1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-vadd/sse-mul32-ld32.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 #ifdef __GNUC__
13   #include <x86intrin.h>
14 #else
15   #include <immintrin.h>
16   #include <ammintrin.h>
17 #endif
18 
19 #include <xnnpack/intrinsics-polyfill.h>
20 #include <xnnpack/vadd.h>
21 
22 
xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_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)])23 void xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x8(
24     size_t n,
25     const int8_t* input_x,
26     const int8_t* input_y,
27     int8_t* output,
28     const union xnn_qs8_add_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
29 {
30   const __m128i vzero_point_product = _mm_load_si128((const __m128i*) params->sse2.zero_point_product);
31   const __m128i vx_multiplier = _mm_load_si128((const __m128i*) params->sse2.x_multiplier);
32   const __m128i vy_multiplier = _mm_load_si128((const __m128i*) params->sse2.y_multiplier);
33   const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
34   const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
35   const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
36   const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
37   const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
38   const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
39 
40   for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
41     const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x));
42     const __m128i vy0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y));
43     const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4));
44     const __m128i vy4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 4));
45     input_x += 8;
46     input_y += 8;
47 
48     __m128i vacc0123 = _mm_macc_epi32(vx0123, vx_multiplier, vzero_point_product);
49     __m128i vacc4567 = _mm_macc_epi32(vx4567, vx_multiplier, vzero_point_product);
50 
51     vacc0123 = _mm_macc_epi32(vy0123, vy_multiplier, vacc0123);
52     vacc4567 = _mm_macc_epi32(vy4567, vy_multiplier, vacc4567);
53 
54     const __m128i vrem0123 = _mm_add_epi32(_mm_and_si128(vacc0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0123));
55     const __m128i vrem4567 = _mm_add_epi32(_mm_and_si128(vacc4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567));
56 
57     vacc0123 = _mm_sub_epi32(_mm_sra_epi32(vacc0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_threshold));
58     vacc4567 = _mm_sub_epi32(_mm_sra_epi32(vacc4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold));
59 
60     __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
61 
62     vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
63 
64     vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
65 
66     const __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
67 
68     _mm_storel_epi64((__m128i*) output, vout0123456701234567);
69     output += 8;
70   }
71   if XNN_UNLIKELY(n != 0) {
72     {
73       const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x));
74       const __m128i vy0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y));
75       const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4));
76       const __m128i vy4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 4));
77 
78       __m128i vacc0123 = _mm_macc_epi32(vx0123, vx_multiplier, vzero_point_product);
79       __m128i vacc4567 = _mm_macc_epi32(vx4567, vx_multiplier, vzero_point_product);
80 
81       vacc0123 = _mm_macc_epi32(vy0123, vy_multiplier, vacc0123);
82       vacc4567 = _mm_macc_epi32(vy4567, vy_multiplier, vacc4567);
83 
84       const __m128i vrem0123 = _mm_add_epi32(_mm_and_si128(vacc0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0123));
85       const __m128i vrem4567 = _mm_add_epi32(_mm_and_si128(vacc4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567));
86 
87       vacc0123 = _mm_sub_epi32(_mm_sra_epi32(vacc0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_threshold));
88       vacc4567 = _mm_sub_epi32(_mm_sra_epi32(vacc4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold));
89 
90       __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
91       vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
92       vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
93 
94       __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
95 
96       if (n & (4 * sizeof(int8_t))) {
97         *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
98         vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
99         output += 4;
100       }
101       if (n & (2 * sizeof(int8_t))) {
102         *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
103         vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
104         output += 2;
105       }
106       if (n & (1 * sizeof(int8_t))) {
107         *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
108       }
109     }
110   }
111 }
112