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
|
//===----------- MultiBuffering.cpp ---------------------------------------===//
//
// 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 multi buffering transformation.
//
//===----------------------------------------------------------------------===//
#include "mlir/Dialect/Affine/IR/AffineOps.h"
#include "mlir/Dialect/MemRef/IR/MemRef.h"
#include "mlir/Dialect/MemRef/Transforms/Passes.h"
#include "mlir/IR/Dominance.h"
#include "mlir/Interfaces/LoopLikeInterface.h"
using namespace mlir;
/// Return true if the op fully overwrite the given `buffer` value.
static bool overrideBuffer(Operation *op, Value buffer) {
auto copyOp = dyn_cast<memref::CopyOp>(op);
if (!copyOp)
return false;
return copyOp.getTarget() == buffer;
}
/// Replace the uses of `oldOp` with the given `val` and for subview uses
/// propagate the type change. Changing the memref type may require propagating
/// it through subview ops so we cannot just do a replaceAllUse but need to
/// propagate the type change and erase old subview ops.
static void replaceUsesAndPropagateType(Operation *oldOp, Value val,
OpBuilder &builder) {
SmallVector<Operation *> opToDelete;
SmallVector<OpOperand *> operandsToReplace;
for (OpOperand &use : oldOp->getUses()) {
auto subviewUse = dyn_cast<memref::SubViewOp>(use.getOwner());
if (!subviewUse) {
// Save the operand to and replace outside the loop to not invalidate the
// iterator.
operandsToReplace.push_back(&use);
continue;
}
builder.setInsertionPoint(subviewUse);
Type newType = memref::SubViewOp::inferRankReducedResultType(
subviewUse.getType().getShape(), val.getType().cast<MemRefType>(),
subviewUse.getStaticOffsets(), subviewUse.getStaticSizes(),
subviewUse.getStaticStrides());
Value newSubview = builder.create<memref::SubViewOp>(
subviewUse->getLoc(), newType.cast<MemRefType>(), val,
subviewUse.getMixedOffsets(), subviewUse.getMixedSizes(),
subviewUse.getMixedStrides());
replaceUsesAndPropagateType(subviewUse, newSubview, builder);
opToDelete.push_back(use.getOwner());
}
for (OpOperand *operand : operandsToReplace)
operand->set(val);
// Clean up old subview ops.
for (Operation *op : opToDelete)
op->erase();
}
/// Helper to convert get a value from an OpFoldResult or create it at the
/// builder insert point.
static Value getOrCreateValue(OpFoldResult res, OpBuilder &builder,
Location loc) {
Value value = res.dyn_cast<Value>();
if (value)
return value;
return builder.create<arith::ConstantIndexOp>(
loc, res.dyn_cast<Attribute>().cast<IntegerAttr>().getInt());
}
// Transformation to do multi-buffering/array expansion to remove dependencies
// on the temporary allocation between consecutive loop iterations.
// Returns success if the transformation happened and failure otherwise.
// This is not a pattern as it requires propagating the new memref type to its
// uses and requires updating subview ops.
FailureOr<memref::AllocOp> mlir::memref::multiBuffer(memref::AllocOp allocOp,
unsigned multiplier) {
DominanceInfo dom(allocOp->getParentOp());
LoopLikeOpInterface candidateLoop;
for (Operation *user : allocOp->getUsers()) {
auto parentLoop = user->getParentOfType<LoopLikeOpInterface>();
if (!parentLoop)
return failure();
/// Make sure there is no loop carried dependency on the allocation.
if (!overrideBuffer(user, allocOp.getResult()))
continue;
// If this user doesn't dominate all the other users keep looking.
if (llvm::any_of(allocOp->getUsers(), [&](Operation *otherUser) {
return !dom.dominates(user, otherUser);
}))
continue;
candidateLoop = parentLoop;
break;
}
if (!candidateLoop)
return failure();
std::optional<Value> inductionVar = candidateLoop.getSingleInductionVar();
std::optional<OpFoldResult> lowerBound = candidateLoop.getSingleLowerBound();
std::optional<OpFoldResult> singleStep = candidateLoop.getSingleStep();
if (!inductionVar || !lowerBound || !singleStep)
return failure();
if (!dom.dominates(allocOp.getOperation(), candidateLoop))
return failure();
OpBuilder builder(candidateLoop);
SmallVector<int64_t, 4> newShape(1, multiplier);
ArrayRef<int64_t> oldShape = allocOp.getType().getShape();
newShape.append(oldShape.begin(), oldShape.end());
auto newMemref = MemRefType::get(newShape, allocOp.getType().getElementType(),
MemRefLayoutAttrInterface(),
allocOp.getType().getMemorySpace());
builder.setInsertionPoint(allocOp);
Location loc = allocOp->getLoc();
auto newAlloc = builder.create<memref::AllocOp>(loc, newMemref, ValueRange{},
allocOp->getAttrs());
builder.setInsertionPoint(&candidateLoop.getLoopBody().front(),
candidateLoop.getLoopBody().front().begin());
SmallVector<Value> operands = {*inductionVar};
AffineExpr induc = getAffineDimExpr(0, allocOp.getContext());
unsigned dimCount = 1;
auto getAffineExpr = [&](OpFoldResult e) -> AffineExpr {
if (std::optional<int64_t> constValue = getConstantIntValue(e)) {
return getAffineConstantExpr(*constValue, allocOp.getContext());
}
auto value = getOrCreateValue(e, builder, candidateLoop->getLoc());
operands.push_back(value);
return getAffineDimExpr(dimCount++, allocOp.getContext());
};
auto init = getAffineExpr(*lowerBound);
auto step = getAffineExpr(*singleStep);
AffineExpr expr = ((induc - init).floorDiv(step)) % multiplier;
auto map = AffineMap::get(dimCount, 0, expr);
Value bufferIndex = builder.create<AffineApplyOp>(loc, map, operands);
SmallVector<OpFoldResult> offsets, sizes, strides;
offsets.push_back(bufferIndex);
offsets.append(oldShape.size(), builder.getIndexAttr(0));
strides.assign(oldShape.size() + 1, builder.getIndexAttr(1));
sizes.push_back(builder.getIndexAttr(1));
for (int64_t size : oldShape)
sizes.push_back(builder.getIndexAttr(size));
auto dstMemref =
memref::SubViewOp::inferRankReducedResultType(
allocOp.getType().getShape(), newMemref, offsets, sizes, strides)
.cast<MemRefType>();
Value subview = builder.create<memref::SubViewOp>(loc, dstMemref, newAlloc,
offsets, sizes, strides);
replaceUsesAndPropagateType(allocOp, subview, builder);
allocOp.erase();
return newAlloc;
}
|
