File: int8_swish_op_nnpi.cc

package info (click to toggle)
pytorch 1.13.1%2Bdfsg-4
  • links: PTS, VCS
  • area: main
  • in suites: bookworm
  • size: 139,252 kB
  • sloc: cpp: 1,100,274; python: 706,454; ansic: 83,052; asm: 7,618; java: 3,273; sh: 2,841; javascript: 612; makefile: 323; xml: 269; ruby: 185; yacc: 144; objc: 68; lex: 44
file content (27 lines) | stat: -rw-r--r-- 880 bytes parent folder | download | duplicates (2)
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
#include "caffe2/contrib/fakelowp/int8_swish_op_nnpi.h"

namespace caffe2 {

REGISTER_CPU_OPERATOR(SwishFakeInt8NNPI, int8::SwishInt8NNPIOp);

OPERATOR_SCHEMA(SwishFakeInt8NNPI)
    .IdenticalTypeAndShape()
    .Arg("X_scale", "Inout tensor quantization scale")
    .Arg("X_zero_point", "Input tensor quantization offset")
    .Arg("Y_scale", "Output tensor quantization scale")
    .Arg("Y_zero_point", "Output tensor quantization offset")
    .NumInputs(1)
    .NumOutputs(1)
    .SetDoc(R"DOC(
Apply the Swish function element-wise after dequantizing input tensor.
$$Swish(x) = \frac{x}{1+\exp(-x)}$$
Quantize the Swish function output back to Int8.
The input and output of this operator are converted to fp16 precision
before applying the function.
<details>
</details>
)DOC")
    .Input(0, "X", "Int8 Tensor X.")
    .Output(0, "Y", "Int8 Tensor Y.");

} // namespace caffe2