rocSPARSE

rocSPARSE is a high-performance sparse linear algebra library for AMD GPUs within the ROCm software stack. It provides a collection of kernels and helpers for performing sparse-matrix computations on ROCm-enabled hardware, with a design focused on throughput and scalability for large matrices.

The library offers a broad set of operations common in sparse computations, including sparse matrix-vector multiply (SpMV), sparse matrix-matrix multiply (SpMM), sparse general matrix-matrix multiply (SpGEMM), and sparse triangular solves. It also includes utilities for sorting, indexing, and converting between sparse matrix formats, and supports a range of data types (real and complex, in single and double precision).

rocSPARSE supports multiple standard sparse formats, such as CSR and CSC, and provides mechanisms for format conversion and optimized kernels that leverage AMD GPU architectures. The API is exposed through a HIP-based C/C++ interface and is designed to interoperate with other ROCm libraries and tools, such as ROCm’s BLAS and graph analytics stacks.

In production, rocSPARSE is used for scientific computing, engineering simulations, data analytics, and other applications that require efficient sparse computations on AMD hardware. It is distributed as part of the ROCm ecosystem and is maintained to align with newer GPU architectures and software stacks.