无法使 cublasSgelsbatched 函数起作用

如何解决无法使 cublasSgelsbatched 函数起作用

我目前正在尝试使 cublasSgelsbatched (https://docs.nvidia.com/cuda/cublas/index.html) 版本正常工作。我首先制作了一个小测试用例，以查看确切需要哪些参数以及如何输入它们。然而，经过多次反复试验，我仍然无法让它工作，我得到 13 的状态返回，这对应于 CUBLAS_STATUS_EXECUTION_Failed 这是一个非常模糊的错误，我还尝试了其他一些 cublas 测试用例，它们似乎工作正常。我也在 MATlab 中测试了输入矩阵，它确实有一个 LS 解决方案。

#include "stdafx.h"
#include "device_launch_parameters.h"

#include <iostream>
#include <stdlib.h>
#include <stdio.h>
#include <cuda_runtime.h>
#include "cublas_v2.h"

#include <algorithm>
#include <cmath>
#include <Windows.h>

int main()
{
    //init id,handle and stat
    int id = cudaGetDevice(&id);
    cublasHandle_t m_cuBLAS;
    cublasstatus_t stat;

    // create handle
    stat = cublasCreate(&m_cuBLAS);

    //params
    const int C = 3; 
    const int M = 2;
    long lda = C;
    long ldb = M;

    //init variables
    float *Amat,*Ymat,*Xmat;
    float *gAmat,*gYmat;

    //allocate mem
    Amat = (float*) malloc(M * C * sizeof(float));
    Ymat = (float*) malloc(C *  sizeof(float));
    Xmat = (float*) malloc(M *  sizeof(float));

    srand(100);

    for (int i = 0; i < C * M; i++) {
        Amat[i] = rand() % 10 + 1;
        Amat[i] = (float)Amat[i];
    }

    for (int i = 0; i < C; i++) {
        Ymat[i] =  rand() % 10 + 1;
        Ymat[i] = (float)Ymat[i];
    }

    //allocate mem
    cudamalloc( &gAmat,M * C * sizeof(float));
    cudamalloc( &gYmat,C * sizeof(float));

    //copy mem
    cudamemcpy(gAmat,Amat,M * C * sizeof(float),cudamemcpyHostToDevice);
    cudamemcpy(gYmat,Ymat,C * 1 * sizeof(float),cudamemcpyHostToDevice);

    //init info params
    int info = 0;
    int devInfoArray[1] = { 0 };

    //Synchronize (not necesarry I think,but just to test)
    cudaDeviceSynchronize();
    

    //run cublas
    cublasstatus_t status = cublasSgelsBatched(m_cuBLAS,CUBLAS_OP_N,C,M,1,&gAmat,lda,//or 1
        &gYmat,&info,NULL,1);
    
    //Output info
    std::cout << "status = " << status << std::endl;
    std::cout << "info = " << info << std::endl;
    std::cout << "devInfoArray = " << devInfoArray[0] << std::endl;

    cudamemcpy(Xmat,gYmat,cudamemcpyDevicetoHost);

    //Output printed
    std::cout << Xmat[0] << "," << Xmat[1] << "," << Xmat[2] << std::endl;

    //free memory
    free(Amat);
    free(Ymat);
    free(Xmat);

    cudaFree(gAmat);
    cudaFree(gYmat);

    //destory handle
    cublasDestroy(m_cuBLAS);

    return 0;
}

我在使用 CUDA 9.0 在 MVS 中运行的 Windows 10 上

非常感谢您的帮助

解决方法

正如评论中所指出的，您没有在设备上创建正确的指针数组。 batched function 使用位于设备内存中的指针数组，用于数据参数，例如：

Aarray device 输入/输出指向数组的指针数组，每个数组都是暗的。 m x n 且 lda>=max(1,m)。矩阵 Aarray[i] 不应重叠；否则，会出现未定义的行为。

例如传递 &gAmat 似乎满足类型要求，但是那个指针不指向设备内存。

对您的代码的以下修改侧重于正确处理 gAmat 和 gYmat 对我来说似乎没有错误：

$ cat t130.cu
#include <iostream>
#include <stdlib.h>
#include <stdio.h>
#include <cuda_runtime.h>
#include <cublas_v2.h>

#include <algorithm>
#include <cmath>

int main()
{
    //init id,handle and stat
    int id = cudaGetDevice(&id);
    cublasHandle_t m_cuBLAS;
    cublasStatus_t stat;

    // create handle
    stat = cublasCreate(&m_cuBLAS);

    //params
    const int C = 3;
    const int M = 2;
    long lda = C;
    long ldb = M;

    //init variables
    float *Amat,*Ymat,*Xmat;
    float *gAmat,*gYmat;

