[Software pipeline] Fix hardcoded index in access_ptr rewriting, add a GPU test with depth 4 - #11495
Merged
Merged
Conversation
masahi
changed the title
access_ptr rewriting, add a GPU test with depth 3 access_ptr rewriting, add a GPU test with depth 4
vinx13
approved these changes
May 27, 2022
masahi
force-pushed
the
software-pipe-index-fix
branch
from
May 27, 2022 20:36
b0b3a40 to
853a128
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fix a hardcoded index in
access_ptrrewriting, which assumes that the number of stages is 2.Refactored MMA code in test_tir_schedule_tensorize_ldmatrix_mma.py, so that it can be used by other tests. The new test in
test_tir_transform_inject_software_pipeline.pyapplies software pipelining annotations to the MMA-tensorized schedule withsoftware_pipeline_stage = [0, 0, 3], which makes global to shared load pipelined with depth 4. Without async copy, this is not useful for performance. But it does demonstrate that a multi-stage pipeline with depth > 2 works on a semi-realistic GPU schedule.The test uses large dynamic shared memory, which serves as a test case for #11478.
@vinx13 @junrushao1994 @csullivan