Debian Codegraph @

src/qu8-vaddc/gen/qu8-vaddc-minmax-avx2-mul32-ld64-u8.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vaddc/avx2-mul32-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 <immintrin.h>
13
14 #include "xnnpack/intrinsics-polyfill.h"
15 #include "xnnpack/vbinary.h"
16
17
18 void xnn_qu8_vaddc_minmax_ukernel__avx2_mul32_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 __m256i vbias = _mm256_set1_epi32(params->scalar.b_multiplier * (int32_t) *input_b + params->scalar.bias);
32 const __m256i va_multiplier = _mm256_set1_epi32(params->scalar.a_multiplier);
33 const __m128i vshift = _mm_set1_epi64x(params->scalar.shift);
34 const __m128i voutput_zero_point = _mm_set1_epi16(params->scalar.output_zero_point);
35 const __m128i voutput_min = _mm_set1_epi8(params->scalar.output_min);
36 const __m128i voutput_max = _mm_set1_epi8(params->scalar.output_max);
37
38 XNN_FORCE_REALIZATION(vbias);
39 XNN_FORCE_REALIZATION(va_multiplier);
40 XNN_FORCE_REALIZATION(voutput_zero_point);
41 XNN_FORCE_REALIZATION(voutput_min);
42 XNN_FORCE_REALIZATION(voutput_max);
43
44 for (; batch >= 8 * sizeof(uint8_t); batch -= 8 * sizeof(uint8_t)) {
45 const __m256i va01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) input_a));
46 input_a += 8;
47
48 __m256i vacc01234567 = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va01234567, va_multiplier));
49
50 vacc01234567 = _mm256_sra_epi32(vacc01234567, vshift);
51
52 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
53
54 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
55
56 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
57
58 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
59
60 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
61 output += 8;
62 }
63 if XNN_UNLIKELY(batch != 0) {
64 {
65 const __m256i va01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) input_a));
66
67 __m256i vacc01234567 = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va01234567, va_multiplier));
68
69 vacc01234567 = _mm256_sra_epi32(vacc01234567, vshift);
70
71 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
72 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
73 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
74 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
75
76 if (batch & (4 * sizeof(uint8_t))) {
77 _mm_storeu_si32(output, vout0123456701234567);
78 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
79 output += 4;
80 }
81 if (batch & (2 * sizeof(uint8_t))) {
82 _mm_storeu_si16(output, vout0123456701234567);
83 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
84 output += 2;
85 }
86 if (batch & (1 * sizeof(uint8_t))) {
87 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
88 }
89 }
90 }
91 }