mirror of
https://github.com/pytorch/pytorch.git
synced 2025-10-20 12:54:11 +08:00
9642c7568967ab424c5d0e04ef2cd1e82a54b5f8
10 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
| c17ba69ba5 |
[submodule] Revert "Adds support for accelerated sorting with x86-simd-sort (#127936) (#141901)
Looks like the original PR caused: https://github.com/pytorch/pytorch/issues/140590 Please see comment: https://github.com/pytorch/pytorch/issues/140590#issuecomment-2508704480 Pull Request resolved: https://github.com/pytorch/pytorch/pull/141901 Approved by: https://github.com/andrewor14, https://github.com/malfet |
|||
| 7e65060410 |
Adds support for accelerated sorting with x86-simd-sort (#127936)
Adds x86-simd-sort as a submodule to accelerate sorting for 32-bit and 64-bit datatypes when AVX2 or AVX512 are available. For contiguous data, this can be over a 10x speedup for large arrays. For discontiguous data, it can give over a 4x speedup with larger arrays. These benchmarks were gathered on a Skylake system (7900x), limited to 8 threads. <details> <summary><b>Contiguous Benchmarks</b></summary> ``` float32, normally distributed (in microseconds) size Default AVX2 AVX512 Default/AVX2 Default/AVX512 16 7.150844336 6.886271477 7.132277489 1.038420335 1.002603214 128 9.208030939 8.478154898 7.846915245 1.086089019 1.173458697 1024 37.79037627 23.60707456 16.44122627 1.600807257 2.298513241 10000 714.7355628 203.9921844 105.5683001 3.503739934 6.770361577 100000 8383.074408 721.6333354 465.3709247 11.61680593 18.01374766 1000000 97124.31945 5632.054572 3920.148401 17.24491803 24.77567416 10000000 1161974.907 86070.48988 71533.82301 13.50027063 16.24371323 int32_t, uniformly distributed (in microseconds) size Default AVX2 AVX512 Default/AVX2 Default/AVX512 16 7.203208685 6.92212224 7.014458179 1.040606975 1.026908779 128 8.972388983 8.195516348 7.592543125 1.094792396 1.18173698 1024 32.77489477 23.6874548 15.36617105 1.383639359 2.132925285 10000 607.8824128 193.3402024 99.25090471 3.144107667 6.124703997 100000 523.9384684 608.1836536 442.3166784 0.861480682 1.184532472 1000000 5211.348627 5271.598405 3518.861883 0.988570871 1.480975611 10000000 133853.6263 81463.05084 67852.97394 1.643120714 1.972700952 ``` </details> Note that the int32_t sort is accelerated by FBGEMM's radix sort for larger arrays, but this only handles contiguous data and in one sorting direction. <details> <summary><b>Discontiguous Benchmarks</b></summary> ``` float, normal distributed, discontiguous in sorted dimension (in microseconds) size Default AVX2 AVX512 Default/AVX2 Default/AVX512 16 3.836543679 4.011214256 3.84376061 0.956454439 0.99812243 128 5.755310194 5.755723127 4.820394962 0.999928257 1.193949923 1024 49.46946019 24.78790785 15.47874362 1.995709379 3.195960952 10000 665.2505291 236.6165959 143.9490662 2.811512551 4.621429974 100000 4328.002203 1329.001212 818.3516414 3.256582586 5.288682743 1000000 47651.5018 16693.72045 11827.39551 2.854456677 4.028909133 10000000 556655.1288 236252.6258 184215.9828 2.356185998 3.021752621 int32_t, uniformly distributed, discontiguous in sorted dimension (in microseconds) size Default AVX2 AVX512 Default/AVX2 Default/AVX512 16 3.817994356 3.878117442 3.770039797 0.984496837 1.012719908 128 5.578731397 5.577152082 4.716770534 1.000283176 1.182743862 1024 43.3412619 23.61275801 14.55446819 1.835501887 2.977866408 10000 634.3997478 224.4322851 133.9518324 2.826686667 4.736028889 100000 4084.358152 1292.363303 781.7867576 3.16037924 5.22438902 1000000 46262.20465 16608.35284 11367.51817 2.785478192 4.06968381 10000000 541231.9104 235185.1861 180249.9294 2.301301028 3.002674742 ``` </details> Pull Request resolved: https://github.com/pytorch/pytorch/pull/127936 Approved by: https://github.com/jgong5, https://github.com/peterbell10, https://github.com/sanchitintel |
|||
| 0e19522122 |
Revert "Adds support for accelerated sorting with x86-simd-sort (#127936)"
This reverts commit 239a9ad65eebf93dcf9bb108a5129d4160b12c86.
