[Integrate] Drop LLVM revert of "Remove matmul_transpose variants" by hanhanW · Pull Request #21344 · iree-org/iree

hanhanW · 2025-07-11T17:35:15Z

rengolin · 2025-07-17T21:47:54Z

Awesome @hanhanW, thanks!

All errors seem to be about this particular problem:

external/llvm-project/llvm/include/llvm/Support/Casting.h:64:65: error: cannot initialize a parameter of type 'mlir::Operation *' with an rvalue of type 'const mlir::linalg::LinalgOp *'
  static inline bool doit(const From &Val) { return To::classof(&Val); }
                                                                ^~~~
external/llvm-project/llvm/include/llvm/Support/Casting.h:81:32: note: in instantiation of member function 'llvm::isa_impl<mlir::linalg::BatchMatmulTransposeBOp, mlir::linalg::LinalgOp>::doit' requested here
    return isa_impl<To, From>::doit(Val);
                               ^
external/llvm-project/llvm/include/llvm/Support/Casting.h:137:37: note: in instantiation of member function 'llvm::isa_impl_cl<mlir::linalg::BatchMatmulTransposeBOp, const mlir::linalg::LinalgOp>::doit' requested here
    return isa_impl_cl<To, FromTy>::doit(Val);
                                    ^
external/llvm-project/llvm/include/llvm/Support/Casting.h:257:58: note: in instantiation of member function 'llvm::isa_impl_wrap<mlir::linalg::BatchMatmulTransposeBOp, const mlir::linalg::LinalgOp, const mlir::linalg::LinalgOp>::doit' requested here
        typename simplify_type<const From>::SimpleType>::doit(f);
                                                         ^
external/llvm-project/llvm/include/llvm/Support/Casting.h:549:36: note: in instantiation of member function 'llvm::CastIsPossible<mlir::linalg::BatchMatmulTransposeBOp, const mlir::linalg::LinalgOp>::isPossible' requested here
  return CastInfo<To, const From>::isPossible(Val);
                                   ^
external/llvm-project/llvm/include/llvm/Support/Casting.h:554:29: note: in instantiation of function template specialization 'llvm::isa<mlir::linalg::BatchMatmulTransposeBOp, mlir::linalg::LinalgOp>' requested here
  return isa<First>(Val) || isa<Second, Rest...>(Val);
                            ^
external/llvm-project/llvm/include/llvm/Support/Casting.h:554:29: note: in instantiation of function template specialization 'llvm::isa<mlir::linalg::BatchMatmulOp, mlir::linalg::BatchMatmulTransposeBOp, mlir::linalg::LinalgOp>' requested here
external/llvm-project/llvm/include/llvm/Support/Casting.h:554:29: note: in instantiation of function template specialization 'llvm::isa<mlir::linalg::MatmulTransposeBOp, mlir::linalg::BatchMatmulOp, mlir::linalg::BatchMatmulTransposeBOp, mlir::linalg::LinalgOp>' requested here
external/llvm-project/llvm/include/llvm/Support/Casting.h:678:10: note: in instantiation of function template specialization 'llvm::isa<mlir::linalg::MatmulOp, mlir::linalg::MatmulTransposeBOp, mlir::linalg::BatchMatmulOp, mlir::linalg::BatchMatmulTransposeBOp, mlir::linalg::LinalgOp>' requested here
  return isa<X...>(Val);
         ^
external/llvm-project/llvm/include/llvm/Support/Casting.h:683:10: note: in instantiation of function template specialization 'llvm::isa_and_present<mlir::linalg::MatmulOp, mlir::linalg::MatmulTransposeBOp, mlir::linalg::BatchMatmulOp, mlir::linalg::BatchMatmulTransposeBOp, mlir::linalg::LinalgOp>' requested here
  return isa_and_present<X...>(Val);
         ^
compiler/src/iree/compiler/Preprocessing/Common/GeneralizeLinalgMatMul.cpp:33:11: note: in instantiation of function template specialization 'llvm::isa_and_nonnull<mlir::linalg::MatmulOp, mlir::linalg::MatmulTransposeBOp, mlir::linalg::BatchMatmulOp, mlir::linalg::BatchMatmulTransposeBOp, mlir::linalg::LinalgOp>' requested here
      if (isa_and_nonnull<linalg::MatmulOp, linalg::MatmulTransposeBOp,
          ^
external/llvm-project/mlir/include/mlir/Dialect/Linalg/IR/Linalg.h:264:34: note: passing argument to parameter 'op' here
  static bool classof(Operation *op);
                                 ^
2 errors generated.

