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_x24(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_vadd_minmax_ukernel__avx2_mul32_ld64_x24(
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 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 >= 24 * sizeof(int8_t); n -= 24 * sizeof(int8_t)) {
34 const __m256i va01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_a));
35 const __m256i vb01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_b));
36 const __m256i va89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
37 const __m256i vb89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_b + 8)));
38 const __m256i vaGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_a + 16)));
39 const __m256i vbGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_b + 16)));
40 input_a += 24;
41 input_b += 24;
42
43 __m256i vacc01234567 = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va01234567, va_multiplier));
44 __m256i vacc89ABCDEF = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va89ABCDEF, va_multiplier));
45 __m256i vaccGHIJKLMN = _mm256_add_epi32(vbias, _mm256_mullo_epi32(vaGHIJKLMN, va_multiplier));
46
47 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vb01234567, vb_multiplier));
48 vacc89ABCDEF = _mm256_add_epi32(vacc89ABCDEF, _mm256_mullo_epi32(vb89ABCDEF, vb_multiplier));
49 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vbGHIJKLMN, vb_multiplier));
50
51 vacc01234567 = _mm256_sra_epi32(vacc01234567, vshift);
52 vacc89ABCDEF = _mm256_sra_epi32(vacc89ABCDEF, vshift);
53 vaccGHIJKLMN = _mm256_sra_epi32(vaccGHIJKLMN, vshift);
54
55 __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
56 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vaccGHIJKLMN), _mm256_extracti128_si256(vaccGHIJKLMN, 1)), _mm256_castsi256_si128(voutput_zero_point));
57
58 __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0));
59 __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
60
61 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
62 voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
63
64 vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
65 voutGHIJKLMNGHIJKLMN = _mm_min_epi8(voutGHIJKLMNGHIJKLMN, voutput_max);
66
67 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
68 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
69 output += 24;
70 }
71 if XNN_UNLIKELY(n != 0) {
72 do {
73 const __m256i va01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_a));
74 const __m256i vb01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_b));
75 input_a += 8;
76 input_b += 8;
77
78 __m256i vacc01234567 = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va01234567, va_multiplier));
79
80 vacc01234567 = _mm256_add_epi32(vacc01234567, _mm256_mullo_epi32(vb01234567, vb_multiplier));
81
82 vacc01234567 = _mm256_sra_epi32(vacc01234567, vshift);
83
84 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point));
85 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
86 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
87 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
88
89 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
90 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
91 output += 8;
92 n -= 8 * sizeof(int8_t);
93 } else {
94 if (n & (4 * sizeof(int8_t))) {
95 _mm_storeu_si32(output, vout0123456701234567);
96 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
97 output += 4;
98 }
99 if (n & (2 * sizeof(int8_t))) {
100 _mm_storeu_si16(output, vout0123456701234567);
101 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
102 output += 2;
103 }
104 if (n & (1 * sizeof(int8_t))) {
105 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
106 }
107 n = 0;
108 }
109 } while (n != 0);
110 }
111 }
112