-
Notifications
You must be signed in to change notification settings - Fork 10.7k
/
FusePadOpWithLinalgConsumer.cpp
127 lines (113 loc) · 4.84 KB
/
FusePadOpWithLinalgConsumer.cpp
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
//===- FusePadOpWithLinalgConsumer.cpp ---- Fuse pad with linalg producer -===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file implements patterns that fuses a linalg.generic -> tensor.pad op
// chain into a tensor.extract_slice -> linalg.generic -> tensor.insert_slice
// op chain.
//
//===----------------------------------------------------------------------===//
#include "mlir/Dialect/Linalg/Transforms/Transforms.h"
#include "mlir/Dialect/Linalg/IR/Linalg.h"
#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
using namespace mlir;
namespace {
/// A sequence of operations
///
/// ```mlir
/// %0 = linalg. ...
/// %1 = tensor.pad %0 ...
/// ```
///
/// can be replaced with
///
/// ```mlir
/// %0 = linalg.fill
/// %1 = tensor.extract_slice %0 ...
/// %2 = linalg. .... outs(..., %1, ....) ....
/// %3 = tensor.insert_slice %2 into %1 ...
/// ```
///
/// if the `linalg.generic` has all parallel iterator types.
struct FusePadOp : OpRewritePattern<tensor::PadOp> {
using OpRewritePattern<tensor::PadOp>::OpRewritePattern;
LogicalResult matchAndRewrite(tensor::PadOp padOp,
PatternRewriter &rewriter) const override {
// Only works on padding op that sets the padded value to a constant.
Value padValue = padOp.getConstantPaddingValue();
if (!padValue)
return rewriter.notifyMatchFailure(padOp, "non constant padding");
// This pattern could work for any Linalg op. For now restrict it to generic
// ops.
Value source = padOp.getSource();
auto linalgOp = source.getDefiningOp<linalg::GenericOp>();
if (!linalgOp) {
return rewriter.notifyMatchFailure(
padOp, "expected source to be linalg.generic op");
}
// All iterator types need to be parallel.
if (linalgOp.getNumLoops() != linalgOp.getNumParallelLoops()) {
return rewriter.notifyMatchFailure(
padOp, "only supported for ops with all parallel iterator types");
}
ReifiedRankedShapedTypeDims resultShape;
ReifyRankedShapedTypeOpInterface reifyShapedTypeInterface =
dyn_cast<ReifyRankedShapedTypeOpInterface>(padOp.getOperation());
if (failed(reifyShapedTypeInterface.reifyResultShapes(rewriter,
resultShape)) ||
resultShape.size() != 1) {
return rewriter.notifyMatchFailure(
padOp, "failed to get shape of pad op result");
}
Location loc = padOp.getLoc();
// Create the tensor of same size as output of the pad op.
RankedTensorType padResultType = padOp.getResultType();
auto resultSizes = getAsOpFoldResult(resultShape[0]);
auto initTensor = rewriter.create<linalg::InitTensorOp>(
loc, resultSizes, padResultType.getElementType());
// Fill the tensor with the pad value.
// TODO: There is an option to fill only the boundaries. For now just
// filling the whole tensor.
auto fillTensor =
rewriter.create<linalg::FillOp>(loc, padValue, initTensor.getResult());
// Construct a slice of the fill result that is to be replaced with the
// result of the generic op. The low pad values are the offsets, the size of
// the source is the size of the slice.
// TODO: This insert/extract could be potentially made a utility method.
unsigned resultNumber = source.cast<OpResult>().getResultNumber();
SmallVector<OpFoldResult> offsets = padOp.getMixedLowPad();
SmallVector<OpFoldResult> sizes;
sizes.reserve(offsets.size());
for (const auto &shape : llvm::enumerate(
source.getType().cast<RankedTensorType>().getShape())) {
if (ShapedType::isDynamic(shape.value())) {
sizes.push_back(
rewriter.create<tensor::DimOp>(loc, source, shape.index())
.getResult());
} else {
sizes.push_back(rewriter.getIndexAttr(shape.value()));
}
}
SmallVector<OpFoldResult> strides(offsets.size(), rewriter.getIndexAttr(1));
auto slice = rewriter.create<tensor::ExtractSliceOp>(
loc, fillTensor.getResult(0), offsets, sizes, strides);
// Clone the generic op.
auto clonedOp =
cast<linalg::GenericOp>(rewriter.clone(*linalgOp.getOperation()));
clonedOp.setOutputOperand(resultNumber, slice.getResult());
// Insert it back into the result of the fill.
rewriter.replaceOpWithNewOp<tensor::InsertSliceOp>(
padOp, clonedOp.getResult(resultNumber), fillTensor.getResult(0),
offsets, sizes, strides);
return success();
}
};
} // namespace
void mlir::linalg::populateFuseTensorPadWithProducerLinalgOpPatterns(
RewritePatternSet &patterns) {
patterns.add<FusePadOp>(patterns.getContext());
}