深度学习（六十一）NNPACK 移植与实验

一、Ubuntu下使用：

1、克隆下载NNPACK

git clone https://github.com/Maratyszcza/NNPACK.git

2、安装nijia，并编译

Install ninja build system

sudo apt-get install ninja-build || brew install ninja

Install PeachPy assembler and confu configuration system

[sudo] pip install --upgrade git+https://github.com/Maratyszcza/PeachPy
[sudo] pip install --upgrade git+https://github.com/Maratyszcza/confu

Then clone NNPACK, install dependencies, configure, and build

cd NNPACK
confu setup
python ./configure.py
ninja

3、链接的时候还需要用到pthread库，否在会遇到找不到实现函数。

target_link_libraries(NNPACK libnnpack.a libpthreadpool.a pthread)

需要加入pthread库，否这会报错。

4、测试代码

#include <iostream>
#include "nnpack.h"
#include <ctime>
#include <vector>

using namespace std;
float test_nnpack(){
    //init nnpack
    enum nnp_status init_status = nnp_initialize();
    if (init_status != nnp_status_success) {
        return 0;
    }

    enum nnp_convolution_algorithm algorithm = nnp_convolution_algorithm_auto;
    enum nnp_convolution_transform_strategy strategy=nnp_convolution_transform_strategy_tuple_based;
    const size_t batch_size = 1;
    const size_t input_channels = 128;
    const size_t output_channels = 128;
    const struct nnp_padding input_padding = { 1, 1, 1, 1 };
    const struct nnp_size input_size ={ 256, 256};
    const struct nnp_size kernel_size = { 5, 5 };
    const struct  nnp_size stride={.width=2,.height=2};
    const struct nnp_size output_size = {
            .width = (input_padding.left + input_size.width + input_padding.right - kernel_size.width)/stride.width + 1,
            .height =(input_padding.top + input_size.height + input_padding.bottom - kernel_size.height)/stride.height  + 1
    };


    //malloc memory for input, kernel, output, bias
    float* input = (float*)malloc(batch_size * input_channels *input_size.height *input_size.width * sizeof(float));
    float* kernel = (float*)malloc(input_channels * output_channels * kernel_size.height * kernel_size.width * sizeof(float));
    float* output = (float*)malloc(batch_size* output_channels * output_size.height * output_size.width * sizeof(float));
    float* bias = (float*)malloc(output_channels * sizeof(float));

    pthreadpool_t threadpool= nullptr;


    struct nnp_profile computation_profile;//use for compute time;
    //init input data
    int i,j;
    for(int c=0; c<input_channels;c++ ){
        for(i=0; i<input_size.height; i++){
            for(j=0; j<input_size.width; j++){
                input[c*input_size.height*input_size.width+i*input_size.width+j] = (i*input_size.width+j)*0.1;
            }
        }
    }

    //init kernel data
    for(int i=0; i<output_channels;i++ ){
        for(j=0; j<input_channels*kernel_size.height*kernel_size.width; j++){
            kernel[i*input_channels*kernel_size.height*kernel_size.width+j] = 0.1;
        }
    }

    //init bias data
    for(int i=0; i<output_channels;i++ ){
        bias[i] = 1.0;
    }

    //execute conv

    for(int i=0;i<10;i++)
    {
        nnp_convolution_inference(algorithm,
                                  strategy,
                                  input_channels,
                                  output_channels,
                                  input_size,
                                  input_padding,
                                  kernel_size,
                                  stride,
                                  input,
                                  kernel,
                                  bias,
                                  output,
                                  threadpool,
                                  nullptr);
    }

  std::vector<float>out;
    for(int i=0;i<output_channels*output_size.height*output_size.width;i++){
        out.push_back(output[i]);
    }

    return conv_time_use;
}
int main() {
    cout << test_nnpack()<< endl;
    return 0;
}

在mxnet、tiny-dnn中的引用方法：

https://www.insight.io/github.com/dmlc/mxnet/blob/master/src/operator/nnpack/nnpack_convolution-inl.h

https://github.com/tiny-dnn/tiny-dnn/blob/master/tiny_dnn/core/kernels/conv2d_op_nnpack.h

在tiny-dnn中，底层调用nnpack为啥只采用了一个线程，于是我测试了多线程的速度，经过测试发现nnpack的弱点，原来在于多线程：

二、android使用

1、添加ndk-build环境变量路径：

export NDK_ROOT=/home/hjimce/Android/Sdk/ndk-bundle/

cd到nnpack所在的目录，更改jni APP_PLATFORM := android-21

2、直接运行：

${NDK_ROOT}/ndk-build 

3、代码中引用cmake要指定相关的库文件：

arguments "-DANDROID_ARM_NEON=TRUE", "-DANDROID_TOOLCHAIN=clang"
ndk {
    // Specifies the ABI configurations of your native
    // libraries Gradle should build and package with your APK.
    abiFilters 'x86','armeabi-v7a'
}

手机速度测试：

修改：多线程时间测试不能使用time.clock(),可使用auto begin =std::chrono::high_resolution_clock::now();