OpenCL配置及测试，使用visual studio

依赖库

OpenCL是一套标准，由Khronos Group管理，Khronos在github上有一个仓库，另外各个硬件厂家也都有自己的实现。

KhronosGroup

github地址：https://github.com/KhronosGroup/OpenCL-SDK
点击页面右侧的Releases，根据自己的环境进行下载。
可以下载编译好的压缩包，比如：OpenCL-SDK-v2022.09.30-Win-x64.zip
解压后可以看到bin，include，lib等文件夹，这些是主要要用的，习惯上我会把最外层文件夹名字后缀全部干掉，只留下OpenCL。

Nvidia

Nvidia的OpenCL库在装Cuda的时候会顺带装上，需要去Cuda安装目录里面扒出来。
Cuda下载：https://developer.nvidia.com/cuda-downloads
官方Demo：https://developer.nvidia.com/opencl

安装之后去下面目录中把CL文件夹拷贝出来：
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.4\include
然后去下面两个文件夹中把OpenCL.lib拷贝出来：
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.4\lib\Win32
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.4\lib\x64

最后形成这样一个文件夹结构：

• OpenCL
  • include
    • CL：一堆头文件
  • lib
    • Win32：OpenCL.lib
    • x64：OpenCL.lib

Intel

Intel比较坑，需要下载sdk并安装，但是下载sdk的过程必须要求注册。
Intel OpenCL sdk下载：https://software.intel.com/en-us/intel-opencl/download

安装后的sdk的路径为：
C:\Program Files (x86)\IntelSWTools\system_studio_2020\OpenCL\sdk
这个路径下有include和lib，拷贝出来形成与上述相同的文件夹结构。

在Visual Studio中配置依赖库

下面以KhronosGroup的SDK为例。
创建一个新的VS空项目，平台改成x64，然后把OpenCL依赖库拷贝进去。

配置include和lib目录

右键点击项目 - 属性，按如下方式配置include和lib路径。
请注意配置成相对路径，这样把工程拷贝给别人后，别人不需要更改直接就可以运行，下同。

配置附加依赖项

点击链接器 - 输入 - 附加依赖项，把OpenCL.lib写进去。

配置dll目录（可选）

在调试 - 环境中配置dll路径，注意这时候路径前面要写PATH=。
这一步常常是可选的，为什么呢？因为如果安装过cuda或者其他类似驱动的话，Windows\System32文件夹下常常已经有OpenCL.dll了，这个系统文件夹下的dll可以默认被调用到，所以不用配置也可以使用。（不过System32文件夹下的dll和KhronosGroup的dll不是一套代码编译出来的，可能会存在接口不一致情况。）

测试代码

入门可以找一些公开的代码库学习一下，比如Nvidia的Samples：https://developer.nvidia.com/opencl

但是Nvidia自己搞了一些公用库，并且有些工程项目使用了非常老的vs版本（vs08，vs10之类），不太好配。
因此下面以其中一个DotProduct为例，把代码重新梳理了一下，把下面的两个文件main.cpp和dot_product.cl放入工程目录下，并把main.cpp添加到工程源文件中，即可点F5执行测试。
dot_product.cl是kernel文件，内容是Nvidia sample里面贴过来的，一点没改。注意文件名不要改，因为主程序要通过文件名去找kernel文件。

main.cpp

#include 
#include 
#include 
#include 
#include 


void CheckErrorCode(cl_int error_code, cl_int reference, std::string function_name, int return_code);
template<typename T> T RoundUp(T group_size, T global_size);
void FillArray(float* data, size_t size);
std::string ReadFile(const std::string& filaname);
void DotProductHost(const float* a, const float* b, float* result, size_t array_size);
int CompareResult(const float* reference, const float* data, size_t array_size, const float epsilon);


int main()
{
    // common variables
    cl_int error_code;
    size_t array_size = 9992222;        // Length of float arrays to process
    std::string kernel_souce_file = "dot_product.cl";

    // Get the NVIDIA platform
    std::string platform_name_keyword = "NVIDIA";
    cl_platform_id platform = nullptr;  // output of this block
    cl_uint num_platforms;

    error_code = clGetPlatformIDs(0, NULL, &num_platforms);
    CheckErrorCode(error_code, CL_SUCCESS, "clGetPlatformIDs", -11);
    if (num_platforms == 0) {
        std::cout << "No OpenCL platform found!" << std::endl;
        return -12;
    } else {
        // if there's a platform or more, make space for ID's
        cl_platform_id* platforms = nullptr;
        platforms = (cl_platform_id*)malloc(num_platforms * sizeof(cl_platform_id));
        if (platforms == nullptr) {
            std::cout << "Failed to allocate memory for cl_platform ID's!" << std::endl;
            return -13;
        }

        error_code = clGetPlatformIDs(num_platforms, platforms, NULL);
        CheckErrorCode(error_code, CL_SUCCESS, "clGetPlatformIDs", -14);

        char info_buffer[1024] = { 0 };
        for (cl_uint i = 0; i < num_platforms; ++i) {
            error_code = clGetPlatformInfo(platforms[i], CL_PLATFORM_NAME, sizeof(info_buffer), &info_buffer, NULL);
            CheckErrorCode(error_code, CL_SUCCESS, "clGetPlatformInfo", -15);

            std::string platform_name = info_buffer;
            if (platform_name.find(platform_name_keyword) != std::string::npos) {
                platform = platforms[i];
                break;
            }
        }

        if (platform == nullptr) {
            std::cout << "WARNING: " << platform_name_keyword <<
                " OpenCL platform not found - defaulting to first platform!" << std::endl;
            platform = platforms[0];
        }
        free(platforms);
        platforms = nullptr;
    }

    // Get device
    cl_device_id device = nullptr;      // output of this block
    cl_uint num_devices = 0;            // Number of devices available
    cl_uint target_device_id = 0;       // Default Device to compute on
    cl_uint num_compute_units = 0;      // Number of compute units (SM's on NV GPU)

    error_code = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 0, NULL, &num_devices);
    CheckErrorCode(error_code, CL_SUCCESS, "clGetDeviceIDs", -21);
    if (num_devices == 0) {
        std::cout << "No OpenCL devices found!" << std::endl;
        return -22;
    } else {
        // if there's a device or more, make space for ID's
        cl_device_id* devices = nullptr;
        devices = (cl_device_id*)malloc(num_devices * sizeof(cl_device_id));
        if (devices == nullptr) {
            std::cout << "Failed to allocate memory for cl_device ID's!" << std::endl;
            return -23;
        }

        // clamp target_device_id
        target_device_id = std::min(std::max(target_device_id, static_cast<cl_uint>(0)), num_devices - 1);

        // select device
        error_code = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, num_devices, devices, NULL);
        CheckErrorCode(error_code, CL_SUCCESS, "clGetDeviceIDs", -24);
        device = devices[target_device_id];

        // print device info
        char info_buffer[1024] = { 0 };
        error_code = clGetDeviceInfo(device, CL_DEVICE_NAME, sizeof(info_buffer), &info_buffer, NULL);
        CheckErrorCode(error_code, CL_SUCCESS, "clGetDeviceInfo", -25);
        std::cout << "Using GPU device: " << info_buffer << std::endl;

        // get and print number of compute units
        error_code = clGetDeviceInfo(device, CL_DEVICE_MAX_COMPUTE_UNITS, sizeof(num_compute_units), &num_compute_units, NULL);
        CheckErrorCode(error_code, CL_SUCCESS, "clGetDeviceInfo", -26);
        std::cout << "Compute Units: " << num_compute_units << std::endl;

        free(devices);
        devices = nullptr;
    }

    // Set local and global work size
    size_t local_work_size = 256;
    size_t global_work_size = RoundUp(local_work_size, array_size);
    std::cout << "array_size: " << array_size << std::endl;
    std::cout << "local work size: " << local_work_size << std::endl;
    std::cout << "global_work_size:" << global_work_size << std::endl;

    // Create the context
    cl_context gpu_context = nullptr;
    gpu_context = clCreateContext(0, 1, &device, NULL, NULL, &error_code);
    CheckErrorCode(error_code, CL_SUCCESS, "clCreateContext", -31);

    // Create a command-queue
    std::cout << "clCreateCommandQueue..." << std::endl;
    cl_command_queue command_queue = nullptr;
    //command_queue = clCreateCommandQueue(gpu_context, device, 0, &error_code);    // clCreateCommandQueue is deprecated after CL1.2
    command_queue = clCreateCommandQueueWithProperties(gpu_context, device, 0, &error_code);
    CheckErrorCode(error_code, CL_SUCCESS, "clCreateCommandQueue", -41);

    // Read the OpenCL kernel in from source file
    std::cout << "oclLoadProgSource " << kernel_souce_file << "..." << std::endl;
    std::string kernel_source = ReadFile(kernel_souce_file);
    CheckErrorCode(kernel_source != "", true, "Read kernel source file", -61);
    size_t kernel_length = kernel_source.size();
    const char* kernel_souce_cstr = kernel_source.c_str();

