[mlir][vector] Add unroll patterns for vector.load and vector.store

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

@@ -54,6 +54,33 @@ static SmallVector<Value> sliceTransferIndices(ArrayRef<int64_t> elementOffsets,
   return slicedIndices;
 }
 
+// compute the new indices for vector.load/store by adding offsets to
+// originalIndices.
+// It assumes m <= n (m = offsets.size(), n = originalIndices.size())
+// Last m of originalIndices will be updated.
+static SmallVector<Value> computeIndices(PatternRewriter &rewriter,
+                                         Location loc,
+                                         ArrayRef<Value> originalIndices,
+                                         ArrayRef<int64_t> offsets) {
+  assert(offsets.size() <= originalIndices.size() &&
+         "Offsets should not exceed the number of original indices");
+  SmallVector<Value> indices(originalIndices);
+  auto originalIter = originalIndices.rbegin();
+  auto offsetsIter = offsets.rbegin();
+  auto indicesIter = indices.rbegin();
+  while (offsetsIter != offsets.rend()) {
+    Value original = *originalIter;
+    int64_t offset = *offsetsIter;
+    if (offset != 0)
+      *indicesIter = rewriter.create<arith::AddIOp>(
+          loc, original, rewriter.create<arith::ConstantIndexOp>(loc, offset));
+    originalIter++;
+    offsetsIter++;
+    indicesIter++;
+  }
+  return indices;
+};
+
 // Clones `op` into a new operations that takes `operands` and returns
 // `resultTypes`.
 static Operation *cloneOpWithOperandsAndTypes(OpBuilder &builder, Location loc,
@@ -631,6 +658,98 @@ struct UnrollGatherPattern : public OpRewritePattern<vector::GatherOp> {
   vector::UnrollVectorOptions options;
 };
 
