Adding support for scheduling the epilogue computation when smem_epilogue parameter is true #3581
Conversation
This file contains 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
rdspring1
reviewed
Dec 12, 2024
There was a problem hiding this comment.
Here is my transformation helper. It can be used in scheduleMmaResult. Just trying to avoid some merge conflicts.
void HopperMultipleMatmulScheduler::transformLikeMmaOutput(
TensorView* tv,
bool is_mma_result) {
// TODO Add constraints
auto apply_k_dim_offset = [is_mma_result](int64_t idx) constexpr {
return (is_mma_result) ? idx - 1 : idx;
};
// Original: [..., Mo, No, Mi, Ni]
tv->split(apply_k_dim_offset(-2), getM(params_->mma_macro));
tv->split(apply_k_dim_offset(-1), getN(params_->mma_macro));
// After Split: [..., Mo, No, Mio, Mii, Nio, Nii]
tv->reorder({{apply_k_dim_offset(-3), apply_k_dim_offset(-2)}});
// After Reorder: [..., Mo, No, Mio, Nio, Mii, Nii]
tv->merge(apply_k_dim_offset(-4));
// After Merge: [..., Mo, No, Mio * Nio, Mii, Nii]
tv->axis(apply_k_dim_offset(-3))->parallelize(ParallelType::TIDy);
// After Parallelize: [..., Mo, No, Mio * Nio (TIDy), Mii, Nii]
}
protonu
force-pushed
the
pbasu_mma_epilogue_hopper_no_smem_epilogue
branch
from
94f39a1 to
be5f18b
Compare
protonu
force-pushed
the
pbasu_mma_epilogue_hopper_smem_epilogue
branch
from
b9df449 to
aa86b8b
Compare
|
!build |
protonu
force-pushed
the
pbasu_mma_epilogue_hopper_no_smem_epilogue
branch
from
895df0f to
09cb407
Compare
protonu
force-pushed
the
pbasu_mma_epilogue_hopper_smem_epilogue
branch
2 times, most recently
from
c0cfa4e to
af126a1
Compare
Base automatically changed from
pbasu_mma_epilogue_hopper_no_smem_epilogue
to
main
December 14, 2024 14:46
protonu
force-pushed
the
pbasu_mma_epilogue_hopper_smem_epilogue
branch
from
af126a1 to
8066a54
Compare
protonu
force-pushed
the
pbasu_mma_epilogue_hopper_smem_epilogue
branch
from
8066a54 to
2c19191
Compare
|
!test |
jacobhinkle
reviewed
Dec 16, 2024
Comment on lines
524
to
539
| blockTileTensors({d}); | ||
| parallelizeBlocks({d}); | ||
| transformLikeMmaOutput(d, /*is_mma_result=*/false); | ||
|
|
||
| // Apply mma common transformation | ||
| for (auto tv : tvs_to_schedule) { | ||
| transformLikeMmaOutput(tv, /*is_mma_result=*/false); | ||
| } | ||
| scheduler_utils::BoundedDirectionalTransformPropagator::backward( | ||
| d, | ||
| -1, | ||
| {dc}, | ||
| scheduler_utils::BoundedDirectionalTransformPropagator::Options() | ||
| .propagateParallelType() | ||
| .propagateToBoundary()); | ||
|
|
||
| // Schedule register cache; Output from epilogue | ||
| { | ||
| auto s = mma_utils::MmaSwizzler::scheduleMmaOutputAllocation( | ||
| dc->getLoopDomain()); | ||
| dc->setLoopDomain(s.as<IterDomain*>()); | ||
| dc->setAllocationDomain(s.as<IterDomain*>(), true); | ||
| } | ||
| auto s = mma_utils::MmaSwizzler::scheduleMmaOutputAllocation( | ||
| dc->getLoopDomain()); | ||
| dc->setLoopDomain(s.as<IterDomain*>()); | ||
| dc->setAllocationDomain(s.as<IterDomain*>(), true); |
There was a problem hiding this comment.
Why propagate from d to dc then schedule dc directly like this instead of just directly scheduling dc? What does the propagation provide us? Could use this in a code comment.
There was a problem hiding this comment.
Please take a look at the comment.
jacobhinkle
reviewed
Dec 16, 2024
jacobhinkle
reviewed
Dec 16, 2024
jacobhinkle
approved these changes
Dec 16, 2024
rdspring1
reviewed
Dec 16, 2024
For the test |
|
!build |
|
!test |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Refactoring some code and adding some support for smem_epilogue
TODO: add support for smem_epilogue when the output of the mma op is not cast down to half precision.