Reverted https://github.com/pytorch/pytorch/pull/127936 on behalf of https://github.com/atalman due to test/test_sort_and_select.py::TestSortAndSelectCPU::test_sort_discontiguous_slow_cpu_float32 [GH job link](https://github.com/pytorch/pytorch/actions/runs/10994904767/job/30525578456) [HUD commit link](
|
|||
| 239a9ad65e |
Adds support for accelerated sorting with x86-simd-sort (#127936)
Adds x86-simd-sort as a submodule to accelerate sorting for 32-bit and 64-bit datatypes when AVX2 or AVX512 are available. For contiguous data, this can be over a 10x speedup for large arrays. For discontiguous data, it can give over a 4x speedup with larger arrays. These benchmarks were gathered on a Skylake system (7900x), limited to 8 threads. <details> <summary><b>Contiguous Benchmarks</b></summary> ``` float32, normally distributed (in microseconds) size Default AVX2 AVX512 Default/AVX2 Default/AVX512 16 7.150844336 6.886271477 7.132277489 1.038420335 1.002603214 128 9.208030939 8.478154898 7.846915245 1.086089019 1.173458697 1024 37.79037627 23.60707456 16.44122627 1.600807257 2.298513241 10000 714.7355628 203.9921844 105.5683001 3.503739934 6.770361577 100000 8383.074408 721.6333354 465.3709247 11.61680593 18.01374766 1000000 97124.31945 5632.054572 3920.148401 17.24491803 24.77567416 10000000 1161974.907 86070.48988 71533.82301 13.50027063 16.24371323 int32_t, uniformly distributed (in microseconds) size Default AVX2 AVX512 Default/AVX2 Default/AVX512 16 7.203208685 6.92212224 7.014458179 1.040606975 1.026908779 128 8.972388983 8.195516348 7.592543125 1.094792396 1.18173698 1024 32.77489477 23.6874548 15.36617105 1.383639359 2.132925285 10000 607.8824128 193.3402024 99.25090471 3.144107667 6.124703997 100000 523.9384684 608.1836536 442.3166784 0.861480682 1.184532472 1000000 5211.348627 5271.598405 3518.861883 0.988570871 1.480975611 10000000 133853.6263 81463.05084 67852.97394 1.643120714 1.972700952 ``` </details> Note that the int32_t sort is accelerated by FBGEMM's radix sort for larger arrays, but this only handles contiguous data and in one sorting direction. <details> <summary><b>Discontiguous Benchmarks</b></summary> ``` float, normal distributed, discontiguous in sorted dimension (in microseconds) size Default AVX2 AVX512 Default/AVX2 Default/AVX512 16 3.836543679 4.011214256 3.84376061 0.956454439 0.99812243 128 5.755310194 5.755723127 4.820394962 0.999928257 1.193949923 1024 49.46946019 24.78790785 15.47874362 1.995709379 3.195960952 10000 665.2505291 236.6165959 143.9490662 2.811512551 4.621429974 100000 4328.002203 1329.001212 818.3516414 3.256582586 5.288682743 1000000 47651.5018 16693.72045 11827.39551 2.854456677 4.028909133 10000000 556655.1288 236252.6258 184215.9828 2.356185998 3.021752621 int32_t, uniformly distributed, discontiguous in sorted dimension (in microseconds) size Default AVX2 AVX512 Default/AVX2 Default/AVX512 16 3.817994356 3.878117442 3.770039797 0.984496837 1.012719908 128 5.578731397 5.577152082 4.716770534 1.000283176 1.182743862 1024 43.3412619 23.61275801 14.55446819 1.835501887 2.977866408 10000 634.3997478 224.4322851 133.9518324 2.826686667 4.736028889 100000 4084.358152 1292.363303 781.7867576 3.16037924 5.22438902 1000000 46262.20465 16608.35284 11367.51817 2.785478192 4.06968381 10000000 541231.9104 235185.1861 180249.9294 2.301301028 3.002674742 ``` </details> Pull Request resolved: https://github.com/pytorch/pytorch/pull/127936 Approved by: https://github.com/jgong5, https://github.com/peterbell10 |
|||
| 544c04f2df |
Add uint8 support for interpolate for CPU images (#90771)
Joint work with @vfdev-5
This PR introduces native uint8 support for `interpolate()`, for `bilinear` ~and `bicubic`~ modes for CPU images (`mode=nearest[_exact]` was already supported ).
On a typical torchvision training job on ImageNet, the speedup are ~4X when AVX2 is supported, comparing the uint8 native (this PR) vs torchvision's current `Resize()`:
```
AA = antialias
float = uint8->float->interpolate()->round()->clamp()->uint8 (what Resize() currently does)
input_size output_size channels_last AA mode num_threads speed-up float vs uint8 (this PR)
(1, 3, 270, 268) -> (224, 224) True True bilinear num_threads=1 4X 2.6ms vs 0.7ms
(1, 3, 270, 268) -> (224, 224) True False bilinear num_threads=1 2.1X 1.3ms vs 0.6ms
(1, 3, 270, 268) -> (224, 224) False True bilinear num_threads=1 3X 2.1ms vs 0.7ms
(1, 3, 270, 268) -> (224, 224) False False bilinear num_threads=1 4X 2.4ms vs 0.6ms
(Note: we removed bicubic support for now)
(1, 3, 270, 268) -> (224, 224) True True bicubic num_threads=1 4X 2.9ms vs 0.7ms
(1, 3, 270, 268) -> (224, 224) True False bicubic num_threads=1 5X 3.1ms vs 0.7ms
(1, 3, 270, 268) -> (224, 224) False True bicubic num_threads=1 3X 2.4ms vs 0.7ms
(1, 3, 270, 268) -> (224, 224) False False bicubic num_threads=1 4X 2.8ms vs 0.7ms
```
There is still room for further speed-ups (see TODOs in the code).
#### More benchmark details
with AVX2 support - speedups typically range from 1.5X to 10X. A few edge-cases are slower, worth investigating why.
<details>
```
AA = antialias
float = uint8->float->interpolate()->round()->clamp()->uint8 (what Resize() currently does)
input_size output_size channels_last AA mode num_threads speed-up float vs uint8 (this PR)
(1, 3, 64, 64) -> (224, 224) True True bilinear num_threads=1 5X 1.1ms vs 0.2ms
(1, 3, 64, 64) -> (224, 224) True False bilinear num_threads=1 5X 1.2ms vs 0.2ms
(1, 3, 64, 64) -> (224, 224) False True bilinear num_threads=1 2.8X 0.6ms vs 0.2ms
(1, 3, 64, 64) -> (224, 224) False False bilinear num_threads=1 7X 1.6ms vs 0.2ms
(1, 3, 64, 64) -> (224, 224) True True bicubic num_threads=1 5X 1.2ms vs 0.2ms
(1, 3, 64, 64) -> (224, 224) True False bicubic num_threads=1 12X 2.9ms vs 0.2ms
(1, 3, 64, 64) -> (224, 224) False True bicubic num_threads=1 3X 0.8ms vs 0.2ms
(1, 3, 64, 64) -> (224, 224) False False bicubic num_threads=1 7X 1.8ms vs 0.2ms
(1, 3, 64, 64) -> (224, 224) True True bilinear num_threads=2 2.6X 0.6ms vs 0.2ms
(1, 3, 64, 64) -> (224, 224) True False bilinear num_threads=2 2.8X 0.6ms vs 0.2ms
(1, 3, 64, 64) -> (224, 224) False True bilinear num_threads=2 1.7X 0.4ms vs 0.2ms
(1, 3, 64, 64) -> (224, 224) False False bilinear num_threads=2 1.4X 0.3ms vs 0.2ms
(1, 3, 64, 64) -> (224, 224) True True bicubic num_threads=2 2.7X 0.7ms vs 0.2ms
(1, 3, 64, 64) -> (224, 224) True False bicubic num_threads=2 7X 1.6ms vs 0.2ms
(1, 3, 64, 64) -> (224, 224) False True bicubic num_threads=2 1.8X 0.4ms vs 0.2ms
(1, 3, 64, 64) -> (224, 224) False False bicubic num_threads=2 4X 1.0ms vs 0.2ms
(1, 3, 224, 224) -> (270, 268) True True bilinear num_threads=1 4X 2.5ms vs 0.6ms
(1, 3, 224, 224) -> (270, 268) True False bilinear num_threads=1 3.0X 1.8ms vs 0.6ms
(1, 3, 224, 224) -> (270, 268) False True bilinear num_threads=1 3X 1.8ms vs 0.6ms
(1, 3, 224, 224) -> (270, 268) False False bilinear num_threads=1 4X 2.3ms vs 0.6ms
(1, 3, 224, 224) -> (270, 268) True True bicubic num_threads=1 4X 2.7ms vs 0.6ms
(1, 3, 224, 224) -> (270, 268) True False bicubic num_threads=1 7X 4.3ms vs 0.6ms
(1, 3, 224, 224) -> (270, 268) False True bicubic num_threads=1 3X 2.1ms vs 0.6ms
(1, 3, 224, 224) -> (270, 268) False False bicubic num_threads=1 4X 2.6ms vs 0.6ms
(1, 3, 224, 224) -> (270, 268) True True bilinear num_threads=2 2.7X 1.6ms vs 0.6ms
(1, 3, 224, 224) -> (270, 268) True False bilinear num_threads=2 2.6X 1.5ms vs 0.6ms
(1, 3, 224, 224) -> (270, 268) False True bilinear num_threads=2 2.1X 1.2ms vs 0.6ms
(1, 3, 224, 224) -> (270, 268) False False bilinear num_threads=2 1.6X 0.9ms vs 0.6ms
(1, 3, 224, 224) -> (270, 268) True True bicubic num_threads=2 2.8X 1.7ms vs 0.6ms
(1, 3, 224, 224) -> (270, 268) True False bicubic num_threads=2 5X 2.8ms vs 0.6ms
(1, 3, 224, 224) -> (270, 268) False True bicubic num_threads=2 2.3X 1.4ms vs 0.6ms
(1, 3, 224, 224) -> (270, 268) False False bicubic num_threads=2 3X 1.9ms vs 0.6ms
(1, 3, 256, 256) -> (1024, 1024) True True bilinear num_threads=1 4X 26.6ms vs 6.7ms
(1, 3, 256, 256) -> (1024, 1024) True False bilinear num_threads=1 4X 23.9ms vs 6.8ms
(1, 3, 256, 256) -> (1024, 1024) False True bilinear num_threads=1 2.5X 16.8ms vs 6.8ms
(1, 3, 256, 256) -> (1024, 1024) False False bilinear num_threads=1 5X 33.1ms vs 6.8ms
(1, 3, 256, 256) -> (1024, 1024) True True bicubic num_threads=1 4X 25.9ms vs 7.3ms
(1, 3, 256, 256) -> (1024, 1024) True False bicubic num_threads=1 8X 59.6ms vs 7.3ms
(1, 3, 256, 256) -> (1024, 1024) False True bicubic num_threads=1 1.9X 14.3ms vs 7.4ms
(1, 3, 256, 256) -> (1024, 1024) False False bicubic num_threads=1 5X 35.4ms vs 7.3ms
(1, 3, 256, 256) -> (1024, 1024) True True bilinear num_threads=2 2.0X 13.6ms vs 6.8ms
(1, 3, 256, 256) -> (1024, 1024) True False bilinear num_threads=2 2.2X 14.8ms vs 6.7ms
(1, 3, 256, 256) -> (1024, 1024) False True bilinear num_threads=2 1.3X 8.8ms vs 6.9ms
(1, 3, 256, 256) -> (1024, 1024) False False bilinear num_threads=2 1.2X 8.4ms vs 6.8ms
(1, 3, 256, 256) -> (1024, 1024) True True bicubic num_threads=2 1.8X 12.8ms vs 7.3ms
(1, 3, 256, 256) -> (1024, 1024) True False bicubic num_threads=2 4X 32.1ms vs 7.2ms
(1, 3, 256, 256) -> (1024, 1024) False True bicubic num_threads=2 1.4X 10.1ms vs 7.3ms
(1, 3, 256, 256) -> (1024, 1024) False False bicubic num_threads=2 2.9X 20.9ms vs 7.3ms
(1, 3, 224, 224) -> (64, 64) True True bilinear num_threads=1 1.4X 0.5ms vs 0.3ms
(1, 3, 224, 224) -> (64, 64) True False bilinear num_threads=1 0.7X 0.2ms vs 0.3ms
(1, 3, 224, 224) -> (64, 64) False True bilinear num_threads=1 1.3X 0.4ms vs 0.3ms
(1, 3, 224, 224) -> (64, 64) False False bilinear num_threads=1 1.4X 0.4ms vs 0.3ms
(1, 3, 224, 224) -> (64, 64) True True bicubic num_threads=1 2.1X 0.7ms vs 0.3ms
(1, 3, 224, 224) -> (64, 64) True False bicubic num_threads=1 1.3X 0.4ms vs 0.3ms
(1, 3, 224, 224) -> (64, 64) False True bicubic num_threads=1 1.9X 0.6ms vs 0.3ms
(1, 3, 224, 224) -> (64, 64) False False bicubic num_threads=1 1.0X 0.3ms vs 0.3ms
(1, 3, 224, 224) -> (64, 64) True True bilinear num_threads=2 1.0X 0.3ms vs 0.3ms
(1, 3, 224, 224) -> (64, 64) True False bilinear num_threads=2 0.6X 0.2ms vs 0.3ms
(1, 3, 224, 224) -> (64, 64) False True bilinear num_threads=2 0.8X 0.3ms vs 0.3ms
(1, 3, 224, 224) -> (64, 64) False False bilinear num_threads=2 1.4X 0.4ms vs 0.3ms
(1, 3, 224, 224) -> (64, 64) True True bicubic num_threads=2 1.4X 0.5ms vs 0.3ms
(1, 3, 224, 224) -> (64, 64) True False bicubic num_threads=2 1.2X 0.4ms vs 0.3ms
(1, 3, 224, 224) -> (64, 64) False True bicubic num_threads=2 1.2X 0.4ms vs 0.4ms
(1, 3, 224, 224) -> (64, 64) False False bicubic num_threads=2 0.9X 0.3ms vs 0.3ms
(1, 3, 270, 268) -> (224, 224) True True bilinear num_threads=1 4X 2.6ms vs 0.7ms
(1, 3, 270, 268) -> (224, 224) True False bilinear num_threads=1 2.1X 1.3ms vs 0.6ms
(1, 3, 270, 268) -> (224, 224) False True bilinear num_threads=1 3X 2.1ms vs 0.7ms
(1, 3, 270, 268) -> (224, 224) False False bilinear num_threads=1 4X 2.4ms vs 0.6ms
(1, 3, 270, 268) -> (224, 224) True True bicubic num_threads=1 4X 2.9ms vs 0.7ms
(1, 3, 270, 268) -> (224, 224) True False bicubic num_threads=1 5X 3.1ms vs 0.7ms
(1, 3, 270, 268) -> (224, 224) False True bicubic num_threads=1 3X 2.4ms vs 0.7ms
(1, 3, 270, 268) -> (224, 224) False False bicubic num_threads=1 4X 2.8ms vs 0.7ms
(1, 3, 270, 268) -> (224, 224) True True bilinear num_threads=2 1.5X 1.0ms vs 0.7ms
(1, 3, 270, 268) -> (224, 224) True False bilinear num_threads=2 1.2X 0.8ms vs 0.6ms
(1, 3, 270, 268) -> (224, 224) False True bilinear num_threads=2 2.3X 1.5ms vs 0.7ms
(1, 3, 270, 268) -> (224, 224) False False bilinear num_threads=2 1.9X 1.2ms vs 0.6ms
(1, 3, 270, 268) -> (224, 224) True True bicubic num_threads=2 1.6X 1.2ms vs 0.7ms
(1, 3, 270, 268) -> (224, 224) True False bicubic num_threads=2 4X 2.4ms vs 0.7ms
