opencl - 在多个 GPU 上运行 OpenCL 内核？

Question

现在我编写了几个算法，在一个 GPU 上并行运行，但是当我尝试在多个 GPU(例如 3)上执行它们时，它们都有相同的问题。问题在于，在一个 GPU 上执行的代码在 3 个 GPU 上执行的时间完全相同(不是更快)。我尝试使用更多数据执行，尝试执行不同的任务，但没有任何帮助。最后，我最终尝试运行最简单的任务，如 elements sum，但仍然犯了这个可怕的错误。这就是为什么我不认为这是特定算法的问题，而且我觉得我的代码中存在错误(甚至是我在多个 GPU 上并行化代码的方法)。

这是我的 Parallel.cpp 类的头文件:

#ifndef PARALLEL_H
#define PARALLEL_H

#define __NO_STD_VECTOR // Use cl::vector and cl::string and
#define __NO_STD_STRING // not STL versions, more on this later
#include <CL/cl.h>

class Parallel
{
    public:
        Parallel();
        int executeAttachVectorsKernel(int*, int*, int*, int);
        static void getMaxWorkGroupSize(int*, int*, int*);
        virtual ~Parallel();
    protected:
    private:
        char* file_contents(const char*, int*);
        void getShortInfo(cl_device_id);
        int init(void);
        cl_platform_id platform;
        cl_device_id* devices;
        cl_uint num_devices;
        cl_command_queue* queues;
        int* WGSizes;
        int* WGNumbers;
        cl_context context;
        cl_program program;
        cl_kernel kernel;
        cl_mem input1;
        cl_mem input2;
        cl_mem output;
};

#endif // PARALLEL_H

这里是初始化方法init:

int Parallel::init() {
cl_int err;

//Connect to the first platfrom
err = clGetPlatformIDs(1, &platform, NULL);
if (err != CL_SUCCESS) {
    cerr << "Error occured while executing clGetPlatformIDs" << endl;
    return EXIT_FAILURE;
}

//Get devices number
err = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 0, NULL, &num_devices);
if (err != CL_SUCCESS) {
    cerr << "Error: Failed to create a device group:" << endl;
    return EXIT_FAILURE;
}

cout << "NUM DEVICES =" << num_devices << endl;

devices = new cl_device_id[num_devices];
//Get all the GPU devices
err = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, num_devices, devices, NULL);

//Create one context for all the devices
context = clCreateContext(NULL, num_devices, devices, NULL, NULL, &err);
if (!context) {
    cerr << "Error: Failed to create a compute context!" << endl;
    return EXIT_FAILURE;
}

queues = new cl_command_queue[num_devices];
WGNumbers = new int[num_devices];
WGSizes = new int[num_devices];


for(int i = 0; i < num_devices; i++) {
    //Create a command queue for every device
    queues[i] = clCreateCommandQueue(context, devices[i], 0, &err);
    if (!queues[i]) {
        cerr << "Error: Failed to create a command commands!" << endl;
        return EXIT_FAILURE;
    }

    cl_ulong temp;
    clGetDeviceInfo(devices[i], CL_DEVICE_MAX_WORK_GROUP_SIZE, sizeof(temp), &temp, NULL);
    WGSizes[i] = (int)temp;

    clGetDeviceInfo(devices[i], CL_DEVICE_MAX_WORK_ITEM_SIZES, sizeof(temp), &temp, NULL);
    WGNumbers[i] = (int)temp;
}

//Translate kernel code into chars
int pl;
size_t program_length;
string path = "./kernel/kernel_av.cl";

char* cSourceCL = file_contents(path.c_str(), &pl);
program_length = (size_t)pl;

//Create a program
program = clCreateProgramWithSource(context, 1,
                  (const char **) &cSourceCL, &program_length, &err);

if (!program) {
    cerr << "Error: Failed to create compute program!" << endl;
    return EXIT_FAILURE;
}

//Create an executable
err = clBuildProgram(program, 0, NULL, NULL, NULL, NULL);
if (err != CL_SUCCESS)
{
    size_t len;
    char buffer[2048];

    cerr << "Error: Failed to build program executable!" << endl;
    exit(1);
}

// Create the compute kernel in the program
kernel = clCreateKernel(program, "calculate2dim", &err);
if (err != CL_SUCCESS)
{
    cerr << "Error: Failed to create compute kernel!" << endl;
    exit(1);
}
}

