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
12 #include <tmmintrin.h>
13
14 #include <fp16/bitcasts.h>
15
16 #include <xnnpack/requantization-stubs.h>
17
18
xnn_requantize_precise__ssse3(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_requantize_precise__ssse3(
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 % 16 == 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 uint32_t multiplier = (scale_bits & UINT32_C(0x007FFFFF)) | UINT32_C(0x00800000);
34 const uint32_t shift = 127 + 23 - (scale_bits >> 23);
35 assert(shift >= 24);
36 assert(shift < 56);
37 const uint64_t rounding = UINT64_C(1) << (shift - 1);
38
39 const __m128i vmultiplier = _mm_set1_epi32(multiplier);
40 const __m128i vzero_point = _mm_set1_epi16((short) (uint16_t) zero_point);
41 const __m128i vqmin = _mm_set1_epi8((char) qmin);
42 const __m128i vqmax = _mm_set1_epi8((char) qmax);
43 const __m128i vshift = _mm_cvtsi32_si128((int) shift);
44 const __m128i vrounding = _mm_set1_epi64x(rounding);
45 for (; n != 0; n -= 16) {
46 const __m128i x = _mm_loadu_si128((const __m128i*) input);
47 const __m128i y = _mm_loadu_si128((const __m128i*) (input + 4));
48 const __m128i z = _mm_loadu_si128((const __m128i*) (input + 8));
49 const __m128i w = _mm_loadu_si128((const __m128i*) (input + 12));
50 input += 16;
51
52 const __m128i x_abs0123 = _mm_abs_epi32(x);
53 const __m128i y_abs0123 = _mm_abs_epi32(y);
54 const __m128i z_abs0123 = _mm_abs_epi32(z);
55 const __m128i w_abs0123 = _mm_abs_epi32(w);
56
57 const __m128i x_abs1032 = _mm_shuffle_epi32(x_abs0123, _MM_SHUFFLE(2, 3, 0, 1));
58 const __m128i y_abs1032 = _mm_shuffle_epi32(y_abs0123, _MM_SHUFFLE(2, 3, 0, 1));
59 const __m128i z_abs1032 = _mm_shuffle_epi32(z_abs0123, _MM_SHUFFLE(2, 3, 0, 1));
60 const __m128i w_abs1032 = _mm_shuffle_epi32(w_abs0123, _MM_SHUFFLE(2, 3, 0, 1));
61
62 const __m128i x_absmul02 = _mm_mul_epu32(x_abs0123, vmultiplier);
63 const __m128i y_absmul02 = _mm_mul_epu32(y_abs0123, vmultiplier);
64 const __m128i z_absmul02 = _mm_mul_epu32(z_abs0123, vmultiplier);
65 const __m128i w_absmul02 = _mm_mul_epu32(w_abs0123, vmultiplier);
66
67 const __m128i x_absmul13 = _mm_mul_epu32(x_abs1032, vmultiplier);
68 const __m128i y_absmul13 = _mm_mul_epu32(y_abs1032, vmultiplier);
69 const __m128i z_absmul13 = _mm_mul_epu32(z_abs1032, vmultiplier);
70 const __m128i w_absmul13 = _mm_mul_epu32(w_abs1032, vmultiplier);
71
72 const __m128i x_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(x_absmul02, vrounding), vshift);
73 const __m128i x_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(x_absmul13, vrounding), vshift);
74 const __m128i y_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(y_absmul02, vrounding), vshift);
75 const __m128i y_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(y_absmul13, vrounding), vshift);
76 const __m128i z_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(z_absmul02, vrounding), vshift);
77 const __m128i z_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(z_absmul13, vrounding), vshift);
78 const __m128i w_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(w_absmul02, vrounding), vshift);
79 const __m128i w_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(w_absmul13, vrounding), vshift);
80
81 const __m128i x_abs_scaled0213 = _mm_castps_si128(
82 _mm_shuffle_ps(_mm_castsi128_ps(x_abs_scaled02), _mm_castsi128_ps(x_abs_scaled13), _MM_SHUFFLE(2, 0, 2, 0)));
83 const __m128i y_abs_scaled0213 = _mm_castps_si128(
84 _mm_shuffle_ps(_mm_castsi128_ps(y_abs_scaled02), _mm_castsi128_ps(y_abs_scaled13), _MM_SHUFFLE(2, 0, 2, 0)));
85 const __m128i z_abs_scaled0213 = _mm_castps_si128(
86 _mm_shuffle_ps(_mm_castsi128_ps(z_abs_scaled02), _mm_castsi128_ps(z_abs_scaled13), _MM_SHUFFLE(2, 0, 2, 0)));
87 const __m128i w_abs_scaled0213 = _mm_castps_si128(
88 _mm_shuffle_ps(_mm_castsi128_ps(w_abs_scaled02), _mm_castsi128_ps(w_abs_scaled13), _MM_SHUFFLE(2, 0, 2, 0)));
89
90 const __m128i x_abs_scaled = _mm_shuffle_epi32(x_abs_scaled0213, _MM_SHUFFLE(3, 1, 2, 0));
91 const __m128i y_abs_scaled = _mm_shuffle_epi32(y_abs_scaled0213, _MM_SHUFFLE(3, 1, 2, 0));
92 const __m128i z_abs_scaled = _mm_shuffle_epi32(z_abs_scaled0213, _MM_SHUFFLE(3, 1, 2, 0));
93 const __m128i w_abs_scaled = _mm_shuffle_epi32(w_abs_scaled0213, _MM_SHUFFLE(3, 1, 2, 0));
94
95 const __m128i x_scaled = _mm_sign_epi32(x_abs_scaled, x);
96 const __m128i y_scaled = _mm_sign_epi32(y_abs_scaled, y);
97 const __m128i z_scaled = _mm_sign_epi32(z_abs_scaled, z);
98 const __m128i w_scaled = _mm_sign_epi32(w_abs_scaled, w);
99
100 const __m128i xy_packed = _mm_adds_epi16(_mm_packs_epi32(x_scaled, y_scaled), vzero_point);
101 const __m128i zw_packed = _mm_adds_epi16(_mm_packs_epi32(z_scaled, w_scaled), vzero_point);
102 const __m128i xyzw_packed = _mm_packus_epi16(xy_packed, zw_packed);
103 const __m128i xyzw_clamped = _mm_max_epu8(_mm_min_epu8(xyzw_packed, vqmax), vqmin);
104
105 // 4x PABSD
106 // 8x PSHUFD
107 // 8x PMULUDQ
108 // 8x PSRLQ
109 // 8x PADDQ
110 // 4x SHUFPS
111 // 4x PSIGND
112 // 2x PACKSSDW
113 // 1x PACKUSWB
114 // 2x PADDW
115 // 1x PMAXUB
116 // 1x PMINUB
117 // ---------------------
118 // 51 instructions total
119
120 _mm_storeu_si128((__m128i*) output, xyzw_clamped);
121 output += 16;
122 }
123 }
124