blob: ff6c6eda4b2be1d238af994b586f1406e8e264e6 [file] [log] [blame]
Marat Dukhan0270d9f2020-08-11 00:56:46 -07001// Auto-generated file. Do not edit!
2// Template: src/qs8-vaddc/sse-mul16-ld64.c.in
3// Generator: tools/xngen
4//
5// Copyright 2020 Google LLC
6//
7// This source code is licensed under the BSD-style license found in the
8// LICENSE file in the root directory of this source tree.
9
10#include <assert.h>
11
12#include <smmintrin.h>
13
14#include <xnnpack/vadd.h>
15
16
17void xnn_qs8_vaddc_minmax_ukernel__sse41_mul16_ld64_x32(
18 size_t n,
Marat Dukhan076bcfe2021-07-19 19:24:42 -070019 const int8_t* input_a,
20 const int8_t* input_b,
Marat Dukhan0270d9f2020-08-11 00:56:46 -070021 int8_t* output,
Marat Dukhan6e0fc392021-07-19 18:38:24 -070022 const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
Marat Dukhan0270d9f2020-08-11 00:56:46 -070023{
Marat Dukhan49d90052021-07-19 19:59:30 -070024 const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
25 const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
Marat Dukhan947c2982021-07-20 13:11:29 -070026 const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
Marat Dukhan0270d9f2020-08-11 00:56:46 -070027 const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
28 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
29 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
30 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
31
Marat Dukhana842fef2021-07-19 21:07:40 -070032 __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
33 vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
34 vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
Marat Dukhan0270d9f2020-08-11 00:56:46 -070035 for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
Marat Dukhan076bcfe2021-07-19 19:24:42 -070036 const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
37 const __m128i va89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
38 const __m128i vaGHIJKLMN = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 16)));
39 const __m128i vaOPQRSTUV = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 24)));
40 input_a += 32;
Marat Dukhan0270d9f2020-08-11 00:56:46 -070041
42
Marat Dukhan076bcfe2021-07-19 19:24:42 -070043 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
44 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
45 __m128i vaprod89ABCDEFhi = _mm_mulhi_epu16(va89ABCDEF, va_multiplier_lo);
46 const __m128i vaprod89ABCDEFlo = _mm_mullo_epi16(va89ABCDEF, va_multiplier_lo);
47 __m128i vaprodGHIJKLMNhi = _mm_mulhi_epu16(vaGHIJKLMN, va_multiplier_lo);
48 const __m128i vaprodGHIJKLMNlo = _mm_mullo_epi16(vaGHIJKLMN, va_multiplier_lo);
49 __m128i vaprodOPQRSTUVhi = _mm_mulhi_epu16(vaOPQRSTUV, va_multiplier_lo);
50 const __m128i vaprodOPQRSTUVlo = _mm_mullo_epi16(vaOPQRSTUV, va_multiplier_lo);
Marat Dukhan0270d9f2020-08-11 00:56:46 -070051
Marat Dukhan076bcfe2021-07-19 19:24:42 -070052 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
53 vaprod89ABCDEFhi = _mm_add_epi16(vaprod89ABCDEFhi, _mm_mullo_epi16(va89ABCDEF, va_multiplier_hi));
54 vaprodGHIJKLMNhi = _mm_add_epi16(vaprodGHIJKLMNhi, _mm_mullo_epi16(vaGHIJKLMN, va_multiplier_hi));
55 vaprodOPQRSTUVhi = _mm_add_epi16(vaprodOPQRSTUVhi, _mm_mullo_epi16(vaOPQRSTUV, va_multiplier_hi));
Marat Dukhan0270d9f2020-08-11 00:56:46 -070056
Marat Dukhan076bcfe2021-07-19 19:24:42 -070057 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
58 vaprod89ABCDEFhi = _mm_sub_epi16(vaprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(va89ABCDEF, 15), va_multiplier_lo));
59 vaprodGHIJKLMNhi = _mm_sub_epi16(vaprodGHIJKLMNhi, _mm_and_si128(_mm_srai_epi16(vaGHIJKLMN, 15), va_multiplier_lo));
60 vaprodOPQRSTUVhi = _mm_sub_epi16(vaprodOPQRSTUVhi, _mm_and_si128(_mm_srai_epi16(vaOPQRSTUV, 15), va_multiplier_lo));
Marat Dukhan0270d9f2020-08-11 00:56:46 -070061
Marat Dukhana842fef2021-07-19 21:07:40 -070062 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
63 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
64 __m128i vacc89AB = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
65 __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
66 __m128i vaccGHIJ = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprodGHIJKLMNlo, vaprodGHIJKLMNhi));
