[Mlir-commits] [mlir] f397bdf - [mlir][tensor] Fold consumer linalg transpose with producer tensor pack (#74206)
llvmlistbot at llvm.org
llvmlistbot at llvm.org
Wed Dec 13 14:26:22 PST 2023
Author: Prathamesh Tagore
Date: 2023-12-13T14:26:19-08:00
New Revision: f397bdf5aee331d984d5e41ed39a6834ec9fe0c5
URL: https://github.com/llvm/llvm-project/commit/f397bdf5aee331d984d5e41ed39a6834ec9fe0c5
DIFF: https://github.com/llvm/llvm-project/commit/f397bdf5aee331d984d5e41ed39a6834ec9fe0c5.diff
LOG: [mlir][tensor] Fold consumer linalg transpose with producer tensor pack (#74206)
Partial fix to https://github.com/openxla/iree/issues/15367
Added:
Modified:
mlir/lib/Dialect/Tensor/Transforms/FoldIntoPackAndUnpackPatterns.cpp
mlir/test/Dialect/Tensor/fold-into-pack-and-unpack.mlir
Removed:
################################################################################
diff --git a/mlir/lib/Dialect/Tensor/Transforms/FoldIntoPackAndUnpackPatterns.cpp b/mlir/lib/Dialect/Tensor/Transforms/FoldIntoPackAndUnpackPatterns.cpp
index 9eac3e5c7ef910..e4509b331beeac 100644
--- a/mlir/lib/Dialect/Tensor/Transforms/FoldIntoPackAndUnpackPatterns.cpp
+++ b/mlir/lib/Dialect/Tensor/Transforms/FoldIntoPackAndUnpackPatterns.cpp
@@ -6,6 +6,7 @@
//
//===----------------------------------------------------------------------===//
+#include "mlir/Dialect/Linalg/IR/Linalg.h"
#include "mlir/Dialect/Tensor/IR/Tensor.h"
#include "mlir/Dialect/Tensor/Transforms/Transforms.h"
#include "mlir/IR/PatternMatch.h"
@@ -81,10 +82,71 @@ struct FoldUnpackWithExtractSliceOp : public OpRewritePattern<ExtractSliceOp> {
return success();
}
};
+
+/// Fold 'pack' -> 'transpose' into 'pack' since 'pack' already has transpose
+/// semantics.
+struct FoldProducerPackWithConsumerLinalgTransposeOp
+ : public OpRewritePattern<linalg::TransposeOp> {
+ using OpRewritePattern<linalg::TransposeOp>::OpRewritePattern;
+
+ LogicalResult matchAndRewrite(linalg::TransposeOp transposeOp,
+ PatternRewriter &rewriter) const override {
+ auto packOp = transposeOp.getOperand(0).getDefiningOp<PackOp>();
+
+ if (!packOp)
+ return failure();
+
+ auto innerDimsPos = packOp.getInnerDimsPos();
+ auto mixedInnerTiles = packOp.getMixedTiles();
+ auto outerDimsPerm = packOp.getOuterDimsPerm();
+ auto transposePerm = transposeOp.getPermutation();
+ SmallVector<int64_t> newOuterDimsPermVec;
+ SmallVector<int64_t> newInnerDimsPosVec;
+ SmallVector<OpFoldResult> newMixedInnerTilesVec;
+ int64_t srcRank = packOp.getSourceRank();
+
+ // Process transpose operation for non-tiled outer dimensions
+ for (unsigned int i = 0; i < srcRank; ++i) {
+ int64_t remappedPosition = transposePerm[i];
+
+ // If tensor.pack has outer_dims_perm attribute, then consider it during
+ // index remapping.
