File: fused_nbit_rowwise_conversion.h

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 (39 lines) | stat: -rw-r--r-- 789 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
#pragma once

#include <cstddef>
#include <cstdint>

namespace caffe2 {

void FloatToFused8BitRowwiseQuantized(
    const float* input,
    size_t input_rows,
    int input_columns,
    std::uint8_t* output);

void Fused8BitRowwiseQuantizedToFloat(
    const std::uint8_t* input,
    size_t input_rows,
    int input_columns,
    float* output);

/**
 * Row-wise quantization with fp16 scale and bias
 *
 * @param bit_rate can be 2, 4, or 8
 */
void FloatToFusedNBitRowwiseQuantizedSBHalf(
    int bit_rate,
    const float* input,
    size_t input_rows,
    int input_columns,
    std::uint8_t* output);

void FusedNBitRowwiseQuantizedSBHalfToFloat(
    int bit_rate,
    const std::uint8_t* input,
    size_t input_rows,
    int input_columns,
    float* output);

} // namespace caffe2