ConvertSimQuant.cpp
5.19 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
//===- ConvertSimQuant.cpp - Converts simulated quant ops------------------===//
//
// Part of the MLIR 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/QuantOps/FakeQuantSupport.h"
#include "mlir/Dialect/QuantOps/Passes.h"
#include "mlir/Dialect/QuantOps/QuantOps.h"
#include "mlir/Dialect/QuantOps/UniformSupport.h"
#include "mlir/IR/Attributes.h"
#include "mlir/IR/PatternMatch.h"
#include "mlir/IR/StandardTypes.h"
#include "mlir/Pass/Pass.h"
using namespace mlir;
using namespace mlir::quant;
namespace {
class ConvertSimulatedQuantPass
: public FunctionPass<ConvertSimulatedQuantPass> {
public:
void runOnFunction() override;
};
/// Base class rewrites ConstFakeQuant into a qbarrier/dbarrier pair.
template <typename ConcreteRewriteClass, typename FakeQuantOp>
class FakeQuantRewrite : public OpRewritePattern<FakeQuantOp> {
public:
using OpRewritePattern<FakeQuantOp>::OpRewritePattern;
FakeQuantRewrite(MLIRContext *ctx, bool *hadFailure)
: OpRewritePattern<FakeQuantOp>(ctx), hadFailure(hadFailure) {}
PatternMatchResult matchAndRewrite(FakeQuantOp op,
PatternRewriter &rewriter) const override {
// TODO: If this pattern comes up more frequently, consider adding core
// support for failable rewrites.
if (failableRewrite(op, rewriter)) {
*hadFailure = true;
return Pattern::matchFailure();
}
return Pattern::matchSuccess();
}
private:
bool *hadFailure;
bool failableRewrite(FakeQuantOp op, PatternRewriter &rewriter) const {
auto converter = ExpressedToQuantizedConverter::forInputType(op.getType());
if (!converter) {
return (op.emitError("unsupported quantized type conversion"), true);
}
QuantizedType elementType =
static_cast<const ConcreteRewriteClass *>(this)
->convertFakeQuantAttrsToType(op, converter.expressedType);
if (!elementType) {
// Note that the fakeQuantAttrsToType will have emitted the error.
return true;
}
Type quantizedType = converter.convert(elementType);
assert(quantizedType &&
"Converter accepted a type that it did not convert");
// TODO: Map to a qbarrier with an attribute like [Forced] to signal that
// this is a forced/hard-coded constraint.
auto qbarrier = rewriter.create<QuantizeCastOp>(op.getLoc(), quantizedType,
op.inputs());
rewriter.replaceOpWithNewOp<DequantizeCastOp>(op, converter.inputType,
qbarrier.getResult());
return false;
}
};
class ConstFakeQuantRewrite
: public FakeQuantRewrite<ConstFakeQuantRewrite, ConstFakeQuant> {
public:
using BaseRewrite = FakeQuantRewrite<ConstFakeQuantRewrite, ConstFakeQuant>;
ConstFakeQuantRewrite(MLIRContext *ctx, bool *hadFailure)
: BaseRewrite(ctx, hadFailure) {}
QuantizedType convertFakeQuantAttrsToType(ConstFakeQuant fqOp,
Type expressedType) const {
return fakeQuantAttrsToType(
fqOp.getLoc(), fqOp.num_bits().getSExtValue(),
fqOp.min().convertToFloat(), fqOp.max().convertToFloat(),
fqOp.narrow_range(), expressedType, fqOp.is_signed());
}
};
class ConstFakeQuantPerAxisRewrite
: public FakeQuantRewrite<ConstFakeQuantPerAxisRewrite,
ConstFakeQuantPerAxis> {
public:
using BaseRewrite =
FakeQuantRewrite<ConstFakeQuantPerAxisRewrite, ConstFakeQuantPerAxis>;
ConstFakeQuantPerAxisRewrite(MLIRContext *ctx, bool *hadFailure)
: BaseRewrite(ctx, hadFailure) {}
QuantizedType convertFakeQuantAttrsToType(ConstFakeQuantPerAxis fqOp,
Type expressedType) const {
SmallVector<double, 4> min, max;
min.reserve(fqOp.min().size());
max.reserve(fqOp.max().size());
for (auto m : fqOp.min())
min.push_back(m.cast<FloatAttr>().getValueAsDouble());
for (auto m : fqOp.max())
max.push_back(m.cast<FloatAttr>().getValueAsDouble());
return fakeQuantAttrsToType(fqOp.getLoc(), fqOp.num_bits().getSExtValue(),
fqOp.axis().getSExtValue(), min, max,
fqOp.narrow_range(), expressedType,
fqOp.is_signed());
}
};
} // namespace
void ConvertSimulatedQuantPass::runOnFunction() {
bool hadFailure = false;
OwningRewritePatternList patterns;
auto func = getFunction();
auto ctx = func.getContext();
patterns.insert<ConstFakeQuantRewrite, ConstFakeQuantPerAxisRewrite>(
ctx, &hadFailure);
applyPatternsGreedily(func, patterns);
if (hadFailure)
signalPassFailure();
}
std::unique_ptr<OpPassBase<FuncOp>>
mlir::quant::createConvertSimulatedQuantPass() {
return std::make_unique<ConvertSimulatedQuantPass>();
}
static PassRegistration<ConvertSimulatedQuantPass>
pass("quant-convert-simulated-quantization",
"Converts training-time simulated quantization ops to corresponding "
"quantize/dequantize casts.");