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
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
|
//===- LowerVectorScam.cpp - Lower 'vector.scan' 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.scan' 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/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;
/// This function constructs the appropriate integer or float
/// operation given the vector combining kind and operands. The
/// supported int operations are : add, mul, min (signed/unsigned),
/// max(signed/unsigned), and, or, xor. The supported float
/// operations are : add, mul, min and max.
static Value genOperator(Location loc, Value x, Value y,
vector::CombiningKind kind,
PatternRewriter &rewriter) {
using vector::CombiningKind;
auto elType = cast<VectorType>(x.getType()).getElementType();
bool isInt = elType.isIntOrIndex();
Value combinedResult{nullptr};
switch (kind) {
case CombiningKind::ADD:
if (isInt)
combinedResult = rewriter.create<arith::AddIOp>(loc, x, y);
else
combinedResult = rewriter.create<arith::AddFOp>(loc, x, y);
break;
case CombiningKind::MUL:
if (isInt)
combinedResult = rewriter.create<arith::MulIOp>(loc, x, y);
else
combinedResult = rewriter.create<arith::MulFOp>(loc, x, y);
break;
case CombiningKind::MINUI:
combinedResult = rewriter.create<arith::MinUIOp>(loc, x, y);
break;
case CombiningKind::MINSI:
combinedResult = rewriter.create<arith::MinSIOp>(loc, x, y);
break;
case CombiningKind::MAXUI:
combinedResult = rewriter.create<arith::MaxUIOp>(loc, x, y);
break;
case CombiningKind::MAXSI:
combinedResult = rewriter.create<arith::MaxSIOp>(loc, x, y);
break;
case CombiningKind::AND:
combinedResult = rewriter.create<arith::AndIOp>(loc, x, y);
break;
case CombiningKind::OR:
combinedResult = rewriter.create<arith::OrIOp>(loc, x, y);
break;
case CombiningKind::XOR:
combinedResult = rewriter.create<arith::XOrIOp>(loc, x, y);
break;
case CombiningKind::MINF:
combinedResult = rewriter.create<arith::MinFOp>(loc, x, y);
break;
case CombiningKind::MAXF:
combinedResult = rewriter.create<arith::MaxFOp>(loc, x, y);
break;
}
return combinedResult;
}
/// This function checks to see if the vector combining kind
/// is consistent with the integer or float element type.
static bool isValidKind(bool isInt, vector::CombiningKind kind) {
using vector::CombiningKind;
enum class KindType { FLOAT, INT, INVALID };
KindType type{KindType::INVALID};
switch (kind) {
case CombiningKind::MINF:
case CombiningKind::MAXF:
type = KindType::FLOAT;
break;
case CombiningKind::MINUI:
case CombiningKind::MINSI:
case CombiningKind::MAXUI:
case CombiningKind::MAXSI:
case CombiningKind::AND:
case CombiningKind::OR:
case CombiningKind::XOR:
type = KindType::INT;
break;
case CombiningKind::ADD:
case CombiningKind::MUL:
type = isInt ? KindType::INT : KindType::FLOAT;
break;
}
bool isValidIntKind = (type == KindType::INT) && isInt;
bool isValidFloatKind = (type == KindType::FLOAT) && (!isInt);
return (isValidIntKind || isValidFloatKind);
}
namespace {
/// Convert vector.scan op into arith ops and vector.insert_strided_slice /
/// vector.extract_strided_slice.
///
/// Example:
///
/// ```
/// %0:2 = vector.scan <add>, %arg0, %arg1
/// {inclusive = true, reduction_dim = 1} :
/// (vector<2x3xi32>, vector<2xi32>) to (vector<2x3xi32>, vector<2xi32>)
/// ```
///
/// is converted to:
///
/// ```
/// %cst = arith.constant dense<0> : vector<2x3xi32>
/// %0 = vector.extract_strided_slice %arg0
/// {offsets = [0, 0], sizes = [2, 1], strides = [1, 1]}
/// : vector<2x3xi32> to vector<2x1xi32>
/// %1 = vector.insert_strided_slice %0, %cst
/// {offsets = [0, 0], strides = [1, 1]}
/// : vector<2x1xi32> into vector<2x3xi32>
/// %2 = vector.extract_strided_slice %arg0
/// {offsets = [0, 1], sizes = [2, 1], strides = [1, 1]}
/// : vector<2x3xi32> to vector<2x1xi32>
/// %3 = arith.muli %0, %2 : vector<2x1xi32>
/// %4 = vector.insert_strided_slice %3, %1
/// {offsets = [0, 1], strides = [1, 1]}
/// : vector<2x1xi32> into vector<2x3xi32>
/// %5 = vector.extract_strided_slice %arg0
/// {offsets = [0, 2], sizes = [2, 1], strides = [1, 1]}
/// : vector<2x3xi32> to vector<2x1xi32>
/// %6 = arith.muli %3, %5 : vector<2x1xi32>
/// %7 = vector.insert_strided_slice %6, %4
/// {offsets = [0, 2], strides = [1, 1]}
/// : vector<2x1xi32> into vector<2x3xi32>
/// %8 = vector.shape_cast %6 : vector<2x1xi32> to vector<2xi32>
/// return %7, %8 : vector<2x3xi32>, vector<2xi32>
/// ```
struct ScanToArithOps : public OpRewritePattern<vector::ScanOp> {
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(vector::ScanOp scanOp,
PatternRewriter &rewriter) const override {
auto loc = scanOp.getLoc();
VectorType destType = scanOp.getDestType();
ArrayRef<int64_t> destShape = destType.getShape();
auto elType = destType.getElementType();
bool isInt = elType.isIntOrIndex();
if (!isValidKind(isInt, scanOp.getKind()))
return failure();
VectorType resType = VectorType::get(destShape, elType);
Value result = rewriter.create<arith::ConstantOp>(
loc, resType, rewriter.getZeroAttr(resType));
int64_t reductionDim = scanOp.getReductionDim();
bool inclusive = scanOp.getInclusive();
int64_t destRank = destType.getRank();
VectorType initialValueType = scanOp.getInitialValueType();
int64_t initialValueRank = initialValueType.getRank();
SmallVector<int64_t> reductionShape(destShape.begin(), destShape.end());
reductionShape[reductionDim] = 1;
VectorType reductionType = VectorType::get(reductionShape, elType);
SmallVector<int64_t> offsets(destRank, 0);
SmallVector<int64_t> strides(destRank, 1);
SmallVector<int64_t> sizes(destShape.begin(), destShape.end());
sizes[reductionDim] = 1;
ArrayAttr scanSizes = rewriter.getI64ArrayAttr(sizes);
ArrayAttr scanStrides = rewriter.getI64ArrayAttr(strides);
Value lastOutput, lastInput;
for (int i = 0; i < destShape[reductionDim]; i++) {
offsets[reductionDim] = i;
ArrayAttr scanOffsets = rewriter.getI64ArrayAttr(offsets);
Value input = rewriter.create<vector::ExtractStridedSliceOp>(
loc, reductionType, scanOp.getSource(), scanOffsets, scanSizes,
scanStrides);
Value output;
if (i == 0) {
if (inclusive) {
output = input;
} else {
if (initialValueRank == 0) {
// ShapeCastOp cannot handle 0-D vectors
output = rewriter.create<vector::BroadcastOp>(
loc, input.getType(), scanOp.getInitialValue());
} else {
output = rewriter.create<vector::ShapeCastOp>(
loc, input.getType(), scanOp.getInitialValue());
}
}
} else {
Value y = inclusive ? input : lastInput;
output = genOperator(loc, lastOutput, y, scanOp.getKind(), rewriter);
assert(output != nullptr);
}
result = rewriter.create<vector::InsertStridedSliceOp>(
loc, output, result, offsets, strides);
lastOutput = output;
lastInput = input;
}
Value reduction;
if (initialValueRank == 0) {
Value v = rewriter.create<vector::ExtractOp>(loc, lastOutput, 0);
reduction =
rewriter.create<vector::BroadcastOp>(loc, initialValueType, v);
} else {
reduction = rewriter.create<vector::ShapeCastOp>(loc, initialValueType,
lastOutput);
}
rewriter.replaceOp(scanOp, {result, reduction});
return success();
}
};
} // namespace
void mlir::vector::populateVectorScanLoweringPatterns(
RewritePatternSet &patterns, PatternBenefit benefit) {
patterns.add<ScanToArithOps>(patterns.getContext(), benefit);
}
|