+ if (!outerDimsPerm.empty()) {
+ if (transposePerm[i] >= srcRank) {
+ return rewriter.notifyMatchFailure(
+ transposeOp,
+ "Cannot fold in tensor.pack if a tile dimension was transposed "
+ "with a non-tile dimension in linalg.transpose.");
+ }
+ remappedPosition = outerDimsPerm[remappedPosition];
+ }
+
+ newOuterDimsPermVec.push_back(remappedPosition);
+ }
+
+ // Process transpose operation for tiled inner dimensions
+ for (unsigned int i = srcRank; i < transposePerm.size(); ++i) {
+ int64_t remappedPosition = transposePerm[i] - srcRank;
+ newMixedInnerTilesVec.push_back(mixedInnerTiles[remappedPosition]);
+ newInnerDimsPosVec.push_back(innerDimsPos[remappedPosition]);
+ }
+
+ Value output = packOp.createDestinationTensor(
+ rewriter, transposeOp.getLoc(), packOp.getSource(),
+ newMixedInnerTilesVec, newInnerDimsPosVec, newOuterDimsPermVec);
+
+ rewriter.replaceOpWithNewOp<PackOp>(
+ transposeOp, packOp.getSource(), output, newInnerDimsPosVec,
+ newMixedInnerTilesVec, packOp.getPaddingValue(), newOuterDimsPermVec);
+
+ return success();
+ }
+};
} // namespace
void populateFoldIntoPackAndUnpackPatterns(RewritePatternSet &patterns) {
- patterns.insert<FoldUnpackWithExtractSliceOp, FoldPadWithPackOp>(
+ patterns.insert<FoldUnpackWithExtractSliceOp, FoldPadWithPackOp,
+ FoldProducerPackWithConsumerLinalgTransposeOp>(
patterns.getContext());
}
diff --git a/mlir/test/Dialect/Tensor/fold-into-pack-and-unpack.mlir b/mlir/test/Dialect/Tensor/fold-into-pack-and-unpack.mlir
index 5c757896657427..ca4eb4ff679445 100644
--- a/mlir/test/Dialect/Tensor/fold-into-pack-and-unpack.mlir
+++ b/mlir/test/Dialect/Tensor/fold-into-pack-and-unpack.mlir
@@ -114,3 +114,234 @@ func.func @pad_pack_
diff erent_padding_value(%src: tensor<16641x16xf32>) -> tenso
// CHECK-LABEL: func.func @pad_pack_
diff erent_padding_value
// CHECK: tensor.pad
// CHECK: tensor.pack
+
+// -----
+
+func.func @tensor_pack_linalg_transpose_fold(%arg0: tensor<56x57x1x64xf32>) -> tensor<1x57x56x2x32xf32> {
+ %0 = tensor.empty() : tensor<56x2x1x57x32xf32>
+ %pack = tensor.pack %arg0
+ outer_dims_perm = [0, 3, 2, 1]
+ inner_dims_pos = [3]
+ inner_tiles = [32]
+ into %0 : tensor<56x57x1x64xf32> -> tensor<56x2x1x57x32xf32>
+
+ %1 = tensor.empty() : tensor<1x57x56x2x32xf32>
+ %transposed = linalg.transpose
+ ins(%pack : tensor<56x2x1x57x32xf32>)
+ outs(%1 : tensor<1x57x56x2x32xf32>)
+ permutation = [2, 3, 0, 1, 4]
+ return %transposed : tensor<1x57x56x2x32xf32>
+}
+// CHECK: func @tensor_pack_linalg_transpose_fold(
+// CHECK-SAME: %[[ARG0:.+]]: tensor<56x57x1x64xf32>)
+// CHECK: %[[INIT:.+]] = tensor.empty() : tensor<1x57x56x2x32xf32>
+// CHECK: %[[PACK:.+]] = tensor.pack %[[ARG0]]
+// CHECK-SAME: outer_dims_perm = [2, 1, 0, 3]
+// CHECK-SAME: inner_dims_pos = [3] inner_tiles = [32]
+// CHECK-SAME: into %[[INIT]]
+// CHECK: return %[[PACK]]
+
+// -----
+
+func.func @tensor_pack_linalg_transpose_fold_with_padding(%arg0: tensor<56x57x1x55xf32>, %padding: f32) -> tensor<1x57x56x2x32xf32> {
+ %0 = tensor.empty() : tensor<56x2x1x57x32xf32>
+ %pack = tensor.pack %arg0 padding_value(%padding : f32)
+ outer_dims_perm = [0, 3, 2, 1]
+ inner_dims_pos = [3]
+ inner_tiles = [32]
+ into %0 : tensor<56x57x1x55xf32> -> tensor<56x2x1x57x32xf32>
+
+ %1 = tensor.empty() : tensor<1x57x56x2x32xf32>
+ %transposed = linalg.transpose
+ ins(%pack : tensor<56x2x1x57x32xf32>)
+ outs(%1 : tensor<1x57x56x2x32xf32>)
+ permutation = [2, 3, 0, 1, 4]
+ return %transposed : tensor<1x57x56x2x32xf32>
+}
+// CHECK: func @tensor_pack_linalg_transpose_fold_with_padding(
+// CHECK-SAME: %[[ARG0:.+]]: tensor<56x57x1x55xf32>, %[[PADDING:.+]]: f32)
+// CHECK: %[[INIT:.+]] = tensor.empty() : tensor<1x57x56x2x32xf32>
+// CHECK: %[[PACK:.+]] = tensor.pack %[[ARG0]] padding_value(%[[PADDING]] : f32)
+// CHECK-SAME: outer_dims_perm = [2, 1, 0, 3]
+// CHECK-SAME: inner_dims_pos = [3] inner_tiles = [32]
+// CHECK-SAME: into %[[INIT]]
+// CHECK: return %[[PACK]]
+
+// -----
+
+func.func @tensor_pack_linalg_transpose_fold_no_outer_dims_perm(%arg0: tensor<56x57x1x64xf32>) -> tensor<1x2x56x57x32xf32> {
+ %0 = tensor.empty() : tensor<56x57x1x2x32xf32>
+ %pack = tensor.pack %arg0
+ inner_dims_pos = [3]
+ inner_tiles = [32]
+ into %0 : tensor<56x57x1x64xf32> -> tensor<56x57x1x2x32xf32>
+
+ %1 = tensor.empty() : tensor<1x2x56x57x32xf32>
+ %transposed = linalg.transpose
+ ins(%pack : tensor<56x57x1x2x32xf32>)
+ outs(%1 : tensor<1x2x56x57x32xf32>)
+ permutation = [2, 3, 0, 1, 4]
+ return %transposed : tensor<1x2x56x57x32xf32>
+}
+// CHECK: func @tensor_pack_linalg_transpose_fold_no_outer_dims_perm(
+// CHECK-SAME: %[[ARG0:.+]]: tensor<56x57x1x64xf32>)
+// CHECK: %[[INIT:.+]] = tensor.empty() : tensor<1x2x56x57x32xf32>
+// CHECK: %[[PACK:.+]] = tensor.pack %[[ARG0]]
+// CHECK-SAME: outer_dims_perm = [2, 3, 0, 1]
+// CHECK-SAME: inner_dims_pos = [3] inner_tiles = [32]
+// CHECK-SAME: into %[[INIT]]
+// CHECK: return %[[PACK]]
+
+// -----
+
+func.func @tensor_pack_linalg_transpose_fold_tile_dims_transpose(%arg0: tensor<56x72x24x128xf32>) -> tensor<12x56x4x9x32x8x2xf32> {
