Debian Codegraph @

src/qu8-vaddc/gen/qu8-vaddc-minmax-sse2-mul16-ld64-u8.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_u8(
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 >= 8 * sizeof(uint8_t); batch -= 8 * sizeof(uint8_t)) {
49 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
50 input_a += 8;
51
52 const __m128i vzero = _mm_setzero_si128();
53 va01234567 = _mm_unpacklo_epi8(va01234567, vzero);
54
55 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
56 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
57
58 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
59
60
61 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
62 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
63
64 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
65 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
66
67 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
68
69
70 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
71
72 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
73
74 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
75
76 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
77 output += 8;
78 }
79 if XNN_UNLIKELY(batch != 0) {
80 {
81 __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
82
83 va01234567 = _mm_unpacklo_epi8(va01234567, _mm_setzero_si128());
84
85 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
86 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
87
88 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
89
90
91 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
92 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
93
94 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
95 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
96
97 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
98
99 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
100 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
101 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
102
103 if (batch & (4 * sizeof(uint8_t))) {
104 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
105 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
106 output += 4;
107 }
108 if (batch & (2 * sizeof(uint8_t))) {
109 unaligned_store_u16(output, (uint16_t) _mm_cvtsi128_si32(vout0123456701234567));
110 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
111 output += 2;
112 }
113 if (batch & (1 * sizeof(uint8_t))) {
114 *output = (uint8_t) _mm_cvtsi128_si32(vout0123456701234567);
115 }
116 }
117 }
118 }