Improve Fused8BitRowwiseQuantizedSBFloatToFloatOrHalfNeon by 5%-15% #3860
Conversation
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
Summary: We are using SVE's pred to avoid the trailing loop, this technique shows an improved throughput of 5% to 15%. Benchmarks carry a lot of variance, thus produced assembly was checked: https://godbolt.org/z/K6scf6P8s Results are for the float32 case, used in AdRetriever before: rows, cols, elems_per_usec, GB/Sec 100, 16, 3536.86, 14.15 100, 64, 3363.99, 13.46 100, 128, 4674.71, 18.70 100, 256, 4708.46, 18.83 100, 512, 6247.59, 24.99 100, 1024, 6747.96, 26.99 100, 2048, 8719.72, 34.88 120, 16, 5210.26, 20.84 120, 64, 3850.57, 15.40 120, 128, 3688.19, 14.75 120, 256, 4566.65, 18.27 120, 512, 8684.83, 34.74 120, 1024, 8800.45, 35.20 120, 2048, 9396.97, 37.59 1000, 16, 5324.87, 21.30 1000, 64, 8846.61, 35.39 1000, 128, 8939.13, 35.76 1000, 256, 9537.45, 38.15 1000, 512, 9667.61, 38.67 1000, 1024, 7702.69, 30.81 1000, 2048, 8187.50, 32.75 after: rows, cols, elems_per_usec, GB/Sec 100, 16, 4040.40, 16.16 100, 64, 3600.29, 14.40 100, 128, 4740.05, 18.96 100, 256, 4605.53, 18.42 100, 512, 7603.47, 30.41 100, 1024, 8693.21, 34.77 100, 2048, 9166.19, 36.66 120, 16, 3143.01, 12.57 120, 64, 3919.11, 15.68 120, 128, 4766.29, 19.07 120, 256, 5489.12, 21.96 120, 512, 8569.90, 34.28 120, 1024, 8916.10, 35.66 120, 2048, 9849.85, 39.40 1000, 16, 4903.63, 19.61 1000, 64, 9281.85, 37.13 1000, 128, 10090.42, 40.36 1000, 256, 10314.59, 41.26 1000, 512, 10142.08, 40.57 1000, 1024, 7991.56, 31.97 1000, 2048, 8030.02, 32.12 Differential Revision: D71602944
|
This pull request was exported from Phabricator. Differential Revision: D71602944 |
✅ Deploy Preview for pytorch-fbgemm-docs ready!
To edit notification comments on pull requests, go to your Netlify site configuration. |
|
This pull request has been merged in bc30b06. |
q10
pushed a commit
to q10/FBGEMM
that referenced
this pull request
Apr 10, 2025
…ytorch#950) Summary: Pull Request resolved: facebookresearch/FBGEMM#950 X-link: pytorch#3860 We are using SVE's pred to avoid the trailing loop, this technique shows an improved throughput of 5% to 15%. Benchmarks carry a lot of variance, thus produced assembly was checked: https://godbolt.org/z/K6scf6P8s Results are for the float32 case, used in AdRetriever before: rows, cols, elems_per_usec, GB/Sec 100, 16, 3536.86, 14.15 100, 64, 3363.99, 13.46 100, 128, 4674.71, 18.70 100, 256, 4708.46, 18.83 100, 512, 6247.59, 24.99 100, 1024, 6747.96, 26.99 100, 2048, 8719.72, 34.88 120, 16, 5210.26, 20.84 120, 64, 3850.57, 15.40 120, 128, 3688.19, 14.75 120, 256, 4566.65, 18.27 120, 512, 8684.83, 34.74 120, 1024, 8800.45, 35.20 120, 2048, 9396.97, 37.59 1000, 16, 5324.87, 21.30 1000, 64, 8846.61, 35.39 1000, 128, 8939.13, 35.76 1000, 256, 9537.45, 38.15 1000, 512, 9667.61, 38.67 1000, 1024, 7702.69, 30.81 1000, 2048, 8187.50, 32.75 after: rows, cols, elems_per_usec, GB/Sec 100, 16, 4040.40, 16.16 100, 64, 3600.29, 14.40 100, 128, 4740.05, 18.96 100, 256, 4605.53, 18.42 100, 512, 7603.47, 30.41 100, 1024, 8693.21, 34.77 100, 2048, 9166.19, 36.66 120, 16, 3143.01, 12.57 120, 64, 3919.11, 15.68 120, 128, 4766.29, 19.07 120, 256, 5489.12, 21.96 120, 512, 8569.90, 34.28 120, 1024, 8916.10, 35.66 120, 2048, 9849.85, 39.40 1000, 16, 4903.63, 19.61 1000, 64, 9281.85, 37.13 1000, 128, 10090.42, 40.36 1000, 256, 10314.59, 41.26 1000, 512, 10142.08, 40.57 1000, 1024, 7991.56, 31.97 1000, 2048, 8030.02, 32.12 Reviewed By: embg Differential Revision: D71602944 fbshipit-source-id: 0d466ed954331f9f35b04332c65dcfceb5367e18
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Summary:
We are using SVE's pred to avoid the trailing loop, this technique shows an improved throughput of 5% to 15%.
Benchmarks carry a lot of variance, thus produced assembly was checked: https://godbolt.org/z/K6scf6P8s
Results are for the float32 case, used in AdRetriever
before:
rows, cols, elems_per_usec, GB/Sec
100, 16, 3536.86, 14.15
100, 64, 3363.99, 13.46
100, 128, 4674.71, 18.70
100, 256, 4708.46, 18.83
100, 512, 6247.59, 24.99
100, 1024, 6747.96, 26.99
100, 2048, 8719.72, 34.88
120, 16, 5210.26, 20.84
120, 64, 3850.57, 15.40
120, 128, 3688.19, 14.75
120, 256, 4566.65, 18.27
120, 512, 8684.83, 34.74
120, 1024, 8800.45, 35.20
120, 2048, 9396.97, 37.59
1000, 16, 5324.87, 21.30
1000, 64, 8846.61, 35.39
1000, 128, 8939.13, 35.76
1000, 256, 9537.45, 38.15
1000, 512, 9667.61, 38.67
1000, 1024, 7702.69, 30.81
1000, 2048, 8187.50, 32.75
after:
rows, cols, elems_per_usec, GB/Sec
100, 16, 4040.40, 16.16
100, 64, 3600.29, 14.40
100, 128, 4740.05, 18.96
100, 256, 4605.53, 18.42
100, 512, 7603.47, 30.41
100, 1024, 8693.21, 34.77
100, 2048, 9166.19, 36.66
120, 16, 3143.01, 12.57
120, 64, 3919.11, 15.68
120, 128, 4766.29, 19.07
120, 256, 5489.12, 21.96
120, 512, 8569.90, 34.28
120, 1024, 8916.10, 35.66
120, 2048, 9849.85, 39.40
1000, 16, 4903.63, 19.61
1000, 64, 9281.85, 37.13
1000, 128, 10090.42, 40.36
1000, 256, 10314.59, 41.26
1000, 512, 10142.08, 40.57
1000, 1024, 7991.56, 31.97
1000, 2048, 8030.02, 32.12
Differential Revision: D71602944