67 __m128i vaccKLMN = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprodGHIJKLMNlo, vaprodGHIJKLMNhi));
68 __m128i vaccOPQR = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprodOPQRSTUVlo, vaprodOPQRSTUVhi));
69 __m128i vaccSTUV = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprodOPQRSTUVlo, vaprodOPQRSTUVhi));
Marat Dukhan0270d9f2020-08-11 00:56:46 -070070
Marat Dukhan8a045652021-07-20 14:07:21 -070071 vacc0123 = _mm_sra_epi32(_mm_add_epi32(vacc0123, vrounding), vshift);
72 vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
73 vacc89AB = _mm_sra_epi32(_mm_add_epi32(vacc89AB, vrounding), vshift);
74 vaccCDEF = _mm_sra_epi32(_mm_add_epi32(vaccCDEF, vrounding), vshift);
75 vaccGHIJ = _mm_sra_epi32(_mm_add_epi32(vaccGHIJ, vrounding), vshift);
76 vaccKLMN = _mm_sra_epi32(_mm_add_epi32(vaccKLMN, vrounding), vshift);
77 vaccOPQR = _mm_sra_epi32(_mm_add_epi32(vaccOPQR, vrounding), vshift);
78 vaccSTUV = _mm_sra_epi32(_mm_add_epi32(vaccSTUV, vrounding), vshift);
Marat Dukhan0270d9f2020-08-11 00:56:46 -070079
80 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
81 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
82 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
83 __m128i voutOPQRSTUV = _mm_adds_epi16(_mm_packs_epi32(vaccOPQR, vaccSTUV), voutput_zero_point);
84
85 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
86 vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
87 voutGHIJKLMN = _mm_max_epi16(voutGHIJKLMN, voutput_min);
88 voutOPQRSTUV = _mm_max_epi16(voutOPQRSTUV, voutput_min);
89
90 vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
91 vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
92 voutGHIJKLMN = _mm_min_epi16(voutGHIJKLMN, voutput_max);
93 voutOPQRSTUV = _mm_min_epi16(voutOPQRSTUV, voutput_max);
94
95 const __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
96 const __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
97
98 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
99 _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
100 output += 32;
101 }
102 if XNN_UNLIKELY(n != 0) {
103 do {
Marat Dukhan076bcfe2021-07-19 19:24:42 -0700104 const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
105 input_a += 8;
Marat Dukhan0270d9f2020-08-11 00:56:46 -0700106
107
Marat Dukhan076bcfe2021-07-19 19:24:42 -0700108 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
109 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
Marat Dukhan0270d9f2020-08-11 00:56:46 -0700110
Marat Dukhan076bcfe2021-07-19 19:24:42 -0700111 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
Marat Dukhan0270d9f2020-08-11 00:56:46 -0700112
Marat Dukhan076bcfe2021-07-19 19:24:42 -0700113 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
Marat Dukhan0270d9f2020-08-11 00:56:46 -0700114
Marat Dukhana842fef2021-07-19 21:07:40 -0700115 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
116 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
Marat Dukhan0270d9f2020-08-11 00:56:46 -0700117
Marat Dukhan8a045652021-07-20 14:07:21 -0700118 vacc0123 = _mm_sra_epi32(_mm_add_epi32(vacc0123, vrounding), vshift);
119 vacc4567 = _mm_sra_epi32(_mm_add_epi32(vacc4567, vrounding), vshift);
Marat Dukhan0270d9f2020-08-11 00:56:46 -0700120
121 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
122 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
123 vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
124
125 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
126
127 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
128 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
129 output += 8;
Marat Dukhanba7b2792020-09-02 14:26:45 -0700130 n -= 8 * sizeof(int8_t);
Marat Dukhan0270d9f2020-08-11 00:56:46 -0700131 } else {
132 if (n & (4 * sizeof(int8_t))) {
133 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
134 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
135 output += 4;
136 }
137 if (n & (2 * sizeof(int8_t))) {
138 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
139 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
140 output += 2;
141 }
142 if (n & (1 * sizeof(int8_t))) {
143 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
Marat Dukhan0270d9f2020-08-11 00:56:46 -0700144 }
145 n = 0;
146 }
147 } while (n != 0);
148 }
149}