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
|
#ifndef CAFFE2_OPERATORS_INT8_GIVEN_TENSOR_FILL_OP_H_
#define CAFFE2_OPERATORS_INT8_GIVEN_TENSOR_FILL_OP_H_
#include "caffe2/core/context.h"
#include "caffe2/core/logging.h"
#include "caffe2/core/operator.h"
#include "caffe2/core/tensor_int8.h"
#include "caffe2/operators/filler_op.h"
#include "caffe2/utils/cast.h"
#include "caffe2/utils/math.h"
namespace caffe2 {
namespace int8 {
class Int8GivenTensorFillOp final : public Operator<CPUContext> {
public:
template <class... Args>
explicit Int8GivenTensorFillOp(Args&&... args)
: Operator<CPUContext>(std::forward<Args>(args)...),
scale_(this->template GetSingleArgument<float>("Y_scale", 1.0)),
zero_point_(
this->template GetSingleArgument<int32_t>("Y_zero_point", 0)),
shape_(this->template GetRepeatedArgument<int64_t>("shape")) {
ExtractValues();
}
bool RunOnDevice() override {
auto* output = Outputs()[0]->template GetMutable<Int8TensorCPU>();
ReinitializeTensor(&output->t, shape_, at::dtype<uint8_t>().device(CPU));
output->scale = scale_;
output->zero_point = zero_point_;
return Fill(output);
}
private:
void ExtractValues() {
auto source_values = this->template GetSingleArgument<string>("values", "");
ReinitializeTensor(
&values_,
{static_cast<int64_t>(source_values.size())},
at::dtype<uint8_t>().device(CPU));
uint8_t* values_data = values_.template mutable_data<uint8_t>();
for (const auto i : c10::irange(source_values.size())) {
values_data[i] = static_cast<uint8_t>(source_values[i]);
}
}
bool Fill(Int8TensorCPU* output) {
TORCH_DCHECK_EQ(output->t.numel(), values_.numel())
<< "output size: " << output->t.numel()
<< " given size: " << values_.numel();
auto* data = output->t.template mutable_data<uint8_t>();
const uint8_t* values_data = values_.template data<uint8_t>();
if (output->t.numel()) {
context_.template CopySameDevice<uint8_t>(
output->t.numel(), values_data, data);
}
return true;
}
float scale_;
int32_t zero_point_;
vector<int64_t> shape_;
Tensor values_;
};
class Int8GivenIntTensorFillOp final : public Operator<CPUContext> {
public:
template <class... Args>
explicit Int8GivenIntTensorFillOp(Args&&... args)
: Operator<CPUContext>(std::forward<Args>(args)...),
scale_(this->template GetSingleArgument<float>("Y_scale", 1.0)),
zero_point_(
this->template GetSingleArgument<int32_t>("Y_zero_point", 0)),
shape_(this->template GetRepeatedArgument<int64_t>("shape")) {
ExtractValues();
}
bool RunOnDevice() override {
auto* output = Outputs()[0]->template GetMutable<Int8TensorCPU>();
output->t.Resize(shape_);
output->scale = scale_;
output->zero_point = zero_point_;
return Fill(output);
}
private:
void ExtractValues() {
auto source_values = this->template GetRepeatedArgument<int32_t>("values");
ReinitializeTensor(
&values_,
{static_cast<int64_t>(source_values.size())},
at::dtype<int32_t>().device(CPU));
auto* values_data = values_.template mutable_data<int32_t>();
for (const auto i : c10::irange(source_values.size())) {
values_data[i] = static_cast<int32_t>(source_values[i]);
}
}
bool Fill(Int8TensorCPU* output) {
TORCH_DCHECK_EQ(output->t.numel(), values_.numel())
<< "output size: " << output->t.numel()
<< " given size: " << values_.numel();
auto* data = output->t.template mutable_data<int32_t>();
const auto* values_data = values_.template data<int32_t>();
if (output->t.numel()) {
context_.template CopySameDevice<int32_t>(
output->t.numel(), values_data, data);
}
return true;
}
float scale_;
int32_t zero_point_;
vector<int64_t> shape_;
Tensor values_;
};
} // namespace int8
} // namespace caffe2
#endif // CAFFE2_OPERATORS_INT8_GIVEN_TENSOR_FILL_OP_H_
|
