Make FloatToFloat16 conversion 75x faster using SVE2 instructions #3626
+124
−3
Conversation
This file contains 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
|
This pull request was exported from Phabricator. Differential Revision: D68520774 |
✅ Deploy Preview for pytorch-fbgemm-docs ready!
To edit notification comments on pull requests, go to your Netlify site configuration. |
|
This pull request was exported from Phabricator. Differential Revision: D68520774 |
embg
pushed a commit
to embg/FBGEMM
that referenced
this pull request
Jan 28, 2025
…torch#3626) Summary: Pull Request resolved: pytorch#3626 X-link: facebookresearch/FBGEMM#703 Rounding was previously (1) not vectorized and (2) [implemented in software](https://fburl.com/code/fa1jzpmo), so speeds were less than 1 byte per cycle. That's really slow. With SVE2 instructions, it's 75x faster (see test plan for measurement). That's due to a combination of vectorization + hardware support for rounding. This is a 0.25% CPU win for AdFinder on Grace: {F1974759899} Differential Revision: D68520774
embg
force-pushed
the
export-D68520774
branch
from
59a241d to
85aeb87
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68520774 |
embg
pushed a commit
to embg/FBGEMM
that referenced
this pull request
Jan 28, 2025
…torch#3626) Summary: Pull Request resolved: pytorch#3626 X-link: facebookresearch/FBGEMM#703 Rounding was previously (1) not vectorized and (2) [implemented in software](https://fburl.com/code/fa1jzpmo), so speeds were less than 1 byte per cycle. That's really slow. With SVE2 instructions, it's 75x faster (see test plan for measurement). That's due to a combination of vectorization + hardware support for rounding. This is a 0.25% CPU win for AdFinder on Grace: {F1974759899} Differential Revision: D68520774
embg
force-pushed
the
export-D68520774
branch
from
85aeb87 to
49f9fca
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68520774 |
embg
pushed a commit
to embg/FBGEMM
that referenced
this pull request
Jan 28, 2025
…torch#3626) Summary: Pull Request resolved: pytorch#3626 X-link: facebookresearch/FBGEMM#703 Rounding was previously (1) not vectorized and (2) [implemented in software](https://fburl.com/code/fa1jzpmo), so speeds were less than 1 byte per cycle. That's really slow. With SVE2 instructions, it's 75x faster (see test plan for measurement). That's due to a combination of vectorization + hardware support for rounding. This is a 0.25% CPU win for AdFinder on Grace: {F1974759899} Differential Revision: D68520774
embg
force-pushed
the
export-D68520774
branch
from
49f9fca to
641aa26
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68520774 |
embg
pushed a commit
to embg/FBGEMM
that referenced
this pull request
Jan 28, 2025
…torch#3626) Summary: Pull Request resolved: pytorch#3626 X-link: facebookresearch/FBGEMM#703 Rounding was previously (1) not vectorized and (2) [implemented in software](https://fburl.com/code/fa1jzpmo), so speeds were less than 1 byte per cycle. That's really slow. With SVE2 instructions, it's 75x faster (see test plan for measurement). That's due to a combination of vectorization + hardware support for rounding. This is a 0.25% CPU win for AdFinder on Grace: {F1974759899} Differential Revision: D68520774
embg
force-pushed
the
export-D68520774
branch
from
641aa26 to
0daa1c8
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68520774 |
embg
pushed a commit
to embg/FBGEMM
that referenced
this pull request
Jan 28, 2025
…torch#3626) Summary: Pull Request resolved: pytorch#3626 X-link: facebookresearch/FBGEMM#703 Rounding was previously (1) not vectorized and (2) [implemented in software](https://fburl.com/code/fa1jzpmo), so speeds were less than 1 byte per cycle. That's really slow. With SVE2 instructions, it's 75x faster (see test plan for measurement). That's due to a combination of vectorization + hardware support for rounding. Differential Revision: D68520774
embg
force-pushed
the
export-D68520774
branch
from
0daa1c8 to
bc6aaff
Compare
embg
pushed a commit
to embg/FBGEMM
that referenced
this pull request
Jan 28, 2025
