Revert "Revert " Reapply "[Layouts] Propagate layouts into conditionals (#5610)" (#5725)""

[anmyachev]

This reverts commit 942cf94.

[anmyachev]

New lit test failed due to unimplemented DotOperandEncodingAttr::getSizePerThread:

intel-xpu-backend-for-triton/lib/Dialect/TritonGPU/IR/Dialect.cpp

Lines 2509 to 2511 in 8b95bc5

	llvm::report_fatal_error(
	"DotOperandEncodingAttr non-NvidiaMmaEncodingAttr parent not "
	"supported yet");

It gets into this code branch because of an Intel-specific change in the general code:

intel-xpu-backend-for-triton/lib/Dialect/TritonGPU/IR/LinearLayoutConversions.cpp

Lines 878 to 884 in 8b95bc5

	// The triton generate the homogeneous kernel run on every thread.
	// The multiple threads of the parent layout which are distributed on the
	// sliced dim are squeezed to hold the same value of tensor redundantly. The
	// multiple values of sizePerThreads[dim] of the parent are reduced to the
	// only one. We need to fix up the number of registers in case we just removed
	// all zeros aggressively.
	auto sizePerThreads = triton::gpu::getSizePerThread(getParent());

@whitneywhtsang do you have any idea how to fix this better?

[whitneywhtsang]

@whitneywhtsang do you have any idea how to fix this better?

From the error message, I would think the fix should be to add the missing DotOperandEncodingAttr::getSizePerThread.
@chengjunlu attempted to upstream the change in general code here: triton-lang/triton#5328, which is considered as not needed for long term. You can see the motivation of the change there.
@chengjunlu Any suggestions?

[chengjunlu]

I think the better we need to do two things.

1. To add a new code to the getSizePerThread for the dotOp layout with block layout as parent. (As same as the new code added in getElemsPerThread https://github.com/triton-lang/triton/blob/04669ef6db83ce6448b24869c7df38f992eaeeef/lib/Dialect/TritonGPU/IR/Dialect.cpp#L947)
2. We need to PR the first change to upstream and then rebase/merge the PR [BACKEND] Support the conversion of nested layout "#slice->#dot->#third_party_mma" to LinearLayout triton-lang/triton#5328

This is the clean solution as no extra attribute interface required to be added to legacy layout.

[anmyachev]

Triton PR: triton-lang/triton#5863

[chengjunlu]

LGTM.

[whitneywhtsang]

New lit test failed due to unimplemented DotOperandEncodingAttr::getSizePerThread:

intel-xpu-backend-for-triton/lib/Dialect/TritonGPU/IR/Dialect.cpp

Lines 2509 to 2511 in 8b95bc5

	llvm::report_fatal_error(
	"DotOperandEncodingAttr non-NvidiaMmaEncodingAttr parent not "
	"supported yet");

It gets into this code branch because of an Intel-specific change in the general code:

intel-xpu-backend-for-triton/lib/Dialect/TritonGPU/IR/LinearLayoutConversions.cpp

Lines 878 to 884 in 8b95bc5

	// The triton generate the homogeneous kernel run on every thread.
	// The multiple threads of the parent layout which are distributed on the
	// sliced dim are squeezed to hold the same value of tensor redundantly. The
	// multiple values of sizePerThreads[dim] of the parent are reduced to the
	// only one. We need to fix up the number of registers in case we just removed
	// all zeros aggressively.
	auto sizePerThreads = triton::gpu::getSizePerThread(getParent());

@whitneywhtsang do you have any idea how to fix this better?

Looks like the Intel-specific change is no longer needed after merging 61b5674 in #3391.