[Mlir-commits] [mlir] [mlir][vector] Sink vector.extract/splat into load/store ops (PR #134389)
Ivan Butygin
llvmlistbot at llvm.org
Fri Apr 4 07:39:58 PDT 2025
https://github.com/Hardcode84 created https://github.com/llvm/llvm-project/pull/134389
```
vector.load %arg0[%arg1] : memref<?xf32>, vector<4xf32>
vector.extract %0[1] : f32 from vector<4xf32>
```
Gets converted to:
```
%c1 = arith.constant 1 : index
%0 = arith.addi %arg1, %c1 overflow<nsw> : index
%1 = memref.load %arg0[%0] : memref<?xf32>
```
```
%0 = vector.splat %arg2 : vector<1xf32>
vector.store %0, %arg0[%arg1] : memref<?xf32>, vector<1xf32>
```
Gets converted to:
```
memref.store %arg2, %arg0[%arg1] : memref<?xf32>
```
>From cd97f248a6f85fb7b634508844dfe9e2478f50c1 Mon Sep 17 00:00:00 2001
From: Ivan Butygin <ivan.butygin at gmail.com>
Date: Fri, 4 Apr 2025 15:39:48 +0200
Subject: [PATCH 1/2] [mlir][vector] Sink extract/splat into load/store ops
```
vector.load %arg0[%arg1] : memref<?xf32>, vector<4xf32>
vector.extract %0[1] : f32 from vector<4xf32>
```
Gets converted to:
```
%c1 = arith.constant 1 : index
%0 = arith.addi %arg1, %c1 overflow<nsw> : index
%1 = memref.load %arg0[%0] : memref<?xf32>
```
```
%0 = vector.splat %arg2 : vector<1xf32>
vector.store %0, %arg0[%arg1] : memref<?xf32>, vector<1xf32>
```
Gets converted to:
```
memref.store %arg2, %arg0[%arg1] : memref<?xf32>
```
---
.../Vector/TransformOps/VectorTransformOps.td | 24 ++-
.../Vector/Transforms/VectorRewritePatterns.h | 5 +
.../TransformOps/VectorTransformOps.cpp | 5 +
.../Vector/Transforms/VectorTransforms.cpp | 127 ++++++++++++
.../Dialect/Vector/vector-sink-transform.mlir | 1 +
mlir/test/Dialect/Vector/vector-sink.mlir | 189 ++++++++++++++++++
.../Dialect/Vector/TestVectorTransforms.cpp | 1 +
7 files changed, 350 insertions(+), 2 deletions(-)
diff --git a/mlir/include/mlir/Dialect/Vector/TransformOps/VectorTransformOps.td b/mlir/include/mlir/Dialect/Vector/TransformOps/VectorTransformOps.td
index f46aa0428f12f..7fbb437908866 100644
--- a/mlir/include/mlir/Dialect/Vector/TransformOps/VectorTransformOps.td
+++ b/mlir/include/mlir/Dialect/Vector/TransformOps/VectorTransformOps.td
@@ -469,8 +469,28 @@ def ApplySinkVectorPatternsOp : Op<Transform_Dialect,
%0 = arith.addf %a, %b : vector<4x2xf32>
%r = vector.transpose %0, [1, 0] : vector<2x4xf32>
```
- At the moment, these patterns are limited to vector.broadcast and
- vector.transpose.
+ At the moment, these patterns are limited to vector.broadcast,
+ vector.transpose and vector.extract.
+ }];
+
+ let assemblyFormat = "attr-dict";
+}
+
+def ApplySinkVectorMemPatternsOp : Op<Transform_Dialect,
+ "apply_patterns.vector.sink_mem_ops",
+ [DeclareOpInterfaceMethods<PatternDescriptorOpInterface>]> {
+ let description = [{
+ Patterns that remove redundant Vector Ops by merging them with load/store
+ ops
+ ```
+ vector.load %arg0[%arg1] : memref<?xf32>, vector<4xf32>
+ vector.extract %0[1] : f32 from vector<4xf32>
+ ```
+ Gets converted to:
+ ```
+ %c1 = arith.constant 1 : index
+ %0 = arith.addi %arg1, %c1 overflow<nsw> : index
+ %1 = memref.load %arg0[%0] : memref<?xf32>
}];
let assemblyFormat = "attr-dict";
diff --git a/mlir/include/mlir/Dialect/Vector/Transforms/VectorRewritePatterns.h b/mlir/include/mlir/Dialect/Vector/Transforms/VectorRewritePatterns.h
index 7de4a6a315750..7d3134bdae233 100644
--- a/mlir/include/mlir/Dialect/Vector/Transforms/VectorRewritePatterns.h
+++ b/mlir/include/mlir/Dialect/Vector/Transforms/VectorRewritePatterns.h
@@ -161,6 +161,11 @@ void populateVectorTransferCollapseInnerMostContiguousDimsPatterns(
void populateSinkVectorOpsPatterns(RewritePatternSet &patterns,
PatternBenefit benefit = 1);
+/// Patterns that remove redundant Vector Ops by re-ordering them with
+/// memory Ops:
+void populateSinkVectorMemOpsPatterns(RewritePatternSet &patterns,
+ PatternBenefit benefit = 1);
+
/// Patterns that fold chained vector reductions. These patterns assume that
/// elementwise operations (e.g., `arith.addf` with vector operands) are
/// cheaper than vector reduction.
diff --git a/mlir/lib/Dialect/Vector/TransformOps/VectorTransformOps.cpp b/mlir/lib/Dialect/Vector/TransformOps/VectorTransformOps.cpp
index 12dcf768dd928..a888d745be443 100644
--- a/mlir/lib/Dialect/Vector/TransformOps/VectorTransformOps.cpp
+++ b/mlir/lib/Dialect/Vector/TransformOps/VectorTransformOps.cpp
@@ -212,6 +212,11 @@ void transform::ApplySinkVectorPatternsOp::populatePatterns(
vector::populateSinkVectorOpsPatterns(patterns);
}
+void transform::ApplySinkVectorMemPatternsOp::populatePatterns(
+ RewritePatternSet &patterns) {
+ vector::populateSinkVectorMemOpsPatterns(patterns);
+}
+
//===----------------------------------------------------------------------===//
// Transform op registration
//===----------------------------------------------------------------------===//
diff --git a/mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp b/mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp
index b6fac80d871e6..697a4228b3a53 100644
--- a/mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp
@@ -1103,6 +1103,127 @@ class ExtractOpFromElementwise final
}
};
+/// Pattern to rewrite vector.extract(vector.load) -> vector/memref.load.
+/// ```
+/// vector.load %arg0[%arg1] : memref<?xf32>, vector<4xf32>
+/// vector.extract %0[1] : f32 from vector<4xf32>
+/// ```
+/// Gets converted to:
+/// ```
