What is hipSPARSELt?

Contents

What is hipSPARSELt?#

hipSPARSELt is a SPARSE marshalling library supporting multiple backends. It presents a common interface that provides Basic Linear Algebra Subroutines (BLAS) for sparse computation, implemented on top of the AMD ROCm runtime and toolchains. Through supported backends, hipSPARSELt enables structured sparse matrix multiplication that can take advantage of AMD sparse MFMA Matrix Core acceleration on supported AMD GPUs.

hipSPARSELt sits between the application and a “worker” SPARSE library, marshalling inputs into the backend library and results back to the application. It exports a uniform interface that doesn’t require client-side changes when switching backends. The supported backends are: rocSPARSELt and NVIDIA CUDA cuSPARSELt v0.6.3.

The hipSPARSELt library is created using the HIP programming language and is optimized for the latest AMD discrete GPUs.

Key features#

  • Mixed-precision computation:

    • FP16 input/output with FP32 matrix core accumulate

    • BFLOAT16 input/output with FP32 matrix core accumulate

    • INT8 input/output with INT32 matrix core accumulate

    • INT8 input with FP16 output and INT32 matrix core accumulate

    • FP8 input with FP32 output and FP32 matrix core accumulate

    • BF8 input with FP32 output and FP32 matrix core accumulate

  • Sparse matrix capabilities:

    • Matrix pruning and compression functionalities

    • Auto-tuning functionality (see hipsparseLtMatmulSearch())

  • Batched sparse GEMM:

    • Single sparse matrix/multiple dense matrices (broadcast)

    • Multiple sparse and dense matrices

    • Batched bias vector

  • Fused activation support in SpMM kernels:

    • ReLU

    • ClippedReLU

    • GeLU

    • GeLU scaling

    • Abs

    • LeakyReLU

    • Sigmoid

    • Tanh

Contents