[mlir] Move vector.{to_elements,from_elements} unrolling to `VectorUnroll.cpp` #159118

amd-eochoalo · 2025-09-16T16:01:59Z

This PR moves the patterns that unroll vector.to_elements and vector.from_elements into the file with other vector unrolling operations. This PR also adds these unrolling patterns into the populateVectorUnrollPatterns. And renames populateVectorToElementsLoweringPatterns populateVectorFromElementsLoweringPatterns to populateVectorToElementsUnrollPatterns populateVectorFromElementsUnrollPatterns.

mlir/lib/Dialect/SPIRV/Transforms/SPIRVConversion.cpp

llvmbot · 2025-09-16T16:15:07Z

@llvm/pr-subscribers-mlir-vector

@llvm/pr-subscribers-mlir-spirv

Author: Erick Ochoa Lopez (amd-eochoalo)

Changes

This patch adds the unrolling patterns for vector.to_elements and vector.from_elements to the action unrollVectorsInFuncBodies. This affects the passes test-convert-to-spriv (when running with option run-vector-unrolling=true ) and test-spirv-vector-unrolling.

Full diff: https://github.com/llvm/llvm-project/pull/159118.diff

2 Files Affected:

(modified) mlir/lib/Dialect/SPIRV/Transforms/SPIRVConversion.cpp (+2)
(modified) mlir/test/Conversion/ConvertToSPIRV/vector-unroll.mlir (+44)

diff --git a/mlir/lib/Dialect/SPIRV/Transforms/SPIRVConversion.cpp b/mlir/lib/Dialect/SPIRV/Transforms/SPIRVConversion.cpp
index 49f4ce8de7c76..98e294b40456f 100644
--- a/mlir/lib/Dialect/SPIRV/Transforms/SPIRVConversion.cpp
+++ b/mlir/lib/Dialect/SPIRV/Transforms/SPIRVConversion.cpp
@@ -1495,6 +1495,8 @@ LogicalResult mlir::spirv::unrollVectorsInFuncBodies(Operation *op) {
     RewritePatternSet patterns(context);
     auto options = vector::UnrollVectorOptions().setNativeShapeFn(
         [](auto op) { return mlir::spirv::getNativeVectorShape(op); });
+    vector::populateVectorFromElementsLoweringPatterns(patterns);
+    vector::populateVectorToElementsLoweringPatterns(patterns);
     populateVectorUnrollPatterns(patterns, options);
     if (failed(applyPatternsGreedily(op, std::move(patterns))))
       return failure();
diff --git a/mlir/test/Conversion/ConvertToSPIRV/vector-unroll.mlir b/mlir/test/Conversion/ConvertToSPIRV/vector-unroll.mlir
index c85f4334ff2e5..0957f67690b97 100644
--- a/mlir/test/Conversion/ConvertToSPIRV/vector-unroll.mlir
+++ b/mlir/test/Conversion/ConvertToSPIRV/vector-unroll.mlir
@@ -96,3 +96,47 @@ func.func @transpose(%arg0 : vector<2x3xi32>) -> (vector<3x2xi32>) {
   %0 = vector.transpose %arg0, [1, 0] : vector<2x3xi32> to vector<3x2xi32>
   return %0 : vector<3x2xi32>
 }
+
+// -----
+
+// In order to verify that the pattern is applied,
+// we need to make sure that the the 2d vector does not
+// come from the parameters. Otherwise, the pattern
+// in unrollVectorsInSignatures which splits the 2d vector
+// parameter will take precedent. Similarly, let's avoid
+// returning a vector as another pattern would take precendence.
+
+// CHECK-LABEL: @unroll_to_elements_2d
+func.func @unroll_to_elements_2d() -> (f32, f32, f32, f32) {
+  %1 = "test.op"() : () -> (vector<2x2xf32>)
+  // CHECK: %[[VEC2D:.+]] = "test.op"
+  // CHECK: %[[VEC0:.+]] = vector.extract %[[VEC2D]][0] : vector<2xf32> from vector<2x2xf32>
+  // CHECK: %[[VEC1:.+]] = vector.extract %[[VEC2D]][1] : vector<2xf32> from vector<2x2xf32>
+  // CHECK: %[[RES0:.+]]:2 = vector.to_elements %[[VEC0]]
+  // CHECK: %[[RES1:.+]]:2 = vector.to_elements %[[VEC1]]
+  %2:4 = vector.to_elements %1 : vector<2x2xf32>
+  return %2#0, %2#1, %2#2, %2#3 : f32, f32, f32, f32
+}
+
+// -----
+
+// In order to verify that the pattern is applied,
+// we need to make sure that the the 2d vector is used
+// by an operation and that extracts are not folded away.
+// In other words we can't use "test.op" nor return the
+// value `%0 = vector.from_elements`
+
+// CHECK-LABEL: @unroll_from_elements_2d
+// CHECK-SAME: (%[[ARG0:.+]]: f32, %[[ARG1:.+]]: f32, %[[ARG2:.+]]: f32, %[[ARG3:.+]]: f32)
+func.func @unroll_from_elements_2d(%arg0: f32, %arg1: f32, %arg2: f32, %arg3: f32) -> (vector<2x2xf32>) {
+  // CHECK: %[[VEC0:.+]] = vector.from_elements %[[ARG0]], %[[ARG1]] : vector<2xf32>
+  // CHECK: %[[VEC1:.+]] = vector.from_elements %[[ARG2]], %[[ARG3]] : vector<2xf32>
+  %0 = vector.from_elements %arg0, %arg1, %arg2, %arg3 : vector<2x2xf32>
+
+  // CHECK: %[[RES0:.+]] = arith.addf %[[VEC0]], %[[VEC0]]
+  // CHECK: %[[RES1:.+]] = arith.addf %[[VEC1]], %[[VEC1]]
+  %1 = arith.addf %0, %0 : vector<2x2xf32>
+
+  // return %[[RES0]], %%[[RES1]] : vector<2xf32>, vector<2xf32>
+  return %1 : vector<2x2xf32>
+}

llvmbot · 2025-09-16T16:15:07Z

@llvm/pr-subscribers-mlir

Author: Erick Ochoa Lopez (amd-eochoalo)

Changes

This patch adds the unrolling patterns for vector.to_elements and vector.from_elements to the action unrollVectorsInFuncBodies. This affects the passes test-convert-to-spriv (when running with option run-vector-unrolling=true ) and test-spirv-vector-unrolling.

Full diff: https://github.com/llvm/llvm-project/pull/159118.diff

2 Files Affected:

(modified) mlir/lib/Dialect/SPIRV/Transforms/SPIRVConversion.cpp (+2)
(modified) mlir/test/Conversion/ConvertToSPIRV/vector-unroll.mlir (+44)

diff --git a/mlir/lib/Dialect/SPIRV/Transforms/SPIRVConversion.cpp b/mlir/lib/Dialect/SPIRV/Transforms/SPIRVConversion.cpp
index 49f4ce8de7c76..98e294b40456f 100644
--- a/mlir/lib/Dialect/SPIRV/Transforms/SPIRVConversion.cpp
+++ b/mlir/lib/Dialect/SPIRV/Transforms/SPIRVConversion.cpp
@@ -1495,6 +1495,8 @@ LogicalResult mlir::spirv::unrollVectorsInFuncBodies(Operation *op) {
     RewritePatternSet patterns(context);
     auto options = vector::UnrollVectorOptions().setNativeShapeFn(
         [](auto op) { return mlir::spirv::getNativeVectorShape(op); });
+    vector::populateVectorFromElementsLoweringPatterns(patterns);
+    vector::populateVectorToElementsLoweringPatterns(patterns);
     populateVectorUnrollPatterns(patterns, options);
     if (failed(applyPatternsGreedily(op, std::move(patterns))))
       return failure();
diff --git a/mlir/test/Conversion/ConvertToSPIRV/vector-unroll.mlir b/mlir/test/Conversion/ConvertToSPIRV/vector-unroll.mlir
index c85f4334ff2e5..0957f67690b97 100644
--- a/mlir/test/Conversion/ConvertToSPIRV/vector-unroll.mlir
+++ b/mlir/test/Conversion/ConvertToSPIRV/vector-unroll.mlir
@@ -96,3 +96,47 @@ func.func @transpose(%arg0 : vector<2x3xi32>) -> (vector<3x2xi32>) {
   %0 = vector.transpose %arg0, [1, 0] : vector<2x3xi32> to vector<3x2xi32>
   return %0 : vector<3x2xi32>
 }
+
+// -----
+
+// In order to verify that the pattern is applied,
+// we need to make sure that the the 2d vector does not
+// come from the parameters. Otherwise, the pattern
+// in unrollVectorsInSignatures which splits the 2d vector
+// parameter will take precedent. Similarly, let's avoid
+// returning a vector as another pattern would take precendence.
+
+// CHECK-LABEL: @unroll_to_elements_2d
+func.func @unroll_to_elements_2d() -> (f32, f32, f32, f32) {
+  %1 = "test.op"() : () -> (vector<2x2xf32>)
+  // CHECK: %[[VEC2D:.+]] = "test.op"
+  // CHECK: %[[VEC0:.+]] = vector.extract %[[VEC2D]][0] : vector<2xf32> from vector<2x2xf32>
+  // CHECK: %[[VEC1:.+]] = vector.extract %[[VEC2D]][1] : vector<2xf32> from vector<2x2xf32>
+  // CHECK: %[[RES0:.+]]:2 = vector.to_elements %[[VEC0]]
+  // CHECK: %[[RES1:.+]]:2 = vector.to_elements %[[VEC1]]
+  %2:4 = vector.to_elements %1 : vector<2x2xf32>
+  return %2#0, %2#1, %2#2, %2#3 : f32, f32, f32, f32
+}
+
+// -----
+
+// In order to verify that the pattern is applied,
+// we need to make sure that the the 2d vector is used
+// by an operation and that extracts are not folded away.
+// In other words we can't use "test.op" nor return the
+// value `%0 = vector.from_elements`
+
+// CHECK-LABEL: @unroll_from_elements_2d
+// CHECK-SAME: (%[[ARG0:.+]]: f32, %[[ARG1:.+]]: f32, %[[ARG2:.+]]: f32, %[[ARG3:.+]]: f32)
+func.func @unroll_from_elements_2d(%arg0: f32, %arg1: f32, %arg2: f32, %arg3: f32) -> (vector<2x2xf32>) {
+  // CHECK: %[[VEC0:.+]] = vector.from_elements %[[ARG0]], %[[ARG1]] : vector<2xf32>
+  // CHECK: %[[VEC1:.+]] = vector.from_elements %[[ARG2]], %[[ARG3]] : vector<2xf32>
+  %0 = vector.from_elements %arg0, %arg1, %arg2, %arg3 : vector<2x2xf32>
+
+  // CHECK: %[[RES0:.+]] = arith.addf %[[VEC0]], %[[VEC0]]
+  // CHECK: %[[RES1:.+]] = arith.addf %[[VEC1]], %[[VEC1]]
+  %1 = arith.addf %0, %0 : vector<2x2xf32>
+
+  // return %[[RES0]], %%[[RES1]] : vector<2xf32>, vector<2xf32>
+  return %1 : vector<2x2xf32>
+}

amd-eochoalo · 2025-09-16T16:16:52Z

mlir/test/Conversion/ConvertToSPIRV/vector-unroll.mlir

  %0 = vector.transpose %arg0, [1, 0] : vector<2x3xi32> to vector<3x2xi32>
  return %0 : vector<3x2xi32>
 }
+


It looks like these are the only tests that check for unrolling, although it does not check the conversion from vector to spirv, it just checks that the unrolling was applied. We could also start a new file that runs test-convert-to-spirv with run-vector-unrolling=true as an integration test.

I'm not sure if we need specific vector unrolling test for SPIR-V but we should add independent test to test/Dialect/Vector/.... I think we have vector-unroll-options.mlir for unrolling.

I'm tasked with making sure that vector.to_elements and vector.from_elements are unrolled during the test-convert-to-spirv lowering, so I think that adding tests here makes sense.

amd-eochoalo · 2025-09-16T20:29:26Z

mlir/lib/Dialect/Vector/Transforms/VectorUnroll.cpp

+               UnrollStorePattern, UnrollBroadcastPattern, UnrollFromElements,
+               UnrollToElements>(patterns.getContext(), options, benefit);


Suggested change

UnrollStorePattern, UnrollBroadcastPattern, UnrollFromElements,

UnrollToElements>(patterns.getContext(), options, benefit);

UnrollStorePattern, UnrollBroadcastPattern>(

patterns.getContext(), options, benefit);

If it is more convenient I can make this a non-functional change first by removing these patterns and removing the test. Then I can make another PR which adds the patterns.

mlir/lib/Dialect/Vector/Transforms/VectorUnroll.cpp

dcaballe

LG, a few comments. Thanks!

mlir/include/mlir/Dialect/Vector/Transforms/VectorRewritePatterns.h

dcaballe · 2025-09-17T18:54:46Z

mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp

      // Transform N-D vector.from_elements to 1-D vector.from_elements before
      // conversion.
-      vector::populateVectorFromElementsLoweringPatterns(patterns);
+      vector::populateVectorFromElementsUnrollPatterns(patterns);


Why do we need to add just the unroll patterns for vector.from_elements here? It sounds a more generic unrolling for all the ops should have happen before calling this pass?

I kept these since we have transform dialect operations that call these functions. I think I can follow up this PR and remove them, but I would rather keep this PR simple.

mlir/lib/Dialect/Vector/Transforms/VectorUnroll.cpp

dcaballe · 2025-09-17T19:04:43Z

mlir/test/Conversion/ConvertToSPIRV/vector-unroll.mlir

  %0 = vector.transpose %arg0, [1, 0] : vector<2x3xi32> to vector<3x2xi32>
  return %0 : vector<3x2xi32>
 }
+


I'm not sure if we need specific vector unrolling test for SPIR-V but we should add independent test to test/Dialect/Vector/.... I think we have vector-unroll-options.mlir for unrolling.

banach-space · 2025-09-17T19:26:03Z

This PR moves the patterns that unroll vector.to_elements and vector.from_elements into the file with other vector unrolling operations.

+1, thanks!

llvm-ci · 2025-09-18T17:59:09Z

LLVM Buildbot has detected a new failure on builder ppc64le-mlir-rhel-clang running on ppc64le-mlir-rhel-test while building mlir at step 6 "test-build-check-mlir-build-only-check-mlir".

Full details are available at: https://lab.llvm.org/buildbot/#/builders/129/builds/29793

Here is the relevant piece of the build log for the reference

Step 6 (test-build-check-mlir-build-only-check-mlir) failure: 1200 seconds without output running [b'ninja', b'check-mlir'], attempting to kill
...
PASS: MLIR :: mlir-tblgen/llvm-intrinsics.td (3487 of 3498)
PASS: MLIR :: Pass/pipeline-options-parsing.mlir (3488 of 3498)
PASS: MLIR-Unit :: IR/./MLIRIRTests/0/129 (3489 of 3498)
PASS: MLIR-Unit :: IR/./MLIRIRTests/38/129 (3490 of 3498)
PASS: MLIR-Unit :: Interfaces/./MLIRInterfacesTests/13/22 (3491 of 3498)
PASS: MLIR-Unit :: Interfaces/./MLIRInterfacesTests/11/22 (3492 of 3498)
PASS: MLIR-Unit :: Pass/./MLIRPassTests/10/13 (3493 of 3498)
PASS: MLIR-Unit :: Interfaces/./MLIRInterfacesTests/12/22 (3494 of 3498)
PASS: MLIR-Unit :: IR/./MLIRIRTests/39/129 (3495 of 3498)
PASS: MLIR :: mlir-reduce/dce-test.mlir (3496 of 3498)
command timed out: 1200 seconds without output running [b'ninja', b'check-mlir'], attempting to kill
process killed by signal 9
program finished with exit code -1
elapsedTime=2342.098277

…roll.cpp` (llvm#159118) This PR moves the patterns that unroll vector.to_elements and vector.from_elements into the file with other vector unrolling operations. This PR also adds these unrolling patterns into the `populateVectorUnrollPatterns`. And renames `populateVectorToElementsLoweringPatterns` `populateVectorFromElementsLoweringPatterns` to `populateVectorToElementsUnrollPatterns` `populateVectorFromElementsUnrollPatterns`.

[mlir] Add vector.{to_elements,from_elements} unrolling to VectorToSPIRV

08655a9

amd-eochoalo changed the title ~~[mlir] Add vector.{to_elements,from_elements} unrolling to VectorToSPIRV~~ [mlir] Add vector.{to_elements,from_elements} unrolling to test-spirv-vector-unrolling Sep 16, 2025

amd-eochoalo changed the title ~~[mlir] Add vector.{to_elements,from_elements} unrolling to test-spirv-vector-unrolling~~ [mlir] Add vector.{to_elements,from_elements} unrolling to unrollVectorsInFuncBodies Sep 16, 2025

amd-eochoalo commented Sep 16, 2025

View reviewed changes

mlir/lib/Dialect/SPIRV/Transforms/SPIRVConversion.cpp Outdated Show resolved Hide resolved

amd-eochoalo marked this pull request as ready for review September 16, 2025 16:14

amd-eochoalo requested review from antiagainst and kuhar as code owners September 16, 2025 16:14

llvmbot added mlir:spirv mlir labels Sep 16, 2025

amd-eochoalo commented Sep 16, 2025

View reviewed changes

amd-eochoalo added 6 commits September 16, 2025 11:25

populate patterns inside populateVectorUnrollPatterns

c891a27

Copy over UnrollToElements

e1f9605

Removes reference to previous UnrollToElements pattern

5c3e7d5

Copy over UnrollFromElements

04431c5

Removes reference to previous UnrollFromElements pattern

70a667e

Adds UnrollVectorOptions to ToElements and ForElements patterns

c054f16

amd-eochoalo requested review from fabianmcg, ftynse, nicolasvasilache, banach-space, dcaballe, hanhanW and Groverkss as code owners September 16, 2025 20:12

llvmbot added mlir:gpu mlir:vectorops mlir:vector labels Sep 16, 2025

amd-eochoalo changed the title ~~[mlir] Add vector.{to_elements,from_elements} unrolling to unrollVectorsInFuncBodies~~ [mlir] Move vector.{to_elements,from_elements} unrolling to VectorUnroll.cpp Sep 16, 2025

amd-eochoalo commented Sep 16, 2025

View reviewed changes

kuhar approved these changes Sep 16, 2025

View reviewed changes

mlir/lib/Dialect/Vector/Transforms/VectorUnroll.cpp Outdated Show resolved Hide resolved

amd-eochoalo added 2 commits September 17, 2025 10:49

Merge branch 'main' into eochoa/2025-09-16/unroll-to-from-elements

28ca33b

Address review comments

5a266fb

amd-eochoalo mentioned this pull request Sep 17, 2025

Rename functions iree-org/iree#22020

Closed

dcaballe reviewed Sep 17, 2025

View reviewed changes

amd-eochoalo added 4 commits September 18, 2025 08:43

Update documentation

f0283ca

Reword documentation 'exhaustive' -> 'fixed-point'

5d7afb1

Add documentation to UnrollToElements

44145b5

Merge branch 'main' into eochoa/2025-09-16/unroll-to-from-elements

d3af4ce

amd-eochoalo merged commit 8b9c70d into llvm:main Sep 18, 2025
9 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[mlir] Move vector.{to_elements,from_elements} unrolling to `VectorUnroll.cpp` #159118

[mlir] Move vector.{to_elements,from_elements} unrolling to `VectorUnroll.cpp` #159118

Uh oh!

amd-eochoalo commented Sep 16, 2025 •

edited

Loading

Uh oh!

Uh oh!

llvmbot commented Sep 16, 2025 •

edited

Loading

Uh oh!

llvmbot commented Sep 16, 2025

Uh oh!

amd-eochoalo Sep 16, 2025

Uh oh!

dcaballe Sep 17, 2025

Uh oh!

amd-eochoalo Sep 18, 2025

Uh oh!

amd-eochoalo Sep 16, 2025 •

edited

Loading

Uh oh!

Uh oh!

dcaballe left a comment

Uh oh!

Uh oh!

dcaballe Sep 17, 2025

Uh oh!

amd-eochoalo Sep 17, 2025

Uh oh!

Uh oh!

Uh oh!

dcaballe Sep 17, 2025

Uh oh!

banach-space commented Sep 17, 2025

Uh oh!

Uh oh!

llvm-ci commented Sep 18, 2025

Uh oh!

Uh oh!

		UnrollStorePattern, UnrollBroadcastPattern, UnrollFromElements,
		UnrollToElements>(patterns.getContext(), options, benefit);

[mlir] Move vector.{to_elements,from_elements} unrolling to VectorUnroll.cpp #159118

[mlir] Move vector.{to_elements,from_elements} unrolling to VectorUnroll.cpp #159118

Uh oh!

Conversation

amd-eochoalo commented Sep 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

llvmbot commented Sep 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmbot commented Sep 16, 2025

Uh oh!

amd-eochoalo Sep 16, 2025

Choose a reason for hiding this comment

Uh oh!

dcaballe Sep 17, 2025

Choose a reason for hiding this comment

Uh oh!

amd-eochoalo Sep 18, 2025

Choose a reason for hiding this comment

Uh oh!

amd-eochoalo Sep 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

dcaballe left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

dcaballe Sep 17, 2025

Choose a reason for hiding this comment

Uh oh!

amd-eochoalo Sep 17, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

dcaballe Sep 17, 2025

Choose a reason for hiding this comment

Uh oh!

banach-space commented Sep 17, 2025

Uh oh!

Uh oh!

llvm-ci commented Sep 18, 2025

Uh oh!

Uh oh!

[mlir] Move vector.{to_elements,from_elements} unrolling to `VectorUnroll.cpp` #159118

[mlir] Move vector.{to_elements,from_elements} unrolling to `VectorUnroll.cpp` #159118

amd-eochoalo commented Sep 16, 2025 •

edited

Loading

llvmbot commented Sep 16, 2025 •

edited

Loading

amd-eochoalo Sep 16, 2025 •

edited

Loading