Question

我试图将矩阵与其转置相乘，但我无法设法正确进行sgemm调用。 Sgemm需要很多参数。像lda，ldb这样的一些人对我来说很困惑。如果我使用方形矩阵调用下面的函数，则无法正常工作。

/*param inMatrix: contains the matrix data in major order like [1 2 3 1 2 3]
  param rowNum: Number of rows in a matrix eg if matrix is 
                 |1  1|
                 |2  2|
                 |3  3| than rowNum should be 3*/ 
void matrixtTransposeMult(std::vector<float>& inMatrix, int rowNum)
{
    cublasHandle_t handle;
    cublasCreate(&handle);

    int colNum = (int)inMatrix.size() / rowNum;
    thrust::device_vector<float> d_InMatrix(inMatrix);
    thrust::device_vector<float> d_outputMatrix(rowNum*rowNum);
    float alpha = 1.0f;
    float beta = 0.0f;

    cublasSgemm(handle, CUBLAS_OP_N, CUBLAS_OP_T, rowNum, rowNum, colNum, &alpha,
        thrust::raw_pointer_cast(d_InMatrix.data()), colNum, thrust::raw_pointer_cast(d_InMatrix.data()), colNum, &beta,
        thrust::raw_pointer_cast(d_outputMatrix.data()), rowNum);

    thrust::host_vector<float> result = d_outputMatrix;
    for (auto elem : result)
        std::cout << elem << ",";
    std::cout << std::endl;

    cublasDestroy(handle);
}

我错过了什么？如何为矩阵* matrixTranspose进行正确的sgemm调用？

Answer 1

下面的设置对我有用，如果我遗失了什么，请警告我。我希望它对某些人有用

void matrixtTransposeMult(std::vector<float>& inMatrix, int rowNum)
{
    cublasHandle_t handle;
    cublasCreate(&handle);

    int colNum = (int)inMatrix.size() / rowNum;
    thrust::device_vector<float> d_InMatrix(inMatrix);
    thrust::device_vector<float> d_outputMatrix(rowNum*rowNum);
    float alpha = 1.0f;
    float beta = 0.0f;

    cublasSgemm(handle, CUBLAS_OP_N, CUBLAS_OP_T, rowNum, rowNum, colNum, &alpha,
        thrust::raw_pointer_cast(d_InMatrix.data()), rowNum, thrust::raw_pointer_cast(d_InMatrix.data()), rowNum, &beta,
        thrust::raw_pointer_cast(d_outputMatrix.data()), rowNum);

    thrust::host_vector<float> result = d_outputMatrix;
    for (auto elem : result)
        std::cout << elem << ",";
    std::cout << std::endl;

    cublasDestroy(handle);
}

使用cuBlas将矩阵与其转置相乘

1 个答案: