• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright (c) Facebook, Inc. and its affiliates.
2 // All rights reserved.
3 //
4 // Copyright 2019 Google LLC
5 //
6 // This source code is licensed under the BSD-style license found in the
7 // LICENSE file in the root directory of this source tree.
8 
9 #include <assert.h>
10 #include <math.h>
11 #include <stdint.h>
12 #include <stddef.h>
13 
14 #include <xnnpack/math.h>
15 #include <xnnpack/requantization-stubs.h>
16 
17 
xnn_qs8_requantize_fp32__scalar_fmagic(size_t n,const int32_t * input,float scale,int8_t zero_point,int8_t qmin,int8_t qmax,int8_t * output)18 void xnn_qs8_requantize_fp32__scalar_fmagic(
19     size_t n,
20     const int32_t* input,
21     float scale,
22     int8_t zero_point,
23     int8_t qmin,
24     int8_t qmax,
25     int8_t* output)
26 {
27   assert(n % 4 == 0);
28   assert(scale < 1.0f);
29   assert(scale >= 0x1.0p-32f);
30 
31   const float fmin = (float) ((int32_t) qmin - (int32_t) zero_point);
32   const float fmax = (float) ((int32_t) qmax - (int32_t) zero_point);
33   const float fmagic = 12582912.0f;
34   const int32_t imagic = INT32_C(0x4B400000) - (int32_t) zero_point;
35   for (; n != 0; n -= 4) {
36     const int32_t x = input[0];
37     const int32_t y = input[1];
38     const int32_t z = input[2];
39     const int32_t w = input[3];
40     input += 4;
41 
42     const float x_scaled = (float) x * scale;
43     const float y_scaled = (float) y * scale;
44     const float z_scaled = (float) z * scale;
45     const float w_scaled = (float) w * scale;
46 
47     const float x_clamped = math_min_f32(math_max_f32(x_scaled, fmin), fmax);
48     const float y_clamped = math_min_f32(math_max_f32(y_scaled, fmin), fmax);
49     const float z_clamped = math_min_f32(math_max_f32(z_scaled, fmin), fmax);
50     const float w_clamped = math_min_f32(math_max_f32(w_scaled, fmin), fmax);
51 
52     const int32_t x_biased = (int32_t) float_as_uint32(x_clamped + fmagic) - imagic;
53     const int32_t y_biased = (int32_t) float_as_uint32(y_clamped + fmagic) - imagic;
54     const int32_t z_biased = (int32_t) float_as_uint32(z_clamped + fmagic) - imagic;
55     const int32_t w_biased = (int32_t) float_as_uint32(w_clamped + fmagic) - imagic;
56 
57     output[0] = (int8_t) x_biased;
58     output[1] = (int8_t) y_biased;
59     output[2] = (int8_t) z_biased;
60     output[3] = (int8_t) w_biased;
61     output += 4;
62   }
63 }
64