commit eb15493eac2746668e237b7ac1ddaca88fdea75e
author Benoit Jacob <jacob.benoit.1@gmail.com> Wed Oct 09 10:07:20 2024 -0400
committer GitHub <noreply@github.com> Wed Oct 09 14:07:20 2024 +0000
tree 77856db15e346b21e0227651fca027585b488dcc
parent 5270093401cea05a548352fce312e39d0291024c

e2e matmul test improvements (#18725)

This PR is made of individual commits for review convenience and so we
can drop anything that causes problems on CI.

* Add default shapes set, combining small and large.
* The need to specify "small" or "large" is a real need in only a
minority of cases. That's a difference from when these tests were first
added.
* Enable dynamic sizes in large shapes, leaving only gpu_large_aligned
out.
* Who remembered that large shapes weren't tested as dynamic shapes,
unlike small shapes... and unlike "gpu_large" shapes?!
* Rename gpu_large_aligned -> easy_large_static.
* This is only needed in sketchy GPU codegen pipelines that can't deal
with sizes that aren't multiples of some internal tile size.
* Fold gpu_large into large and tolerate fuzzy bf16 accumulators.
* Retaining the evidently more curated set of shapes from "gpu_large".
The larger sizes ran into new issues with the mostly artificial case of
bf16 accumulators.
* Use default shapes and reenable sanitizers.
* This simplifies the build, reduces the number of targets and increases
coverage as "default" combines small and large shapes. And this
reenables sanitizers that hard been disabled on large sizes due to
timeouts. As tests at some point started verifying only a subset of
result matrix elements, the timeouts should be avoided now.
* Enable default shapes for most rocm tests.
  * The motivation for this PR. The rest just bubbled up from there.
* Make large shapes more diverse (including odd and rectangular kinds of
shapes).

---------

Signed-off-by: Benoit Jacob <jacob.benoit.1@gmail.com>