(1, 3, 270, 268) -> (224, 224) False True bicubic num_threads=2 2.4X 1.6ms vs 0.7ms
(1, 3, 270, 268) -> (224, 224) False False bicubic num_threads=2 2.8X 1.8ms vs 0.6ms
(1, 3, 1024, 1024) -> (256, 256) True True bilinear num_threads=1 2.1X 12.8ms vs 6.1ms
(1, 3, 1024, 1024) -> (256, 256) True False bilinear num_threads=1 0.6X 3.8ms vs 5.9ms
(1, 3, 1024, 1024) -> (256, 256) False True bilinear num_threads=1 1.2X 7.1ms vs 6.1ms
(1, 3, 1024, 1024) -> (256, 256) False False bilinear num_threads=1 1.9X 11.0ms vs 5.9ms
(1, 3, 1024, 1024) -> (256, 256) True True bicubic num_threads=1 2.0X 12.6ms vs 6.4ms
(1, 3, 1024, 1024) -> (256, 256) True False bicubic num_threads=1 1.0X 6.1ms vs 6.0ms
(1, 3, 1024, 1024) -> (256, 256) False True bicubic num_threads=1 1.8X 11.3ms vs 6.4ms
(1, 3, 1024, 1024) -> (256, 256) False False bicubic num_threads=1 0.8X 4.6ms vs 6.0ms
(1, 3, 1024, 1024) -> (256, 256) True True bilinear num_threads=2 1.6X 9.3ms vs 6.0ms
(1, 3, 1024, 1024) -> (256, 256) True False bilinear num_threads=2 0.3X 2.0ms vs 5.8ms
(1, 3, 1024, 1024) -> (256, 256) False True bilinear num_threads=2 1.2X 7.2ms vs 6.0ms
(1, 3, 1024, 1024) -> (256, 256) False False bilinear num_threads=2 0.3X 1.6ms vs 5.8ms
(1, 3, 1024, 1024) -> (256, 256) True True bicubic num_threads=2 1.1X 7.1ms vs 6.5ms
(1, 3, 1024, 1024) -> (256, 256) True False bicubic num_threads=2 0.6X 3.3ms vs 5.9ms
(1, 3, 1024, 1024) -> (256, 256) False True bicubic num_threads=2 0.9X 5.9ms vs 6.3ms
(1, 3, 1024, 1024) -> (256, 256) False False bicubic num_threads=2 0.4X 2.4ms vs 5.9ms
```
</details>
without AVX2 support - no significant speed-up, but there are various possible improvements (see TODOs)
<details>
```
AA = antialias
float = uint8->float->interpolate()->round()->clamp()->uint8 (what Resize() currently does)
input_size output_size channels_last AA mode num_threads speed-up float vs uint8 (this PR)
(1, 3, 64, 64) -> (224, 224) True True bilinear num_threads=1 0.9X 1.5ms vs 1.6ms
(1, 3, 64, 64) -> (224, 224) True False bilinear num_threads=1 0.9X 1.5ms vs 1.6ms
(1, 3, 64, 64) -> (224, 224) False True bilinear num_threads=1 0.8X 0.9ms vs 1.1ms
(1, 3, 64, 64) -> (224, 224) False False bilinear num_threads=1 1.5X 1.7ms vs 1.1ms
(1, 3, 64, 64) -> (224, 224) True True bicubic num_threads=1 0.9X 1.6ms vs 1.8ms
(1, 3, 64, 64) -> (224, 224) True False bicubic num_threads=1 2.1X 3.9ms vs 1.9ms
(1, 3, 64, 64) -> (224, 224) False True bicubic num_threads=1 0.8X 1.1ms vs 1.4ms
(1, 3, 64, 64) -> (224, 224) False False bicubic num_threads=1 1.7X 2.4ms vs 1.5ms
(1, 3, 64, 64) -> (224, 224) True True bilinear num_threads=2 0.9X 0.8ms vs 0.8ms
(1, 3, 64, 64) -> (224, 224) True False bilinear num_threads=2 0.9X 0.8ms vs 0.8ms
(1, 3, 64, 64) -> (224, 224) False True bilinear num_threads=2 0.9X 0.5ms vs 0.6ms
(1, 3, 64, 64) -> (224, 224) False False bilinear num_threads=2 0.7X 0.5ms vs 0.7ms
(1, 3, 64, 64) -> (224, 224) True True bicubic num_threads=2 0.9X 0.9ms vs 1.0ms
(1, 3, 64, 64) -> (224, 224) True False bicubic num_threads=2 2.1X 2.0ms vs 1.0ms
(1, 3, 64, 64) -> (224, 224) False True bicubic num_threads=2 0.8X 0.6ms vs 0.8ms
(1, 3, 64, 64) -> (224, 224) False False bicubic num_threads=2 1.7X 1.3ms vs 0.8ms
(1, 3, 224, 224) -> (270, 268) True True bilinear num_threads=1 1.0X 3.0ms vs 3.0ms
(1, 3, 224, 224) -> (270, 268) True False bilinear num_threads=1 1.0X 2.8ms vs 2.9ms
(1, 3, 224, 224) -> (270, 268) False True bilinear num_threads=1 1.0X 2.3ms vs 2.2ms
(1, 3, 224, 224) -> (270, 268) False False bilinear num_threads=1 1.4X 3.3ms vs 2.3ms
(1, 3, 224, 224) -> (270, 268) True True bicubic num_threads=1 1.0X 3.5ms vs 3.5ms
(1, 3, 224, 224) -> (270, 268) True False bicubic num_threads=1 1.7X 6.1ms vs 3.5ms
(1, 3, 224, 224) -> (270, 268) False True bicubic num_threads=1 0.9X 2.6ms vs 2.9ms
(1, 3, 224, 224) -> (270, 268) False False bicubic num_threads=1 1.4X 4.2ms vs 2.9ms
(1, 3, 224, 224) -> (270, 268) True True bilinear num_threads=2 1.0X 1.7ms vs 1.7ms
(1, 3, 224, 224) -> (270, 268) True False bilinear num_threads=2 0.9X 1.6ms vs 1.8ms
(1, 3, 224, 224) -> (270, 268) False True bilinear num_threads=2 0.9X 1.3ms vs 1.4ms
(1, 3, 224, 224) -> (270, 268) False False bilinear num_threads=2 0.7X 1.1ms vs 1.6ms
(1, 3, 224, 224) -> (270, 268) True True bicubic num_threads=2 1.0X 2.0ms vs 2.0ms
(1, 3, 224, 224) -> (270, 268) True False bicubic num_threads=2 1.7X 3.2ms vs 1.9ms
(1, 3, 224, 224) -> (270, 268) False True bicubic num_threads=2 0.8X 1.5ms vs 1.9ms
(1, 3, 224, 224) -> (270, 268) False False bicubic num_threads=2 1.2X 2.3ms vs 1.9ms
(1, 3, 256, 256) -> (1024, 1024) True True bilinear num_threads=1 1.1X 34.7ms vs 32.4ms
(1, 3, 256, 256) -> (1024, 1024) True False bilinear num_threads=1 1.0X 31.2ms vs 32.4ms
(1, 3, 256, 256) -> (1024, 1024) False True bilinear num_threads=1 1.0X 23.5ms vs 22.7ms
(1, 3, 256, 256) -> (1024, 1024) False False bilinear num_threads=1 1.9X 42.5ms vs 22.7ms
(1, 3, 256, 256) -> (1024, 1024) True True bicubic num_threads=1 0.9X 33.9ms vs 37.4ms
(1, 3, 256, 256) -> (1024, 1024) True False bicubic num_threads=1 2.2X 84.0ms vs 37.5ms
(1, 3, 256, 256) -> (1024, 1024) False True bicubic num_threads=1 1.0X 28.4ms vs 28.8ms
(1, 3, 256, 256) -> (1024, 1024) False False bicubic num_threads=1 2.0X 56.7ms vs 28.8ms
(1, 3, 256, 256) -> (1024, 1024) True True bilinear num_threads=2 1.1X 17.5ms vs 16.4ms
(1, 3, 256, 256) -> (1024, 1024) True False bilinear num_threads=2 1.1X 17.7ms vs 16.4ms
(1, 3, 256, 256) -> (1024, 1024) False True bilinear num_threads=2 0.8X 8.8ms vs 11.4ms
(1, 3, 256, 256) -> (1024, 1024) False False bilinear num_threads=2 1.0X 11.1ms vs 11.4ms
(1, 3, 256, 256) -> (1024, 1024) True True bicubic num_threads=2 1.1X 19.9ms vs 18.8ms
(1, 3, 256, 256) -> (1024, 1024) True False bicubic num_threads=2 2.3X 42.5ms vs 18.7ms
(1, 3, 256, 256) -> (1024, 1024) False True bicubic num_threads=2 1.0X 14.1ms vs 14.5ms
(1, 3, 256, 256) -> (1024, 1024) False False bicubic num_threads=2 2.0X 28.4ms vs 14.5ms
(1, 3, 224, 224) -> (64, 64) True True bilinear num_threads=1 1.0X 0.6ms vs 0.6ms
(1, 3, 224, 224) -> (64, 64) True False bilinear num_threads=1 0.7X 0.3ms vs 0.4ms
(1, 3, 224, 224) -> (64, 64) False True bilinear num_threads=1 0.9X 0.5ms vs 0.6ms
(1, 3, 224, 224) -> (64, 64) False False bilinear num_threads=1 1.7X 0.6ms vs 0.4ms
(1, 3, 224, 224) -> (64, 64) True True bicubic num_threads=1 1.0X 0.8ms vs 0.8ms
(1, 3, 224, 224) -> (64, 64) True False bicubic num_threads=1 1.1X 0.5ms vs 0.5ms
(1, 3, 224, 224) -> (64, 64) False True bicubic num_threads=1 0.9X 0.7ms vs 0.8ms
(1, 3, 224, 224) -> (64, 64) False False bicubic num_threads=1 0.9X 0.4ms vs 0.4ms
(1, 3, 224, 224) -> (64, 64) True True bilinear num_threads=2 1.0X 0.4ms vs 0.4ms
(1, 3, 224, 224) -> (64, 64) True False bilinear num_threads=2 0.8X 0.2ms vs 0.3ms
