optimize Half performance for maxpool on CPU#101379
Closed
CaoE wants to merge 3 commits intogh/CaoE/25/basefrom
Closed
optimize Half performance for maxpool on CPU#101379CaoE wants to merge 3 commits intogh/CaoE/25/basefrom
CaoE wants to merge 3 commits intogh/CaoE/25/basefrom
Conversation
[ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/101379
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 263f398: This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This was referenced May 15, 2023
### Testing Single socket (28 cores): shape | fp32 forward / ms | fp16 forward / ms | bf16 forward / ms | fp32 backward / ms | fp16 backward / ms | bf16 backward / ms -- | -- | -- | -- | -- | -- | -- size: (1, 56, 264, 264), kernel: 3, stride: 1, mem_format: contig | 3.959584 | 4.774427414 | 5.493402 | 0.557232 | 0.59731 | 0.568485 size: (1, 56, 264, 264), kernel: 3, stride: 1, mem_format: CL| 0.815511 | 1.413416862 | 1.351261 | 5.710506 | 9.366035 | 10.57506 size: (32, 16, 200, 200), kernel: 3, stride: 1, mem_format: contig | 21.28074 | 25.78539848 | 29.55482 | 6.921601 | 3.817368 | 3.869653 size: (32, 16, 200, 200), kernel: 3, stride: 1, mem_format: CL | 5.755067 | 4.850935936 | 4.793715 | 6.737041 | 4.509115 | 5.024385 size: (32, 32, 100, 100), kernel: 3, stride: 1, mem_format: contig | 10.63426 | 13.64016 | 15.28637 | 2.67656 | 1.67952 | 1.71365 size: (32, 32, 100, 100), kernel: 3, stride: 1, mem_format: CL | 2.63570 | 2.05289 | 2.05532 | 2.55452 | 2.11025 | 2.33923 size: (4, 19, 10, 16, 16), kernel: 3, stride: 1, mem_format: contig | 0.375469 | 0.421462059 | 0.479748 | 0.066364 | 0.065937 | 0.065155 size: (4, 19, 10, 16, 16), kernel: 3, stride: 1, mem_format: CL3d | 0.112197 | 0.117073059 | 0.112326 | 0.111697 | 0.141022 | 0.145364 Single core: shape | fp32 forward / ms | fp16 forward / ms | bf16 forward / ms | fp32 backward / ms | fp16 backward / ms | bf16 backward / ms -- | -- | -- | -- | -- | -- | -- size: (1, 56, 264, 264), kernel: 3, stride: 1, mem_format: contig | 92.16582 | 111.3258338 | 128.6513 | 6.684325 | 10.9745 | 12.21541 size: (1, 56, 264, 264), kernel: 3, stride: 1, mem_format: CL | 10.14318 | 28.9168644 | 29.80297 | 7.350142 | 9.718766 | 11.25323 size: (32, 16, 200, 200), kernel: 3, stride: 1, mem_format: contig | 485.4677 | 586.6346097 | 674.8961 | 44.93304 | 61.85198 | 68.8631 size: (32, 16, 200, 200), kernel: 3, stride: 1, mem_format: CL | 81.45644 | 86.70579195 | 89.43031 | 46.44922 | 53.36975 | 62.60659 size: (32, 32, 100, 100), kernel: 3, stride: 1, mem_format: contig | 238.55453 | 290.02304 | 331.89967 | 19.694657 | 29.75552 | 32.78853 size: (32, 32, 100, 100), kernel: 3, stride: 1, mem_format: CL | 30.17079 | 31.60022 | 32.75628 | 22.44543 | 25.69056 |30.17796 size: (4, 19, 10, 16, 16), kernel: 3, stride: 1, mem_format: contig | 7.474389 | 8.705227375 | 9.937217 | 0.236015 | 0.410647 | 0.434229 size: (4, 19, 10, 16, 16), kernel: 3, stride: 1, mem_format: CL3d | 2.318954 | 2.456078529 | 2.469444 | 0.262125 | 0.353301 | 0.401361 cc jgong5 mingfeima XiaobingSuper sanchitintel ashokei jingxu10 [ghstack-poisoned]
Collaborator
|
Please elaborate how the changes contribute to the performance in the PR description. |
jgong5
approved these changes
May 16, 2023
