| commit | 93b2d991581db86074dd8011fdc903bd554466b1 | [log] [tgz] |
|---|---|---|
| author | Nicolas Hug <nicolashug@fb.com> | Thu Oct 06 11:32:29 2022 +0000 |
| committer | PyTorch MergeBot <pytorchmergebot@users.noreply.github.com> | Fri Oct 07 07:52:36 2022 +0000 |
| tree | 9ceec6c530ff988503c025d799ee97bd8fea1f18 | |
| parent | 70c6a988d6b5cabed84c686316b6bbeb235cc05c [diff] |
Improve interpolate() speed for channels_last images and masks (#86361) This PR improves the speed of `interpolate()`: - on images and masks (`num_channels < 4`, `channels_last=True`) - for the following modes: linear (antialias=False), nearest (int and float), and nearest-exact (int and float) - for both upsampling and downsampling The actual speed-up ranges from 1.1X to 110X, but this depends on various factors like number of threads and of course input_size/output_size. In a typical torchvision ImageNet training job (where num_threads=1 because of DataLoader multi-processing), the following speed-ups should be expected (I ran much more benchmarks than this one, see below for more details): ``` (1, 3, 600, 400) -> (224, 224) linear float32 num_threads=1 1.0X 1.0ms vs 1.0ms (1, 3, 600, 400) -> (224, 224) nearest float32 num_threads=1 1.9X 0.9ms vs 0.5ms (1, 3, 600, 400) -> (224, 224) nearest uint8 num_threads=1 1.7X 0.9ms vs 0.5ms (1, 3, 600, 400) -> (224, 224) nearest-exact float32 num_threads=1 2.1X 1.0ms vs 0.5ms (1, 3, 600, 400) -> (224, 224) nearest-exact uint8 num_threads=1 1.8X 0.9ms vs 0.5ms (1, 1, 600, 400) -> (224, 224) linear float32 num_threads=1 7X 0.8ms vs 0.1ms (1, 1, 600, 400) -> (224, 224) nearest float32 num_threads=1 14X 0.852ms vs 0.061ms (1, 1, 600, 400) -> (224, 224) nearest uint8 num_threads=1 9X 0.828ms vs 0.087ms (1, 1, 600, 400) -> (224, 224) nearest-exact float32 num_threads=1 15X 0.922ms vs 0.061ms (1, 1, 600, 400) -> (224, 224) nearest-exact uint8 num_threads=1 10X 0.897ms vs 0.087ms ``` An immediate follow-up to this PR would be to do the same changes for the 3D kernels. Thanks a ton @fmassa for the help! ### Speedup benchmarks: Results: <details> ``` ---------------------------------------------------------------------------------------------------- (1, 3, 64, 64) -> (224, 224) linear float32 num_threads=1 0.9X 0.9ms vs 1.1ms (1, 3, 64, 64) -> (224, 224) nearest float32 num_threads=1 1.6X 0.9ms vs 0.5ms (1, 3, 64, 64) -> (224, 224) nearest uint8 num_threads=1 1.7X 0.9ms vs 0.5ms (1, 3, 64, 64) -> (224, 224) nearest-exact float32 num_threads=1 1.7X 1.0ms vs 0.5ms (1, 3, 64, 64) -> (224, 224) nearest-exact uint8 num_threads=1 1.9X 0.9ms vs 0.5ms (1, 1, 64, 64) -> (224, 224) linear float32 num_threads=1 8X 0.806ms vs 0.097ms (1, 1, 64, 64) -> (224, 224) nearest float32 num_threads=1 15X 0.848ms vs 0.056ms (1, 1, 64, 64) -> (224, 224) nearest uint8 num_threads=1 10X 0.828ms vs 0.084ms (1, 1, 64, 64) -> (224, 224) nearest-exact float32 num_threads=1 16X 0.914ms vs 0.057ms (1, 1, 64, 64) -> (224, 224) nearest-exact uint8 num_threads=1 10X 0.900ms vs 0.086ms (1, 3, 64, 64) -> (224, 224) linear float32 num_threads=2 1.6X 1.1ms vs 0.7ms (1, 3, 64, 64) -> (224, 224) nearest float32 num_threads=2 1.6X 0.6ms vs 0.4ms (1, 3, 64, 64) -> (224, 224) nearest uint8 num_threads=2 1.7X 0.4ms vs 0.3ms (1, 3, 64, 64) -> (224, 224) nearest-exact float32 num_threads=2 1.7X 0.6ms vs 0.4ms (1, 3, 64, 64) -> (224, 224) nearest-exact uint8 num_threads=2 1.7X 0.5ms vs 0.3ms (1, 1, 64, 64) -> (224, 224) linear float32 num_threads=2 9X 0.800ms vs 0.088ms (1, 1, 64, 64) -> (224, 224) nearest float32 num_threads=2 11X 0.459ms vs 0.043ms (1, 1, 64, 64) -> (224, 224) nearest uint8 num_threads=2 7X 0.424ms vs 0.064ms (1, 1, 64, 64) -> (224, 224) nearest-exact float32 num_threads=2 12X 0.503ms vs 0.043ms (1, 1, 64, 64) -> (224, 224) nearest-exact uint8 num_threads=2 8X 0.461ms vs 0.059ms (1, 3, 64, 64) -> (224, 224) linear float32 num_threads=12 3X 1.1ms vs 0.3ms (1, 3, 64, 64) -> (224, 224) nearest float32 num_threads=12 1.6X 0.3ms vs 0.2ms (1, 3, 64, 64) -> (224, 224) nearest uint8 num_threads=12 1.5X 0.2ms vs 0.1ms (1, 3, 64, 64) -> (224, 224) nearest-exact float32 num_threads=12 1.5X 0.3ms vs 0.2ms (1, 3, 64, 64) -> (224, 224) nearest-exact uint8 num_threads=12 1.5X 0.2ms vs 0.1ms (1, 1, 64, 64) -> (224, 224) linear float32 num_threads=12 5X 0.8ms vs 0.2ms (1, 1, 64, 64) -> (224, 224) nearest float32 num_threads=12 10X 0.445ms vs 0.047ms (1, 1, 64, 64) -> (224, 224) nearest uint8 num_threads=12 7X 0.432ms vs 0.062ms (1, 1, 64, 64) -> (224, 224) nearest-exact float32 num_threads=12 10X 0.478ms vs 0.046ms (1, 1, 64, 64) -> (224, 224) nearest-exact uint8 num_threads=12 7X 0.470ms vs 0.063ms (1, 3, 64, 64) -> (224, 224) linear float32 num_threads=32 3X 1.1ms vs 0.4ms (1, 3, 64, 64) -> (224, 224) nearest float32 num_threads=32 1.8X 0.3ms vs 0.2ms (1, 3, 64, 64) -> (224, 224) nearest uint8 num_threads=32 1.5X 0.2ms vs 0.1ms (1, 3, 64, 64) -> (224, 224) nearest-exact float32 num_threads=32 1.4X 0.3ms vs 0.2ms (1, 3, 64, 64) -> (224, 224) nearest-exact uint8 num_threads=32 1.5X 0.2ms vs 0.1ms (1, 1, 64, 64) -> (224, 224) linear float32 num_threads=32 11X 0.815ms vs 0.074ms (1, 1, 64, 64) -> (224, 224) nearest float32 num_threads=32 10X 0.443ms vs 0.045ms (1, 1, 64, 64) -> (224, 224) nearest uint8 num_threads=32 7X 0.436ms vs 0.061ms (1, 1, 64, 64) -> (224, 224) nearest-exact float32 num_threads=32 10X 0.478ms vs 0.046ms (1, 1, 64, 64) -> (224, 224) nearest-exact uint8 num_threads=32 8X 0.470ms vs 0.061ms ---------------------------------------------------------------------------------------------------- (1, 3, 128, 128) -> (224, 224) linear float32 num_threads=1 0.9X 0.9ms vs 1.1ms (1, 3, 128, 128) -> (224, 224) nearest float32 num_threads=1 1.5X 0.9ms vs 0.6ms (1, 3, 128, 128) -> (224, 224) nearest uint8 num_threads=1 1.7X 0.9ms vs 0.5ms (1, 3, 128, 128) -> (224, 224) nearest-exact float32 num_threads=1 1.6X 1.0ms vs 0.6ms (1, 3, 128, 128) -> (224, 224) nearest-exact uint8 num_threads=1 1.8X 0.9ms vs 0.5ms (1, 1, 128, 128) -> (224, 224) linear float32 num_threads=1 8X 0.808ms vs 0.099ms (1, 1, 128, 128) -> (224, 224) nearest float32 num_threads=1 15X 0.848ms vs 0.058ms (1, 1, 128, 128) -> (224, 224) nearest uint8 num_threads=1 9X 0.820ms vs 0.087ms (1, 1, 128, 128) -> (224, 224) nearest-exact float32 num_threads=1 