I have fixed an issue with the classof yesterday and this is the commit that you do not seem to have it here.

Can you cherry pick again on the latest commit of that PR and see if the issues go away?

rengolin · 2025-07-17T22:44:32Z

@kuhar this is what you meant on the llvm pr, I guess.

hanhanW · 2025-07-17T22:52:57Z

Thanks! I'll also rebase to the main branch, which has fresh LLVM commits.

rengolin · 2025-07-19T12:29:46Z

I have rebase my pr on a recent branch, so should be easy to cherry pick.

kuhar · 2025-08-15T15:22:39Z

This needs to be rebased

…#147961)"" Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>

Signed-off-by: Abhishek Varma <abhvarma@amd.com>

Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>

…71716e88bd4af8f22c Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>

Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>

Yu-Zhewen · 2025-08-15T16:32:45Z

This needs to be rebased

Done

hanhanW · 2025-08-15T19:44:25Z

+  Operation *op = bmmOp.getOperation();
+  if (!IREE::LinalgExt::isPureBatchMatmul(op)) {


I think this can just be !IREE::LinalgExt::isPureBatchMatmul(bmmOp)/

hanhanW · 2025-08-15T19:44:31Z

+  Operation *op = matmulOp.getOperation();
+  if (!IREE::LinalgExt::isPureMatmul(op)) {


hanhanW · 2025-08-15T19:50:03Z

+bool isPureMatmul(Operation *op) {
+  return dyn_cast_or_null<linalg::MatmulOp>(op) &&
+         linalg::MatmulOp::isDefaultIndexingMaps(op->getAttr("indexing_maps"));
+}


I'd do it this way, which avoids magic string.

auto matmulOp = dyn_cast_or_null<linalg::MatmulOp>(op); return matmulOp && linalg::MatmulOp::isDefaultIndexingMaps(matmulOp.getIndexingMaps());

hanhanW · 2025-08-15T19:56:47Z

I think we don't need to check maps for this file, because it is a test for matching. We should just check match_status.

hanhanW · 2025-08-15T21:22:38Z

+bool isPureBatchMatmul(Operation *op) {
+  return dyn_cast_or_null<linalg::BatchMatmulOp>(op) &&
+         linalg::BatchMatmulOp::isDefaultIndexingMaps(
+             op->getAttr("indexing_maps"));
+}


It needs to be updated as well. Sorry I forgot to point this out

Thank. Done

hanhanW · 2025-08-15T21:26:08Z

We can do it in a follow-up, because I really want to drop the revert first. Functionality-wise, it is okay.

The better implementation in my mind is that we bail out if the op is not a pure matmul/batch_matmul or with transpose variants.

Then we look at the indexing maps to create transpose op + matmul/batch_matmul ops.

Oh, I misread the code. never mind.

Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>

hanhanW · 2025-08-15T21:30:18Z

I just realized that I can't approve the PR because I created the PR. More eyes are better, @kuhar can you review and stamp?

kuhar

GeneralizeNamedOps may also need to be updated: #21716

cc: @newling

…21344) Issue: #21349 --------- Signed-off-by: Yu-Zhewen <zhewenyu@amd.com> Signed-off-by: Abhishek Varma <abhvarma@amd.com> Co-authored-by: Yu-Zhewen <zhewenyu@amd.com> Co-authored-by: Abhishek Varma <abhvarma@amd.com> Co-authored-by: Ean Garvey <ean.garvey@amd.com>

qedawkins mentioned this pull request Jul 11, 2025

Tracking issue for #21344 tasks (named matmul op removal) #21349

Closed

6 tasks

rengolin mentioned this pull request Jul 17, 2025

[MLIR][Linalg] Remove matmul_transpose variants llvm/llvm-project#147961

Merged

hanhanW force-pushed the integrates/llvm-cherry-pick-147961 branch from d68c020 to 32640f8 Compare July 17, 2025 22:58

hanhanW closed this Jul 21, 2025

hanhanW force-pushed the integrates/llvm-cherry-pick-147961 branch from 32640f8 to 28c2301 Compare July 21, 2025 18:15

hanhanW reopened this Jul 21, 2025

Abhishek-Varma requested review from MaheshRavishankar, antiagainst, pashu123 and qedawkins as code owners July 22, 2025 08:33

