Better handle types and overflows in softmax (especially lagacy and stable)#1476
Draft
jmitrevs wants to merge 6 commits into
Draft
Better handle types and overflows in softmax (especially lagacy and stable)#1476jmitrevs wants to merge 6 commits into
jmitrevs wants to merge 6 commits into
Conversation
There was a problem hiding this comment.
Pull request overview
This PR focuses on making the Vivado softmax implementations more robust against overflow by improving internal type handling (notably for legacy and latency variants), while also improving default type choices for stable softmax and adding auto-precision inference support for Softmax accumulators.
Changes:
- Updated Vivado softmax templates to use accumulator types for summation / inversion-table addressing and to split exp/inv table types (legacy).
- Updated Vivado stream softmax templates accordingly (latency/stable/legacy).
- Added Softmax-specific
accum_tinference inInferPrecisionTypesand adjusted FPGA backend Softmax default type attributes (unsigned table defaults, newinp_norm_tattribute).
Reviewed changes
Copilot reviewed 4 out of 4 changed files in this pull request and generated 4 comments.
| File | Description |
|---|---|
| hls4ml/templates/vivado/nnet_utils/nnet_activation.h | Aligns softmax (latency/stable/legacy) internal types to reduce overflow risk and clarifies exp/inv table type usage. |
| hls4ml/templates/vivado/nnet_utils/nnet_activation_stream.h | Mirrors softmax type changes for streaming implementations (but currently has inconsistencies that need fixing). |
| hls4ml/model/optimizer/passes/infer_precision.py | Adds Softmax-specific accumulator type inference when precision is set to auto. |
| hls4ml/backends/fpga/fpga_backend.py | Updates Softmax type attribute defaults (unsigned tables, adds inp_norm_t, adjusts accumulator handling). |
Comments suppressed due to low confidence (1)
hls4ml/templates/vivado/nnet_utils/nnet_activation_stream.h:295
- This legacy stream implementation still uses
CONFIG_T::table_sizeto compute/clampindex, butexp_tableis now allocated withCONFIG_T::exp_table_size. If these differ (and they can, since softmax exposes separate exp/inv table sizes),indexcan go out of bounds or the scaling will be wrong. Useexp_table_sizeconsistently for exp table addressing here.
auto data_round = (data_cache[j] - data_cache[i]) * CONFIG_T::table_size / 16;
int index = data_round + 8 * CONFIG_T::table_size / 16;
if (index < 0)
index = 0;
if (index > CONFIG_T::table_size - 1)
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Contributor
Author
|
I should also update the oneAPI backend. Truthfully that is a bit more involved. I will look into it this month, but have to work some hardware tests the next two weeks, so this for the moment will be lower priority. |
jmitrevs
added
please test
jmitrevs
added
please test
Contributor
Author
|
On the oneAPI (and Quartus) side there are already some softmax changes in PR #1432 so we need to coordinate. |
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
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.
Description
The legacy softmax was prone to overflow, so this PR made it use an accumulator type. Also, for stable, the default types are better chosen. There are some changes in latency, as well, to avoid overflow, but my feeling is that latency can often behave badly, so this doesn't fix that.
Type of change
Tests
The standard softmax tests should still pass. One could consider adding specific overflow tests, like we have in the current version of the brevitas tutorial.
Checklist
pre-commiton the files I edited or added.