+/// %c1 = arith.constant 1 : index
+/// %0 = arith.addi %arg1, %c1 overflow<nsw> : index
+/// %1 = memref.load %arg0[%0] : memref<?xf32>
+/// ```
+class ExtractOpFromLoad final : public OpRewritePattern<vector::ExtractOp> {
+public:
+ using OpRewritePattern::OpRewritePattern;
+
+ LogicalResult matchAndRewrite(vector::ExtractOp op,
+ PatternRewriter &rewriter) const override {
+ auto loadOp = op.getVector().getDefiningOp<vector::LoadOp>();
+ if (!loadOp)
+ return rewriter.notifyMatchFailure(op, "not a load op");
+
+ if (!loadOp->hasOneUse())
+ return rewriter.notifyMatchFailure(op, "expected single op use");
+
+ VectorType memVecType = loadOp.getVectorType();
+ if (memVecType.isScalable())
+ return rewriter.notifyMatchFailure(op,
+ "scalable vectors are not supported");
+
+ MemRefType memType = loadOp.getMemRefType();
+ if (isa<VectorType>(memType.getElementType()))
+ return rewriter.notifyMatchFailure(
+ op, "memrefs of vectors are not supported");
+
+ int64_t rankOffset = memType.getRank() - memVecType.getRank();
+ if (rankOffset < 0)
+ return rewriter.notifyMatchFailure(op, "unsupported ranks combination");
+
+ auto resVecType = dyn_cast<VectorType>(op.getResult().getType());
+ int64_t finalRank = 0;
+ if (resVecType)
+ finalRank = resVecType.getRank();
+
+ SmallVector<Value> indices = loadOp.getIndices();
+ SmallVector<OpFoldResult> extractPos = op.getMixedPosition();
+
+ OpBuilder::InsertionGuard g(rewriter);
+ rewriter.setInsertionPoint(loadOp);
+ Location loc = loadOp.getLoc();
+ for (auto i : llvm::seq<int64_t>(rankOffset, indices.size() - finalRank)) {
+ OpFoldResult pos = extractPos[i - rankOffset];
+ if (isConstantIntValue(pos, 0))
+ continue;
+
+ Value offset = getValueOrCreateConstantIndexOp(rewriter, loc, pos);
+
+ auto ovf = arith::IntegerOverflowFlags::nsw;
+ indices[i] = rewriter.create<arith::AddIOp>(loc, indices[i], offset, ovf);
+ }
+
+ Value base = loadOp.getBase();
+ if (resVecType) {
+ rewriter.replaceOpWithNewOp<vector::LoadOp>(op, resVecType, base,
+ indices);
+ } else {
+ rewriter.replaceOpWithNewOp<memref::LoadOp>(op, base, indices);
+ }
+ rewriter.eraseOp(loadOp);
+ return success();
+ }
+};
+
+/// Pattern to rewrite vector.store(vector.splat) -> vector/memref.store.
+/// ```
+/// %0 = vector.splat %arg2 : vector<1xf32>
+/// vector.store %0, %arg0[%arg1] : memref<?xf32>, vector<1xf32>
+/// ```
+/// Gets converted to:
+/// ```
+/// memref.store %arg2, %arg0[%arg1] : memref<?xf32>
+/// ```
+class StoreFromSplat final : public OpRewritePattern<vector::StoreOp> {
+public:
+ using OpRewritePattern::OpRewritePattern;
+
+ LogicalResult matchAndRewrite(vector::StoreOp op,
+ PatternRewriter &rewriter) const override {
+ VectorType vecType = op.getVectorType();
+ if (vecType.isScalable())
+ return rewriter.notifyMatchFailure(op,
+ "scalable vectors are not supported");
+
+ if (isa<VectorType>(op.getMemRefType().getElementType()))
+ return rewriter.notifyMatchFailure(
+ op, "memrefs of vectors are not supported");
+
+ if (vecType.getNumElements() != 1)
+ return rewriter.notifyMatchFailure(
+ op, "only 1-element, vectors are supported");
+
+ Operation *splat = op.getValueToStore().getDefiningOp();
+ if (!isa_and_present<vector::BroadcastOp, vector::SplatOp>(splat))
+ return rewriter.notifyMatchFailure(op, "not a splat");
+
+ if (!splat->hasOneUse())
+ return rewriter.notifyMatchFailure(op, "expected single op use");
+
+ Value source = splat->getOperand(0);
+ Value base = op.getBase();
+ ValueRange indices = op.getIndices();
+
+ if (isa<VectorType>(source.getType())) {
+ rewriter.replaceOpWithNewOp<vector::StoreOp>(op, source, base, indices);
+ } else {
+ rewriter.replaceOpWithNewOp<memref::StoreOp>(op, source, base, indices);
+ }
+ rewriter.eraseOp(splat);
+ return success();
+ }
+};
+
// Helper that returns a vector comparison that constructs a mask:
// mask = [0,1,..,n-1] + [o,o,..,o] < [b,b,..,b]
//
@@ -2175,6 +2296,12 @@ void mlir::vector::populateSinkVectorOpsPatterns(RewritePatternSet &patterns,
patterns.getContext(), benefit);
}
+void mlir::vector::populateSinkVectorMemOpsPatterns(RewritePatternSet &patterns,
+ PatternBenefit benefit) {
+ patterns.add<ExtractOpFromLoad, StoreFromSplat>(patterns.getContext(),
+ benefit);
+}
+
void mlir::vector::populateChainedVectorReductionFoldingPatterns(
RewritePatternSet &patterns, PatternBenefit benefit) {
patterns.add<ChainedReduction>(patterns.getContext(), benefit);
diff --git a/mlir/test/Dialect/Vector/vector-sink-transform.mlir b/mlir/test/Dialect/Vector/vector-sink-transform.mlir
index ef17b69b2444c..4d04276742164 100644
--- a/mlir/test/Dialect/Vector/vector-sink-transform.mlir
+++ b/mlir/test/Dialect/Vector/vector-sink-transform.mlir
@@ -7,6 +7,7 @@ module attributes {transform.with_named_sequence} {
%func = transform.structured.match ops{["func.func"]} in %module_op : (!transform.any_op) -> !transform.any_op
transform.apply_patterns to %func {
transform.apply_patterns.vector.sink_ops
+ transform.apply_patterns.vector.sink_mem_ops
} : !transform.any_op
transform.yield
}
diff --git a/mlir/test/Dialect/Vector/vector-sink.mlir b/mlir/test/Dialect/Vector/vector-sink.mlir
index 8c8f1797aaab6..ad4fdbe0a7b5a 100644
--- a/mlir/test/Dialect/Vector/vector-sink.mlir
+++ b/mlir/test/Dialect/Vector/vector-sink.mlir
@@ -513,3 +513,192 @@ func.func @negative_extract_vec_fma(%arg0: vector<4xf32>, %arg1: vector<4xf32>,
%1 = vector.extract %0[1] : f32 from vector<4xf32>
return %1 : f32
}
+
+//-----------------------------------------------------------------------------
+// [Pattern: ExtractOpFromLoad]
+//-----------------------------------------------------------------------------
+
+// CHECK-LABEL: @extract_load_scalar
+// CHECK-SAME: (%[[ARG0:.*]]: memref<?xf32>, %[[ARG1:.*]]: index)
