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
|
//===- ParallelLoopTiling.cpp - Tiles scf.parallel ------------------------===//
//
// 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 loop tiling on parallel loops.
//
//===----------------------------------------------------------------------===//
#include "mlir/Dialect/SCF/Transforms/Passes.h"
#include "mlir/Dialect/Affine/IR/AffineOps.h"
#include "mlir/Dialect/Arithmetic/IR/Arithmetic.h"
#include "mlir/Dialect/SCF/IR/SCF.h"
#include "mlir/Dialect/SCF/Transforms/Transforms.h"
#include "mlir/Dialect/SCF/Utils/Utils.h"
namespace mlir {
#define GEN_PASS_DEF_SCFPARALLELLOOPTILING
#include "mlir/Dialect/SCF/Transforms/Passes.h.inc"
} // namespace mlir
using namespace mlir;
using namespace mlir::scf;
/// Tile a parallel loop of the form
/// scf.parallel (%i0, %i1) = (%arg0, %arg1) to (%arg2, %arg3)
/// step (%arg4, %arg5)
///
/// into
/// scf.parallel (%i0, %i1) = (%arg0, %arg1) to (%arg2, %arg3)
/// step (%arg4*tileSize[0],
/// %arg5*tileSize[1])
/// scf.parallel (%j0, %j1) = (0, 0) to (min(%arg4*tileSize[0], %arg2-%i0)
/// min(%arg5*tileSize[1], %arg3-%i1))
/// step (%arg4, %arg5)
///
/// or, when no-min-max-bounds is true, into
/// scf.parallel (%i0, %i1) = (%arg0, %arg1) to (%arg2, %arg3)
/// step (%arg4*tileSize[0],
/// %arg5*tileSize[1])
/// scf.parallel (%j0, %j1) = (0, 0) to (%arg4*tileSize[0],
/// %arg5*tileSize[1])
/// step (%arg4, %arg5)
/// %inbound = (%j0 * %arg4 + %i0 < %arg2) &&
/// (%j1 * %arg5 + %i1 < %arg3)
/// scf.if (%inbound)
/// ....
///
/// where the uses of %i0 and %i1 in the loop body are replaced by
/// %i0 + j0 and %i1 + %j1.
///
/// The old loop is replaced with the new one.
std::pair<ParallelOp, ParallelOp>
mlir::scf::tileParallelLoop(ParallelOp op, ArrayRef<int64_t> tileSizes,
bool noMinMaxBounds) {
OpBuilder b(op);
auto zero = b.create<arith::ConstantIndexOp>(op.getLoc(), 0);
SmallVector<Value, 2> tileSizeConstants;
tileSizeConstants.reserve(op.getUpperBound().size());
for (size_t i = 0, end = op.getUpperBound().size(); i != end; ++i) {
if (i < tileSizes.size())
tileSizeConstants.push_back(
b.create<arith::ConstantIndexOp>(op.getLoc(), tileSizes[i]));
else
// Just pick 1 for the remaining dimensions.
tileSizeConstants.push_back(
b.create<arith::ConstantIndexOp>(op.getLoc(), 1));
}
// Create the outer loop with adjusted steps.
SmallVector<Value, 2> newSteps;
newSteps.reserve(op.getStep().size());
for (auto step : llvm::zip(op.getStep(), tileSizeConstants)) {
newSteps.push_back(b.create<arith::MulIOp>(op.getLoc(), std::get<0>(step),
std::get<1>(step)));
}
auto outerLoop = b.create<ParallelOp>(op.getLoc(), op.getLowerBound(),
op.getUpperBound(), newSteps);
b.setInsertionPointToStart(outerLoop.getBody());
// Compute min(size, dim - offset) to avoid out-of-bounds accesses.
auto minMap = AffineMap::get(
/*dimCount=*/3, /*symbolCount=*/0,
{getAffineDimExpr(/*position=*/0, b.getContext()),
getAffineDimExpr(/*position=*/1, b.getContext()) -
getAffineDimExpr(/*position=*/2, b.getContext())},
b.getContext());
// Create the inner loop with adjusted bounds.
SmallVector<Value, 2> newBounds;
newBounds.reserve(op.getUpperBound().size());
bool needInboundCheck = false;
for (auto [lowerBound, upperBound, newStep, iv, step, tileSizeConstant] :
llvm::zip(outerLoop.getLowerBound(), outerLoop.getUpperBound(),
outerLoop.getStep(), outerLoop.getInductionVars(),
op.getStep(), tileSizeConstants)) {
// Collect the statically known loop bounds
auto lowerBoundConstant =
dyn_cast_or_null<arith::ConstantIndexOp>(lowerBound.getDefiningOp());
auto upperBoundConstant =
dyn_cast_or_null<arith::ConstantIndexOp>(upperBound.getDefiningOp());
auto stepConstant =
dyn_cast_or_null<arith::ConstantIndexOp>(step.getDefiningOp());
auto tileSize =
cast<arith::ConstantIndexOp>(tileSizeConstant.getDefiningOp()).value();
// If the loop bounds and the loop step are constant and if the number of
// loop iterations is an integer multiple of the tile size, we use a static
// bound for the inner loop.
if (lowerBoundConstant && upperBoundConstant && stepConstant) {
auto numIterations = llvm::divideCeil(upperBoundConstant.value() -
lowerBoundConstant.value(),
stepConstant.value());
if (numIterations % tileSize == 0) {
newBounds.push_back(newStep);
continue;
}
}
// For InboundCheck mode, just use the variable outer step
if (noMinMaxBounds) {
newBounds.push_back(newStep);
needInboundCheck = true;
continue;
}
// Otherwise, we dynamically compute the bound for
// each iteration of the outer loop.
newBounds.push_back(
b.create<AffineMinOp>(op.getLoc(), b.getIndexType(), minMap,
ValueRange{newStep, upperBound, iv}));
}
auto innerLoop = b.create<ParallelOp>(
op.getLoc(), SmallVector<Value, 2>(newBounds.size(), zero), newBounds,
op.getStep());
if (noMinMaxBounds && needInboundCheck) {
b.setInsertionPointToStart(innerLoop.getBody());
// Insert in-bound check
Value inbound =
b.create<arith::ConstantIntOp>(op.getLoc(), 1, b.getIntegerType(1));
for (auto [outerUpperBound, outerIV, innerIV, innerStep] :
llvm::zip(outerLoop.getUpperBound(), outerLoop.getInductionVars(),
innerLoop.getInductionVars(), innerLoop.getStep())) {
// %in_bound = %in_bound &&
// (%inner_iv * %inner_step + %outer_iv < %outer_upper_bound)
Value index = b.create<arith::AddIOp>(
op.getLoc(), b.create<arith::MulIOp>(op.getLoc(), innerIV, innerStep),
outerIV);
Value dimInbound = b.create<arith::CmpIOp>(
op.getLoc(), arith::CmpIPredicate::ult, index, outerUpperBound);
inbound = b.create<arith::AndIOp>(op.getLoc(), inbound, dimInbound);
}
auto ifInbound = b.create<IfOp>(op.getLoc(),
/*resultTypes*/ ArrayRef<Type>{}, inbound,
/*hasElseRegion*/ false);
ifInbound.getThenRegion().takeBody(op.getRegion());
Block &thenBlock = ifInbound.getThenRegion().front();
b.setInsertionPointToStart(innerLoop.getBody());
for (const auto &ivs : llvm::enumerate(llvm::zip(
innerLoop.getInductionVars(), outerLoop.getInductionVars()))) {
auto newIndex = b.create<arith::AddIOp>(
op.getLoc(), std::get<0>(ivs.value()), std::get<1>(ivs.value()));
thenBlock.getArgument(ivs.index())
.replaceAllUsesExcept(newIndex, newIndex);
}
thenBlock.eraseArguments(0, thenBlock.getNumArguments());
} else {
innerLoop.getRegion().takeBody(op.getRegion());
b.setInsertionPointToStart(innerLoop.getBody());
for (auto ivs : llvm::zip(innerLoop.getInductionVars(),
outerLoop.getInductionVars())) {
Value innerIndex = std::get<0>(ivs);
auto newIndex = b.create<arith::AddIOp>(op.getLoc(), std::get<0>(ivs),
std::get<1>(ivs));
innerIndex.replaceAllUsesExcept(newIndex, newIndex);
}
}
op.erase();
return std::make_pair(outerLoop, innerLoop);
}
namespace {
struct ParallelLoopTiling
: public impl::SCFParallelLoopTilingBase<ParallelLoopTiling> {
ParallelLoopTiling() = default;
explicit ParallelLoopTiling(ArrayRef<int64_t> tileSizes,
bool noMinMaxBounds = false) {
this->tileSizes = tileSizes;
this->noMinMaxBounds = noMinMaxBounds;
}
void runOnOperation() override {
auto *parentOp = getOperation();
SmallVector<ParallelOp, 2> innermostPloops;
getInnermostParallelLoops(parentOp, innermostPloops);
for (ParallelOp ploop : innermostPloops) {
// FIXME: Add reduction support.
if (ploop.getNumReductions() == 0)
tileParallelLoop(ploop, tileSizes, noMinMaxBounds);
}
}
};
} // namespace
std::unique_ptr<Pass>
mlir::createParallelLoopTilingPass(ArrayRef<int64_t> tileSizes,
bool noMinMaxBounds) {
return std::make_unique<ParallelLoopTiling>(tileSizes, noMinMaxBounds);
}
|
