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/vadd.h>
16
17
xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x8(size_t n,const int8_t * input_a,const int8_t * input_b,int8_t * output,const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x8(
19 size_t n,
20 const int8_t* input_a,
21 const int8_t* input_b,
22 int8_t* output,
23 const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 const __m256i va_multiplier = _mm256_load_si256((const __m256i*) params->avx2.a_multiplier);
26 const __m128i vshift = _mm_load_si128((const __m128i*) params->avx2.shift);
27 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->avx2.output_zero_point);
28 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->avx2.output_min);
29 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->avx2.output_max);
30
31 const __m256i vbias = _mm256_add_epi32(
32 _mm256_broadcastd_epi32(_mm_cvtsi32_si128(params->avx2.b_multiplier[0] * (int32_t) *input_b)),
33 _mm256_load_si256((const __m256i*) params->avx2.bias));
34 for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
35 const __m256i va01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_a));
36 input_a += 8;
37
38 __m256i vacc01234567 = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va01234567, va_multiplier));
39
40 vacc01234567 = _mm256_sra_epi32(vacc01234567, vshift);
41
42 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
43
44 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
45
46 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
47
48 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
49
50 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
51 output += 8;
52 }
53 if XNN_UNLIKELY(n != 0) {
54 {
55 const __m256i va01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_a));
56
57 __m256i vacc01234567 = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va01234567, va_multiplier));
58
59 vacc01234567 = _mm256_sra_epi32(vacc01234567, vshift);
60
61 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), voutput_zero_point);
62 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
63 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
64 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
65
66 if (n & (4 * sizeof(int8_t))) {
67 _mm_storeu_si32(output, vout0123456701234567);
68 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
69 output += 4;
70 }
71 if (n & (2 * sizeof(int8_t))) {
72 _mm_storeu_si16(output, vout0123456701234567);
73 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
74 output += 2;
75 }
76 if (n & (1 * sizeof(int8_t))) {
77 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
78 }
79 }
80 }
81 }
82