Arm backend: regression test for equal non-simple IFM scale Add on Ethos-U55 (#21151)#21151
Arm backend: regression test for equal non-simple IFM scale Add on Ethos-U55 (#21151)#211513l1 wants to merge 1 commit into
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21151
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ✅ No FailuresAs of commit fa7dd6c with merge base 3802831 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
|
@3l1 has exported this pull request. If you are a Meta employee, you can view the originating Diff in D113307975. |
This PR needs a
|
…hos-U55 (#21151) Summary: Add EqualScaleAdd (x + x) tests on Ethos-U55 and Ethos-U85. When both Add/Sub operands share one equal, non-simple (non-power-of-2) scale, the Ethos-U55 equal-scale code path writes both scales into 16-bit OPA/OPB registers and the wide value can overflow to zero, silently dropping an operand (x + x returns x, i.e. the output is halved; softmax's Sub(x, max) loses an operand). Correct lowering keeps the operand as an explicit rescale so the FVP result matches the quantized reference. The check is Ethos-U55-codegen-specific (it reproduces on the FVP, not on the TOSA reference path). Non-power-of-2 input magnitudes force the non-simple equal scale that triggers the overflow. Reviewed By: rascani Differential Revision: D113307975
meta-codesync
Bot
changed the title
d2d9cf5 to
fa7dd6c
Compare
2 participants
Summary:
Add EqualScaleAdd (x + x) tests on Ethos-U55 and Ethos-U85. When both Add/Sub
operands share one equal, non-simple (non-power-of-2) scale, the Ethos-U55
equal-scale code path writes both scales into 16-bit OPA/OPB registers and the
wide value can overflow to zero, silently dropping an operand (x + x returns x,
i.e. the output is halved; softmax's Sub(x, max) loses an operand). Correct
lowering keeps the operand as an explicit rescale so the FVP result matches the
quantized reference.
The check is Ethos-U55-codegen-specific (it reproduces on the FVP, not on the
TOSA reference path). Non-power-of-2 input magnitudes force the non-simple equal
scale that triggers the overflow.
Reviewed By: rascani
Differential Revision: D113307975