cusparseLt - `cusparseMatMul`

Description

This sample demonstrates the usage of cuSPARSELt library and cusparseMatMul APIs for performing structured matrix - dense matrix multiplication by exploiting NVIDIA Sparse Tensor Cores, where the structured matrix is compressed with 50% sparsity ratio. The sample also demonstrates the usage of batched computation, Split-K, ReLU activation function, and bias.

cusparseLt Documentation

C_i = ReLU(A_i * B_i + C_i + bias)

where A is an structured matrix, and B, C, D are dense matrices

Building

Linux

make CUSPARSELT_PATH=<cusparseLt_path> CUDA_TOOLKIT_PATH=<cuda_toolkit_path>

or in alternative:

mkdir build
cd build
cmake -DCUSPARSELT_PATH=<cusparseLt_path> -DCMAKE_CUDA_COMPILER=<nvcc_path> ..
make

Support

Supported SM Architectures: SM 8.0, SM 8.6, SM 8.7, SM 8.9, SM 9.0, SM10.0, SM10.1, SM110, SM 12.0, SM 12.1
Supported OSes: Linux, Windows
Supported CPU Architectures: x86_64, arm64
Supported Compilers: gcc, clang, Microsoft msvc, Nvidia HPC SDK nvc
Language: C++14

Prerequisites

CUDA 12.9 toolkit (or above) and compatible driver (see CUDA Driver Release Notes).
cusparseLt 0.8.0 or above
CMake 3.18 or above