Utils.cpp
17.2 KB
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
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
//===- Utils.cpp - Utilities to support the Linalg dialect ----------------===//
//
// 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 utilities for the Linalg dialect.
//
//===----------------------------------------------------------------------===//
#include "mlir/Dialect/Linalg/Utils/Utils.h"
#include "mlir/Dialect/Affine/EDSC/Intrinsics.h"
#include "mlir/Dialect/Affine/IR/AffineOps.h"
#include "mlir/Dialect/Linalg/IR/LinalgOps.h"
#include "mlir/Dialect/Linalg/IR/LinalgTypes.h"
#include "mlir/Dialect/SCF/EDSC/Builders.h"
#include "mlir/Dialect/SCF/SCF.h"
#include "mlir/Dialect/StandardOps/IR/Ops.h"
#include "mlir/IR/AffineExpr.h"
#include "mlir/IR/AffineMap.h"
#include "mlir/IR/Matchers.h"
#include "mlir/IR/OpImplementation.h"
#include "mlir/Pass/Pass.h"
#include "mlir/Transforms/FoldUtils.h"
using namespace mlir;
using namespace mlir::linalg;
using namespace mlir::scf;
Optional<RegionMatcher::BinaryOpKind>
RegionMatcher::matchAsScalarBinaryOp(GenericOp op) {
auto ®ion = op.region();
if (!llvm::hasSingleElement(region))
return llvm::None;
Block &block = region.front();
if (block.getNumArguments() != 2 ||
!block.getArgument(0).getType().isSignlessIntOrFloat() ||
!block.getArgument(1).getType().isSignlessIntOrFloat())
return llvm::None;
auto &ops = block.getOperations();
if (!llvm::hasSingleElement(block.without_terminator()))
return llvm::None;
using mlir::matchers::m_Val;
auto a = m_Val(block.getArgument(0));
auto b = m_Val(block.getArgument(1));
auto addPattern = m_Op<linalg::YieldOp>(m_Op<AddIOp>(a, b));
if (addPattern.match(&ops.back()))
return BinaryOpKind::IAdd;
return llvm::None;
}
static Value emitOrFoldComposedAffineApply(OpBuilder &b, Location loc,
AffineMap map,
ValueRange operandsRef,
OperationFolder *folder) {
SmallVector<Value, 4> operands(operandsRef.begin(), operandsRef.end());
fullyComposeAffineMapAndOperands(&map, &operands);
canonicalizeMapAndOperands(&map, &operands);
return folder ? folder->create<AffineApplyOp>(b, loc, map, operands)
: b.create<AffineApplyOp>(loc, map, operands);
}
SmallVector<Value, 4> mlir::linalg::applyMapToValues(OpBuilder &b, Location loc,
AffineMap map,
ValueRange values,
OperationFolder *folder) {
SmallVector<Value, 4> res;
res.reserve(map.getNumResults());
unsigned numDims = map.getNumDims(), numSym = map.getNumSymbols();
// For each `expr` in `map`, applies the `expr` to the values extracted from
// ranges. If the resulting application can be folded into a Value, the
// folding occurs eagerly. Otherwise, an affine.apply operation is emitted.
for (auto expr : map.getResults()) {
AffineMap map = AffineMap::get(numDims, numSym, expr);
res.push_back(emitOrFoldComposedAffineApply(b, loc, map, values, folder));
}
return res;
}
/// Returns all the operands of `linalgOp` that are not views.
/// Asserts that these operands are value types to allow transformations like
/// tiling to just use the values when cloning `linalgOp`.
SmallVector<Value, 4>
mlir::linalg::getAssumedNonViewOperands(LinalgOp linalgOp) {
auto *op = linalgOp.getOperation();
unsigned numViews = linalgOp.getNumInputsAndOutputs();
unsigned nOperands = op->getNumOperands() - numViews;
SmallVector<Value, 4> res;
res.reserve(nOperands);
for (unsigned i = 0; i < nOperands; ++i) {
res.push_back(op->getOperand(numViews + i));
auto t = res.back().getType();
(void)t;
assert((t.isSignlessIntOrIndexOrFloat() || t.isa<VectorType>()) &&
"expected scalar or vector type");
}
return res;
}
bool mlir::linalg::isParallelIteratorType(Attribute attr) {
if (auto strAttr = attr.dyn_cast<StringAttr>()) {
return strAttr.getValue() == getParallelIteratorTypeName();
}
return false;
}
bool mlir::linalg::isReductionIteratorType(Attribute attr) {
if (auto strAttr = attr.dyn_cast<StringAttr>()) {
return strAttr.getValue() == getReductionIteratorTypeName();
}
return false;
}
bool mlir::linalg::isWindowIteratorType(Attribute attr) {
if (auto strAttr = attr.dyn_cast<StringAttr>()) {
return strAttr.getValue() == getWindowIteratorTypeName();
}
return false;
}
/// Explicit instantiation of loop nest generator for different loop types.
template struct mlir::linalg::GenerateLoopNest<scf::ForOp>;
template struct mlir::linalg::GenerateLoopNest<scf::ParallelOp>;
template struct mlir::linalg::GenerateLoopNest<AffineForOp>;
/// Given a list of subview ranges, extract individual values for lower, upper
/// bounds and steps and put them into the corresponding vectors.
static void unpackRanges(ArrayRef<Range> ranges, SmallVectorImpl<Value> &lbs,
SmallVectorImpl<Value> &ubs,
SmallVectorImpl<Value> &steps) {
for (Range range : ranges) {
lbs.emplace_back(range.offset);
ubs.emplace_back(range.size);
steps.emplace_back(range.stride);
}
}
namespace mlir {
namespace linalg {
/// Return the linearized list of all view dimensions in a linalgOp.
SmallVector<Value, 8> getViewSizes(OpBuilder &builder, LinalgOp linalgOp) {
auto loc = linalgOp.getLoc();
SmallVector<Value, 8> res;
SmallVector<unsigned, 4> ranks;
for (auto v : linalgOp.getInputsAndOutputBuffers()) {
MemRefType t = v.getType().template cast<MemRefType>();
ranks.push_back(t.getRank());
for (unsigned i = 0; i < t.getRank(); ++i)
res.push_back(builder.create<DimOp>(loc, v, i));
}
auto attr = linalgOp.template getAttrOfType<IntegerAttr>("symbol_source");
if (attr) {
// Find the correct position for inserting values for symbols.
unsigned numSymb = ranks[attr.getInt()], symbolsPos = 0;
for (unsigned idx = 0; idx < attr.getInt(); idx++)
symbolsPos += ranks[idx];
// Append the end of the value list that corresponds to the
// values mapping to symbols. Since inside concatinated map symbols are
// repeated we have to repeat the sizes as well.
// Reserve is mandatory to avoid a potential undefined behavior with
// pushing back to smallvector from itself.
res.reserve(res.size() + ranks.size() * numSymb);
for (unsigned idx = 0, s = ranks.size(); idx < s; ++idx)
for (unsigned idx2 = 0; idx2 < numSymb; ++idx2)
res.push_back(res[symbolsPos + idx2]);
}
return res;
}
Optional<SmallVector<Value, 4>>
getLoopRanges(OpBuilder &builder, LinalgOp linalgOp, OperationFolder *folder) {
SmallVector<Value, 8> viewSizes = getViewSizes(builder, linalgOp);
AffineMap invertedMap =
inversePermutation(concatAffineMaps(linalgOp.getIndexingMaps()));
if (!invertedMap)
return {};
return applyMapToValues(builder, linalgOp.getLoc(), invertedMap, viewSizes,
folder);
}
/// Specialization to build an scf "for" nest.
template <>
void GenerateLoopNest<scf::ForOp>::doit(
ArrayRef<Range> loopRanges, ValueRange iterArgInitValues,
ArrayRef<Attribute> iteratorTypes,
function_ref<scf::ValueVector(ValueRange, ValueRange)> bodyBuilderFn,
Optional<LinalgLoopDistributionOptions>) {
SmallVector<Value, 4> lbs, ubs, steps;
unpackRanges(loopRanges, lbs, ubs, steps);
edsc::loopNestBuilder(lbs, ubs, steps, iterArgInitValues, bodyBuilderFn);
}
/// Specialization to build affine "for" nest.
template <>
void GenerateLoopNest<AffineForOp>::doit(
ArrayRef<Range> loopRanges, ValueRange iterArgInitValues,
ArrayRef<Attribute> iteratorTypes,
function_ref<scf::ValueVector(ValueRange, ValueRange)> bodyBuilderFn,
Optional<LinalgLoopDistributionOptions>) {
assert(iterArgInitValues.empty() && "unexpected AffineForOp init values");
SmallVector<Value, 4> lbs, ubs, steps;
unpackRanges(loopRanges, lbs, ubs, steps);
// Affine loops require constant steps.
SmallVector<int64_t, 4> constantSteps;
constantSteps.reserve(steps.size());
for (Value v : steps) {
auto op = v.getDefiningOp<ConstantIndexOp>();
assert(op && "Affine loops require constant steps");
constantSteps.push_back(op.getValue());
}
auto bodyBuilderWithoutIterArgsFn = [&](ValueRange ivs) {
bodyBuilderFn(ivs, {});
};
edsc::affineLoopNestBuilder(lbs, ubs, constantSteps,
bodyBuilderWithoutIterArgsFn);
}
/// Update the `lb`, `ub` and `step` to get per processor `lb`, `ub` and `step`.
static void updateBoundsForCyclicDistribution(OpBuilder &builder, Location loc,
Value procId, Value nprocs,
Value &lb, Value &ub,
Value &step) {
using edsc::op::operator+;
using edsc::op::operator*;
lb = lb + (procId * step);
step = nprocs * step;
}
/// Generates a loop nest consisting of scf.parallel and scf.for, depending
/// on the `iteratorTypes.` Consecutive parallel loops create a single
/// scf.parallel operation; each sequential loop creates a new scf.for
/// operation. The body of the innermost loop is populated by
/// `bodyBuilderFn` that accepts a range of induction variables for all
/// loops. `ivStorage` is used to store the partial list of induction
/// variables.
// TODO: this function can be made iterative instead. However, it
// will have at most as many recursive calls as nested loops, which rarely
// exceeds 10.
static void
generateParallelLoopNest(ValueRange lbs, ValueRange ubs, ValueRange steps,
ArrayRef<Attribute> iteratorTypes,
function_ref<void(ValueRange)> bodyBuilderFn,
SmallVectorImpl<Value> &ivStorage,
ArrayRef<DistributionMethod> distributionMethod = {}) {
assert(lbs.size() == ubs.size());
assert(lbs.size() == steps.size());
assert(lbs.size() == iteratorTypes.size());
// If there are no (more) loops to be generated, generate the body and be
// done with it.
if (iteratorTypes.empty())
return bodyBuilderFn(ivStorage);
// Find the outermost parallel loops and drop their types from the list.
unsigned nLoops = iteratorTypes.size();
unsigned nOuterPar =
nLoops - iteratorTypes.drop_while(isParallelIteratorType).size();
// If there are no outer parallel loops, generate one sequential loop and
// recurse. Note that we wouldn't have dropped anything from `iteratorTypes`
// in this case.
if (nOuterPar == 0) {
edsc::loopNestBuilder(lbs[0], ubs[0], steps[0], [&](Value iv) {
ivStorage.push_back(iv);
generateParallelLoopNest(lbs.drop_front(), ubs.drop_front(),
steps.drop_front(), iteratorTypes.drop_front(),
bodyBuilderFn, ivStorage, distributionMethod);
});
return;
}
if (distributionMethod.empty()) {
// Generate a single parallel loop-nest operation for all outermost
// parallel loops and recurse.
edsc::OperationBuilder<scf::ParallelOp>(
lbs.take_front(nOuterPar), ubs.take_front(nOuterPar),
steps.take_front(nOuterPar),
[&](OpBuilder &nestedBuilder, Location nestedLoc, ValueRange localIvs) {
edsc::ScopedContext context(nestedBuilder, nestedLoc);
ivStorage.append(localIvs.begin(), localIvs.end());
generateParallelLoopNest(
lbs.drop_front(nOuterPar), ubs.drop_front(nOuterPar),
steps.drop_front(nOuterPar), iteratorTypes.drop_front(nOuterPar),
bodyBuilderFn, ivStorage,
(distributionMethod.size() < nOuterPar)
? ArrayRef<DistributionMethod>()
: distributionMethod.drop_front(nOuterPar));
});
return;
}
// Process all consecutive similarly distributed loops simultaneously.
DistributionMethod methodToUse = distributionMethod[0];
unsigned numProcessed = 1;
for (unsigned i = 1; i < nOuterPar && i < distributionMethod.size(); ++i) {
if (distributionMethod[i] != methodToUse)
break;
numProcessed++;
}
switch (methodToUse) {
case DistributionMethod::Cyclic: {
// Generate a single parallel loop-nest operation for all outermost
// parallel loops and recurse.
edsc::OperationBuilder<scf::ParallelOp>(
lbs.take_front(numProcessed), ubs.take_front(numProcessed),
steps.take_front(numProcessed),
[&](OpBuilder &nestedBuilder, Location nestedLoc, ValueRange localIvs) {
edsc::ScopedContext context(nestedBuilder, nestedLoc);
ivStorage.append(localIvs.begin(), localIvs.end());
generateParallelLoopNest(
lbs.drop_front(numProcessed), ubs.drop_front(numProcessed),
steps.drop_front(numProcessed),
iteratorTypes.drop_front(numProcessed), bodyBuilderFn, ivStorage,
(distributionMethod.size() < numProcessed)
? ArrayRef<DistributionMethod>()
: distributionMethod.drop_front(numProcessed));
});
return;
}
case DistributionMethod::CyclicNumProcsGeNumIters: {
// Check (for the processed loops) that the iteration is in-bounds.
using edsc::op::slt;
using edsc::op::operator&&;
Value cond = slt(lbs[0], ubs[0]);
for (unsigned i = 1; i < numProcessed; ++i)
cond = cond && slt(lbs[i], ubs[i]);
ivStorage.append(lbs.begin(), std::next(lbs.begin(), numProcessed));
edsc::conditionBuilder(cond, [&]() {
generateParallelLoopNest(
lbs.drop_front(numProcessed), ubs.drop_front(numProcessed),
steps.drop_front(numProcessed),
iteratorTypes.drop_front(numProcessed), bodyBuilderFn, ivStorage,
distributionMethod.drop_front(numProcessed));
});
return;
}
case DistributionMethod::CyclicNumProcsEqNumIters:
// No check/loops needed here. Set the `%iv` to be the `%lb` and proceed
// with inner loop generation.
ivStorage.append(lbs.begin(), std::next(lbs.begin(), numProcessed));
generateParallelLoopNest(
lbs.drop_front(numProcessed), ubs.drop_front(numProcessed),
steps.drop_front(numProcessed), iteratorTypes.drop_front(numProcessed),
bodyBuilderFn, ivStorage, distributionMethod.drop_front(numProcessed));
return;
}
}
/// Specialization for generating a mix of parallel and sequential scf loops.
template <>
void GenerateLoopNest<scf::ParallelOp>::doit(
ArrayRef<Range> loopRanges, ValueRange iterArgInitValues,
ArrayRef<Attribute> iteratorTypes,
function_ref<scf::ValueVector(ValueRange, ValueRange)> bodyBuilderFn,
Optional<LinalgLoopDistributionOptions> distributionOptions) {
assert(iterArgInitValues.empty() && "unexpected ParallelOp init values");
// This function may be passed more iterator types than ranges.
assert(iteratorTypes.size() >= loopRanges.size() &&
"expected iterator type for all ranges");
iteratorTypes = iteratorTypes.take_front(loopRanges.size());
SmallVector<Value, 8> lbsStorage, ubsStorage, stepsStorage, ivs;
unsigned numLoops = iteratorTypes.size();
ivs.reserve(numLoops);
lbsStorage.reserve(numLoops);
ubsStorage.reserve(numLoops);
stepsStorage.reserve(numLoops);
// Get the loop lb, ub, and step.
unpackRanges(loopRanges, lbsStorage, ubsStorage, stepsStorage);
// Modify the lb, ub, and step based on the distribution options.
SmallVector<DistributionMethod, 0> distributionMethod;
if (distributionOptions) {
auto &options = distributionOptions.getValue();
OpBuilder &builder = edsc::ScopedContext::getBuilderRef();
Location loc = edsc::ScopedContext::getLocation();
distributionMethod.assign(distributionOptions->distributionMethod.begin(),
distributionOptions->distributionMethod.end());
SmallVector<Range, 2> parallelLoopRanges;
for (auto iteratorType : enumerate(iteratorTypes)) {
if (isParallelIteratorType(iteratorType.value()))
parallelLoopRanges.push_back(loopRanges[iteratorType.index()]);
}
if (distributionMethod.size() < parallelLoopRanges.size())
parallelLoopRanges.resize(distributionMethod.size());
SmallVector<ProcInfo, 2> procInfo =
options.procInfo(builder, loc, parallelLoopRanges);
unsigned index = 0;
for (auto iteratorType : enumerate(iteratorTypes)) {
if (index >= procInfo.size())
break;
if (isParallelIteratorType(iteratorType.value())) {
unsigned i = iteratorType.index();
updateBoundsForCyclicDistribution(builder, loc, procInfo[index].procId,
procInfo[index].nprocs, lbsStorage[i],
ubsStorage[i], stepsStorage[i]);
index++;
}
}
}
ValueRange lbs(lbsStorage), ubs(ubsStorage), steps(stepsStorage);
auto bodyBuilderWithoutIterArgsFn = [&](ValueRange ivs) {
bodyBuilderFn(ivs, {});
};
generateParallelLoopNest(lbs, ubs, steps, iteratorTypes,
bodyBuilderWithoutIterArgsFn, ivs,
distributionMethod);
assert(ivs.size() == iteratorTypes.size() && "did not generate enough loops");
}
} // namespace linalg
} // namespace mlir