16X 0.909ms vs 0.059ms (1, 1, 128, 128) -> (224, 224) nearest-exact uint8 num_threads=1 10X 0.898ms vs 0.088ms (1, 3, 128, 128) -> (224, 224) linear float32 num_threads=2 1.4X 0.9ms vs 0.7ms (1, 3, 128, 128) -> (224, 224) nearest float32 num_threads=2 1.5X 0.5ms vs 0.3ms (1, 3, 128, 128) -> (224, 224) nearest uint8 num_threads=2 1.7X 0.4ms vs 0.3ms (1, 3, 128, 128) -> (224, 224) nearest-exact float32 num_threads=2 1.5X 0.5ms vs 0.4ms (1, 3, 128, 128) -> (224, 224) nearest-exact uint8 num_threads=2 1.8X 0.5ms vs 0.3ms (1, 1, 128, 128) -> (224, 224) linear float32 num_threads=2 9X 0.799ms vs 0.090ms (1, 1, 128, 128) -> (224, 224) nearest float32 num_threads=2 10X 0.459ms vs 0.045ms (1, 1, 128, 128) -> (224, 224) nearest uint8 num_threads=2 7X 0.427ms vs 0.059ms (1, 1, 128, 128) -> (224, 224) nearest-exact float32 num_threads=2 11X 0.501ms vs 0.044ms (1, 1, 128, 128) -> (224, 224) nearest-exact uint8 num_threads=2 8X 0.460ms vs 0.060ms (1, 3, 128, 128) -> (224, 224) linear float32 num_threads=12 2.9X 1.0ms vs 0.3ms (1, 3, 128, 128) -> (224, 224) nearest float32 num_threads=12 1.2X 0.2ms vs 0.2ms (1, 3, 128, 128) -> (224, 224) nearest uint8 num_threads=12 1.5X 0.2ms vs 0.1ms (1, 3, 128, 128) -> (224, 224) nearest-exact float32 num_threads=12 1.1X 0.2ms vs 0.2ms (1, 3, 128, 128) -> (224, 224) nearest-exact uint8 num_threads=12 1.6X 0.2ms vs 0.1ms (1, 1, 128, 128) -> (224, 224) linear float32 num_threads=12 12X 0.809ms vs 0.068ms (1, 1, 128, 128) -> (224, 224) nearest float32 num_threads=12 11X 0.438ms vs 0.041ms (1, 1, 128, 128) -> (224, 224) nearest uint8 num_threads=12 8X 0.432ms vs 0.055ms (1, 1, 128, 128) -> (224, 224) nearest-exact float32 num_threads=12 12X 0.480ms vs 0.041ms (1, 1, 128, 128) -> (224, 224) nearest-exact uint8 num_threads=12 8X 0.464ms vs 0.056ms (1, 3, 128, 128) -> (224, 224) linear float32 num_threads=32 3X 1.1ms vs 0.3ms (1, 3, 128, 128) -> (224, 224) nearest float32 num_threads=32 1.3X 0.3ms vs 0.2ms (1, 3, 128, 128) -> (224, 224) nearest uint8 num_threads=32 1.5X 0.2ms vs 0.1ms (1, 3, 128, 128) -> (224, 224) nearest-exact float32 num_threads=32 1.4X 0.3ms vs 0.2ms (1, 3, 128, 128) -> (224, 224) nearest-exact uint8 num_threads=32 1.6X 0.2ms vs 0.1ms (1, 1, 128, 128) -> (224, 224) linear float32 num_threads=32 11X 0.813ms vs 0.075ms (1, 1, 128, 128) -> (224, 224) nearest float32 num_threads=32 10X 0.443ms vs 0.046ms (1, 1, 128, 128) -> (224, 224) nearest uint8 num_threads=32 7X 0.433ms vs 0.061ms (1, 1, 128, 128) -> (224, 224) nearest-exact float32 num_threads=32 10X 0.478ms vs 0.046ms (1, 1, 128, 128) -> (224, 224) nearest-exact uint8 num_threads=32 8X 0.470ms vs 0.062ms ---------------------------------------------------------------------------------------------------- (1, 3, 224, 224) -> (600, 400) linear float32 num_threads=1 0.9X 4.5ms vs 5.2ms (1, 3, 224, 224) -> (600, 400) nearest float32 num_threads=1 1.5X 4.2ms vs 2.8ms (1, 3, 224, 224) -> (600, 400) nearest uint8 num_threads=1 1.8X 4.1ms vs 2.3ms (1, 3, 224, 224) -> (600, 400) nearest-exact float32 num_threads=1 1.6X 4.5ms vs 2.8ms (1, 3, 224, 224) -> (600, 400) nearest-exact uint8 num_threads=1 1.9X 4.4ms vs 2.3ms (1, 1, 224, 224) -> (600, 400) linear float32 num_threads=1 9X 3.8ms vs 0.4ms (1, 1, 224, 224) -> (600, 400) nearest float32 num_threads=1 17X 4.0ms vs 0.2ms (1, 1, 224, 224) -> (600, 400) nearest uint8 num_threads=1 11X 3.9ms vs 0.4ms (1, 1, 224, 224) -> (600, 400) nearest-exact float32 num_threads=1 19X 4.4ms vs 0.2ms (1, 1, 224, 224) -> (600, 400) nearest-exact uint8 num_threads=1 12X 4.3ms vs 0.4ms (1, 3, 224, 224) -> (600, 400) linear float32 num_threads=2 1.5X 4.5ms vs 3.1ms (1, 3, 224, 224) -> (600, 400) nearest float32 num_threads=2 1.4X 2.3ms vs 1.6ms (1, 3, 224, 224) -> (600, 400) nearest uint8 num_threads=2 1.7X 2.1ms vs 1.2ms (1, 3, 224, 224) -> (600, 400) nearest-exact float32 num_threads=2 1.6X 2.5ms vs 1.6ms (1, 3, 224, 224) -> (600, 400) nearest-exact uint8 num_threads=2 1.8X 2.2ms vs 1.2ms (1, 1, 224, 224) -> (600, 400) linear float32 num_threads=2 15X 3.8ms vs 0.3ms (1, 1, 224, 224) -> (600, 400) nearest float32 num_threads=2 15X 2.2ms vs 0.1ms (1, 1, 224, 224) -> (600, 400) nearest uint8 num_threads=2 7X 2.0ms vs 0.3ms (1, 1, 224, 224) -> (600, 400) nearest-exact float32 num_threads=2 16X 2.4ms vs 0.1ms (1, 1, 224, 224) -> (600, 400) nearest-exact uint8 num_threads=2 8X 2.2ms vs 0.3ms (1, 3, 224, 224) -> (600, 400) linear float32 num_threads=12 8X 5.2ms vs 0.7ms (1, 3, 224, 224) -> (600, 400) nearest float32 num_threads=12 1.3X 0.6ms vs 0.4ms (1, 3, 224, 224) -> (600, 400) nearest uint8 num_threads=12 1.7X 0.4ms vs 0.2ms (1, 3, 224, 224) -> (600, 400) nearest-exact float32 num_threads=12 1.4X 0.6ms vs 0.4ms (1, 3, 224, 224) -> (600, 400) nearest-exact uint8 num_threads=12 1.8X 0.4ms vs 0.2ms (1, 1, 224, 224) -> (600, 400) linear float32 num_threads=12 36X 3.9ms vs 0.1ms (1, 1, 224, 224) -> (600, 400) nearest float32 num_threads=12 10X 0.526ms vs 0.051ms (1, 1, 224, 224) -> (600, 400) nearest uint8 num_threads=12 7X 0.514ms vs 0.069ms (1, 1, 224, 224) -> (600, 400) nearest-exact float32 num_threads=12 11X 0.569ms vs 0.052ms (1, 1, 224, 224) -> (600, 400) nearest-exact uint8 num_threads=12 8X 0.557ms vs 0.070ms (1, 3, 224, 224) -> (600, 400) linear float32 num_threads=32 9X 4.5ms vs 0.5ms (1, 3, 224, 224) -> (600, 400) nearest float32 num_threads=32 0.5X 0.2ms vs 0.5ms (1, 3, 224, 224) -> (600, 400) nearest uint8 num_threads=32 1.5X 0.2ms vs 0.1ms (1, 3, 224, 224) -> (600, 400) nearest-exact float32 num_threads=32 1.0X 0.5ms vs 0.5ms (1, 3, 224, 224) -> (600, 400) nearest-exact uint8 num_threads=32 1.6X 0.2ms vs 0.1ms (1, 1, 224, 224) -> (600, 400) linear float32 num_threads=32 44X 3.864ms vs 0.087ms (1, 1, 224, 224) -> (600, 400) nearest float32 num_threads=32 10X 0.527ms vs 0.053ms (1, 1, 224, 224) -> (600, 400) nearest uint8 num_threads=32 7X 0.516ms vs 0.070ms (1, 1, 224, 224) -> (600, 400) nearest-exact float32 num_threads=32 10X 0.567ms vs 0.055ms (1, 1, 224, 224) -> (600, 400) nearest-exact uint8 num_threads=32 8X 0.558ms vs 0.072ms ---------------------------------------------------------------------------------------------------- (1, 3, 256, 256) -> (320, 320) linear float32 num_threads=1 1.0X 1.9ms vs 1.9ms (1, 3, 256, 256) -> (320, 320) nearest float32 