CUBLAS矩阵乘法

时间:2011-04-05 11:23:32

标签: cuda matrix-multiplication blas cublas

用CUDA实现矩阵乘法后。我试图用CUBLAS实现它(感谢论坛中一些人的建议)。

我可以乘以平方矩阵但是(是的再次......)我在使用非方矩阵时遇到了困难。唯一可行的非方矩阵乘法类型是改变矩阵A的宽度(A * B = C)。

我没有得到任何错误,但结果矩阵返回错误的值。这是我的代码(它基本上是simpleCUBLAS SDK示例的改编版):

#include <stdlib.h>
#include <stdio.h>
#include "cublas.h"
#define HA 2
#define WA 9
#define WB 2
#define HB WA 
#define WC WB   
#define HC HA  
#define index(i,j,ld) (((j)*(ld))+(i))

void printMat(float*P,int uWP,int uHP){
//printf("\n %f",P[1]);
int i,j;
for(i=0;i<uHP;i++){

    printf("\n");

    for(j=0;j<uWP;j++)
        printf("%f ",P[index(i,j,uHP)]);
        //printf("%f ",P[i*uWP+j]);
}
}




 int  main (int argc, char** argv) {
    cublasStatus status;
        int i,j;
        cublasInit();

        float *A = (float*)malloc(HA*WA*sizeof(float));
        float *B = (float*)malloc(HB*WB*sizeof(float));
        float *C = (float*)malloc(HC*WC*sizeof(float));
    if (A == 0) {
        fprintf (stderr, "!!!! host memory allocation error (A)\n");
        return EXIT_FAILURE;
    }
    if (B == 0) {
        fprintf (stderr, "!!!! host memory allocation error (A)\n");
        return EXIT_FAILURE;
    }
    if (C == 0) {
        fprintf (stderr, "!!!! host memory allocation error (A)\n");
        return EXIT_FAILURE;
        }


        for (i=0;i<HA;i++)
     for (j=0;j<WA;j++)
        A[index(i,j,HA)] = (float) index(i,j,HA);   
        for (i=0;i<HB;i++)
     for (j=0;j<WB;j++)
        B[index(i,j,HB)] = (float) index(i,j,HB); 
    /*
    for (i=0;i<HA*WA;i++)
    A[i]=(float) i;
    for (i=0;i<HB*WB;i++)
    B[i]=(float) i;         */  


        float* AA; float* BB; float* CC;

    /*ALLOCATE ON THE DEVICE*/
    status=cublasAlloc(HA*WA,sizeof(float),(void**)&AA);
        if (status != CUBLAS_STATUS_SUCCESS) {
        fprintf (stderr, "!!!! device memory allocation error (A)\n");
        return EXIT_FAILURE;
        }

        status=cublasAlloc(HB*WB,sizeof(float),(void**)&BB);
        if (status != CUBLAS_STATUS_SUCCESS) {
        fprintf (stderr, "!!!! device memory allocation error (A)\n");
        return EXIT_FAILURE;
        }

        status=cublasAlloc(HC*WC,sizeof(float),(void**)&CC);
        if (status != CUBLAS_STATUS_SUCCESS) {
        fprintf (stderr, "!!!! device memory allocation error (A)\n");
        return EXIT_FAILURE;
        }

    /*SET MATRIX*/
        status=cublasSetMatrix(HA,WA,sizeof(float),A,HA,AA,HA);
        if (status != CUBLAS_STATUS_SUCCESS) {
        fprintf (stderr, "!!!! device memory allocation error (A)\n");
        return EXIT_FAILURE;
        }

        status=cublasSetMatrix(HB,WB,sizeof(float),B,HB,BB,HB);
        if (status != CUBLAS_STATUS_SUCCESS) {
        fprintf (stderr, "!!!! device memory allocation error (A)\n");
        return EXIT_FAILURE;
        }

    /*KERNEL*/
        cublasSgemm('n','n',HA,WB,WA,1,AA,HA,BB,HB,0,CC,HC);

        status = cublasGetError();
        if (status != CUBLAS_STATUS_SUCCESS) {
        fprintf (stderr, "!!!! kernel execution error.\n");
        return EXIT_FAILURE;
        }
        cublasGetMatrix(HC,WC,sizeof(float),CC,HC,C,HC);
        if (status != CUBLAS_STATUS_SUCCESS) {
        fprintf (stderr, "!!!! device read error (A)\n");
        return EXIT_FAILURE;
        }


    /* PERFORMANCE OUTPUT*/

    printf("\nMatriz A:\n");
    printMat(A,WA,HA);
    printf("\nMatriz B:\n");
    printMat(B,WB,HB);
    printf("\nMatriz C:\n");
    printMat(C,WC,HC);

        free( A );  free( B );  free ( C );
        status = cublasFree(AA);
        if (status != CUBLAS_STATUS_SUCCESS) {
        fprintf (stderr, "!!!! memory free error (A)\n");
        return EXIT_FAILURE;
        }
        status = cublasFree(BB);
        if (status != CUBLAS_STATUS_SUCCESS) {
        fprintf (stderr, "!!!! memory free error (B)\n");
        return EXIT_FAILURE;
        }
        status = cublasFree(CC);
        if (status != CUBLAS_STATUS_SUCCESS) {
        fprintf (stderr, "!!!! memory free error (C)\n");
    return EXIT_FAILURE;
    }

        /* Shutdown */
        status = cublasShutdown();
        if (status != CUBLAS_STATUS_SUCCESS) {
        fprintf (stderr, "!!!! shutdown error (A)\n");
        return EXIT_FAILURE;
        }

    if (argc > 1) {
        if (!strcmp(argv[1], "-noprompt") ||!strcmp(argv[1], "-qatest") ) 
        {
    return EXIT_SUCCESS;
        }
        } 
        else
        {
            printf("\nPress ENTER to exit...\n");
            getchar();
        }

return EXIT_SUCCESS;


  }

有什么想法?此外,是否有人在CUBLAS中有一个矩阵乘法实现正在工作,所以我可以比较?提前谢谢。

1 个答案:

答案 0 :(得分:7)

我不明白为什么你认为你发布的代码不起作用。当我编译并运行它时,如果我在matlab中输入相同的矩阵并计算它们的乘积,那么生成的可执行文件会产生相同的输出。

CUBLAS是一个FORTRAN BLAS,它期望以列主要顺序输入(并且您的代码是列主要有序)。如果结果与您想要的结果不符,您必须在某处混淆列和行主要排序。