Batched Refactorization in cuSOLVER

A complete example of batched refactorization in CUDA cuSOLVER.

Batched refactorization module in cuSOLVER provides an efficient method to solve batches of linear systems with fixed left-hand side sparse matrix (or matrices with fixed sparsity pattern but varying coefficients) and varying right-hand sides, based on LU decomposition. Only some partially completed code snippets can be found in the official documentation (as of CUDA 11.1.0) that are related to it. In this repository, a complete example is provided to demonstrate the usage of related functions including

• cusolverRfBatchSetupHost,
• cusolverRfBatchAnalyze,
• cusolverRfBatchResetValues,
• cusolverRfBatchRefactor,
• and cusolverRfBatchSolve.

Build

This project is developed based on the unbatched refactorization example 7_CUDALibraries/cuSolverRf from NVIDIA's official CUDA samples for Linux. You may modify CUDA_PATH in Makefile to build this project.

Usage

A 60x speedup was observed with the commands below, on a GeForce RTX 3080.

./cuSolverRf -P=symamd -file=./lap2D_5pt_n100.mtx # 0.000641 sec for each linear system
./cuSolverRfBatch -P=symamd -file=./lap2D_5pt_n100.mtx -bs=256 # 0.000010 sec for each linear system

License

This software contains source code provided by NVIDIA Corporation. See End User License Agreement for more information.