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
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
|
//===- IndexingUtils.cpp - Helpers related to index computations ----------===//
//
// 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
//
//===----------------------------------------------------------------------===//
#include "mlir/Dialect/Utils/IndexingUtils.h"
#include "mlir/Dialect/Utils/StaticValueUtils.h"
#include "mlir/IR/AffineExpr.h"
#include "mlir/IR/Builders.h"
#include "mlir/IR/BuiltinAttributes.h"
#include "mlir/IR/MLIRContext.h"
#include "llvm/ADT/STLExtras.h"
#include <numeric>
#include <optional>
using namespace mlir;
template <typename ExprType>
SmallVector<ExprType> computeSuffixProductImpl(ArrayRef<ExprType> sizes,
ExprType unit) {
if (sizes.empty())
return {};
SmallVector<ExprType> strides(sizes.size(), unit);
for (int64_t r = strides.size() - 2; r >= 0; --r)
strides[r] = strides[r + 1] * sizes[r + 1];
return strides;
}
template <typename ExprType>
SmallVector<ExprType> computeElementwiseMulImpl(ArrayRef<ExprType> v1,
ArrayRef<ExprType> v2) {
// Early exit if both are empty, let zip_equal fail if only 1 is empty.
if (v1.empty() && v2.empty())
return {};
SmallVector<ExprType> result;
for (auto it : llvm::zip_equal(v1, v2))
result.push_back(std::get<0>(it) * std::get<1>(it));
return result;
}
template <typename ExprType>
ExprType linearizeImpl(ArrayRef<ExprType> offsets, ArrayRef<ExprType> basis,
ExprType zero) {
assert(offsets.size() == basis.size());
ExprType linearIndex = zero;
for (unsigned idx = 0, e = basis.size(); idx < e; ++idx)
linearIndex = linearIndex + offsets[idx] * basis[idx];
return linearIndex;
}
template <typename ExprType, typename DivOpTy>
SmallVector<ExprType> delinearizeImpl(ExprType linearIndex,
ArrayRef<ExprType> strides,
DivOpTy divOp) {
int64_t rank = strides.size();
SmallVector<ExprType> offsets(rank);
for (int64_t r = 0; r < rank; ++r) {
offsets[r] = divOp(linearIndex, strides[r]);
linearIndex = linearIndex % strides[r];
}
return offsets;
}
//===----------------------------------------------------------------------===//
// Utils that operate on static integer values.
//===----------------------------------------------------------------------===//
SmallVector<int64_t> mlir::computeSuffixProduct(ArrayRef<int64_t> sizes) {
assert(llvm::all_of(sizes, [](int64_t s) { return s >= 0; }) &&
"sizes must be nonnegative");
int64_t unit = 1;
return ::computeSuffixProductImpl(sizes, unit);
}
SmallVector<int64_t> mlir::computeElementwiseMul(ArrayRef<int64_t> v1,
ArrayRef<int64_t> v2) {
return computeElementwiseMulImpl(v1, v2);
}
int64_t mlir::computeSum(ArrayRef<int64_t> basis) {
assert(llvm::all_of(basis, [](int64_t s) { return s > 0; }) &&
"basis must be nonnegative");
if (basis.empty())
return 0;
return std::accumulate(basis.begin(), basis.end(), 1, std::plus<int64_t>());
}
int64_t mlir::computeProduct(ArrayRef<int64_t> basis) {
assert(llvm::all_of(basis, [](int64_t s) { return s > 0; }) &&
"basis must be nonnegative");
if (basis.empty())
return 0;
return std::accumulate(basis.begin(), basis.end(), 1,
std::multiplies<int64_t>());
}
int64_t mlir::linearize(ArrayRef<int64_t> offsets, ArrayRef<int64_t> basis) {
assert(llvm::all_of(basis, [](int64_t s) { return s > 0; }) &&
"basis must be nonnegative");
int64_t zero = 0;
return linearizeImpl(offsets, basis, zero);
}
SmallVector<int64_t> mlir::delinearize(int64_t linearIndex,
ArrayRef<int64_t> strides) {
assert(llvm::all_of(strides, [](int64_t s) { return s > 0; }) &&
"strides must be nonnegative");
return delinearizeImpl(linearIndex, strides,
[](int64_t e1, int64_t e2) { return e1 / e2; });
}
std::optional<SmallVector<int64_t>>
mlir::computeShapeRatio(ArrayRef<int64_t> shape, ArrayRef<int64_t> subShape) {
if (shape.size() < subShape.size())
return std::nullopt;
assert(llvm::all_of(shape, [](int64_t s) { return s > 0; }) &&
"shape must be nonnegative");
assert(llvm::all_of(subShape, [](int64_t s) { return s > 0; }) &&
"subShape must be nonnegative");
// Starting from the end, compute the integer divisors.
std::vector<int64_t> result;
result.reserve(shape.size());
for (auto [size, subSize] :
llvm::zip(llvm::reverse(shape), llvm::reverse(subShape))) {
// If integral division does not occur, return and let the caller decide.
if (size % subSize != 0)
return std::nullopt;
result.push_back(size / subSize);
}
// At this point we computed the ratio (in reverse) for the common size.
