUSION_CONV_ADD_BN_RELU_REGISTER
+static framework::FusionOpRegistrar fusion_conv_add_bn_relu_registrar(
+    new FusionConvAddBNReluMatcher());
+#define FUSION_CONV_ADD_BN_RELU_REGISTER
+#endif
+
+#endif
+
+#ifdef PADDLE_MOBILE_FPGA
+#endif
+
+}  // namespace operators
+}  // namespace paddle_mobile
+
+#endif

src/operators/fusion_conv_add_relu_op.h

@@ -36,7 +36,7 @@ class FusionConvAddReluOpMatcher : public framework::FusionOpMatcher {
       framework::Node *node,
       std::vector<std::shared_ptr<framework::Node>> *removed_nodes) {
     node->Folder(node_.Depth(), Type(),
-                 {{G_OP_TYPE_ELEMENTWISE_ADD, {"Y", "Y"}}}, removed_nodes);
+                 {{G_OP_TYPE_ELEMENTWISE_ADD, {{"Y", "Y"}}}}, removed_nodes);
   }
   std::string Type() { return G_OP_TYPE_FUSION_CONV_ADD_RELU; }
 };
@@ -65,11 +65,11 @@ class FusionConvAddReluOp : public framework::OperatorWithKernel<
 
 #ifdef PADDLE_MOBILE_CPU
 
-#ifndef CONV_ADD_RELU_REGISTER
-#define CONV_ADD_RELU_REGISTER
+//#ifndef CONV_ADD_RELU_REGISTER
+//#define CONV_ADD_RELU_REGISTER
 // static framework::FusionOpRegistrar fusion_conv_add_relu_registrar(new
 // FusionConvAddReluOpMatcher());
-#endif
+//#endif
 
 #endif
 #ifdef PADDLE_MOBILE_MALI_GPU

src/operators/fusion_fc_op.cpp

@@ -55,8 +55,8 @@ template class FusionFcOp<CPU, float>;
 
 namespace ops = paddle_mobile::operators;
 #ifdef PADDLE_MOBILE_CPU
-USE_OP_CPU(fc);
-REGISTER_OPERATOR_CPU(fc, ops::FusionFcOp);
+USE_OP_CPU(fusion_fc);
+REGISTER_OPERATOR_CPU(fusion_fc, ops::FusionFcOp);
 #endif
 #ifdef PADDLE_MOBILE_MALI_GPU
 USE_OP_MALI_GPU(fc);

src/operators/fusion_fc_op.h

@@ -38,7 +38,7 @@ class FusionFcMatcher : public framework::FusionOpMatcher {
       framework::Node *node,
       std::vector<std::shared_ptr<framework::Node>> *removed_nodes) {
     node->Folder(node_.Depth(), Type(),
-                 {{G_OP_TYPE_ELEMENTWISE_ADD, {"Y", "Z"}}}, removed_nodes);
+                 {{G_OP_TYPE_ELEMENTWISE_ADD, {{"Y", "Z"}}}}, removed_nodes);
   }
 
   std::string Type() { return G_OP_TYPE_FC; }
@@ -66,17 +66,21 @@ class FusionFcOp
 };
 
 #ifdef PADDLE_MOBILE_CPU
+
 #ifndef CONV_CPU_REGISTER
 #define CONV_CPU_REGISTER
 static framework::FusionOpRegistrar fc_registrar(new FusionFcMatcher());
 #endif
+
 #endif
 
 #ifdef PADDLE_MOBILE_MALI_GPU
+
 #ifndef CONV_CPU_REGISTER
 #define CONV_CPU_REGISTER
 static framework::FusionOpRegistrar fc_registrar(new FusionFcMatcher());
 #endif
+
 #endif
 
 #ifdef PADDLE_MOBILE_FPGA

src/operators/kernel/arm/batchnorm_kernel.cpp

@@ -21,7 +21,7 @@ namespace paddle_mobile {
 namespace operators {
 
 template <>
-bool BatchNormKernel<CPU, float>::Init(const BatchNormParam &para) const {
+bool BatchNormKernel<CPU, float>::Init(BatchNormParam *param) {
   return true;
 }
 

src/operators/kernel/arm/box_coder_kernel.cpp

@@ -111,7 +111,7 @@ void DecodeCenterSize(const framework::Tensor& target_box,
 }
 
 template <>
-bool BoxCoderKernel<CPU, float>::Init(const BoxCoderParam& para) const {
+bool BoxCoderKernel<CPU, float>::Init(BoxCoderParam* param) {
   return true;
 }
 

src/operators/kernel/arm/concat_kernel.cpp

@@ -53,7 +53,7 @@ class ConcatFunctor {
 };
 
 template <>
-bool ConcatKernel<CPU, float>::Init(const ConcatParam &para) const {
+bool ConcatKernel<CPU, float>::Init(ConcatParam *param) {
   return true;
 }
 

src/operators/kernel/arm/conv_add_bn_relu_kernel.cpp

+/* 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 FUSION_CONVADDBNRELU_OP
+
+#include "operators/kernel/conv_add_bn_relu_kernel.h"
+#include "operators/kernel/central-arm-func/conv_add_bn_relu_func.h"
+
+namespace paddle_mobile {
+namespace operators {
+
+template <>
+bool ConvAddBNReluKernel<CPU, float>::Init(FusionConvAddBNReluParam *param) {
+  const Tensor *mean = (*param).InputMean();
+  const Tensor *variance = (*param).InputVariance();
+  const Tensor *scale = (*param).InputScale();
+  const Tensor *bias = (*param).InputBias();
+  const float epsilon = (*param).Epsilon();
+
+  auto mean_ptr = mean->data<float>();
+  auto variance_ptr = variance->data<float>();
+  auto scale_ptr = scale->data<float>();
+  auto bias_ptr = bias->data<float>();
+
+  const int C = mean->numel();
+  float inv_std_ptr[C];
+  for (int i = 0; i < C; i++) {
+    inv_std_ptr[i] =
+        1 / static_cast<float>(pow((variance_ptr[i] + epsilon), 0.5));
+  }
+  Tensor *new_scale = new Tensor();
+  Tensor *new_bias = new Tensor();
+  auto new_scale_ptr = new_scale->mutable_data<float>({C});
+  auto new_bias_ptr = new_bias->mutable_data<float>({C});
+  for (int i = 0; i < C; i++) {
+    new_scale_ptr[i] = inv_std_ptr[i] * scale_ptr[i];
+    new_bias_ptr[i] = bias_ptr[i] - mean_ptr[i] * inv_std_ptr[i] * scale_ptr[i];
+  }
+  (*param).SetNewScale(new_scale);
+  (*param).SetNewBias(new_bias);
+  return true;
+}
+
+template <>
+void ConvAddBNReluKernel<CPU, float>::Compute(
+    const FusionConvAddBNReluParam &param) const {
+  ConvAddBNReluCompute<float>(param);
+}
+template class ConvAddBNReluKernel<CPU, float>;
+
+}  // namespace operators
+}  // namespace paddle_mobile
+
+#endif