(1, 3, 224, 224) -> (64, 64) False True bilinear num_threads=2 0.9X 0.3ms vs 0.3ms
(1, 3, 224, 224) -> (64, 64) False False bilinear num_threads=2 1.3X 0.3ms vs 0.2ms
(1, 3, 224, 224) -> (64, 64) True True bicubic num_threads=2 1.0X 0.5ms vs 0.5ms
(1, 3, 224, 224) -> (64, 64) True False bicubic num_threads=2 1.3X 0.4ms vs 0.3ms
(1, 3, 224, 224) -> (64, 64) False True bicubic num_threads=2 0.9X 0.5ms vs 0.5ms
(1, 3, 224, 224) -> (64, 64) False False bicubic num_threads=2 1.2X 0.3ms vs 0.3ms
(1, 3, 270, 268) -> (224, 224) True True bilinear num_threads=1 0.8X 2.1ms vs 2.5ms
(1, 3, 270, 268) -> (224, 224) True False bilinear num_threads=1 0.7X 1.6ms vs 2.4ms
(1, 3, 270, 268) -> (224, 224) False True bilinear num_threads=1 1.2X 2.4ms vs 2.1ms
(1, 3, 270, 268) -> (224, 224) False False bilinear num_threads=1 1.3X 2.6ms vs 2.0ms
(1, 3, 270, 268) -> (224, 224) True True bicubic num_threads=1 1.1X 3.4ms vs 3.0ms
(1, 3, 270, 268) -> (224, 224) True False bicubic num_threads=1 1.7X 4.8ms vs 2.8ms
(1, 3, 270, 268) -> (224, 224) False True bicubic num_threads=1 1.1X 2.9ms vs 2.7ms
(1, 3, 270, 268) -> (224, 224) False False bicubic num_threads=1 1.4X 3.5ms vs 2.4ms
(1, 3, 270, 268) -> (224, 224) True True bilinear num_threads=2 0.9X 1.2ms vs 1.3ms
(1, 3, 270, 268) -> (224, 224) True False bilinear num_threads=2 1.3X 1.6ms vs 1.2ms
(1, 3, 270, 268) -> (224, 224) False True bilinear num_threads=2 0.8X 0.9ms vs 1.1ms
(1, 3, 270, 268) -> (224, 224) False False bilinear num_threads=2 1.3X 1.3ms vs 1.0ms
(1, 3, 270, 268) -> (224, 224) True True bicubic num_threads=2 1.4X 2.2ms vs 1.6ms
(1, 3, 270, 268) -> (224, 224) True False bicubic num_threads=2 1.9X 2.8ms vs 1.5ms
(1, 3, 270, 268) -> (224, 224) False True bicubic num_threads=2 0.8X 1.1ms vs 1.4ms
(1, 3, 270, 268) -> (224, 224) False False bicubic num_threads=2 1.7X 2.1ms vs 1.3ms
(1, 3, 1024, 1024) -> (256, 256) True True bilinear num_threads=1 1.0X 10.0ms vs 9.9ms
(1, 3, 1024, 1024) -> (256, 256) True False bilinear num_threads=1 0.7X 4.6ms vs 6.2ms
(1, 3, 1024, 1024) -> (256, 256) False True bilinear num_threads=1 0.9X 9.1ms vs 9.8ms
(1, 3, 1024, 1024) -> (256, 256) False False bilinear num_threads=1 1.7X 9.4ms vs 5.7ms
(1, 3, 1024, 1024) -> (256, 256) True True bicubic num_threads=1 1.0X 15.2ms vs 14.8ms
(1, 3, 1024, 1024) -> (256, 256) True False bicubic num_threads=1 1.0X 7.6ms vs 7.5ms
(1, 3, 1024, 1024) -> (256, 256) False True bicubic num_threads=1 0.9X 13.3ms vs 14.4ms
(1, 3, 1024, 1024) -> (256, 256) False False bicubic num_threads=1 0.8X 5.9ms vs 7.0ms
(1, 3, 1024, 1024) -> (256, 256) True True bilinear num_threads=2 1.2X 6.0ms vs 5.2ms
(1, 3, 1024, 1024) -> (256, 256) True False bilinear num_threads=2 0.7X 2.3ms vs 3.2ms
(1, 3, 1024, 1024) -> (256, 256) False True bilinear num_threads=2 1.0X 4.8ms vs 5.0ms
(1, 3, 1024, 1024) -> (256, 256) False False bilinear num_threads=2 0.7X 1.9ms vs 2.9ms
(1, 3, 1024, 1024) -> (256, 256) True True bicubic num_threads=2 1.6X 12.3ms vs 7.5ms
(1, 3, 1024, 1024) -> (256, 256) True False bicubic num_threads=2 1.0X 3.9ms vs 3.9ms
(1, 3, 1024, 1024) -> (256, 256) False True bicubic num_threads=2 1.0X 7.0ms vs 7.3ms
(1, 3, 1024, 1024) -> (256, 256) False False bicubic num_threads=2 0.9X 3.0ms vs 3.5ms
```
</details>
Benchmark code
<details>
```py
import operator_benchmark as op_bench
import torch
"""Microbenchmarks for interpolate operator."""