num_threads=1 2.0X 1.8ms vs 0.9ms (1, 3, 256, 256) -> (320, 320) nearest uint8 num_threads=1 1.7X 1.8ms vs 1.0ms (1, 3, 256, 256) -> (320, 320) nearest-exact float32 num_threads=1 2.1X 1.9ms vs 0.9ms (1, 3, 256, 256) -> (320, 320) nearest-exact uint8 num_threads=1 1.9X 1.9ms vs 1.0ms (1, 1, 256, 256) -> (320, 320) linear float32 num_threads=1 9X 1.6ms vs 0.2ms (1, 1, 256, 256) -> (320, 320) nearest float32 num_threads=1 16X 1.7ms vs 0.1ms (1, 1, 256, 256) -> (320, 320) nearest uint8 num_threads=1 10X 1.7ms vs 0.2ms (1, 1, 256, 256) -> (320, 320) nearest-exact float32 num_threads=1 17X 1.9ms vs 0.1ms (1, 1, 256, 256) -> (320, 320) nearest-exact uint8 num_threads=1 11X 1.8ms vs 0.2ms (1, 3, 256, 256) -> (320, 320) linear float32 num_threads=2 1.7X 1.9ms vs 1.1ms (1, 3, 256, 256) -> (320, 320) nearest float32 num_threads=2 2.0X 1.0ms vs 0.5ms (1, 3, 256, 256) -> (320, 320) nearest uint8 num_threads=2 1.7X 0.9ms vs 0.5ms (1, 3, 256, 256) -> (320, 320) nearest-exact float32 num_threads=2 2.3X 1.1ms vs 0.5ms (1, 3, 256, 256) -> (320, 320) nearest-exact uint8 num_threads=2 1.8X 1.0ms vs 0.5ms (1, 1, 256, 256) -> (320, 320) linear float32 num_threads=2 8X 1.6ms vs 0.2ms (1, 1, 256, 256) -> (320, 320) nearest float32 num_threads=2 14X 0.931ms vs 0.067ms (1, 1, 256, 256) -> (320, 320) nearest uint8 num_threads=2 7X 0.9ms vs 0.1ms (1, 1, 256, 256) -> (320, 320) nearest-exact float32 num_threads=2 15X 1.016ms vs 0.069ms (1, 1, 256, 256) -> (320, 320) nearest-exact uint8 num_threads=2 9X 0.9ms vs 0.1ms (1, 3, 256, 256) -> (320, 320) linear float32 num_threads=12 8X 1.9ms vs 0.3ms (1, 3, 256, 256) -> (320, 320) nearest float32 num_threads=12 1.7X 0.2ms vs 0.1ms (1, 3, 256, 256) -> (320, 320) nearest uint8 num_threads=12 1.5X 0.2ms vs 0.1ms (1, 3, 256, 256) -> (320, 320) nearest-exact float32 num_threads=12 1.9X 0.2ms vs 0.1ms (1, 3, 256, 256) -> (320, 320) nearest-exact uint8 num_threads=12 1.6X 0.2ms vs 0.1ms (1, 1, 256, 256) -> (320, 320) linear float32 num_threads=12 20X 1.630ms vs 0.081ms (1, 1, 256, 256) -> (320, 320) nearest float32 num_threads=12 10X 0.457ms vs 0.044ms (1, 1, 256, 256) -> (320, 320) nearest uint8 num_threads=12 7X 0.439ms vs 0.060ms (1, 1, 256, 256) -> (320, 320) nearest-exact float32 num_threads=12 11X 0.485ms vs 0.045ms (1, 1, 256, 256) -> (320, 320) nearest-exact uint8 num_threads=12 8X 0.474ms vs 0.061ms (1, 3, 256, 256) -> (320, 320) linear float32 num_threads=32 8X 1.9ms vs 0.3ms (1, 3, 256, 256) -> (320, 320) nearest float32 num_threads=32 2.0X 0.2ms vs 0.1ms (1, 3, 256, 256) -> (320, 320) nearest uint8 num_threads=32 1.6X 0.2ms vs 0.1ms (1, 3, 256, 256) -> (320, 320) nearest-exact float32 num_threads=32 1.4X 0.2ms vs 0.2ms (1, 3, 256, 256) -> (320, 320) nearest-exact uint8 num_threads=32 1.4X 0.2ms vs 0.1ms (1, 1, 256, 256) -> (320, 320) linear float32 num_threads=32 21X 1.628ms vs 0.078ms (1, 1, 256, 256) -> (320, 320) nearest float32 num_threads=32 9X 0.453ms vs 0.048ms (1, 1, 256, 256) -> (320, 320) nearest uint8 num_threads=32 7X 0.445ms vs 0.063ms (1, 1, 256, 256) -> (320, 320) nearest-exact float32 num_threads=32 11X 0.535ms vs 0.048ms (1, 1, 256, 256) -> (320, 320) nearest-exact uint8 num_threads=32 8X 0.502ms vs 0.063ms ---------------------------------------------------------------------------------------------------- (1, 3, 500, 500) -> (800, 800) linear float32 num_threads=1 1.0X 13.8ms vs 14.0ms (1, 3, 500, 500) -> (800, 800) nearest float32 num_threads=1 1.8X 13.1ms vs 7.4ms (1, 3, 500, 500) -> (800, 800) nearest uint8 num_threads=1 1.8X 11.1ms vs 6.1ms (1, 3, 500, 500) -> (800, 800) nearest-exact float32 num_threads=1 1.9X 13.9ms vs 7.4ms (1, 3, 500, 500) -> (800, 800) nearest-exact uint8 num_threads=1 1.9X 11.8ms vs 6.1ms (1, 1, 500, 500) -> (800, 800) linear float32 num_threads=1 10X 10.2ms vs 1.1ms (1, 1, 500, 500) -> (800, 800) nearest float32 num_threads=1 19X 10.8ms vs 0.6ms (1, 1, 500, 500) -> (800, 800) nearest uint8 num_threads=1 11X 10.4ms vs 0.9ms (1, 1, 500, 500) -> (800, 800) nearest-exact float32 num_threads=1 20X 11.6ms vs 0.6ms (1, 1, 500, 500) -> (800, 800) nearest-exact uint8 num_threads=1 12X 11.4ms vs 0.9ms (1, 3, 500, 500) -> (800, 800) linear float32 num_threads=2 1.8X 13.7ms vs 7.7ms (1, 3, 500, 500) -> (800, 800) nearest float32 num_threads=2 2.6X 7.3ms vs 2.8ms (1, 3, 500, 500) -> (800, 800) nearest uint8 num_threads=2 1.8X 5.6ms vs 3.1ms (1, 3, 500, 500) -> (800, 800) nearest-exact float32 num_threads=2 1.9X 7.9ms vs 4.1ms (1, 3, 500, 500) -> (800, 800) nearest-exact uint8 num_threads=2 1.9X 6.0ms vs 3.1ms (1, 1, 500, 500) -> (800, 800) linear float32 num_threads=2 18X 10.1ms vs 0.6ms (1, 1, 500, 500) -> (800, 800) nearest float32 num_threads=2 19X 5.8ms vs 0.3ms (1, 1, 500, 500) -> (800, 800) nearest uint8 num_threads=2 10X 5.3ms vs 0.5ms (1, 1, 500, 500) -> (800, 800) nearest-exact float32 num_threads=2 20X 6.3ms vs 0.3ms (1, 1, 500, 500) -> (800, 800) nearest-exact uint8 num_threads=2 11X 5.7ms vs 0.5ms (1, 3, 500, 500) -> (800, 800) linear float32 num_threads=12 8X 13.8ms vs 1.6ms (1, 3, 500, 500) -> (800, 800) nearest float32 num_threads=12 2.9X 1.5ms vs 0.5ms (1, 3, 500, 500) -> (800, 800) nearest uint8 num_threads=12 1.7X 1.0ms vs 0.5ms (1, 3, 500, 500) -> (800, 800) nearest-exact float32 num_threads=12 1.5X 1.5ms vs 1.0ms (1, 3, 500, 500) -> (800, 800) nearest-exact uint8 num_threads=12 1.8X 1.0ms vs 0.6ms (1, 1, 500, 500) -> (800, 800) linear float32 num_threads=12 80X 10.1ms vs 0.1ms (1, 1, 500, 500) -> (800, 800) nearest float32 num_threads=12 13X 0.928ms vs 0.072ms (1, 1, 500, 500) -> (800, 800) nearest uint8 num_threads=12 8X 0.9ms vs 0.1ms (1, 1, 500, 500) -> (800, 800) nearest-exact float32 num_threads=12 13X 1.001ms vs 0.074ms (1, 1, 500, 500) -> (800, 800) nearest-exact uint8 num_threads=12 9X 1.0ms vs 0.1ms (1, 3, 500, 500) -> (800, 800) linear float32 num_threads=32 18X 14.0ms vs 0.8ms (1, 3, 500, 500) -> (800, 800) nearest float32 num_threads=32 1.9X 1.0ms vs 0.6ms (1, 3, 500, 500) -> (800, 800) nearest uint8 num_threads=32 2.9X 0.7ms vs 0.2ms (1, 3, 500, 500) -> (800, 800) nearest-exact float32 num_threads=32 1.7X 0.9ms vs 0.6ms (1, 3, 500, 500) -> (800, 800) nearest-exact uint8 num_threads=32 1.8X 0.4ms vs 0.2ms (1, 1, 500, 500) -> (800, 800) linear float32 num_threads=32 111X 10.254ms vs 0.092ms (1, 