src/operators/kernel/arm/conv_add_kernel.cpp

@@ -19,7 +19,7 @@ namespace paddle_mobile {
 namespace operators {
 
 template <>
-bool ConvAddKernel<CPU, float>::Init(const FusionConvAddParam &para) const {
+bool ConvAddKernel<CPU, float>::Init(FusionConvAddParam *param) {
   return true;
 }
 

src/operators/kernel/arm/conv_add_relu_kernel.cpp

@@ -21,8 +21,7 @@ namespace paddle_mobile {
 namespace operators {
 
 template <>
-bool ConvAddReluKernel<CPU, float>::Init(
-    const FusionConvAddReluParam &para) const {
+bool ConvAddReluKernel<CPU, float>::Init(FusionConvAddReluParam *param) {
   return true;
 }
 

src/operators/kernel/arm/conv_kernel.cpp

@@ -21,7 +21,7 @@ namespace paddle_mobile {
 namespace operators {
 
 template <>
-bool ConvKernel<CPU, float>::Init(const ConvParam &para) const {
+bool ConvKernel<CPU, float>::Init(ConvParam *param) {
   return true;
 }
 

src/operators/kernel/arm/depthwise_conv_kernel.cpp

@@ -21,7 +21,7 @@ namespace paddle_mobile {
 namespace operators {
 
 template <>
-bool DepthwiseConvKernel<CPU, float>::Init(const ConvParam &para) const {
+bool DepthwiseConvKernel<CPU, float>::Init(ConvParam *param) {
   return true;
 }
 

src/operators/kernel/arm/elementwise_add_kernel.cpp

@@ -27,8 +27,7 @@ struct AddFunctor {
 };
 
 template <>
-bool ElementwiseAddKernel<CPU, float>::Init(
-    const ElementwiseAddParam &para) const {
+bool ElementwiseAddKernel<CPU, float>::Init(ElementwiseAddParam *param) {
   return true;
 }
 

src/operators/kernel/arm/fusion_fc_kernel.cpp

@@ -22,7 +22,7 @@ namespace paddle_mobile {
 namespace operators {
 
 template <>
-bool FusionFcKernel<CPU, float>::Init(const FusionFcParam &para) const {
+bool FusionFcKernel<CPU, float>::Init(FusionFcParam *param) {
   return true;
 }
 

src/operators/kernel/arm/lrn_kernel.cpp

@@ -22,7 +22,7 @@ namespace paddle_mobile {
 namespace operators {
 
 template <>
-bool LrnKernel<CPU, float>::Init(const LrnParam &para) const {
+bool LrnKernel<CPU, float>::Init(LrnParam *param) {
   return true;
 }
 

src/operators/kernel/arm/mul_kernel.cpp

@@ -22,7 +22,7 @@ namespace paddle_mobile {
 namespace operators {
 
 template <>
-bool MulKernel<CPU, float>::Init(const MulParam &para) const {
+bool MulKernel<CPU, float>::Init(MulParam *param) {
   return true;
 }
 

src/operators/kernel/arm/multiclass_nms_kernel.cpp

@@ -204,8 +204,7 @@ void MultiClassOutput(const Tensor& scores, const Tensor& bboxes,
 }
 
 template <>
-bool MultiClassNMSKernel<CPU, float>::Init(
-    const MultiClassNMSParam& para) const {
+bool MultiClassNMSKernel<CPU, float>::Init(MultiClassNMSParam* param) {
   return true;
 }
 

src/operators/kernel/arm/pool_kernel.cpp

@@ -36,7 +36,7 @@ inline void PoolBasic(std::string pooling_type, std::vector<int> ksize,
 }
 
 template <>
-bool PoolKernel<CPU, float>::Init(const PoolParam &para) const {
+bool PoolKernel<CPU, float>::Init(PoolParam *param) {
   return true;
 }
 

src/operators/kernel/arm/prior_box_kernel.cpp

@@ -27,7 +27,7 @@ struct ClipFunctor {
 };
 
 template <>
-bool PriorBoxKernel<CPU, float>::Init(const PriorBoxParam &para) const {
+bool PriorBoxKernel<CPU, float>::Init(PriorBoxParam *param) {
   return true;
 }
 

src/operators/kernel/arm/relu_kernel.cpp

@@ -26,7 +26,7 @@ struct ReluFunctor {
 };
 
 template <>
-bool ReluKernel<CPU, float>::Init(const ReluParam &para) const {
+bool ReluKernel<CPU, float>::Init(ReluParam *param) {
   return true;
 }
 

src/operators/kernel/arm/reshape_kernel.cpp

@@ -20,7 +20,7 @@ namespace paddle_mobile {
 namespace operators {
 
 template <>
-bool ReshapeKernel<CPU, float>::Init(const ReshapeParam &para) const {
+bool ReshapeKernel<CPU, float>::Init(ReshapeParam *param) {
   return true;
 }
 

src/operators/kernel/arm/sigmoid_kernel.cpp

@@ -72,7 +72,7 @@ void sigmoid(const Tensor *X, Tensor *Y) {
 }
 
 template <>
-bool SigmoidKernel<CPU, float>::Init(const SigmoidParam &para) const {
+bool SigmoidKernel<CPU, float>::Init(SigmoidParam *param) {
   return true;
 }
 

src/operators/kernel/arm/softmax_kernel.cpp

@@ -20,7 +20,7 @@ namespace paddle_mobile {
 namespace operators {
 
 template <>
-bool SoftmaxKernel<CPU, float>::Init(const SoftmaxParam &para) const {
+bool SoftmaxKernel<CPU, float>::Init(SoftmaxParam *param) {
   return true;
 }
 

src/operators/kernel/arm/transpose_kernel.cpp

@@ -35,7 +35,7 @@ namespace operators {
 // }
 
 template <>
-bool TransposeKernel<CPU, float>::Init(const TransposeParam& para) const {
+bool TransposeKernel<CPU, float>::Init(TransposeParam* param) {
   return true;
 }
 

src/operators/kernel/batchnorm_kernel.h

@@ -29,7 +29,7 @@ class BatchNormKernel
     : public framework::OpKernelBase<DeviceType, BatchNormParam> {
  public:
   void Compute(const BatchNormParam &param) const;
-  bool Init(const BatchNormParam &para) const;
+  bool Init(BatchNormParam *param);
 };
 
