Skip to content

[mlir][tensor] Fold producer linalg transpose with consumer tensor pack #75658

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

Merged
merged 2 commits into from
Jan 10, 2024
Merged

[mlir][tensor] Fold producer linalg transpose with consumer tensor pack #75658

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
aedabce
Fix merge conflicts
meshtag Jan 4, 2024
976c4bb
Fix multiple tiles bug
meshtag Jan 9, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
44 changes: 43 additions & 1 deletion mlir/lib/Dialect/Tensor/Transforms/PackAndUnpackPatterns.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,6 +9,7 @@
#include "mlir/Dialect/Linalg/IR/Linalg.h"
#include "mlir/Dialect/Tensor/IR/Tensor.h"
#include "mlir/Dialect/Tensor/Transforms/Transforms.h"
#include "mlir/Dialect/Utils/IndexingUtils.h"
#include "mlir/IR/PatternMatch.h"
#include "llvm/Support/Debug.h"

Expand Down Expand Up @@ -223,11 +224,52 @@ struct FoldProducerPackWithConsumerLinalgTransposeOp
return success();
}
};

/// Fold 'transpose' -> 'pack' into 'pack' since 'pack' already has transpose
/// semantics.
struct FoldConsumerPackWithProducerLinalgTransposeOp
: public OpRewritePattern<PackOp> {
using OpRewritePattern<PackOp>::OpRewritePattern;

LogicalResult matchAndRewrite(PackOp packOp,
PatternRewriter &rewriter) const override {
auto transposeOp = packOp.getSource().getDefiningOp<linalg::TransposeOp>();

if (!transposeOp)
return failure();

auto transposePermutation = transposeOp.getPermutation();
auto outerDimsPerm = packOp.getOuterDimsPerm();
auto innerDimsPos = packOp.getInnerDimsPos();
SmallVector<int64_t> newInnerDimsPosVec;
SmallVector<int64_t> newOuterDimsPermVec =
llvm::to_vector(transposePermutation);

if (!outerDimsPerm.empty())
applyPermutationToVector(newOuterDimsPermVec, outerDimsPerm);

// Can't use applyPermutationToVector for newInnerDimsPosVec since input and
// permutation rank won't necessarily be equal in all cases.
for (auto dim : innerDimsPos)
newInnerDimsPosVec.push_back(transposePermutation[dim]);

Value output = packOp.createDestinationTensor(
rewriter, packOp.getLoc(), transposeOp.getOperand(0),
packOp.getMixedTiles(), newInnerDimsPosVec, newOuterDimsPermVec);

rewriter.replaceOpWithNewOp<PackOp>(
packOp, transposeOp.getOperand(0), output, newInnerDimsPosVec,
packOp.getMixedTiles(), packOp.getPaddingValue(), newOuterDimsPermVec);

return success();
}
};
} // namespace

void populateFoldIntoPackAndUnpackPatterns(RewritePatternSet &patterns) {
patterns.insert<FoldUnpackWithExtractSliceOp, FoldPadWithPackOp,
FoldProducerPackWithConsumerLinalgTransposeOp>(
FoldProducerPackWithConsumerLinalgTransposeOp,
FoldConsumerPackWithProducerLinalgTransposeOp>(
patterns.getContext());
}

Expand Down
177 changes: 177 additions & 0 deletions mlir/test/Dialect/Tensor/fold-into-pack-and-unpack.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -345,3 +345,180 @@ func.func @tensor_pack_linalg_transpose_fold_dynamic_outer_dims_tile_dims_tile_s
// 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: }

// -----

func.func @linalg_transpose_tensor_pack_fold(%arg0: tensor<56x57x1x64xf32>) -> tensor<1x57x56x2x32xf32> {
%0 = tensor.empty() : tensor<1x56x57x64xf32>
%transposed = linalg.transpose
ins(%arg0 : tensor<56x57x1x64xf32>)
outs(%0 : tensor<1x56x57x64xf32>)
permutation = [2, 0, 1, 3]

%1 = tensor.empty() : tensor<1x57x56x2x32xf32>
%pack = tensor.pack %transposed
outer_dims_perm = [0, 2, 1, 3]
inner_dims_pos = [3]
inner_tiles = [32]
into %1 : tensor<1x56x57x64xf32> -> tensor<1x57x56x2x32xf32>
return %pack : tensor<1x57x56x2x32xf32>
}
//CHECK-LABEL: func @linalg_transpose_tensor_pack_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 @linalg_transpose_tensor_pack_fold_with_padding(%arg0: tensor<56x57x1x55xf32>, %padding: f32) -> tensor<1x57x56x2x32xf32> {
%0 = tensor.empty() : tensor<1x56x57x55xf32>
%transpose = linalg.transpose
ins(%arg0 : tensor<56x57x1x55xf32>)
outs(%0 : tensor<1x56x57x55xf32>)
permutation = [2, 0, 1, 3]