    // Create the program
    std::cout << "clCreateProgramWithSource..." << std::endl;
    cl_program program = nullptr;
    program = clCreateProgramWithSource(gpu_context, 1, (const char**)&kernel_souce_cstr, &kernel_length, &error_code);
    CheckErrorCode(error_code, CL_SUCCESS, "clCreateProgramWithSource", -62);

    error_code = clBuildProgram(program, 0, NULL, NULL, NULL, NULL);
    CheckErrorCode(error_code, CL_SUCCESS, "clBuildProgram", -63);

    // Create the kernel
    std::cout << "clCreateKernel (dot_product)..." << std::endl;
    cl_kernel kernel = nullptr;
    kernel = clCreateKernel(program, "DotProduct", &error_code);    // kernel_name (2nd param) should be same as the function name in kernel source file
    CheckErrorCode(error_code, CL_SUCCESS, "clCreateKernel", -64);

    // Allocate and initialize host arrays
    std::cout << "Allocate and Init Host Mem..." << std::endl;
    void* src_a;
    void* src_b;
    void* dst;
    void* golden;
    src_a = (void*)malloc(sizeof(cl_float4) * global_work_size);
    src_b = (void*)malloc(sizeof(cl_float4) * global_work_size);
    dst = (void*)malloc(sizeof(cl_float) * global_work_size);
    golden = (void*)malloc(sizeof(cl_float) * array_size);
    FillArray(static_cast<float*>(src_a), 4 * array_size);
    FillArray(static_cast<float*>(src_b), 4 * array_size);

    // Allocate the OpenCL buffer memory objects for source and result on the device GMEM
    std::cout << "clCreateBuffer on the Device GMEM..." << std::endl;
    cl_mem device_src_a = nullptr;
    cl_mem device_src_b = nullptr;
    cl_mem device_dst = nullptr;
    device_src_a = clCreateBuffer(gpu_context, CL_MEM_READ_ONLY, sizeof(cl_float) * global_work_size * 4, NULL, &error_code);
    CheckErrorCode(error_code, CL_SUCCESS, "clCreateBuffer", -51);
    device_src_b = clCreateBuffer(gpu_context, CL_MEM_READ_ONLY, sizeof(cl_float) * global_work_size * 4, NULL, &error_code);
    CheckErrorCode(error_code, CL_SUCCESS, "clCreateBuffer", -52);
    device_dst = clCreateBuffer(gpu_context, CL_MEM_READ_ONLY, sizeof(cl_float) * global_work_size, NULL, &error_code);
    CheckErrorCode(error_code, CL_SUCCESS, "clCreateBuffer", -53);

    // Set the Argument values
    std::cout << "clSetKernelArg 0 - 3..." << std::endl;
    error_code = clSetKernelArg(kernel, 0, sizeof(cl_mem), (void*)&device_src_a);
    error_code |= clSetKernelArg(kernel, 1, sizeof(cl_mem), (void*)&device_src_b);
    error_code |= clSetKernelArg(kernel, 2, sizeof(cl_mem), (void*)&device_dst);
    error_code |= clSetKernelArg(kernel, 3, sizeof(cl_int), (void*)&array_size);
    CheckErrorCode(error_code, CL_SUCCESS, "clSetKernelArg", -65);

    // Asynchronous write of data to GPU device
    std::cout << "clEnqueueWriteBuffer (src_a and src_b)..." << std::endl;
    error_code = clEnqueueWriteBuffer(command_queue, device_src_a, CL_FALSE, 0, sizeof(cl_float) * global_work_size * 4, src_a, 0, NULL, NULL);
    error_code |= clEnqueueWriteBuffer(command_queue, device_src_b, CL_FALSE, 0, sizeof(cl_float) * global_work_size * 4, src_b, 0, NULL, NULL);
    CheckErrorCode(error_code, CL_SUCCESS, "clEnqueueWriteBuffer", -66);

    // Launch kernel
    clock_t device_time_beg;
    clock_t device_time_end;
    std::cout << "clEnqueueNDRangeKernel (DotProduct)..." << std::endl;
    device_time_beg = clock();
    error_code = clEnqueueNDRangeKernel(command_queue, kernel, 1, NULL, &global_work_size, &local_work_size, 0, NULL, NULL);
    device_time_end = clock();
    float device_time_cost = (float)(device_time_end - device_time_beg);
    CheckErrorCode(error_code, CL_SUCCESS, "clEnqueueNDRangeKernel", -71);


    // Read back results and check accumulated errors
    std::cout << "clEnqueueReadBuffer (dst)..." << std::endl;
    error_code = clEnqueueReadBuffer(command_queue, device_dst, CL_TRUE, 0, sizeof(cl_float) * global_work_size, dst, 0, NULL, NULL);
    CheckErrorCode(error_code, CL_SUCCESS, "clSetKernelArg", -81);

    // Compute and compare results for golden-host and report errors and pass/fail
    std::cout << "Comparing against Host/C++ computation..." << std::endl;
    clock_t host_time_beg;
    clock_t host_time_end;
    host_time_beg = clock();
    DotProductHost((const float*)src_a, (const float*)src_b, (float*)golden, array_size);
    host_time_end = clock();
    float host_time_cost = (float)(host_time_end - host_time_beg);
    int error_count = CompareResult((const float*)golden, (const float*)dst, array_size, 0.000001f);
    std::cout << "error value count: " << error_count << std::endl;


    // Print time cost for host and device
    std::cout << "Host time cost (ms): " << host_time_cost << std::endl;
    std::cout << "Device time cost (ms):" << device_time_cost << std::endl;


    // Release and clean
    free(src_a);
    free(src_b);
    free(dst);
    free(golden);
    src_a = nullptr;
    src_b = nullptr;
    dst = nullptr;
    golden = nullptr;
    kernel_souce_cstr = nullptr;

    if (gpu_context != nullptr) {
        clReleaseContext(gpu_context);
        gpu_context = nullptr;
    }
    if (command_queue != nullptr) {
        clReleaseCommandQueue(command_queue);
        command_queue = nullptr;
    }
    if (program != nullptr) {
        clReleaseProgram(program);
        program = nullptr;
    }
    if (kernel != nullptr) {
        clReleaseKernel(kernel);
        kernel = nullptr;
    }
    if (device_src_a != nullptr) {
        clReleaseMemObject(device_src_a);
        device_src_a = nullptr;
    }
    if (device_src_b != nullptr) {
        clReleaseMemObject(device_src_b);
        device_src_b = nullptr;
    }
    if (device_dst != nullptr) {
        clReleaseMemObject(device_dst);
        device_dst = nullptr;
    }

    return 0;
}


void CheckErrorCode(cl_int error_code, cl_int reference, std::string function_name, int return_code)
{
    if (error_code != reference) {
        std::cout << "Error in " << function_name << " Call, error code = " << error_code << std::endl;
        exit(return_code);
    }
    return;
}

template<typename T>
T RoundUp(T group_size, T global_size)
{
    T r = global_size % group_size;
    if (r == 0) {
        return global_size;
    } else {
        return global_size + group_size - r;
    }
}


void FillArray(float* data, size_t size)
{
    const float scale = 1.0f / static_cast<float>(RAND_MAX);
    for (size_t i = 0; i < size; ++i) {
        data[i] = scale * rand();
    }
}


std::string ReadFile(const std::string& filaname)
{
    std::string filestring;
    std::fstream ifs;
    ifs.open(filaname, std::ios::in);
    if (ifs) {
        std::istreambuf_iterator<char> begin(ifs);
        std::istreambuf_iterator<char> end;
        filestring = std::string(begin, end);
    }
    ifs.close();
    return filestring;
}


void DotProductHost(const float* a, const float* b, float* result, size_t array_size)
{
    for (size_t i = 0; i < array_size; i++) {
        result[i] = 0.0f;
        for (size_t j = 0; j < 4; j++) {
            size_t k = i * 4 + j;
            result[i] += a[k] * b[k];
        }
    }
}


int CompareResult(const float* reference, const float* data, size_t array_size, const float epsilon)
{
    int error_count = 0;
    float diff;
    for (size_t i = 0; i < array_size; ++i) {
        diff = fabs(reference[i] - data[i]);
        if (diff > epsilon) {
            ++error_count;
        }
    }
    return error_count;
}

dot_product.cl

/*
 * Copyright 1993-2010 NVIDIA Corporation.  All rights reserved.
 *
 * Please refer to the NVIDIA end user license agreement (EULA) associated
 * with this source code for terms and conditions that govern your use of
 * this software. Any use, reproduction, disclosure, or distribution of
 * this software and related documentation outside the terms of the EULA
 * is strictly prohibited.
 *
 */
 
 __kernel void DotProduct (__global float* a, __global float* b, __global float* c, int iNumElements)
{
    // find position in global arrays
    int iGID = get_global_id(0);

    // bound check (equivalent to the limit on a 'for' loop for standard/serial C code
    if (iGID >= iNumElements)
    {   
        return; 
    }

    // process 
    int iInOffset = iGID << 2;
    c[iGID] = a[iInOffset] * b[iInOffset] 
               + a[iInOffset + 1] * b[iInOffset + 1]
               + a[iInOffset + 2] * b[iInOffset + 2]
               + a[iInOffset + 3] * b[iInOffset + 3];
}