Compute input min/max with a single vectorized pass in DynamicQuantizeLinear #531
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
There was already a `max` method, so this fills in a gap.
This is useful for reductions which need to compute multiple values in one pass over the data.
This allows for computing the minimum and maximum values in a slice of floats with one pass over the slice.
Previously two separate passes over the data were used to compute the min/max values. Use the `MinMax` op from rten-vecmath to compute this in one vectorized pass. In a benchmark with a quantized ModernBERT model this made DynamicQuantizeLinear 2.5-3x faster.
|
For reference, ORT is using the same intrinsics I am here, or the SSE version with the same behavior in the case of x64 - https://github.com/microsoft/onnxruntime/blob/ecdeecae617d1b37b42bca51e1ade979dd260961/onnxruntime/core/mlas/lib/mlasi.h#L2230. |
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Outdated suggestions cannot be applied.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.
Combine the separate passes over the input to compute the min/max in DynamicQuantizeLinear with a single vectorized pass.
There is a caveat that the new implementation doesn't guarantee the same handling of NANs in the input as before, and this will vary by architecture. The ReduceMin / ReduceMax ops always propagate NANs, whereas this implementation just uses the obvious min/max intrinsic (eg.
_m256_min_ps) which may do something else.