OpenCL基础示例 - 向量加法解释

# 对于NVIDIA GPU
sudo apt-get install nvidia-opencl-dev
 
# 对于AMD GPU
sudo apt-get install amdgpu-pro-opencl-dev
 
# 对于Intel GPU/CPU
sudo apt-get install intel-opencl-icd
 
# 通用OpenCL头文件
sudo apt-get install opencl-headers ocl-icd-opencl-dev

# 安装OpenCL头文件和加载器
sudo yum install opencl-headers ocl-icd-devel
 
# 对于NVIDIA GPU，安装CUDA工具包
# 对于AMD GPU，安装ROCm

make 15-opencl-vector-addition

./15-opencl-vector-addition

#ifdef __APPLE__
#include <OpenCL/opencl.h>
#else
#include <CL/cl.h>
#endif

const char* kernelSource = 
"__kernel void vectorAdd(__global const float* A,\n"
"                       __global const float* B,\n"
"                       __global float* C,\n"
"                       const int numElements) {\n"
"    int i = get_global_id(0);\n"
"    if (i < numElements) {\n"
"        C[i] = A[i] + B[i];\n"
"    }\n"
"}\n";

void checkError(cl_int error, const char* operation) {
    if (error != CL_SUCCESS) {
        printf("Error during %s: %d\n", operation, error);
        exit(1);
    }
}

// 获取平台
ret = clGetPlatformIDs(1, &platform_id, &ret_num_platforms);
checkError(ret, "getting platform IDs");
 
// 获取设备（优先选择GPU，回退到任何设备类型）
ret = clGetDeviceIDs(platform_id, CL_DEVICE_TYPE_GPU, 1, &device_id, &ret_num_devices);
if (ret != CL_SUCCESS) {
    printf("未找到GPU，尝试任何设备类型...\n");
    ret = clGetDeviceIDs(platform_id, CL_DEVICE_TYPE_ALL, 1, &device_id, &ret_num_devices);
    checkError(ret, "getting device IDs");
}

// 创建OpenCL上下文
cl_context context = clCreateContext(NULL, 1, &device_id, NULL, NULL, &ret);
checkError(ret, "creating context");
 
// 创建命令队列
cl_command_queue command_queue = clCreateCommandQueue(context, device_id, 0, &ret);
checkError(ret, "creating command queue");

// 在设备上创建内存缓冲区
cl_mem d_A = clCreateBuffer(context, CL_MEM_READ_ONLY, dataSize, NULL, &ret);
cl_mem d_B = clCreateBuffer(context, CL_MEM_READ_ONLY, dataSize, NULL, &ret);
cl_mem d_C = clCreateBuffer(context, CL_MEM_WRITE_ONLY, dataSize, NULL, &ret);
 
// 将数据复制到设备缓冲区
ret = clEnqueueWriteBuffer(command_queue, d_A, CL_TRUE, 0, dataSize, h_A, 0, NULL, NULL);
ret = clEnqueueWriteBuffer(command_queue, d_B, CL_TRUE, 0, dataSize, h_B, 0, NULL, NULL);

// 从源码创建程序
cl_program program = clCreateProgramWithSource(context, 1, &kernelSource, NULL, &ret);
checkError(ret, "creating program");
 
// 构建程序
ret = clBuildProgram(program, 1, &device_id, NULL, NULL, NULL);
if (ret != CL_SUCCESS) {
    // 获取构建日志以调试
    size_t log_size;
    clGetProgramBuildInfo(program, device_id, CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size);
    char *log = (char*)malloc(log_size);
    clGetProgramBuildInfo(program, device_id, CL_PROGRAM_BUILD_LOG, log_size, log, NULL);
    printf("Build log:\n%s\n", log);
    free(log);
    exit(1);
}

// 创建内核
cl_kernel kernel = clCreateKernel(program, "vectorAdd", &ret);
 
// 设置内核参数
ret = clSetKernelArg(kernel, 0, sizeof(cl_mem), (void*)&d_A);
ret = clSetKernelArg(kernel, 1, sizeof(cl_mem), (void*)&d_B);
ret = clSetKernelArg(kernel, 2, sizeof(cl_mem), (void*)&d_C);
ret = clSetKernelArg(kernel, 3, sizeof(int), (void*)&numElements);
 
// 执行内核
size_t globalWorkSize = numElements;
size_t localWorkSize = 256; // 工作组大小
 
// 调整全局工作大小为局部工作大小的倍数
if (globalWorkSize % localWorkSize != 0) {
    globalWorkSize = ((globalWorkSize / localWorkSize) + 1) * localWorkSize;
}
 
ret = clEnqueueNDRangeKernel(command_queue, kernel, 1, NULL, &globalWorkSize, &localWorkSize, 0, NULL, NULL);

// 等待内核完成
ret = clFinish(command_queue);
checkError(ret, "waiting for kernel to finish");
 
// 将结果读回主机
ret = clEnqueueReadBuffer(command_queue, d_C, CL_TRUE, 0, dataSize, h_C, 0, NULL, NULL);