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
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
|
//===- LowerVectorBroadcast.cpp - Lower 'vector.broadcast' operation ------===//
//
// 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 target-independent rewrites and utilities to lower the
// 'vector.broadcast' operation.
//
//===----------------------------------------------------------------------===//
#include "mlir/Dialect/Affine/IR/AffineOps.h"
#include "mlir/Dialect/Arith/IR/Arith.h"
#include "mlir/Dialect/Arith/Utils/Utils.h"
#include "mlir/Dialect/Linalg/IR/Linalg.h"
#include "mlir/Dialect/MemRef/IR/MemRef.h"
#include "mlir/Dialect/SCF/IR/SCF.h"
#include "mlir/Dialect/Tensor/IR/Tensor.h"
#include "mlir/Dialect/Utils/IndexingUtils.h"
#include "mlir/Dialect/Utils/StructuredOpsUtils.h"
#include "mlir/Dialect/Vector/IR/VectorOps.h"
#include "mlir/Dialect/Vector/Transforms/LoweringPatterns.h"
#include "mlir/Dialect/Vector/Transforms/VectorRewritePatterns.h"
#include "mlir/Dialect/Vector/Utils/VectorUtils.h"
#include "mlir/IR/BuiltinAttributeInterfaces.h"
#include "mlir/IR/BuiltinTypes.h"
#include "mlir/IR/ImplicitLocOpBuilder.h"
#include "mlir/IR/Location.h"
#include "mlir/IR/Matchers.h"
#include "mlir/IR/PatternMatch.h"
#include "mlir/IR/TypeUtilities.h"
#include "mlir/Interfaces/VectorInterfaces.h"
#include "mlir/Support/LogicalResult.h"
#define DEBUG_TYPE "vector-broadcast-lowering"
using namespace mlir;
using namespace mlir::vector;
namespace {
/// Progressive lowering of BroadcastOp.
class BroadcastOpLowering : public OpRewritePattern<vector::BroadcastOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(vector::BroadcastOp op,
PatternRewriter &rewriter) const override {
auto loc = op.getLoc();
VectorType dstType = op.getResultVectorType();
VectorType srcType = dyn_cast<VectorType>(op.getSourceType());
Type eltType = dstType.getElementType();
// Scalar to any vector can use splat.
if (!srcType) {
rewriter.replaceOpWithNewOp<vector::SplatOp>(op, dstType, op.getSource());
return success();
}
// Determine rank of source and destination.
int64_t srcRank = srcType.getRank();
int64_t dstRank = dstType.getRank();
// Stretching scalar inside vector (e.g. vector<1xf32>) can use splat.
if (srcRank <= 1 && dstRank == 1) {
Value ext;
if (srcRank == 0)
ext = rewriter.create<vector::ExtractElementOp>(loc, op.getSource());
else
ext = rewriter.create<vector::ExtractOp>(loc, op.getSource(), 0);
rewriter.replaceOpWithNewOp<vector::SplatOp>(op, dstType, ext);
return success();
}
// Duplicate this rank.
// For example:
// %x = broadcast %y : k-D to n-D, k < n
// becomes:
// %b = broadcast %y : k-D to (n-1)-D
// %x = [%b,%b,%b,%b] : n-D
// becomes:
// %b = [%y,%y] : (n-1)-D
// %x = [%b,%b,%b,%b] : n-D
if (srcRank < dstRank) {
// Duplication.
VectorType resType =
VectorType::get(dstType.getShape().drop_front(), eltType);
Value bcst =
rewriter.create<vector::BroadcastOp>(loc, resType, op.getSource());
Value result = rewriter.create<arith::ConstantOp>(
loc, dstType, rewriter.getZeroAttr(dstType));
for (int64_t d = 0, dim = dstType.getDimSize(0); d < dim; ++d)
result = rewriter.create<vector::InsertOp>(loc, bcst, result, d);
rewriter.replaceOp(op, result);
return success();
}
// Find non-matching dimension, if any.
assert(srcRank == dstRank);
int64_t m = -1;
for (int64_t r = 0; r < dstRank; r++)
if (srcType.getDimSize(r) != dstType.getDimSize(r)) {
m = r;
break;
}
// All trailing dimensions are the same. Simply pass through.
if (m == -1) {
rewriter.replaceOp(op, op.getSource());
return success();
}
// Any non-matching dimension forces a stretch along this rank.
// For example:
// %x = broadcast %y : vector<4x1x2xf32> to vector<4x2x2xf32>
// becomes:
// %a = broadcast %y[0] : vector<1x2xf32> to vector<2x2xf32>
// %b = broadcast %y[1] : vector<1x2xf32> to vector<2x2xf32>
// %c = broadcast %y[2] : vector<1x2xf32> to vector<2x2xf32>
// %d = broadcast %y[3] : vector<1x2xf32> to vector<2x2xf32>
// %x = [%a,%b,%c,%d]
// becomes:
// %u = broadcast %y[0][0] : vector<2xf32> to vector <2x2xf32>
// %v = broadcast %y[1][0] : vector<2xf32> to vector <2x2xf32>
// %a = [%u, %v]
// ..
// %x = [%a,%b,%c,%d]
VectorType resType =
VectorType::get(dstType.getShape().drop_front(), eltType);
Value result = rewriter.create<arith::ConstantOp>(
loc, dstType, rewriter.getZeroAttr(dstType));
if (m == 0) {
// Stetch at start.
Value ext = rewriter.create<vector::ExtractOp>(loc, op.getSource(), 0);
Value bcst = rewriter.create<vector::BroadcastOp>(loc, resType, ext);
for (int64_t d = 0, dim = dstType.getDimSize(0); d < dim; ++d)
result = rewriter.create<vector::InsertOp>(loc, bcst, result, d);
} else {
// Stetch not at start.
for (int64_t d = 0, dim = dstType.getDimSize(0); d < dim; ++d) {
Value ext = rewriter.create<vector::ExtractOp>(loc, op.getSource(), d);
Value bcst = rewriter.create<vector::BroadcastOp>(loc, resType, ext);
result = rewriter.create<vector::InsertOp>(loc, bcst, result, d);
}
}
rewriter.replaceOp(op, result);
return success();
}
};
} // namespace
void mlir::vector::populateVectorBroadcastLoweringPatterns(
RewritePatternSet &patterns, PatternBenefit benefit) {
patterns.add<BroadcastOpLowering>(patterns.getContext(), benefit);
}
|
