Debian Codegraph @

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

raw
1 // 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/common.h"
15 #include "xnnpack/unaligned.h"
16 #include "xnnpack/vbinary.h"
17
18
19 void xnn_qu8_vaddc_minmax_ukernel__sse41_mul16_ld64_u16(
20 size_t batch,
21 const uint8_t* input_a,
22 const uint8_t* input_b,
23 uint8_t* output,
24 const struct xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26 assert(batch != 0);
27 assert(batch % sizeof(uint8_t) == 0);
28 assert(input_a != NULL);
29 assert(input_b != NULL);
30 assert(output != NULL);
31
32 const __m128i vbias = _mm_set1_epi32(params->scalar.b_multiplier * (int32_t) *input_b + params->scalar.bias);
33 const __m128i va_multiplier_lo = _mm_set1_epi16(params->scalar.a_multiplier);
34 const __m128i va_multiplier_hi = _mm_set1_epi16((uint32_t)params->scalar.a_multiplier >> 16);
35 const __m128i vshift = _mm_cvtsi32_si128((int) params->scalar.shift);
36 const __m128i voutput_zero_point = _mm_set1_epi16(params->scalar.output_zero_point);
37 const __m128i voutput_min = _mm_set1_epi8(params->scalar.output_min);
38 const __m128i voutput_max = _mm_set1_epi8(params->scalar.output_max);
39
40 XNN_FORCE_REALIZATION(vbias);
41 XNN_FORCE_REALIZATION(va_multiplier_lo);
42 XNN_FORCE_REALIZATION(va_multiplier_hi);
43 XNN_FORCE_REALIZATION(vshift);
44 XNN_FORCE_REALIZATION(voutput_zero_point);
45 XNN_FORCE_REALIZATION(voutput_min);
46 XNN_FORCE_REALIZATION(voutput_max);
47
48 for (; batch >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) {
49 const __m128i va01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
50 const __m128i va89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
51 input_a += 16;
52
53
54 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
55 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
56 __m128i vaprod89ABCDEFhi = _mm_mulhi_epu16(va89ABCDEF, va_multiplier_lo);
57 const __m128i vaprod89ABCDEFlo = _mm_mullo_epi16(va89ABCDEF, va_multiplier_lo);
58
59 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
60 vaprod89ABCDEFhi = _mm_add_epi16(vaprod89ABCDEFhi, _mm_mullo_epi16(va89ABCDEF, va_multiplier_hi));
61
62
63 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
64 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
65 __m128i vacc89AB = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
66 __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
67
68 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
69 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
70 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
71 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
72
73 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
74 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
75
76
77 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
78
79 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
80
81 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, voutput_max);
82
83 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
84 output += 16;
85 }
86 if XNN_UNLIKELY(batch != 0) {
87 do {
88 const __m128i va01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
89 input_a += 8;
90
91
92 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
93 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
94
95 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
96
97
98 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
99 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
100
101 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
102 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
103
104 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
105
106 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
107 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
108 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
109
110 if XNN_LIKELY(batch >= (8 * sizeof(uint8_t))) {
111 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
112 output += 8;
113 batch -= 8 * sizeof(uint8_t);
114 } else {
115 if (batch & (4 * sizeof(uint8_t))) {
116 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
117 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
118 output += 4;
119 }
120 if (batch & (2 * sizeof(uint8_t))) {
121 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
122 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
123 output += 2;
124 }
125 if (batch & (1 * sizeof(uint8_t))) {
126 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
127 }
128 batch = 0;
129 }
130 } while (batch != 0);
131 }
132 }