+func.func @extract_load_scalar(%arg0: memref<?xf32>, %arg1: index) -> f32 {
+// CHECK: %[[RES:.*]] = memref.load %[[ARG0]][%[[ARG1]]] : memref<?xf32>
+// CHECK: return %[[RES]] : f32
+ %0 = vector.load %arg0[%arg1] : memref<?xf32>, vector<4xf32>
+ %1 = vector.extract %0[0] : f32 from vector<4xf32>
+ return %1 : f32
+}
+
+// CHECK-LABEL: @extract_load_scalar_non_zero_off
+// CHECK-SAME: (%[[ARG0:.*]]: memref<?xf32>, %[[ARG1:.*]]: index)
+func.func @extract_load_scalar_non_zero_off(%arg0: memref<?xf32>, %arg1: index) -> f32 {
+// CHECK: %[[C1:.*]] = arith.constant 1 : index
+// CHECK: %[[OFF:.*]] = arith.addi %[[ARG1]], %[[C1]] overflow<nsw> : index
+// CHECK: %[[RES:.*]] = memref.load %[[ARG0]][%[[OFF]]] : memref<?xf32>
+// CHECK: return %[[RES]] : f32
+ %0 = vector.load %arg0[%arg1] : memref<?xf32>, vector<4xf32>
+ %1 = vector.extract %0[1] : f32 from vector<4xf32>
+ return %1 : f32
+}
+
+// CHECK-LABEL: @extract_load_scalar_dyn_off
+// CHECK-SAME: (%[[ARG0:.*]]: memref<?xf32>, %[[ARG1:.*]]: index, %[[ARG2:.*]]: index)
+func.func @extract_load_scalar_dyn_off(%arg0: memref<?xf32>, %arg1: index, %arg2: index) -> f32 {
+// CHECK: %[[OFF:.*]] = arith.addi %[[ARG1]], %[[ARG2]] overflow<nsw> : index
+// CHECK: %[[RES:.*]] = memref.load %[[ARG0]][%[[OFF]]] : memref<?xf32>
+// CHECK: return %[[RES]] : f32
+ %0 = vector.load %arg0[%arg1] : memref<?xf32>, vector<4xf32>
+ %1 = vector.extract %0[%arg2] : f32 from vector<4xf32>
+ return %1 : f32
+}
+
+// CHECK-LABEL: @extract_load_vec
+// CHECK-SAME: (%[[ARG0:.*]]: memref<?x?xf32>, %[[ARG1:.*]]: index, %[[ARG2:.*]]: index)
+func.func @extract_load_vec(%arg0: memref<?x?xf32>, %arg1: index, %arg2: index) -> vector<4xf32> {
+// CHECK: %[[C1:.*]] = arith.constant 1 : index
+// CHECK: %[[OFF:.*]] = arith.addi %[[ARG1]], %[[C1]] overflow<nsw> : index
+// CHECK: %[[RES:.*]] = vector.load %[[ARG0]][%[[OFF]], %[[ARG2]]] : memref<?x?xf32>, vector<4xf32>
+// CHECK: return %[[RES]] : vector<4xf32>
+ %0 = vector.load %arg0[%arg1, %arg2] : memref<?x?xf32>, vector<2x4xf32>
+ %1 = vector.extract %0[1] : vector<4xf32> from vector<2x4xf32>
+ return %1 : vector<4xf32>
+}
+
+// CHECK-LABEL: @extract_load_scalar_high_rank
+// CHECK-SAME: (%[[ARG0:.*]]: memref<?x?xf32>, %[[ARG1:.*]]: index, %[[ARG2:.*]]: index)
+func.func @extract_load_scalar_high_rank(%arg0: memref<?x?xf32>, %arg1: index, %arg2: index) -> f32 {
+// CHECK: %[[C1:.*]] = arith.constant 1 : index
+// CHECK: %[[OFF:.*]] = arith.addi %[[ARG2]], %[[C1]] overflow<nsw> : index
+// CHECK: %[[RES:.*]] = memref.load %[[ARG0]][%[[ARG1]], %[[OFF]]] : memref<?x?xf32>
+// CHECK: return %[[RES]] : f32
+ %0 = vector.load %arg0[%arg1, %arg2] : memref<?x?xf32>, vector<4xf32>
+ %1 = vector.extract %0[1] : f32 from vector<4xf32>
+ return %1 : f32
+}
+
+// CHECK-LABEL: @extract_load_vec_high_rank