1, 500, 500) -> (800, 800) nearest float32 num_threads=32 14X 0.784ms vs 0.056ms (1, 1, 500, 500) -> (800, 800) nearest uint8 num_threads=32 7X 0.551ms vs 0.075ms (1, 1, 500, 500) -> (800, 800) nearest-exact float32 num_threads=32 11X 0.607ms vs 0.057ms (1, 1, 500, 500) -> (800, 800) nearest-exact uint8 num_threads=32 8X 0.596ms vs 0.076ms ---------------------------------------------------------------------------------------------------- (1, 3, 224, 224) -> (64, 64) linear float32 num_threads=1 1.0X 0.084ms vs 0.084ms (1, 3, 224, 224) -> (64, 64) nearest float32 num_threads=1 1.0X 0.077ms vs 0.078ms (1, 3, 224, 224) -> (64, 64) nearest uint8 num_threads=1 1.0X 0.076ms vs 0.076ms (1, 3, 224, 224) -> (64, 64) nearest-exact float32 num_threads=1 1.0X 0.083ms vs 0.083ms (1, 3, 224, 224) -> (64, 64) nearest-exact uint8 num_threads=1 1.0X 0.081ms vs 0.082ms (1, 1, 224, 224) -> (64, 64) linear float32 num_threads=1 1.0X 0.071ms vs 0.071ms (1, 1, 224, 224) -> (64, 64) nearest float32 num_threads=1 1.0X 0.074ms vs 0.074ms (1, 1, 224, 224) -> (64, 64) nearest uint8 num_threads=1 1.0X 0.072ms vs 0.072ms (1, 1, 224, 224) -> (64, 64) nearest-exact float32 num_threads=1 1.0X 0.080ms vs 0.080ms (1, 1, 224, 224) -> (64, 64) nearest-exact uint8 num_threads=1 0.9X 0.078ms vs 0.084ms (1, 3, 224, 224) -> (64, 64) linear float32 num_threads=2 1.0X 0.083ms vs 0.084ms (1, 3, 224, 224) -> (64, 64) nearest float32 num_threads=2 1.0X 0.076ms vs 0.077ms (1, 3, 224, 224) -> (64, 64) nearest uint8 num_threads=2 1.0X 0.075ms vs 0.074ms (1, 3, 224, 224) -> (64, 64) nearest-exact float32 num_threads=2 1.0X 0.082ms vs 0.083ms (1, 3, 224, 224) -> (64, 64) nearest-exact uint8 num_threads=2 1.0X 0.080ms vs 0.083ms (1, 1, 224, 224) -> (64, 64) linear float32 num_threads=2 1.0X 0.070ms vs 0.071ms (1, 1, 224, 224) -> (64, 64) nearest float32 num_threads=2 1.0X 0.073ms vs 0.075ms (1, 1, 224, 224) -> (64, 64) nearest uint8 num_threads=2 1.0X 0.071ms vs 0.072ms (1, 1, 224, 224) -> (64, 64) nearest-exact float32 num_threads=2 1.0X 0.079ms vs 0.080ms (1, 1, 224, 224) -> (64, 64) nearest-exact uint8 num_threads=2 1.0X 0.077ms vs 0.079ms (1, 3, 224, 224) -> (64, 64) linear float32 num_threads=12 1.0X 0.083ms vs 0.084ms (1, 3, 224, 224) -> (64, 64) nearest float32 num_threads=12 1.0X 0.080ms vs 0.078ms (1, 3, 224, 224) -> (64, 64) nearest uint8 num_threads=12 1.0X 0.077ms vs 0.075ms (1, 3, 224, 224) -> (64, 64) nearest-exact float32 num_threads=12 1.0X 0.083ms vs 0.083ms (1, 3, 224, 224) -> (64, 64) nearest-exact uint8 num_threads=12 1.0X 0.083ms vs 0.082ms (1, 1, 224, 224) -> (64, 64) linear float32 num_threads=12 1.0X 0.071ms vs 0.071ms (1, 1, 224, 224) -> (64, 64) nearest float32 num_threads=12 1.0X 0.076ms vs 0.074ms (1, 1, 224, 224) -> (64, 64) nearest uint8 num_threads=12 1.0X 0.073ms vs 0.071ms (1, 1, 224, 224) -> (64, 64) nearest-exact float32 num_threads=12 1.0X 0.080ms vs 0.080ms (1, 1, 224, 224) -> (64, 64) nearest-exact uint8 num_threads=12 1.0X 0.080ms vs 0.078ms (1, 3, 224, 224) -> (64, 64) linear float32 num_threads=32 1.0X 0.084ms vs 0.084ms (1, 3, 224, 224) -> (64, 64) nearest float32 num_threads=32 1.0X 0.078ms vs 0.077ms (1, 3, 224, 224) -> (64, 64) nearest uint8 num_threads=32 1.0X 0.076ms vs 0.076ms (1, 3, 224, 224) -> (64, 64) nearest-exact float32 num_threads=32 1.0X 0.083ms vs 0.083ms (1, 3, 224, 224) -> (64, 64) nearest-exact uint8 num_threads=32 1.0X 0.081ms vs 0.082ms (1, 1, 224, 224) -> (64, 64) linear float32 num_threads=32 1.0X 0.072ms vs 0.072ms (1, 1, 224, 224) -> (64, 64) nearest float32 num_threads=32 1.0X 0.074ms vs 0.075ms (1, 1, 224, 224) -> (64, 64) nearest uint8 num_threads=32 1.0X 0.072ms vs 0.072ms (1, 1, 224, 224) -> (64, 64) nearest-exact float32 num_threads=32 1.0X 0.077ms vs 0.080ms (1, 1, 224, 224) -> (64, 64) nearest-exact uint8 num_threads=32 1.0X 0.076ms vs 0.079ms ---------------------------------------------------------------------------------------------------- (1, 3, 224, 224) -> (128, 128) linear float32 num_threads=1 1.0X 0.3ms vs 0.3ms (1, 3, 224, 224) -> (128, 128) nearest float32 num_threads=1 1.8X 0.3ms vs 0.2ms (1, 3, 224, 224) -> (128, 128) nearest uint8 num_threads=1 1.6X 0.3ms vs 0.2ms (1, 3, 224, 224) -> (128, 128) nearest-exact float32 num_threads=1 2.0X 0.3ms vs 0.2ms (1, 3, 224, 224) -> (128, 128) nearest-exact uint8 num_threads=1 1.7X 0.3ms vs 0.2ms (1, 1, 224, 224) -> (128, 128) linear float32 num_threads=1 6X 0.265ms vs 0.044ms (1, 1, 224, 224) -> (128, 128) nearest float32 num_threads=1 10X 0.280ms vs 0.028ms (1, 1, 224, 224) -> (128, 128) nearest uint8 num_threads=1 7X 0.273ms vs 0.037ms (1, 1, 224, 224) -> (128, 128) nearest-exact float32 num_threads=1 11X 0.303ms vs 0.028ms (1, 1, 224, 224) -> (128, 128) nearest-exact uint8 num_threads=1 8X 0.297ms vs 0.038ms (1, 3, 224, 224) -> (128, 128) linear float32 num_threads=2 1.5X 0.3ms vs 0.2ms (1, 3, 224, 224) -> (128, 128) nearest float32 num_threads=2 1.8X 0.163ms vs 0.093ms (1, 3, 224, 224) -> (128, 128) nearest uint8 num_threads=2 1.5X 0.2ms vs 0.1ms (1, 3, 224, 224) -> (128, 128) nearest-exact float32 num_threads=2 1.9X 0.180ms vs 0.096ms (1, 3, 224, 224) -> (128, 128) nearest-exact uint8 num_threads=2 1.6X 0.2ms vs 0.1ms (1, 1, 224, 224) -> (128, 128) linear float32 num_threads=2 6X 0.264ms vs 0.044ms (1, 1, 224, 224) -> (128, 128) nearest float32 num_threads=2 10X 0.278ms vs 0.028ms (1, 1, 224, 224) -> (128, 128) nearest uint8 num_threads=2 7X 0.270ms vs 0.037ms (1, 1, 224, 224) -> (128, 128) nearest-exact float32 num_threads=2 11X 0.298ms vs 0.028ms (1, 1, 224, 224) -> (128, 128) nearest-exact uint8 num_threads=2 8X 0.293ms vs 0.037ms (1, 3, 224, 224) -> (128, 128) linear float32 num_threads=12 1.5X 0.3ms vs 0.2ms (1, 3, 224, 224) -> (128, 128) nearest float32 num_threads=12 1.7X 0.158ms vs 0.095ms (1, 3, 224, 224) -> (128, 128) nearest uint8 num_threads=12 1.5X 0.2ms vs 0.1ms (1, 3, 224, 224) -> (128, 128) nearest-exact float32 num_threads=12 1.7X 0.170ms vs 0.100ms (1, 3, 224, 224) -> (128, 128) nearest-exact uint8 num_threads=12 1.6X 0.2ms vs 0.1ms (1, 1, 224, 224) -> (128, 128) linear float32 num_threads=12 6X 0.269ms vs 0.043ms (1, 1, 224, 224) -> (128, 128) nearest float32 num_threads=12 11X 0.291ms vs 0.027ms (1, 1, 224, 224) -> (128, 128) nearest uint8 num_threads=12 8X 0.281ms vs 0.037ms (1, 