 }  // namespace operators

src/operators/kernel/box_coder_kernel.h

@@ -30,7 +30,7 @@ class BoxCoderKernel
     : public framework::OpKernelBase<DeviceType, BoxCoderParam> {
  public:
   void Compute(const BoxCoderParam& param) const;
-  bool Init(const BoxCoderParam& para) const;
+  bool Init(BoxCoderParam* param);
 };
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/central-arm-func/conv_add_bn_relu_func.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. */
+
+#ifdef FUSION_CONVADDBNRELU_OP
+
+#pragma once
+#include "operators/kernel/conv_add_bn_relu_kernel.h"
+#include "operators/math/depthwise_conv_3x3.h"
+#include "operators/op_param.h"
+namespace paddle_mobile {
+namespace operators {
+
+template <typename P>
+void ConvAddBNReluCompute(const FusionConvAddBNReluParam &param) {
+  const Tensor *input = param.Input();
+  Tensor filter = *param.Filter();
+  Tensor bias = *param.Bias();
+  Tensor new_bias = *param.NewBias();
+  Tensor new_scale = *param.NewScale();
+  auto new_bias_ptr = new_bias.data<float>();
+  auto new_scale_ptr = new_scale.data<float>();
+  int axis = param.Axis();
+  int groups = param.Groups();
+  std::vector<int> strides = param.Strides();
+  std::vector<int> paddings = param.Paddings();
+  std::vector<int> dilations = param.Dilations();
+  Tensor *output = param.Output();
+  std::vector<int64_t> filter_shape_vec(framework::vectorize(filter.dims()));
+
+  if (filter_shape_vec[2] == 3 && strides[0] == 1 && groups > 1) {
+    math::DepthwiseConvAddBNRelu3x3s1p1(input, filter, output, &bias, 1,
+                                        &new_scale, &new_bias, 1, 1);
+  } else {
+    const int batch_size = static_cast<int>(input->dims()[0]);
+
+    math::expand_bias(bias, axis, output->dims());
+    output->ShareDataWith(bias);
+
+    std::vector<int64_t> output_shape_vec(framework::vectorize(output->dims()));
+    size_t data_dim = filter_shape_vec.size() - 2;
+    std::vector<int64_t> col_shape_vec(1 + 2 * data_dim);
+    col_shape_vec[0] = input->dims()[1] / groups;
+    for (size_t j = 0; j < data_dim; ++j) {
+      col_shape_vec[j + 1] = filter_shape_vec[j + 2];
+      col_shape_vec[j + 1 + data_dim] = output_shape_vec[j + 2];
+    }
+    framework::DDim col_shape(framework::make_ddim(col_shape_vec));
+
+    framework::DDim col_matrix_shape =
+        framework::flatten_to_2d(col_shape, data_dim + 1);
+
+    bool is_expand =
+        math::IsExpand(filter_shape_vec, strides, paddings, dilations);
+    Tensor col;
+    Tensor col_matrix;
+    if (is_expand) {
+      col.mutable_data<float>(col_shape);
+      col_matrix.ShareDataWith(col);
+      col_matrix.Resize(col_matrix_shape);
+    }
+
+    framework::DDim input_shape = framework::slice_ddim(
+        input->dims(), 1, static_cast<int>(input->dims().size()));
+
+    framework::DDim filter_matrix_shape = {filter.dims()[0],
+                                           filter.numel() / filter.dims()[0]};
+    filter.Resize(filter_matrix_shape);
+    framework::DDim output_matrix_shape = {
+        output->dims()[1],
+        output->numel() / (output->dims()[0] * output->dims()[1])};
+
+    // convolution operator: im2col(or vol2col) + gemm
+    int in_step = static_cast<int>(input->dims()[1]) / groups;
+    int out_step = static_cast<int>(output->dims()[1]) / groups;
+
+    math::Vol2ColFunctor<CPU, float> vol2col;
+    math::Im2ColFunctor<math::ColFormat::kCFO, CPU, float> im2col;
+
+    for (int i = 0; i < batch_size; i++) {
+      Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape);
+      Tensor out_batch = output->Slice(i, i + 1).Resize(output_matrix_shape);
+
+      for (int g = 0; g < groups; g++) {
+        Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step);
+
+        if (!is_expand) {
+          col.ShareDataWith(in_slice);
+          col_matrix.ShareDataWith(col);
+          col_matrix.Resize(col_matrix_shape);
+        } else if (data_dim == 2U) {
+          // im2col
+          im2col(in_slice, dilations, strides,
+                 std::vector<int>{paddings[0], paddings[1], paddings[0],
+                                  paddings[1]},
+                 &col);
+        } else if (data_dim == 3U) {
+          // vol2col
+          vol2col(in_slice, dilations, strides, paddings, &col);
+        }
+
+        // gemm
+        Tensor out_slice = out_batch.Slice(g * out_step, (g + 1) * out_step);
+        Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step);
+        math::matmul<float>(filter_slice, false, col_matrix, false,
+                            static_cast<float>(1), &out_slice,
+                            static_cast<float>(1), false);
+      }
+    }
+
+    auto output_ptr = output->data<float>();
+    for (int c = 0; c < output_matrix_shape[0]; c++) {
+      int start = c * output_matrix_shape[1];
+      for (int j = 0; j < output_matrix_shape[1]; j++) {
+        output_ptr[start + j] =
+            output_ptr[start + j] * new_scale_ptr[c] + new_bias_ptr[c];
+        output_ptr[start + j] =
+            output_ptr[start + j] < 0 ? 0 : output_ptr[start + j];
+      }
+    }
+  }
+}
+}  // namespace operators
+}  // namespace paddle_mobile
+
+#endif

src/operators/kernel/concat_kernel.h

@@ -27,7 +27,7 @@ template <typename DeviceType, typename T>
 class ConcatKernel : public framework::OpKernelBase<DeviceType, ConcatParam> {
  public:
   void Compute(const ConcatParam &param) const;
-  bool Init(const ConcatParam &para) const;
+  bool Init(ConcatParam *param);
 };
 
 }  // namespace operators

src/operators/math/depthwiseconv3x3s1p1.h → src/operators/kernel/conv_add_bn_relu_kernel.h

@@ -13,15 +13,33 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
-#include "framework/tensor.h"
+
+#ifdef FUSION_CONVADDBNRELU_OP
+
+#include <vector>
+#include "framework/ddim.h"
+#include "framework/operator.h"
+#include "operators/math/conv_func.h"
+#include "operators/math/im2col.h"
+#include "operators/math/math_function.h"
+#include "operators/math/vol2col.h"
+#include "operators/op_param.h"
 
 namespace paddle_mobile {
 namespace operators {
-namespace math {
-using framework::Tensor;
 
-void DepthwiseConv3x3s1p1(const Tensor *input, Tensor filter, Tensor *output,
-                          Tensor bias, bool if_bias);
-}  // namespace math
+using framework::DDim;
+using framework::OpKernelBase;
+
+template <typename DeviceType, typename T>
+class ConvAddBNReluKernel
+    : public OpKernelBase<DeviceType, FusionConvAddBNReluParam> {
+ public:
+  void Compute(const FusionConvAddBNReluParam &param) const;
+  bool Init(FusionConvAddBNReluParam *param);
+};
+
 }  // namespace operators
 }  // namespace paddle_mobile
+
+#endif

src/operators/kernel/conv_add_kernel.h

@@ -40,7 +40,7 @@ template <typename DeviceType, typename T>
 class ConvAddKernel : public OpKernelBase<DeviceType, FusionConvAddParam> {
  public:
   void Compute(const FusionConvAddParam &param) const;
-  bool Init(const FusionConvAddParam &para) const;
+  bool Init(FusionConvAddParam *param);
 };
 
