Debian Codegraph @

src/qu8-vadd/gen/qu8-vadd-minmax-avx2-mul32-ld64-u16.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vadd/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/common.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/vbinary.h"
17
18
19 void xnn_qu8_vadd_minmax_ukernel__avx2_mul32_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 __m256i vbias = _mm256_set1_epi32(params->scalar.bias);
33 const __m256i va_multiplier = _mm256_set1_epi32(params->scalar.a_multiplier);
34 const __m256i vb_multiplier = _mm256_set1_epi32(params->scalar.b_multiplier);
35 const __m128i vshift = _mm_set1_epi64x(params->scalar.shift);
36 const __m256i voutput_zero_point = _mm256_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);
42 XNN_FORCE_REALIZATION(vb_multiplier);
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 __m256i va01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) input_a));
50 const __m256i vb01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) input_b));
51 const __m256i va89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
52 const __m256i vb89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (input_b + 8)));
53 input_a += 16;
54 input_b += 16;
55
56 __m256i vacc01234567 = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va01234567, va_multiplier));
57 __m256i vacc89ABCDEF = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va89ABCDEF, va_multiplier));
58
59 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vb01234567, vb_multiplier));
60 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vb89ABCDEF, vb_multiplier));
61
62 vacc01234567 = _mm256_sra_epi32(vacc01234567, vshift);
63 vacc89ABCDEF = _mm256_sra_epi32(vacc89ABCDEF, vshift);
64
65 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
66
67 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packus_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
68
69 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
70
71 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, voutput_max);
72
73 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
74 output += 16;
75 }
76 if XNN_UNLIKELY(batch != 0) {
77 do {
78 const __m256i va01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) input_a));
79 const __m256i vb01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) input_b));
80 input_a += 8;
81 input_b += 8;
82
83 __m256i vacc01234567 = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va01234567, va_multiplier));
84
85 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vb01234567, vb_multiplier));
86
87 vacc01234567 = _mm256_sra_epi32(vacc01234567, vshift);
88
89 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point));
90 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
91 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
92 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
93
94 if XNN_LIKELY(batch >= (8 * sizeof(uint8_t))) {
95 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
96 output += 8;
97 batch -= 8 * sizeof(uint8_t);
98 } else {
99 if (batch & (4 * sizeof(uint8_t))) {
100 _mm_storeu_si32(output, vout0123456701234567);
101 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
102 output += 4;
103 }
104 if (batch & (2 * sizeof(uint8_t))) {
105 _mm_storeu_si16(output, vout0123456701234567);
106 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
107 output += 2;
108 }
109 if (batch & (1 * sizeof(uint8_t))) {
110 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
111 }
112 batch = 0;
113 }
114 } while (batch != 0);
115 }
116 }