    //allocate mem
    Amat = (float*) malloc(M * C * sizeof(float));
    Ymat = (float*) malloc(C *  sizeof(float));
    Xmat = (float*) malloc(M *  sizeof(float));

    srand(100);

    for (int i = 0; i < C * M; i++) {
        Amat[i] = rand() % 10 + 1;
        Amat[i] = (float)Amat[i];
    }

    for (int i = 0; i < C; i++) {
        Ymat[i] =  rand() % 10 + 1;
        Ymat[i] = (float)Ymat[i];
    }

    //allocate mem
    cudaMalloc( &gAmat,M * C * sizeof(float));
    cudaMalloc( &gYmat,C * sizeof(float));

    //copy mem
    cudaMemcpy(gAmat,Amat,M * C * sizeof(float),cudaMemcpyHostToDevice);
    cudaMemcpy(gYmat,Ymat,C * 1 * sizeof(float),cudaMemcpyHostToDevice);
    float **ggAmat,**ggYmat;
    cudaMalloc(&ggAmat,sizeof(float*));
    cudaMalloc(&ggYmat,sizeof(float*));
    cudaMemcpy(ggAmat,&gAmat,sizeof(float*),cudaMemcpyHostToDevice);
    cudaMemcpy(ggYmat,&gYmat,cudaMemcpyHostToDevice);
    //init info params
    int info = 0;
    int devInfoArray[1] = { 0 };

    //Synchronize (not necesarry I think,but just to test)
    cudaDeviceSynchronize();


    //run cublas
    cublasStatus_t status = cublasSgelsBatched(m_cuBLAS,CUBLAS_OP_N,C,M,1,ggAmat,lda,//or 1
        ggYmat,&info,NULL,1);

    //Output info
    std::cout << "status = " << status << std::endl;
    std::cout << "info = " << info << std::endl;
    std::cout << "devInfoArray = " << devInfoArray[0] << std::endl;

    cudaMemcpy(Xmat,gYmat,cudaMemcpyDeviceToHost);

    //Output printed
    std::cout << Xmat[0] << "," << Xmat[1] << "," << Xmat[2] << std::endl;

    //free memory
    free(Amat);
    free(Ymat);
    free(Xmat);

    cudaFree(gAmat);
    cudaFree(gYmat);

    //destory handle
    cublasDestroy(m_cuBLAS);

    return 0;
}
$ nvcc -o t130 t130.cu -lcublas
t130.cu(15): warning: variable "stat" was set but never used

t130.cu(24): warning: variable "ldb" was declared but never referenced

$ cuda-memcheck ./t130
========= CUDA-MEMCHECK
status = 0
info = 0
devInfoArray = 0
-0.0226168,0.514827,-4.29722
========= ERROR SUMMARY: 0 errors
$

您的代码仅显示一个数组。如果您有一批数组，您将为 A 和 Y 中的每一个传递一个实际的设备分配指针数组。

根据下面的评论，这里是使用非随机输入的代码版本：

$ cat t130.cu
#include <iostream>
#include <stdlib.h>
#include <stdio.h>
#include <cuda_runtime.h>
#include <cublas_v2.h>

#include <algorithm>
#include <cmath>

int main()
{
    //init id,handle and stat
    int id = cudaGetDevice(&id);
    cublasHandle_t m_cuBLAS;
    cublasStatus_t status;
    // create handle
    status = cublasCreate(&m_cuBLAS);
    std::cout << "status = " << status << std::endl;

    //params
    const int C = 3;
    const int M = 2;
    long lda = C;

    //init variables
    float *Amat,*gYmat;

    //allocate mem
    Amat = (float*) malloc(M * C * sizeof(float));
    Ymat = (float*) malloc(C *  sizeof(float));
    Xmat = (float*) malloc(M *  sizeof(float));

    srand(100);
#if 0
    for (int i = 0; i < C * M; i++) {
        Amat[i] = rand() % 10 + 1;
        Amat[i] = (float)Amat[i];
    }

    for (int i = 0; i < C; i++) {
        Ymat[i] =  rand() % 10 + 1;
        Ymat[i] = (float)Ymat[i];
    }
#endif
    Amat[0] = 6;
    Amat[1] = 7;
    Amat[2] = 6;
    Amat[3] = 5;
    Amat[4] = 5;
    Amat[5] = 5;
    Ymat[0] = 9;
    Ymat[1] = 3;
    Ymat[2] = 10;
    //allocate mem
    cudaMalloc( &gAmat,but just to test)
    cudaDeviceSynchronize();


    //run cublas
    status = cublasSgelsBatched(m_cuBLAS," << Xmat[2] << std::endl;

    //free memory
    free(Amat);
    free(Ymat);
    free(Xmat);

    cudaFree(gAmat);
    cudaFree(gYmat);

    //destory handle
    cublasDestroy(m_cuBLAS);

    return 0;
}
$ nvcc -o t130 t130.cu -lcublas
$ cuda-memcheck ./t130
========= CUDA-MEMCHECK
status = 0
status = 0
info = 0
devInfoArray = 0
-6.5,9.7,0.707106
========= ERROR SUMMARY: 0 errors
$