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/intrinsics-polyfill.h>
15 #include <xnnpack/vadd.h>
16
17
xnn_qu8_vadd_minmax_ukernel__avx2_mul32_ld64_x16(size_t n,const uint8_t * input_a,const uint8_t * input_b,uint8_t * output,const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_vadd_minmax_ukernel__avx2_mul32_ld64_x16(
19 size_t n,
20 const uint8_t* input_a,
21 const uint8_t* input_b,
22 uint8_t* output,
23 const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 const __m256i vbias = _mm256_load_si256((const __m256i*) params->avx2.bias);
26 const __m256i va_multiplier = _mm256_load_si256((const __m256i*) params->avx2.a_multiplier);
27 const __m256i vb_multiplier = _mm256_load_si256((const __m256i*) params->avx2.b_multiplier);
28 const __m128i vshift = _mm_load_si128((const __m128i*) params->avx2.shift);
29 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->avx2.output_zero_point);
30 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->avx2.output_min);
31 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->avx2.output_max);
32
33 for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
34 const __m256i va01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) input_a));
35 const __m256i vb01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) input_b));
36 const __m256i va89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
37 const __m256i vb89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (input_b + 8)));
38 input_a += 16;
39 input_b += 16;
40
41 __m256i vacc01234567 = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va01234567, va_multiplier));
42 __m256i vacc89ABCDEF = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va89ABCDEF, va_multiplier));
43
44 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vb01234567, vb_multiplier));
45 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vb89ABCDEF, vb_multiplier));
46
47 vacc01234567 = _mm256_sra_epi32(vacc01234567, vshift);
48 vacc89ABCDEF = _mm256_sra_epi32(vacc89ABCDEF, vshift);
49
50 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
51
52 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packus_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
53
54 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
55
56 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, voutput_max);
57
58 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
59 output += 16;
60 }
61 if XNN_UNLIKELY(n != 0) {
62 do {
63 const __m256i va01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) input_a));
64 const __m256i vb01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) input_b));
65 input_a += 8;
66 input_b += 8;
67
68 __m256i vacc01234567 = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va01234567, va_multiplier));
69
70 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vb01234567, vb_multiplier));
71
72 vacc01234567 = _mm256_sra_epi32(vacc01234567, vshift);
73
74 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point));
75 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
76 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
77 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
78
79 if XNN_LIKELY(n >= (8 * sizeof(uint8_t))) {
80 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
81 output += 8;
82 n -= 8 * sizeof(uint8_t);
83 } else {
84 if (n & (4 * sizeof(uint8_t))) {
85 _mm_storeu_si32(output, vout0123456701234567);
86 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
87 output += 4;
88 }
89 if (n & (2 * sizeof(uint8_t))) {
90 _mm_storeu_si16(output, vout0123456701234567);
91 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
92 output += 2;
93 }
94 if (n & (1 * sizeof(uint8_t))) {
95 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
96 }
97 n = 0;
98 }
99 } while (n != 0);
100 }
101 }
102