rocBLAS

rocBLAS is AMD’s high-performance Basic Linear Algebra Subprograms (BLAS) library for the ROCm software platform. It provides accelerated implementations of BLAS Level 1, 2, and 3 routines for AMD GPUs in a HIP-based environment, with a focus on enabling fast matrix and vector operations for scientific computing, data analytics, and machine learning workloads on AMD hardware.

The library uses a handle-based API. Applications create a rocblas_handle, initialize it, and call functions prefixed with rocblas_ to perform operations. Example routines include rocblas_sgemm, rocblas_dgemm, rocblas_cgemm, and rocblas_zgemm, which implement single- and double-precision real and complex matrix-matrix multiplication. Batched and strided_batched variants such as rocblas_sgemm_strided_batched enable processing many matrices efficiently. Other routines cover vectors and matrices (for example, dot, axpy, scal, gemv). rocBLAS supports multiple data types, including float, double, complex float, complex double, and, on supported hardware, half precision. Computations use device pointers, and a HIP stream can be attached to a handle via rocblas_set_stream.

rocBLAS is a core component of ROCm, optimized for AMD GPUs across generations and designed to interoperate with other ROCm libraries and tools. Its API and usage resemble cuBLAS in naming and structure to facilitate porting from CUDA-based code, while maintaining architecture-specific optimizations. The library is maintained by AMD and distributed as part of the open-source ROCm stack, serving a wide range of high-performance and scientific computing applications that require efficient BLAS routines on AMD hardware.