Method MultTransAdd

ngscuda/cuda_linalg.cpp:271–308

  void DevSparseMatrix :: MultTransAdd (double s, const BaseVector & x, BaseVector & y) const
  {
    static Timer tmv("DevSparseMatrix :: MultTransAdd");
    CudaRegionTimer reg(tmv);

    UnifiedVectorWrapper ux(x);
    UnifiedVectorWrapper uy(y);

    ux.UpdateDevice();
    uy.UpdateDevice();

    double alpha= s;
    double beta = 1;

    size_t bufferSize = 0;
    void* dBuffer = NULL;

    cusparseDnVecDescr_t descr_x, descr_y;
    cusparseCreateDnVec (&descr_x, ux.Size(), ux.DevData(), CUDA_R_64F);
    cusparseCreateDnVec (&descr_y, uy.Size(), uy.DevData(), CUDA_R_64F);

    cusparseSpMV_bufferSize(Get_CuSparse_Handle(), CUSPARSE_OPERATION_TRANSPOSE,
                            &alpha, descr, descr_x, &beta, descr_y, CUDA_R_64F,
                            CUSPARSE_SPMV_ALG_DEFAULT, &bufferSize);
    cudaMalloc(&dBuffer, bufferSize);

    cusparseSpMV(Get_CuSparse_Handle(), 
                 CUSPARSE_OPERATION_TRANSPOSE, &alpha, descr,
                 descr_x, &beta, descr_y, CUDA_R_64F,
                 CUSPARSE_SPMV_ALG_DEFAULT, dBuffer);

    cudaFree(dBuffer);

    cusparseDestroyDnVec(descr_x);
    cusparseDestroyDnVec(descr_y);

    uy.InvalidateHost();
  }