Debian Codegraph @

src/qu8-vaddc/gen/qu8-vaddc-minmax-avx2-mul32-ld64-u16.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_u16(
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 __m256i voutput_zero_point = _mm256_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 >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) {
45 const __m256i va01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) input_a));
46 const __m256i va89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
47 input_a += 16;
48
49 __m256i vacc01234567 = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va01234567, va_multiplier));
50 __m256i vacc89ABCDEF = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va89ABCDEF, va_multiplier));
51
52 vacc01234567 = _mm256_sra_epi32(vacc01234567, vshift);
53 vacc89ABCDEF = _mm256_sra_epi32(vacc89ABCDEF, vshift);
54
55 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
56
57 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packus_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
58
59 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
60
61 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, voutput_max);
62
63 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
64 output += 16;
65 }
66 if XNN_UNLIKELY(batch != 0) {
67 do {
68 const __m256i va01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) input_a));
69 input_a += 8;
70
71 __m256i vacc01234567 = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va01234567, va_multiplier));
72
73 vacc01234567 = _mm256_sra_epi32(vacc01234567, vshift);
74
75 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point));
76 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
77 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
78 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
79
80 if XNN_LIKELY(batch >= (8 * sizeof(uint8_t))) {
81 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
82 output += 8;
83 batch -= 8 * sizeof(uint8_t);
84 } else {
85 if (batch & (4 * sizeof(uint8_t))) {
86 _mm_storeu_si32(output, vout0123456701234567);
87 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
88 output += 4;
89 }
90 if (batch & (2 * sizeof(uint8_t))) {
91 _mm_storeu_si16(output, vout0123456701234567);
92 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
93 output += 2;
94 }
95 if (batch & (1 * sizeof(uint8_t))) {
96 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
97 }
98 batch = 0;
99 }
100 } while (batch != 0);
101 }
102 }