Fix linalg.matmul_transpose_b for big tiles
#410
Merged
Conversation
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
Signed-off-by: dchigarev <[email protected]>
dchigarev
commented
Nov 15, 2024
Comment on lines
-709
to
-711
| if (transpose) { | ||
| std::swap(newRowOffs, newColOffs); | ||
| } |
Contributor
Author
There was a problem hiding this comment.
placing std::swap inside a nested for-loop was a bad idea since it swaps the values each iteration producing non-sense offsets at the end
Signed-off-by: dchigarev <[email protected]>
dchigarev
commented
Nov 15, 2024
| %subview_1 = memref.subview %arg0[%arg3, 0] [32, 1024] [1, 1] : memref<1024x1024xf16> to memref<32x1024xf16, strided<[1024, 1], offset: ?>> | ||
| %subview_2 = memref.subview %arg1[%arg4, 0] [32, 1024] [1, 1] : memref<1024x1024xf16> to memref<32x1024xf16, strided<[1024, 1], offset: ?>> | ||
| linalg.matmul_transpose_b ins(%subview_1, %subview_2 : memref<32x1024xf16, strided<[1024, 1], offset: ?>>, memref<32x1024xf16, strided<[1024, 1], offset: ?>>) outs(%subview_0 : memref<32x32xf16, strided<[1024, 1], offset: ?>>) | ||
| scf.parallel (%arg3, %arg4) = (%c0, %c0) to (%c1024, %c1024) step (%c16, %c64) { |
Contributor
Author
There was a problem hiding this comment.
increased the tile size for Y axis to test the problematic case
dchigarev
commented
Nov 15, 2024
Comment on lines
45
to
+48
| // CHECK: %[[tB:.+]] = xegpu.update_nd_offset %[[rootB]], [%c0, %c0] | ||
| // CHECK: %[[tB1:.+]] = xegpu.update_nd_offset %[[rootB]], [%c16, %c0] | ||
| // CHECK: %[[tB2:.+]] = xegpu.update_nd_offset %[[rootB]], [%c32, %c0] | ||
| // CHECK: %[[tB3:.+]] = xegpu.update_nd_offset %[[rootB]], [%c48, %c0] |
Contributor
Author
There was a problem hiding this comment.
it used to do something like:
xegpu.update_nd_offset %[[rootB]], [%c0, %c0]
xegpu.update_nd_offset %[[rootB]], [%c16, %c0]
xegpu.update_nd_offset %[[rootB]], [%c32, %c16]
xegpu.update_nd_offset %[[rootB]], [%c16, %c32]
AndreyPavlenko
approved these changes
Nov 15, 2024
kurapov-peter
approved these changes
Nov 16, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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.
It was discovered that
linalg.matmul_transpose_bis lowered to xegpu incorrectly in the case of large tiles. The issue is caused by astd::swapplaced inside a nested for-loop, which, instead of swappingrowOffsandcolOffsonly once, performs the swap in every iteration, resulting in incorrect offsets.