+// CHECK-SAME: (%[[ARG0:.*]]: memref<?x?x?xf32>, %[[ARG1:.*]]: index, %[[ARG2:.*]]: index, %[[ARG3:.*]]: index)
+func.func @extract_load_vec_high_rank(%arg0: memref<?x?x?xf32>, %arg1: index, %arg2: index, %arg3: index) -> vector<4xf32> {
+// CHECK: %[[C1:.*]] = arith.constant 1 : index
+// CHECK: %[[OFF:.*]] = arith.addi %[[ARG2]], %[[C1]] overflow<nsw> : index
+// CHECK: %[[RES:.*]] = vector.load %[[ARG0]][%[[ARG1]], %[[OFF]], %[[ARG3]]] : memref<?x?x?xf32>, vector<4xf32>
+// CHECK: return %[[RES]] : vector<4xf32>
+ %0 = vector.load %arg0[%arg1, %arg2, %arg3] : memref<?x?x?xf32>, vector<2x4xf32>
+ %1 = vector.extract %0[1] : vector<4xf32> from vector<2x4xf32>
+ return %1 : vector<4xf32>
+}
+
+// CHECK-LABEL: @negative_load_scalar_from_vec_memref
+// CHECK-SAME: (%[[ARG0:.*]]: memref<?xvector<4xf32>>, %[[ARG1:.*]]: index)
+func.func @negative_load_scalar_from_vec_memref(%arg0: memref<?xvector<4xf32>>, %arg1: index) -> f32 {
+// CHECK: %[[RES:.*]] = vector.load %[[ARG0]][%[[ARG1]]] : memref<?xvector<4xf32>>, vector<4xf32>
+// CHECK: %[[EXT:.*]] = vector.extract %[[RES]][0] : f32 from vector<4xf32>
+// CHECK: return %[[EXT]] : f32
+ %0 = vector.load %arg0[%arg1] : memref<?xvector<4xf32>>, vector<4xf32>
+ %1 = vector.extract %0[0] : f32 from vector<4xf32>
+ return %1 : f32
+}
+
+// CHECK-LABEL: @negative_extract_load_no_single_use
+// CHECK-SAME: (%[[ARG0:.*]]: memref<?xf32>, %[[ARG1:.*]]: index)
+func.func @negative_extract_load_no_single_use(%arg0: memref<?xf32>, %arg1: index) -> (f32, vector<4xf32>) {
+// CHECK: %[[RES:.*]] = vector.load %[[ARG0]][%[[ARG1]]] : memref<?xf32>, vector<4xf32>
+// CHECK: %[[EXT:.*]] = vector.extract %[[RES]][0] : f32 from vector<4xf32>
+// CHECK: return %[[EXT]], %[[RES]] : f32, vector<4xf32>
+ %0 = vector.load %arg0[%arg1] : memref<?xf32>, vector<4xf32>
+ %1 = vector.extract %0[0] : f32 from vector<4xf32>
+ return %1, %0 : f32, vector<4xf32>
+}
+
+// CHECK-LABEL: @negative_load_scalable
+// CHECK-SAME: (%[[ARG0:.*]]: memref<?xf32>, %[[ARG1:.*]]: index)
+func.func @negative_load_scalable(%arg0: memref<?xf32>, %arg1: index) -> f32 {
+// CHECK: %[[RES:.*]] = vector.load %[[ARG0]][%[[ARG1]]] : memref<?xf32>, vector<[1]xf32>
+// CHECK: %[[EXT:.*]] = vector.extract %[[RES]][0] : f32 from vector<[1]xf32>
+// CHECK: return %[[EXT]] : f32
+ %0 = vector.load %arg0[%arg1] : memref<?xf32>, vector<[1]xf32>
+ %1 = vector.extract %0[0] : f32 from vector<[1]xf32>
+ return %1 : f32
+}
+
+// CHECK-LABEL: @negative_extract_load_unsupported_ranks
+// CHECK-SAME: (%[[ARG0:.*]]: memref<?xf32>, %[[ARG1:.*]]: index)
+func.func @negative_extract_load_unsupported_ranks(%arg0: memref<?xf32>, %arg1: index) -> vector<4xf32> {
+// CHECK: %[[RES:.*]] = vector.load %[[ARG0]][%[[ARG1]]] : memref<?xf32>, vector<2x4xf32>
+// CHECK: %[[EXT:.*]] = vector.extract %[[RES]][1] : vector<4xf32> from vector<2x4xf32>
+// CHECK: return %[[EXT]] : vector<4xf32>
