• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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/vaddsub.h>
15 
16 
xnn_qs8_vadd_minmax_ukernel__sse2_mul16_ld64_x32(size_t n,const int8_t * input_a,const int8_t * input_b,int8_t * output,const union xnn_qs8_addsub_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_vadd_minmax_ukernel__sse2_mul16_ld64_x32(
18     size_t n,
19     const int8_t* input_a,
20     const int8_t* input_b,
21     int8_t* output,
22     const union xnn_qs8_addsub_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24   const __m128i vbias = _mm_load_si128((const __m128i*) params->sse2.bias);
25   const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
26   const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
27   const __m128i vb_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.b_multiplier_lo);
28   const __m128i vb_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.b_multiplier_hi);
29   const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
30   const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
31   const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
32   const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
33 
34   for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
35     __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
36     __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b);
37     __m128i va89ABCDEF = _mm_loadl_epi64((const __m128i*) (input_a + 8));
38     __m128i vb89ABCDEF = _mm_loadl_epi64((const __m128i*) (input_b + 8));
39     __m128i vaGHIJKLMN = _mm_loadl_epi64((const __m128i*) (input_a + 16));
40     __m128i vbGHIJKLMN = _mm_loadl_epi64((const __m128i*) (input_b + 16));
41     __m128i vaOPQRSTUV = _mm_loadl_epi64((const __m128i*) (input_a + 24));
42     __m128i vbOPQRSTUV = _mm_loadl_epi64((const __m128i*) (input_b + 24));
43     input_a += 32;
44     input_b += 32;
45 
46     va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8);
47     vb01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vb01234567, vb01234567), 8);
48     va89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(va89ABCDEF, va89ABCDEF), 8);
49     vb89ABCDEF = _mm_srai_epi16(_mm_unpacklo_epi8(vb89ABCDEF, vb89ABCDEF), 8);
50     vaGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vaGHIJKLMN, vaGHIJKLMN), 8);
51     vbGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vbGHIJKLMN, vbGHIJKLMN), 8);
52     vaOPQRSTUV = _mm_srai_epi16(_mm_unpacklo_epi8(vaOPQRSTUV, vaOPQRSTUV), 8);
53     vbOPQRSTUV = _mm_srai_epi16(_mm_unpacklo_epi8(vbOPQRSTUV, vbOPQRSTUV), 8);
54 
55     __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
56     __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
57     const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
58     const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
59     __m128i vaprod89ABCDEFhi = _mm_mulhi_epu16(va89ABCDEF, va_multiplier_lo);
60     __m128i vbprod89ABCDEFhi = _mm_mulhi_epu16(vb89ABCDEF, vb_multiplier_lo);
61     const __m128i vaprod89ABCDEFlo = _mm_mullo_epi16(va89ABCDEF, va_multiplier_lo);
62     const __m128i vbprod89ABCDEFlo = _mm_mullo_epi16(vb89ABCDEF, vb_multiplier_lo);
63     __m128i vaprodGHIJKLMNhi = _mm_mulhi_epu16(vaGHIJKLMN, va_multiplier_lo);
64     __m128i vbprodGHIJKLMNhi = _mm_mulhi_epu16(vbGHIJKLMN, vb_multiplier_lo);
65     const __m128i vaprodGHIJKLMNlo = _mm_mullo_epi16(vaGHIJKLMN, va_multiplier_lo);
66     const __m128i vbprodGHIJKLMNlo = _mm_mullo_epi16(vbGHIJKLMN, vb_multiplier_lo);
67     __m128i vaprodOPQRSTUVhi = _mm_mulhi_epu16(vaOPQRSTUV, va_multiplier_lo);
68     __m128i vbprodOPQRSTUVhi = _mm_mulhi_epu16(vbOPQRSTUV, vb_multiplier_lo);
69     const __m128i vaprodOPQRSTUVlo = _mm_mullo_epi16(vaOPQRSTUV, va_multiplier_lo);
70     const __m128i vbprodOPQRSTUVlo = _mm_mullo_epi16(vbOPQRSTUV, vb_multiplier_lo);
71 
72     vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
73     vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
74     vaprod89ABCDEFhi = _mm_add_epi16(vaprod89ABCDEFhi, _mm_mullo_epi16(va89ABCDEF, va_multiplier_hi));
75     vbprod89ABCDEFhi = _mm_add_epi16(vbprod89ABCDEFhi, _mm_mullo_epi16(vb89ABCDEF, vb_multiplier_hi));
76     vaprodGHIJKLMNhi = _mm_add_epi16(vaprodGHIJKLMNhi, _mm_mullo_epi16(vaGHIJKLMN, va_multiplier_hi));
77     vbprodGHIJKLMNhi = _mm_add_epi16(vbprodGHIJKLMNhi, _mm_mullo_epi16(vbGHIJKLMN, vb_multiplier_hi));
78     vaprodOPQRSTUVhi = _mm_add_epi16(vaprodOPQRSTUVhi, _mm_mullo_epi16(vaOPQRSTUV, va_multiplier_hi));
79     vbprodOPQRSTUVhi = _mm_add_epi16(vbprodOPQRSTUVhi, _mm_mullo_epi16(vbOPQRSTUV, vb_multiplier_hi));
80 
81     vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
82     vbprod01234567hi = _mm_sub_epi16(vbprod01234567hi, _mm_and_si128(_mm_srai_epi16(vb01234567, 15), vb_multiplier_lo));
83     vaprod89ABCDEFhi = _mm_sub_epi16(vaprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(va89ABCDEF, 15), va_multiplier_lo));
84     vbprod89ABCDEFhi = _mm_sub_epi16(vbprod89ABCDEFhi, _mm_and_si128(_mm_srai_epi16(vb89ABCDEF, 15), vb_multiplier_lo));
85     vaprodGHIJKLMNhi = _mm_sub_epi16(vaprodGHIJKLMNhi, _mm_and_si128(_mm_srai_epi16(vaGHIJKLMN, 15), va_multiplier_lo));
86     vbprodGHIJKLMNhi = _mm_sub_epi16(vbprodGHIJKLMNhi, _mm_and_si128(_mm_srai_epi16(vbGHIJKLMN, 15), vb_multiplier_lo));
87     vaprodOPQRSTUVhi = _mm_sub_epi16(vaprodOPQRSTUVhi, _mm_and_si128(_mm_srai_epi16(vaOPQRSTUV, 15), va_multiplier_lo));
88     vbprodOPQRSTUVhi = _mm_sub_epi16(vbprodOPQRSTUVhi, _mm_and_si128(_mm_srai_epi16(vbOPQRSTUV, 15), vb_multiplier_lo));
89 
90     __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
91     __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
92     __m128i vacc89AB = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
93     __m128i vaccCDEF = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod89ABCDEFlo, vaprod89ABCDEFhi));
94     __m128i vaccGHIJ = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprodGHIJKLMNlo, vaprodGHIJKLMNhi));
95     __m128i vaccKLMN = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprodGHIJKLMNlo, vaprodGHIJKLMNhi));
96     __m128i vaccOPQR = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprodOPQRSTUVlo, vaprodOPQRSTUVhi));
97     __m128i vaccSTUV = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprodOPQRSTUVlo, vaprodOPQRSTUVhi));
98 
99     vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
100     vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
101     vacc89AB = _mm_add_epi32(vacc89AB, _mm_unpacklo_epi16(vbprod89ABCDEFlo, vbprod89ABCDEFhi));
102     vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_unpackhi_epi16(vbprod89ABCDEFlo, vbprod89ABCDEFhi));
103     vaccGHIJ = _mm_add_epi32(vaccGHIJ, _mm_unpacklo_epi16(vbprodGHIJKLMNlo, vbprodGHIJKLMNhi));
104     vaccKLMN = _mm_add_epi32(vaccKLMN, _mm_unpackhi_epi16(vbprodGHIJKLMNlo, vbprodGHIJKLMNhi));
105     vaccOPQR = _mm_add_epi32(vaccOPQR, _mm_unpacklo_epi16(vbprodOPQRSTUVlo, vbprodOPQRSTUVhi));
106     vaccSTUV = _mm_add_epi32(vaccSTUV, _mm_unpackhi_epi16(vbprodOPQRSTUVlo, vbprodOPQRSTUVhi));
107 
108     vacc0123 = _mm_sra_epi32(vacc0123, vshift);
109     vacc4567 = _mm_sra_epi32(vacc4567, vshift);
110     vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
111     vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
112     vaccGHIJ = _mm_sra_epi32(vaccGHIJ, vshift);
113     vaccKLMN = _mm_sra_epi32(vaccKLMN, vshift);
114     vaccOPQR = _mm_sra_epi32(vaccOPQR, vshift);
115     vaccSTUV = _mm_sra_epi32(vaccSTUV, vshift);
116 
117     __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
118     __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
119     __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
