Question

我正在尝试学习如何在本教程中使用OpenCL https://anteru.net/blog/2012/11/04/2016/index.html但是我不认为浮动缓冲区中的值被设置为任何值。当我在最后读取缓冲区时，它全部为0，带有科学记数法的十进制数字，就像它充满了随机存储器。我将发布以下代码。内核做的是接受3个参数，float buffer x，float buffer y和float a。 const int i = get_global_id(0)和y[i] += a * x[i]; 问题是（我认为）我从来没有在aBuffer或bBuffer中放任何数字，因此乘法和加法毫无意义。但奇怪的是，当我让内核执行此操作y[i] += a;时，我认为它会给出相同的输出。

main.cpp中：

#include <iostream>
#include <vector>


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


using namespace std;




int main(int argc, const char * argv[]) {
    cl_uint platformIdCount = 0;
    clGetPlatformIDs(0, nullptr, &platformIdCount);

    vector<cl_platform_id> platformIds(platformIdCount);
    clGetPlatformIDs(platformIdCount, platformIds.data(), nullptr);
    cout << "Platforms " << platformIdCount << endl;

    cl_uint deviceIdCount = 0;


    clGetDeviceIDs(platformIds[0], CL_DEVICE_TYPE_GPU, 0, nullptr, &deviceIdCount);
    cout << "Devices " << deviceIdCount << endl;

    vector<cl_device_id> deviceIds(deviceIdCount);
    clGetDeviceIDs(platformIds[0], CL_DEVICE_TYPE_GPU, deviceIdCount, deviceIds.data(), nullptr);


    const cl_context_properties contextProperties[] = {

        CL_CONTEXT_PLATFORM,
        reinterpret_cast<cl_context_properties>(platformIds[0]),
    0,0
};
    cl_int error = 0;
    cl_context context = clCreateContext(contextProperties, deviceIdCount, deviceIds.data(), nullptr, nullptr, &error);

    error = 0;

    cl_mem aBuffer = clCreateBuffer(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, sizeof(float) * (64), nullptr, &error);
    cl_mem bBuffer = clCreateBuffer(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, sizeof(float) * (64), nullptr, &error);



    cl_program program;

    clBuildProgram(program, deviceIdCount, deviceIds.data(), nullptr, nullptr, nullptr);

    cl_kernel kernel1 = clCreateKernel(program, "SAXPY", &error);

    clSetKernelArg(kernel1, 0, sizeof(cl_mem), aBuffer);
    clSetKernelArg(kernel1, 1, sizeof(cl_mem), bBuffer);
    static const float two = 2.0f;
    clSetKernelArg(kernel1, 2, sizeof(float),&two);

    const size_t globalWorkSize [] = {64,0,0};
    cl_command_queue queue;
    clEnqueueNDRangeKernel(queue, kernel1, 1, nullptr, globalWorkSize, nullptr, 0, nullptr, nullptr);



    float done[64];
    clEnqueueReadBuffer(queue, bBuffer, CL_TRUE, 0, sizeof(float)*64, done, 0, nullptr, nullptr);
    for (int a = 0; a < 64; a++) {
        cout << done[a] << endl;
    }




    clReleaseContext(context);
    return 0;
}

.cl文件：

kernel void SAXPY(__global float* x,__global float* y, float a){
    const int i = get_global_id(0);
    //y[i] = 2.0f;
    y[i] += a * x[i];
}

Answer 1

首先，在设置内核参数时，必须pass pointer到内存对象：

clSetKernelArg(kernel1, 0, sizeof(cl_mem), &aBuffer); // &aBuffer, not aBuffer
clSetKernelArg(kernel1, 1, sizeof(cl_mem), &bBuffer); // &bBuffer, not bBuffer

其次，你没有create command queue：

cl_command_queue queue = clCreateCommandQueue(context, deviceIds[0], 0, nullptr);

第三，你没有在clCreateProgramWithSource()之前致电clBuildProgram()。

此外，尝试初始化cl_mem对象：

cl_float* mem = (cl_float*) malloc(sizeof(cl_float)*64);
for(int i=0; i<64; i++)
    mem[i] = i;

cl_mem aBuffer = clCreateBuffer(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, sizeof(cl_float) * (64), mem, &error);
cl_mem bBuffer = clCreateBuffer(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, sizeof(cl_float) * (64), mem, &error);

free(mem);

固定代码：

#include <iostream>
#include <vector>

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

using namespace std;


int main(int argc, const char * argv[]) {
    cl_uint platformIdCount = 0;
    clGetPlatformIDs(0, nullptr, &platformIdCount);

    vector<cl_platform_id> platformIds(platformIdCount);
    clGetPlatformIDs(platformIdCount, platformIds.data(), nullptr);

    cl_uint deviceIdCount = 0;
    clGetDeviceIDs(platformIds[0], CL_DEVICE_TYPE_GPU, 0, nullptr, &deviceIdCount);

    vector<cl_device_id> deviceIds(deviceIdCount);
    clGetDeviceIDs(platformIds[0], CL_DEVICE_TYPE_GPU, deviceIdCount, deviceIds.data(), nullptr);

    const cl_context_properties contextProperties[] = {

        CL_CONTEXT_PLATFORM,
        (cl_context_properties)platformIds[0],
        0
    };

    cl_int error = 0;
    cl_context context = clCreateContext(contextProperties, 1, &deviceIds[0], [](const char* errinfo, const void* private_info, size_t cb, void* user_data) -> void {
        /* context-creation and runtime error handler */
        cout << "Context error: " << errinfo << endl;
    }, nullptr, &error);


    cl_float* mem = (cl_float*) malloc(sizeof(cl_float)*64);
    for(int i=0; i<64; i++)
        mem[i] = i;

    cl_mem aBuffer = clCreateBuffer(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, sizeof(cl_float) * (64), mem, &error);
    cl_mem bBuffer = clCreateBuffer(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, sizeof(cl_float) * (64), mem, &error);

    free(mem);

    cl_program program;
    string src = "__kernel void SAXPY(__global float* x, __global float* y, float a){"
                    "size_t i=get_global_id(0);"
                    "y[i]=a*x[i];"
                 "}";

    const char* sources[] = {src.c_str()};
    const size_t lens[] = {src.length()};

    program = clCreateProgramWithSource(context, 1, sources, lens, &error);
    clBuildProgram(program, 1, &deviceIds[0], nullptr, nullptr, nullptr);

    cl_kernel kernel1 = clCreateKernel(program, "SAXPY", &error);
    clSetKernelArg(kernel1, 0, sizeof(cl_mem), &aBuffer);
    clSetKernelArg(kernel1, 1, sizeof(cl_mem), &bBuffer);
    static const float two = 2.0f;
    clSetKernelArg(kernel1, 2, sizeof(float),&two);

    const size_t globalWorkSize [] = {64,0,0};
    cl_command_queue queue = clCreateCommandQueue(context, deviceIds[0], 0, nullptr);

    clEnqueueNDRangeKernel(queue, kernel1, 1, nullptr, globalWorkSize, nullptr, 0, nullptr, nullptr);

    float done[64];
    clEnqueueReadBuffer(queue, bBuffer, CL_TRUE, 0, sizeof(float)*64, done, 0, nullptr, nullptr);
    for (int a = 0; a < 64; a++)
        cout << done[a] << endl;

    clReleaseContext(context);
    return 0;
}

OpenCL缓冲区为空

1 个答案: