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_x16(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_vaddc_minmax_ukernel__avx2_mul32_ld64_x16(
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 vx_multiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.x_multiplier));
26   const __m256i vremainder_mask = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_mask));
27   const __m256i vremainder_threshold = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_threshold));
28   const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
29   const __m256i voutput_zero_point = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
30   const __m256i voutput_min = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_min));
31   const __m256i voutput_max = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_max));
32 
33   __m256i vzero_point_product = _mm256_broadcastsi128_si256(_mm_add_epi32(
34     _mm_broadcastd_epi32(_mm_cvtsi32_si128(params->sse2.y_multiplier[0] * (int32_t) *input_y)),
35     _mm_load_si128((const __m128i*) params->sse2.zero_point_product)));
36   for (; n >= 16 * sizeof(int8_t); n -= 16 * sizeof(int8_t)) {
37     const __m256i vx01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_x));
38     const __m256i vx89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_x + 8)));
39     input_x += 16;
40 
41     __m256i vacc01234567 = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vx01234567, vx_multiplier));
42     __m256i vacc89ABCDEF = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vx89ABCDEF, vx_multiplier));
43 
44     const __m256i vrem01234567 = _mm256_add_epi32(_mm256_and_si256(vacc01234567, vremainder_mask), _mm256_srai_epi32(vacc01234567, 31));
45     const __m256i vrem89ABCDEF = _mm256_add_epi32(_mm256_and_si256(vacc89ABCDEF, vremainder_mask), _mm256_srai_epi32(vacc89ABCDEF, 31));
46 
47     vacc01234567 = _mm256_sub_epi32(_mm256_sra_epi32(vacc01234567, vshift), _mm256_cmpgt_epi32(vrem01234567, vremainder_threshold));
48     vacc89ABCDEF = _mm256_sub_epi32(_mm256_sra_epi32(vacc89ABCDEF, vshift), _mm256_cmpgt_epi32(vrem89ABCDEF, vremainder_threshold));
49 
50     __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point);
51 
52     vout012389AB4567CDEF = _mm256_min_epi16(_mm256_max_epi16(vout012389AB4567CDEF, voutput_min), voutput_max);
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 
56     _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
57     output += 16;
58   }
59   if XNN_UNLIKELY(n != 0) {
60     do {
61       const __m256i vx01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_x));
62       input_x += 8;
63 
64       __m256i vacc01234567 = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vx01234567, vx_multiplier));
65 
66       const __m256i vrem01234567 = _mm256_add_epi32(_mm256_and_si256(vacc01234567, vremainder_mask), _mm256_srai_epi32(vacc01234567, 31));
67 
68       vacc01234567 = _mm256_sub_epi32(_mm256_sra_epi32(vacc01234567, vshift), _mm256_cmpgt_epi32(vrem01234567, vremainder_threshold));
69 
70       __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point));
71       vout01234567 = _mm_min_epi16(_mm_max_epi16(vout01234567, _mm256_castsi256_si128(voutput_min)), _mm256_castsi256_si128(voutput_max));
72       __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
73 
74       if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
75         _mm_storel_epi64((__m128i*) output, vout0123456701234567);
76         output += 8;
77         n -= 8 * sizeof(int8_t);
78       } else {
79         if (n & (4 * sizeof(int8_t))) {
80           *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
81           vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
82           output += 4;
83         }
84         if (n & (2 * sizeof(int8_t))) {
85           *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
86           vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
87           output += 2;
88         }
89         if (n & (1 * sizeof(int8_t))) {
90           *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
91         }
92         n = 0;
93       }
94     } while (n != 0);
95   }
96 }
97