+ %0 = tensor.empty() : tensor<4x9x12x56x8x2x32xf32>
+ %pack = tensor.pack %arg0
+ outer_dims_perm = [3, 1, 2, 0]
+ inner_dims_pos = [1, 2, 3]
+ inner_tiles = [8, 2, 32]
+ into %0 : tensor<56x72x24x128xf32> -> tensor<4x9x12x56x8x2x32xf32>
+
+ %1 = tensor.empty() : tensor<12x56x4x9x32x8x2xf32>
+ %transposed = linalg.transpose
+ ins(%pack : tensor<4x9x12x56x8x2x32xf32>)
+ outs(%1 : tensor<12x56x4x9x32x8x2xf32>)
+ permutation = [2, 3, 0, 1, 6, 4, 5]
+ return %transposed : tensor<12x56x4x9x32x8x2xf32>
+}
+// CHECK: func @tensor_pack_linalg_transpose_fold_tile_dims_transpose(
+// CHECK-SAME: %[[ARG0:.+]]: tensor<56x72x24x128xf32>)
+// CHECK: %[[INIT:.+]] = tensor.empty() : tensor<12x56x4x9x32x8x2xf32>
+// CHECK: %[[PACK:.+]] = tensor.pack %[[ARG0]]
+// CHECK-SAME: outer_dims_perm = [2, 0, 3, 1]
+// CHECK-SAME: inner_dims_pos = [3, 1, 2] inner_tiles = [32, 8, 2]
+// CHECK-SAME: into %[[INIT]]
+// CHECK: return %[[PACK]]
+
+// -----
+
+func.func @tensor_pack_linalg_transpose_fold_tile_dims_outer_dims_transpose(%arg0: tensor<56x72x24x128xf32>) -> tensor<9x56x2x12x32x8x4xf32> {
+ %0 = tensor.empty() : tensor<4x12x9x56x8x2x32xf32>
+ %pack = tensor.pack %arg0
+ outer_dims_perm = [3, 2, 1, 0]
+ inner_dims_pos = [1, 2, 3]
+ inner_tiles = [8, 2, 32]
+ into %0 : tensor<56x72x24x128xf32> -> tensor<4x12x9x56x8x2x32xf32>
+
+ %1 = tensor.empty() : tensor<9x56x2x12x32x8x4xf32>
+ %transposed = linalg.transpose
+ ins(%pack : tensor<4x12x9x56x8x2x32xf32>)
+ outs(%1 : tensor<9x56x2x12x32x8x4xf32>)
+ permutation = [2, 3, 5, 1, 6, 4, 0]
+ return %transposed : tensor<9x56x2x12x32x8x4xf32>
+}
+// CHECK: func @tensor_pack_linalg_transpose_fold_tile_dims_outer_dims_transpose(
+// CHECK-SAME: %[[ARG0:.+]]: tensor<56x72x24x128xf32>)
+// CHECK: tensor.pack
+// CHECK: linalg.transpose
+
+// -----
+
+func.func @tensor_pack_linalg_transpose_fold_dynamic_outer_dims(%arg0: tensor<56x?x?x64xf32>) -> tensor<?x?x56x2x32xf32> {
+ %0 = tensor.empty() : tensor<56x2x1x57x32xf32>
+ %pack = tensor.pack %arg0
+ outer_dims_perm = [0, 3, 2, 1]
+ inner_dims_pos = [3]
+ inner_tiles = [32]
+ into %0 : tensor<56x?x?x64xf32> -> tensor<56x2x1x57x32xf32>
+
+ %1 = tensor.empty() : tensor<1x57x56x2x32xf32>
+ %transposed = linalg.transpose
+ ins(%pack : tensor<56x2x1x57x32xf32>)
+ outs(%1 : tensor<1x57x56x2x32xf32>)
+ permutation = [2, 3, 0, 1, 4]
+
+ %return_value = tensor.cast %transposed : tensor<1x57x56x2x32xf32> to tensor<?x?x56x2x32xf32>
+ return %return_value : tensor<?x?x56x2x32xf32>
+}
+// CHECK: func @tensor_pack_linalg_transpose_fold_dynamic_outer_dims(
+// CHECK-SAME: %[[ARG0:.+]]: tensor<56x?x?x64xf32>)
+// CHECK: %[[c1:.+]] = arith.constant 1 : index
+// CHECK: %[[c2:.+]] = arith.constant 2 : index
+// CHECK: %[[dim:.+]] = tensor.dim %[[ARG0]], %[[c1]] : tensor<56x?x?x64xf32>
+// CHECK: %[[dim_0:.+]] = tensor.dim %[[ARG0]], %[[c2]] : tensor<56x?x?x64xf32>
+// CHECK: %[[INIT:.+]] = tensor.empty(%[[dim_0]], %[[dim]]) : tensor<?x?x56x2x32xf32>
+// CHECK: %[[PACK:.+]] = tensor.pack %[[ARG0]]
+// CHECK-SAME: outer_dims_perm = [2, 1, 0, 3]
+// CHECK-SAME: inner_dims_pos = [3] inner_tiles = [32]
+// CHECK-SAME: into %[[INIT]]
+// CHECK: return %[[PACK]]
+
+// -----
+
+func.func @tensor_pack_linalg_transpose_fold_dynamic_outer_and_tile_dims(%arg0: tensor<56x?x?x128xf32>) -> tensor<?x?x56x9x32x8x2xf32> {
+ %0 = tensor.empty() : tensor<56x9x12x4x8x2x32xf32>
+ %pack = tensor.pack %arg0
+ inner_dims_pos = [1, 2, 3]
+ inner_tiles = [8, 2, 32]
+ into %0 : tensor<56x?x?x128xf32> -> tensor<56x9x12x4x8x2x32xf32>
+
+ %1 = tensor.empty() : tensor<12x4x56x9x32x8x2xf32>
+ %transposed = linalg.transpose
+ ins(%pack : tensor<56x9x12x4x8x2x32xf32>)
+ outs(%1 : tensor<12x4x56x9x32x8x2xf32>)
+ permutation = [2, 3, 0, 1, 6, 4, 5]
+
+ %return_value = tensor.cast %transposed : tensor<12x4x56x9x32x8x2xf32> to tensor<?x?x56x9x32x8x2xf32>
+ return %return_value : tensor<?x?x56x9x32x8x2xf32>
+}
+// CHECK: #[[map:.+]] = affine_map<()[s0] -> (s0 ceildiv 8)>
+// CHECK: #[[map1:.+]] = affine_map<()[s0] -> (s0 ceildiv 2)>
+// CHECK: module {
+// CHECK: func.func @tensor_pack_linalg_transpose_fold_dynamic_outer_and_tile_dims(
+// CHECK-SAME: %[[ARG0:.+]]: tensor<56x?x?x128xf32>)
+// CHECK: %[[c1:.+]] = arith.constant 1 : index
+// CHECK: %[[c2:.+]] = arith.constant 2 : index
+// CHECK: %[[dim:.+]] = tensor.dim %[[ARG0]], %[[c1]] : tensor<56x?x?x128xf32>
+// CHECK: %[[dim_0:.+]] = tensor.dim %[[ARG0]], %[[c2]] : tensor<56x?x?x128xf32>
+// CHECK: %[[mapped_dim1:.+]] = affine.apply #[[map:.+]]()[%[[dim]]]
+// CHECK: %[[mapped_dim2:.+]] = affine.apply #[[map1:.+]]()[%[[dim_0]]]
+// CHECK: %[[INIT:.+]] = tensor.empty(%[[mapped_dim2]], %[[mapped_dim1]]) : tensor<?x4x56x?x32x8x2xf32>
+// CHECK: %[[PACK:.+]] = tensor.pack %[[ARG0]] outer_dims_perm = [2, 3, 0, 1] inner_dims_pos = [3, 1, 2] inner_tiles = [32, 8, 2] into %[[INIT]] : tensor<56x?x?x128xf32> -> tensor<?x4x56x?x32x8x2xf32>
+// CHECK: %[[CAST:.+]] = tensor.cast %[[PACK]] : tensor<?x4x56x?x32x8x2xf32> to tensor<?x?x56x9x32x8x2xf32>
