[TOSA] Fix output size calculation for pool ops #4125
Conversation
justin-ngo-arm
force-pushed
the
pool-output-size-fix
branch
from
d2ef123 to
63ab68e
Compare
|
@sahas3, I have updated my code according to your comments as well as added some e2e tests. |
TOSA requires (inputDim + padBefore + padAfter - kernel) to be fully divisible by stride. This update adds pad and input size modifications for pooling ops (AvgPool2d and MaxPool2d) to satisfy that requirement by TOSA. Signed-off-by: Justin Ngo <[email protected]> Change-Id: Iab4021f2dda87cb87e54e4e9ca20bd3688dc1c50
justin-ngo-arm
force-pushed
the
pool-output-size-fix
branch
from
63ab68e to
1c9c78c
Compare
|
@vivekkhandelwal1, can you help me retrigger the CI for this PR please? CI failed because of time out due to unavailable resources. |
Done |
|
@vivekkhandelwal1 can you help me check the CI please? I'm not sure what's wrong with it.
|
|
@justin-ngo-arm, this change caused a LIT failure in Windows CI https://github.com/llvm/torch-mlir/actions/runs/14649870432/job/41112829976#step:8:5957 likely because |
|
After this PR was merged, the Windows build started failing https://github.com/llvm/torch-mlir/actions/runs/14649870432/job/41112829976. @justin-ngo-arm can you please take a look? |
|
@sahas3, @vivekkhandelwal1, I'm working on a PR to fix this. I will submit it ASAP. |
TOSA requires (inputDim + padBefore + padAfter - kernel) to be fully divisible by stride. This update adds pad and input size modifications for pooling ops (AvgPool2d and MaxPool2d) to satisfy that requirement by TOSA.