uint8_t* readbmpfile(const char* filepath) {FILE* img_pFile;size_t img_size;size_t result;img_pFile = fopen(filepath, "rb");if (img_pFile == NULL) { fputs("File error", stderr); exit(1); }// 读bmp大小fseek(img_pFile, 0, SEEK_END);//把指针放去末尾img_size = ftell(img_pFile);//读出指针的位置，也就是这个文件的byte数量rewind(img_pFile);//将指针放回开头// 开块动态数组uint8_t* ptr_buffer;ptr_buffer = new uint8_t[img_size - BMP_OFFSET];if (ptr_buffer == NULL) { fputs("Memory error", stderr); exit(2); }//跳1078fseek(img_pFile, BMP_OFFSET, SEEK_SET);//将bmp从1078位以后的内容读去动态数组ptr_bufferresult = fread(ptr_buffer, sizeof(uint8_t), img_size - BMP_OFFSET, img_pFile);//result是正确读到的元素数量if (result != img_size - BMP_OFFSET) { fputs("Reading error", stderr); exit(3); }cout << result << endl;fclose(img_pFile);return ptr_buffer;
}

int main(int argc, char* argv[]) {cout<<"setting address of xclbin"<<endl<<endl;/**********annotation*********** ↓ 通过argv把xo文件的xclbin传进来* ********************/if (argc != 2) {std::cout << "Usage: " << argv[0] << " <XCLBIN File>" << std::endl;return EXIT_FAILURE;}std::string binaryFile = argv[1];//传入卷积kernel的地址//std::string binaryFile_2 = argv[2];//传入池化kernel的地址cout<<argv[0]<<endl;//地址    cout<<argv[1]<<endl;//第一个xclbin

cout<<"#######read w_conv1#######"<<endl;size_t size_w_conv1_byte = sizeoffile("../data/W_conv1.bin");cout<<"size_w_conv1_byte"<<size_w_conv1_byte<<endl;size_t size_w_conv1_float = size_w_conv1_byte/4;cout<<"size_w_conv1_float"<<size_w_conv1_float<<endl;std::vector<float,aligned_allocator<float>> din_w_conv1(size_w_conv1_float);cout<<"allocation of memory finished"<<endl;float* ptr_buffer_w_conv1 = readfilterfile_to_float("../data/W_conv1.bin",IN_HEIGHT1,IN_WIDTH1,IN_CH1,OUT_CH1);//把数据从我们的读取函数取出来，放到vector里面cout<<"readfilterfile_to_float (W_conv1.bin) finished"<<endl;for(size_t i = 0 ; i < size_w_conv1_float ; i++){din_w_conv1[i] = ptr_buffer_w_conv1[i];}cout<<"setting din_w_conv1 finished"<<endl;delete[] ptr_buffer_w_conv1;for (int i = 0; i < 16; i++) {for (int j = 0; j < 3; j++) {for (int k = 0; k < 3; k++) {cout << setw(5) << (din_w_conv1[i*3*3 + j*3+k]);}//          cout << endl;}}

/**********annotation*********** ↓分配输出第一次卷积后，dout_featureout1的内存，并且初始化为0* ********************/size_t size_featureout1 = OUT_CH1* IN_WIDTH1* IN_HEIGHT1;//16*28*28std::vector<float,aligned_allocator<float>> dout_featureout1(size_featureout1);for(size_t i = 0 ; i < size_featureout1 ; i++){dout_featureout1[i] = 0;}cout<<"allocation and setting of dou_featureout1"<<endl<<endl;

cout << "##########step3:OPENCL configuration##############" << endl << endl;/**********annotation**** ↓OPENCL的操作:* one device* one context* one queue* two binary files* two programs* two kernels* ********************/// Step 0: Device //配置器件std::vector<cl::Device> devices = get_devices("Xilinx");devices.resize(1);cl::Device device = devices[0];// Step 1: Create Context//配个contextOCL_CHECK(err, cl::Context context(device, NULL, NULL, NULL, &err));cout<<"create context"<<endl;// Step 2: Create Command Queue//生产指令列表//OCL_CHECK(err, cl::CommandQueue q(context, device, CL_QUEUE_PROFILING_ENABLE, &err));OCL_CHECK(err, cl::CommandQueue q(context, device, 0, &err));cout<<"create commandqueue"<<endl;// Step 3: Load Binary File from disk//读取二进制xo文件unsigned fileBufSize;//fileBufSize在下面的read_binary_file函数里面被赋了文件大小值char* fileBuf = read_binary_file(binaryFile, fileBufSize);//这里用到了最开始的argv，所以把上面的语句放这应该会更好点cl::Program::Binaries bins{{fileBuf, fileBufSize}};cout<<"read conv xclbin"<<endl;// Step 4: Create the program object from the binary and program the FPGA device with it//个人理解：相当于烧bitstreamOCL_CHECK(err, cl::Program program(context, devices, bins, NULL, &err));cout<<"program xclbin"<<endl;// Step 5: Create Kernels//个人理解：相当于给烧了bitstream的PL部分配上PS的寄存器OCL_CHECK(err, cl::Kernel krnl_Conv(program,"Conv", &err));//kernel1//这里用的名字应该是hls选的那个TOP函数的函数名cout<<"create kernel:Conv finished"<<endl<<endl;...其它ip核同理

