c - 如何创建OpenCL命令队列？

Question

我正在尝试学习 OpenCL，但我什至无法制作一个简单的内核来工作。

下面的代码来自《OpenCL 示例编程》一书，我修改、修改、修改......但我仍然不知道问题出在哪里。

每次我在 PC(AMD Athlon 5350 APU 和 Radeon R3)中执行该程序时，它都会将结果打印为“0.0000”。如果我在另一台带有 NVIDIA 1080 TI 的计算机(该 HD 的克隆，因此一切都相同)中运行相同的可执行文件，程序将输出“3.000”作为结果。

我注意到编译器输出中出现警告，因此我将过时的 clCreateCommandQueue 调用更改为 clCreateCommandQueueWithProperties()。

现在......它只是段错误(通过 printf() 测试，我知道它在 clCreateCommandQueueWithProperties 期间/之后出现段错误)。

在配备 NVIDIA GPU 的系统上，它就可以正常工作。

我错过了什么？

#include <stdint.h>
#include <string.h>
#include <stdlib.h>
#include <stdio.h>
#include <unistd.h>
#include <time.h>
#include <malloc.h>

#include <CL/cl.h>

#define VECTOR_NS 4096
#define VECTOR_SIZE (VECTOR_NS*sizeof(float))

static const char* saxpy_kernel =
    "__kernel void saxpy_kernel(__global float *A, __global float *B, __global float *C)\n"
    "{\n"
        "int index = get_global_id(0);\n"
        "C[0] = 3;\n"
    "}\n"
    ;

int main(void)
{
    int i;
    float total;
    float* A;
    float* B;
    float* C;
    float* Cmapped;
    cl_mem Acl;
    cl_mem Bcl;
    cl_mem Ccl;
    cl_context context;
    cl_platform_id* platforms;
    cl_uint num_platforms;
    cl_uint num_devices;
    cl_command_queue queue;
    cl_kernel kernel;
    cl_int clStatus;

    // Get platform and device information
    platforms = NULL;
    //Set up the Platform
    clStatus = clGetPlatformIDs(0, NULL, &num_platforms);
    platforms = (cl_platform_id *)malloc(sizeof(cl_platform_id)*num_platforms);
    clStatus = clGetPlatformIDs(num_platforms, platforms, NULL);
    //Get the devices list and choose the device you want to run on
    cl_device_id* device_list = NULL;

    clStatus = clGetDeviceIDs( platforms[0], CL_DEVICE_TYPE_GPU, 0, NULL, &num_devices);
    device_list = (cl_device_id *) malloc(sizeof(cl_device_id)*num_devices);

    clStatus = clGetDeviceIDs( platforms[0], CL_DEVICE_TYPE_GPU, num_devices, device_list, NULL);
    // Create one OpenCL context for each device in the platform
    context = clCreateContext( NULL, num_devices, device_list, NULL, NULL, &clStatus);

    /* Create the command queue */
    //queue = clCreateCommandQueue(context, device_list[0], 0, &clStatus);
    queue = clCreateCommandQueueWithProperties(context, device_list[0], NULL, &clStatus);

    if(clStatus != CL_SUCCESS){
        fprintf(stderr, "ERROR: failed to execute the kernel: %d.\n", clStatus);
        exit(1);
    }

    /* */
    if((A = aligned_alloc(sysconf(_SC_PAGESIZE), VECTOR_SIZE)) == NULL){
        fprintf(stderr, "ERROR: failed to allocate memory.\n");
        exit(1);
    }
    if((B = aligned_alloc(sysconf(_SC_PAGESIZE), VECTOR_SIZE)) == NULL){
        fprintf(stderr, "ERROR: failed to allocate memory.\n");
        exit(1);
    }
    if((C = aligned_alloc(sysconf(_SC_PAGESIZE), VECTOR_SIZE)) == NULL){
        fprintf(stderr, "ERROR: failed to allocate memory.\n");
        exit(1);
    }

    /* Initialize it */
    i = 0;
    do {
        A[i] = 1;
        B[i] = 2;
        C[i] = 0;
    } while(++i != VECTOR_NS);