 }  // namespace operators

src/operators/kernel/conv_add_relu_kernel.h

@@ -36,7 +36,7 @@ class ConvAddReluKernel
     : public OpKernelBase<DeviceType, FusionConvAddReluParam> {
  public:
   void Compute(const FusionConvAddReluParam &param) const;
-  bool Init(const FusionConvAddReluParam &para) const;
+  bool Init(FusionConvAddReluParam *param);
 };
 
 }  // namespace operators

src/operators/kernel/conv_kernel.h

@@ -32,7 +32,7 @@ template <typename DeviceType, typename T>
 class ConvKernel : public OpKernelBase<DeviceType, ConvParam> {
  public:
   void Compute(const ConvParam &param) const;
-  bool Init(const ConvParam &para) const;
+  bool Init(ConvParam *param);
 };
 
 }  // namespace operators

src/operators/kernel/depthwise_conv_kernel.h

@@ -31,7 +31,7 @@ template <typename DeviceType, typename T>
 class DepthwiseConvKernel : public OpKernelBase<DeviceType, ConvParam> {
  public:
   void Compute(const ConvParam &param) const;
-  bool Init(const ConvParam &para) const;
+  bool Init(ConvParam *param);
 };
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/elementwise_add_kernel.h

@@ -30,7 +30,7 @@ class ElementwiseAddKernel
     : public framework::OpKernelBase<DeviceType, ElementwiseAddParam> {
  public:
   void Compute(const ElementwiseAddParam &param) const;
-  bool Init(const ElementwiseAddParam &para) const;
+  bool Init(ElementwiseAddParam *param);
 };
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/fpga/conv_kernel.cpp

@@ -20,7 +20,7 @@ namespace paddle_mobile {
 namespace operators {
 
 template <>
-bool ConvKernel<FPGA, float>::Init(const ConvParam &para) const {
+bool ConvKernel<FPGA, float>::Init(ConvParam *param) {
   return true;
 }
 

src/operators/kernel/fusion_fc_kernel.h

@@ -28,7 +28,7 @@ class FusionFcKernel
     : public framework::OpKernelBase<DeviceType, FusionFcParam> {
  public:
   void Compute(const FusionFcParam& param) const;
-  bool Init(const FusionFcParam& para) const;
+  bool Init(FusionFcParam* param);
 };
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/lrn_kernel.h

@@ -170,7 +170,7 @@ template <typename DeviceType, typename T>
 class LrnKernel : public framework::OpKernelBase<DeviceType, LrnParam> {
  public:
   void Compute(const LrnParam &param) const;
-  bool Init(const LrnParam &para) const;
+  bool Init(LrnParam *param);
 };
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/mali/batchnorm_kernel.cpp

@@ -128,7 +128,7 @@ class AclBatchNormOp : public acl::ACLOperator {
 };
 
 template <>
-bool BatchNormKernel<GPU_MALI, float>::Init(const BatchNormParam& param) const {
+bool BatchNormKernel<GPU_MALI, float>::Init(BatchNormParam* param) {
   AclBatchNormOp<GPU_MALI, float>* acl_op =
       reinterpret_cast<AclBatchNormOp<GPU_MALI, float>*>(this->GetAclOp());
   if (acl_op == nullptr) {

src/operators/kernel/mali/conv_kernel.cpp

@@ -195,7 +195,7 @@ class AclConvOp : public acl::ACLOperator {
 };
 
 template <>
-bool ConvKernel<GPU_MALI, float>::Init(const ConvParam& param) const {
+bool ConvKernel<GPU_MALI, float>::Init(ConvParam* param) {
   AclConvOp<GPU_MALI, float>* acl_op =
       reinterpret_cast<AclConvOp<GPU_MALI, float>*>(this->GetAclOp());
   if (acl_op == nullptr) {

src/operators/kernel/mali/elementwise_add_kernel.cpp

@@ -27,8 +27,7 @@ struct AddFunctor {
 };
 
 template <>
-bool ElementwiseAddKernel<GPU_MALI, float>::Init(
-    const ElementwiseAddParam &para) const {
+bool ElementwiseAddKernel<GPU_MALI, float>::Init(ElementwiseAddParam *param) {
   return true;
 }
 

src/operators/kernel/mali/fushion_fc_kernel.cpp

@@ -22,7 +22,7 @@ namespace paddle_mobile {
 namespace operators {
 
 template <>
-bool FusionFcKernel<GPU_MALI, float>::Init(const FusionFcParam &para) const {
+bool FusionFcKernel<GPU_MALI, float>::Init(FusionFcParam *param) {
   return true;
 }
 

src/operators/kernel/mali/mul_kernel.cpp

@@ -22,7 +22,7 @@ namespace paddle_mobile {
 namespace operators {
 
 template <>
-bool MulKernel<GPU_MALI, float>::Init(const MulParam &para) const {
+bool MulKernel<GPU_MALI, float>::Init(MulParam *param) {
   return true;
 }
 

src/operators/kernel/mali/reshape_kernel.cpp

@@ -22,7 +22,7 @@ namespace paddle_mobile {
 namespace operators {
 
 template <>
-bool ReshapeKernel<GPU_MALI, float>::Init(const ReshapeParam &para) const {
+bool ReshapeKernel<GPU_MALI, float>::Init(ReshapeParam *param) {
   return true;
 }
 

src/operators/kernel/mul_kernel.h

@@ -29,7 +29,7 @@ template <typename DeviceType, typename T>
 class MulKernel : public framework::OpKernelBase<DeviceType, MulParam> {
  public:
   void Compute(const MulParam &param) const;
-  bool Init(const MulParam &para) const;
+  bool Init(MulParam *param);
 };
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/multiclass_nms_kernel.h

@@ -28,7 +28,7 @@ class MultiClassNMSKernel
     : public framework::OpKernelBase<DeviceType, MultiClassNMSParam> {
  public:
   void Compute(const MultiClassNMSParam& param) const;
-  bool Init(const MultiClassNMSParam& para) const;
+  bool Init(MultiClassNMSParam* param);
 };
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/pool_kernel.h

@@ -28,7 +28,7 @@ template <typename DeviceType, typename T>
 class PoolKernel : public OpKernelBase<DeviceType, PoolParam> {
  public:
   void Compute(const PoolParam &param) const override;
-  bool Init(const PoolParam &para) const;
+  bool Init(PoolParam *param);
 };
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/prior_box_kernel.h

@@ -55,7 +55,7 @@ class PriorBoxKernel
     : public framework::OpKernelBase<DeviceType, PriorBoxParam> {
  public:
   void Compute(const PriorBoxParam& param) const;
-  bool Init(const PriorBoxParam& para) const;
+  bool Init(PriorBoxParam* param);
 };
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/relu_kernel.h

@@ -27,7 +27,7 @@ template <typename DeviceType, typename T>
 class ReluKernel : public framework::OpKernelBase<DeviceType, ReluParam> {
  public:
   void Compute(const ReluParam& param) const;
-  bool Init(const ReluParam& para) const;
+  bool Init(ReluParam* param);
 };
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/reshape_kernel.h

