Add microbenchmark for layer normalization and improve latency #22223
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
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
amarin16
changed the title
This reverts commit 6aece95.
|
please add a summary of benchmark results. before vs. after |
fajin-corp
reviewed
Oct 11, 2024
fajin-corp
reviewed
Oct 11, 2024
fajin-corp
reviewed
Oct 11, 2024
fajin-corp
reviewed
Oct 11, 2024
fajin-corp
reviewed
Oct 11, 2024
|
instead of converting fp16 to fp32, do you plan to implement fp16 kernels? |
that can be done in a separate PR |
fajin-corp
reviewed
Oct 14, 2024
fajin-corp
reviewed
Oct 14, 2024
fajin-corp
approved these changes
Oct 14, 2024
fs-eire
added a commit
that referenced
this pull request
Oct 18, 2024
### Description The recent PR #22223 introduced 2 bugs in implementation of CPU LayerNorm f16: - possible access to nullptr for bias `const TensorShape& bias_shape = bias->Shape();` will crash when `bias` does not exist. (amazingly seems this one is not coverred by any test case) - fix: guard with pointer check - a racing condition inside ComputeJob `ComputeJob()` is dispatched to threadpool and it internally tries to modify `LayerNormImpl::scale_fp32_` and `LayerNormImpl::bias_fp32_`, which are `std::unique_ptr`s and are not thread-safe. - fix: move the modification of `LayerNormImpl::scale_fp32_` and `LayerNormImpl::bias_fp32_` out of `ComputeJob()` and put into `LayerNormImpl::ComputeWithoutContext()`. It may still have racing condition because `ConcurrentRunSupported` is set to `true` for CPU EP. Added an OrtMutex. This should fixes the recent flaky tests as well.
guschmue
pushed a commit
that referenced
this pull request
Oct 18, 2024
- Added a microbenchmark for the `LayerNormalization` MLFloat16 support added in #22063. - Updated the `LayerNormalization` MLFloat16 implementation to improve the latency. ``` ---------------------------------------------------------------------------------------------- Original MLFloat16 support Time CPU Iterations ---------------------------------------------------------------------------------------------- BM_LayerNormalization<MLFloat16, float>/1/real_time 15599 us 15625 us 47 BM_LayerNormalization<MLFloat16, float>/1/real_time 14714 us 14824 us 39 BM_LayerNormalization<MLFloat16, float>/1/real_time 14634 us 14688 us 50 ---------------------------------------------------------------------------------------------- Updated MLFloat16 support Time CPU Iterations ---------------------------------------------------------------------------------------------- BM_LayerNormalization<MLFloat16, float>/1/real_time 7276 us 7254 us 84 BM_LayerNormalization<MLFloat16, float>/1/real_time 6820 us 6720 us 93 BM_LayerNormalization<MLFloat16, float>/1/real_time 6840 us 6882 us 84 ```
guschmue
pushed a commit
that referenced
this pull request
Oct 18, 2024
### Description The recent PR #22223 introduced 2 bugs in implementation of CPU LayerNorm f16: - possible access to nullptr for bias `const TensorShape& bias_shape = bias->Shape();` will crash when `bias` does not exist. (amazingly seems this one is not coverred by any test case) - fix: guard with pointer check - a racing condition inside ComputeJob `ComputeJob()` is dispatched to threadpool and it internally tries to modify `LayerNormImpl::scale_fp32_` and `LayerNormImpl::bias_fp32_`, which are `std::unique_ptr`s and are not thread-safe. - fix: move the modification of `LayerNormImpl::scale_fp32_` and `LayerNormImpl::bias_fp32_` out of `ComputeJob()` and put into `LayerNormImpl::ComputeWithoutContext()`. It may still have racing condition because `ConcurrentRunSupported` is set to `true` for CPU EP. Added an OrtMutex. This should fixes the recent flaky tests as well.
tianleiwu
pushed a commit
that referenced
this pull request
Oct 18, 2024
### Description The recent PR #22223 introduced 2 bugs in implementation of CPU LayerNorm f16: - possible access to nullptr for bias `const TensorShape& bias_shape = bias->Shape();` will crash when `bias` does not exist. (amazingly seems this one is not coverred by any test case) - fix: guard with pointer check - a racing condition inside ComputeJob `ComputeJob()` is dispatched to threadpool and it internally tries to modify `LayerNormImpl::scale_fp32_` and `LayerNormImpl::bias_fp32_`, which are `std::unique_ptr`s and are not thread-safe. - fix: move the modification of `LayerNormImpl::scale_fp32_` and `LayerNormImpl::bias_fp32_` out of `ComputeJob()` and put into `LayerNormImpl::ComputeWithoutContext()`. It may still have racing condition because `ConcurrentRunSupported` is set to `true` for CPU EP. Added an OrtMutex. This should fixes the recent flaky tests as well.
apsonawane
pushed a commit
that referenced
this pull request
Oct 22, 2024
### Description The recent PR #22223 introduced 2 bugs in implementation of CPU LayerNorm f16: - possible access to nullptr for bias `const TensorShape& bias_shape = bias->Shape();` will crash when `bias` does not exist. (amazingly seems this one is not coverred by any test case) - fix: guard with pointer check - a racing condition inside ComputeJob `ComputeJob()` is dispatched to threadpool and it internally tries to modify `LayerNormImpl::scale_fp32_` and `LayerNormImpl::bias_fp32_`, which are `std::unique_ptr`s and are not thread-safe. - fix: move the modification of `LayerNormImpl::scale_fp32_` and `LayerNormImpl::bias_fp32_` out of `ComputeJob()` and put into `LayerNormImpl::ComputeWithoutContext()`. It may still have racing condition because `ConcurrentRunSupported` is set to `true` for CPU EP. Added an OrtMutex. This should fixes the recent flaky tests as well.
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.
LayerNormalizationMLFloat16 support added in Add MLFloat16 support for LayerNormalization, SkipLayerNormalization #22063.LayerNormalizationMLFloat16 implementation to improve the latency.