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 <smmintrin.h>
13
14 #include <xnnpack/unaligned.h>
15 #include <xnnpack/vadd.h>
16
17
xnn_qs8_vadd_minmax_ukernel__sse41_mul16_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__sse41_mul16_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 __m128i vbias = _mm_load_si128((const __m128i*) params->sse4_mul16.bias);
26 const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_lo);
27 const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse4_mul16.a_multiplier_hi);
28 const __m128i vb_multiplier_lo = _mm_load_si128((const __m128i*) params->sse4_mul16.b_multiplier_lo);
29 const __m128i vb_multiplier_hi = _mm_load_si128((const __m128i*) params->sse4_mul16.b_multiplier_hi);
30 const __m128i vshift = _mm_cvtsi32_si128((int) params->sse4_mul16.shift);
31 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul16.output_zero_point);
32 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul16.output_min);
33 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul16.output_max);
34
35 for (; n >= 24 * sizeof(int8_t); n -= 24 * sizeof(int8_t)) {
36 const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
37 const __m128i vb01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_b));
38 const __m128i va89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
39 const __m128i vb89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_b + 8)));
40 const __m128i vaGHIJKLMN = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 16)));
41 const __m128i vbGHIJKLMN = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_b + 16)));
42 input_a += 24;
43 input_b += 24;
44
45
46 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
47 __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
48 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
49 const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
50 __m128i vaprod89ABCDEFhi = _mm_mulhi_epu16(va89ABCDEF, va_multiplier_lo);
51 __m128i vbprod89ABCDEFhi = _mm_mulhi_epu16(vb89ABCDEF, vb_multiplier_lo);
52 const __m128i vaprod89ABCDEFlo = _mm_mullo_epi16(va89ABCDEF, va_multiplier_lo);
53 const __m128i vbprod89ABCDEFlo = _mm_mullo_epi16(vb89ABCDEF, vb_multiplier_lo);
54 __m128i vaprodGHIJKLMNhi = _mm_mulhi_epu16(vaGHIJKLMN, va_multiplier_lo);
55 __m128i vbprodGHIJKLMNhi = _mm_mulhi_epu16(vbGHIJKLMN, vb_multiplier_lo);
56 const __m128i vaprodGHIJKLMNlo = _mm_mullo_epi16(vaGHIJKLMN, va_multiplier_lo);
57 const __m128i vbprodGHIJKLMNlo = _mm_mullo_epi16(vbGHIJKLMN, vb_multiplier_lo);
58
59 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
60 vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
61 vaprod89ABCDEFhi = _mm_add_epi16(vaprod89ABCDEFhi, _mm_mullo_epi16(va89ABCDEF, va_multiplier_hi));
62 vbprod89ABCDEFhi = _mm_add_epi16(vbprod89ABCDEFhi, _mm_mullo_epi16(vb89ABCDEF, vb_multiplier_hi));
63 vaprodGHIJKLMNhi = _mm_add_epi16(vaprodGHIJKLMNhi, _mm_mullo_epi16(vaGHIJKLMN, va_multiplier_hi));
64 vbprodGHIJKLMNhi = _mm_add_epi16(vbprodGHIJKLMNhi, _mm_mullo_epi16(vbGHIJKLMN, vb_multiplier_hi));
65
66 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
67 vbprod01234567hi = _mm_sub_epi16(vbprod01234567hi, _mm_and_si128(_mm_srai_epi16(vb01234567, 15), vb_multiplier_lo));
68 vaprod89ABCDEFhi = _mm_sub_epi16(vaprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(va89ABCDEF, 15), va_multiplier_lo));
69 vbprod89ABCDEFhi = _mm_sub_epi16(vbprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(vb89ABCDEF, 15), vb_multiplier_lo));
70 vaprodGHIJKLMNhi = _mm_sub_epi16(vaprodGHIJKLMNhi, _mm_and_si128(_mm_srai_epi16(vaGHIJKLMN, 15), va_multiplier_lo));
71 vbprodGHIJKLMNhi = _mm_sub_epi16(vbprodGHIJKLMNhi, _mm_and_si128(_mm_srai_epi16(vbGHIJKLMN, 15), vb_multiplier_lo));
72
73 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
74 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
75 __m128i vacc89AB = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
76 __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
77 __m128i vaccGHIJ = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprodGHIJKLMNlo, vaprodGHIJKLMNhi));
78 __m128i vaccKLMN = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprodGHIJKLMNlo, vaprodGHIJKLMNhi));
79
80 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
81 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
82 vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vbprod89ABCDEFlo, vbprod89ABCDEFhi));
83 vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vbprod89ABCDEFlo, vbprod89ABCDEFhi));
84 vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vbprodGHIJKLMNlo, vbprodGHIJKLMNhi));
85 vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vbprodGHIJKLMNlo, vbprodGHIJKLMNhi));
86
87 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
88 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
89 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
90 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
91 vaccGHIJ = _mm_sra_epi32(vaccGHIJ, vshift);
92 vaccKLMN = _mm_sra_epi32(vaccKLMN, vshift);
93
94 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
95 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
96 __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
97
98
99 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
100 __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN);
101
102 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
103 voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min);
104
105 vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
106 voutGHIJKLMNGHIJKLMN = _mm_min_epi8(voutGHIJKLMNGHIJKLMN, voutput_max);
107
108 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
109 _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN);
110 output += 24;
111 }
112 if XNN_UNLIKELY(n != 0) {
113 do {
114 const __m128i va01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
115 const __m128i vb01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_b));
116 input_a += 8;
117 input_b += 8;
118
119
120 __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
121 __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
122 const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
123 const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
124
125 vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
126 vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
127
128 vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
129 vbprod01234567hi = _mm_sub_epi16(vbprod01234567hi, _mm_and_si128(_mm_srai_epi16(vb01234567, 15), vb_multiplier_lo));
130
131 __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
132 __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
133
134 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
135 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
136
137 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
138 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
139
140 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
141
142 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
143 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
144 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
145
146 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
147 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
148 output += 8;
149 n -= 8 * sizeof(int8_t);
150 } else {
151 if (n & (4 * sizeof(int8_t))) {
152 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
153 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
154 output += 4;
155 }
156 if (n & (2 * sizeof(int8_t))) {
157 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
158 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
159 output += 2;
160 }
161 if (n & (1 * sizeof(int8_t))) {
162 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
163 }
164 n = 0;
165 }
166 } while (n != 0);
167 }
168 }
169