@@ -71,7 +71,7 @@ template <typename DeviceType, typename T>
 class ReshapeKernel : public framework::OpKernelBase<DeviceType, ReshapeParam> {
  public:
   void Compute(const ReshapeParam& param) const;
-  bool Init(const ReshapeParam& para) const;
+  bool Init(ReshapeParam* param);
 };
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/sigmoid_kernel.h

@@ -26,7 +26,7 @@ template <typename DeviceType, typename T>
 class SigmoidKernel : public OpKernelBase<DeviceType, SigmoidParam> {
  public:
   void Compute(const SigmoidParam& param) const override;
-  bool Init(const SigmoidParam& para) const;
+  bool Init(SigmoidParam* param);
 };
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/softmax_kernel.h

@@ -29,7 +29,7 @@ template <typename DeviceType, typename T>
 class SoftmaxKernel : public OpKernelBase<DeviceType, SoftmaxParam> {
  public:
   void Compute(const SoftmaxParam &param) const override;
-  bool Init(const SoftmaxParam &para) const;
+  bool Init(SoftmaxParam *param);
 };
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/transpose_kernel.h

@@ -29,7 +29,7 @@ class TransposeKernel
     : public framework::OpKernelBase<DeviceType, TransposeParam> {
  public:
   void Compute(const TransposeParam& param) const;
-  bool Init(const TransposeParam& para) const;
+  bool Init(TransposeParam* param);
 };
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/math/depthwise_conv_3x3.cpp

@@ -12,6 +12,7 @@ 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 "operators/math/depthwise_conv_3x3.h"
+#include <arm_neon.h>
 #include <vector>
 