执行内核的方法在这里:

int Parallel::executeAttachVectorsKernel(int* data1, int* data2, int* results, int vectors_num) {

cl_int err;
size_t global;  // global domain size for our calculation
size_t local;   // local domain size for our calculation

int partition = vectors_num/num_devices;
unsigned int count = partition;
input1 = clCreateBuffer(context, CL_MEM_READ_ONLY, sizeof(int) * count, NULL, NULL);
input2 = clCreateBuffer(context, CL_MEM_READ_ONLY, sizeof(int) * count, NULL, NULL);
output = clCreateBuffer(context, CL_MEM_WRITE_ONLY, sizeof(int) * count, NULL, NULL);
if (!input1 || !input2 || !output) {
    cerr << "Error: Failed to allocate device memory!" << endl;
    exit(1);
}

int** data1_apart = new int*[num_devices];
int** data2_apart = new int*[num_devices];
int** results_apart = new int*[num_devices];

for(int i = 0; i < num_devices; i++) {
    cout << "Executing parallel part on GPU " << i + 1 << endl;
    cout << "Partition size = " << partition << endl;
    data1_apart[i] = new int[partition];
    data2_apart[i] = new int[partition];
    results_apart[i] = new int[partition];

    for(int j = i*partition, k = 0; k < partition; j++, k++) {
        data1_apart[i][k] = data1[j];
        data2_apart[i][k] = data2[j];
    }

    //Transfer the input vector into device memory
    err = clEnqueueWriteBuffer(queues[i], input1,
                               CL_TRUE, 0, sizeof(int) * count,
                               data1_apart[i], 0, NULL, NULL);

    err = clEnqueueWriteBuffer(queues[i], input2,
                               CL_TRUE, 0, sizeof(int) * count,
                               data2_apart[i], 0, NULL, NULL);

    if (err != CL_SUCCESS)
    {
        cerr << "Error: Failed to write to source array!" << endl;
        exit(1);
    }

    int parameter4 = count/WGNumbers[i];

     //Set the arguments to the compute kernel
    err = 0;
    err  = clSetKernelArg(kernel, 0, sizeof(cl_mem), &input1);
    err |= clSetKernelArg(kernel, 1, sizeof(cl_mem), &input2);
    err |= clSetKernelArg(kernel, 2, sizeof(cl_mem), &output);
    err |= clSetKernelArg(kernel, 3, sizeof(int), &parameter4);
    if (err != CL_SUCCESS)
    {
        cerr << "Error: Failed to set kernel arguments! " << err << endl;
        exit(1);
    }

    global = WGNumbers[i];
    local = WGSizes[i];

    if(local > global) {
        local = global;
    }
    cout << "global = " << global << " local = " << local << endl;

    err = clEnqueueNDRangeKernel(queues[i], kernel,
                                 1, NULL, &global, &local,
                                 0, NULL, NULL);
    if (err)
    {
        cerr << "Error: Failed to execute kernel!" << endl;
        return EXIT_FAILURE;
    }
}

for(int i = 0; i < num_devices; i++) {
    //Wait for all commands to complete
    clFinish(queues[i]);

    //Read back the results from the device to verify the output

    err = clEnqueueReadBuffer(queues[i], output,
                               CL_TRUE, 0, sizeof(int) * count,
                               results_apart[i], 0, NULL, NULL );
    if (err != CL_SUCCESS)
    {
        cerr << "Error: Failed to read output array! " <<  err << endl;
        exit(1);
    }

    for(int j = 0; j < partition; j++) {
        results[i*partition + j] = results_apart[i][j];
    }

    delete [] data1_apart[i];
    delete [] data2_apart[i];
    delete [] results_apart[i];
}

clReleaseMemObject(input1);
clReleaseMemObject(input2);
clReleaseMemObject(output);
delete [] data1_apart;
delete [] data2_apart;
}

在将这个问题发布到 stackoverflow 之前，我已经与这个问题抗争了 2-3 周，现在我真的需要有人的帮助，所以我将非常感谢任何想法和答案!

Answer 1

这是我认为正在发生的事情。您为每个参与的 opencl 设备调用一次 clEnqueueNDRangeKernel。此时，没有一个内核开始执行，因为还没有调用 clFlush。接下来，您为每个队列创建一个 clFinish。第一个 clFinish 调用导致第一个排队的工作组运行。它还等待它完成。第一个工作组完成后，clFinish 会将控制权返回给您的应用程序。然后，您的应用会为下一个队列调用 clFinish。这会触发第二个工作灌浆运行，并等待它完成。所以工作是按顺序进行的。解决方案可能与在每次调用 clEnqueueNDRangeKernel 后立即调用 clFush 一样简单。这就是我的 AMD 系统的行为方式。我将很快发布一个工作示例。