Cortex-M backend: Add mul and linear tests #14746
Conversation
Minor included fixes: - Make quantized_linear_fusion_pass an XNNPACK pass to initialize it with an exported program - Add TO_EXECUTORCH as a valid stage after RUN_PASSES - Add ramp_tensor function to simplify creating dummy data Signed-off-by: Adrian Lundell <adrian.lundell@arm.com> Change-Id: Id13be6427390483aa1df1b76fc363ae4d0eae876
AdrianLundell
added
partner: arm
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14746
Note: Links to docs will display an error until the docs builds have been completed. ❌ 4 New FailuresAs of commit 048b0c0 with merge base 75f968d ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
bot
added
the
CLA Signed
|
@digantdesai @psiddh What do you think? |
|
Looks like "int32 NameError" specifically happens in the dialect transformation stage for these two test cases: linear_rank2_pos and linear_bias. This appears more fundamental issue , isn't ? |
|
I think it has to do with how meta-data is propagated in the pass if I remember correctly, the error is fixed in an upcoming patch at least! |
Signed-off-by: Adrian Lundell <adrian.lundell@arm.com>
| def __init__(self): | ||
| super().__init__( | ||
| XNNPACKPassManager, pass_list=[QuantizedOpFusionPass, ReplaceQuantNodesPass] | ||
| XNNPACKPassManager, |
There was a problem hiding this comment.
Let's fork it and put it under backends/cortex_m
I was thinking s/XNNPACK/MCU but we can do the rebranding later and if you want to stick to cortex_m then s/XNNPACK/CortexM ?
There was a problem hiding this comment.
I went with CortexM in here 57a7903 so it will have to be rebranded later
| @@ -0,0 +1,211 @@ | |||
| # Copyright 2025 Arm Limited and/or its affiliates. | |||
| self.check_count(ops_before_transforms) | ||
| self.run_passes() | ||
| self.check_count(ops_after_transforms) | ||
| self.run_method_and_compare_outputs(inputs=self.example_inputs, qtol=qtol) |
There was a problem hiding this comment.
should we merge test_dialect and test_impl?
There was a problem hiding this comment.
I kind of like the split since the python implementation is so separate from the c++ implementation and it's nice to get feedback on only the thing you are working on. On the other hand you can simply modify the stages of the tester during development to get this so it should be fine!
There was a problem hiding this comment.
Given, (1) python eager dq->op->q, (2) python impl of the cortex-m op, and (3) C++ impl of the cortex-m op, I guess it makes sense to do A: (1) vs. (2) and B: (1) vs. (3) but when I suggested to merge I assumed we don't really care about (2) besides unit-tests, unless the second comparison (i.e. B) is (2) vs. (3).
digantdesai
added
the
module: microcontrollers
3 participants
Minor included fixes:
cc @freddan80 @per @zingo @oscarandersson8218 @digantdesai @psiddh