+struct UnrollLoadPattern : public OpRewritePattern<vector::LoadOp> {
+  UnrollLoadPattern(MLIRContext *context,
+                    const vector::UnrollVectorOptions &options,
+                    PatternBenefit benefit = 1)
+      : OpRewritePattern<vector::LoadOp>(context, benefit), options(options) {}
+
+  LogicalResult matchAndRewrite(vector::LoadOp loadOp,
+                                PatternRewriter &rewriter) const override {
+    VectorType vecType = loadOp.getVectorType();
+    // Only unroll >1D loads
+    if (vecType.getRank() <= 1)
+      return failure();
+
+    Location loc = loadOp.getLoc();
+    ArrayRef<int64_t> originalShape = vecType.getShape();
+
+    // Target type is a 1D vector of the innermost dimension.
+    auto targetType =
+        VectorType::get(originalShape.back(), vecType.getElementType());
+
+    // Extend the targetShape to the same rank of original shape by padding 1s
+    // for leading dimensions for convenience of computing offsets
+    SmallVector<int64_t> targetShape(originalShape.size(), 1);
+    targetShape.back() = originalShape.back();
+
+    Value result = rewriter.create<arith::ConstantOp>(
+        loc, vecType, rewriter.getZeroAttr(vecType));
+
+    SmallVector<Value> originalIndices(loadOp.getIndices().begin(),
+                                       loadOp.getIndices().end());
+
+    for (SmallVector<int64_t> offsets :
+         StaticTileOffsetRange(originalShape, targetShape)) {
+      SmallVector<Value> indices =
+          computeIndices(rewriter, loc, originalIndices, offsets);
+      Value slice = rewriter.create<vector::LoadOp>(loc, targetType,
+                                                    loadOp.getBase(), indices);
+      // Insert the slice into the result at the correct position.
+      result = rewriter.createOrFold<vector::InsertStridedSliceOp>(
+          loc, slice, result, offsets, SmallVector<int64_t>({1}));
+    }
+    rewriter.replaceOp(loadOp, result);
+    return success();
+  }
+
+private:
+  vector::UnrollVectorOptions options;
+};
+
+struct UnrollStorePattern : public OpRewritePattern<vector::StoreOp> {
+  UnrollStorePattern(MLIRContext *context,
+                     const vector::UnrollVectorOptions &options,
+                     PatternBenefit benefit = 1)
+      : OpRewritePattern<vector::StoreOp>(context, benefit), options(options) {}
+
+  LogicalResult matchAndRewrite(vector::StoreOp storeOp,
+                                PatternRewriter &rewriter) const override {
+    VectorType vecType = storeOp.getVectorType();
+    // Only unroll >1D stores.
+    if (vecType.getRank() <= 1)
+      return failure();
+
+    Location loc = storeOp.getLoc();
+    ArrayRef<int64_t> originalShape = vecType.getShape();
+
+    // Extend the targetShape to the same rank of original shape by padding 1s
+    // for leading dimensions for convenience of computing offsets
+    SmallVector<int64_t> targetShape(originalShape.size(), 1);
+    targetShape.back() = originalShape.back();
+
+    Value base = storeOp.getBase();
+    Value vector = storeOp.getValueToStore();
+
+    SmallVector<Value> originalIndices(storeOp.getIndices().begin(),
+                                       storeOp.getIndices().end());
+
+    for (SmallVector<int64_t> offsets :
+         StaticTileOffsetRange(originalShape, targetShape)) {
+      SmallVector<Value> indices =
+          computeIndices(rewriter, loc, originalIndices, offsets);
+      offsets.pop_back();
+      Value slice = rewriter.create<vector::ExtractOp>(loc, vector, offsets);
+      rewriter.create<vector::StoreOp>(loc, slice, base, indices);
+    }
+    rewriter.eraseOp(storeOp);
+    return success();
+  }
+
+private:
+  vector::UnrollVectorOptions options;
+};
+
 struct UnrollBroadcastPattern : public OpRewritePattern<vector::BroadcastOp> {
   UnrollBroadcastPattern(MLIRContext *context,
                          const vector::UnrollVectorOptions &options,
@@ -699,10 +818,10 @@ struct UnrollBroadcastPattern : public OpRewritePattern<vector::BroadcastOp> {
 void mlir::vector::populateVectorUnrollPatterns(
     RewritePatternSet &patterns, const UnrollVectorOptions &options,
     PatternBenefit benefit) {
-  patterns
-      .add<UnrollTransferReadPattern, UnrollTransferWritePattern,
-           UnrollContractionPattern, UnrollElementwisePattern,
-           UnrollReductionPattern, UnrollMultiReductionPattern,
-           UnrollTransposePattern, UnrollGatherPattern, UnrollBroadcastPattern>(
-          patterns.getContext(), options, benefit);
+  patterns.add<UnrollTransferReadPattern, UnrollTransferWritePattern,
+               UnrollContractionPattern, UnrollElementwisePattern,
+               UnrollReductionPattern, UnrollMultiReductionPattern,
+               UnrollTransposePattern, UnrollGatherPattern, UnrollLoadPattern,
+               UnrollStorePattern, UnrollBroadcastPattern>(
+      patterns.getContext(), options, benefit);
 }

mlir/test/Dialect/Vector/vector-unroll-options.mlir

@@ -378,3 +378,76 @@ func.func @vector_broadcast_with_tailing_unit_dim(%v: vector<4x1xf32>) -> vector
 //       CHECK: [[b3:%.+]] = vector.broadcast [[s3]] : vector<2x1xf32> to vector<2x2xf32>
 //       CHECK: [[r3:%.+]] = vector.insert_strided_slice [[b3]], [[r2]] {offsets = [2, 2], strides = [1, 1]} : vector<2x2xf32> into vector<4x4xf32>
 //       CHECK: return [[r3]] : vector<4x4xf32>
+
+
+// CHECK-LABEL: func.func @unroll_2D_vector_load(
+// CHECK-SAME:  %[[ARG:.*]]: memref<4x4xf16>) -> vector<4x4xf16> {
+func.func @unroll_2D_vector_load(%arg0: memref<4x4xf16>) -> vector<4x4xf16> {
+  // CHECK: %[[C3:.*]] = arith.constant 3 : index
+  // CHECK: %[[C2:.*]] = arith.constant 2 : index
+  // CHECK: %[[C1:.*]] = arith.constant 1 : index
+  // CHECK: %[[C0:.*]] = arith.constant 0 : index
+  // CHECK: %[[CST:.*]] = arith.constant dense<0.000000e+00> : vector<4x4xf16>
+  // CHECK: %[[V0:.*]] = vector.load %[[ARG]][%[[C0]], %[[C0]]] : memref<4x4xf16>, vector<4xf16>
+  // CHECK: %[[V1:.*]] = vector.insert_strided_slice %[[V0]], %[[CST]] {offsets = [0, 0], strides = [1]} : vector<4xf16> into vector<4x4xf16>
+  // CHECK: %[[V2:.*]] = vector.load %[[ARG]][%[[C1]], %[[C0]]] : memref<4x4xf16>, vector<4xf16>
+  // CHECK: %[[V3:.*]] = vector.insert_strided_slice %[[V2]], %[[V1]] {offsets = [1, 0], strides = [1]} : vector<4xf16> into vector<4x4xf16>
+  // CHECK: %[[V4:.*]] = vector.load %[[ARG]][%[[C2]], %[[C0]]] : memref<4x4xf16>, vector<4xf16>
+  // CHECK: %[[V5:.*]] = vector.insert_strided_slice %[[V4]], %[[V3]] {offsets = [2, 0], strides = [1]} : vector<4xf16> into vector<4x4xf16>
+  // CHECK: %[[V6:.*]] = vector.load %[[ARG]][%[[C3]], %[[C0]]] : memref<4x4xf16>, vector<4xf16>
+  // CHECK: %[[V7:.*]] = vector.insert_strided_slice %[[V6]], %[[V5]] {offsets = [3, 0], strides = [1]} : vector<4xf16> into vector<4x4xf16>
+  // CHECK: return %[[V7]] : vector<4x4xf16>
+  %c0 = arith.constant 0 : index
+  %0 = vector.load %arg0[%c0, %c0] : memref<4x4xf16>, vector<4x4xf16>
+  return %0 : vector<4x4xf16>
+}
+
+// CHECK-LABEL: func.func @unroll_2D_vector_store(
+// CHECK-SAME:  %[[ARG0:.*]]: memref<4x4xf16>, %[[ARG1:.*]]: vector<4x4xf16>) {
+func.func @unroll_2D_vector_store(%arg0: memref<4x4xf16>, %arg1: vector<4x4xf16>) {
+  // CHECK: %[[C3:.*]] = arith.constant 3 : index
+  // CHECK: %[[C2:.*]] = arith.constant 2 : index
+  // CHECK: %[[C1:.*]] = arith.constant 1 : index
+  // CHECK: %[[C0:.*]] = arith.constant 0 : index
+  // CHECK: %[[V0:.*]] = vector.extract %[[ARG1]][0] : vector<4xf16> from vector<4x4xf16>
+  // CHECK: vector.store %[[V0]], %[[ARG0]][%[[C0]], %[[C0]]] : memref<4x4xf16>, vector<4xf16>
+  // CHECK: %[[V1:.*]] = vector.extract %[[ARG1]][1] : vector<4xf16> from vector<4x4xf16>
+  // CHECK: vector.store %[[V1]], %[[ARG0]][%[[C1]], %[[C0]]] : memref<4x4xf16>, vector<4xf16>
+  // CHECK: %[[V2:.*]] = vector.extract %[[ARG1]][2] : vector<4xf16> from vector<4x4xf16>
+  // CHECK: vector.store %[[V2]], %[[ARG0]][%[[C2]], %[[C0]]] : memref<4x4xf16>, vector<4xf16>
+  // CHECK: %[[V3:.*]] = vector.extract %[[ARG1]][3] : vector<4xf16> from vector<4x4xf16>
+  // CHECK: vector.store %[[V3]], %[[ARG0]][%[[C3]], %[[C0]]] : memref<4x4xf16>, vector<4xf16>
+  %c0 = arith.constant 0 : index
+  vector.store %arg1, %arg0[%c0, %c0] : memref<4x4xf16>, vector<4x4xf16>
+  return
+}
+
+// CHECK-LABEL: func.func @unroll_vector_load(
+// CHECK-SAME:  %[[ARG:.*]]: memref<4x4x4x4xf16>) -> vector<2x2xf16> {
+func.func @unroll_vector_load(%arg0: memref<4x4x4x4xf16>) -> vector<2x2xf16> {
+  // CHECK: %[[C2:.*]] = arith.constant 2 : index
+  // CHECK: %[[C1:.*]] = arith.constant 1 : index
+  // CHECK: %[[CST:.*]] = arith.constant dense<0.000000e+00> : vector<2x2xf16>
+  // CHECK: %[[V0:.*]] = vector.load %[[ARG]][%[[C1]], %[[C1]], %[[C1]], %[[C1]]] : memref<4x4x4x4xf16>, vector<2xf16>
+  // CHECK: %[[V1:.*]] = vector.insert_strided_slice %[[V0]], %[[CST]] {offsets = [0, 0], strides = [1]} : vector<2xf16> into vector<2x2xf16>
+  // CHECK: %[[V2:.*]] = vector.load %[[ARG]][%[[C1]], %[[C1]], %[[C2]], %[[C1]]] : memref<4x4x4x4xf16>, vector<2xf16>
+  // CHECK: %[[V3:.*]] = vector.insert_strided_slice %[[V2]], %[[V1]] {offsets = [1, 0], strides = [1]} : vector<2xf16> into vector<2x2xf16>
+  // CHECK: return %[[V3]] : vector<2x2xf16>
+  %c1 = arith.constant 1 : index
+  %0 = vector.load %arg0[%c1, %c1, %c1, %c1] : memref<4x4x4x4xf16>, vector<2x2xf16>
+  return %0 : vector<2x2xf16>
+}
+
+// CHECK-LABEL: func.func @unroll_vector_store(
+// CHECK-SAME:  %[[ARG0:.*]]: memref<4x4x4x4xf16>, %[[ARG1:.*]]: vector<2x2xf16>) {
+func.func @unroll_vector_store(%arg0: memref<4x4x4x4xf16>, %arg1: vector<2x2xf16>) {
+  // CHECK: %[[C2:.*]] = arith.constant 2 : index
+  // CHECK: %[[C1:.*]] = arith.constant 1 : index
+  // CHECK: %[[V0:.*]] = vector.extract %[[ARG1]][0] : vector<2xf16> from vector<2x2xf16>
+  // CHECK: vector.store %[[V0]], %[[ARG0]][%[[C1]], %[[C1]], %[[C1]], %[[C1]]] : memref<4x4x4x4xf16>, vector<2xf16>
+  // CHECK: %[[V1:.*]] = vector.extract %[[ARG1]][1] : vector<2xf16> from vector<2x2xf16>
+  // CHECK: vector.store %[[V1]], %[[ARG0]][%[[C1]], %[[C1]], %[[C2]], %[[C1]]] : memref<4x4x4x4xf16>, vector<2xf16>
+  %c1 = arith.constant 1 : index
+  vector.store %arg1, %arg0[%c1, %c1, %c1, %c1] : memref<4x4x4x4xf16>, vector<2x2xf16>
+  return
+}

mlir/test/lib/Dialect/Vector/TestVectorTransforms.cpp

@@ -178,6 +178,16 @@ struct TestVectorUnrollingPatterns
                         return success(isa<vector::TransposeOp>(op));
                       }));
 
+    populateVectorUnrollPatterns(
+        patterns, UnrollVectorOptions()
+                      .setNativeShape(ArrayRef<int64_t>{2, 2})
+                      .setFilterConstraint([](Operation *op) {
+                        if (auto loadOp = dyn_cast<vector::LoadOp>(op))
+                          return success(loadOp.getType().getRank() > 1);
+                        if (auto storeOp = dyn_cast<vector::StoreOp>(op))
+                          return success(storeOp.getVectorType().getRank() > 1);
+                        return failure();
+                      }));
     if (unrollBasedOnType) {
       UnrollVectorOptions::NativeShapeFnType nativeShapeFn =
           [](Operation *op) -> std::optional<SmallVector<int64_t>> {