cublasSgemm无效__global__读取

时间:2015-10-21 16:30:06

标签: cuda cublas

当尝试使用cublasSgemm例程执行张量矩阵产品时,会发生地址超出范围错误,下面提供了一个示例: -

========= Invalid __global__ read of size 4
=========     at 0x000019f8 in sgemm_sm35_ldg_nn_64x16x64x16x16
=========     by thread (6,3,0) in block (6,3,0)
=========     Address 0x7ffc059064a8 is out of bounds
=========     Saved host backtrace up to driver entry point at kernel launch time
=========     Host Frame:/lib64/libcuda.so.1 (cuLaunchKernel + 0x2cd) [0x15859d]
=========     Host Frame:/usr/local/cuda-7.5/lib64/libcublas.so.7.5 [0x21fb31]
=========     Host Frame:/usr/local/cuda-7.5/lib64/libcublas.so.7.5 [0x23a343]
=========     Host Frame:/usr/local/cuda-7.5/lib64/libcublas.so.7.5 [0x1d4e92]
=========     Host Frame:/usr/local/cuda-7.5/lib64/libcublas.so.7.5 [0x1d17b4]
=========     Host Frame:/usr/local/cuda-7.5/lib64/libcublas.so.7.5 [0x1d2c5e]
=========     Host Frame:/usr/local/cuda-7.5/lib64/libcublas.so.7.5 [0x1d37b2]
=========     Host Frame:/usr/local/cuda-7.5/lib64/libcublas.so.7.5 [0xecd31]
=========     Host Frame:./test [0x2c0e]
=========     Host Frame:./test [0x2a99]
=========     Host Frame:/lib64/libc.so.6 (__libc_start_main + 0xf5) [0x21af5]
=========     Host Frame:./test [0x2749]

在我的应用程序中多次检查维度并确定这不是问题之后,我写了一个最小的工作示例。下面是一个简单的例子,它将两个平方矩阵相乘: -

#include "stdlib.h"
#include "time.h"
#include "stdio.h"
#include "cuda.h"
#include <cuda_runtime.h>
#include "cublas_v2.h"
#include <math.h>
#include "cuda_error.h"

void matrixMult(cublasOperation_t transA, cublasOperation_t transB, int M, int N,
            int K, float alpha, float *A, float *B, float beta, float *C,
                cublasHandle_t *cb_handle);

int main(){
    int i, j, idx;
    int D = 500;

    int len = D*D;
    float *A_h, *B_h, *C_h;
    float *A_d, *B_d, *C_d;

    A_h = (float*)malloc(len*sizeof(float));
    B_h = (float*)malloc(len*sizeof(float));
    C_h = (float*)malloc(len*sizeof(float));

    srand48(time(NULL));
    for(i=0; i<D; i++){
        for(j=0; j<D; j++){
            A_h[i*D + j] = drand48();
            B_h[i*D + j] = drand48();
        }
    }

    cudaCheck(cudaMalloc((void**)&A_d, len*sizeof(float)));
    cudaCheck(cudaMalloc((void**)&B_d, len*sizeof(float)));
    cudaCheck(cudaMalloc((void**)&C_d, len*sizeof(float)));
    cudaCheck(cudaMemcpy(A_d, A_h, len*sizeof(float),  cudaMemcpyHostToDevice));
    cudaCheck(cudaMemcpy(B_d, B_h, len*sizeof(float), cudaMemcpyHostToDevice));

    cublasHandle_t cb_handle;
    cublasCheck(cublasCreate(&cb_handle));
    cublasSetPointerMode(cb_handle, CUBLAS_POINTER_MODE_DEVICE);
    matrixMult(CUBLAS_OP_N, CUBLAS_OP_N, D, D, D, 1.0, B_d, A_d, 0.0, C_d, &cb_handle);
    cublasDestroy(cb_handle);

    cudaCheck(cudaMemcpy(C_h, C_d, len*sizeof(float), cudaMemcpyDeviceToHost));
    cudaCheck(cudaFree(A_d));
    cudaCheck(cudaFree(B_d));
    cudaCheck(cudaFree(C_d));

    free(A_h);
    free(B_h);
    free(C_h);
}

void matrixMult(cublasOperation_t transA, cublasOperation_t transB, int M, int N,
            int K, float alpha, float *A, float *B, float beta, float *C,
            cublasHandle_t *cb_handle){
    int lda = (transA == CUBLAS_OP_N) ? K : M;
    int ldb = (transB == CUBLAS_OP_N) ? N : K;
    int ldc = N;
    cublasCheck(cublasSgemm(*cb_handle, transB, transA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, ldc));
}

使用以下简单的错误捕获标题: -

#ifndef CUDA_ERROR_CHECK
#define CUDA_ERROR_CHECK

#include <cuda_runtime.h>
#include "cublas_v2.h"

#define cudaCheck(ans){cuda_assert((ans), __FILE__, __LINE__);}
#define cublasCheck(ans){cublas_assert((ans), __FILE__, __LINE__);}

inline void cuda_assert(cudaError_t code, const char *file, int line){
   if(code != cudaSuccess){
      fprintf(stderr,"CUDA Error: %s %s %d\n", cudaGetErrorString(code), file, line);
      exit(code);
   }
}

inline void cublas_assert(cublasStatus_t code, const char *file, int line){
    if(code != CUBLAS_STATUS_SUCCESS){
        fprintf(stderr, "CUBLAS Error! %s line: %d error code: %d\n", file, line, code);
        exit(code);
    }
}

#endif

请注意,上述误差输出由上述方阵示例产生。我的张量积应用得到了类似的输出。

我正在使用带有Titan Black卡的CUDA 7.5。我做了一些根本错误的事情,还是可能是我的cuBLAS安装问题?

1 个答案:

答案 0 :(得分:1)

如果你消除了这个:

cublasSetPointerMode(cb_handle, CUBLAS_POINTER_MODE_DEVICE);

您的代码将正常运行。不清楚为什么要将指针模式设置为CUBLAS_POINTER_MODE_DEVICEdocumentation表示:

  

有两类使用标量参数的函数:

     
      
  • 通过主机或设备上的引用获取alpha和/或beta参数作为缩放因子的函数,例如gemm

  •   
  • 在主机或设备上返回标量结果的函数,如amax(),amin,asum(),rotg(),rotmg(),dot()和nrm2()。

  •   
     

对于第一类的函数,当指针模式设置为CUBLAS_POINTER_MODE_HOST时,标量参数alpha和/或beta可以在堆栈上或在堆上分配。

CUBLAS_POINTER_MODE_HOST默认设置,在您的情况下,它是正确的设置,其中&alpha&beta是指向主机内存的指针:< / p>

cublasCheck(cublasSgemm(*cb_handle, transB, transA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, ldc));