[wave2water] E2E execution of matmul kernel via water middle-end #672
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: tyb0807 <[email protected]>
ftynse
reviewed
Jan 5, 2026
Comment on lines
+167
to
+183
|
|
||
| @tkw.wave(constraints) | ||
| def matmul( | ||
| a: tkl.Memory[M, K, ADDRESS_SPACE, dtype], | ||
| b: tkl.Memory[N, K, ADDRESS_SPACE, dtype], | ||
| c: tkl.Memory[M, N, GLOBAL_ADDRESS_SPACE, tkl.f32], | ||
| ): | ||
| c_reg = tkl.Register[M, N, tkl.f32](0.0) | ||
|
|
||
| @tkw.iterate(K, init_args=[c_reg]) | ||
| def repeat(acc: tkl.Register[M, N, tkl.f32]) -> tkl.Register[M, N, tkl.f32]: | ||
| a_reg = tkw.read(a) | ||
| b_reg = tkw.read(b) | ||
| acc = tkw.mma(a_reg, b_reg, acc) | ||
| return acc | ||
|
|
||
| tkw.write(repeat, c) |
Contributor
There was a problem hiding this comment.
Can we use one from templates?
| # Apply Water PassManager lowering | ||
| lowered_mlir = apply_water_middle_end_passes(wave_dialect_mlir) | ||
|
|
||
| print(lowered_mlir) |
Contributor
There was a problem hiding this comment.
Let's have FileCheck comments here so it doesn't look like forgotten debug output.
| compile_to_mlir=True, | ||
| location_capture_config=LocationCaptureConfig(level=LocationCaptureLevel.NONE), | ||
| enforce_locations=False, | ||
| print_mlir=True, |
| b_tensor = device_randn(n, k, dtype=torch.float16) # Note: transposed in matmul | ||
| c_tensor = device_zeros(m, n, dtype=torch.float32) | ||
|
|
||
| # Expected result (CPU computation) |
Contributor
There was a problem hiding this comment.
The code above creates tensors on device, why is this called CPU computation?
| c_tensor = device_zeros(m, n, dtype=torch.float32) | ||
|
|
||
| # Expected result (CPU computation) | ||
| expected = torch.matmul(a_tensor.float(), b_tensor.T.float()) |
Contributor
There was a problem hiding this comment.
It is a bad idea to compute expected values with a higher precision than actual values.
|
|
||
| compiled_e2e(a_tensor, b_tensor, c_tensor) | ||
|
|
||
| assert_close(c_tensor, expected, rtol=1e-3, atol=1e-3) |
Contributor
There was a problem hiding this comment.
1e-3 looks a bit too lax, do we really need it?
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.
Fixes #600. Requires #667.