%1 = tensor.empty() : tensor<1x57x56x2x32xf32>
%pack = tensor.pack %transpose padding_value(%padding : f32)
outer_dims_perm = [0, 2, 1, 3]
inner_dims_pos = [3]
inner_tiles = [32]
into %1 : tensor<1x56x57x55xf32> -> tensor<1x57x56x2x32xf32>
return %pack : tensor<1x57x56x2x32xf32>
}
//CHECK-LABEL: func @linalg_transpose_tensor_pack_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 @linalg_transpose_tensor_pack_fold_no_outer_dims_perm(%arg0: tensor<56x57x1x64xf32>) -> tensor<1x56x57x2x32xf32> {
%0 = tensor.empty() : tensor<1x56x57x64xf32>
%transposed = linalg.transpose
ins(%arg0 : tensor<56x57x1x64xf32>)
outs(%0 : tensor<1x56x57x64xf32>)
permutation = [2, 0, 1, 3]

%1 = tensor.empty() : tensor<1x56x57x2x32xf32>
%pack = tensor.pack %transposed
inner_dims_pos = [3]
inner_tiles = [32]
into %1 : tensor<1x56x57x64xf32> -> tensor<1x56x57x2x32xf32>
return %pack : tensor<1x56x57x2x32xf32>
}
//CHECK-LABEL: func @linalg_transpose_tensor_pack_fold_no_outer_dims_perm(
// CHECK-SAME: %[[ARG0:.+]]: tensor<56x57x1x64xf32>)
// CHECK: %[[INIT:.+]] = tensor.empty() : tensor<1x56x57x2x32xf32>
// CHECK: %[[PACK:.+]] = tensor.pack %[[ARG0]]
// CHECK-SAME: outer_dims_perm = [2, 0, 1, 3]
// CHECK-SAME: inner_dims_pos = [3] inner_tiles = [32]
// CHECK-SAME: into %[[INIT]]
// CHECK: return %[[PACK]]

// -----

func.func @linalg_transpose_tensor_pack_fold_complex_inner_dims_change(%arg0: tensor<25x30x35x40xf32>, %transpose_dest: tensor<35x40x25x30xf32>, %pack_dest: tensor<3x35x5x8x5x10x5xf32>) -> tensor<3x35x5x8x5x10x5xf32> {
%transposed = linalg.transpose
ins(%arg0 : tensor<25x30x35x40xf32>)
outs(%transpose_dest : tensor<35x40x25x30xf32>)
permutation = [2, 3, 0, 1]

%pack = tensor.pack %transposed
outer_dims_perm = [3, 0, 2, 1]
inner_dims_pos = [1, 3, 2]
inner_tiles = [5, 10, 5]
into %pack_dest : tensor<35x40x25x30xf32> -> tensor<3x35x5x8x5x10x5xf32>
return %pack : tensor<3x35x5x8x5x10x5xf32>
}
//CHECK-LABEL: func.func @linalg_transpose_tensor_pack_fold_complex_inner_dims_change(
// CHECK-SAME: %[[ARG0:.+]]: tensor<25x30x35x40xf32>,
// CHECK-SAME: %[[ARG1:.+]]: tensor<35x40x25x30xf32>,
// CHECK-SAME: %[[ARG2:.+]]: tensor<3x35x5x8x5x10x5xf32>) -> tensor<3x35x5x8x5x10x5xf32> {
// CHECK: %[[VAL0:.+]] = tensor.empty() : tensor<3x35x5x8x5x10x5xf32>
// CHECK: %[[PACK:.+]] = tensor.pack %[[ARG0]]
// CHECK-SAME: outer_dims_perm = [1, 2, 0, 3]
// CHECK-SAME: inner_dims_pos = [3, 1, 0]
// CHECK-SAME: inner_tiles = [5, 10, 5]
// CHECK-SAME: into %[[VAL0]]
// CHECK: return %[[PACK]]

// -----

func.func @linalg_transpose_tensor_pack_fold_dynamic_outer_dims_tile_dims_tile_sizes(%arg0: tensor<?x?x?x?xf32>, %transpose_dest: tensor<?x?x?x?xf32>, %pack_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> {
%transposed = linalg.transpose
ins(%arg0 : tensor<?x?x?x?xf32>)
outs(%transpose_dest : tensor<?x?x?x?xf32>)
permutation = [2, 3, 0, 1]