1, 224, 224) -> (128, 128) nearest-exact float32 num_threads=12 11X 0.305ms vs 0.028ms (1, 1, 224, 224) -> (128, 128) nearest-exact uint8 num_threads=12 8X 0.306ms vs 0.038ms (1, 3, 224, 224) -> (128, 128) linear float32 num_threads=32 1.5X 0.3ms vs 0.2ms (1, 3, 224, 224) -> (128, 128) nearest float32 num_threads=32 1.6X 0.160ms vs 0.098ms (1, 3, 224, 224) -> (128, 128) nearest uint8 num_threads=32 1.5X 0.2ms vs 0.1ms (1, 3, 224, 224) -> (128, 128) nearest-exact float32 num_threads=32 1.7X 0.171ms vs 0.099ms (1, 3, 224, 224) -> (128, 128) nearest-exact uint8 num_threads=32 1.6X 0.2ms vs 0.1ms (1, 1, 224, 224) -> (128, 128) linear float32 num_threads=32 6X 0.269ms vs 0.044ms (1, 1, 224, 224) -> (128, 128) nearest float32 num_threads=32 10X 0.282ms vs 0.028ms (1, 1, 224, 224) -> (128, 128) nearest uint8 num_threads=32 7X 0.276ms vs 0.037ms (1, 1, 224, 224) -> (128, 128) nearest-exact float32 num_threads=32 11X 0.305ms vs 0.028ms (1, 1, 224, 224) -> (128, 128) nearest-exact uint8 num_threads=32 8X 0.299ms vs 0.038ms ---------------------------------------------------------------------------------------------------- (1, 3, 320, 320) -> (256, 256) linear float32 num_threads=1 1.0X 1.2ms vs 1.3ms (1, 3, 320, 320) -> (256, 256) nearest float32 num_threads=1 2.0X 1.2ms vs 0.6ms (1, 3, 320, 320) -> (256, 256) nearest uint8 num_threads=1 1.7X 1.1ms vs 0.7ms (1, 3, 320, 320) -> (256, 256) nearest-exact float32 num_threads=1 2.1X 1.2ms vs 0.6ms (1, 3, 320, 320) -> (256, 256) nearest-exact uint8 num_threads=1 1.9X 1.2ms vs 0.7ms (1, 1, 320, 320) -> (256, 256) linear float32 num_threads=1 8X 1.1ms vs 0.1ms (1, 1, 320, 320) -> (256, 256) nearest float32 num_threads=1 15X 1.109ms vs 0.073ms (1, 1, 320, 320) -> (256, 256) nearest uint8 num_threads=1 10X 1.1ms vs 0.1ms (1, 1, 320, 320) -> (256, 256) nearest-exact float32 num_threads=1 16X 1.192ms vs 0.074ms (1, 1, 320, 320) -> (256, 256) nearest-exact uint8 num_threads=1 11X 1.2ms vs 0.1ms (1, 3, 320, 320) -> (256, 256) linear float32 num_threads=2 1.7X 1.2ms vs 0.7ms (1, 3, 320, 320) -> (256, 256) nearest float32 num_threads=2 2.0X 0.6ms vs 0.3ms (1, 3, 320, 320) -> (256, 256) nearest uint8 num_threads=2 1.7X 0.6ms vs 0.3ms (1, 3, 320, 320) -> (256, 256) nearest-exact float32 num_threads=2 2.2X 0.7ms vs 0.3ms (1, 3, 320, 320) -> (256, 256) nearest-exact uint8 num_threads=2 1.8X 0.6ms vs 0.3ms (1, 1, 320, 320) -> (256, 256) linear float32 num_threads=2 9X 1.0ms vs 0.1ms (1, 1, 320, 320) -> (256, 256) nearest float32 num_threads=2 11X 0.598ms vs 0.052ms (1, 1, 320, 320) -> (256, 256) nearest uint8 num_threads=2 8X 0.556ms vs 0.072ms (1, 1, 320, 320) -> (256, 256) nearest-exact float32 num_threads=2 12X 0.649ms vs 0.053ms (1, 1, 320, 320) -> (256, 256) nearest-exact uint8 num_threads=2 8X 0.598ms vs 0.073ms (1, 3, 320, 320) -> (256, 256) linear float32 num_threads=12 5X 1.2ms vs 0.3ms (1, 3, 320, 320) -> (256, 256) nearest float32 num_threads=12 1.5X 0.2ms vs 0.1ms (1, 3, 320, 320) -> (256, 256) nearest uint8 num_threads=12 1.3X 0.2ms vs 0.1ms (1, 3, 320, 320) -> (256, 256) nearest-exact float32 num_threads=12 1.6X 0.2ms vs 0.1ms (1, 3, 320, 320) -> (256, 256) nearest-exact uint8 num_threads=12 1.4X 0.2ms vs 0.1ms (1, 1, 320, 320) -> (256, 256) linear float32 num_threads=12 9X 1.0ms vs 0.1ms (1, 1, 320, 320) -> (256, 256) nearest float32 num_threads=12 12X 0.572ms vs 0.048ms (1, 1, 320, 320) -> (256, 256) nearest uint8 num_threads=12 8X 0.560ms vs 0.068ms (1, 1, 320, 320) -> (256, 256) nearest-exact float32 num_threads=12 13X 0.617ms vs 0.049ms (1, 1, 320, 320) -> (256, 256) nearest-exact uint8 num_threads=12 9X 0.604ms vs 0.068ms (1, 3, 320, 320) -> (256, 256) linear float32 num_threads=32 5X 1.2ms vs 0.3ms (1, 3, 320, 320) -> (256, 256) nearest float32 num_threads=32 1.5X 0.2ms vs 0.1ms (1, 3, 320, 320) -> (256, 256) nearest uint8 num_threads=32 1.4X 0.2ms vs 0.1ms (1, 3, 320, 320) -> (256, 256) nearest-exact float32 num_threads=32 1.6X 0.2ms vs 0.1ms (1, 3, 320, 320) -> (256, 256) nearest-exact uint8 num_threads=32 1.4X 0.2ms vs 0.1ms (1, 1, 320, 320) -> (256, 256) linear float32 num_threads=32 13X 1.042ms vs 0.081ms (1, 1, 320, 320) -> (256, 256) nearest float32 num_threads=32 12X 0.586ms vs 0.050ms (1, 1, 320, 320) -> (256, 256) nearest uint8 num_threads=32 8X 0.562ms vs 0.069ms (1, 1, 320, 320) -> (256, 256) nearest-exact float32 num_threads=32 12X 0.621ms vs 0.051ms (1, 1, 320, 320) -> (256, 256) nearest-exact uint8 num_threads=32 9X 0.609ms vs 0.070ms ---------------------------------------------------------------------------------------------------- (1, 3, 600, 400) -> (224, 224) linear float32 num_threads=1 1.0X 1.0ms vs 1.0ms (1, 3, 600, 400) -> (224, 224) nearest float32 num_threads=1 1.9X 0.9ms vs 0.5ms (1, 3, 600, 400) -> (224, 224) nearest uint8 num_threads=1 1.7X 0.9ms vs 0.5ms (1, 3, 600, 400) -> (224, 224) nearest-exact float32 num_threads=1 2.1X 1.0ms vs 0.5ms (1, 3, 600, 400) -> (224, 224) nearest-exact uint8 num_threads=1 1.8X 0.9ms vs 0.5ms (1, 1, 600, 400) -> (224, 224) linear float32 num_threads=1 7X 0.8ms vs 0.1ms (1, 1, 600, 400) -> (224, 224) nearest float32 num_threads=1 14X 0.852ms vs 0.061ms (1, 1, 600, 400) -> (224, 224) nearest uint8 num_threads=1 9X 0.828ms vs 0.087ms (1, 1, 600, 400) -> (224, 224) nearest-exact float32 num_threads=1 15X 0.922ms vs 0.061ms (1, 1, 600, 400) -> (224, 224) nearest-exact uint8 num_threads=1 10X 0.897ms vs 0.087ms (1, 3, 600, 400) -> (224, 224) linear float32 num_threads=2 1.6X 0.9ms vs 0.6ms (1, 3, 600, 400) -> (224, 224) nearest float32 num_threads=2 1.9X 0.5ms vs 0.2ms (1, 3, 600, 400) -> (224, 224) nearest uint8 num_threads=2 1.7X 0.4ms vs 0.3ms (1, 3, 600, 400) -> (224, 224) nearest-exact float32 num_threads=2 2.1X 0.5ms vs 0.3ms (1, 3, 600, 400) -> (224, 224) nearest-exact uint8 num_threads=2 1.8X 0.5ms vs 0.3ms (1, 1, 600, 400) -> (224, 224) linear float32 num_threads=2 10X 0.808ms vs 0.084ms (1, 1, 600, 400) -> (224, 224) nearest float32 num_threads=2 10X 0.462ms vs 0.046ms (1, 1, 600, 400) -> (224, 224) nearest uint8 num_threads=2 7X 0.429ms vs 0.062ms (1, 1, 600, 400) -> (224, 224) nearest-exact float32 num_threads=2 12X 0.504ms vs 0.044ms (1, 1, 600, 400) -> (224, 224) nearest-exact uint8 num_threads=2 7X 0.461ms vs 0.063ms (1, 3, 600, 400) -> (224, 224) linear float32 