wsl-inspur edited a comment on pull request #5485: URL: https://github.com/apache/incubator-tvm/pull/5485#issuecomment-625091214
@FrozenGene We made the following tests to go through all the layouts, Layout_1(the same in this PR) input_tile = (P, CI, alpha, alpha) data_pack = (alpha, alpha, P, CI) bgemm = (alpha, alpha, P, CO) inverse = (P, CO, m, m) output = (N, H, W, CO) kernel = (alpha, alpha, CI, CO) Layout_2 input_tile = (P, CI, alpha, alpha) data_pack = (alpha, alpha, P, CI) bgemm = (alpha, alpha, P, CO) inverse = (P, CO, m, m) output = (N, H, W, CO) kernel = (alpha, alpha, CO, CI) Layout_3 input_tile = (alpha, alpha, P, CI) data_pack = (alpha, alpha, P, CI) bgemm = (alpha, alpha, P, CO) inverse = (P, CO, m, m) output = (N, H, W, CO) kernel = (alpha, alpha, CI, CO) Layout_4 input_tile = (alpha, alpha, P, CI) data_pack = (alpha, alpha, P, CI) bgemm = (alpha, alpha, P, CO) inverse = (m, m, P, CO) output = (N, H, W, CO) kernel = (alpha, alpha, CO, CI) The results are listed below. kernel:3x3x64x64 feature maps:56x56x64: batch | layout_1 | layout_2 | layout_3 | layout_4 -- | -- | -- | -- | -- 1 | 0.0762 | 0.0766 | 0.0758 | 0.0907 2 | 0.0911 | 0.0931 | 0.0957 | 0.0939 4 | 0.1197 | 0.124 | 0.1188 | 0.1257 8 | 0.1979 | 0.1942 | 0.2026 | 0.2208 16 | 0.3453 | 0.3577 | 0.3427 | 0.3833 32 | 0.6613 | 0.7161 | 0.6615 | 0.7574 256 | 5.5837 | 5.3269 | 5.772 | 5.7058 kernel:3x3x256x256 feature maps:14x14x256: batch | layout_1 | layout_2 | layout_3 | layout_4 -- | -- | -- | -- | -- 1 | 0.0633 | 0.0694 | 0.0703 | 0.656 2 | 0.0825 | 0.0835 | 0.0822 | 0.851 4 | 0.1417 | 0.1562 | 0.1382 | 0.1502 8 | 0.1829 | 0.1853 | 0.1833 | 0.182 16 | 0.264 | 0.277 | 0.2832 | 0.2621 32 | 0.4506 | 0.4799 | 0.4507 | 0.4773 256 | 3.9432 | 4.0867 | 3.5088 | 4.5032 Note: weight transform was pre-computed. The benchmarks were running on T4 GPU (16GB, 70W). Latency is reported with unit of ms. We can see that the performance of all layouts are in the same level. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected]
