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 #pragma once
10
11 #if defined(__cplusplus) && (__cplusplus >= 201103L)
12 #include <cstdint>
13 #include <cstddef>
14 #include <cassert>
15 #include <cmath>
16 #else
17 #include <stdint.h>
18 #include <stddef.h>
19 #include <assert.h>
20 #include <math.h>
21 #endif
22
23 #include <fp16.h>
24
25 #include <xnnpack/common.h>
26 #include <xnnpack/params.h>
27 #include <xnnpack/scalar-utils.h>
28
29
xnn_q31_requantize(int32_t n,union xnn_q31_requantization_params params)30 static inline uint8_t xnn_q31_requantize(
31 int32_t n,
32 union xnn_q31_requantization_params params)
33 {
34 const int64_t product = (int64_t) n * (int64_t) params.scalar.multiplier;
35 const int32_t q31product = (int32_t) (uint32_t) ((uint64_t) (product + INT64_C(0x40000000)) >> 31);
36 const int32_t remainder = (q31product & params.scalar.remainder_mask) - (int32_t) (n < 0);
37 n = asr_s32(q31product, params.scalar.shift) + (int32_t) (remainder > params.scalar.remainder_threshold);
38 if (n < params.scalar.min_less_zero_point) {
39 n = params.scalar.min_less_zero_point;
40 }
41 if (n > params.scalar.max_less_zero_point) {
42 n = params.scalar.max_less_zero_point;
43 }
44
45 return (uint8_t) (n + params.scalar.zero_point);
46 }
47
xnn_avgpool_quantize(int32_t n,union xnn_q8_avgpool_params params)48 static inline uint8_t xnn_avgpool_quantize(
49 int32_t n,
50 union xnn_q8_avgpool_params params)
51 {
52 const int64_t product = (int64_t) n * (int64_t) params.scalar.multiplier;
53 const int64_t adjusted_product = product - (int64_t) (n < 0);
54
55 n = (int32_t) asr_s64(adjusted_product + params.scalar.rounding, params.scalar.right_shift);
56 if (n < params.scalar.output_min_less_zero_point) {
57 n = params.scalar.output_min_less_zero_point;
58 }
59 if (n > params.scalar.output_max_less_zero_point) {
60 n = params.scalar.output_max_less_zero_point;
61 }
62
63 return (uint8_t) (n + params.scalar.output_zero_point);
64 }
65
xnn_add_quantize(uint8_t a,uint8_t b,union xnn_q8_add_params params)66 static inline uint8_t xnn_add_quantize(
67 uint8_t a, uint8_t b,
68 union xnn_q8_add_params params)
69 {
70 // Multiply by factors and accumulate products.
71 int32_t acc = params.scalar.zero_point_product +
72 (int32_t) ((uint32_t) a * params.scalar.a_multiplier) +
73 (int32_t) ((uint32_t) b * params.scalar.b_multiplier);
74
75 // Shift right and round.
76 const int32_t rem = (acc & params.scalar.remainder_mask) - (int32_t) (acc < 0);
77 acc = asr_s32(acc, params.scalar.shift) + (int32_t) (rem > params.scalar.remainder_threshold);
78
79 // Clamp and add output zero point.
80 int32_t y = acc + params.scalar.y_zero_point;
81 if (y >= params.scalar.y_max) {
82 y = params.scalar.y_max;
83 }
84 if (y <= params.scalar.y_min) {
85 y = params.scalar.y_min;
86 }
87 return (uint8_t) y;
88 }
89