|
An effective implementation of Strassen's algorithmusing AVX intrinsics for a multicore architecture |
|---|---|
| รหัสดีโอไอ | |
| Creator | 1. Nwe Zin Oo 2. Panyayot Chaikan |
| Title | An effective implementation of Strassen's algorithmusing AVX intrinsics for a multicore architecture |
| Publisher | Research and Development Office, Prince of Songkla University |
| Publication Year | 2563 |
| Journal Title | Songklanakarin Journal of Science and Technology |
| Journal Vol. | 42 |
| Journal No. | 6 |
| Page no. | 1368-1376 |
| Keyword | advanced vector extension, AVX, AVX-2, matrix-matrix multiplication, FMA, Strassen's algorithm |
| URL Website | https://rdo.psu.ac.th/sjstweb/index.php |
| ISSN | 0125-3395 |
| Abstract | This paper proposes an effective implementation of Strassen's algorithm with AVX intrinsics to augment matrix-matrixmultiplication in a multicore system. AVX-2 and FMA3 intrinsic functions are utilized, along with OpenMP, to implement themultiplication kernel of Strassen's algorithm. Loop tiling and unrolling techniques are also utilized to increase the cacheutilization. A systematic method is proposed for determining the best stop condition for the recursion to achieve maximumperformance on specific matrix sizes. In addition, an analysis method makes fine-tuning possible when our algorithm is adaptedto another machine with a different hardware configuration. Performance comparisons between our algorithm and the latestversions of two well-known open-source libraries have been carried out. Our algorithm is, on average, 1.52 and 1.87 times fasterthan the Eigen and the OpenBLAS libraries, respectively, and can be scaled efficiently when the matrix becomes larger. |
