| 1 | | = A fast GEMM implementation on a Cypress GPU = |
| 2 | | by N.Nakasato (University of Aizu), submitted September 7, 2010. |
| 3 | | |
| 4 | | This paper will be presented at 1st International Workshop on |
| 5 | | Performance Modeling, Benchmarking and Simulation of High Performance Computing Systems (PMBS 10) |
| 6 | | held as part of SC10, New Orleans, November 13-19, 2010 |
| 7 | | |
| 8 | | == abstract == |
| 9 | | We present benchmark results of optimized dense matrix multiplication |
| 10 | | kernels for a Cypress GPU. We write general matrix multiply (GEMM) kernels |
| 11 | | for single (SP), double (DP) and double-double (DDP) precision. |
| 12 | | Our SGEMM and DGEMM kernels show 73% and 87% of |
| 13 | | the theoretical performance of the GPU, respectively. |
| 14 | | Currently, our SGEMM and DGEMM kernels are fastest |
| 15 | | with one GPU chip to our knowledge. |
| 16 | | Furthermore, the performance of our matrix multiply kernel in DDP is 31 Gflop/s. |
| 17 | | It is more than 200 times faster than the performance |
| 18 | | results on single core of a recent CPU (with mpack version 0.6.5). |
| 19 | | We describe our GEMM kernels with main focus on the SGEMM implementation |
| 20 | | since all GEMM kernels share common programming and optimization techniques. |
| 21 | | While a conventional wisdom of GPU programming recommends us |
| 22 | | to heavily use shared memory on GPUs, |
| 23 | | we show that texture cache is very effective on the Cypress architecture. |
| 24 | | |
| 25 | | == preliminary results == |
| 26 | | * [wiki:"GEMM_Performance_Cypress"] |
| 27 | | * [wiki:"MatrixMultiply"] |
| 28 | | |
| 29 | | == preprint == |
| 30 | | Posted later. |
| 31 | | |
| 32 | | == Sample program for DGEMM == |
| 33 | | |
| 34 | | |
| | 1 | See [wiki:"Fast_GEMM_implementation_On_Cypress"] since we slightly change the title ;-). |