…torch#3626) Summary: Pull Request resolved: pytorch#3626 X-link: facebookresearch/FBGEMM#703 Rounding was previously (1) not vectorized and (2) [implemented in software](https://fburl.com/code/fa1jzpmo), so speeds were less than 1 byte per cycle. That's really slow. With SVE2 instructions, it's 75x faster (see test plan for measurement). That's due to a combination of vectorization + hardware support for rounding. Differential Revision: D68520774
|
This pull request was exported from Phabricator. Differential Revision: D68520774 |
embg
force-pushed
the
export-D68520774
branch
from
bc6aaff to
8fb04c2
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68520774 |
embg
pushed a commit
to embg/FBGEMM
that referenced
this pull request
Jan 28, 2025
…torch#3626) Summary: Pull Request resolved: pytorch#3626 X-link: facebookresearch/FBGEMM#703 Rounding was previously (1) not vectorized and (2) [implemented in software](https://fburl.com/code/fa1jzpmo), so speeds were less than 1 byte per cycle. That's really slow. With SVE2 instructions, it's 75x faster (see test plan for measurement). That's due to a combination of vectorization + hardware support for rounding. Differential Revision: D68520774
embg
force-pushed
the
export-D68520774
branch
from
8fb04c2 to
dc5d7d6
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68520774 |
embg
pushed a commit
to embg/FBGEMM
that referenced
this pull request
Jan 28, 2025
…torch#3626) Summary: Pull Request resolved: pytorch#3626 X-link: facebookresearch/FBGEMM#703 Rounding was previously (1) not vectorized and (2) [implemented in software](https://fburl.com/code/fa1jzpmo), so speeds were less than 1 byte per cycle. That's really slow. With SVE2 instructions, it's 75x faster (see test plan for measurement). That's due to a combination of vectorization + hardware support for rounding. Differential Revision: D68520774
embg
force-pushed
the
export-D68520774
branch
from
dc5d7d6 to
a34a91d
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68520774 |
embg
pushed a commit
to embg/FBGEMM
that referenced
this pull request
Jan 28, 2025
…torch#3626) Summary: Pull Request resolved: pytorch#3626 X-link: facebookresearch/FBGEMM#703 Rounding was previously (1) not vectorized and (2) [implemented in software](https://fburl.com/code/fa1jzpmo), so speeds were less than 1 byte per cycle. That's really slow. With SVE2 instructions, it's 75x faster (see test plan for measurement). That's due to a combination of vectorization + hardware support for rounding. Differential Revision: D68520774
embg
force-pushed
the
export-D68520774
branch
from
a34a91d to
e460390
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68520774 |
embg
pushed a commit
to embg/FBGEMM
that referenced
this pull request
Jan 28, 2025
…torch#3626) Summary: Pull Request resolved: pytorch#3626 X-link: facebookresearch/FBGEMM#703 Rounding was previously (1) not vectorized and (2) [implemented in software](https://fburl.com/code/fa1jzpmo), so speeds were less than 1 byte per cycle. That's really slow. With SVE2 instructions, it's 75x faster (see test plan for measurement). That's due to a combination of vectorization + hardware support for rounding. Differential Revision: D68520774
embg
force-pushed
the
export-D68520774
branch
from
e460390 to
0815817
Compare
…torch#3626) Summary: Pull Request resolved: pytorch#3626 X-link: facebookresearch/FBGEMM#703 Rounding was previously (1) not vectorized and (2) [implemented in software](https://fburl.com/code/fa1jzpmo), so speeds were less than 1 byte per cycle. That's really slow. With SVE2 instructions, it's 75x faster (see test plan for measurement). That's due to a combination of vectorization + hardware support for rounding. Reviewed By: q10 Differential Revision: D68520774
|
This pull request was exported from Phabricator. Differential Revision: D68520774 |
embg
force-pushed
the
export-D68520774
branch
from
0815817 to
38dbb37
Compare
|
This pull request has been merged in c5e1cde. |
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:
X-link: https://github.com/facebookresearch/FBGEMM/pull/703
Rounding was previously (1) not vectorized and (2) implemented in software, so speeds were less than 1 byte per cycle. That's really slow.
With SVE2 instructions, it's 75x faster (see test plan for measurement). That's due to a combination of vectorization + hardware support for rounding.
Differential Revision: D68520774