[Snippets][CPU] Make ExtractUnsupportedTransposes HW dependent #32676
Conversation
aobolensk
changed the title
There was a problem hiding this comment.
Please revert changes in this file
| ::testing::Values(MHA::default_thread_count), | ||
| ::testing::Values(2), // decomposed Transpose + MHA | ||
| ::testing::Values(MHA::default_thread_count), | ||
| ::testing::Values(expected_nodes_mha_4d_f32), // decomposed Transpose + MHA |
There was a problem hiding this comment.
Let's remove the comment from here since it is not actual
| static constexpr size_t expected_nodes_mha_4d_f32 = 4; | ||
| #else | ||
| static constexpr size_t expected_nodes_mha_4d_f32 = 2; |
There was a problem hiding this comment.
Let's add an explanatory comment here
There was a problem hiding this comment.
Done, added for all ifdefs
There was a problem hiding this comment.
Same as above: let's leave a comment for expected_nodes_mha_mul_add , and remove the comment in INSTANTIATE_TEST_SUITE_P
The same for other test files
| static CommonOptimizations::Config conf(1, true); | ||
| static bool initialized = false; | ||
| if (!initialized) { | ||
| conf.set_transpose_support_callback([](const std::shared_ptr<const ov::Node>& node) -> bool { |
There was a problem hiding this comment.
Ideally, we need to share the callback code between tests and plugin transformation pipeline. Maybe we can introduce some helper for that? The helper could have e.g. bool parameter which defines if the "is_brgemm_case" check should be applied as well
What do you think?
There was a problem hiding this comment.
Good idea, implemented
aobolensk
force-pushed
the
snippets-extract-transpose-config
branch
from
c40d374 to
beebc20
Compare
aobolensk
force-pushed
the
snippets-extract-transpose-config
branch
from
2e50d87 to
107df70
Compare
2 participants
Details:
Introduce a callback function for
ExtractUnsupportedTransposespass as a part ofCommonOptimizations::Configto customize pass behavior depending on Transpose support.For example, ARM64 platform supports transpose decomposition, but MatMul with Transpose A/B is not supported so far. Rest of the (potential) platforms mark Transpose as not supported completely
An alternative approach for #32592
Tickets: