File: int8_dequantize_op.h

package info (click to toggle)
pytorch 1.7.1-7
  • links: PTS, VCS
  • area: main
  • in suites: bullseye
  • size: 80,340 kB
  • sloc: cpp: 670,830; python: 343,991; ansic: 67,845; asm: 5,503; sh: 2,924; java: 2,888; xml: 266; makefile: 244; ruby: 148; yacc: 144; objc: 51; lex: 44
file content (52 lines) | stat: -rw-r--r-- 1,148 bytes parent folder | download
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
#ifndef CAFFE2_OPERATORS_INT8_DEQUANTIZE_OP_H_
#define CAFFE2_OPERATORS_INT8_DEQUANTIZE_OP_H_

#include "caffe2/core/context.h"
#include "caffe2/core/operator.h"
#include "caffe2/core/tensor_int8.h"
#include "caffe2/operators/quantized/int8_utils.h"

namespace caffe2 {

namespace int8 {

namespace {

void Int8Dequantize(
    const uint8_t* in,
    float* out,
    const int64_t N,
    const float X_scale,
    const int32_t X_offset) {
  for (auto i = 0; i < N; ++i) {
    out[i] = (static_cast<int32_t>(in[i]) - X_offset) * X_scale;
  }
}

} // namespace

class Int8DequantizeOp final : public Operator<CPUContext> {
 public:
  using Operator<CPUContext>::Operator;

  bool RunOnDevice() override {
    const auto& X = Inputs()[0]->template Get<Int8TensorCPU>();

    auto* Y = Output(0, X.t.sizes(), at::dtype<float>());
    int32_t X_offset = X.zero_point;
    auto X_scale = X.scale;
    Int8Dequantize(
        X.t.data<uint8_t>(),
        Y->mutable_data<float>(),
        X.t.numel(),
        X_scale,
        X_offset);
    return true;
  }
};

} // namespace int8

} // namespace caffe2

#endif // CAFFE2_OPERATORS_INT8_DEQUANTIZE_OP_H_