Debian Codegraph @

src/qu8-vadd/gen/qu8-vadd-minmax-sse2-mul16-ld64-u8.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 <emmintrin.h>
13
14 #include "xnnpack/unaligned.h"
15 #include "xnnpack/vbinary.h"
16
17
18 void xnn_qu8_vadd_minmax_ukernel__sse2_mul16_ld64_u8(
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 >= 8 * sizeof(uint8_t); batch -= 8 * sizeof(uint8_t)) {
52 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
53 __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b);
54 input_a += 8;
55 input_b += 8;
56
57 const __m128i vzero = _mm_setzero_si128();
58 va01234567 = _mm_unpacklo_epi8(va01234567, vzero);
59 vb01234567 = _mm_unpacklo_epi8(vb01234567, vzero);
60
61 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
62 __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
63 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
64 const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
65
66 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
67 vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
68
69
70 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
71 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
72
73 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
74 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
75
76 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
77 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
78
79 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
80
81
82 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
83
84 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
85
86 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
87
88 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
89 output += 8;
90 }
91 if XNN_UNLIKELY(batch != 0) {
92 {
93 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
94 __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b);
95
96 const __m128i vzero = _mm_setzero_si128();
97 va01234567 = _mm_unpacklo_epi8(va01234567, vzero);
98 vb01234567 = _mm_unpacklo_epi8(vb01234567, vzero);
99
100 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
101 __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
102 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
103 const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
104
105 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
106 vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
107
108
109 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
110 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
111
112 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
113 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
114
115 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
116 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
117
118 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
119
120 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
121 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
122 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
123
124 if (batch & (4 * sizeof(uint8_t))) {
125 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
126 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
127 output += 4;
128 }
129 if (batch & (2 * sizeof(uint8_t))) {
130 unaligned_store_u16(output, (uint16_t) _mm_cvtsi128_si32(vout0123456701234567));
131 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
132 output += 2;
133 }
134 if (batch & (1 * sizeof(uint8_t))) {
135 *output = (uint8_t) _mm_cvtsi128_si32(vout0123456701234567);
136 }
137 }
138 }
139 }