/
StraightLineVectorizer.cpp
164 lines (141 loc) · 5.36 KB
/
StraightLineVectorizer.cpp
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
#include "include/Transforms/StraightLineVectorizer/StraightLineVectorizer.h"
#include "include/Graph/Graph.h"
#include "llvm/include/llvm/ADT/TypeSwitch.h" // from @llvm-project
#include "llvm/include/llvm/Support/Debug.h" // from @llvm-project
#include "mlir/include/mlir/Analysis/SliceAnalysis.h" // from @llvm-project
#include "mlir/include/mlir/Dialect/Arith/IR/Arith.h" // from @llvm-project
#include "mlir/include/mlir/Dialect/Tensor/IR/Tensor.h" // from @llvm-project
#include "mlir/include/mlir/IR/Visitors.h" // from @llvm-project
#include "mlir/include/mlir/Support/LLVM.h" // from @llvm-project
#include "mlir/include/mlir/Transforms/GreedyPatternRewriteDriver.h" // from @llvm-project
#include "mlir/include/mlir/Transforms/Passes.h" // from @llvm-project
#include "mlir/include/mlir/Transforms/TopologicalSortUtils.h" // from @llvm-project
#define DEBUG_TYPE "straight-line-vectorizer"
namespace mlir {
namespace heir {
#define GEN_PASS_DEF_STRAIGHTLINEVECTORIZER
#include "include/Transforms/StraightLineVectorizer/StraightLineVectorizer.h.inc"
/// Returns true if the two operations can be combined into a single vectorized
/// operation.
bool areCompatible(Operation *lhs, Operation *rhs) {
if (lhs->getName() != rhs->getName() ||
lhs->getDialect() != rhs->getDialect() ||
lhs->getResultTypes() != rhs->getResultTypes()) {
return false;
}
return OpTrait::hasElementwiseMappableTraits(lhs);
}
bool tryVectorizeBlock(Block *block) {
graph::Graph<Operation *> graph;
for (auto &op : block->getOperations()) {
if (!op.hasTrait<OpTrait::Elementwise>()) {
continue;
}
graph.addVertex(&op);
SetVector<Operation *> backwardSlice;
BackwardSliceOptions options;
options.omitBlockArguments = true;
getBackwardSlice(&op, &backwardSlice, options);
for (auto *upstreamDep : backwardSlice) {
// An edge from upstreamDep to `op` means that upstreamDep must be
// computed before `op`.
graph.addEdge(upstreamDep, &op);
}
}
if (graph.empty()) {
return false;
}
auto result = graph.sortGraphByLevels();
assert(succeeded(result) &&
"Only possible failure is a cycle in the SSA graph!");
auto levels = result.value();
LLVM_DEBUG({
llvm::dbgs()
<< "Found operations to vectorize. In topo-sorted level order:\n";
int level_num = 0;
for (const auto &level : levels) {
llvm::dbgs() << "\nLevel " << level_num++ << ":\n";
for (auto op : level) {
llvm::dbgs() << " - " << *op << "\n";
}
}
});
bool madeReplacement = false;
for (const auto &level : levels) {
DenseMap<Operation *, SmallVector<Operation *, 4>> compatibleOps;
for (auto *op : level) {
bool foundCompatible = false;
for (auto &[key, bucket] : compatibleOps) {
if (areCompatible(key, op)) {
compatibleOps[key].push_back(op);
foundCompatible = true;
}
}
if (!foundCompatible) {
compatibleOps[op].push_back(op);
}
}
LLVM_DEBUG(llvm::dbgs()
<< "Partitioned level of size " << level.size() << " into "
<< compatibleOps.size() << " groups of compatible ops\n");
for (auto &[key, bucket] : compatibleOps) {
if (bucket.size() < 2) {
continue;
}
LLVM_DEBUG({
llvm::dbgs() << "Vectorizing ops:\n";
for (auto op : bucket) {
llvm::dbgs() << " - " << *op << "\n";
}
});
OpBuilder builder(bucket.back());
// relies on CGGI ops having a single result type
Type elementType = key->getResultTypes()[0];
RankedTensorType tensorType = RankedTensorType::get(
{static_cast<int64_t>(bucket.size())}, elementType);
SmallVector<Value, 4> vectorizedOperands;
for (int operandIndex = 0; operandIndex < key->getNumOperands();
++operandIndex) {
SmallVector<Value, 4> operands;
operands.reserve(bucket.size());
for (auto *op : bucket) {
operands.push_back(op->getOperand(operandIndex));
}
auto fromElementsOp = builder.create<tensor::FromElementsOp>(
key->getLoc(), tensorType, operands);
vectorizedOperands.push_back(fromElementsOp.getResult());
}
Operation *vectorizedOp = builder.clone(*key);
vectorizedOp->setOperands(vectorizedOperands);
vectorizedOp->getResult(0).setType(tensorType);
int bucketIndex = 0;
for (auto *op : bucket) {
auto extractionIndex = builder.create<arith::ConstantOp>(
op->getLoc(), builder.getIndexAttr(bucketIndex));
auto extractOp = builder.create<tensor::ExtractOp>(
op->getLoc(), elementType, vectorizedOp->getResult(0),
extractionIndex.getResult());
op->replaceAllUsesWith(ValueRange{extractOp.getResult()});
bucketIndex++;
}
for (auto *op : bucket) {
op->erase();
}
madeReplacement = true;
}
}
return madeReplacement;
}
struct StraightLineVectorizer
: impl::StraightLineVectorizerBase<StraightLineVectorizer> {
using StraightLineVectorizerBase::StraightLineVectorizerBase;
void runOnOperation() override {
getOperation()->walk<WalkOrder::PreOrder>([&](Block *block) {
if (tryVectorizeBlock(block)) {
sortTopologically(block);
}
});
}
};
} // namespace heir
} // namespace mlir