1

我有一个内核执行问题，当我使用大型数组（1000 x 10000）时，它不会在正确的位置写入信息。但是对于小数组，没有问题，我检索到正确的结果。对于内核执行，我使用 ATI Mobility RADEON HD 4300 系列的 GPU。

C代码示例是：

#include <stdio.h>
#include <stdlib.h>

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

#define MAX_SOURCE_SIZE (0x100000)
#define MAX_SIZE 108
#define NCOLS 1000
#define NROWS 10000

int main(void) {
    char* source_name = "mykernel.cl";
    char* source_code;
    size_t source_size;
    cl_platform_id platformId = NULL;
    cl_uint nbplatforms;
    cl_device_id deviceId = NULL;
    cl_uint nbdevices;
    cl_context context = NULL;
    cl_int errcode;
    cl_command_queue commandQueue = NULL;
    cl_program program;
    size_t global_work_size[2];
    size_t local_work_size[2];

    FILE* fh;

    //Retrieving platform information
    errcode = clGetPlatformIDs(1, &platformId, &nbplatforms);

    //Retrieving device (GPU) information
    errcode = clGetDeviceIDs(platformId, CL_DEVICE_TYPE_GPU, 1, &deviceId, &nbdevices);

    //Creation of a working context
    context = clCreateContext(NULL, 1, &deviceId, NULL, NULL, &errcode);

    commandQueue = clCreateCommandQueue(context, deviceId, 0, &errcode);

    //Opening and reading the kernel source file
    if((fh = fopen(source_name, "r")) == NULL){
        fprintf(stderr, "Failed to open the file containing the kernel source !\n");
        exit(EXIT_FAILURE);
    }
    source_code = (char*) malloc (MAX_SOURCE_SIZE * sizeof(char));
    source_size = fread(source_code, sizeof(char), MAX_SOURCE_SIZE, fh);
    fclose(fh);

    program = clCreateProgramWithSource(context, 1, (const char**) &source_code, (const size_t*) &source_size, &errcode);

    //Building kernel
    errcode = clBuildProgram(program, 1, &deviceId, NULL, NULL, NULL);

    //Creation of the kernel program
    cl_kernel kernel = clCreateKernel(program, "mykernel", &errcode);
    unsigned int *op1 = (unsigned int*) malloc (NCOLS * NROWS * sizeof(unsigned int));

    cl_mem op1buff = clCreateBuffer(context, CL_MEM_WRITE_ONLY, NCOLS * NROWS * sizeof(unsigned int), NULL, &errcode);

    clSetKernelArg(kernel, 0, sizeof(cl_mem), (void*) &op1buff);

    global_work_size[0] =  NCOLS;
    global_work_size[1] =  NROWS;
    local_work_size[0] = NCOLS;
    local_work_size[1] = 1;

    clEnqueueNDRangeKernel(commandQueue, kernel, 2, NULL, global_work_size, local_work_size, 0, NULL, NULL);

    errcode = clEnqueueReadBuffer(commandQueue, op1buff, CL_TRUE, 0, NCOLS * NROWS * sizeof(unsigned int), (void*)op1, 0, NULL, NULL);

    for(int i = 0; i < NROWS; i++){
        for(int j = 0; j < NCOLS; j++)
            printf("[index:%d - %u] ", i*NCOLS+j, op1[i*NCOLS+j]);
        printf("\n");
    }

    return EXIT_SUCCESS;
}

内核源代码放在名为 mykernel.cl 的文件中，如下所示：

__kernel void mykernel(__global unsigned int* op1buf){
    unsigned int index = get_group_id(1) * get_global_size(0) + get_local_id(0);
    op1buf[index] = index;
}

当我使用大型数组时，该程序的执行会返回从数组中读取的意外值。例如 ：

[index:0 - 16777215] [index:1 - 16777215] [index:2 - 16777215] [index:3 - 16777215] ...
[index:1000 - 3438339071] [index:1001 - 3941660159] [index:1002 - 1650092117] [index:1003 - 2529976771] ...
[index:1000 - 3438339071] [index:1001 - 3941660159] [index:1002 - 1650092117] [index:1003 - 2529976771] ...
[index:3000 - 16777215] [index:3001 - 16777215] [index:3002 - 16777215] [index:3003 - 16777215] ...
[index:4000 - 3438339071] [index:4001 - 3941660159] [index:4002 - 1650092117] [index:4003 - 2529976771] ...
....

我的代码可能有什么问题，或者我没有考虑到使用 GPU 的问题？

提前致谢。

4

1 回答 1

1

1000 显然对于您的设备来说太大了。使用 clGetDeviceInfo 和 CL_DEVICE_MAX_WORK_GROUP_SIZE 来确定可以使用的最大值。

于 2012-05-30T20:32:51.480 回答