Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Modify pass pipeline to allow lowering tt.load to 2DBlockRead #1061

Merged
merged 11 commits into from
May 10, 2024
Merged

Modify pass pipeline to allow lowering tt.load to 2DBlockRead #1061

merged 11 commits into from
May 10, 2024

Conversation

etiotto
Copy link
Contributor

@etiotto etiotto commented May 7, 2024

No description provided.

@etiotto etiotto self-assigned this May 7, 2024
@etiotto etiotto linked an issue May 7, 2024 that may be closed by this pull request
Add passes to enable lowering tt.load using block ptrs to 2DBlockRead #1060
Closed
@etiotto etiotto requested a review from chengjunlu May 7, 2024 19:14
etiotto and others added 9 commits May 7, 2024 21:27
@etiotto
Modify pass pipeline to allow lowering tt.load to 2DBlockRead
6e3cc2c 
Signed-off-by: Tiotto, Ettore <[email protected]>
@etiotto
Fix some test failures
2e5c328 
Signed-off-by: Tiotto, Ettore <[email protected]>
@etiotto
Fix build
42e360e 
Signed-off-by: Tiotto, Ettore <[email protected]>
@whitneywhtsang
Resolves test_block_pointer.py failures
35c1a4c
@whitneywhtsang
Fix test_trans_reshape failure
5d48ffc
@etiotto
Merge remote-tracking branch 'origin/llvm-target' into etiotto.tt.loa…
15d9d19 
…d.to.2DBlockRead
@etiotto
Enable a few more tests in test_flash_attention and relax precision s…
3392d95 
…lightly

Signed-off-by: Tiotto, Ettore <[email protected]>
@etiotto
Remove naked traces
7ac0858 
Signed-off-by: Tiotto, Ettore <[email protected]>
@etiotto
Remove unnecessary code changes
0731031 
Signed-off-by: Tiotto, Ettore <[email protected]>
@etiotto etiotto marked this pull request as ready for review May 9, 2024 16:37
etiotto
etiotto commented May 9, 2024
View reviewed changes
third_party/intel/backend/compiler.py
@@ -69,20 +69,6 @@ class XPUBackend(BaseBackend):
# Experimental pass pipeline for kernels using block pointers.
class Experimental:

@staticmethod
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Note: Now the experimental pass has the same implementation as the default path for make_ttir.

lib/Dialect/TritonGPU/Transforms/Coalesce.cpp Show resolved Hide resolved
@etiotto
Copy link
Contributor Author

etiotto commented May 9, 2024

Now, with this PR, the code generated for 09-experimental-block-pointer.py shows that the 2D block read operation is used:

TRITON_DISABLE_LINE_INFO=1  MLIR_ENABLE_DUMP=1  python 09-experimental-block-pointer.py >! log 2>&1 &  
grep "BlockRead" log| wc                                                                                                      ✔  10361  17:46:33 
    712   22416  158600

@etiotto
Fix merge
f83efec 
Signed-off-by: Tiotto, Ettore <[email protected]>
@etiotto etiotto mentioned this pull request May 9, 2024
Closed
@etiotto
Copy link
Contributor Author

etiotto commented May 10, 2024 •
edited
Loading

@whitneywhtsang @chengjunlu I'd like to land this PR asap. If there are concerns with relaxing the precision of the tests for flash attention we can open a issue defect to track that.

Opened #1085 to track the issue.

@etiotto etiotto mentioned this pull request May 10, 2024
Closed
@etiotto etiotto merged commit 1e4fb64 into llvm-target May 10, 2024
2 checks passed
@etiotto etiotto deleted the etiotto.tt.load.to.2DBlockRead branch May 10, 2024 15:26
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@whitneywhtsang whitneywhtsang whitneywhtsang approved these changes

@chengjunlu chengjunlu Awaiting requested review from chengjunlu

Assignees

@etiotto etiotto

Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.

Add passes to enable lowering tt.load using block ptrs to 2DBlockRead
2 participants
@etiotto @whitneywhtsang