%pack = tensor.pack %transposed
outer_dims_perm = [3, 0, 2, 1]
inner_dims_pos = [1, 3, 2]
inner_tiles = [%tile_p, %tile_q, %tile_r]
into %pack_dest : tensor<?x?x?x?xf32> -> tensor<?x?x?x?x?x?x?xf32>
return %pack : tensor<?x?x?x?x?x?x?xf32>
}
// CHECK: #[[map:.+]] = affine_map<()[s0, s1] -> (s0 ceildiv s1)>
//CHECK-LABEL: func.func @linalg_transpose_tensor_pack_fold_dynamic_outer_dims_tile_dims_tile_sizes(
// CHECK-SAME: %[[ARG0:.+]]: tensor<?x?x?x?xf32>, %[[ARG1:.+]]: tensor<?x?x?x?xf32>,
// CHECK-SAME: %[[ARG2:.+]]: tensor<?x?x?x?x?x?x?xf32>, %[[ARG3:.+]]: index, %[[ARG4:.+]]: index, %[[ARG5:.+]]: index) -> tensor<?x?x?x?x?x?x?xf32> {
// 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: %[[DIM0:.+]] = tensor.dim %[[ARG0]], %[[C1]] : tensor<?x?x?x?xf32>
// CHECK: %[[DIM1:.+]] = tensor.dim %[[ARG0]], %[[C2]] : tensor<?x?x?x?xf32>
// CHECK: %[[DIM2:.+]] = tensor.dim %[[ARG0]], %[[C3]] : tensor<?x?x?x?xf32>
// CHECK: %[[VAL0:.+]] = affine.apply #[[map:.+]]()[%[[DIM2]], %[[ARG3]]]
// CHECK: %[[VAL1:.+]] = affine.apply #[[map:.+]]()[%[[DIM0]], %[[ARG4]]]
// CHECK: %[[VAL2:.+]] = affine.apply #[[map:.+]]()[%[[DIM]], %[[ARG5]]]
// CHECK: %[[VAL3:.+]] = tensor.empty(%[[VAL1]], %[[DIM1]], %[[VAL2]], %[[VAL0]], %[[ARG3]], %[[ARG4]], %[[ARG5]]) : tensor<?x?x?x?x?x?x?xf32>
// CHECK: %[[PACK:.+]] = tensor.pack %[[ARG0]] outer_dims_perm = [1, 2, 0, 3] inner_dims_pos = [3, 1, 0] inner_tiles = [%[[ARG3]], %[[ARG4]], %[[ARG5]]] into %[[VAL3]] : tensor<?x?x?x?xf32> -> tensor<?x?x?x?x?x?x?xf32>
// CHECK: return %[[PACK]] : tensor<?x?x?x?x?x?x?xf32>

// -----

func.func @linalg_transpose_tensor_pack_multiple_tiles(%arg0: tensor<?x32x128xbf16>) -> tensor<32x?x64x16x2xbf16> {
%c0 = arith.constant 0 : index
%cst = arith.constant 0.000000e+00 : bf16
%dim = tensor.dim %arg0, %c0 : tensor<?x32x128xbf16>

%0 = tensor.empty(%dim) : tensor<32x128x?xbf16>
%transposed = linalg.transpose
ins(%arg0 : tensor<?x32x128xbf16>)
outs(%0 : tensor<32x128x?xbf16>)
permutation = [1, 2, 0]

%2 = tensor.empty(%dim) : tensor<32x?x64x16x2xbf16>
%pack = tensor.pack %transposed
padding_value(%cst : bf16)
outer_dims_perm = [0, 2, 1]
inner_dims_pos = [2, 1]
inner_tiles = [16, 2]
into %2 : tensor<32x128x?xbf16> -> tensor<32x?x64x16x2xbf16>
return %pack : tensor<32x?x64x16x2xbf16>
}
// CHECK: #[[map:.+]] = affine_map<()[s0] -> (s0 ceildiv 16)>
//CHECK-LABEL: func.func @linalg_transpose_tensor_pack_multiple_tiles(
// CHECK-SAME: %[[ARG0:.+]]: tensor<?x32x128xbf16>) -> tensor<32x?x64x16x2xbf16> {
// CHECK: %[[C0:.+]] = arith.constant 0 : index
// CHECK: %[[CST:.+]] = arith.constant 0.000000e+00 : bf16
// CHECK: %[[DIM:.+]] = tensor.dim %[[ARG0]], %[[C0]] : tensor<?x32x128xbf16>
// CHECK: %[[VAL0:.+]] = affine.apply #[[map:.+]]()[%[[DIM]]]
// CHECK: %[[VAL1:.+]] = tensor.empty(%[[VAL0]]) : tensor<32x?x64x16x2xbf16>
// CHECK: %[[PACK:.+]] = tensor.pack %[[ARG0]]
// CHECK-SAME: padding_value(%[[CST]] : bf16)
// CHECK-SAME: outer_dims_perm = [1, 0, 2]
// CHECK-SAME: inner_dims_pos = [0, 2]
// CHECK-SAME: inner_tiles = [16, 2]
// CHECK-SAME: into %[[VAL1]] : tensor<?x32x128xbf16> -> tensor<32x?x64x16x2xbf16>
// CHECK: return %[[PACK]] : tensor<32x?x64x16x2xbf16>
// CHECK: }