Question

对于我第一次尝试CUDA / cuBLAS的工作，我试图编写一个简单的函数，将MxN矩阵（用向量std::vector表示）乘以Nx1个“ ones”向量，因此以获取矩阵的rowwise（？）和。这将利用cublas_gemv()以及其他基本的CUDA操作，我认为这是一个不错的起点。

在处理设置问题并读取/复制示例代码之后，这就是我所拥有的：

std::vector<double> test(std::vector<std::vector<double>> in)
{
    std::vector<double> out;
    long in_m = in.size();
    long in_n = in[0].size();
    cudaError_t cudaStat;
    cublasStatus_t stat;
    cublasHandle_t handle;
    // This just converts a vector-of-vectors into a col-first array
    double* p_in = vec2d_to_colfirst_array(in);
    double* p_ones = new double[in_n];
    double* p_out = new double[in_m];
    std::fill(p_ones, p_ones + in_n, 1.0);
    double* dev_in;
    double* dev_ones;
    double* dev_out;
    cudaStat = cudaMalloc((void**)&dev_in, in_m * in_n * sizeof(double));
    cudaStat = cudaMalloc((void**)&dev_ones, in_n * sizeof(double));
    cudaStat = cudaMalloc((void**)&dev_out, in_m * sizeof(double));
    stat = cublasCreate(&handle);
    cudaStat = cudaMemcpy(dev_in, p_in, in_m*in_n * sizeof(double), cudaMemcpyHostToDevice);
    cudaStat = cudaMemcpy(dev_ones, p_ones, in_n * sizeof(double), cudaMemcpyHostToDevice);
    double alpha = 1.0;
    double beta = 0.0;
    stat = cublasDgemv(handle, CUBLAS_OP_N, in_m, in_n, &alpha, dev_in, in_m, dev_ones, 1, &beta, dev_ones, 1);
    cudaStat = cudaMemcpy(p_out, dev_out, in_m * sizeof(double), cudaMemcpyDeviceToHost);
    out.assign(p_out, p_out + in_m);
    cudaFree(dev_in);
    cudaFree(dev_ones);
    cudaFree(dev_out);
    cublasDestroy(handle);
    free(p_in);
    free(p_ones);
    free(p_out);
    return out;
}

它与我阅读的样本看起来并没有太大区别，因此我希望它可以“正常工作”。但是，当我检查p_out时，它们全为零。当然，我没有输入零in矩阵。

我证实vec2d_to_colfirst_array()可以很好地完成工作，并且还可以通过将数据从设备复制回主机然后读取来正确填充dev_in / dev_ones。也许问题出在cublasDgemv()的调用之内，但是由于我是新手（而且与Eigen相比，BLAS语法也不太直观），经过无奈之后，我什至看不到出什么问题了。 / p>

任何帮助表示赞赏！

Answer 1

该错误似乎非常简单。您希望从dev_out复制结果：

cudaStat = cudaMemcpy(p_out, dev_out, in_m * sizeof(double), cudaMemcpyDeviceToHost);

但是您绝不会在cublas通话中使用dev_out：

stat = cublasDgemv(handle, CUBLAS_OP_N, in_m, in_n, &alpha, dev_in, in_m, dev_ones, 1, &beta, dev_ones, 1);

这似乎只是复制粘贴错误。如果将cublas调用中的dev_ones的最后一个实例替换为dev_out，则您的代码对我有用：

stat = cublasDgemv(handle, CUBLAS_OP_N, in_m, in_n, &alpha, dev_in, in_m, dev_ones, 1, &beta, dev_out, 1);

这是一个完全有效的示例，其中进行了更改：

$ cat t315.cu
#include <vector>
#include <cublas_v2.h>
#include <iostream>

const long idim1 = 8;
const long idim2 = 8;

double* vec2d_to_colfirst_array(std::vector<std::vector<double>> in){
    long dim1 = in.size();
    long dim2 = in[0].size();
    long k = 0;
    double *res = new double[dim1*dim2];
    for (int i = 0; i < dim1; i++)
      for (int j = 0; j < dim2; j++) res[k++] = in[i][j];
    return res;
}


std::vector<double> test(std::vector<std::vector<double>> in)
{
    std::vector<double> out;
    long in_m = in.size();
    long in_n = in[0].size();
    cudaError_t cudaStat;
    cublasStatus_t stat;
    cublasHandle_t handle;
    // This just converts a vector-of-vectors into a col-first array
    double* p_in = vec2d_to_colfirst_array(in);
    double* p_ones = new double[in_n];
    double* p_out = new double[in_m];
    std::fill(p_ones, p_ones + in_n, 1.0);
    double* dev_in;
    double* dev_ones;
    double* dev_out;
    cudaStat = cudaMalloc((void**)&dev_in, in_m * in_n * sizeof(double));
    cudaStat = cudaMalloc((void**)&dev_ones, in_n * sizeof(double));
    cudaStat = cudaMalloc((void**)&dev_out, in_m * sizeof(double));
    stat = cublasCreate(&handle);
    cudaStat = cudaMemcpy(dev_in, p_in, in_m*in_n * sizeof(double), cudaMemcpyHostToDevice);
    cudaStat = cudaMemcpy(dev_ones, p_ones, in_n * sizeof(double), cudaMemcpyHostToDevice);
    double alpha = 1.0;
    double beta = 0.0;
    stat = cublasDgemv(handle, CUBLAS_OP_N, in_m, in_n, &alpha, dev_in, in_m, dev_ones, 1, &beta, dev_out, 1);
    cudaStat = cudaMemcpy(p_out, dev_out, in_m * sizeof(double), cudaMemcpyDeviceToHost);
    out.assign(p_out, p_out + in_m);
    cudaFree(dev_in);
    cudaFree(dev_ones);
    cudaFree(dev_out);
    cublasDestroy(handle);

    free(p_in);
    free(p_ones);
    free(p_out);
    return out;
}

int main(){

  std::vector<double> a(idim2, 1.0);
  std::vector<std::vector<double>> b;
  for (int i = 0; i <  idim1; i++) b.push_back(a);
  std::vector<double> c = test(b);
  for (int i = 0; i < c.size(); i++) std::cout << c[i] << ",";
  std::cout << std::endl;
}

$ nvcc -std=c++11 -o t315 t315.cu -lcublas
t315.cu(24): warning: variable "cudaStat" was set but never used

t315.cu(25): warning: variable "stat" was set but never used

$ cuda-memcheck ./t315
========= CUDA-MEMCHECK
8,8,8,8,8,8,8,8,
========= ERROR SUMMARY: 0 errors
$

请注意，我认为free()不是与new一起使用的正确API，但这似乎并不是您所遇到问题的症结所在。

矩阵向量乘法（cublasDgemv）返回零

1 个答案: