Skip to content

[mlir][vector] Use vector.broadcast in place of vector.splat #148028

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

[mlir][vector] Use vector.broadcast in place of vector.splat #148028

wants to merge 1 commit into from
+7 −6

Conversation

newling
Copy link
Contributor

@newling newling commented Jul 10, 2025

Part of deprecation of vector.splat

RFC: https://discourse.llvm.org/t/rfc-mlir-vector-deprecate-then-remove-vector-splat/87143/4
More complete deprecation: #147818

@newling newling requested a review from grypp as a code owner July 10, 2025 18:45
@llvmbot
Copy link
Member

llvmbot commented Jul 10, 2025 •
edited
Loading

@llvm/pr-subscribers-mlir

@llvm/pr-subscribers-mlir-nvgpu

Author: James Newling (newling)

Changes

Part of deprecation of vector.splat

RFC: https://discourse.llvm.org/t/rfc-mlir-vector-deprecate-then-remove-vector-splat/87143/4
More complete deprecation: #147818

Full diff: https://github.com/llvm/llvm-project/pull/148028.diff

4 Files Affected:

  • (modified) mlir/lib/Dialect/Arith/Transforms/EmulateUnsupportedFloats.cpp (+1-1)
  • (modified) mlir/lib/Dialect/NVGPU/TransformOps/NVGPUTransformOps.cpp (+1-1)
  • (modified) mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp (+2-1)
  • (modified) mlir/test/Dialect/NVGPU/transform-matmul-to-nvvm.mlir (+3-3)
diff --git a/mlir/lib/Dialect/Arith/Transforms/EmulateUnsupportedFloats.cpp b/mlir/lib/Dialect/Arith/Transforms/EmulateUnsupportedFloats.cpp
index 62022bfb7df1e..22dd3bd0ea98f 100644
--- a/mlir/lib/Dialect/Arith/Transforms/EmulateUnsupportedFloats.cpp
+++ b/mlir/lib/Dialect/Arith/Transforms/EmulateUnsupportedFloats.cpp
@@ -123,7 +123,7 @@ void mlir::arith::populateEmulateUnsupportedFloatsLegality(
       vector::FMAOp, vector::OuterProductOp, vector::MatmulOp, vector::ScanOp>(
       [&](Operation *op) { return converter.isLegal(op); });
   target.addLegalOp<arith::BitcastOp, arith::ExtFOp, arith::TruncFOp,
-                    arith::ConstantOp, vector::SplatOp>();
+                    arith::ConstantOp, vector::SplatOp, vector::BroadcastOp>();
 }
 
 void EmulateUnsupportedFloatsPass::runOnOperation() {
diff --git a/mlir/lib/Dialect/NVGPU/TransformOps/NVGPUTransformOps.cpp b/mlir/lib/Dialect/NVGPU/TransformOps/NVGPUTransformOps.cpp
index d2c94b124cdfb..bcd62acf6b9ce 100644
--- a/mlir/lib/Dialect/NVGPU/TransformOps/NVGPUTransformOps.cpp
+++ b/mlir/lib/Dialect/NVGPU/TransformOps/NVGPUTransformOps.cpp
@@ -688,7 +688,7 @@ Value MmaSyncBuilder::buildMmaSyncMemRefLoadOperand(
 
   Type elementType = getElementTypeOrSelf(memref.getType());
   auto vt = VectorType::get(vectorShape, elementType);
-  Value res = b.create<vector::SplatOp>(loc, vt, loads[0]);
+  Value res = b.create<vector::BroadcastOp>(loc, vt, loads[0]);
   foreachIndividualVectorElement(
       res,
       /*applyFn=*/
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
index 2c48a735bf956..9ebac21770136 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
@@ -438,7 +438,8 @@ struct UnrollCreateDescOp : public UnrollPattern<xegpu::CreateDescOp> {
           // Compute the offset
           Value inc = rewriter.create<arith::ConstantIndexOp>(
               loc, i * blockedChunkSize);
-          Value incVec = rewriter.create<vector::SplatOp>(loc, indiceType, inc);
+          Value incVec =
+              rewriter.create<vector::BroadcastOp>(loc, indiceType, inc);
           Value offsetIndice =
               rewriter.create<arith::AddIOp>(loc, indice, incVec);
 
diff --git a/mlir/test/Dialect/NVGPU/transform-matmul-to-nvvm.mlir b/mlir/test/Dialect/NVGPU/transform-matmul-to-nvvm.mlir
index 07e03f3b8473d..bbe27fe1b99d9 100644
--- a/mlir/test/Dialect/NVGPU/transform-matmul-to-nvvm.mlir
+++ b/mlir/test/Dialect/NVGPU/transform-matmul-to-nvvm.mlir
@@ -20,14 +20,14 @@ func.func @matmul_16x8x4xf32_global(
 // CHECK:           %[[VAL_7:.*]] = affine.apply #[[$div4p8]]()[%[[TIDX]]]
 // CHECK:           %[[VAL_8:.*]] = affine.apply #[[$mod4]]()[%[[TIDX]]]
 // CHECK:           %[[VAL_9:.*]] = memref.load %[[VAL_0]][%[[VAL_7]], %[[VAL_8]]] : memref<16x4xf32>
-// CHECK:           %[[VAL_10:.*]] = vector.splat %[[VAL_6]] : vector<2x1xf32>
+// CHECK:           %[[VAL_10:.*]] = vector.broadcast %[[VAL_6]] : f32 to vector<2x1xf32>
 // CHECK:           %[[VAL_11:.*]] = vector.insert %[[VAL_6]], %[[VAL_10]] [0, 0] : f32 into vector<2x1xf32>
 // CHECK:           %[[LHS:.*]] = vector.insert %[[VAL_9]], %[[VAL_11]] [1, 0] : f32 into vector<2x1xf32>
 //
 // CHECK:           %[[VAL_13:.*]] = affine.apply #[[$mod4]]()[%[[TIDX]]]
 // CHECK:           %[[VAL_14:.*]] = affine.apply #[[$div4]]()[%[[TIDX]]]
 // CHECK:           %[[VAL_15:.*]] = memref.load %[[VAL_1]][%[[VAL_13]], %[[VAL_14]]] : memref<4x8xf32>
-// CHECK:           %[[VAL_16:.*]] = vector.splat %[[VAL_15]] : vector<1x1xf32>
+// CHECK:           %[[VAL_16:.*]] = vector.broadcast %[[VAL_15]] : f32 to vector<1x1xf32>
 // CHECK:           %[[RHS:.*]] = vector.insert %[[VAL_15]], %[[VAL_16]] [0, 0] : f32 into vector<1x1xf32>
 //
 // CHECK:           %[[VAL_18:.*]] = affine.apply #[[$div4]]()[%[[TIDX]]]
@@ -42,7 +42,7 @@ func.func @matmul_16x8x4xf32_global(
 // CHECK:           %[[VAL_27:.*]] = affine.apply #[[$div4p8]]()[%[[TIDX]]]
 // CHECK:           %[[VAL_28:.*]] = affine.apply #[[$map4]]()[%[[TIDX]]]
 // CHECK:           %[[VAL_29:.*]] = memref.load %[[VAL_2]][%[[VAL_27]], %[[VAL_28]]] : memref<16x8xf32>
-// CHECK:           %[[VAL_30:.*]] = vector.splat %[[VAL_20]] : vector<2x2xf32>
+// CHECK:           %[[VAL_30:.*]] = vector.broadcast %[[VAL_20]] : f32 to vector<2x2xf32>
 // CHECK:           %[[VAL_31:.*]] = vector.insert %[[VAL_20]], %[[VAL_30]] [0, 0] : f32 into vector<2x2xf32>
 // CHECK:           %[[VAL_32:.*]] = vector.insert %[[VAL_23]], %[[VAL_31]] [0, 1] : f32 into vector<2x2xf32>
 // CHECK:           %[[VAL_33:.*]] = vector.insert %[[VAL_26]], %[[VAL_32]] [1, 0] : f32 into vector<2x2xf32>

@llvmbot
Copy link
Member

llvmbot commented Jul 10, 2025

@llvm/pr-subscribers-mlir-arith

Author: James Newling (newling)

Changes

Part of deprecation of vector.splat

RFC: https://discourse.llvm.org/t/rfc-mlir-vector-deprecate-then-remove-vector-splat/87143/4
More complete deprecation: #147818

Full diff: https://github.com/llvm/llvm-project/pull/148028.diff

4 Files Affected:

  • (modified) mlir/lib/Dialect/Arith/Transforms/EmulateUnsupportedFloats.cpp (+1-1)
  • (modified) mlir/lib/Dialect/NVGPU/TransformOps/NVGPUTransformOps.cpp (+1-1)
  • (modified) mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp (+2-1)
  • (modified) mlir/test/Dialect/NVGPU/transform-matmul-to-nvvm.mlir (+3-3)
diff --git a/mlir/lib/Dialect/Arith/Transforms/EmulateUnsupportedFloats.cpp b/mlir/lib/Dialect/Arith/Transforms/EmulateUnsupportedFloats.cpp
index 62022bfb7df1e..22dd3bd0ea98f 100644
--- a/mlir/lib/Dialect/Arith/Transforms/EmulateUnsupportedFloats.cpp
+++ b/mlir/lib/Dialect/Arith/Transforms/EmulateUnsupportedFloats.cpp
@@ -123,7 +123,7 @@ void mlir::arith::populateEmulateUnsupportedFloatsLegality(
       vector::FMAOp, vector::OuterProductOp, vector::MatmulOp, vector::ScanOp>(
       [&](Operation *op) { return converter.isLegal(op); });
   target.addLegalOp<arith::BitcastOp, arith::ExtFOp, arith::TruncFOp,
-                    arith::ConstantOp, vector::SplatOp>();
+                    arith::ConstantOp, vector::SplatOp, vector::BroadcastOp>();
 }
 
 void EmulateUnsupportedFloatsPass::runOnOperation() {
diff --git a/mlir/lib/Dialect/NVGPU/TransformOps/NVGPUTransformOps.cpp b/mlir/lib/Dialect/NVGPU/TransformOps/NVGPUTransformOps.cpp
index d2c94b124cdfb..bcd62acf6b9ce 100644
--- a/mlir/lib/Dialect/NVGPU/TransformOps/NVGPUTransformOps.cpp
+++ b/mlir/lib/Dialect/NVGPU/TransformOps/NVGPUTransformOps.cpp
@@ -688,7 +688,7 @@ Value MmaSyncBuilder::buildMmaSyncMemRefLoadOperand(
 
   Type elementType = getElementTypeOrSelf(memref.getType());
   auto vt = VectorType::get(vectorShape, elementType);
-  Value res = b.create<vector::SplatOp>(loc, vt, loads[0]);
+  Value res = b.create<vector::BroadcastOp>(loc, vt, loads[0]);
   foreachIndividualVectorElement(
       res,
       /*applyFn=*/
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
index 2c48a735bf956..9ebac21770136 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
@@ -438,7 +438,8 @@ struct UnrollCreateDescOp : public UnrollPattern<xegpu::CreateDescOp> {
           // Compute the offset
           Value inc = rewriter.create<arith::ConstantIndexOp>(
               loc, i * blockedChunkSize);
-          Value incVec = rewriter.create<vector::SplatOp>(loc, indiceType, inc);
+          Value incVec =
+              rewriter.create<vector::BroadcastOp>(loc, indiceType, inc);
           Value offsetIndice =
               rewriter.create<arith::AddIOp>(loc, indice, incVec);
 
diff --git a/mlir/test/Dialect/NVGPU/transform-matmul-to-nvvm.mlir b/mlir/test/Dialect/NVGPU/transform-matmul-to-nvvm.mlir
index 07e03f3b8473d..bbe27fe1b99d9 100644
--- a/mlir/test/Dialect/NVGPU/transform-matmul-to-nvvm.mlir
+++ b/mlir/test/Dialect/NVGPU/transform-matmul-to-nvvm.mlir
@@ -20,14 +20,14 @@ func.func @matmul_16x8x4xf32_global(
 // CHECK:           %[[VAL_7:.*]] = affine.apply #[[$div4p8]]()[%[[TIDX]]]
 // CHECK:           %[[VAL_8:.*]] = affine.apply #[[$mod4]]()[%[[TIDX]]]
 // CHECK:           %[[VAL_9:.*]] = memref.load %[[VAL_0]][%[[VAL_7]], %[[VAL_8]]] : memref<16x4xf32>
-// CHECK:           %[[VAL_10:.*]] = vector.splat %[[VAL_6]] : vector<2x1xf32>
+// CHECK:           %[[VAL_10:.*]] = vector.broadcast %[[VAL_6]] : f32 to vector<2x1xf32>
 // CHECK:           %[[VAL_11:.*]] = vector.insert %[[VAL_6]], %[[VAL_10]] [0, 0] : f32 into vector<2x1xf32>
 // CHECK:           %[[LHS:.*]] = vector.insert %[[VAL_9]], %[[VAL_11]] [1, 0] : f32 into vector<2x1xf32>
 //
 // CHECK:           %[[VAL_13:.*]] = affine.apply #[[$mod4]]()[%[[TIDX]]]
 // CHECK:           %[[VAL_14:.*]] = affine.apply #[[$div4]]()[%[[TIDX]]]
 // CHECK:           %[[VAL_15:.*]] = memref.load %[[VAL_1]][%[[VAL_13]], %[[VAL_14]]] : memref<4x8xf32>
-// CHECK:           %[[VAL_16:.*]] = vector.splat %[[VAL_15]] : vector<1x1xf32>
+// CHECK:           %[[VAL_16:.*]] = vector.broadcast %[[VAL_15]] : f32 to vector<1x1xf32>
 // CHECK:           %[[RHS:.*]] = vector.insert %[[VAL_15]], %[[VAL_16]] [0, 0] : f32 into vector<1x1xf32>
 //
 // CHECK:           %[[VAL_18:.*]] = affine.apply #[[$div4]]()[%[[TIDX]]]
@@ -42,7 +42,7 @@ func.func @matmul_16x8x4xf32_global(
 // CHECK:           %[[VAL_27:.*]] = affine.apply #[[$div4p8]]()[%[[TIDX]]]
 // CHECK:           %[[VAL_28:.*]] = affine.apply #[[$map4]]()[%[[TIDX]]]
 // CHECK:           %[[VAL_29:.*]] = memref.load %[[VAL_2]][%[[VAL_27]], %[[VAL_28]]] : memref<16x8xf32>
-// CHECK:           %[[VAL_30:.*]] = vector.splat %[[VAL_20]] : vector<2x2xf32>
+// CHECK:           %[[VAL_30:.*]] = vector.broadcast %[[VAL_20]] : f32 to vector<2x2xf32>
 // CHECK:           %[[VAL_31:.*]] = vector.insert %[[VAL_20]], %[[VAL_30]] [0, 0] : f32 into vector<2x2xf32>
 // CHECK:           %[[VAL_32:.*]] = vector.insert %[[VAL_23]], %[[VAL_31]] [0, 1] : f32 into vector<2x2xf32>
 // CHECK:           %[[VAL_33:.*]] = vector.insert %[[VAL_26]], %[[VAL_32]] [1, 0] : f32 into vector<2x2xf32>

@llvmbot
Copy link
Member

llvmbot commented Jul 10, 2025

@llvm/pr-subscribers-mlir-gpu

Author: James Newling (newling)

Changes

Part of deprecation of vector.splat

RFC: https://discourse.llvm.org/t/rfc-mlir-vector-deprecate-then-remove-vector-splat/87143/4
More complete deprecation: #147818

Full diff: https://github.com/llvm/llvm-project/pull/148028.diff

4 Files Affected:

  • (modified) mlir/lib/Dialect/Arith/Transforms/EmulateUnsupportedFloats.cpp (+1-1)
  • (modified) mlir/lib/Dialect/NVGPU/TransformOps/NVGPUTransformOps.cpp (+1-1)
  • (modified) mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp (+2-1)
  • (modified) mlir/test/Dialect/NVGPU/transform-matmul-to-nvvm.mlir (+3-3)
diff --git a/mlir/lib/Dialect/Arith/Transforms/EmulateUnsupportedFloats.cpp b/mlir/lib/Dialect/Arith/Transforms/EmulateUnsupportedFloats.cpp
index 62022bfb7df1e..22dd3bd0ea98f 100644
--- a/mlir/lib/Dialect/Arith/Transforms/EmulateUnsupportedFloats.cpp
+++ b/mlir/lib/Dialect/Arith/Transforms/EmulateUnsupportedFloats.cpp
@@ -123,7 +123,7 @@ void mlir::arith::populateEmulateUnsupportedFloatsLegality(
       vector::FMAOp, vector::OuterProductOp, vector::MatmulOp, vector::ScanOp>(
       [&](Operation *op) { return converter.isLegal(op); });
   target.addLegalOp<arith::BitcastOp, arith::ExtFOp, arith::TruncFOp,
-                    arith::ConstantOp, vector::SplatOp>();
+                    arith::ConstantOp, vector::SplatOp, vector::BroadcastOp>();
 }
 
 void EmulateUnsupportedFloatsPass::runOnOperation() {
diff --git a/mlir/lib/Dialect/NVGPU/TransformOps/NVGPUTransformOps.cpp b/mlir/lib/Dialect/NVGPU/TransformOps/NVGPUTransformOps.cpp
index d2c94b124cdfb..bcd62acf6b9ce 100644
--- a/mlir/lib/Dialect/NVGPU/TransformOps/NVGPUTransformOps.cpp
+++ b/mlir/lib/Dialect/NVGPU/TransformOps/NVGPUTransformOps.cpp
@@ -688,7 +688,7 @@ Value MmaSyncBuilder::buildMmaSyncMemRefLoadOperand(
 
   Type elementType = getElementTypeOrSelf(memref.getType());
   auto vt = VectorType::get(vectorShape, elementType);
-  Value res = b.create<vector::SplatOp>(loc, vt, loads[0]);
+  Value res = b.create<vector::BroadcastOp>(loc, vt, loads[0]);
   foreachIndividualVectorElement(
       res,
       /*applyFn=*/
diff --git a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
index 2c48a735bf956..9ebac21770136 100644
--- a/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
+++ b/mlir/lib/Dialect/XeGPU/Transforms/XeGPUUnroll.cpp
@@ -438,7 +438,8 @@ struct UnrollCreateDescOp : public UnrollPattern<xegpu::CreateDescOp> {
           // Compute the offset
           Value inc = rewriter.create<arith::ConstantIndexOp>(
               loc, i * blockedChunkSize);
-          Value incVec = rewriter.create<vector::SplatOp>(loc, indiceType, inc);
+          Value incVec =
+              rewriter.create<vector::BroadcastOp>(loc, indiceType, inc);
           Value offsetIndice =
               rewriter.create<arith::AddIOp>(loc, indice, incVec);
 
diff --git a/mlir/test/Dialect/NVGPU/transform-matmul-to-nvvm.mlir b/mlir/test/Dialect/NVGPU/transform-matmul-to-nvvm.mlir
index 07e03f3b8473d..bbe27fe1b99d9 100644
--- a/mlir/test/Dialect/NVGPU/transform-matmul-to-nvvm.mlir
+++ b/mlir/test/Dialect/NVGPU/transform-matmul-to-nvvm.mlir
@@ -20,14 +20,14 @@ func.func @matmul_16x8x4xf32_global(
 // CHECK:           %[[VAL_7:.*]] = affine.apply #[[$div4p8]]()[%[[TIDX]]]
 // CHECK:           %[[VAL_8:.*]] = affine.apply #[[$mod4]]()[%[[TIDX]]]
 // CHECK:           %[[VAL_9:.*]] = memref.load %[[VAL_0]][%[[VAL_7]], %[[VAL_8]]] : memref<16x4xf32>
-// CHECK:           %[[VAL_10:.*]] = vector.splat %[[VAL_6]] : vector<2x1xf32>
+// CHECK:           %[[VAL_10:.*]] = vector.broadcast %[[VAL_6]] : f32 to vector<2x1xf32>
 // CHECK:           %[[VAL_11:.*]] = vector.insert %[[VAL_6]], %[[VAL_10]] [0, 0] : f32 into vector<2x1xf32>
 // CHECK:           %[[LHS:.*]] = vector.insert %[[VAL_9]], %[[VAL_11]] [1, 0] : f32 into vector<2x1xf32>
 //
 // CHECK:           %[[VAL_13:.*]] = affine.apply #[[$mod4]]()[%[[TIDX]]]
 // CHECK:           %[[VAL_14:.*]] = affine.apply #[[$div4]]()[%[[TIDX]]]
 // CHECK:           %[[VAL_15:.*]] = memref.load %[[VAL_1]][%[[VAL_13]], %[[VAL_14]]] : memref<4x8xf32>
-// CHECK:           %[[VAL_16:.*]] = vector.splat %[[VAL_15]] : vector<1x1xf32>
+// CHECK:           %[[VAL_16:.*]] = vector.broadcast %[[VAL_15]] : f32 to vector<1x1xf32>
 // CHECK:           %[[RHS:.*]] = vector.insert %[[VAL_15]], %[[VAL_16]] [0, 0] : f32 into vector<1x1xf32>
 //
 // CHECK:           %[[VAL_18:.*]] = affine.apply #[[$div4]]()[%[[TIDX]]]
@@ -42,7 +42,7 @@ func.func @matmul_16x8x4xf32_global(
 // CHECK:           %[[VAL_27:.*]] = affine.apply #[[$div4p8]]()[%[[TIDX]]]
 // CHECK:           %[[VAL_28:.*]] = affine.apply #[[$map4]]()[%[[TIDX]]]
 // CHECK:           %[[VAL_29:.*]] = memref.load %[[VAL_2]][%[[VAL_27]], %[[VAL_28]]] : memref<16x8xf32>
-// CHECK:           %[[VAL_30:.*]] = vector.splat %[[VAL_20]] : vector<2x2xf32>
+// CHECK:           %[[VAL_30:.*]] = vector.broadcast %[[VAL_20]] : f32 to vector<2x2xf32>
 // CHECK:           %[[VAL_31:.*]] = vector.insert %[[VAL_20]], %[[VAL_30]] [0, 0] : f32 into vector<2x2xf32>
 // CHECK:           %[[VAL_32:.*]] = vector.insert %[[VAL_23]], %[[VAL_31]] [0, 1] : f32 into vector<2x2xf32>
 // CHECK:           %[[VAL_33:.*]] = vector.insert %[[VAL_26]], %[[VAL_32]] [1, 0] : f32 into vector<2x2xf32>

@adam-smnk adam-smnk requested a review from chencha3 July 10, 2025 19:40
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@grypp grypp Awaiting requested review from grypp grypp is a code owner

@chencha3 chencha3 Awaiting requested review from chencha3

Assignees
No one assigned
Labels
mlir:arith mlir:gpu mlir:nvgpu mlir
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
2 participants
@newling @llvmbot