+ %0 = vector.load %arg0[%arg1] : memref<?xf32>, vector<2x4xf32>
+ %1 = vector.extract %0[1] : vector<4xf32> from vector<2x4xf32>
+ return %1 : vector<4xf32>
+}
+
+//-----------------------------------------------------------------------------
+// [Pattern: StoreFromSplat]
+//-----------------------------------------------------------------------------
+
+// CHECK-LABEL: @store_splat
+// CHECK-SAME: (%[[ARG0:.*]]: memref<?xf32>, %[[ARG1:.*]]: index, %[[ARG2:.*]]: f32)
+func.func @store_splat(%arg0: memref<?xf32>, %arg1: index, %arg2: f32) {
+// CHECK: memref.store %[[ARG2]], %[[ARG0]][%[[ARG1]]] : memref<?xf32>
+ %0 = vector.splat %arg2 : vector<1xf32>
+ vector.store %0, %arg0[%arg1] : memref<?xf32>, vector<1xf32>
+ return
+}
+
+// CHECK-LABEL: @store_broadcast
+// CHECK-SAME: (%[[ARG0:.*]]: memref<?xf32>, %[[ARG1:.*]]: index, %[[ARG2:.*]]: f32)
+func.func @store_broadcast(%arg0: memref<?xf32>, %arg1: index, %arg2: f32) {
+// CHECK: memref.store %[[ARG2]], %[[ARG0]][%[[ARG1]]] : memref<?xf32>
+ %0 = vector.broadcast %arg2 : f32 to vector<1xf32>
+ vector.store %0, %arg0[%arg1] : memref<?xf32>, vector<1xf32>
+ return
+}
+
+// CHECK-LABEL: @store_broadcast_1d_2d
+// CHECK-SAME: (%[[ARG0:.*]]: memref<?x?xf32>, %[[ARG1:.*]]: index, %[[ARG2:.*]]: index, %[[ARG3:.*]]: vector<1xf32>)
+func.func @store_broadcast_1d_2d(%arg0: memref<?x?xf32>, %arg1: index, %arg2: index, %arg3: vector<1xf32>) {
+// CHECK: vector.store %[[ARG3]], %[[ARG0]][%[[ARG1]], %[[ARG2]]] : memref<?x?xf32>, vector<1xf32>
+ %0 = vector.broadcast %arg3 : vector<1xf32> to vector<1x1xf32>
+ vector.store %0, %arg0[%arg1, %arg2] : memref<?x?xf32>, vector<1x1xf32>
+ return
+}
+
+// CHECK-LABEL: @negative_store_scalable
+// CHECK-SAME: (%[[ARG0:.*]]: memref<?xf32>, %[[ARG1:.*]]: index, %[[ARG2:.*]]: f32)
+func.func @negative_store_scalable(%arg0: memref<?xf32>, %arg1: index, %arg2: f32) {
+// CHECK: %[[RES:.*]] = vector.splat %[[ARG2]] : vector<[1]xf32>
+// CHECK: vector.store %[[RES]], %[[ARG0]][%[[ARG1]]] : memref<?xf32>, vector<[1]xf32>
+ %0 = vector.splat %arg2 : vector<[1]xf32>
+ vector.store %0, %arg0[%arg1] : memref<?xf32>, vector<[1]xf32>
+ return
+}
+
+// CHECK-LABEL: @negative_store_vec_memref
+// CHECK-SAME: (%[[ARG0:.*]]: memref<?xvector<1xf32>>, %[[ARG1:.*]]: index, %[[ARG2:.*]]: f32)
+func.func @negative_store_vec_memref(%arg0: memref<?xvector<1xf32>>, %arg1: index, %arg2: f32) {
+// CHECK: %[[RES:.*]] = vector.splat %[[ARG2]] : vector<1xf32>
+// CHECK: vector.store %[[RES]], %[[ARG0]][%[[ARG1]]] : memref<?xvector<1xf32>>, vector<1xf32>
+ %0 = vector.splat %arg2 : vector<1xf32>
+ vector.store %0, %arg0[%arg1] : memref<?xvector<1xf32>>, vector<1xf32>
+ return
+}
+
+// CHECK-LABEL: @negative_store_non_1
+// CHECK-SAME: (%[[ARG0:.*]]: memref<?xf32>, %[[ARG1:.*]]: index, %[[ARG2:.*]]: f32)
+func.func @negative_store_non_1(%arg0: memref<?xf32>, %arg1: index, %arg2: f32) {
