rv-sparse

Sparse linear algebra for RISC-V. RV-Sparse computes CSR sparse matrix-matrix multiplication (SpGEMM), C = A × B, in FP32. It includes a scalar kernel and three RISC-V Vector (RVV) kernels.

Build

make clean
make

To cross-compile for RISC-V:

make TARGET_ARCH=riscv

This selects the RISC-V toolchain and builds with -march=rv64gcv -mabi=lp64d.

You can also select the target explicitly with ARCH_FLAGS:

make ARCH_FLAGS="-march=rv64gcv -mabi=lp64d"   # scalar and RVV kernels
make ARCH_FLAGS="-march=rv64gc -mabi=lp64d"    # scalar kernel only

The RVV kernels are built when the target enables the RISC-V Vector extension.

Usage

#include "rv_sparse.h"

rvsp_csr_matrix_t A, B, C;

rvsp_csr_create(
    &A,
    M, K, A_nnz,
    A_row_ptr,
    A_col_idx,
    A_values,
    RVSP_DTYPE_FP32
);

rvsp_csr_create(
    &B,
    K, N, B_nnz,
    B_row_ptr,
    B_col_idx,
    B_values,
    RVSP_DTYPE_FP32
);

rvsp_spgemm_options_t options = {
    .backend = RVSP_BACKEND_SCALAR,
    .input_dtype = RVSP_DTYPE_FP32,
    .output_dtype = RVSP_DTYPE_FP32,
};

rvsp_spgemm_csr(&A, &B, &C, &options);

rvsp_csr_destroy(&C);

rvsp_spgemm_csr allocates the output matrix C. The caller is responsible for releasing it with rvsp_csr_destroy.

For repeated multiplies with the same sparsity pattern, use the descriptor API to reuse the structure analysis. It also allows a specific accumulation strategy to be selected.

Documentation

• docs/api_design.md for the API reference
• examples/ for worked examples

License

GPL-3.0-only. See LICENSE.

Project by Talha Ahmed (@talha-ahmed-1) and mentees

[C. Ramirez] (@crissmath), [Ayokunle Amodu] (@ayokunle321), [Hamza Mateen] (@HamzaMateen) under the Linux Foundation Mentorship Program, sponsored by RISC-V International.

Hardware evaluation was conducted on RISC-V development boards provided by Cloud-V and the DC-ROMA II laptop provided by DeepComputing.