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_x32(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_x32(
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 __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) 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 >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
35 const __m256i va01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_a));
36 const __m256i va89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
37 const __m256i vaGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_a + 16)));
38 const __m256i vaOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_a + 24)));
39 input_a += 32;
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 __m256i vaccGHIJKLMN = _mm256_add_epi32(vbias, _mm256_mullo_epi32(vaGHIJKLMN, va_multiplier));
44 __m256i vaccOPQRSTUV = _mm256_add_epi32(vbias, _mm256_mullo_epi32(vaOPQRSTUV, va_multiplier));
45
46 vacc01234567 = _mm256_sra_epi32(vacc01234567, vshift);
47 vacc89ABCDEF = _mm256_sra_epi32(vacc89ABCDEF, vshift);
48 vaccGHIJKLMN = _mm256_sra_epi32(vaccGHIJKLMN, vshift);
49 vaccOPQRSTUV = _mm256_sra_epi32(vaccOPQRSTUV, vshift);
50
51 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
52 __m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(vaccGHIJKLMN, vaccOPQRSTUV), voutput_zero_point);
53
54 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
55 __m128i voutGHIJKLMNOPQRSTUV = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(voutGHIJOPQRKLMNSTUV), _mm256_extracti128_si256(voutGHIJOPQRKLMNSTUV, 1)), _MM_SHUFFLE(3, 1, 2, 0));
56
57 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
58 voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
59
60 vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
61 voutGHIJKLMNOPQRSTUV = _mm_min_epi8(voutGHIJKLMNOPQRSTUV, voutput_max);
62
63 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
64 _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
65 output += 32;
66 }
67 if XNN_UNLIKELY(n != 0) {
68 do {
69 const __m256i va01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_a));
70 input_a += 8;
71
72 __m256i vacc01234567 = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va01234567, va_multiplier));
73
74 vacc01234567 = _mm256_sra_epi32(vacc01234567, vshift);
75
76 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point));
77 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
78 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
79 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
80
81 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
82 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
83 output += 8;
84 n -= 8 * sizeof(int8_t);
85 } else {
86 if (n & (4 * sizeof(int8_t))) {
87 _mm_storeu_si32(output, vout0123456701234567);
88 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
89 output += 4;
90 }
91 if (n & (2 * sizeof(int8_t))) {
92 _mm_storeu_si16(output, vout0123456701234567);
93 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
94 output += 2;
95 }
96 if (n & (1 * sizeof(int8_t))) {
97 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
98 }
99 n = 0;
100 }
101 } while (n != 0);
102 }
103 }
104