    Acl = clCreateBuffer(context, CL_MEM_USE_HOST_PTR | CL_MEM_READ_WRITE, VECTOR_SIZE, A, &clStatus); // CL_MEM_READ_ONLY
    Bcl = clCreateBuffer(context, CL_MEM_USE_HOST_PTR | CL_MEM_READ_WRITE, VECTOR_SIZE, B, &clStatus); // CL_MEM_READ_ONLY
    Ccl = clCreateBuffer(context, CL_MEM_USE_HOST_PTR | CL_MEM_READ_WRITE, VECTOR_SIZE, C, &clStatus); // CL_MEM_WRITE_ONLY

    // Create a program from the kernel source
    // Build the program
    cl_program program = clCreateProgramWithSource(context, 1, (const char**)&saxpy_kernel, NULL, &clStatus);

    if(clStatus != CL_SUCCESS){
        fprintf(stderr, "ERROR: failed to compile the OpenCL code.\n");
        exit(1);
    }

    clStatus = clBuildProgram(program, 1, device_list, NULL, NULL, NULL);

    if(clStatus != CL_SUCCESS){
        fprintf(stderr, "ERROR: failed to compile the OpenCL code.\n");
        exit(1);
    }

    // Create the OpenCL kernel
    kernel = clCreateKernel(program, "saxpy_kernel", &clStatus);
    // Set the arguments of the kernel
    clSetKernelArg(kernel, 0, sizeof(cl_mem), (void*)&Acl);
    clSetKernelArg(kernel, 1, sizeof(cl_mem), (void*)&Bcl);
    clSetKernelArg(kernel, 2, sizeof(cl_mem), (void*)&Ccl);

    // Execute the OpenCL kernel on the list
    size_t global_size = VECTOR_NS; // Process the entire lists
    size_t local_size = 1;

    // Process one item at a time
    clStatus = clEnqueueNDRangeKernel(queue, kernel, 1, NULL, &global_size, &local_size, 0, NULL, NULL);

    if(clStatus != CL_SUCCESS){
        fprintf(stderr, "ERROR: failed to execute the kernel: %d.\n", clStatus);
        exit(1);
    }

    //* Clean up and wait for all the comands to complete. */
    clFlush(queue);

    /* Display the result to the screen */
    Cmapped = (float*) clEnqueueMapBuffer(queue, Ccl, CL_TRUE, CL_MAP_READ, 0, VECTOR_SIZE, 0, NULL, NULL, &clStatus); // CL_MEM_USE_HOST_PTR

    if(clStatus != CL_SUCCESS){
        fprintf(stderr, "ERROR: failed to execute the kernel: %d.\n", clStatus);
        exit(1);
    }

    total = 0;
    for(i = 0; i < VECTOR_NS; i++)
        total += C[i];

    printf("TOTAL: %f\n", total);

    /* Clean up and wait for all the comands to complete. */
    clFlush(queue);
    clFinish(queue);

    /* Finally release all OpenCL allocated objects and host buffers. */
    clReleaseKernel(kernel);
    clReleaseProgram(program);
    clReleaseMemObject(Acl);
    clReleaseMemObject(Bcl);
    clReleaseMemObject(Ccl);
    clReleaseCommandQueue(queue);
    clReleaseContext(context);

    free(platforms);
    free(device_list);

    return 0;
}

Answer 1

虽然您的代码在我的机器上按原样工作，但我认为可能存在导致您出现问题的问题。通话

Cmapped = (float*)clEnqueueMapBuffer(queue, Ccl, CL_TRUE, CL_MAP_READ, 0, VECTOR_SIZE, 0, NULL, NULL, &clStatus);

更改了Cmapped，但您尝试从原始C缓冲区中读取

total = 0;
for(i = 0; i < VECTOR_NS; i++)
    total += C[i];

这可能应该是这样的

total = 0;
for(i = 0; i < VECTOR_NS; i++)
    total += Cmapped[i];

但是，由于您使用 CL_MEM_USE_HOST_PTR 标志创建了缓冲区，如果您使用 clEnqueueReadBuffer 来代替，您的 OpenCL 驱动程序也应该能够优化复制操作，如果您使用原始指针作为目的地:

clEnqueueReadBuffer(queue, Ccl, CL_TRUE, 0, VECTOR_SIZE, C, 0, NULL, NULL);

如果 OpenCL 实现尚未将数据缓存到设备内存，则这应该是无操作。