num_threads=12 4X 1.0ms vs 0.2ms (1, 3, 600, 400) -> (224, 224) nearest float32 num_threads=12 1.7X 0.2ms vs 0.1ms (1, 3, 600, 400) -> (224, 224) nearest uint8 num_threads=12 1.5X 0.2ms vs 0.1ms (1, 3, 600, 400) -> (224, 224) nearest-exact float32 num_threads=12 1.9X 0.2ms vs 0.1ms (1, 3, 600, 400) -> (224, 224) nearest-exact uint8 num_threads=12 1.6X 0.2ms vs 0.1ms (1, 1, 600, 400) -> (224, 224) linear float32 num_threads=12 12X 0.820ms vs 0.067ms (1, 1, 600, 400) -> (224, 224) nearest float32 num_threads=12 11X 0.438ms vs 0.041ms (1, 1, 600, 400) -> (224, 224) nearest uint8 num_threads=12 8X 0.431ms vs 0.056ms (1, 1, 600, 400) -> (224, 224) nearest-exact float32 num_threads=12 12X 0.482ms vs 0.041ms (1, 1, 600, 400) -> (224, 224) nearest-exact uint8 num_threads=12 8X 0.467ms vs 0.056ms (1, 3, 600, 400) -> (224, 224) linear float32 num_threads=32 4X 1.0ms vs 0.3ms (1, 3, 600, 400) -> (224, 224) nearest float32 num_threads=32 1.7X 0.2ms vs 0.1ms (1, 3, 600, 400) -> (224, 224) nearest uint8 num_threads=32 1.5X 0.2ms vs 0.1ms (1, 3, 600, 400) -> (224, 224) nearest-exact float32 num_threads=32 1.8X 0.2ms vs 0.1ms (1, 3, 600, 400) -> (224, 224) nearest-exact uint8 num_threads=32 1.6X 0.2ms vs 0.1ms (1, 1, 600, 400) -> (224, 224) linear float32 num_threads=32 12X 0.824ms vs 0.070ms (1, 1, 600, 400) -> (224, 224) nearest float32 num_threads=32 10X 0.443ms vs 0.044ms (1, 1, 600, 400) -> (224, 224) nearest uint8 num_threads=32 7X 0.438ms vs 0.059ms (1, 1, 600, 400) -> (224, 224) nearest-exact float32 num_threads=32 11X 0.479ms vs 0.045ms (1, 1, 600, 400) -> (224, 224) nearest-exact uint8 num_threads=32 8X 0.470ms vs 0.059ms ---------------------------------------------------------------------------------------------------- (1, 3, 800, 800) -> (500, 500) linear float32 num_threads=1 1.0X 4.7ms vs 4.7ms (1, 3, 800, 800) -> (500, 500) nearest float32 num_threads=1 2.0X 4.4ms vs 2.2ms (1, 3, 800, 800) -> (500, 500) nearest uint8 num_threads=1 1.8X 4.3ms vs 2.5ms (1, 3, 800, 800) -> (500, 500) nearest-exact float32 num_threads=1 2.1X 4.7ms vs 2.2ms (1, 3, 800, 800) -> (500, 500) nearest-exact uint8 num_threads=1 1.9X 4.6ms vs 2.5ms (1, 1, 800, 800) -> (500, 500) linear float32 num_threads=1 9X 4.0ms vs 0.4ms (1, 1, 800, 800) -> (500, 500) nearest float32 num_threads=1 17X 4.2ms vs 0.2ms (1, 1, 800, 800) -> (500, 500) nearest uint8 num_threads=1 11X 4.1ms vs 0.4ms (1, 1, 800, 800) -> (500, 500) nearest-exact float32 num_threads=1 19X 4.6ms vs 0.2ms (1, 1, 800, 800) -> (500, 500) nearest-exact uint8 num_threads=1 12X 4.5ms vs 0.4ms (1, 3, 800, 800) -> (500, 500) linear float32 num_threads=2 1.7X 4.7ms vs 2.7ms (1, 3, 800, 800) -> (500, 500) nearest float32 num_threads=2 2.1X 2.4ms vs 1.1ms (1, 3, 800, 800) -> (500, 500) nearest uint8 num_threads=2 1.8X 2.2ms vs 1.3ms (1, 3, 800, 800) -> (500, 500) nearest-exact float32 num_threads=2 2.3X 2.6ms vs 1.1ms (1, 3, 800, 800) -> (500, 500) nearest-exact uint8 num_threads=2 1.9X 2.3ms vs 1.3ms (1, 1, 800, 800) -> (500, 500) linear float32 num_threads=2 15X 4.0ms vs 0.3ms (1, 1, 800, 800) -> (500, 500) nearest float32 num_threads=2 16X 2.3ms vs 0.1ms (1, 1, 800, 800) -> (500, 500) nearest uint8 num_threads=2 9X 2.1ms vs 0.2ms (1, 1, 800, 800) -> (500, 500) nearest-exact float32 num_threads=2 17X 2.5ms vs 0.1ms (1, 1, 800, 800) -> (500, 500) nearest-exact uint8 num_threads=2 10X 2.3ms vs 0.2ms (1, 3, 800, 800) -> (500, 500) linear float32 num_threads=12 10X 4.7ms vs 0.5ms (1, 3, 800, 800) -> (500, 500) nearest float32 num_threads=12 1.9X 0.4ms vs 0.2ms (1, 3, 800, 800) -> (500, 500) nearest uint8 num_threads=12 1.7X 0.4ms vs 0.2ms (1, 3, 800, 800) -> (500, 500) nearest-exact float32 num_threads=12 1.9X 0.4ms vs 0.2ms (1, 3, 800, 800) -> (500, 500) nearest-exact uint8 num_threads=12 1.8X 0.4ms vs 0.2ms (1, 1, 800, 800) -> (500, 500) linear float32 num_threads=12 41X 3.969ms vs 0.096ms (1, 1, 800, 800) -> (500, 500) nearest float32 num_threads=12 11X 0.545ms vs 0.051ms (1, 1, 800, 800) -> (500, 500) nearest uint8 num_threads=12 8X 0.532ms vs 0.070ms (1, 1, 800, 800) -> (500, 500) nearest-exact float32 num_threads=12 11X 0.590ms vs 0.052ms (1, 1, 800, 800) -> (500, 500) nearest-exact uint8 num_threads=12 8X 0.578ms vs 0.071ms (1, 3, 800, 800) -> (500, 500) linear float32 num_threads=32 17X 4.7ms vs 0.3ms (1, 3, 800, 800) -> (500, 500) nearest float32 num_threads=32 1.8X 0.2ms vs 0.1ms (1, 3, 800, 800) -> (500, 500) nearest uint8 num_threads=32 2.0X 0.3ms vs 0.1ms (1, 3, 800, 800) -> (500, 500) nearest-exact float32 num_threads=32 1.9X 0.2ms vs 0.1ms (1, 3, 800, 800) -> (500, 500) nearest-exact uint8 num_threads=32 1.6X 0.2ms vs 0.1ms (1, 1, 800, 800) -> (500, 500) linear float32 num_threads=32 45X 4.028ms vs 0.090ms (1, 1, 800, 800) -> (500, 500) nearest float32 num_threads=32 10X 0.549ms vs 0.053ms (1, 1, 800, 800) -> (500, 500) nearest uint8 num_threads=32 7X 0.536ms vs 0.072ms (1, 1, 800, 800) -> (500, 500) nearest-exact float32 num_threads=32 11X 0.592ms vs 0.055ms (1, 1, 800, 800) -> (500, 500) nearest-exact uint8 num_threads=32 8X 0.581ms vs 0.074ms ``` </details> Code: <details> I used this file which is adapted from https://github.com/pytorch/pytorch/blob/master/benchmarks/operator_benchmark/pt/interpolate_test.py ```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', dtype=torch.float): input_image = torch.randint(0, 256, size=input_size, dtype=dtype, device='cpu', requires_grad=self.auto_set()) if channels_last: if input_image.ndim == 4: input_image = input_image.contiguous(memory_format=torch.channels_last) elif input_image.ndim == 5: input_image = input_image.contiguous(memory_format=torch.channels_last_3d) else: raise ValueError( f"Can not set channels_last to the input of {input_image.ndim} dims" ) align_corners = None if "nearest" in mode else False if mode == "linear": mode = { 3: 'linear', 4: 'bilinear', 5: 'trilinear', }[input_image.ndim] self.inputs = { "input_image": input_image, "output_size": output_size, "mode": mode, "align_corners": align_corners, } self.set_module_name("interpolate") def forward(self, input_image, output_size, mode, align_corners): return torch.nn.functional.interpolate(input_image, size=output_size, mode=mode, align_corners=align_corners) def make_config(): sizes = ( ((224, 224), (64, 64)), ((224, 224), (128, 