1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vadd/sse-mul16-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 <emmintrin.h>
13
14 #include <xnnpack/vadd.h>
15
16
xnn_qs8_vadd_minmax_ukernel__sse2_mul16_ld64_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)])17 void xnn_qs8_vadd_minmax_ukernel__sse2_mul16_ld64_x8(
18 size_t n,
19 const int8_t* input_x,
20 const int8_t* input_y,
21 int8_t* output,
22 const union xnn_qs8_add_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
23 {
24 const __m128i vzero_point_product = _mm_load_si128((const __m128i*) params->sse2.zero_point_product);
25 const __m128i vx_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.x_multiplier_lo);
26 const __m128i vx_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.x_multiplier_hi);
27 const __m128i vy_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.y_multiplier_lo);
28 const __m128i vy_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.y_multiplier_hi);
29 const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
30 const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
31 const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
32 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
33 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
34 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
35
36 for (; n >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
37 __m128i vx01234567 = _mm_loadl_epi64((const __m128i*) input_x);
38 __m128i vy01234567 = _mm_loadl_epi64((const __m128i*) input_y);
39 input_x += 8;
40 input_y += 8;
41
42 vx01234567 = _mm_unpacklo_epi8(vx01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vx01234567));
43 vy01234567 = _mm_unpacklo_epi8(vy01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vy01234567));
44
45 __m128i vxprod01234567hi = _mm_mulhi_epu16(vx01234567, vx_multiplier_lo);
46 __m128i vyprod01234567hi = _mm_mulhi_epu16(vy01234567, vy_multiplier_lo);
47 const __m128i vxprod01234567lo = _mm_mullo_epi16(vx01234567, vx_multiplier_lo);
48 const __m128i vyprod01234567lo = _mm_mullo_epi16(vy01234567, vy_multiplier_lo);
49
50 vxprod01234567hi = _mm_add_epi16(vxprod01234567hi, _mm_mullo_epi16(vx01234567, vx_multiplier_hi));
51 vyprod01234567hi = _mm_add_epi16(vyprod01234567hi, _mm_mullo_epi16(vy01234567, vy_multiplier_hi));
52
53 vxprod01234567hi = _mm_sub_epi16(vxprod01234567hi, _mm_and_si128(_mm_srai_epi16(vx01234567, 15), vx_multiplier_lo));
54 vyprod01234567hi = _mm_sub_epi16(vyprod01234567hi, _mm_and_si128(_mm_srai_epi16(vy01234567, 15), vy_multiplier_lo));
55
56 __m128i vacc0123 = _mm_add_epi32(vzero_point_product, _mm_unpacklo_epi16(vxprod01234567lo, vxprod01234567hi));
57 __m128i vacc4567 = _mm_add_epi32(vzero_point_product, _mm_unpackhi_epi16(vxprod01234567lo, vxprod01234567hi));
58
59 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vyprod01234567lo, vyprod01234567hi));
60 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vyprod01234567lo, vyprod01234567hi));
61
62 const __m128i vrem0123 = _mm_add_epi32(_mm_and_si128(vacc0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0123));
63 const __m128i vrem4567 = _mm_add_epi32(_mm_and_si128(vacc4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567));
64
65 vacc0123 = _mm_sub_epi32(_mm_sra_epi32(vacc0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_threshold));
66 vacc4567 = _mm_sub_epi32(_mm_sra_epi32(vacc4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold));
67
68 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
69
70 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
71
72 vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
73
74 const __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
75
76 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
77 output += 8;
78 }
79 if XNN_UNLIKELY(n != 0) {
80 {
81 __m128i vx01234567 = _mm_loadl_epi64((const __m128i*) input_x);
82 __m128i vy01234567 = _mm_loadl_epi64((const __m128i*) input_y);
83
84 vx01234567 = _mm_unpacklo_epi8(vx01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vx01234567));
85 vy01234567 = _mm_unpacklo_epi8(vy01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vy01234567));
86
87 __m128i vxprod01234567hi = _mm_mulhi_epu16(vx01234567, vx_multiplier_lo);
88 __m128i vyprod01234567hi = _mm_mulhi_epu16(vy01234567, vy_multiplier_lo);
89 const __m128i vxprod01234567lo = _mm_mullo_epi16(vx01234567, vx_multiplier_lo);
90 const __m128i vyprod01234567lo = _mm_mullo_epi16(vy01234567, vy_multiplier_lo);
91
92 vxprod01234567hi = _mm_add_epi16(vxprod01234567hi, _mm_mullo_epi16(vx01234567, vx_multiplier_hi));
93 vyprod01234567hi = _mm_add_epi16(vyprod01234567hi, _mm_mullo_epi16(vy01234567, vy_multiplier_hi));
94
95 vxprod01234567hi = _mm_sub_epi16(vxprod01234567hi, _mm_and_si128(_mm_srai_epi16(vx01234567, 15), vx_multiplier_lo));
96 vyprod01234567hi = _mm_sub_epi16(vyprod01234567hi, _mm_and_si128(_mm_srai_epi16(vy01234567, 15), vy_multiplier_lo));
97
98 __m128i vacc0123 = _mm_add_epi32(vzero_point_product, _mm_unpacklo_epi16(vxprod01234567lo, vxprod01234567hi));
99 __m128i vacc4567 = _mm_add_epi32(vzero_point_product, _mm_unpackhi_epi16(vxprod01234567lo, vxprod01234567hi));
100
101 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vyprod01234567lo, vyprod01234567hi));
102 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vyprod01234567lo, vyprod01234567hi));
103
104 const __m128i vrem0123 = _mm_add_epi32(_mm_and_si128(vacc0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0123));
105 const __m128i vrem4567 = _mm_add_epi32(_mm_and_si128(vacc4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567));
106
107 vacc0123 = _mm_sub_epi32(_mm_sra_epi32(vacc0123, vshift), _mm_cmpgt_epi32(vrem0123, vremainder_threshold));
108 vacc4567 = _mm_sub_epi32(_mm_sra_epi32(vacc4567, vshift), _mm_cmpgt_epi32(vrem4567, vremainder_threshold));
109
110 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
111 vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
112 vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
113
114 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
115
116 if (n & (4 * sizeof(int8_t))) {
117 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
118 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
119 output += 4;
120 }
121 if (n & (2 * sizeof(int8_t))) {
122 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
123 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
124 output += 2;
125 }
126 if (n & (1 * sizeof(int8_t))) {
127 *output = (int32_t) _mm_cvtsi128_si32(vout0123456701234567);
128 }
129 }
130 }
131 }
132