A Thread-Level Synchronization-Free Sparse Triangular Solve on GPUs
This is the source code of a paper entitled "CapelliniSpTRSV: A Thread-Level Synchronization-Free Sparse Triangular Solve on GPUs" by Jiya Su, Feng Zhang, Weifeng Liu, Bingsheng He, Ruofan Wu, Xiaoyong Du, Rujia Wang, 2020.
There are two versions, namely CUDA and openCL.
In CUDA version, we show 6 SpTRSV algorithms.Among them, SyncFree_csc, cuSP and cuSP-layer are already proposed algorithms, which are used as benchmark algorithms. The other three algorithms are our proposed algorithms. mix has the best performance, but has a short preprocessing time. ourWrtFst has better performance than our2Part, and there is no preprocessing time for both algorithms.
In OpenCL version, we only give 4 algorithms. Among them, SyncFree_csc and SyncFree_csc2 are already proposed algorithms, which are used as benchmark algorithms. The other two algorithms are our proposed algorithms. mix has the best performance, but has a short preprocessing time. ourWrtFst does not require preprocessing time.
We will continue to improve this project to make CapelliniSpTRSV easier to use.
Our video introducing CapelliniSpTRSV can be seen on https://www.youtube.com/watch?v=HMNeLWq7D48.
Our paper can be downloaded from ACM (https://dl.acm.org/doi/abs/10.1145/3404397.3404400).
Our slide can be downloaded from https://jnamaral.github.io/icpp20/slides/Su_CapelliniSpTRSV.pdf.
Sparse triangular solves (SpTRSVs) have been extensively used in linear algebra fields, and many GPU SpTRSV algorithms have been proposed. Synchronization-free SpTRSVs, due to their short preprocessing time and high performance, are currently the most popular SpTRSV algorithms. However, we observe that the performance of the current synchronization-free SpTRSV algorithms on different matrices varies greatly. Specifically, we find that when the average number of components per level is high and the average number of nonzero elements per row is low, these SpTRSVs exhibit low performance. The reason is that current SpTRSVs use a warp in GPUs to process a row in sparse matrices, and such warp-level designs increase the dependencies in SpTRSVs; this problem becomes serious in these cases. To solve this problem, we propose CapelliniSpTRSV, a thread-level synchronization-free SpTRSV algorithm. Particularly, CapelliniSpTRSV has three desirable features. First, CapelliniSpTRSV does not need preprocessing to calculate levels. Second, our algorithm exhibits high performance on matrices that previous SpTRSVs cannot handle efficiently. Third, CapelliniSpTRSV is based on the most popular sparse matrix storage, compressed sparse row (CSR) format, which implies that users do not need to conduct format conversion. We evaluate CapelliniSpTRSV with 245 matrices from the Florida Sparse Matrix Collection on three GPU platforms, and experiments show that our SpTRSV exhibits 3.41 GFLOPS/s, which is 5.26x speedup over the state-of-the-art synchronization-free SpTRSV algorithm, and 4.00x speedup over the SpTRSV in cuSPARSE. CapelliniSpTRSV can be downloaded from https://github.com/JiyaSu/CapelliniSpTRSV.
The source code is download from https://github.com/bhSPARSE/Benchmark_SpTRSM_using_CSC. In OpenCL, SyncFree_csc2 removes the division of warp in SyncFree_csc (sptrsv_syncfree_opencl.cl).
The code uses the SpTRSV function from the cuSPARSE library. The function has two options CUSPARSE_SOLVE_POLICY_NO_LEVEL and CUSPARSE_SOLVE_POLICY_USE_LEVEL, corresponding to cuSP and cuSP-layer respectively.
A novel thread-level synchronization-free SpTRSV algorithm, targeting the sparse matrices that have large number of components per level and small number of nonzero elements per row.
An optimization SpTRSV algorithm based on our2Part, which has better performance and targets the sparse matrices that have large number of components per level and small number of nonzero elements per row.
An integrated algorithm with SyncFree and ourWrtFst for all sparse matrices.
- Choose the algorithm you want to run, and enter the folder corresponding to the algorithm.
- Adjust the common.h file according to the GPU hardware, the repeated times, and the accuracy of the calculation (single or double precision).
- Set CUDA path in the Makefile.
- Run
make. - Run
./main example.mtx. (kernel is in the spts_ .h) - The result is saved in result.csv as
matrix path, row number, the number of total nonzero elements, the average number of nonzero elements per row, level number, the average number of rows per level, pre_time, solve_time, gflops, bandwith.
CUDA:
- nvidia GTX 1080 (Pascal) GPU in a host with CUDA v8.0 and Ubuntu 16.04.4 Linux installed.
- nvidia Tesla V100 (Volta) GPU in a host with CUDA v9.0 and Ubuntu 16.04.1 Linux installed.
- nvidia GeForce RTX 2080 Ti (Turing) GPU in a host with CUDA v10.2 and Ubuntu 18.04.4 Linux installed.
OpenCL:
- Radeon Vega 11 APU with ROCm compiler and Ubuntu 18.04.3 Linux installed.
CapelliniSpTRSV is developed by Renmin University of China, China University of Petroleum, National University of Singapore, and Illinois Institute of Technology.
Jiya Su, Feng Zhang, Ruofan Wu and Xiaoyong Du are with the Key Laboratory of Data Engineering and Knowledge Engineering (MOE), and School of Information, Renmin University of China.
Weifeng Liu is with the Super Scientific Software Laboratory, Department of Computer Science and Technology, China University of Petroleum.
Bingsheng He is with the School of Computing, National University of Singapore.
Rujia Wang is with the Computer Science Department, Illinois Institute of Technology.
If you have any questions, please contact us ([email protected] or [email protected]).
If you use our code, please cite our paper:
@inproceedings{su2020sptrsv,
author = {Jiya Su and Feng Zhang and Weifeng Liu and Bingsheng He and Ruofan Wu and Xiaoyong Du and Rujia Wang},
title = {{CapelliniSpTRSV: A Thread-Level Synchronization-Free Sparse Triangular Solve on GPUs}},
year = {2020},
booktitle = {Proceedings of the 49th International Conference on Parallel Processing},
series = {ICPP '20}
}