rocblasdgemm

rocblas_dgemm is the double-precision general matrix-matrix multiplication routine in the ROCm library rocBLAS. It computes C = alpha * op(A) * op(B) + beta * C, where A, B, and C reside on the device. The operation op(X) is determined by transA and transB (none, transpose, or conjugate transpose; the latter is equivalent to transpose for real numbers). Following BLAS conventions, matrices are treated as column-major and the leading dimensions lda, ldb, and ldc specify the physical strides between columns.

The API is exposed as a C-style function in rocBLAS. A typical prototype is: rocblas_status rocblas_dgemm(rocblas_handle handle, rocblas_operation transA, rocblas_operation transB, rocblas_int m, rocblas_int n, rocblas_int k, const double* alpha, const double* A, rocblas_int lda, const double* B, rocblas_int ldb, const double* beta, double* C, rocblas_int ldc). A, B, and C point to device memory; alpha and beta are scalars whose memory location is determined by the current pointer mode of the handle (host or device). The handle controls the execution context and stream association.

Dimensional rules follow op(A) being m-by-k and op(B) being k-by-n, so the result C is m-by-n. Therefore lda ≥ max(1, number of rows of op(A)), ldb ≥ max(1, number of rows of op(B)), and ldc ≥ max(1, m).

rocblas_dgemm is used as a building block in dense linear algebra on AMD GPUs. In addition to the standard routine, rocBLAS provides batched variants (for multiple GEMMs) and strided variants to support higher throughput in applications requiring many small to moderate GEMMs. It is part of the ROCm software stack and integrates with other rocBLAS routines and the ROCm runtime.