Repository navigation
Fix NaN sign comparison for Sum/Average/Product in TensorValueReductionsPreserveScalarOrder - #136650
Merged
Merged
Conversation
…rder test For the [-Inf, +Inf, -0.0, 0.0, NaN, -3] input, Sum (-Inf + +Inf) and Product (Inf * 0) construct a canonical NaN that then combines with the input NaN. Which NaN propagates depends on operand order, which JitStress perturbs on arm64. Average is Sum / n. Use AssertBitsOrNaN for these, as was done for Dot and CosineSimilarity; keep exact-bit checks for Norm, Distance, StdDev and the min/max family, which never construct a NaN from this input (StdDev applies Abs before squaring). Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
|
Azure Pipelines: Successfully started running 3 pipeline(s). 13 pipeline(s) were filtered out due to trigger conditions. There may be pipelines that require an authorized user to comment /azp run to run. |
Contributor
|
Tagging subscribers to this area: @dotnet/area-system-numerics-tensors |
Contributor
There was a problem hiding this comment.
🟢 Approval recommended
No active review issues were identified for the test-only change to tolerate NaN sign and payload differences.
No open issues 🎉
What changed in this PR
Updates the tensor scalar-order reduction test to tolerate NaN sign and payload differences for Sum, Average, and Product, extending the earlier fixes for Dot and CosineSimilarity.
Changes:
- Uses the existing
AssertBitsOrNaNhelper for the three affected assertions while preserving exact-bit comparisons for non-NaN results. - Leaves other reduction assertions and production code unchanged.
| File | Description |
|---|---|
src/libraries/System.Numerics.Tensors/tests/TensorTests.cs |
Adjusts three assertions to tolerate NaN differences caused by operand ordering under JIT stress. |
Review effort: Balanced
Member
Author
|
@tannergooding I think this is correct? |
tannergooding
approved these changes
Oct 29, 2026
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.
Follow-up to #136438 (Dot) and #136552 (CosineSimilarity).
TensorValueReductionsPreserveScalarOrderfails on linux-arm64libraries-jitstresslegs at theTensor.Sumassertion (TensorTests.cs:1207): expected0x7FF8000000000000(+NaN), actual0xFFF8000000000000(-NaN).For the input
[-Inf, +Inf, -0.0, 0.0, NaN, -3],-Inf + +Infconstructs a canonical NaN that is later combined with the inputdouble.NaN. When both operands are NaN, which one propagates depends on operand order, and JitStress changes that order.Following the guidance in #136438 (keep exact NaN payload checks except where a canonical NaN may be constructed), this switches only these assertions to the existing
AssertBitsOrNaNhelper:Tensor.Sum:-Inf + +Inf(the failing assertion)Tensor.Average:Sum / n. It fails next once Sum is relaxed.Tensor.Product:Inf * 0constructs a canonical NaN that then combines with the input NaN. It did not fail locally, but it has the same order sensitivity.Unchanged (exact bits):
NormandDistancenever construct a NaN from this input, so only the input NaN propagates.StdDevappliesAbsbefore squaring in both the product code and the test, which normalizes the sign. The min/max family only selects input values.Validation (osx-arm64, Release runtime/libs with a Checked
libclrjitswapped into testhost)DOTNET_JitStress=1or2withDOTNET_TieredCompilation=0reproduces the CI failure exactly (layouts 0, 1, 2, 4, 5 fail at line 1207 with the same bits).TailcallStress,ReadyToRun=0, and with no stress.System.Numerics.Tensors.Tests: 6505 passed, 0 failed, underJitStress=2+TieredCompilation=0and again with the Release JIT.Resolves #136647
Note
This PR description was generated with GitHub Copilot.