Add NEON-accelerated int8mm for bfloat16 #125290
Closed
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
Apparently `vshlq_u32` is faster than `vcvt_f32_f16` I.e. the same stories110M run at 60 tokens/sec with f16, but at 66 tokens/sec with bf16
malfet
added
release notes: performance_as_product
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/125290
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 954400e with merge base ea347fa ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
pytorch-bot
bot
added
the
module: cpu
|
@pytorchbot merge |
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
andoorve
pushed a commit
to andoorve/pytorch
that referenced
this pull request
May 1, 2024
As apparently `vshlq_u32` is faster than `vcvt_f32_f16` Refactor NEON `tinygemm_kernel` to rely on `load_as_float32x4` and `load_as_float32x4x2` and implement them for float16 (using vcvt), bfloat16 (using left shift) and plain float32 (not using anything) As result stories110M run at 60 tokens/sec with f16, but at 66 tokens/sec with bf16 and 75 tokens/sec with f32, though more bandwith demand starts to favor reduced floating types as model size gets bigger. Pull Request resolved: pytorch#125290 Approved by: https://github.com/mikekgfb
petrex
pushed a commit
to petrex/pytorch
that referenced
this pull request
May 3, 2024
As apparently `vshlq_u32` is faster than `vcvt_f32_f16` Refactor NEON `tinygemm_kernel` to rely on `load_as_float32x4` and `load_as_float32x4x2` and implement them for float16 (using vcvt), bfloat16 (using left shift) and plain float32 (not using anything) As result stories110M run at 60 tokens/sec with f16, but at 66 tokens/sec with bf16 and 75 tokens/sec with f32, though more bandwith demand starts to favor reduced floating types as model size gets bigger. Pull Request resolved: pytorch#125290 Approved by: https://github.com/mikekgfb
malfet
added a commit
that referenced
this pull request
May 17, 2024
It used to be vectorized only for f16, but no reason not to do the same for bf16 or f32 Spiritual followup of #125290
pytorchmergebot
pushed a commit
that referenced
this pull request
May 17, 2024
It used to be vectorized only for f16, but no reason not to do the same for bf16 or f32 Spiritual followup of #125290 Pull Request resolved: #126512 Approved by: https://github.com/Skylion007
ZelboK
pushed a commit
to ZelboK/pytorch
that referenced
this pull request
May 19, 2024
It used to be vectorized only for f16, but no reason not to do the same for bf16 or f32 Spiritual followup of pytorch#125290 Pull Request resolved: pytorch#126512 Approved by: https://github.com/Skylion007
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.
As apparently
vshlq_u32is faster thanvcvt_f32_f16Refactor NEON
tinygemm_kernelto rely onload_as_float32x4andload_as_float32x4x2and implement them for float16 (using vcvt), bfloat16 (using left shift) and plain float32 (not using anything)As result stories110M run at 60 tokens/sec with f16, but at 66 tokens/sec with bf16 and 75 tokens/sec with f32, though more bandwith demand starts to favor reduced floating types as model size gets bigger.
cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10