rocblassgemm

rocBLAS is AMD's GPU-accelerated library for basic linear algebra on the ROCm platform. It provides high-performance implementations of BLAS routines, including the general matrix-matrix multiply, GEMM. The GEMM routines in rocBLAS cover multiple precisions, such as single-precision (sgemm), double-precision (dgemm), and complex variants, with batched and strided-batched options for many small or large problems.

The fundamental operation performed by the rocBLAS GEMM routines is C = alpha op(A) op(B) + beta C, where op(X) denotes either X or its transpose, and A, B, and C have compatible dimensions (subject to the chosen transpositions). The routine accepts typical GEMM parameters: the matrices A, B, and C; their leading dimensions; the dimensions m, n, k; the scalar multipliers alpha and beta; and transposition flags for A and B.

Usage is handle-based: a program creates a rocblas_handle, allocates and transfers memory for the matrices on the device, and then calls the GEMM routine with the appropriate pointers and dimensions. rocBLAS also provides batched and strided-batched variants to perform many GEMMs efficiently within a single call, as well as specialized variants for different data layouts and precisions. Execution is asynchronous with respect to the host, enabling overlap of computation and communication on supported ROCm platforms.

RocBLAS GEMM is a core component of AMD’s ROCm linear algebra stack, commonly used in high-performance computing and machine learning on AMD GPUs. It is analogous in purpose to cuBLAS sgemm/dgemm on NVIDIA hardware and is distributed as part of the ROCm open-source ecosystem.