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 <fp16/bitcasts.h>
15
16 #include <xnnpack/math.h>
17 #include <xnnpack/requantization-stubs.h>
18
19
xnn_qu8_requantize_fp32__scalar_fmagic(size_t n,const int32_t * input,float scale,uint8_t zero_point,uint8_t qmin,uint8_t qmax,uint8_t * output)20 void xnn_qu8_requantize_fp32__scalar_fmagic(
21 size_t n,
22 const int32_t* input,
23 float scale,
24 uint8_t zero_point,
25 uint8_t qmin,
26 uint8_t qmax,
27 uint8_t* output)
28 {
29 assert(n % 4 == 0);
30 assert(scale < 1.0f);
31 assert(scale >= 0x1.0p-32f);
32
33 const float fmin = (float) ((int32_t) (uint32_t) qmin - (int32_t) (uint32_t) zero_point);
34 const float fmax = (float) ((int32_t) (uint32_t) qmax - (int32_t) (uint32_t) zero_point);
35 const float fmagic = 12582912.0f;
36 const int32_t imagic = INT32_C(0x4B400000) - (int32_t)(uint32_t) zero_point;
37 for (; n != 0; n -= 4) {
38 const int32_t x = input[0];
39 const int32_t y = input[1];
40 const int32_t z = input[2];
41 const int32_t w = input[3];
42 input += 4;
43
44 const float x_scaled = (float) x * scale;
45 const float y_scaled = (float) y * scale;
46 const float z_scaled = (float) z * scale;
47 const float w_scaled = (float) w * scale;
48
49 const float x_clamped = math_min_f32(math_max_f32(x_scaled, fmin), fmax);
50 const float y_clamped = math_min_f32(math_max_f32(y_scaled, fmin), fmax);
51 const float z_clamped = math_min_f32(math_max_f32(z_scaled, fmin), fmax);
52 const float w_clamped = math_min_f32(math_max_f32(w_scaled, fmin), fmax);
53
54 const int32_t x_biased = (int32_t) fp32_to_bits(x_clamped + fmagic) - imagic;
55 const int32_t y_biased = (int32_t) fp32_to_bits(y_clamped + fmagic) - imagic;
56 const int32_t z_biased = (int32_t) fp32_to_bits(z_clamped + fmagic) - imagic;
57 const int32_t w_biased = (int32_t) fp32_to_bits(w_clamped + fmagic) - imagic;
58
59 output[0] = (uint8_t) x_biased;
60 output[1] = (uint8_t) y_biased;
61 output[2] = (uint8_t) z_biased;
62 output[3] = (uint8_t) w_biased;
63 output += 4;
64 }
65 }
66