0

我有一个愚蠢的问题,我希望有人向我解释一下。我有一个简单的 OpenCL 内核,它只接收一个 double,在内核中将其打印出来,然后将其复制回主机。我注意到,当我在设备上 printf(显然使用 CPU 作为设备而不是 GPU)时,打印的值是浮点精度而不是双精度。OpenCL 中有关于 printf 的内容吗?(我想知道它是否对浮动进行了隐式转换?)这是一些测试代码。(这是在 Macbook Pro(视网膜)、OSX 10.8.4 上运行的)。

#include <stdio.h>
#include <OpenCL/opencl.h>
#include <math.h>
#define CL_CHECK(_expr)                                                      \
do {                                                                         \
    cl_int _err = _expr;                                                     \
    if (_err == CL_SUCCESS)                                                  \
        break;                                                               \
    fprintf(stderr, "OpenCL Error: '%s' returned %d!\n", #_expr, (int)_err); \
    abort();                                                                 \
 } while (0)

 #define CL_CHECK_ERR(_expr)                                                      \
 ({                                                                               \
     cl_int _err = CL_INVALID_VALUE;                                              \
     typeof(_expr) _ret = _expr;                                                  \
     if (_err != CL_SUCCESS) {                                                    \
         fprintf(stderr, "OpenCL Error: '%s' returned %d!\n", #_expr, (int)_err); \
         abort();                                                                 \
     }                                                                            \
     _ret;                                                                        \
 })

void pfn_notify(const char *errinfo, const void *private_info, size_t cb, void *user_data{
    fprintf(stderr, "OpenCL Error (via pfn_notify): %s\n", errinfo);
}

int main(int argc, const char * argv[]){

    cl_platform_id platforms[100];
    cl_uint platforms_n = 0;
    CL_CHECK(clGetPlatformIDs(100, platforms, &platforms_n));
    if (platforms_n == 0)return 1;

    cl_device_id devices[100];
    cl_uint devices_n = 0;
    CL_CHECK(clGetDeviceIDs(platforms[0], CL_DEVICE_TYPE_CPU, 100, devices, &devices_n));
    if (devices_n == 0)return 1;

    cl_context context;
    context = CL_CHECK_ERR(clCreateContext(NULL, 1, devices, &pfn_notify, NULL, &_err));

    double num = M_PI ;
    printf("number before is : %1.17e\n",num);

    const char *program_source[] = {
        "#pragma OPENCL EXTENSION cl_khr_fp64 : enable\n",
        "__kernel void simple_demo(__global double *src, __global double *dst)\n",
        "{\n",
        "   int i = get_global_id(0);\n",
        "   printf(\"src on device is : %1.17e\\n\",src[i]);\n",
        "   dst[i] = src[i];\n",
        "}\n"
     };

     cl_program program;
     program = CL_CHECK_ERR(clCreateProgramWithSource(context, 
         sizeof(program_source)/sizeof(*program_source), program_source, NULL, &_err));
     if (clBuildProgram(program, 1, devices, "", NULL, NULL) != CL_SUCCESS) {
            char buffer[10240];
            clGetProgramBuildInfo(program, devices[0], CL_PROGRAM_BUILD_LOG,
                sizeof(buffer), buffer, NULL);
            fprintf(stderr, "CL Compilation failed:\n%s", buffer);
            abort();
     }

    cl_mem input_buffer;
    input_buffer = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_ONLY, sizeof(double), NULL, &_err));

    cl_mem output_buffer;
    output_buffer = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_WRITE_ONLY, sizeof(double), NULL, &_err));

    cl_kernel kernel;
    kernel = CL_CHECK_ERR(clCreateKernel(program, "simple_demo", &_err));
    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(input_buffer), &input_buffer));
    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(output_buffer), &output_buffer));

    cl_command_queue queue;
    queue = CL_CHECK_ERR(clCreateCommandQueue(context, devices[0], 0, &_err));

    CL_CHECK(clEnqueueWriteBuffer(queue, input_buffer, CL_TRUE, 0, sizeof(double), &num, 0, NULL, NULL));

    cl_event kernel_completion;
    size_t global_work_size[1] = { 1 };
    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 1, NULL, global_work_size, NULL, 0, NULL, &kernel_completion));
    CL_CHECK(clWaitForEvents(1, &kernel_completion));
    CL_CHECK(clReleaseEvent(kernel_completion));

    printf("number after is  :");
    double data;
    CL_CHECK(clEnqueueReadBuffer(queue, output_buffer, CL_TRUE, 0, sizeof(double), &data, 0, NULL, NULL));
    printf(" %1.17e", data);
    printf("\n");

    CL_CHECK(clReleaseMemObject(input_buffer));
    CL_CHECK(clReleaseMemObject(output_buffer));

    CL_CHECK(clReleaseKernel(kernel));
    CL_CHECK(clReleaseProgram(program));
    CL_CHECK(clReleaseContext(context));

    return 0;
}

如果你复制它并编译并运行它,你应该得到:

number before is : 3.14159265358979312e+00
src is           : 3.14159274101257324e+00
number after is  : 3.14159265358979312e+00

有任何想法吗?

4

1 回答 1

1

它看起来像是您正在使用的 OpenCL 库或运行时环境中的错误。它(基本上)在带有 Intel Core-i7 和 AMD A6-3650 的 Windows 上正常工作:

OpenCL 1.2 AMD-APP (1124.2) + AMD A8-3650:
number before is : 3.14159265358979312e+000
src on device is : 3.14159265358979310e+000
number after is  : 3.14159265358979312e+000

Intel OpenCL 1.2 + Intel Core-i7:
number before is : 3.14159265358979312e+000
src on device is : 3.14159265358979310e+00
number after is  : 3.14159265358979312e+000

不知道为什么有些库使用 3 位数字表示“e”格式。C99 规范声明“指数始终包含至少两位数字,并且仅包含表示指数所需的更多数字。”

也不确定为什么打印的值不完全匹配,但至少它们很接近。

于 2013-08-17T03:19:40.930 回答