Abhishek-Varma requested a review from Yu-Zhewen July 22, 2025 08:34

Abhishek-Varma force-pushed the integrates/llvm-cherry-pick-147961 branch 3 times, most recently from 7120f2a to 42d2e6e Compare July 22, 2025 09:39

Yu-Zhewen force-pushed the integrates/llvm-cherry-pick-147961 branch 5 times, most recently from 6aa0f5f to 061c693 Compare July 22, 2025 20:07

hanhanW force-pushed the integrates/llvm-cherry-pick-147961 branch 2 times, most recently from 538cad9 to ef13595 Compare July 23, 2025 17:06

Yu-Zhewen force-pushed the integrates/llvm-cherry-pick-147961 branch from ef13595 to a825464 Compare July 23, 2025 19:34

Yu-Zhewen requested review from Groverkss, IanWood1 and kuhar as code owners July 23, 2025 19:34

Yu-Zhewen force-pushed the integrates/llvm-cherry-pick-147961 branch from a825464 to 8e3828a Compare July 23, 2025 19:36

Yu-Zhewen force-pushed the integrates/llvm-cherry-pick-147961 branch from 20b06b2 to b060031 Compare August 14, 2025 17:24

Yu-Zhewen changed the title ~~DNS - Cherry-pick llvm "Remove matmul_transpose variants"~~ [Integrate] Drop LLVM revert of "Remove matmul_transpose variants" Aug 14, 2025

Yu-Zhewen mentioned this pull request Aug 15, 2025

Update punet_fp8.mlir link #21698

Closed

Yu-Zhewen and others added 12 commits August 15, 2025 16:31

Revert "Revert "[MLIR][Linalg] Remove matmul_transpose variants (llvm…

5f789bb

…#147961)"" Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>

[CI] Fix build issues

58a52bc

Signed-off-by: Abhishek Varma <abhvarma@amd.com>

fix PropagateLinalgTranspose

8b9e689

Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>

fix e2e matmul

e9c9ed4

Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>

lit tests

62dfe88

Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>

more fixes to plain matmul check

b6cf08f

Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>

update transpose_matmul test

cda6edd

Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>

use tag dispatch

099cf54

Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>

getOperation

e568400

Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>

add utility functions to check matmul without transpose

6593782

Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>

Update punet MLIR with artifacts from shark-ai@3629692abec38d13fbe0bd…

f07fc80

…71716e88bd4af8f22c Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>

use manual updated mlir file for punet fp8

efe5104

Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>

Yu-Zhewen force-pushed the integrates/llvm-cherry-pick-147961 branch from b060031 to efe5104 Compare August 15, 2025 16:32

hanhanW commented Aug 15, 2025

View reviewed changes

resolving review comments

6f33b05

Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>

Yu-Zhewen force-pushed the integrates/llvm-cherry-pick-147961 branch from 9e28846 to 6f33b05 Compare August 15, 2025 21:29

kuhar approved these changes Aug 15, 2025

View reviewed changes

kuhar reviewed Aug 16, 2025

View reviewed changes

Yu-Zhewen merged commit 3df06b7 into main Aug 16, 2025
48 checks passed

Yu-Zhewen deleted the integrates/llvm-cherry-pick-147961 branch August 16, 2025 08:35

newling mentioned this pull request Aug 18, 2025

[Codegen] Add matmul and batched matmul to list of ops to generalize #21720

Merged

		Operation *op = bmmOp.getOperation();
		if (!IREE::LinalgExt::isPureBatchMatmul(op)) {

		Operation *op = matmulOp.getOperation();
		if (!IREE::LinalgExt::isPureMatmul(op)) {

Conversation

hanhanW commented Jul 11, 2025 • edited by kuhar Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

rengolin commented Jul 17, 2025

Uh oh!

rengolin commented Jul 17, 2025

Uh oh!

hanhanW commented Jul 17, 2025

Uh oh!

rengolin commented Jul 19, 2025

Uh oh!

kuhar commented Aug 15, 2025

Uh oh!

Yu-Zhewen commented Aug 15, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

hanhanW commented Aug 15, 2025

Uh oh!

kuhar left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

hanhanW commented Jul 11, 2025 •

edited by kuhar

Loading