Update base for Update on "Improved perfs for vectorized bilinear interpolate cpu uint8 RGB-case (channels last)"

vfdev-5 · vfdev-5 · commit 368c89984b82 · 2023-03-23T21:14:55.000Z
## Description - Based on #96651 - Improved perfs for vectorized **bilinear** interpolate uint8 RGB-case, **channels last** - unified RGB and RGBA processing code such that RGB input is not copied into RGBA - Performances are more close to Pillow-SIMD (labeled as `Pillow (9.0.0.post1)` in the results) - RGBA case perfs are the same after refactoring (see Source link below) - Fixed mem pointer alignment, added more comments (reviews from #96651) ## Results - `Pillow (9.0.0.post1)` == Pillow-SIMD ``` [-------------------------------------------------------------------------------------------------- Resize -------------------------------------------------------------------------------------------------] | Pillow (9.0.0.post1) | torch (2.1.0a0+gitce4be01) PR | torch (2.1.0a0+git5309c44) nightly | Speed-up: PR vs nightly 1 threads: -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- 3 torch.uint8 channels_last bilinear (256, 256) -> (32, 32) aa=True | 38.548 (+-0.280) | 57.536 (+-0.210) | 132.147 (+-1.236) | 2.297 (+-0.000) 3 torch.uint8 channels_last bilinear (256, 256) -> (32, 32) aa=False | | 38.532 (+-0.219) | 111.789 (+-1.175) | 2.901 (+-0.000) 3 torch.uint8 channels_last bilinear (256, 256) -> (224, 224) aa=True | 127.689 (+-1.348) | 156.262 (+-1.213) | 302.518 (+-2.632) | 1.936 (+-0.000) 3 torch.uint8 channels_last bilinear (256, 256) -> (224, 224) aa=False | | 145.483 (+-1.077) | 286.663 (+-2.494) | 1.970 (+-0.000) 3 torch.uint8 channels_last bilinear (256, 256) -> (320, 320) aa=True | 178.117 (+-1.956) | 215.053 (+-1.470) | 439.375 (+-4.014) | 2.043 (+-0.000) 3 torch.uint8 channels_last bilinear (256, 256) -> (320, 320) aa=False | | 211.340 (+-2.239) | 438.537 (+-4.143) | 2.075 (+-0.000) 3 torch.uint8 channels_last bilinear (520, 520) -> (32, 32) aa=True | 112.593 (+-1.266) | 130.414 (+-1.633) | 446.804 (+-3.283) | 3.426 (+-0.000) 3 torch.uint8 channels_last bilinear (520, 520) -> (32, 32) aa=False | | 58.767 (+-0.203) | 374.244 (+-13.598) | 6.368 (+-0.000) 3 torch.uint8 channels_last bilinear (520, 520) -> (224, 224) aa=True | 283.210 (+-2.937) | 324.157 (+-1.895) | 720.197 (+-3.467) | 2.222 (+-0.000) 3 torch.uint8 channels_last bilinear (520, 520) -> (224, 224) aa=False | | 239.800 (+-2.492) | 592.834 (+-3.903) | 2.472 (+-0.000) 3 torch.uint8 channels_last bilinear (712, 712) -> (32, 32) aa=True | 186.255 (+-1.629) | 204.834 (+-1.496) | 787.868 (+-3.648) | 3.846 (+-0.000) 3 torch.uint8 channels_last bilinear (712, 712) -> (32, 32) aa=False | | 77.335 (+-0.341) | 651.016 (+-3.926) | 8.418 (+-0.000) 3 torch.uint8 channels_last bilinear (712, 712) -> (224, 224) aa=True | 410.286 (+-2.439) | 443.934 (+-2.899) | 1123.923 (+-14.988) | 2.532 (+-0.000) 3 torch.uint8 channels_last bilinear (712, 712) -> (224, 224) aa=False | | 312.220 (+-2.307) | 915.347 (+-4.486) | 2.932 (+-0.000) # More test-cases from #90771 3 torch.uint8 channels_last bilinear (64, 64) -> (224, 224) aa=True | 60.611 (+-0.337) | 80.849 (+-1.780) | 170.465 (+-1.830) | 2.108 (+-0.000) 3 torch.uint8 channels_last bilinear (224, 224) -> (270, 268) aa=True | 132.971 (+-1.624) | 164.892 (+-1.426) | 330.971 (+-3.249) | 2.007 (+-0.000) 3 torch.uint8 channels_last bilinear (256, 256) -> (1024, 1024) aa=True | 948.467 (+-3.179) | 891.414 (+-5.282) | 2805.510 (+-25.503) | 3.147 (+-0.000) 3 torch.uint8 channels_last bilinear (224, 224) -> (64, 64) aa=True | 52.539 (+-0.327) | 72.471 (+-0.367) | 135.933 (+-1.625) | 1.876 (+-0.000) 3 torch.uint8 channels_last bilinear (270, 268) -> (224, 224) aa=True | 138.669 (+-1.867) | 168.628 (+-1.213) | 321.112 (+-2.904) | 1.904 (+-0.000) 3 torch.uint8 channels_last bilinear (1024, 1024) -> (256, 256) aa=True | 689.933 (+-3.175) | 746.911 (+-2.985) | 2050.880 (+-22.188) | 2.746 (+-0.000) 3 torch.uint8 channels_last bilinear (64, 64) -> (224, 224) aa=False | | 78.347 (+-0.338) | 169.646 (+-1.640) | 2.165 (+-0.000) 3 torch.uint8 channels_last bilinear (224, 224) -> (270, 268) aa=False | | 162.194 (+-1.089) | 329.754 (+-2.590) | 2.033 (+-0.000) 3 torch.uint8 channels_last bilinear (256, 256) -> (1024, 1024) aa=False | | 894.476 (+-2.738) | 2815.870 (+-22.589) | 3.148 (+-0.000) 3 torch.uint8 channels_last bilinear (224, 224) -> (64, 64) aa=False | | 52.728 (+-0.406) | 112.024 (+-1.225) | 2.125 (+-0.000) 3 torch.uint8 channels_last bilinear (270, 268) -> (224, 224) aa=False | | 151.560 (+-1.128) | 299.152 (+-3.353) | 1.974 (+-0.000) 3 torch.uint8 channels_last bilinear (1024, 1024) -> (256, 256) aa=False | | 500.053 (+-4.288) | 1698.601 (+-16.785) | 3.397 (+-0.000) ``` Note: There is no perf regression for other case. There some cases (see Source below) with small speed-ups, for the rest it is roughly around 1.0 +/- 0.1 which may be attributed to noisy measurements ... [Source](https://gist.github.com/vfdev-5/1c0778904a07ce40401306548b9525e8#file-20230322-132441-pr_vs_nightly-speedup-md) ## Context - #90771 cc jgong5 mingfeima XiaobingSuper sanchitintel ashokei jingxu10 [ghstack-poisoned]
diff --git a/aten/src/ATen/native/cpu/UpSampleKernelAVXAntialias.h b/aten/src/ATen/native/cpu/UpSampleKernelAVXAntialias.h
@@ -325,8 +325,8 @@ void upsample_avx_bilinear_uint8(
     buffer_vert = at::empty({4, yout, xout}, input.options());
   }
 
-  // TODO: The unpack / pack operations create a
-  // copy of the original input and output tensor. There should be a way to avoid these copies by instead
+  // TODO: The unpack / pack operations create a copy of the original input and
+  // output tensor. There should be a way to avoid these copies by instead
   // modifying the low-level kernels. Or maybe at least avoid copying the entire
   // tensors and just copy part of them (line by line).
   for (const auto i : c10::irange(batch_size)) {