class InterpolateBenchmark(op_bench.TorchBenchmarkBase):
def init(self, input_size, output_size, channels_last=False, mode='linear', antialias=False, dtype=torch.float):
input_image = torch.randint(0, 256, size=input_size, dtype=torch.uint8, device='cpu')
if channels_last:
input_image = input_image.contiguous(memory_format=torch.channels_last)
self.inputs = {
"input_image": input_image,
"output_size": output_size,
"mode": mode,
"antialias": antialias,
"dtype":dtype,
}
self.set_module_name("interpolate")
def forward(self, input_image, output_size, mode, antialias, dtype):
if dtype == torch.float:
input_image = input_image.float()
out = torch.nn.functional.interpolate(input_image, size=output_size, mode=mode, align_corners=False, antialias=antialias)
if dtype == torch.float:
out = out.round().clamp(min=0, max=256).to(torch.uint8)
def make_config():
sizes = (
((224, 224), (64, 64)),
((270, 268), (224, 224)),
((256, 256), (1024, 1024)),
)
attrs = []
for (HW1, HW2) in sizes:
attrs.append([(1, 3, *HW1), HW2]) # 3 channels
# attrs.append([(1, 1, *HW1), HW2]) # 1 channel
attrs.append([(1, 3, *HW2), HW1]) # 3 channels
# attrs.append([(1, 1, *HW2), HW1]) # 1 channel
config = op_bench.config_list(
attr_names=["input_size", "output_size"],
attrs=attrs,
cross_product_configs={
'channels_last': [True, False],
'mode': ["bilinear", "bicubic"],
'antialias': [True, False],
# 'dtype': [torch.float, torch.uint8]
# 'dtype': [torch.uint8]
'dtype': [torch.float]
},
tags=["short"],
)
return config
config = make_config()
op_bench.generate_pt_test(config, InterpolateBenchmark)
if __name__ == "__main__":
op_bench.benchmark_runner.main()
```
```py
import re
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("f1", nargs="?", default="main")
parser.add_argument("f2", nargs="?", default="new")
args = parser.parse_args()
with open(args.f1) as f:
main = f.readlines()
with open(args.f2) as f:
new = f.readlines()
out = []
for main_line, new_line in zip(main, new):
# num_threads=1 # TODO: remove
if main_line.startswith("num_threads="):
num_threads = int(main_line.split("=")[-1])
if main_line.startswith("# Input"):
deets = f"{main_line.strip()}, {num_threads=}"
if main_line.startswith("Forward"):
main_time = float(main_line.split()[-1])
new_time = float(new_line.split()[-1])
ratio = main_time / new_time
fmt = ".1f" if ratio < 3 else ".0f"
improv = f"{ratio:{fmt}}X"
time_fmt = ",.3f" if new_time < 100 else ",.1f"
deets = deets.strip().replace("# Input: ", "")
deets = deets.replace(": ", "=")
deets = deets.replace("input_size=", "")
deets = deets.replace(", output_size=", " -> ")
deets = deets.replace("dtype=torch.", "")
deets = deets.replace("mode=", "")
deets = deets.replace("antialias=", "")
deets = deets.replace("channels_last=", "")
# deets = deets.replace("channels_last=True, ", "")
split = deets.split(",")
# size = ','.join(split[:-3])
# mode, dtype, threads = split[-3:]
# deets = f"{size:<30} {mode:<15} {dtype:<10} {threads:<15}"
size = ','.join(split[:-5])
channels_last, mode, antialias, dtype, threads= split[-5:]
deets = f"{size:<33} {channels_last:<7} {antialias:<7} {mode:<10} {threads:<15}"
l = f"{deets} {improv:<5} {main_time / 1000:{time_fmt}}ms vs {new_time / 1000:{time_fmt}}ms"
out.append(l)
def key(s):
# s = ''.join(s.split()[1:]) # remove "N.nX" part
num_threads = (int(re.findall(r"num_threads=(\d+)", s)[0]),)
input_shape, output_shape = re.findall("\(.*?\)", s)
input_shape = input_shape[1:-1] # remove parenthesis
input_HW = tuple(int(x) for x in input_shape.split(",")[-2:])
input_C = (-int(input_shape.split(",")[1]),)
output_HW = tuple(int(x) for x in output_shape[1:-1].split(","))
is_downsample = (output_HW[0] < input_HW[0],)
if "linear" in s:
mode = "linear"
elif "nearest-exact" in s:
mode = "nearest-exact"
else:
# assert "nearest" in s
mode = "nearest"
mode = (mode,)
return is_downsample + input_HW + output_HW + num_threads + input_C + mode
for i, l in enumerate(sorted(out, key=key)):
if i % 8 == 0:
print()
# if i % 10 == 0 and i % 40 != 0:
# print()
# if i % 40 == 0:
# print("-" * 100)
print(l)
```
</details>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/90771
Approved by: https://github.com/peterbell10, https://github.com/ngimel
|
|||
| 6eaa324c9f |
Implement torch.igamma (#46183)
Summary: Fixes https://github.com/pytorch/pytorch/issues/41637 This is regularized lower incomplete gamma function, equivalent to scipy's `gammainc` and tensorflow `igamma`. cc fritzo mruberry Pull Request resolved: https://github.com/pytorch/pytorch/pull/46183 Reviewed By: gchanan Differential Revision: D24479126 Pulled By: mruberry fbshipit-source-id: fdf8ea289fe4ca1b408810732192411e948fcdfe |
|||
| 06d50b5eb0 |
Pull in fairscale.nn.Pipe into PyTorch. (#44090)
Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/44090 This is an initial commit pulling in the torchgpipe fork at https://github.com/facebookresearch/fairscale. The purpose of this commit is to just pull in the code and ensure all tests and builds work fine. We will slowly modify this to match our intended API mentioned in https://fb.quip.com/txurAV3zIFox#RPZACAfAKMq. Follow up PRs would address further changes needed on top of the initial commit.. We're pulling the code into the `torch.distributed._pipeline.sync` package. The package is private on purpose since there is a lot of work (ex: docs, API changes etc.) that needs to go in before we can actually officially support this. ghstack-source-id: 114864254 Test Plan: 1) waitforbuildbot 2) Ran all tests on my devgpu Reviewed By: mrshenli Differential Revision: D23493316 fbshipit-source-id: fe3c8b7dadeeb86abdc00e8a8652491b0b16743a |
|||
| a161639fcd |
Move copyright lines back to NOTICE file, fixes #6911 (#8310)
Signed-off-by: Edward Z. Yang <ezyang@cs.stanford.edu> |
|||
| 90afedb6e2 | Merge caffe2 with pytorch. | |||
| 8286ce1e3a |
Re-license to Apache
Summary: Closes https://github.com/caffe2/caffe2/pull/1260 Differential Revision: D5906739 Pulled By: Yangqing fbshipit-source-id: e482ba9ba60b5337d9165f28f7ec68d4518a0902 |
