Debian Codegraph @

src/qu8-vaddc/gen/qu8-vaddc-minmax-sse2-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 <emmintrin.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__sse2_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 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
50 __m128i va89ABCDEF = _mm_loadl_epi64((const __m128i*) (input_a + 8));
51 input_a += 16;
52
53 const __m128i vzero = _mm_setzero_si128();
54 va01234567 = _mm_unpacklo_epi8(va01234567, vzero);
55 va89ABCDEF = _mm_unpacklo_epi8(va89ABCDEF, vzero);
56
57 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
58 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
59 __m128i vaprod89ABCDEFhi = _mm_mulhi_epu16(va89ABCDEF, va_multiplier_lo);
60 const __m128i vaprod89ABCDEFlo = _mm_mullo_epi16(va89ABCDEF, va_multiplier_lo);
61
62 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
63 vaprod89ABCDEFhi = _mm_add_epi16(vaprod89ABCDEFhi, _mm_mullo_epi16(va89ABCDEF, va_multiplier_hi));
64
65
66 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
67 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
68 __m128i vacc89AB = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
69 __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
70
71 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
72 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
73 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
74 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
75
76 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
77 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
78
79
80 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
81
82 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
83
84 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, voutput_max);
85
86 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
87 output += 16;
88 }
89 if XNN_UNLIKELY(batch != 0) {
90 do {
91 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
92 input_a += 8;
93
94 va01234567 = _mm_unpacklo_epi8(va01234567, _mm_setzero_si128());
95
96 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
97 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
98
99 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
100
101
102 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
103 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
104
105 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
106 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
107
108 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
109
110 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
111 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
112 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
113
114 if XNN_LIKELY(batch >= (8 * sizeof(uint8_t))) {
115 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
116 output += 8;
117 batch -= 8 * sizeof(uint8_t);
118 } else {
119 if (batch & (4 * sizeof(uint8_t))) {
120 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
121 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
122 output += 4;
123 }
124 if (batch & (2 * sizeof(uint8_t))) {
125 unaligned_store_u16(output, (uint16_t) _mm_cvtsi128_si32(vout0123456701234567));
126 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
127 output += 2;
128 }
129 if (batch & (1 * sizeof(uint8_t))) {
130 *output = (uint8_t) _mm_cvtsi128_si32(vout0123456701234567);
131 }
132 batch = 0;
133 }
134 } while (batch != 0);
135 }
136 }