+// CHECK: return %[[CAST]] : tensor<?x?x56x9x32x8x2xf32>
+// CHECK: }
+
+// -----
+
+func.func @tensor_pack_linalg_transpose_fold_dynamic_outer_dims_tile_dims_tile_sizes(%arg0: tensor<?x?x?x?xf32>, %pack_dest: tensor<?x?x?x?x?x?x?xf32>, %transpose_dest: tensor<?x?x?x?x?x?x?xf32>, %tile_p : index, %tile_q : index, %tile_r : index) -> tensor<?x?x?x?x?x?x?xf32> {
+ %pack = tensor.pack %arg0
+ outer_dims_perm = [3, 0, 2, 1]
+ inner_dims_pos = [1, 2, 3]
+ inner_tiles = [%tile_p, %tile_q, %tile_r]
+ into %pack_dest : tensor<?x?x?x?xf32> -> tensor<?x?x?x?x?x?x?xf32>
+
+ %transposed = linalg.transpose
+ ins(%pack : tensor<?x?x?x?x?x?x?xf32>)
+ outs(%transpose_dest : tensor<?x?x?x?x?x?x?xf32>)
+ permutation = [2, 3, 0, 1, 6, 4, 5]
+
+ return %transposed : tensor<?x?x?x?x?x?x?xf32>
+}
+// CHECK: #[[map:.+]] = affine_map<()[s0, s1] -> (s0 ceildiv s1)>
+// CHECK: module {
+// CHECK: func.func @tensor_pack_linalg_transpose_fold_dynamic_outer_dims_tile_dims_tile_sizes(
+// CHECK-SAME: %[[ARG0:.+]]: tensor<?x?x?x?xf32>,
+// CHECK-SAME: %[[PACK_DEST:.+]]: tensor<?x?x?x?x?x?x?xf32>, %[[TRANSPOSE_DEST:.+]]: tensor<?x?x?x?x?x?x?xf32>,
+// CHECK-SAME: %[[ARG1:.+]]: index, %[[ARG2:.+]]: index,
+// CHECK-SAME: %[[ARG3:.+]]: index)
+// CHECK: %[[c0:.+]] = arith.constant 0 : index
+// CHECK: %[[c1:.+]] = arith.constant 1 : index
+// CHECK: %[[c2:.+]] = arith.constant 2 : index
+// CHECK: %[[c3:.+]] = arith.constant 3 : index
+// CHECK: %[[dim:.+]] = tensor.dim %[[ARG0]], %[[c0]] : tensor<?x?x?x?xf32>
+// CHECK: %[[dim_0:.+]] = tensor.dim %[[ARG0]], %[[c1]] : tensor<?x?x?x?xf32>
+// CHECK: %[[dim_1:.+]] = tensor.dim %[[ARG0]], %[[c2]] : tensor<?x?x?x?xf32>
+// CHECK: %[[dim_2:.+]] = tensor.dim %[[ARG0]], %[[c3]] : tensor<?x?x?x?xf32>
+// CHECK: %[[mapped_dim0:.+]] = affine.apply #[[map:.+]]()[%[[dim_2]], %[[ARG3]]]
+// CHECK: %[[mapped_dim1:.+]] = affine.apply #[[map:.+]]()[%[[dim_0]], %[[ARG1]]]
+// CHECK: %[[mapped_dim2:.+]] = affine.apply #[[map:.+]]()[%[[dim_1]], %[[ARG2]]]
+// CHECK: %[[INIT:.+]] = tensor.empty(%[[mapped_dim2]], %[[mapped_dim1]], %[[mapped_dim0]], %[[dim]], %[[ARG3]], %[[ARG1]], %[[ARG2]]) : tensor<?x?x?x?x?x?x?xf32>
+// CHECK: %[[PACK:.+]] = tensor.pack %[[ARG0]] outer_dims_perm = [2, 1, 3, 0] inner_dims_pos = [3, 1, 2] inner_tiles = [%[[ARG3]], %[[ARG1]], %[[ARG2]]] into %[[INIT]] : tensor<?x?x?x?xf32> -> tensor<?x?x?x?x?x?x?xf32>
+// CHECK: return %[[PACK]] : tensor<?x?x?x?x?x?x?xf32>
+// CHECK: }
More information about the Mlir-commits
mailing list