120     __m128i voutOPQRSTUV = _mm_adds_epi16(_mm_packs_epi32(vaccOPQR, vaccSTUV), voutput_zero_point);
121 
122     vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
123     vout89ABCDEF = _mm_max_epi16(vout89ABCDEF, voutput_min);
124     voutGHIJKLMN = _mm_max_epi16(voutGHIJKLMN, voutput_min);
125     voutOPQRSTUV = _mm_max_epi16(voutOPQRSTUV, voutput_min);
126 
127     vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
128     vout89ABCDEF = _mm_min_epi16(vout89ABCDEF, voutput_max);
129     voutGHIJKLMN = _mm_min_epi16(voutGHIJKLMN, voutput_max);
130     voutOPQRSTUV = _mm_min_epi16(voutOPQRSTUV, voutput_max);
131 
132     __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
133     __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
134 
135 
136     _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
137     _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
138     output += 32;
139   }
140   if XNN_UNLIKELY(n != 0) {
141     do {
142       __m128i va01234567 = _mm_loadl_epi64((const __m128i*) input_a);
143       __m128i vb01234567 = _mm_loadl_epi64((const __m128i*) input_b);
144       input_a += 8;
145       input_b += 8;
146 
147       va01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(va01234567, va01234567), 8);
148       vb01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vb01234567, vb01234567), 8);
149 
150       __m128i vaprod01234567hi = _mm_mulhi_epu16(va01234567, va_multiplier_lo);
151       __m128i vbprod01234567hi = _mm_mulhi_epu16(vb01234567, vb_multiplier_lo);
152       const __m128i vaprod01234567lo = _mm_mullo_epi16(va01234567, va_multiplier_lo);
153       const __m128i vbprod01234567lo = _mm_mullo_epi16(vb01234567, vb_multiplier_lo);
154 
155       vaprod01234567hi = _mm_add_epi16(vaprod01234567hi, _mm_mullo_epi16(va01234567, va_multiplier_hi));
156       vbprod01234567hi = _mm_add_epi16(vbprod01234567hi, _mm_mullo_epi16(vb01234567, vb_multiplier_hi));
157 
158       vaprod01234567hi = _mm_sub_epi16(vaprod01234567hi, _mm_and_si128(_mm_srai_epi16(va01234567, 15), va_multiplier_lo));
159       vbprod01234567hi = _mm_sub_epi16(vbprod01234567hi, _mm_and_si128(_mm_srai_epi16(vb01234567, 15), vb_multiplier_lo));
160 
161       __m128i vacc0123 = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod01234567lo, vaprod01234567hi));
162       __m128i vacc4567 = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod01234567lo, vaprod01234567hi));
163 
164       vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vbprod01234567lo, vbprod01234567hi));
165       vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vbprod01234567lo, vbprod01234567hi));
166 
167       vacc0123 = _mm_sra_epi32(vacc0123, vshift);
168       vacc4567 = _mm_sra_epi32(vacc4567, vshift);
169 
170       __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
171       vout01234567 = _mm_max_epi16(vout01234567, voutput_min);
172       vout01234567 = _mm_min_epi16(vout01234567, voutput_max);
173 
174       __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
175 
176       if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
177         _mm_storel_epi64((__m128i*) output, vout0123456701234567);
178         output += 8;
179         n -= 8 * sizeof(int8_t);
180       } else {
181         if (n & (4 * sizeof(int8_t))) {
182           *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
183           vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
184           output += 4;
185         }
186         if (n & (2 * sizeof(int8_t))) {
187           *((uint16_t*) output) = (uint16_t) _mm_cvtsi128_si32(vout0123456701234567);
188           vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
189           output += 2;
190         }
191         if (n & (1 * sizeof(int8_t))) {
192           *output = (int32_t) _mm_cvtsi128_si32(vout0123456701234567);
193         }
194         n = 0;
195       }
196     } while (n != 0);
197   }
198 }
199