// Fill with the remaining entries from the shape (still in reverse).
int commonSize = subShape.size();
std::copy(shape.rbegin() + commonSize, shape.rend(),
std::back_inserter(result));
// Reverse again to get it back in the proper order and return.
return SmallVector<int64_t>{result.rbegin(), result.rend()};
}
//===----------------------------------------------------------------------===//
// Utils that operate on AffineExpr.
//===----------------------------------------------------------------------===//
SmallVector<AffineExpr> mlir::computeSuffixProduct(ArrayRef<AffineExpr> sizes) {
if (sizes.empty())
return {};
AffineExpr unit = getAffineConstantExpr(1, sizes.front().getContext());
return ::computeSuffixProductImpl(sizes, unit);
}
SmallVector<AffineExpr> mlir::computeElementwiseMul(ArrayRef<AffineExpr> v1,
ArrayRef<AffineExpr> v2) {
return computeElementwiseMulImpl(v1, v2);
}
AffineExpr mlir::computeSum(MLIRContext *ctx, ArrayRef<AffineExpr> basis) {
if (basis.empty())
return getAffineConstantExpr(0, ctx);
return std::accumulate(basis.begin(), basis.end(),
getAffineConstantExpr(0, ctx),
std::plus<AffineExpr>());
}
AffineExpr mlir::computeProduct(MLIRContext *ctx, ArrayRef<AffineExpr> basis) {
if (basis.empty())
return getAffineConstantExpr(1, ctx);
return std::accumulate(basis.begin(), basis.end(),
getAffineConstantExpr(1, ctx),
std::multiplies<AffineExpr>());
}
AffineExpr mlir::linearize(MLIRContext *ctx, ArrayRef<AffineExpr> offsets,
ArrayRef<AffineExpr> basis) {
AffineExpr zero = getAffineConstantExpr(0, ctx);
return linearizeImpl(offsets, basis, zero);
}
AffineExpr mlir::linearize(MLIRContext *ctx, ArrayRef<AffineExpr> offsets,
ArrayRef<int64_t> basis) {
return linearize(ctx, offsets, getAffineConstantExprs(basis, ctx));
}
SmallVector<AffineExpr> mlir::delinearize(AffineExpr linearIndex,
ArrayRef<AffineExpr> strides) {
return delinearizeImpl(
linearIndex, strides,
[](AffineExpr e1, AffineExpr e2) { return e1.floorDiv(e2); });
}
SmallVector<AffineExpr> mlir::delinearize(AffineExpr linearIndex,
ArrayRef<int64_t> strides) {
MLIRContext *ctx = linearIndex.getContext();
return delinearize(linearIndex, getAffineConstantExprs(strides, ctx));
}
//===----------------------------------------------------------------------===//
// Permutation utils.
//===----------------------------------------------------------------------===//
SmallVector<int64_t>
mlir::invertPermutationVector(ArrayRef<int64_t> permutation) {
assert(llvm::all_of(permutation, [](int64_t s) { return s >= 0; }) &&
"permutation must be non-negative");
SmallVector<int64_t> inversion(permutation.size());
for (const auto &pos : llvm::enumerate(permutation)) {
inversion[pos.value()] = pos.index();
}
return inversion;
}
bool mlir::isIdentityPermutation(ArrayRef<int64_t> permutation) {
for (auto i : llvm::seq<int64_t>(0, permutation.size()))
if (permutation[i] != i)
return false;
return true;
}
bool mlir::isPermutationVector(ArrayRef<int64_t> interchange) {
assert(llvm::all_of(interchange, [](int64_t s) { return s >= 0; }) &&
"permutation must be non-negative");
llvm::SmallDenseSet<int64_t, 4> seenVals;
for (auto val : interchange) {
if (seenVals.count(val))
return false;
seenVals.insert(val);
}
return seenVals.size() == interchange.size();
}
SmallVector<int64_t>
mlir::computePermutationVector(int64_t permSize, ArrayRef<int64_t> positions,
ArrayRef<int64_t> desiredPositions) {
SmallVector<int64_t> res(permSize, -1);
DenseSet<int64_t> seen;
for (auto [pos, desiredPos] : llvm::zip_equal(positions, desiredPositions)) {
res[desiredPos] = pos;
seen.insert(pos);
}
int64_t nextPos = 0;
for (int64_t &entry : res) {
if (entry != -1)
continue;
while (seen.contains(nextPos))
++nextPos;
entry = nextPos;
++nextPos;
}
return res;
}
SmallVector<int64_t> mlir::getI64SubArray(ArrayAttr arrayAttr,
unsigned dropFront,
unsigned dropBack) {
assert(arrayAttr.size() > dropFront + dropBack && "Out of bounds");
auto range = arrayAttr.getAsRange<IntegerAttr>();
SmallVector<int64_t> res;
res.reserve(arrayAttr.size() - dropFront - dropBack);
for (auto it = range.begin() + dropFront, eit = range.end() - dropBack;
it != eit; ++it)
res.push_back((*it).getValue().getSExtValue());
return res;
}
// TODO: do we have any common utily for this?
static MLIRContext *getContext(OpFoldResult val) {
assert(val && "Invalid value");
if (auto attr = dyn_cast<Attribute>(val)) {
return attr.getContext();
}
return cast<Value>(val).getContext();
}
std::pair<AffineExpr, SmallVector<OpFoldResult>>
mlir::computeLinearIndex(OpFoldResult sourceOffset,
ArrayRef<OpFoldResult> strides,
ArrayRef<OpFoldResult> indices) {
assert(strides.size() == indices.size());
auto sourceRank = static_cast<unsigned>(strides.size());
// Hold the affine symbols and values for the computation of the offset.
SmallVector<OpFoldResult> values(2 * sourceRank + 1);
SmallVector<AffineExpr> symbols(2 * sourceRank + 1);
bindSymbolsList(getContext(sourceOffset), MutableArrayRef{symbols});
AffineExpr expr = symbols.front();
values[0] = sourceOffset;
for (unsigned i = 0; i < sourceRank; ++i) {
// Compute the stride.
OpFoldResult origStride = strides[i];
// Build up the computation of the offset.
unsigned baseIdxForDim = 1 + 2 * i;
unsigned subOffsetForDim = baseIdxForDim;
unsigned origStrideForDim = baseIdxForDim + 1;
expr = expr + symbols[subOffsetForDim] * symbols[origStrideForDim];
values[subOffsetForDim] = indices[i];
values[origStrideForDim] = origStride;
}
return {expr, values};
}
std::pair<AffineExpr, SmallVector<OpFoldResult>>
mlir::computeLinearIndex(OpFoldResult sourceOffset, ArrayRef<int64_t> strides,
ArrayRef<Value> indices) {
return computeLinearIndex(
sourceOffset, getAsIndexOpFoldResult(sourceOffset.getContext(), strides),
getAsOpFoldResult(ValueRange(indices)));
}
//===----------------------------------------------------------------------===//
// TileOffsetRange
//===----------------------------------------------------------------------===//
/// Apply left-padding by 1 to the tile shape if required.
static SmallVector<int64_t> padTileShapeToSize(ArrayRef<int64_t> tileShape,
unsigned paddedSize) {
assert(tileShape.size() <= paddedSize &&
"expected tileShape to <= paddedSize");
if (tileShape.size() == paddedSize)
return to_vector(tileShape);
SmallVector<int64_t> result(paddedSize - tileShape.size(), 1);
llvm::append_range(result, tileShape);
return result;
}
mlir::detail::TileOffsetRangeImpl::TileOffsetRangeImpl(
ArrayRef<int64_t> shape, ArrayRef<int64_t> tileShape,
ArrayRef<int64_t> loopOrder)
: tileShape(padTileShapeToSize(tileShape, shape.size())),
inverseLoopOrder(invertPermutationVector(loopOrder)),
sliceStrides(shape.size()) {
// Divide the shape by the tile shape.
std::optional<SmallVector<int64_t>> shapeRatio =
mlir::computeShapeRatio(shape, tileShape);
assert(shapeRatio && shapeRatio->size() == shape.size() &&
"target shape does not evenly divide the original shape");
assert(isPermutationVector(loopOrder) && loopOrder.size() == shape.size() &&
"expected loop order to be a permutation of rank equal to outer "
"shape");
maxLinearIndex = mlir::computeMaxLinearIndex(*shapeRatio);
mlir::applyPermutationToVector(*shapeRatio, loopOrder);
sliceStrides = mlir::computeStrides(*shapeRatio);
}
SmallVector<int64_t> mlir::detail::TileOffsetRangeImpl::getStaticTileOffsets(
int64_t linearIndex) const {
SmallVector<int64_t> tileCoords = applyPermutation(
delinearize(linearIndex, sliceStrides), inverseLoopOrder);
return computeElementwiseMul(tileCoords, tileShape);
}
SmallVector<AffineExpr>
mlir::detail::TileOffsetRangeImpl::getDynamicTileOffsets(
AffineExpr linearIndex) const {
MLIRContext *ctx = linearIndex.getContext();
SmallVector<AffineExpr> tileCoords = applyPermutation(
delinearize(linearIndex, sliceStrides), inverseLoopOrder);
return mlir::computeElementwiseMul(tileCoords,
getAffineConstantExprs(tileShape, ctx));
}
|