+// CHECK: %[[RES:.*]] = vector.splat %[[ARG2]] : vector<4xf32>
+// CHECK: vector.store %[[RES]], %[[ARG0]][%[[ARG1]]] : memref<?xf32>, vector<4xf32>
+ %0 = vector.splat %arg2 : vector<4xf32>
+ vector.store %0, %arg0[%arg1] : memref<?xf32>, vector<4xf32>
+ return
+}
+
+// CHECK-LABEL: @negative_store_no_single_use
+// CHECK-SAME: (%[[ARG0:.*]]: memref<?xf32>, %[[ARG1:.*]]: index, %[[ARG2:.*]]: f32)
+func.func @negative_store_no_single_use(%arg0: memref<?xf32>, %arg1: index, %arg2: f32) -> vector<1xf32> {
+// CHECK: %[[RES:.*]] = vector.splat %[[ARG2]] : vector<1xf32>
+// CHECK: vector.store %[[RES]], %[[ARG0]][%[[ARG1]]] : memref<?xf32>, vector<1xf32>
+// CHECK: return %[[RES:.*]] : vector<1xf32>
+ %0 = vector.splat %arg2 : vector<1xf32>
+ vector.store %0, %arg0[%arg1] : memref<?xf32>, vector<1xf32>
+ return %0 : vector<1xf32>
+}
diff --git a/mlir/test/lib/Dialect/Vector/TestVectorTransforms.cpp b/mlir/test/lib/Dialect/Vector/TestVectorTransforms.cpp
index a54ae816570a8..03f907e46c2c6 100644
--- a/mlir/test/lib/Dialect/Vector/TestVectorTransforms.cpp
+++ b/mlir/test/lib/Dialect/Vector/TestVectorTransforms.cpp
@@ -395,6 +395,7 @@ struct TestVectorSinkPatterns
void runOnOperation() override {
RewritePatternSet patterns(&getContext());
populateSinkVectorOpsPatterns(patterns);
+ populateSinkVectorMemOpsPatterns(patterns);
(void)applyPatternsGreedily(getOperation(), std::move(patterns));
}
};
>From 6363446be1ee87e1ba12076cdce28762774607f1 Mon Sep 17 00:00:00 2001
From: Ivan Butygin <ivan.butygin at gmail.com>
Date: Fri, 4 Apr 2025 16:38:13 +0200
Subject: [PATCH 2/2] comment
---
.../Vector/Transforms/VectorRewritePatterns.h | 13 +++++++++++--
1 file changed, 11 insertions(+), 2 deletions(-)
diff --git a/mlir/include/mlir/Dialect/Vector/Transforms/VectorRewritePatterns.h b/mlir/include/mlir/Dialect/Vector/Transforms/VectorRewritePatterns.h
index 7d3134bdae233..2d8b12c871be7 100644
--- a/mlir/include/mlir/Dialect/Vector/Transforms/VectorRewritePatterns.h
+++ b/mlir/include/mlir/Dialect/Vector/Transforms/VectorRewritePatterns.h
@@ -161,8 +161,17 @@ void populateVectorTransferCollapseInnerMostContiguousDimsPatterns(
void populateSinkVectorOpsPatterns(RewritePatternSet &patterns,
PatternBenefit benefit = 1);
-/// Patterns that remove redundant Vector Ops by re-ordering them with
-/// memory Ops:
+/// Patterns that remove redundant Vector Ops by merging them with load/store
+/// ops
+/// ```
+/// vector.load %arg0[%arg1] : memref<?xf32>, vector<4xf32>
+/// vector.extract %0[1] : f32 from vector<4xf32>
+/// ```
+/// Gets converted to:
+/// ```
+/// %c1 = arith.constant 1 : index
+/// %0 = arith.addi %arg1, %c1 overflow<nsw> : index
+/// %1 = memref.load %arg0[%0] : memref<?xf32>
void populateSinkVectorMemOpsPatterns(RewritePatternSet &patterns,
PatternBenefit benefit = 1);
More information about the Mlir-commits
mailing list