 cout<<"step4:setting parameters for kernel"<<endl;// ================================================================// Setup Buffers and run Kernels// Allocate Global Memory for source_in1 //在device上面分配内存,分配的时候，这里应该.data()只是返回指针，所以这里没有真的在做搬运，而只是标了真正在queue里面要搬运时候的地址// 自己做个规范，host的内存，用下划线，global的内存，用驼峰命名法/ 在device上配读img的位置，读第一次卷积核与偏置的位置，读输出doutFeatureOut1和输出doutFeatureOut11的位置OCL_CHECK(err, cl::Buffer dinImage   (context,CL_MEM_USE_HOST_PTR | CL_MEM_READ_ONLY,img_size, din_img.data(), &err));OCL_CHECK(err, cl::Buffer dinWeightConv1   (context,CL_MEM_USE_HOST_PTR | CL_MEM_READ_ONLY,size_w_conv1_byte, din_w_conv1.data(), &err));OCL_CHECK(err, cl::Buffer dinBiasConv1   (context,CL_MEM_USE_HOST_PTR | CL_MEM_READ_ONLY,size_b_conv1_byte, din_b_conv1.data(), &err));OCL_CHECK(err, cl::Buffer doutFeatureOut1   (context,CL_MEM_USE_HOST_PTR | CL_MEM_READ_ONLY,size_featureout1*4, dout_featureout1.data(), &err));OCL_CHECK(err, cl::Buffer doutFeatureOut11   (context,CL_MEM_USE_HOST_PTR | CL_MEM_READ_ONLY,size_featureout11*4, dout_featureout11.data(), &err));....读取其它参数同理

// 设置第一次卷积和池化的参数OCL_CHECK(err, err = krnl_Conv.setArg(0, IN_CH1 ));//OCL_CHECK(err, err = krnl_Conv.setArg(1,IN_HEIGHT1 ));//OCL_CHECK(err, err = krnl_Conv.setArg(2,IN_WIDTH1 ));//OCL_CHECK(err, err = krnl_Conv.setArg(3,OUT_CH1 ));//OCL_CHECK(err, err = krnl_Conv.setArg(4, KERNEL_WIDTH1));//OCL_CHECK(err, err = krnl_Conv.setArg(5,KERNEL_HEIGHT1 ));//OCL_CHECK(err, err = krnl_Conv.setArg(6,X_STRIDE1 ));//OCL_CHECK(err, err = krnl_Conv.setArg(7,Y_STRIDE1 ));//OCL_CHECK(err, err = krnl_Conv.setArg(8,MODE1 ));//OCL_CHECK(err, err = krnl_Conv.setArg(9,RELU_EN1 ));//OCL_CHECK(err, err = krnl_Conv.setArg(10, dinImage));//OCL_CHECK(err, err = krnl_Conv.setArg(11, dinWeightConv1));//OCL_CHECK(err, err = krnl_Conv.setArg(12, dinBiasConv1));//OCL_CHECK(err, err = krnl_Conv.setArg(13, doutFeatureOut1));//cout<<"setting parameters for first krnl_Conv finished"<<endl;OCL_CHECK(err, err = krnl_Pool.setArg(0, IN_CH11));//OCL_CHECK(err, err = krnl_Pool.setArg(1, IN_HEIGHT11));//OCL_CHECK(err, err = krnl_Pool.setArg(2, IN_WIDTH11));//OCL_CHECK(err, err = krnl_Pool.setArg(3, KERNEL_WIDTH11));//OCL_CHECK(err, err = krnl_Pool.setArg(4, KERNEL_HEIGHT11));//OCL_CHECK(err, err = krnl_Pool.setArg(5, MODE11));//它娘的参数粘帖错了 mmpOCL_CHECK(err, err = krnl_Pool.setArg(6, doutFeatureOut1));//读入上面的结果doutFeatureOut1OCL_CHECK(err, err = krnl_Pool.setArg(7, doutFeatureOut11));//输出为doutFeatureOut11cout << "setting parameters for first krnl_Pool finished" << endl;
...其它核配置方法同理

/**********annotation**********
* 准备执行第一次卷积池化
* ****************/cout<<"开始压指令"<<endl;//从stackoverflow上看，在这把kernel压入队列之后，压入的操作就定死了，后面就可以去改argument重新压新操作进队列了cout<<"1.把img,第一次卷积用的filter和bias从host搬去device"<<endl;OCL_CHECK(err, err = q.enqueueMigrateMemObjects({dinImage, dinWeightConv1,dinBiasConv1},0/* 0 means from host*/));    //把device上的global内容给到kernel里面，话说这里应该是写错了，0标志这里做的是从global传去给kernel里面cout<<"2.执行krnl_Conv"<<endl;OCL_CHECK(err, err = q.enqueueTask(krnl_Conv));cout<<"3.搬输出回host"<<endl;OCL_CHECK(err, err = q.enqueueMigrateMemObjects({doutFeatureOut1},CL_MIGRATE_MEM_OBJECT_HOST));//把算完的doutFeatureOut1拿出来
//cout<<"#######再看一遍host上现在到三组数据#######"<<endl;cout<<"转成float的host上到图像内容"<<endl;for (int k = 0; k < 100; k++) {cout << setw(5) << (din_img[k]);}cout<<endl;cout<<"host上的din_w_conv1"<<endl;for (int k = 0; k < 100; k++) {cout << setw(5) << (din_w_conv1[k]);}cout<<endl;cout<<"host上的din_b_conv1"<<endl;for (int k = 0; k < 100; k++) {cout << setw(5) << (din_b_conv1[k]);}cout<<endl;cout<<"#######第一次conv前host上dout_featureout1前100内容#######"<<endl;for (int k = 0; k < 100; k++) {cout << setw(5) << (float)(dout_featureout1[k]);}cout<<endl;cout<<"执行q.finish()"<<endl;q.finish();cout<<"#######第一次conv后host上dout_featureout1前100内容#######"<<endl;for (int k = 0; k < 100; k++) {cout << setw(5) << (dout_featureout1[k]);}cout<<endl;