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
|
#pragma once
#include "graph/node.h"
#include "tensors/cpu/sharp/int_gemm.h"
namespace marian {
namespace cpu {
namespace int16 {
struct QuantizeNodeOp : public UnaryNodeOp {
float clipValue_;
QuantizeNodeOp(Expr a, float clipValue)
: UnaryNodeOp(a, Type::int16), clipValue_{clipValue} {}
NodeOps forwardOps() override {
return {NodeOp(Quantize16(val_, child(0)->val(), clipValue_))};
}
NodeOps backwardOps() override {
ABORT("Only used for inference");
}
const std::string type() override { return "quantizeInt16"; }
};
class DotNodeOp : public NaryNodeOp {
private:
float scalar_;
public:
DotNodeOp(Expr a, Expr b, float scalar)
: NaryNodeOp({a, b}, newShape(a, b), Type::float32), scalar_(scalar) {}
Shape newShape(Expr a, Expr b) {
auto shapeA = a->shape();
auto shapeB = b->shape();
// Computing A * B^T
shapeB.set(-2, b->shape()[-1]);
shapeB.set(-1, b->shape()[-2]);
Shape outShape = shapeA;
outShape.set(-1, shapeB[-1]);
ABORT_IF(shapeA[-1] != shapeB[-2],
"matrix product requires dimensions to match");
return outShape;
}
NodeOps forwardOps() override {
return {NodeOp(ProdInt16(val_, child(0)->val(), child(1)->val(), scalar_))};
}
NodeOps backwardOps() override {
ABORT("Only used for inference");
}
const std::string type() override { return "dotInt16"; }
};
class AffineNodeOp : public NaryNodeOp {
private:
float scalar_;
public:
AffineNodeOp(const std::vector<Expr>& nodes, float scalar)
: NaryNodeOp(nodes, newShape(nodes[0], nodes[1]), Type::float32), scalar_(scalar) {}
Shape newShape(Expr a, Expr b) {
auto shapeA = a->shape();
auto shapeB = b->shape();
// Computing A * B^T
shapeB.set(-2, b->shape()[-1]);
shapeB.set(-1, b->shape()[-2]);
Shape outShape = shapeA;
outShape.set(-1, shapeB[-1]);
ABORT_IF(shapeA[-1] != shapeB[-2],
"matrix product requires dimensions to match");
return outShape;
}
NodeOps forwardOps() override {
return {
NodeOp(ProdInt16(val_, child(0)->val(), child(1)->val(), scalar_);
AddBias(val_, child(2)->val()))
};
}
NodeOps backwardOps() override {
ABORT("Only used for inference");
}
const std::string type() override { return "affineInt16"; }
};
static inline Expr dot(Expr a, Expr b, float scalar) {
return Expression<cpu::int16::DotNodeOp>(a, b, scalar);
}
static inline Expr affine(Expr a, Expr b, Expr c, float scalar) {
std::vector<Expr> nodes = {a, b, c};
return Expression<cpu::int16::AffineNodeOp>(nodes, scalar);
}
static inline Expr quantize(Expr a, float clipValue) {
return Expression<cpu::int16::QuantizeNodeOp>(a, clipValue);
}
} // namespace int16
} // namespace cpu
} // namespace marian
|