### Motivation Scalar conversion between Half and Float on CPU is more time consuming compared to BFloat16 <-> Float. Use scalar conversion with avx instructions for Half to speed up the Half performance of maxpool on CPU. ### Testing Single socket (28 cores): shape | fp16 forward / ms | bf16 forward / ms | fp16 backward / ms | bf16 backward / ms | speedup ratio (fp16 forward) | speedup ratio (fp16 backward) -- | -- | -- | -- | -- | -- | -- size: (1, 56, 264, 264), kernel: 3, stride: 1, mem_format: contig | 4.774427 | 5.493402 | 0.59731 | 0.568485 | 1.459212 | 3.330214 size: (1, 56, 264, 264), kernel: 3, stride: 1, mem_format: CL | 1.413417 | 1.351261 | 9.366035 | 10.57506 | 1.335862 | 3.897978 size: (32, 16, 200, 200), kernel: 3, stride: 1, mem_format: contig | 25.7854 | 29.55482 | 3.817368 | 3.869653 | 1.469973 | 2.779821 size: (32, 16, 200, 200), kernel: 3, stride: 1, mem_format: CL | 4.850936 | 4.793715 | 4.509115 | 5.024385 | 0.971149 | 3.263856 size: (32, 32, 100, 100), kernel: 3, stride: 1, mem_format: contig | 13.64016 | 15.28637 | 1.67952 | 1.71365 | 1.323064 | 2.954249 size: (32, 32, 100, 100), kernel: 3, stride: 1, mem_format: CL | 2.05289 | 2.05532 | 2.11025 | 2.33923 | 0.978913 | 3.416325 size: (4, 19, 10, 16, 16), kernel: 3, stride: 1, mem_format: contig | 0.421462 | 0.479748 | 0.065937 | 0.065155 | 2.097413 | 1.751505 size: (4, 19, 10, 16, 16), kernel: 3, stride: 1, mem_format: CL3d | 0.117073 | 0.112326 | 0.141022 | 0.145364 | 2.160446 | 3.780899 Single core: shape | fp16 forward / ms | bf16 forward / ms | fp16 backward / ms | bf16 backward / ms | speedup ratio (fp16 forward) | speedup ratio (fp16 backward) -- | -- | -- | -- | -- | -- | -- size: (1, 56, 264, 264), kernel: 3, stride: 1, mem_format: contig | 111.3258 | 128.6513 | 10.9745 | 12.21541 | 1.468419 | 3.77314 size: (1, 56, 264, 264), kernel: 3, stride: 1, mem_format: CL | 28.91686 | 29.80297 | 9.718766 | 11.25323 | 1.327646 | 3.804531 size: (32, 16, 200, 200), kernel: 3, stride: 1, mem_format: contig | 586.6346 | 674.8961 | 61.85198 | 68.8631 | 1.458279 | 3.544892 size: (32, 16, 200, 200), kernel: 3, stride: 1, mem_format: CL | 86.70579 | 89.43031 | 53.36975 | 62.60659 | 1.053431 | 3.822342 size: (32, 32, 100, 100), kernel: 3, stride: 1, mem_format: contig | 290.023 | 331.8997 | 29.75552 | 32.78853 | 1.444988 | 3.601094 size: (32, 32, 100, 100), kernel: 3, stride: 1, mem_format: CL | 31.60022 | 32.75628 | 25.69056 | 30.17796 | 1.061671 | 3.836374 size: (4, 19, 10, 16, 16), kernel: 3, stride: 1, mem_format: contig | 8.705227 | 9.937217 | 0.410647 | 0.434229 | 1.599131 | 3.294046 size: (4, 19, 10, 16, 16), kernel: 3, stride: 1, mem_format: CL3d | 2.456079 | 2.469444 | 0.353301 | 0.401361 | 1.371674 | 3.521077 cc jgong5 mingfeima XiaobingSuper sanchitintel ashokei jingxu10 [ghstack-poisoned]
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stack from ghstack (oldest at bottom):
Motivation
Scalar conversion between Half and Float on CPU is more time consuming compared to BFloat16 <-> Float. Use scalar conversion with avx instructions for Half to speed up the Half performance of maxpool on CPU.
Testing
Single socket (28 cores):
Single core:
cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10