[Mlir-commits] [mlir] [mlir][linalg] Restrict linalg.pack to not have extra padding sizes. (PR #149624)
Han-Chung Wang
llvmlistbot at llvm.org
Fri Jul 18 17:49:01 PDT 2025
https://github.com/hanhanW created https://github.com/llvm/llvm-project/pull/149624
None
>From 59aa0793dc0a7b93a2a5e6cab207143a43d0e930 Mon Sep 17 00:00:00 2001
From: hanhanW <hanhan0912 at gmail.com>
Date: Fri, 18 Jul 2025 17:46:08 -0700
Subject: [PATCH] [mlir][linalg] Restrict linalg.pack to not have extra padding
sizes.
Signed-off-by: hanhanW <hanhan0912 at gmail.com>
---
mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp | 17 ++--
.../Transforms/PackAndUnpackPatterns.cpp | 29 +++++++
mlir/test/Dialect/Linalg/canonicalize.mlir | 25 +++---
.../Linalg/data-layout-propagation.mlir | 22 ++---
mlir/test/Dialect/Linalg/invalid.mlir | 15 ++--
.../Dialect/Linalg/transform-lower-pack.mlir | 16 ++--
.../Tensor/fold-into-pack-and-unpack.mlir | 30 +++++--
.../tile-and-fuse-consumer.mlir | 81 -------------------
8 files changed, 102 insertions(+), 133 deletions(-)
diff --git a/mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp b/mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp
index 3aa6ac3ea0918..1d34d64b2198e 100644
--- a/mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp
+++ b/mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp
@@ -4601,8 +4601,8 @@ static bool isInvalidPackingPosSpecification(ArrayRef<int64_t> dimsPos,
/// Returns true if the dimension of `sourceShape` is smaller than the dimension
/// of the `limitShape`.
-static bool areAllInBound(ArrayRef<int64_t> sourceShape,
- ArrayRef<int64_t> limitShape) {
+static bool isCompatibleShape(ArrayRef<int64_t> sourceShape,
+ ArrayRef<int64_t> limitShape) {
assert(
sourceShape.size() == limitShape.size() &&
"expected source shape rank, and limit of the shape to have same rank");
@@ -4611,7 +4611,7 @@ static bool areAllInBound(ArrayRef<int64_t> sourceShape,
int64_t sourceExtent = std::get<0>(it);
int64_t limit = std::get<1>(it);
return ShapedType::isDynamic(sourceExtent) ||
- ShapedType::isDynamic(limit) || sourceExtent <= limit;
+ ShapedType::isDynamic(limit) || sourceExtent == limit;
});
}
@@ -4673,11 +4673,6 @@ static LogicalResult commonVerifierPackAndUnPackOp(OpTy packOrUnPack) {
// represents full tiles.
RankedTensorType expectedPackedType = PackOp::inferPackedType(
unpackedType, packOrUnPack.getStaticTiles(), innerDimsPos, outerDimPerm);
- if (!areAllInBound(expectedPackedType.getShape(), packedType.getShape())) {
- return op->emitError("the shape of output is not large enough to hold the "
- "packed data. Expected at least ")
- << expectedPackedType << ", got " << packedType;
- }
if (!llvm::all_of(
llvm::zip(packedType.getShape().take_back(mixedTiles.size()),
mixedTiles),
@@ -4694,6 +4689,12 @@ static LogicalResult commonVerifierPackAndUnPackOp(OpTy packOrUnPack) {
return op->emitError("mismatch in inner tile sizes specified and shaped of "
"tiled dimension in the packed type");
}
