• 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 <stdint.h>
11 #include <stddef.h>
12 
13 #include <fp16/bitcasts.h>
14 
15 #include <xnnpack/scalar-utils.h>
16 #include <xnnpack/requantization-stubs.h>
17 
18 
xnn_qu8_requantize_precise__scalar_signed64(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_precise__scalar_signed64(
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 uint32_t scale_bits = fp32_to_bits(scale);
33   const int32_t multiplier = ((int32_t) scale_bits & INT32_C(0x007FFFFF)) | INT32_C(0x00800000);
34   const uint32_t shift = 127 + 23 - (scale_bits >> 23);
35   assert(shift >= 24);
36   assert(shift < 56);
37 
38   const int64_t rounding = INT64_C(1) << (shift - 1);
39   const int32_t smin = (int32_t)(uint32_t) qmin - (int32_t)(uint32_t) zero_point;
40   const int32_t smax = (int32_t)(uint32_t) qmax - (int32_t)(uint32_t) zero_point;
41   for (; n != 0; n -= 4) {
42     const int32_t x = input[0];
43     const int32_t y = input[1];
44     const int32_t z = input[2];
45     const int32_t w = input[3];
46     input += 4;
47 
48     // Compute full 64-bit product of signed 32-bit factors.
49     //
50     // Note: multiplier can be treated as either signed or unsigned.
51     const int64_t x_product = (int64_t) x * (int64_t) multiplier;
52     const int64_t y_product = (int64_t) y * (int64_t) multiplier;
53     const int64_t z_product = (int64_t) z * (int64_t) multiplier;
54     const int64_t w_product = (int64_t) w * (int64_t) multiplier;
55 
56     // Adjust product before subsequent shift with rounding up to simulate shift with rounding away from zero.
57     const int64_t x_adjusted_product = x_product - (int64_t)(x < 0);
58     const int64_t y_adjusted_product = y_product - (int64_t)(y < 0);
59     const int64_t z_adjusted_product = z_product - (int64_t)(z < 0);
60     const int64_t w_adjusted_product = w_product - (int64_t)(w < 0);
61 
62     // Arithmetically shift the full 64-bit product right with rounding.
63     // Rounding is performed towards closest integer, with midpoints rounded up.
64     //
65     // Note that although rounding is precomputed, it is dependent on shift value, and on processors with 64-bit
66     // "right shift with rounding" instruction each line below can be represented by just one such instruction
67     // (e.g. VRSHL.S64 on ARM NEON, SRSHL in ARM64 Advanced SIMD).
68     const int32_t x_scaled = (int32_t) asr_s64(x_adjusted_product + rounding, shift);
69     const int32_t y_scaled = (int32_t) asr_s64(y_adjusted_product + rounding, shift);
70     const int32_t z_scaled = (int32_t) asr_s64(z_adjusted_product + rounding, shift);
71     const int32_t w_scaled = (int32_t) asr_s64(w_adjusted_product + rounding, shift);
72 
73     // Clamp scaled value with zero point between (qmin - zero point) and (qmax - zero point).
74     const int32_t x_clamped = x_scaled < smin ? smin : x_scaled > smax ? smax : x_scaled;
75     const int32_t y_clamped = y_scaled < smin ? smin : y_scaled > smax ? smax : y_scaled;
76     const int32_t z_clamped = z_scaled < smin ? smin : z_scaled > smax ? smax : z_scaled;
77     const int32_t w_clamped = w_scaled < smin ? smin : w_scaled > smax ? smax : w_scaled;
78 
79     // Add zero point to clamped value.
80     // The result is guaranteed to be in [qmin, qmax] range.
81     //
82     // This addition can not be safely done before clamping, because scaled values are in [-2147483520, 2147483519]
83     // range, so addition of zero point (which can be up to 255) can overflow signed 32-bit integer.
84     const int32_t x_biased = x_clamped + zero_point;
85     const int32_t y_biased = y_clamped + zero_point;
86     const int32_t z_biased = z_clamped + zero_point;
87     const int32_t w_biased = w_clamped + zero_point;
88 
89     output[0] = (uint8_t) x_biased;
90     output[1] = (uint8_t) y_biased;
91     output[2] = (uint8_t) z_biased;
92     output[3] = (uint8_t) w_biased;
93     output += 4;
94   }
95 }
96