128)), ((600, 400), (224, 224)), ((320, 320), (256, 256)), ((800, 800), (500, 500)), ) 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], 'mode': ["linear", "nearest", "nearest-exact"], 'dtype': [torch.float, torch.uint8] }, tags=["short"], ) # Need to remove instances with both torch.int and linear # Note: this is naaaasty def get_mode(l): for d in l: if "mode" in d: return d["mode"] def get_dtype(l): for d in l: if "dtype" in d: return d["dtype"] config = [l for l in config if not(get_mode(l) == "linear" and get_dtype(l) == torch.uint8)] return config config = make_config() op_bench.generate_pt_test(config, InterpolateBenchmark) if __name__ == "__main__": op_bench.benchmark_runner.main() ``` with ``` for num_threads in 1 2 12 32; do echo "num_threads=$num_threads" && python -m pt.my_interpolate_test --iterations 1000 --omp_num_threads $num_threads ; done > $out_file ``` and this very ugly helper ```py import re with open("main") as f: main = f.readlines() with open("new") as f: new = f.readlines() out = [] for main_line, new_line in zip(main, new): 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("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}" 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 % 10 == 0 and i % 40 != 0: print() if i % 40 == 0: print("-" * 100) print(l) ``` </details> Closes https://github.com/pytorch/pytorch/issues/83840 When this is merged we should be able to remove some hack in vision as well https://github.com/pytorch/vision/pull/6661 (CC @vfdev-5 @datumbox ) Pull Request resolved: https://github.com/pytorch/pytorch/pull/86361 Approved by: https://github.com/vfdev-5, https://github.com/datumbox, https://github.com/fmassa
PyTorch is a Python package that provides two high-level features:
You can reuse your favorite Python packages such as NumPy, SciPy, and Cython to extend PyTorch when needed.
Our trunk health (Continuous Integration signals) can be found at hud.pytorch.org.
At a granular level, PyTorch is a library that consists of the following components:
| Component | Description |
|---|---|
| torch | A Tensor library like NumPy, with strong GPU support |
| torch.autograd | A tape-based automatic differentiation library that supports all differentiable Tensor operations in torch |
| torch.jit | A compilation stack (TorchScript) to create serializable and optimizable models from PyTorch code |
| torch.nn | A neural networks library deeply integrated with autograd designed for maximum flexibility |
| torch.multiprocessing | Python multiprocessing, but with magical memory sharing of torch Tensors across processes. Useful for data loading and Hogwild training |
| torch.utils | DataLoader and other utility functions for convenience |
Usually, PyTorch is used either as:
Elaborating Further:
If you use NumPy, then you have used Tensors (a.k.a. ndarray).
PyTorch provides Tensors that can live either on the CPU or the GPU and accelerates the computation by a huge amount.
We provide a wide variety of tensor routines to accelerate and fit your scientific computation needs such as slicing, indexing, mathematical operations, linear algebra, reductions. And they are fast!
PyTorch has a unique way of building neural networks: using and replaying a tape recorder.
Most frameworks such as TensorFlow, Theano, Caffe, and CNTK have a static view of the world. One has to build a neural network and reuse the same structure again and again. Changing the way the network behaves means that one has to start from scratch.
With PyTorch, we use a technique called reverse-mode auto-differentiation, which allows you to change the way your network behaves arbitrarily with zero lag or overhead. Our inspiration comes from several research papers on this topic, as well as current and past work such as torch-autograd, autograd, Chainer, etc.
While this technique is not unique to PyTorch, it's one of the fastest implementations of it to date. You get the best of speed and flexibility for your crazy research.
PyTorch is not a Python binding into a monolithic C++ framework. It is built to be deeply integrated into Python. You can use it naturally like you would use NumPy / SciPy / scikit-learn etc. You can write your new neural network layers in Python itself, using your favorite libraries and use packages such as Cython and Numba. Our goal is to not reinvent the wheel where appropriate.
PyTorch is designed to be intuitive, linear in thought, and easy to use. When you execute a line of code, it gets executed. There isn't an asynchronous view of the world. When you drop into a debugger or receive error messages and stack traces, understanding them is straightforward. The stack trace points to exactly where your code was defined. We hope you never spend hours debugging your code because of bad stack traces or asynchronous and opaque execution engines.
PyTorch has minimal framework overhead. We integrate acceleration libraries such as Intel MKL and NVIDIA (cuDNN, NCCL) to maximize speed. At the core, its CPU and GPU Tensor and neural network backends are mature and have been tested for years.
Hence, PyTorch is quite fast – whether you run small or large neural networks.
The memory usage in PyTorch is extremely efficient compared to Torch or some of the alternatives. We've written custom memory allocators for the GPU to make sure that your deep learning models are maximally memory efficient. This enables you to train bigger deep learning models than before.
Writing new neural network modules, or interfacing with PyTorch's Tensor API was designed to be straightforward and with minimal abstractions.
You can write new neural network layers in Python using the torch API or your favorite NumPy-based libraries such as SciPy.
If you want to write your layers in C/C++, we provide a convenient extension API that is efficient and with minimal boilerplate. No wrapper code needs to be written. You can see a tutorial here and an example here.
Commands to install binaries via Conda or pip wheels are on our website: https://pytorch.org/get-started/locally/
Python wheels for NVIDIA's Jetson Nano, Jetson TX1/TX2, Jetson Xavier NX/AGX, and Jetson AGX Orin are provided here and the L4T container is published here
They require JetPack 4.2 and above, and @dusty-nv and @ptrblck are maintaining them.
