cublasDgemv

cublasDgemv is a routine in NVIDIA’s cuBLAS library that performs a general matrix-vector multiply with double-precision real numbers. It computes y = alpha * op(A) * x + beta * y, where op(A) is either A or its transpose, depending on the trans parameter. The operation is carried out on data stored in device memory, and the function uses a cuBLAS handle and the CUDA stream associated with that handle.

The function signature includes: a cuBLAS handle, a trans option, the dimensions m and n, pointers to the scalar multipliers alpha and beta, pointers to the matrix A and vectors x and y, and the leading dimension and increments. Specifically, A is an m-by-n matrix when trans is not transposed (CUBLAS_OP_N), and A is treated as transposed when trans is CUBLAS_OP_T. In all cases, op(A) has dimensions m-by-n. The leading dimension lda must be at least max(1, m) for trans=N or max(1, n) for trans=T. The vector x has length n when trans=N or length m when trans=T, and the resulting vector y has length m. incx and incy are the increments for stepping through x and y, typically 1.

cuBLAS uses column-major storage by default, consistent with traditional BLAS. If your data uses row-major layout, you can adapt by adjusting the trans argument or layout accordingly. All input and output data reside in device memory, and alpha, beta, and the scalars may be supplied from host memory (via pointers) or device memory, with cuBLAS handling the transfer as needed. The routine returns a cuBLASStatus_t status code indicating success or error. cuBLAS provides analogous routines for single-precision (cublasSgemv) and complex data types as appropriate.