 namespace paddle_mobile {
@@ -501,6 +502,322 @@ void DepthwiseConv3x3s1p1(const Tensor *input, const Tensor *filter,
     }
   }
 }
+void DepthwiseConvAddBNRelu3x3s1p1(const Tensor *input, Tensor filter,
+                                   Tensor *output, Tensor *bias, bool if_bias,
+                                   Tensor *new_scale, Tensor *new_bias,
+                                   bool if_bn, bool if_relu) {
+  const float *input_data = input->data<float>();
+  const float *filter_data = filter.data<float>();
+  float *output_data = output->data<float>();
+  const float *bias_data = bias->data<float>();
+  const float *newscale_data = new_scale->data<float>();
+  const float *newbias_data = new_bias->data<float>();
+
+  const int h = static_cast<int>(input->dims()[2]);
+  const int w = static_cast<int>(input->dims()[3]);
+  const int l = h;
+
+  const int batch_size = static_cast<int>(input->dims()[0]);
+  const int c = static_cast<int>(input->dims()[1]);
+  const int hxw = h * w;
+  float32x4_t vbias = vdupq_n_f32(0.0);
+  float32x4_t vnewbias = vdupq_n_f32(0.0);
+  float32x4_t vnewscale = vdupq_n_f32(1.0);
+  float32x4_t vzero = vdupq_n_f32(0);
+
+  for (int b = 0; b < batch_size; ++b) {
+    const float *filter_data_tmp = filter_data;
+
+    for (int j = 0; j < c; ++j) {
+      if (if_bias) {
+        vbias = vdupq_n_f32(bias_data[j]);
+      }
+      if (if_bn) {
+        vnewbias = vdupq_n_f32(newbias_data[j]);
+        vnewscale = vdupq_n_f32(newscale_data[j]);
+      }
+      int l_mid = l - 2;  // l=1->l_mid=-1,l=2->l_mid=0
+      float w00 = filter_data_tmp[0];
+      float w01 = filter_data_tmp[1];
+      float w02 = filter_data_tmp[2];
+      float w10 = filter_data_tmp[3];
+      float w11 = filter_data_tmp[4];
+      float w12 = filter_data_tmp[5];
+      float w20 = filter_data_tmp[6];
+      float w21 = filter_data_tmp[7];
+      float w22 = filter_data_tmp[8];
+
+      output_data[0] =
+          (w11 * input_data[0] + w12 * input_data[1] + w21 * input_data[l] +
+           w22 * input_data[l + 1] + bias_data[j]) *
+              newscale_data[j] +
+          newbias_data[j];
+      output_data[l - 1] = (w10 * input_data[l - 2] + w11 * input_data[l - 1] +
+                            w20 * input_data[2 * l - 2] +
+                            w21 * input_data[2 * l - 1] + bias_data[j]) *
+                               newscale_data[j] +
+                           newbias_data[j];
+
+      output_data[(l - 1) * l] =
+          (w01 * input_data[(l - 2) * l] + w02 * input_data[(l - 2) * l + 1] +
+           w11 * input_data[(l - 1) * l] + w12 * input_data[(l - 1) * l + 1] +
+           bias_data[j]) *
+              newscale_data[j] +
+          newbias_data[j];
+      output_data[l * l - 1] = (w00 * input_data[(l - 2) * (l + 1)] +
+                                w01 * input_data[(l - 2) * (l + 1) + 1] +
+                                w10 * input_data[l * l - 2] +
+                                w11 * input_data[l * l - 1] + bias_data[j]) *
+                                   newscale_data[j] +
+                               newbias_data[j];
+      if (if_relu) {
+        output_data[0] = output_data[0] < 0 ? 0 : output_data[0];
+        output_data[l - 1] = output_data[l - 1] < 0 ? 0 : output_data[l - 1];
+        output_data[(l - 1) * l] =
+            output_data[(l - 1) * l] < 0 ? 0 : output_data[(l - 1) * l];
+        output_data[l * l - 1] =
+            output_data[l * l - 1] < 0 ? 0 : output_data[l * l - 1];
+      }
+      for (int i = 1; i < l - 1; ++i) {
+        output_data[i * l] =
+            (w01 * input_data[i * l - l] + w02 * input_data[i * l - l + 1] +
+             w11 * input_data[i * l] + w12 * input_data[i * l + 1] +
+             w21 * input_data[i * l + l] + w22 * input_data[i * l + l + 1] +
+             bias_data[j]) *
+                newscale_data[j] +
+            newbias_data[j];
+        output_data[i * l + l - 1] =
+            (w00 * input_data[i * l + l - 1 - l - 1] +
+             w01 * input_data[i * l + l - 1 - l] +
+             w10 * input_data[i * l + l - 1 - 1] +
+             w11 * input_data[i * l + l - 1] +
+             w20 * input_data[i * l + l - 1 + l - 1] +
+             w21 * input_data[i * l + l - 1 + l] + bias_data[j]) *
+                newscale_data[j] +
+            newbias_data[j];
+        if (if_relu) {
+          output_data[i * l] = output_data[i * l] < 0 ? 0 : output_data[i * l];
+          output_data[i * l + l - 1] =
+              output_data[i * l + l - 1] < 0 ? 0 : output_data[i * l + l - 1];
+        }
+      }
+
+      // top 1 row and bottom 1 row
+      const float *input_tmp = input_data;
+
+      float32x4_t in0, in1, in2, in3, in4, in5, in6, in7, tmp0, tmp1, tmp2,
+          tmp3, tmp4, tmp5, out0;
+      in0 = vld1q_f32(input_tmp);
+      in2 = vld1q_f32(input_tmp + l);
+      const float *input_tmp_end = input_tmp + (l - 2) * l;
+      in4 = vld1q_f32(input_tmp_end);
+      in6 = vld1q_f32(input_tmp_end + l);
+      int c_mid = l_mid;
+      auto output_ptr = output_data + 1;
+      for (; c_mid > 3; c_mid -= 4) {
+        in1 = vld1q_f32(input_tmp + 4);
+        in3 = vld1q_f32(input_tmp + l + 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);
+
+        out0 = vmulq_n_f32(in0, w10);
+        out0 = vmlaq_n_f32(out0, tmp0, w11);
+        out0 = vmlaq_n_f32(out0, tmp1, w12);
+        out0 = vmlaq_n_f32(out0, in2, w20);
+        out0 = vmlaq_n_f32(out0, tmp2, w21);
+        out0 = vmlaq_n_f32(out0, tmp3, w22);
+        out0 = vaddq_f32(out0, vbias);
+        out0 = vmlaq_f32(vnewbias, vnewscale, out0);
+        if (if_relu) {
+          out0 = vmaxq_f32(out0, vzero);
+        }
+        vst1q_f32(output_ptr, out0);
+
+        in5 = vld1q_f32(input_tmp_end + 4);
+        in7 = vld1q_f32(input_tmp_end + l + 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);
+
+        out0 = vmulq_n_f32(in4, w00);
+        out0 = vmlaq_n_f32(out0, tmp0, w01);
+        out0 = vmlaq_n_f32(out0, tmp1, w02);
+        out0 = vmlaq_n_f32(out0, in6, w10);
+        out0 = vmlaq_n_f32(out0, tmp2, w11);
+        out0 = vmlaq_n_f32(out0, tmp3, w12);
+        out0 = vaddq_f32(out0, vbias);
+        out0 = vmlaq_f32(vnewbias, vnewscale, out0);
+        if (if_relu) {
+          out0 = vmaxq_f32(out0, vzero);
+        }
+        vst1q_f32(output_ptr + (l - 1) * l, out0);
+
+        // 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 pad
+      float32x4_t pad0 = vdupq_n_f32(input_data[l - 1]);
+      float32x4_t pad1 = vdupq_n_f32(input_data[2 * l - 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);
+
+      out0 = vmulq_n_f32(in0, w10);
+      out0 = vmlaq_n_f32(out0, tmp0, w11);
+      out0 = vmlaq_n_f32(out0, tmp1, w12);
+      out0 = vmlaq_n_f32(out0, in2, w20);
+      out0 = vmlaq_n_f32(out0, tmp2, w21);
+      out0 = vmlaq_n_f32(out0, tmp3, w22);
+      out0 = vaddq_f32(out0, vbias);
+      out0 = vmlaq_f32(vnewbias, vnewscale, out0);
+      if (if_relu) {
+        out0 = vmaxq_f32(out0, vzero);
+      }
+      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 pad
+      float32x4_t pad2 = vdupq_n_f32(input_data[l * l - 1 - l]);
+      float32x4_t pad3 = vdupq_n_f32(input_data[l * l - 1]);
+
+      tmp0 = vextq_f32(in4, pad2, 1);
+      tmp1 = vextq_f32(in4, pad2, 2);
+      tmp2 = vextq_f32(in6, pad3, 1);
+      tmp3 = vextq_f32(in6, pad3, 2);
+
+      out0 = vmulq_n_f32(in4, w00);
+      out0 = vmlaq_n_f32(out0, tmp0, w01);
+      out0 = vmlaq_n_f32(out0, tmp1, w02);
+      out0 = vmlaq_n_f32(out0, in6, w10);
+      out0 = vmlaq_n_f32(out0, tmp2, w11);
+      out0 = vmlaq_n_f32(out0, tmp3, w12);
+      out0 = vaddq_f32(out0, vbias);
+      out0 = vmlaq_f32(vnewbias, vnewscale, out0);
+      if (if_relu) {
+        out0 = vmaxq_f32(out0, vzero);
+      }
+      for (int i = 0; i < c_mid; ++i) {
+        if (i == 0) {
+          vst1q_lane_f32(output_ptr + (l - 1) * l + i, out0, 0);
+        }
+        if (i == 1) {
+          vst1q_lane_f32(output_ptr + (l - 1) * l + i, out0, 1);
+        }
+        if (i == 2) {
+          vst1q_lane_f32(output_ptr + (l - 1) * l + i, out0, 2);
+        }
+      }
+      // mid
+
+      for (int i = 0; i < l - 2; ++i) {
+        auto output_ptr = output_data + (i + 1) * l + 1;
+        input_tmp = input_data + i * l;
+        auto in0_tmp = vld1q_f32(input_tmp);
+        auto in2_tmp = vld1q_f32(input_tmp + l);
+        auto in4_tmp = vld1q_f32(input_tmp + l + l);
+        c_mid = l_mid;
+        for (; c_mid > 3; c_mid -= 4) {
+          auto in1_tmp = vld1q_f32(input_tmp + 4);
+          auto in3_tmp = vld1q_f32(input_tmp + l + 4);
+          auto in5_tmp = vld1q_f32(input_tmp + l + l + 4);
+
+          tmp0 = vextq_f32(in0_tmp, in1_tmp, 1);
+          tmp1 = vextq_f32(in0_tmp, in1_tmp, 2);
+          tmp2 = vextq_f32(in2_tmp, in3_tmp, 1);
+          tmp3 = vextq_f32(in2_tmp, in3_tmp, 2);
+          tmp4 = vextq_f32(in4_tmp, in5_tmp, 1);
+          tmp5 = vextq_f32(in4_tmp, in5_tmp, 2);
+
+          out0 = vmulq_n_f32(in0_tmp, w00);
+          out0 = vmlaq_n_f32(out0, tmp0, w01);
+          out0 = vmlaq_n_f32(out0, tmp1, w02);
+          out0 = vmlaq_n_f32(out0, in2_tmp, w10);
+          out0 = vmlaq_n_f32(out0, tmp2, w11);
+          out0 = vmlaq_n_f32(out0, tmp3, w12);
+          out0 = vmlaq_n_f32(out0, in4_tmp, w20);
+          out0 = vmlaq_n_f32(out0, tmp4, w21);
+          out0 = vmlaq_n_f32(out0, tmp5, w22);
+          out0 = vaddq_f32(out0, vbias);
+          out0 = vmlaq_f32(vnewbias, vnewscale, out0);
+          if (if_relu) {
+            out0 = vmaxq_f32(out0, vzero);
+          }
+          vst1q_f32(output_ptr, out0);
+
+          output_ptr += 4;
+          input_tmp += 4;
+          in0_tmp = in1_tmp;
+          in2_tmp = in3_tmp;
+          in4_tmp = in5_tmp;
+        }
+
+        float32x4_t pad0 = vdupq_n_f32(input_data[i * l + l - 1]);
+        float32x4_t pad1 = vdupq_n_f32(input_data[i * l + l - 1 + l]);
+        float32x4_t pad2 = vdupq_n_f32(input_data[i * l + l - 1 + l + l]);
+
+        tmp0 = vextq_f32(in0_tmp, pad0, 1);
+        tmp1 = vextq_f32(in0_tmp, pad0, 2);
+        tmp2 = vextq_f32(in2_tmp, pad1, 1);
+        tmp3 = vextq_f32(in2_tmp, pad1, 2);
+        tmp4 = vextq_f32(in4_tmp, pad2, 1);
+        tmp5 = vextq_f32(in4_tmp, pad2, 2);
+
+        out0 = vmulq_n_f32(in0_tmp, w00);
+        out0 = vmlaq_n_f32(out0, tmp0, w01);
+        out0 = vmlaq_n_f32(out0, tmp1, w02);
+        out0 = vmlaq_n_f32(out0, in2_tmp, w10);
+        out0 = vmlaq_n_f32(out0, tmp2, w11);
+        out0 = vmlaq_n_f32(out0, tmp3, w12);
+        out0 = vmlaq_n_f32(out0, in4_tmp, w20);
+        out0 = vmlaq_n_f32(out0, tmp4, w21);
+        out0 = vmlaq_n_f32(out0, tmp5, w22);
+        out0 = vaddq_f32(out0, vbias);
+        out0 = vmlaq_f32(vnewbias, vnewscale, out0);
+        if (if_relu) {
+          out0 = vmaxq_f32(out0, vzero);
+        }
+        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);
+          }
+        }
+      }
+      output_data += hxw;
+      input_data += hxw;
+      filter_data_tmp += 9;
+    }
+  }
+}
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/math/depthwise_conv_3x3.h

@@ -32,6 +32,10 @@ void DepthwiseConv3x3(const Tensor *input, vector<int> strides,
                       Tensor *output, bool if_bias);
 void DepthwiseConv3x3s1p1(const Tensor *input, const Tensor *filter,
                           Tensor *output, Tensor *bias, bool if_bias);
+void DepthwiseConvAddBNRelu3x3s1p1(const Tensor *input, Tensor filter,
+                                   Tensor *output, Tensor *bias, bool if_bias,
+                                   Tensor *new_scale, Tensor *new_bias,
+                                   bool if_bn, bool if_relu);
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/math/depthwiseconv3x3s1p1.cpp

-/* 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 "operators/math/depthwiseconv3x3s1p1.h"
-#include <arm_neon.h>
-
-namespace paddle_mobile {
-namespace operators {
-namespace math {
-
-using framework::Tensor;
-
-void DepthwiseConv3x3s1p1(const Tensor *input, Tensor filter, Tensor *output,
-                          Tensor bias, bool if_bias) {
-  const float *input_data = input->data<float>();
-  const float *filter_data = filter.data<float>();
-  float *output_data = output->data<float>();
-  const float *bias_data = bias.data<float>();
-
-  const int h = static_cast<int>(input->dims()[2]);
-  const int w = static_cast<int>(input->dims()[3]);
-  const int l = h;
-
-  const int batch_size = static_cast<int>(input->dims()[0]);
-  const int c = static_cast<int>(input->dims()[1]);
-  const int hxw = h * w;
-  float32x4_t vbias = vdupq_n_f32(0.0);
-  for (int b = 0; b < batch_size; ++b) {
-    const float *filter_data_tmp = filter_data;
-
-    for (int j = 0; j < c; ++j) {
-      if (if_bias) {
-        vbias = vdupq_n_f32(bias_data[j]);
-      }
-
-      int l_mid = l - 2;  // l=1->l_mid=-1,l=2->l_mid=0
-      float w00 = filter_data_tmp[0];
-      float w01 = filter_data_tmp[1];
-      float w02 = filter_data_tmp[2];
-      float w10 = filter_data_tmp[3];
-      float w11 = filter_data_tmp[4];
-      float w12 = filter_data_tmp[5];
-      float w20 = filter_data_tmp[6];
-      float w21 = filter_data_tmp[7];
-      float w22 = filter_data_tmp[8];
-
-      output_data[0] = w11 * input_data[0] + w12 * input_data[1] +
-                       w21 * input_data[l] + w22 * input_data[l + 1] +
-                       bias_data[j];
-      output_data[l - 1] = w10 * input_data[l - 2] + w11 * input_data[l - 1] +
-                           w20 * input_data[2 * l - 2] +
-                           w21 * input_data[2 * l - 1] + bias_data[j];
-      output_data[(l - 1) * l] =
-          w01 * input_data[(l - 2) * l] + w02 * input_data[(l - 2) * l + 1] +
-          w11 * input_data[(l - 1) * l] + w12 * input_data[(l - 1) * l + 1] +
-          bias_data[j];
-      output_data[l * l - 1] = w00 * input_data[(l - 2) * (l + 1)] +
-                               w01 * input_data[(l - 2) * (l + 1) + 1] +
-                               w10 * input_data[l * l - 2] +
-                               w11 * input_data[l * l - 1] + bias_data[j];
-
-      for (int i = 1; i < l - 1; ++i) {
-        output_data[i * l] =
-            w01 * input_data[i * l - l] + w02 * input_data[i * l - l + 1] +
-            w11 * input_data[i * l] + w12 * input_data[i * l + 1] +
-            w21 * input_data[i * l + l] + w22 * input_data[i * l + l + 1] +
-            bias_data[j];
-        output_data[i * l + l - 1] = w00 * input_data[i * l + l - 1 - l - 1] +
-                                     w01 * input_data[i * l + l - 1 - l] +
-                                     w10 * input_data[i * l + l - 1 - 1] +
-                                     w11 * input_data[i * l + l - 1] +
-                                     w20 * input_data[i * l + l - 1 + l - 1] +
-                                     w21 * input_data[i * l + l - 1 + l] +
-                                     bias_data[j];
-      }
-
-      // top 1 row and bottom 1 row
-      const float *input_tmp = input_data;
-
-      float32x4_t in0, in1, in2, in3, in4, in5, in6, in7, tmp0, tmp1, tmp2,
-          tmp3, tmp4, tmp5, out0;
-      in0 = vld1q_f32(input_tmp);
-      in2 = vld1q_f32(input_tmp + l);
-      const float *input_tmp_end = input_tmp + (l - 2) * l;
-      in4 = vld1q_f32(input_tmp_end);
-      in6 = vld1q_f32(input_tmp_end + l);
-      int c_mid = l_mid;
-      auto output_ptr = output_data + 1;
-      for (; c_mid > 3; c_mid -= 4) {
-        in1 = vld1q_f32(input_tmp + 4);
-        in3 = vld1q_f32(input_tmp + l + 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);
-
-        out0 = vmulq_n_f32(in0, w10);
-        out0 = vmlaq_n_f32(out0, tmp0, w11);
-        out0 = vmlaq_n_f32(out0, tmp1, w12);
-        out0 = vmlaq_n_f32(out0, in2, w20);
-        out0 = vmlaq_n_f32(out0, tmp2, w21);
-        out0 = vmlaq_n_f32(out0, tmp3, w22);
-        out0 = vaddq_f32(out0, vbias);
-
-        vst1q_f32(output_ptr, out0);
-
-        in5 = vld1q_f32(input_tmp_end + 4);
-        in7 = vld1q_f32(input_tmp_end + l + 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);
-
-        out0 = vmulq_n_f32(in4, w00);
-        out0 = vmlaq_n_f32(out0, tmp0, w01);
-        out0 = vmlaq_n_f32(out0, tmp1, w02);
-        out0 = vmlaq_n_f32(out0, in6, w10);
-        out0 = vmlaq_n_f32(out0, tmp2, w11);
-        out0 = vmlaq_n_f32(out0, tmp3, w12);
-        out0 = vaddq_f32(out0, vbias);
-
-        vst1q_f32(output_ptr + (l - 1) * l, out0);
-
-        // 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 pad
-      float32x4_t pad0 = vdupq_n_f32(input_data[l - 1]);
-      float32x4_t pad1 = vdupq_n_f32(input_data[2 * l - 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);
-
-      out0 = vmulq_n_f32(in0, w10);
-      out0 = vmlaq_n_f32(out0, tmp0, w11);
-      out0 = vmlaq_n_f32(out0, tmp1, w12);
-      out0 = vmlaq_n_f32(out0, in2, w20);
-      out0 = vmlaq_n_f32(out0, tmp2, w21);
-      out0 = vmlaq_n_f32(out0, tmp3, w22);
-      out0 = vaddq_f32(out0, vbias);
-
-      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 pad
-      float32x4_t pad2 = vdupq_n_f32(input_data[l * l - 1 - l]);
-      float32x4_t pad3 = vdupq_n_f32(input_data[l * l - 1]);
-
-      tmp0 = vextq_f32(in4, pad2, 1);
-      tmp1 = vextq_f32(in4, pad2, 2);
-      tmp2 = vextq_f32(in6, pad3, 1);
-      tmp3 = vextq_f32(in6, pad3, 2);
-
-      out0 = vmulq_n_f32(in4, w00);
-      out0 = vmlaq_n_f32(out0, tmp0, w01);
-      out0 = vmlaq_n_f32(out0, tmp1, w02);
-      out0 = vmlaq_n_f32(out0, in6, w10);
-      out0 = vmlaq_n_f32(out0, tmp2, w11);
-      out0 = vmlaq_n_f32(out0, tmp3, w12);
-      out0 = vaddq_f32(out0, vbias);
-
-      for (int i = 0; i < c_mid; ++i) {
-        if (i == 0) {
-          vst1q_lane_f32(output_ptr + (l - 1) * l + i, out0, 0);
-        }
-        if (i == 1) {
-          vst1q_lane_f32(output_ptr + (l - 1) * l + i, out0, 1);
-        }
-        if (i == 2) {
-          vst1q_lane_f32(output_ptr + (l - 1) * l + i, out0, 2);
-        }
-      }
-      // mid
-
-      for (int i = 0; i < l - 2; ++i) {
-        auto output_ptr = output_data + (i + 1) * l + 1;
-        input_tmp = input_data + i * l;
-        auto in0_tmp = vld1q_f32(input_tmp);
-        auto in2_tmp = vld1q_f32(input_tmp + l);
-        auto in4_tmp = vld1q_f32(input_tmp + l + l);
-        c_mid = l_mid;
-        for (; c_mid > 3; c_mid -= 4) {
-          auto in1_tmp = vld1q_f32(input_tmp + 4);
-          auto in3_tmp = vld1q_f32(input_tmp + l + 4);
-          auto in5_tmp = vld1q_f32(input_tmp + l + l + 4);
-
-          tmp0 = vextq_f32(in0_tmp, in1_tmp, 1);
-          tmp1 = vextq_f32(in0_tmp, in1_tmp, 2);
-          tmp2 = vextq_f32(in2_tmp, in3_tmp, 1);
-          tmp3 = vextq_f32(in2_tmp, in3_tmp, 2);
-          tmp4 = vextq_f32(in4_tmp, in5_tmp, 1);
-          tmp5 = vextq_f32(in4_tmp, in5_tmp, 2);
-
-          out0 = vmulq_n_f32(in0_tmp, w00);
-          out0 = vmlaq_n_f32(out0, tmp0, w01);
-          out0 = vmlaq_n_f32(out0, tmp1, w02);
-          out0 = vmlaq_n_f32(out0, in2_tmp, w10);
-          out0 = vmlaq_n_f32(out0, tmp2, w11);
-          out0 = vmlaq_n_f32(out0, tmp3, w12);
-          out0 = vmlaq_n_f32(out0, in4_tmp, w20);
-          out0 = vmlaq_n_f32(out0, tmp4, w21);
-          out0 = vmlaq_n_f32(out0, tmp5, w22);
-          out0 = vaddq_f32(out0, vbias);
-
-          vst1q_f32(output_ptr, out0);
-
-          output_ptr += 4;
-          input_tmp += 4;
-          in0_tmp = in1_tmp;
-          in2_tmp = in3_tmp;
-          in4_tmp = in5_tmp;
-        }
-
-        float32x4_t pad0 = vdupq_n_f32(input_data[i * l + l - 1]);
-        float32x4_t pad1 = vdupq_n_f32(input_data[i * l + l - 1 + l]);
-        float32x4_t pad2 = vdupq_n_f32(input_data[i * l + l - 1 + l + l]);
-
-        tmp0 = vextq_f32(in0_tmp, pad0, 1);
-        tmp1 = vextq_f32(in0_tmp, pad0, 2);
-        tmp2 = vextq_f32(in2_tmp, pad1, 1);
-        tmp3 = vextq_f32(in2_tmp, pad1, 2);
-        tmp4 = vextq_f32(in4_tmp, pad2, 1);
-        tmp5 = vextq_f32(in4_tmp, pad2, 2);
-
-        out0 = vmulq_n_f32(in0_tmp, w00);
-        out0 = vmlaq_n_f32(out0, tmp0, w01);
-        out0 = vmlaq_n_f32(out0, tmp1, w02);
-        out0 = vmlaq_n_f32(out0, in2_tmp, w10);
-        out0 = vmlaq_n_f32(out0, tmp2, w11);
-        out0 = vmlaq_n_f32(out0, tmp3, w12);
-        out0 = vmlaq_n_f32(out0, in4_tmp, w20);
-        out0 = vmlaq_n_f32(out0, tmp4, w21);
-        out0 = vmlaq_n_f32(out0, tmp5, w22);
-        out0 = vaddq_f32(out0, vbias);
-
-        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);
-          }
-        }
-      }
-      output_data += hxw;
-      input_data += hxw;
-      filter_data_tmp += 9;
-    }
-  }
-}
-}  // namespace math
-}  // namespace operators
-}  // namespace paddle_mobile