+ if (!isCompatibleShape(expectedPackedType.getShape(),
+ packedType.getShape())) {
+ return op->emitError("the shape of output is not large enough to hold the "
+ "packed data. Expected at least ")
+ << expectedPackedType << ", got " << packedType;
+ }
return success();
}
diff --git a/mlir/lib/Dialect/Linalg/Transforms/PackAndUnpackPatterns.cpp b/mlir/lib/Dialect/Linalg/Transforms/PackAndUnpackPatterns.cpp
index 2afa2f9b71c2a..02fdd01ed548b 100644
--- a/mlir/lib/Dialect/Linalg/Transforms/PackAndUnpackPatterns.cpp
+++ b/mlir/lib/Dialect/Linalg/Transforms/PackAndUnpackPatterns.cpp
@@ -10,6 +10,7 @@
#include "mlir/Dialect/Linalg/Transforms/Transforms.h"
#include "mlir/Dialect/Tensor/IR/Tensor.h"
#include "mlir/Dialect/Utils/IndexingUtils.h"
+#include "mlir/Dialect/Utils/StaticValueUtils.h"
#include "mlir/IR/PatternMatch.h"
namespace mlir {
@@ -220,6 +221,34 @@ struct FoldPadWithPackOp : public OpRewritePattern<PackOp> {
if (!isEqualConstantIntOrValue(paddingValue, constantPaddingValue))
return failure();
+ RankedTensorType srcType = packOp.getSourceType();
+ RankedTensorType destType = packOp.getDestType();
+ SmallVector<int64_t> outerShapeWithoutTranspose(
+ destType.getShape().take_front(srcType.getRank()));
+ if (!packOp.getOuterDimsPerm().empty()) {
+ applyPermutationToVector(
+ outerShapeWithoutTranspose,
+ invertPermutationVector(packOp.getOuterDimsPerm()));
+ }
+ for (auto [pos, tileSize, high] :
+ llvm::zip_equal(packOp.getInnerDimsPos(), packOp.getStaticInnerTiles(),
+ padOp.getMixedHighPad())) {
+ if (srcType.isDynamicDim(pos))
+ return failure();
+ if (ShapedType::isDynamic(outerShapeWithoutTranspose[pos]))
+ return failure();
+ if (ShapedType::isDynamic(tileSize))
+ return failure();
+ std::optional<int64_t> cstHigh = getConstantIntValue(high);
+ if (!cstHigh)
+ return failure();
+ int64_t paddingSize =
+ outerShapeWithoutTranspose[pos] * tileSize - srcType.getDimSize(pos);
+ // Do not fold the ops if it requires extra padding sizes.
+ if (paddingSize + cstHigh.value() >= tileSize)
+ return failure();
+ }
+
rewriter.replaceOpWithNewOp<PackOp>(
packOp, padOp.getSource(), packOp.getDest(), packOp.getInnerDimsPos(),
packOp.getMixedTiles(), constantPaddingValue,
diff --git a/mlir/test/Dialect/Linalg/canonicalize.mlir b/mlir/test/Dialect/Linalg/canonicalize.mlir
index 7284ae7dbd673..dfe3bfd4a967a 100644
--- a/mlir/test/Dialect/Linalg/canonicalize.mlir
+++ b/mlir/test/Dialect/Linalg/canonicalize.mlir
@@ -1387,42 +1387,43 @@ func.func @recursive_effect(%arg : tensor<1xf32>) {
// CHECK-LABEL: @recursive_effect
// CHECK: linalg.map
+// -----
+
//===----------------------------------------------------------------------===//
// linalg.pack
//===----------------------------------------------------------------------===//
// CHECK-LABEL: func @fold_pack_constant_splat
// CHECK-NOT: linalg.pack
-// CHECK: arith.constant dense<1.000000e-01> : tensor<8x16x8x32xf32>
-func.func @fold_pack_constant_splat(%dest : tensor<8x16x8x32xf32>) -> tensor<8x16x8x32xf32> {
