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_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x8(size_t n,const int8_t * input_x,const int8_t * input_y,int8_t * output,const union xnn_qs8_add_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x8(
19 size_t n,
20 const int8_t* input_x,
21 const int8_t* input_y,
22 int8_t* output,
23 const union xnn_qs8_add_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
24 {
25 const __m256i vzero_point_product = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.zero_point_product));
26 const __m256i vx_multiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.x_multiplier));
27 const __m256i vy_multiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.y_multiplier));
28 const __m256i vremainder_mask = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_mask));
29 const __m256i vremainder_threshold = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_threshold));
30 const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
31 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
32 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
33 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
34
35 for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
36 const __m256i vx01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_x));
37 const __m256i vy01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_y));
38 input_x += 8;
39 input_y += 8;
40
41 __m256i vacc01234567 = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vx01234567, vx_multiplier));
42
43 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vy01234567, vy_multiplier));
44
45 const __m256i vrem01234567 = _mm256_add_epi32(_mm256_and_si256(vacc01234567, vremainder_mask), _mm256_srai_epi32(vacc01234567, 31));
46
47 vacc01234567 = _mm256_sub_epi32(_mm256_sra_epi32(vacc01234567, vshift), _mm256_cmpgt_epi32(vrem01234567, vremainder_threshold));
48
49 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
50
51 vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, voutput_min), voutput_max);
52
53 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
54
55 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
56 output += 8;
57 }
58 if XNN_UNLIKELY(n != 0) {
59 {
60 const __m256i vx01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_x));
61 const __m256i vy01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_y));
62
63 __m256i vacc01234567 = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vx01234567, vx_multiplier));
64
65 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vy01234567, vy_multiplier));
66
67 const __m256i vrem01234567 = _mm256_add_epi32(_mm256_and_si256(vacc01234567, vremainder_mask), _mm256_srai_epi32(vacc01234567, 31));
68
69 vacc01234567 = _mm256_sub_epi32(_mm256_sra_epi32(vacc01234567, vshift), _mm256_cmpgt_epi32(vrem01234567, vremainder_threshold));
70
71 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
72 vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, voutput_min), voutput_max);
73 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
74
75 if (n & (4 * sizeof(int8_t))) {
76 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
77 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
78 output += 4;
79 }
80 if (n & (2 * sizeof(int8_t))) {
81 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
82 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
83 output += 2;
84 }
85 if (n & (1 * sizeof(int8_t))) {
86 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
87 }
88 }
89 }
90 }
91