Debian Codegraph @

src/qu8-vadd/gen/qu8-vadd-minmax-sse41-mul16-ld64-u16.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vadd/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/unaligned.h"
15 #include "xnnpack/vbinary.h"
16
17
18 void xnn_qu8_vadd_minmax_ukernel__sse41_mul16_ld64_u16(
19 size_t batch,
20 const uint8_t* input_a,
21 const uint8_t* input_b,
22 uint8_t* output,
23 const struct xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(batch != 0);
26 assert(batch % sizeof(uint8_t) == 0);
27 assert(input_a != NULL);
28 assert(input_b != NULL);
29 assert(output != NULL);
30
31 const __m128i vbias = _mm_set1_epi32(params->scalar.bias);
32 const __m128i va_multiplier_lo = _mm_set1_epi16(params->scalar.a_multiplier);
33 const __m128i va_multiplier_hi = _mm_set1_epi16((uint32_t)params->scalar.a_multiplier >> 16);
34 const __m128i vb_multiplier_lo = _mm_set1_epi16(params->scalar.b_multiplier);
35 const __m128i vb_multiplier_hi = _mm_set1_epi16((uint32_t)params->scalar.b_multiplier >> 16);
36 const __m128i vshift = _mm_cvtsi32_si128((int) params->scalar.shift);
37 const __m128i voutput_zero_point = _mm_set1_epi16(params->scalar.output_zero_point);
38 const __m128i voutput_min = _mm_set1_epi8(params->scalar.output_min);
39 const __m128i voutput_max = _mm_set1_epi8(params->scalar.output_max);
40
41 XNN_FORCE_REALIZATION(vbias);
42 XNN_FORCE_REALIZATION(va_multiplier_lo);
43 XNN_FORCE_REALIZATION(va_multiplier_hi);
44 XNN_FORCE_REALIZATION(vb_multiplier_lo);
45 XNN_FORCE_REALIZATION(vb_multiplier_hi);
46 XNN_FORCE_REALIZATION(vshift);
47 XNN_FORCE_REALIZATION(voutput_zero_point);
48 XNN_FORCE_REALIZATION(voutput_min);
49 XNN_FORCE_REALIZATION(voutput_max);
50
51 for (; batch >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) {
52 const __m128i va01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
53 const __m128i vb01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_b));
54 const __m128i va89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
55 const __m128i vb89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (input_b + 8)));
56 input_a += 16;
57 input_b += 16;
58
59
60 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
61 __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
62 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
63 const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
64 __m128i vaprod89ABCDEFhi = _mm_mulhi_epu16(va89ABCDEF, va_multiplier_lo);
65 __m128i vbprod89ABCDEFhi = _mm_mulhi_epu16(vb89ABCDEF, vb_multiplier_lo);
66 const __m128i vaprod89ABCDEFlo = _mm_mullo_epi16(va89ABCDEF, va_multiplier_lo);
67 const __m128i vbprod89ABCDEFlo = _mm_mullo_epi16(vb89ABCDEF, vb_multiplier_lo);
68
69 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
70 vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
71 vaprod89ABCDEFhi = _mm_add_epi16(vaprod89ABCDEFhi, _mm_mullo_epi16(va89ABCDEF, va_multiplier_hi));
72 vbprod89ABCDEFhi = _mm_add_epi16(vbprod89ABCDEFhi, _mm_mullo_epi16(vb89ABCDEF, vb_multiplier_hi));
73
74
75 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
76 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
77 __m128i vacc89AB = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
78 __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
79
80 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
81 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
82 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vbprod89ABCDEFlo, vbprod89ABCDEFhi));
83 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vbprod89ABCDEFlo, vbprod89ABCDEFhi));
84
85 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
86 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
87 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
88 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
89
90 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
91 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
92
93
94 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
95
96 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
97
98 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, voutput_max);
99
100 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
101 output += 16;
102 }
103 if XNN_UNLIKELY(batch != 0) {
104 do {
105 const __m128i va01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
106 const __m128i vb01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_b));
107 input_a += 8;
108 input_b += 8;
109
110
111 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
112 __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
113 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
114 const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
115
116 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
117 vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
118
119
120 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
121 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
122
123 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
124 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
125
126 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
127 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
128
129 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
130
131 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
132 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
133 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
134
135 if XNN_LIKELY(batch >= (8 * sizeof(uint8_t))) {
136 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
137 output += 8;
138 batch -= 8 * sizeof(uint8_t);
139 } else {
140 if (batch & (4 * sizeof(uint8_t))) {
141 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
142 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
143 output += 4;
144 }
145 if (batch & (2 * sizeof(uint8_t))) {
146 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
147 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
148 output += 2;
149 }
150 if (batch & (1 * sizeof(uint8_t))) {
151 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
152 }
153 batch = 0;
154 }
155 } while (batch != 0);
156 }
157 }