+// CHECK: arith.constant dense<1.000000e-01> : tensor<4x8x8x32xf32>
+func.func @fold_pack_constant_splat(%dest : tensor<4x8x8x32xf32>) -> tensor<4x8x8x32xf32> {
%cst = arith.constant dense<1.000000e-01> : tensor<64x128xf32>
%0 = linalg.pack %cst outer_dims_perm = [1, 0] inner_dims_pos = [0, 1]
- inner_tiles = [8, 32] into %dest : tensor<64x128xf32> -> tensor<8x16x8x32xf32>
- return %0 : tensor<8x16x8x32xf32>
+ inner_tiles = [8, 32] into %dest : tensor<64x128xf32> -> tensor<4x8x8x32xf32>
+ return %0 : tensor<4x8x8x32xf32>
}
// -----
// CHECK-LABEL: func @fold_padding_value_pack_constant_splat
// CHECK-NOT: linalg.pack
-// CHECK: arith.constant dense<1.000000e-01> : tensor<8x16x8x32xf32>
-func.func @fold_padding_value_pack_constant_splat(%dest : tensor<8x16x8x32xf32>) -> tensor<8x16x8x32xf32> {
+// CHECK: arith.constant dense<1.000000e-01> : tensor<4x8x8x32xf32>
+func.func @fold_padding_value_pack_constant_splat(%dest : tensor<4x8x8x32xf32>) -> tensor<4x8x8x32xf32> {
%pad = arith.constant 1.000000e-01 : f32
%cst = arith.constant dense<1.000000e-01> : tensor<63x127xf32>
%0 = linalg.pack %cst
padding_value(%pad : f32)
outer_dims_perm = [1, 0] inner_dims_pos = [0, 1]
- inner_tiles = [8, 32] into %dest : tensor<63x127xf32> -> tensor<8x16x8x32xf32>
- return %0 : tensor<8x16x8x32xf32>
+ inner_tiles = [8, 32] into %dest : tensor<63x127xf32> -> tensor<4x8x8x32xf32>
+ return %0 : tensor<4x8x8x32xf32>
}
-
// -----
// CHECK-LABEL: func @nofold_padding_value_pack_constant_splat
// CHECK: arith.constant dense<1.000000e-01> : tensor<63x127xf32>
// CHECK: linalg.pack
-func.func @nofold_padding_value_pack_constant_splat(%dest : tensor<8x16x8x32xf32>) -> tensor<8x16x8x32xf32> {
+func.func @nofold_padding_value_pack_constant_splat(%dest : tensor<4x8x8x32xf32>) -> tensor<4x8x8x32xf32> {
%pad = arith.constant 0.0 : f32
%cst = arith.constant dense<1.000000e-01> : tensor<63x127xf32>
%0 = linalg.pack %cst
@@ -1430,8 +1431,8 @@ func.func @nofold_padding_value_pack_constant_splat(%dest : tensor<8x16x8x32xf32
outer_dims_perm = [1, 0]
inner_dims_pos = [0, 1]
inner_tiles = [8, 32]
- into %dest : tensor<63x127xf32> -> tensor<8x16x8x32xf32>
- return %0 : tensor<8x16x8x32xf32>
+ into %dest : tensor<63x127xf32> -> tensor<4x8x8x32xf32>
+ return %0 : tensor<4x8x8x32xf32>
}
// -----
diff --git a/mlir/test/Dialect/Linalg/data-layout-propagation.mlir b/mlir/test/Dialect/Linalg/data-layout-propagation.mlir
index 6fc8d9f152f4e..ae87fffd1af02 100644
--- a/mlir/test/Dialect/Linalg/data-layout-propagation.mlir
+++ b/mlir/test/Dialect/Linalg/data-layout-propagation.mlir
@@ -1295,21 +1295,21 @@ func.func @no_bubble_up_pack_expanded_padding_through_expand_cannot_reassociate(
// -----
-func.func @no_bubble_up_pack_extending_dimension_through_expand_cannot_reassociate(%arg0: tensor<32x64xf32>) -> tensor<8x4x16x8xf32> {
- %empty = tensor.empty() : tensor<8x4x16x8xf32>
+func.func @bubble_up_pack_extending_dimension_through_expand_can_reassociate(%arg0: tensor<32x64xf32>) -> tensor<4x4x16x8xf32> {
