• 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 #pragma once
10 
11 #if defined(__cplusplus) && (__cplusplus >= 201103L)
12   #include <climits>
13   #include <cstdint>
14   #include <cstdbool>
15   #include <cassert>
16 #else
17   #include <limits.h>
18   #include <stdint.h>
19   #include <stdbool.h>
20   #include <assert.h>
21 #endif
22 
23 #include <fp16.h>
24 
25 #include <xnnpack/common.h>
26 
27 
28 #if defined(__clang__) && !defined(__pnacl__)
29   #if __clang_major__ == 3 && __clang_minor__ >= 7 || __clang_major__ > 3
30     #define XNN_IGNORE_SHIFT_BASE_UB __attribute__((__no_sanitize__("shift-base")))
31   #else
32     #define XNN_IGNORE_SHIFT_BASE_UB
33   #endif
34 #elif defined(__GNUC__)
35   #if __GNUC__ >= 8
36     #define XNN_IGNORE_SHIFT_BASE_UB __attribute__((__no_sanitize__("shift-base")))
37   #elif __GNUC__ == 4 && __GNUC_MINOR__ >= 9 || __GNUC__ > 4
38     // 4.9 <= gcc < 8 support ubsan, but doesn't support no_sanitize attribute
39     #define XNN_IGNORE_SHIFT_BASE_UB
40     #ifndef XNN_USE_SHIFT_BASE_UB_WORKAROUND
41       #define XNN_USE_SHIFT_BASE_UB_WORKAROUND 1
42     #endif
43   #else
44     #define XNN_IGNORE_SHIFT_BASE_UB
45   #endif
46 #else
47   #define XNN_IGNORE_SHIFT_BASE_UB
48 #endif
49 
50 XNN_IGNORE_SHIFT_BASE_UB
asr_s32(int32_t x,uint32_t n)51 inline static int32_t asr_s32(int32_t x, uint32_t n) {
52   #ifdef XNN_USE_SHIFT_BASE_UB_WORKAROUND
53     #if XNN_ARCH_X86_64 || XNN_ARCH_ARM64
54       return (int32_t) ((uint64_t) (int64_t) x >> n);
55     #else
56       return x >= 0 ? x >> n : ~(~x >> n);
57     #endif
58   #else
59     return x >> n;
60   #endif
61 }
62 
63 XNN_IGNORE_SHIFT_BASE_UB
asr_s64(int64_t x,uint32_t n)64 inline static int64_t asr_s64(int64_t x, uint32_t n) {
65   #ifdef XNN_USE_SHIFT_BASE_UB_WORKAROUND
66     return x >= 0 ? x >> n : ~(~x >> n);
67   #else
68     return x >> n;
69   #endif
70 }
71 
scalar_requantize_precise(int32_t value,float scale,uint8_t zero_point,uint8_t qmin,uint8_t qmax)72 inline static uint8_t scalar_requantize_precise(
73   int32_t value,
74   float scale,
75   uint8_t zero_point,
76   uint8_t qmin,
77   uint8_t qmax)
78 {
79   assert(scale < 1.0f);
80   assert(scale >= 0x1.0p-32f);
81 
82   const uint32_t scale_bits = fp32_to_bits(scale);
83   const uint32_t multiplier = (scale_bits & UINT32_C(0x007FFFFF)) | UINT32_C(0x00800000);
84   const uint32_t shift = 127 + 23 - (scale_bits >> 23);
85   assert(shift >= 24);
86   assert(shift < 56);
87 
88   // Compute absolute value of input as unsigned 32-bit int.
89   // All further computations will work with unsigned values to avoid undefined behaviour on signed operations.
90   const uint32_t abs_value = (value >= 0) ? (uint32_t) value : -(uint32_t) value;
91 
92   // Compute full 64-bit product of 32-bit factors
93   const uint64_t product = (uint64_t) abs_value * (uint64_t) multiplier;
94 
95   // Shift the full 64-bit product right with rounding.
96   // Rounding is performed towards closest integer, with midpoints rounded up (same as away from zero).
97   const uint64_t rounding = UINT64_C(1) << (shift - 1);
98   const uint32_t abs_scaled_value = (uint32_t) ((product + rounding) >> shift);
99 
100   // Copy the sign of input to scaled absolute input value.
101   const int32_t scaled_value = (int32_t) (value >= 0 ? abs_scaled_value : -abs_scaled_value);
102 
103   // Clamp scaled value with zero point between smin and smax.
104   int32_t clamped_value = scaled_value;
105   const int32_t smin = (int32_t) (uint32_t) qmin - (int32_t) (uint32_t) zero_point;
106   if (clamped_value < smin) {
107     clamped_value = smin;
108   }
109   const int32_t smax = (int32_t) (uint32_t) qmax - (int32_t) (uint32_t) zero_point;
110   if (clamped_value > smax) {
111     clamped_value = smax;
112   }
113 
114   // Add zero point to clamped value.
115   const int32_t biased_value = clamped_value + (int32_t) (uint32_t) zero_point;
116 
117   return biased_value;
118 }
119