If you are installing from source, you will need:
We highly recommend installing an Anaconda environment. You will get a high-quality BLAS library (MKL) and you get controlled dependency versions regardless of your Linux distro.
If you want to compile with CUDA support, install the following (note that CUDA is not supported on macOS)
Note: You could refer to the cuDNN Support Matrix for cuDNN versions with the various supported CUDA, CUDA driver and NVIDIA hardware
If you want to disable CUDA support, export the environment variable USE_CUDA=0. Other potentially useful environment variables may be found in setup.py.
If you are building for NVIDIA's Jetson platforms (Jetson Nano, TX1, TX2, AGX Xavier), Instructions to install PyTorch for Jetson Nano are available here
If you want to compile with ROCm support, install
If you want to disable ROCm support, export the environment variable USE_ROCM=0. Other potentially useful environment variables may be found in setup.py.
Common
conda install astunparse numpy ninja pyyaml setuptools cmake cffi typing_extensions future six requests dataclasses
On Linux
conda install mkl mkl-include # CUDA only: Add LAPACK support for the GPU if needed conda install -c pytorch magma-cuda110 # or the magma-cuda* that matches your CUDA version from https://anaconda.org/pytorch/repo
On MacOS
# Add this package on intel x86 processor machines only conda install mkl mkl-include # Add these packages if torch.distributed is needed conda install pkg-config libuv
On Windows
conda install mkl mkl-include # Add these packages if torch.distributed is needed. # Distributed package support on Windows is a prototype feature and is subject to changes. conda install -c conda-forge libuv=1.39
git clone --recursive https://github.com/pytorch/pytorch cd pytorch # if you are updating an existing checkout git submodule sync git submodule update --init --recursive --jobs 0
On Linux
If you're compiling for AMD ROCm then first run this command:
# Only run this if you're compiling for ROCm python tools/amd_build/build_amd.py
Install PyTorch
export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"} python setup.py install
Note that if you are using Anaconda, you may experience an error caused by the linker:
build/temp.linux-x86_64-3.7/torch/csrc/stub.o: file not recognized: file format not recognized collect2: error: ld returned 1 exit status error: command 'g++' failed with exit status 1
This is caused by ld from the Conda environment shadowing the system ld. You should use a newer version of Python that fixes this issue. The recommended Python version is 3.7.6+ and 3.8.1+.
On macOS
export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"} MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py install
On Windows
Choose Correct Visual Studio Version.
Sometimes there are regressions in new versions of Visual Studio, so it‘s best to use the same Visual Studio Version 16.8.5 as Pytorch CI’s.
PyTorch CI uses Visual C++ BuildTools, which come with Visual Studio Enterprise, Professional, or Community Editions. You can also install the build tools from https://visualstudio.microsoft.com/visual-cpp-build-tools/. The build tools do not come with Visual Studio Code by default.
If you want to build legacy python code, please refer to Building on legacy code and CUDA
CPU-only builds
In this mode PyTorch computations will run on your CPU, not your GPU
conda activate python setup.py install
Note on OpenMP: The desired OpenMP implementation is Intel OpenMP (iomp). In order to link against iomp, you'll need to manually download the library and set up the building environment by tweaking CMAKE_INCLUDE_PATH and LIB. The instruction here is an example for setting up both MKL and Intel OpenMP. Without these configurations for CMake, Microsoft Visual C OpenMP runtime (vcomp) will be used.
CUDA based build
In this mode PyTorch computations will leverage your GPU via CUDA for faster number crunching
NVTX is needed to build Pytorch with CUDA. NVTX is a part of CUDA distributive, where it is called “Nsight Compute”. To install it onto an already installed CUDA run CUDA installation once again and check the corresponding checkbox. Make sure that CUDA with Nsight Compute is installed after Visual Studio.
Currently, VS 2017 / 2019, and Ninja are supported as the generator of CMake. If ninja.exe is detected in PATH, then Ninja will be used as the default generator, otherwise, it will use VS 2017 / 2019.
If Ninja is selected as the generator, the latest MSVC will get selected as the underlying toolchain.
Additional libraries such as Magma, oneDNN, a.k.a MKLDNN or DNNL, and Sccache are often needed. Please refer to the installation-helper to install them.
You can refer to the build_pytorch.bat script for some other environment variables configurations
cmd :: Set the environment variables after you have downloaded and unzipped the mkl package, :: else CMake would throw an error as `Could NOT find OpenMP`. set CMAKE_INCLUDE_PATH={Your directory}\mkl\include set LIB={Your directory}\mkl\lib;%LIB% :: Read the content in the previous section carefully before you proceed. :: [Optional] If you want to override the underlying toolset used by Ninja and Visual Studio with CUDA, please run the following script block. :: "Visual Studio 2019 Developer Command Prompt" will be run automatically. :: Make sure you have CMake >= 3.12 before you do this when you use the Visual Studio generator. set CMAKE_GENERATOR_TOOLSET_VERSION=14.27 set DISTUTILS_USE_SDK=1 for /f "usebackq tokens=*" %i in (`"%ProgramFiles(x86)%\Microsoft Visual Studio\Installer\vswhere.exe" -version [15^,17^) -products * -latest -property installationPath`) do call "%i\VC\Auxiliary\Build\vcvarsall.bat" x64 -vcvars_ver=%CMAKE_GENERATOR_TOOLSET_VERSION% :: [Optional] If you want to override the CUDA host compiler set CUDAHOSTCXX=C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.27.29110\bin\HostX64\x64\cl.exe python setup.py install
You can adjust the configuration of cmake variables optionally (without building first), by doing the following. For example, adjusting the pre-detected directories for CuDNN or BLAS can be done with such a step.
On Linux
export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"} python setup.py build --cmake-only ccmake build # or cmake-gui build
On macOS
export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"} MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py build --cmake-only ccmake build # or cmake-gui build
You can also pull a pre-built docker image from Docker Hub and run with docker v19.03+
docker run --gpus all --rm -ti --ipc=host pytorch/pytorch:latest
Please note that PyTorch uses shared memory to share data between processes, so if torch multiprocessing is used (e.g. for multithreaded data loaders) the default shared memory segment size that container runs with is not enough, and you should increase shared memory size either with --ipc=host or --shm-size command line options to nvidia-docker run.
NOTE: Must be built with a docker version > 18.06
The Dockerfile is supplied to build images with CUDA 11.1 support and cuDNN v8. You can pass PYTHON_VERSION=x.y make variable to specify which Python version is to be used by Miniconda, or leave it unset to use the default.
make -f docker.Makefile # images are tagged as docker.io/${your_docker_username}/pytorch
To build documentation in various formats, you will need Sphinx and the readthedocs theme.
cd docs/ pip install -r requirements.txt
You can then build the documentation by running make <format> from the docs/ folder. Run make to get a list of all available output formats.
If you get a katex error run npm install katex. If it persists, try npm install -g katex
Note: if you installed
nodejswith a different package manager (e.g.,conda) thennpmwill probably install a version ofkatexthat is not compatible with your version ofnodejsand doc builds will fail. A combination of versions that is known to work isnode@6.13.1andkatex@0.13.18. To install the latter withnpmyou can runnpm install -g katex@0.13.18
Installation instructions and binaries for previous PyTorch versions may be found on our website.
Three-pointers to get you started:
PyTorch has a 90-day release cycle (major releases). Please let us know if you encounter a bug by filing an issue.
We appreciate all contributions. If you are planning to contribute back bug-fixes, please do so without any further discussion.
If you plan to contribute new features, utility functions, or extensions to the core, please first open an issue and discuss the feature with us. Sending a PR without discussion might end up resulting in a rejected PR because we might be taking the core in a different direction than you might be aware of.
To learn more about making a contribution to Pytorch, please see our Contribution page.
PyTorch is a community-driven project with several skillful engineers and researchers contributing to it.
PyTorch is currently maintained by Adam Paszke, Sam Gross, Soumith Chintala and Gregory Chanan with major contributions coming from hundreds of talented individuals in various forms and means. A non-exhaustive but growing list needs to mention: Trevor Killeen, Sasank Chilamkurthy, Sergey Zagoruyko, Adam Lerer, Francisco Massa, Alykhan Tejani, Luca Antiga, Alban Desmaison, Andreas Koepf, James Bradbury, Zeming Lin, Yuandong Tian, Guillaume Lample, Marat Dukhan, Natalia Gimelshein, Christian Sarofeen, Martin Raison, Edward Yang, Zachary Devito.
Note: This project is unrelated to hughperkins/pytorch with the same name. Hugh is a valuable contributor to the Torch community and has helped with many things Torch and PyTorch.
PyTorch has a BSD-style license, as found in the LICENSE file.