+ %empty = tensor.empty() : tensor<4x4x16x8xf32>
%expanded = tensor.expand_shape %arg0 [[0], [1, 2]] output_shape [32, 4, 16] : tensor<32x64xf32> into tensor<32x4x16xf32>
- %pack = linalg.pack %expanded inner_dims_pos = [0] inner_tiles = [8] into %empty : tensor<32x4x16xf32> -> tensor<8x4x16x8xf32>
- return %pack : tensor<8x4x16x8xf32>
+ %pack = linalg.pack %expanded inner_dims_pos = [0] inner_tiles = [8] into %empty : tensor<32x4x16xf32> -> tensor<4x4x16x8xf32>
+ return %pack : tensor<4x4x16x8xf32>
}
-// CHECK-LABEL: func.func @no_bubble_up_pack_extending_dimension_through_expand_cannot_reassociate(
+// CHECK-LABEL: func.func @bubble_up_pack_extending_dimension_through_expand_can_reassociate(
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9]+]]
-// CHECK: %[[EMPTY:.+]] = tensor.empty() : tensor<8x4x16x8xf32>
-// CHECK: %[[EXPANDED:.+]] = tensor.expand_shape %[[ARG0]] {{\[}}[0], [1, 2]]
-// CHECK-SAME: output_shape [32, 4, 16] : tensor<32x64xf32> into tensor<32x4x16xf32>
-// CHECK: %[[PACK:.+]] = linalg.pack %[[EXPANDED]]
+// CHECK: %[[EMPTY:.+]] = tensor.empty() : tensor<4x64x8xf32>
+// CHECK: %[[PACK:.+]] = linalg.pack %[[ARG0]]
// CHECK-SAME: inner_dims_pos = [0] inner_tiles = [8] into %[[EMPTY]]
-// CHECK-SAME: : tensor<32x4x16xf32> -> tensor<8x4x16x8xf32>
-// CHECK: return %[[PACK]] : tensor<8x4x16x8xf32>
+// CHECK-SAME: : tensor<32x64xf32> -> tensor<4x64x8xf32>
+// CHECK: %[[EXPANDED:.+]] = tensor.expand_shape %[[PACK]] {{\[}}[0], [1, 2], [3]]
+// CHECK-SAME: output_shape [4, 4, 16, 8] : tensor<4x64x8xf32> into tensor<4x4x16x8xf32>
+// CHECK: return %[[EXPANDED]] : tensor<4x4x16x8xf32>
// -----
diff --git a/mlir/test/Dialect/Linalg/invalid.mlir b/mlir/test/Dialect/Linalg/invalid.mlir
index da1dfc7b6a624..83611a217f652 100644
--- a/mlir/test/Dialect/Linalg/invalid.mlir
+++ b/mlir/test/Dialect/Linalg/invalid.mlir
@@ -1760,6 +1760,7 @@ func.func @pack_invalid(%input: tensor<256x128xf32>, %output: tensor<8x8x32x16xf
}
// -----
+
func.func @pack_mismatch_inner_tile_size_and_output_shape(
%input : tensor<?x?xf32>, %output : tensor<?x?x8x8xf32>) -> tensor<?x?x8x8xf32> {
// expected-error at +1 {{mismatch in inner tile sizes specified and shaped of tiled dimension in the packed type}}
@@ -1834,17 +1835,17 @@ func.func @pack_invalid_result_shape(%input: tensor<256x128xf32>, %output: tenso
// -----
-func.func @pack_invalid(%input: tensor<256x128xf32>, %output: tensor<8x8x32x16xf32>) -> tensor<8x8x32x16xf32> {
- // expected-error at +1 {{the shape of output is not large enough to hold the packed data. Expected at least 'tensor<8x8x16x32xf32>', got 'tensor<8x8x32x16xf32>'}}
- %0 = linalg.pack %input inner_dims_pos = [1, 0] inner_tiles = [16, 32] into %output : tensor<256x128xf32> -> tensor<8x8x32x16xf32>
- return %0 : tensor<8x8x32x16xf32>
+func.func @pack_invalid(%input: tensor<256x128xf32>, %output: tensor<8x7x16x32xf32>) -> tensor<8x7x16x32xf32> {
+ // expected-error at +1 {{the shape of output is not large enough to hold the packed data. Expected at least 'tensor<8x8x16x32xf32>', got 'tensor<8x7x16x32xf32>'}}
+ %0 = linalg.pack %input inner_dims_pos = [1, 0] inner_tiles = [16, 32] into %output : tensor<256x128xf32> -> tensor<8x7x16x32xf32>
+ return %0 : tensor<8x7x16x32xf32>
}
// -----
-func.func @unpack_invalid(%output: tensor<256x128xf32>, %input: tensor<8x8x32x16xf32>) -> tensor<256x128xf32> {
- // expected-error at +1 {{the shape of output is not large enough to hold the packed data. Expected at least 'tensor<8x32x4x32xf32>', got 'tensor<8x8x32x16xf32>'}}
- %0 = linalg.unpack %input inner_dims_pos = [1, 0] inner_tiles = [4, 32] into %output : tensor<8x8x32x16xf32> -> tensor<256x128xf32>
+func.func @unpack_invalid(%output: tensor<256x128xf32>, %input: tensor<8x8x4x32xf32>) -> tensor<256x128xf32> {
+ // expected-error at +1 {{the shape of output is not large enough to hold the packed data. Expected at least 'tensor<8x32x4x32xf32>', got 'tensor<8x8x4x32xf32>'}}
+ %0 = linalg.unpack %input inner_dims_pos = [1, 0] inner_tiles = [4, 32] into %output : tensor<8x8x4x32xf32> -> tensor<256x128xf32>
return %0 : tensor<256x128xf32>
}
diff --git a/mlir/test/Dialect/Linalg/transform-lower-pack.mlir b/mlir/test/Dialect/Linalg/transform-lower-pack.mlir
index 81fd7a8a947d7..9e7681d1a1b7d 100644
--- a/mlir/test/Dialect/Linalg/transform-lower-pack.mlir
+++ b/mlir/test/Dialect/Linalg/transform-lower-pack.mlir
@@ -326,23 +326,23 @@ module attributes {transform.with_named_sequence} {
// -----
// CHECK-LABEL: func.func @pack_with_pad(
-func.func @pack_with_pad(%src: tensor<4225x12xf32>, %dest: tensor<265x16x16x1xf32>)
- -> tensor<265x16x16x1xf32> {
+func.func @pack_with_pad(%src: tensor<4225x12xf32>, %dest: tensor<265x12x16x1xf32>)
+ -> tensor<265x12x16x1xf32> {
// CHECK: tensor.pad {{.*}} low[0, 0]
- // CHECK: : tensor<4225x12xf32> to tensor<4240x16xf32>
+ // CHECK: : tensor<4225x12xf32> to tensor<4240x12xf32>
// CHECK: tensor.expand_shape %{{.*}} {{\[}}[0, 1], [2, 3]]
- // CHECK-SAME: : tensor<4240x16xf32> into tensor<265x16x16x1xf32>
+ // CHECK-SAME: : tensor<4240x12xf32> into tensor<265x16x12x1xf32>
// CHECK: linalg.transpose
- // CHECK-SAME: ins(%{{[a-zA-Z0-9]*}} : tensor<265x16x16x1xf32>)
- // CHECK-SAME: outs(%{{[a-zA-Z0-9]*}} : tensor<265x16x16x1xf32>)
+ // CHECK-SAME: ins(%{{[a-zA-Z0-9]*}} : tensor<265x16x12x1xf32>)
+ // CHECK-SAME: outs(%{{[a-zA-Z0-9]*}} : tensor<265x12x16x1xf32>)
// CHECK-SAME: permutation = [0, 2, 1, 3]
%cst = arith.constant 0.000000e+00 : f32
%0 = linalg.pack %src
padding_value(%cst : f32)
inner_dims_pos = [0, 1]
inner_tiles = [16, 1] into %dest
- : tensor<4225x12xf32> -> tensor<265x16x16x1xf32>
- return %0 : tensor<265x16x16x1xf32>
+ : tensor<4225x12xf32> -> tensor<265x12x16x1xf32>
+ return %0 : tensor<265x12x16x1xf32>
}
module attributes {transform.with_named_sequence} {
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 16efa73f87a2a..eb62de13ebc94 100644
--- a/mlir/test/Dialect/Tensor/fold-into-pack-and-unpack.mlir
+++ b/mlir/test/Dialect/Tensor/fold-into-pack-and-unpack.mlir
@@ -59,13 +59,13 @@ func.func @nofold_unpack_slice_rank_reduced(%arg0 : tensor<?x?x8x4xf32>, %arg1 :
// -----
-func.func @pad_pack(%src: tensor<16641x16xf32>) -> tensor<2082x1x8x32xf32> {
+func.func @pad_pack(%src: tensor<16649x16xf32>) -> tensor<2082x1x8x32xf32> {
%c0 = arith.constant 0 : index
%cst = arith.constant 0.000000e+00 : f32
- %padded = tensor.pad %src low[0, 0] high[15, 0] {
+ %padded = tensor.pad %src low[0, 0] high[7, 0] {
^bb0(%arg0: index, %arg1: index):
tensor.yield %cst : f32
- } : tensor<16641x16xf32> to tensor<16656x16xf32>
+ } : tensor<16649x16xf32> to tensor<16656x16xf32>
%empty = tensor.empty() : tensor<2082x1x8x32xf32>
%pack = linalg.pack %padded padding_value(%cst : f32) inner_dims_pos = [0, 1] inner_tiles = [8, 32] into %empty
: tensor<16656x16xf32> -> tensor<2082x1x8x32xf32>
@@ -81,10 +81,10 @@ func.func @pad_pack(%src: tensor<16641x16xf32>) -> tensor<2082x1x8x32xf32> {
// -----
-func.func @nofold_pad_pack(%src: tensor<16641x16xf32>) -> tensor<2082x1x8x32xf32> {
+func.func @nofold_pad_pack_extra_padding(%src: tensor<16641x16xf32>) -> tensor<2082x1x8x32xf32> {
%c0 = arith.constant 0 : index
%cst = arith.constant 0.000000e+00 : f32
- %padded = tensor.pad %src nofold low[0, 0] high[15, 0] {
+ %padded = tensor.pad %src low[0, 0] high[15, 0] {
^bb0(%arg0: index, %arg1: index):
tensor.yield %cst : f32
} : tensor<16641x16xf32> to tensor<16656x16xf32>
@@ -93,7 +93,25 @@ func.func @nofold_pad_pack(%src: tensor<16641x16xf32>) -> tensor<2082x1x8x32xf32
: tensor<16656x16xf32> -> tensor<2082x1x8x32xf32>
return %pack : tensor<2082x1x8x32xf32>
}
-// CHECK-LABEL: func.func @nofold_pad_pack
+// CHECK-LABLE: func.func @nofold_pad_pack_extra_padding(
+// CHECK: tensor.pad
+// CHECK: linalg.pack
+
+// -----
+
+func.func @nofold_pad_pack(%src: tensor<16649x16xf32>) -> tensor<2082x1x8x32xf32> {
+ %c0 = arith.constant 0 : index
+ %cst = arith.constant 0.000000e+00 : f32
+ %padded = tensor.pad %src nofold low[0, 0] high[7, 0] {
+ ^bb0(%arg0: index, %arg1: index):
+ tensor.yield %cst : f32
+ } : tensor<16649x16xf32> to tensor<16656x16xf32>
+ %empty = tensor.empty() : tensor<2082x1x8x32xf32>
+ %pack = linalg.pack %padded padding_value(%cst : f32) inner_dims_pos = [0, 1] inner_tiles = [8, 32] into %empty
+ : tensor<16656x16xf32> -> tensor<2082x1x8x32xf32>
+ return %pack : tensor<2082x1x8x32xf32>
+}
+// CHECK-LABEL: func.func @nofold_pad_pack(
// CHECK: tensor.pad
// CHECK: linalg.pack
diff --git a/mlir/test/Interfaces/TilingInterface/tile-and-fuse-consumer.mlir b/mlir/test/Interfaces/TilingInterface/tile-and-fuse-consumer.mlir
index cdbca7228ded3..e48e5c6c308be 100644
--- a/mlir/test/Interfaces/TilingInterface/tile-and-fuse-consumer.mlir
+++ b/mlir/test/Interfaces/TilingInterface/tile-and-fuse-consumer.mlir
@@ -646,87 +646,6 @@ module attributes {transform.with_named_sequence} {
// -----
-// It is valid to fuse the pack if the dimension is not tiled even when it needs
-// extra padding.
-
-func.func @fuse_pack_consumer_with_untiled_extra_padding(%arg0: tensor<64x32xf32>, %arg1: tensor<64x32xf32>) -> tensor<33x2x3x16xf32> {
- %0 = scf.forall (%arg2) = (0) to (32) step (16) shared_outs(%arg3 = %arg1) -> (tensor<64x32xf32>) {
- %src = tensor.extract_slice %arg0[0, %arg2] [64, 16] [1, 1] : tensor<64x32xf32> to tensor<64x16xf32>
- %dest = tensor.extract_slice %arg3[0, %arg2] [64, 16] [1, 1] : tensor<64x32xf32> to tensor<64x16xf32>
- %2 = linalg.exp ins(%src : tensor<64x16xf32>) outs(%dest : tensor<64x16xf32>) -> tensor<64x16xf32>
- scf.forall.in_parallel {
- tensor.parallel_insert_slice %2 into %arg3[0, %arg2] [64, 16] [1, 1] : tensor<64x16xf32> into tensor<64x32xf32>
- }
- }
- %1 = tensor.empty() : tensor<33x2x3x16xf32>
- %cst = arith.constant 0.000000e+00 : f32
- %pack = linalg.pack %0 padding_value(%cst : f32) inner_dims_pos = [0, 1] inner_tiles = [3, 16] into %1 : tensor<64x32xf32> -> tensor<33x2x3x16xf32>
- return %pack : tensor<33x2x3x16xf32>
-}
-
-module attributes {transform.with_named_sequence} {
- transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
- %0 = transform.structured.match ops{["tensor.parallel_insert_slice"]} in %arg0 : (!transform.any_op) -> !transform.any_op
- %1 = transform.structured.match ops{["scf.forall"]} in %arg0 : (!transform.any_op) -> !transform.any_op
- %consumer, %fused_consumer = transform.test.fuse_consumer %0 in(%1) : (!transform.any_op, !transform.any_op) -> (!transform.any_op, !transform.any_op)
- transform.yield
- }
-}
-// CHECK: #[[PACK_RESULT_MAP:.*]] = affine_map<(d0) -> (d0 floordiv 16)>
-// CHECK: func.func @fuse_pack_consumer_with_untiled_extra_padding(
-// CHECK-SAME: %[[ARG0:[a-zA-Z0-9]+]]
-// CHECK-SAME: %[[ARG1:[a-zA-Z0-9]+]]
-// CHECK-DAG: %[[OUT_INIT:.*]] = tensor.empty() : tensor<33x2x3x16xf32>
-// CHECK-DAG: %[[PAD_VAL:.*]] = arith.constant 0.000000e+00 : f32
-// CHECK: %{{.*}}:2 = scf.forall (%[[IV:.*]]) = (0) to (32) step (16)
-// CHECK-SAME: shared_outs(%[[FIRST_OUT_ARG:.*]] = %[[ARG1]], %[[PACK_OUT_ARG:.*]] = %[[OUT_INIT]])
-// CHECK: %[[ELEM_SRC:.*]] = tensor.extract_slice %[[ARG0]][0, %[[IV]]] [64, 16] [1, 1]
-// CHECK: %[[ELEM_DEST:.*]] = tensor.extract_slice %[[FIRST_OUT_ARG]][0, %[[IV]]] [64, 16] [1, 1]
-// CHECK: %[[ELEM:.*]] = linalg.exp
-// CHECK-SAME: ins(%[[ELEM_SRC]]
-// CHECK-SAME: outs(%[[ELEM_DEST]]
-// CHECK-DAG: %[[PACK_RESULT_OFFSET:.*]] = affine.apply #[[PACK_RESULT_MAP]](%[[IV]])
-// CHECK-DAG: %[[TILED_PACK_DEST:.*]] = tensor.extract_slice %[[PACK_OUT_ARG]][0, %[[PACK_RESULT_OFFSET]], 0, 0] [33, 1, 3, 16] [1, 1, 1, 1]
-// CHECK: %[[TILED_PACK_OUT:.*]] = linalg.pack %[[ELEM]]
-// CHECK-SAME: padding_value(%[[PAD_VAL]] : f32)
-// CHECK-SAME: inner_dims_pos = [0, 1] inner_tiles = [3, 16]
-// CHECK-SAME: into %[[TILED_PACK_DEST]]
-// CHECK: scf.forall.in_parallel {
-// CHECK: tensor.parallel_insert_slice %[[GENERIC_OUT]] into %[[FIRST_OUT_ARG]][0, %[[IV]]] [64, 16] [1, 1]
-// CHECK: tensor.parallel_insert_slice %[[TILED_PACK_OUT]] into %[[PACK_OUT_ARG]][0, %[[PACK_RESULT_OFFSET]], 0, 0] [33, 1, 3, 16] [1, 1, 1, 1]
-
-// -----
-
-// If the dimension is tiled and it needs extra padding, do not fuse the pack
-// op.
-
-func.func @nofuse_pack_consumer_with_extra_padding(%arg0: tensor<64x32xf32>, %arg1: tensor<64x32xf32>) -> tensor<23x32x3x16xf32> {
- %0 = scf.forall (%arg2) = (0) to (32) step (16) shared_outs(%arg3 = %arg1) -> (tensor<64x32xf32>) {
- %src = tensor.extract_slice %arg0[0, %arg2] [64, 16] [1, 1] : tensor<64x32xf32> to tensor<64x16xf32>
- %dest = tensor.extract_slice %arg3[0, %arg2] [64, 16] [1, 1] : tensor<64x32xf32> to tensor<64x16xf32>
- %2 = linalg.exp ins(%src : tensor<64x16xf32>) outs(%dest : tensor<64x16xf32>) -> tensor<64x16xf32>
- scf.forall.in_parallel {
- // expected-error @below {{failed to fuse consumer of slice}}
- tensor.parallel_insert_slice %2 into %arg3[0, %arg2] [64, 16] [1, 1] : tensor<64x16xf32> into tensor<64x32xf32>
- }
- }
- %1 = tensor.empty() : tensor<23x32x3x16xf32>
- %cst = arith.constant 0.000000e+00 : f32
- %pack = linalg.pack %0 padding_value(%cst : f32) inner_dims_pos = [0, 1] inner_tiles = [3, 16] into %1 : tensor<64x32xf32> -> tensor<23x32x3x16xf32>
- return %pack : tensor<23x32x3x16xf32>
-}
-
-module attributes {transform.with_named_sequence} {
- transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
- %0 = transform.structured.match ops{["tensor.parallel_insert_slice"]} in %arg0 : (!transform.any_op) -> !transform.any_op
- %1 = transform.structured.match ops{["scf.forall"]} in %arg0 : (!transform.any_op) -> !transform.any_op
- %consumer, %fused_consumer = transform.test.fuse_consumer %0 in(%1) : (!transform.any_op, !transform.any_op) -> (!transform.any_op, !transform.any_op)
- transform.yield
- }
-}
-
-// -----
-
// Imperfect tiling is not supported in pack op consumer fusion.
#map = affine_map<(d0) -> (d0 * 5)>
More information about the Mlir-commits
mailing list