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/requantization-stubs.h>
17
18
xnn_qu8_requantize_fp32__scalar_magic(size_t n,const int32_t * input,float scale,uint8_t zero_point,uint8_t qmin,uint8_t qmax,uint8_t * output)19 void xnn_qu8_requantize_fp32__scalar_magic(
20 size_t n,
21 const int32_t* input,
22 float scale,
23 uint8_t zero_point,
24 uint8_t qmin,
25 uint8_t qmax,
26 uint8_t* output)
27 {
28 assert(n % 4 == 0);
29 assert(scale < 1.0f);
30 assert(scale >= 0x1.0p-32f);
31
32 const float fmin = (float) ((int32_t)(uint32_t) qmin - (int32_t)(uint32_t) zero_point);
33 const float fmax = (float) ((int32_t)(uint32_t) qmax - (int32_t)(uint32_t) zero_point);
34 const float fmagic = 12582912.0f;
35 const int32_t imagic = INT32_C(0x4B400000) - (int32_t)(uint32_t) zero_point;
36 for (; n != 0; n -= 4) {
37 const int32_t x = input[0];
38 const int32_t y = input[1];
39 const int32_t z = input[2];
40 const int32_t w = input[3];
41 input += 4;
42
43 const float x_scaled = (float) x * scale;
44 const float y_scaled = (float) y * scale;
45 const float z_scaled = (float) z * scale;
46 const float w_scaled = (float) w * scale;
47
48 const float x_clamped = x_scaled < fmin ? fmin : x_scaled > fmax ? fmax : x_scaled;
49 const float y_clamped = y_scaled < fmin ? fmin : y_scaled > fmax ? fmax : y_scaled;
50 const float z_clamped = z_scaled < fmin ? fmin : z_scaled > fmax ? fmax : z_scaled;
51 const float w_clamped = w_scaled < fmin ? fmin : w_scaled > fmax ? fmax : w_scaled;
52
53 const int32_t x_biased = (int32_t) fp32_to_bits(x_clamped + fmagic) - imagic;
54 const int32_t y_biased = (int32_t) fp32_to_bits(y_clamped + fmagic) - imagic;
55 const int32_t z_biased = (int32_t) fp32_to_bits(z_clamped + fmagic) - imagic;
56 const int32_t w_biased = (int32_t) fp32_to_bits(w_clamped + fmagic) - imagic;
57
58 output[0] = (uint8_t) x_biased;
59 output[1] = (uint8_t) y_biased;
60 output[2] = (uint8_t) z_biased;
61 output[3] = (uint8_t) w_biased;
62 output += 4;
63 }
64 }
65