1 // Auto-generated file. Do not edit!
2 // Template: src/x8-lut/avx2.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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/lut.h>
16 #include <xnnpack/common.h>
17
18
xnn_x8_lut_ukernel__avx2_x96(size_t n,const uint8_t * x,uint8_t * y,const uint8_t t[restrict XNN_MIN_ELEMENTS (256)])19 void xnn_x8_lut_ukernel__avx2_x96(
20 size_t n,
21 const uint8_t* x,
22 uint8_t* y,
23 const uint8_t t[restrict XNN_MIN_ELEMENTS(256)])
24 {
25 assert(n != 0);
26 assert(x != NULL);
27 assert(y != NULL);
28
29 const __m256i vt0 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) t));
30 const __m256i vt1 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 16)));
31 const __m256i vt2 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 32)));
32 const __m256i vt3 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 48)));
33 const __m256i vt4 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 64)));
34 const __m256i vt5 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 80)));
35 const __m256i vt6 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 96)));
36 const __m256i vt7 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 112)));
37 const __m256i vt8 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 128)));
38 const __m256i vt9 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 144)));
39 const __m256i vtA = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 160)));
40 const __m256i vtB = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 176)));
41 const __m256i vtC = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 192)));
42 const __m256i vtD = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 208)));
43 const __m256i vtE = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 224)));
44 const __m256i vtF = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 240)));
45
46 const __m256i vtable0 = vt0;
47 const __m256i vtable1 = _mm256_xor_si256(vt0, vt1);
48 const __m256i vtable2 = _mm256_xor_si256(vt1, vt2);
49 const __m256i vtable3 = _mm256_xor_si256(vt2, vt3);
50 const __m256i vtable4 = _mm256_xor_si256(vt3, vt4);
51 const __m256i vtable5 = _mm256_xor_si256(vt4, vt5);
52 const __m256i vtable6 = _mm256_xor_si256(vt5, vt6);
53 const __m256i vtable7 = _mm256_xor_si256(vt6, vt7);
54 const __m256i vtable8 = _mm256_xor_si256(_mm256_xor_si256(vt7, vt8), vtable0);
55 const __m256i vtable9 = _mm256_xor_si256(_mm256_xor_si256(vt8, vt9), vtable1);
56 const __m256i vtableA = _mm256_xor_si256(_mm256_xor_si256(vt9, vtA), vtable2);
57 const __m256i vtableB = _mm256_xor_si256(_mm256_xor_si256(vtA, vtB), vtable3);
58 const __m256i vtableC = _mm256_xor_si256(_mm256_xor_si256(vtB, vtC), vtable4);
59 const __m256i vtableD = _mm256_xor_si256(_mm256_xor_si256(vtC, vtD), vtable5);
60 const __m256i vtableE = _mm256_xor_si256(_mm256_xor_si256(vtD, vtE), vtable6);
61 const __m256i vtableF = _mm256_xor_si256(_mm256_xor_si256(vtE, vtF), vtable7);
62
63 const __m256i voffset = _mm256_set1_epi8(16);
64 for (; n >= 96 * sizeof(uint8_t); n -= 96 * sizeof(uint8_t)) {
65 __m256i vx0 = _mm256_loadu_si256((const __m256i*) x);
66 __m256i vx1 = _mm256_loadu_si256((const __m256i*) (x + 32));
67 __m256i vx2 = _mm256_loadu_si256((const __m256i*) (x + 64));
68 x += 96;
69
70 __m256i vy0 = _mm256_shuffle_epi8(vtable0, vx0);
71 __m256i vy1 = _mm256_shuffle_epi8(vtable0, vx1);
72 __m256i vy2 = _mm256_shuffle_epi8(vtable0, vx2);
73
74 vx0 = _mm256_sub_epi8(vx0, voffset);
75 vx1 = _mm256_sub_epi8(vx1, voffset);
76 vx2 = _mm256_sub_epi8(vx2, voffset);
77 vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtable1, vx0));
78 vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtable1, vx1));
79 vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtable1, vx2));
80 vx0 = _mm256_sub_epi8(vx0, voffset);
81 vx1 = _mm256_sub_epi8(vx1, voffset);
82 vx2 = _mm256_sub_epi8(vx2, voffset);
83 vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtable2, vx0));
84 vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtable2, vx1));
85 vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtable2, vx2));
86 vx0 = _mm256_sub_epi8(vx0, voffset);
87 vx1 = _mm256_sub_epi8(vx1, voffset);
88 vx2 = _mm256_sub_epi8(vx2, voffset);
89 vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtable3, vx0));
90 vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtable3, vx1));
91 vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtable3, vx2));
92 vx0 = _mm256_sub_epi8(vx0, voffset);
93 vx1 = _mm256_sub_epi8(vx1, voffset);
94 vx2 = _mm256_sub_epi8(vx2, voffset);
95 vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtable4, vx0));
96 vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtable4, vx1));
97 vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtable4, vx2));
98 vx0 = _mm256_sub_epi8(vx0, voffset);
99 vx1 = _mm256_sub_epi8(vx1, voffset);
100 vx2 = _mm256_sub_epi8(vx2, voffset);
101 vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtable5, vx0));
102 vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtable5, vx1));
103 vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtable5, vx2));
104 vx0 = _mm256_sub_epi8(vx0, voffset);
105 vx1 = _mm256_sub_epi8(vx1, voffset);
106 vx2 = _mm256_sub_epi8(vx2, voffset);
107 vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtable6, vx0));
108 vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtable6, vx1));
109 vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtable6, vx2));
110 vx0 = _mm256_sub_epi8(vx0, voffset);
111 vx1 = _mm256_sub_epi8(vx1, voffset);
112 vx2 = _mm256_sub_epi8(vx2, voffset);
113 vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtable7, vx0));
114 vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtable7, vx1));
115 vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtable7, vx2));
116 vx0 = _mm256_sub_epi8(vx0, voffset);
117 vx1 = _mm256_sub_epi8(vx1, voffset);
118 vx2 = _mm256_sub_epi8(vx2, voffset);
119 vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtable8, vx0));
120 vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtable8, vx1));
121 vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtable8, vx2));
122
123 vx0 = _mm256_subs_epi8(vx0, voffset);
124 vx1 = _mm256_subs_epi8(vx1, voffset);
125 vx2 = _mm256_subs_epi8(vx2, voffset);
126 vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtable9, vx0));
127 vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtable9, vx1));
128 vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtable9, vx2));
129 vx0 = _mm256_subs_epi8(vx0, voffset);
130 vx1 = _mm256_subs_epi8(vx1, voffset);
131 vx2 = _mm256_subs_epi8(vx2, voffset);
132 vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtableA, vx0));
133 vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtableA, vx1));
134 vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtableA, vx2));
135 vx0 = _mm256_subs_epi8(vx0, voffset);
136 vx1 = _mm256_subs_epi8(vx1, voffset);
137 vx2 = _mm256_subs_epi8(vx2, voffset);
138 vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtableB, vx0));
139 vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtableB, vx1));
140 vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtableB, vx2));
141 vx0 = _mm256_subs_epi8(vx0, voffset);
142 vx1 = _mm256_subs_epi8(vx1, voffset);
143 vx2 = _mm256_subs_epi8(vx2, voffset);
144 vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtableC, vx0));
145 vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtableC, vx1));
146 vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtableC, vx2));
147 vx0 = _mm256_subs_epi8(vx0, voffset);
148 vx1 = _mm256_subs_epi8(vx1, voffset);
149 vx2 = _mm256_subs_epi8(vx2, voffset);
150 vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtableD, vx0));
151 vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtableD, vx1));
152 vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtableD, vx2));
153 vx0 = _mm256_subs_epi8(vx0, voffset);
154 vx1 = _mm256_subs_epi8(vx1, voffset);
155 vx2 = _mm256_subs_epi8(vx2, voffset);
156 vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtableE, vx0));
157 vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtableE, vx1));
158 vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtableE, vx2));
159 vx0 = _mm256_subs_epi8(vx0, voffset);
160 vx1 = _mm256_subs_epi8(vx1, voffset);
161 vx2 = _mm256_subs_epi8(vx2, voffset);
162 vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtableF, vx0));
163 vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtableF, vx1));
164 vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtableF, vx2));
165
166 _mm256_storeu_si256((__m256i*) y, vy0);
167 _mm256_storeu_si256((__m256i*) (y + 32), vy1);
168 _mm256_storeu_si256((__m256i*) (y + 64), vy2);
169 y += 96;
170 }
171 for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
172 __m128i vx = _mm_loadu_si128((const __m128i*) x);
173 x += 16;
174
175 __m128i vy = _mm_shuffle_epi8(_mm256_castsi256_si128(vtable0), vx);
176
177 vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
178 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable1), vx));
179 vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
180 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable2), vx));
181 vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
182 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable3), vx));
183 vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
184 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable4), vx));
185 vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
186 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable5), vx));
187 vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
188 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable6), vx));
189 vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
190 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable7), vx));
191 vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
192 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable8), vx));
193
194 vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
195 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable9), vx));
196 vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
197 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableA), vx));
198 vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
199 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableB), vx));
200 vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
201 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableC), vx));
202 vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
203 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableD), vx));
204 vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
205 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableE), vx));
206 vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
207 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableF), vx));
208
209 _mm_storeu_si128((__m128i*) y, vy);
210 y += 16;
211 }
212 if XNN_UNLIKELY(n != 0) {
213 __m128i vx = _mm_loadu_si128((const __m128i*) x);
214
215 __m128i vy = _mm_shuffle_epi8(_mm256_castsi256_si128(vtable0), vx);
216
217 vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
218 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable1), vx));
219 vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
220 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable2), vx));
221 vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
222 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable3), vx));
223 vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
224 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable4), vx));
225 vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
226 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable5), vx));
227 vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
228 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable6), vx));
229 vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
230 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable7), vx));
231 vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
232 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable8), vx));
233
234 vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
235 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable9), vx));
236 vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
237 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableA), vx));
238 vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
239 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableB), vx));
240 vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
241 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableC), vx));
242 vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
243 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableD), vx));
244 vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
245 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableE), vx));
246 vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
247 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableF), vx));
248
249 if (n & (8 * sizeof(uint8_t))) {
250 _mm_storel_epi64((__m128i*) y, vy);
251 vy = _mm_unpackhi_epi64(vy, vy);
252 y += 8;
253 }
254 if (n & (4 * sizeof(uint8_t))) {
255 _mm_storeu_si32(y, vy);
256 vy = _mm_srli_epi64(vy, 32);
257 y += 4;
258 }
259 if (n & (2 * sizeof(uint8_t))) {
260 _mm_storeu_si16(y, vy);
261 vy = _mm_srli_epi32(vy, 16);
262 y += 2;
263 }
264 if (n & (1 * sizeof(uint8_t))) {
265 *y = (uint8_t) _mm_extract_epi8(vy, 0);
266 }
267 }
268 }
269