1 // Auto-generated file. Do not edit!
2 // Template: src/x8-lut/avx512skx-vpshufb.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__avx512skx_vpshufb_x192(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__avx512skx_vpshufb_x192(
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 __m512i vt0 = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) t));
30 const __m512i vt1 = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) (t + 16)));
31 const __m512i vt2 = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) (t + 32)));
32 const __m512i vt3 = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) (t + 48)));
33 const __m512i vt4 = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) (t + 64)));
34 const __m512i vt5 = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) (t + 80)));
35 const __m512i vt6 = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) (t + 96)));
36 const __m512i vt7 = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) (t + 112)));
37 const __m512i vt8 = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) (t + 128)));
38 const __m512i vt9 = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) (t + 144)));
39 const __m512i vtA = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) (t + 160)));
40 const __m512i vtB = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) (t + 176)));
41 const __m512i vtC = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) (t + 192)));
42 const __m512i vtD = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) (t + 208)));
43 const __m512i vtE = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) (t + 224)));
44 const __m512i vtF = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) (t + 240)));
45
46 const __m512i vtable0 = vt0;
47 const __m512i vtable1 = _mm512_xor_si512(vt0, vt1);
48 const __m512i vtable2 = _mm512_xor_si512(vt1, vt2);
49 const __m512i vtable3 = _mm512_xor_si512(vt2, vt3);
50 const __m512i vtable4 = _mm512_xor_si512(vt3, vt4);
51 const __m512i vtable5 = _mm512_xor_si512(vt4, vt5);
52 const __m512i vtable6 = _mm512_xor_si512(vt5, vt6);
53 const __m512i vtable7 = _mm512_xor_si512(vt6, vt7);
54 const __m512i vtable8 = _mm512_xor_si512(_mm512_xor_si512(vt7, vt8), vtable0);
55 const __m512i vtable9 = _mm512_xor_si512(_mm512_xor_si512(vt8, vt9), vtable1);
56 const __m512i vtableA = _mm512_xor_si512(_mm512_xor_si512(vt9, vtA), vtable2);
57 const __m512i vtableB = _mm512_xor_si512(_mm512_xor_si512(vtA, vtB), vtable3);
58 const __m512i vtableC = _mm512_xor_si512(_mm512_xor_si512(vtB, vtC), vtable4);
59 const __m512i vtableD = _mm512_xor_si512(_mm512_xor_si512(vtC, vtD), vtable5);
60 const __m512i vtableE = _mm512_xor_si512(_mm512_xor_si512(vtD, vtE), vtable6);
61 const __m512i vtableF = _mm512_xor_si512(_mm512_xor_si512(vtE, vtF), vtable7);
62
63 const __m512i voffset = _mm512_set1_epi8(16);
64 for (; n >= 192 * sizeof(uint8_t); n -= 192 * sizeof(uint8_t)) {
65 __m512i vx0 = _mm512_loadu_si512(x);
66 __m512i vx1 = _mm512_loadu_si512(x + 64);
67 __m512i vx2 = _mm512_loadu_si512(x + 128);
68 x += 192;
69
70 __m512i vy0 = _mm512_shuffle_epi8(vtable0, vx0);
71 __m512i vy1 = _mm512_shuffle_epi8(vtable0, vx1);
72 __m512i vy2 = _mm512_shuffle_epi8(vtable0, vx2);
73
74 vx0 = _mm512_sub_epi8(vx0, voffset);
75 vx1 = _mm512_sub_epi8(vx1, voffset);
76 vx2 = _mm512_sub_epi8(vx2, voffset);
77 vy0 = _mm512_xor_si512(vy0, _mm512_shuffle_epi8(vtable1, vx0));
78 vy1 = _mm512_xor_si512(vy1, _mm512_shuffle_epi8(vtable1, vx1));
79 vy2 = _mm512_xor_si512(vy2, _mm512_shuffle_epi8(vtable1, vx2));
80 vx0 = _mm512_sub_epi8(vx0, voffset);
81 vx1 = _mm512_sub_epi8(vx1, voffset);
82 vx2 = _mm512_sub_epi8(vx2, voffset);
83 vy0 = _mm512_xor_si512(vy0, _mm512_shuffle_epi8(vtable2, vx0));
84 vy1 = _mm512_xor_si512(vy1, _mm512_shuffle_epi8(vtable2, vx1));
85 vy2 = _mm512_xor_si512(vy2, _mm512_shuffle_epi8(vtable2, vx2));
86 vx0 = _mm512_sub_epi8(vx0, voffset);
87 vx1 = _mm512_sub_epi8(vx1, voffset);
88 vx2 = _mm512_sub_epi8(vx2, voffset);
89 vy0 = _mm512_xor_si512(vy0, _mm512_shuffle_epi8(vtable3, vx0));
90 vy1 = _mm512_xor_si512(vy1, _mm512_shuffle_epi8(vtable3, vx1));
91 vy2 = _mm512_xor_si512(vy2, _mm512_shuffle_epi8(vtable3, vx2));
92 vx0 = _mm512_sub_epi8(vx0, voffset);
93 vx1 = _mm512_sub_epi8(vx1, voffset);
94 vx2 = _mm512_sub_epi8(vx2, voffset);
95 vy0 = _mm512_xor_si512(vy0, _mm512_shuffle_epi8(vtable4, vx0));
96 vy1 = _mm512_xor_si512(vy1, _mm512_shuffle_epi8(vtable4, vx1));
97 vy2 = _mm512_xor_si512(vy2, _mm512_shuffle_epi8(vtable4, vx2));
98 vx0 = _mm512_sub_epi8(vx0, voffset);
99 vx1 = _mm512_sub_epi8(vx1, voffset);
100 vx2 = _mm512_sub_epi8(vx2, voffset);
101 vy0 = _mm512_xor_si512(vy0, _mm512_shuffle_epi8(vtable5, vx0));
102 vy1 = _mm512_xor_si512(vy1, _mm512_shuffle_epi8(vtable5, vx1));
103 vy2 = _mm512_xor_si512(vy2, _mm512_shuffle_epi8(vtable5, vx2));
104 vx0 = _mm512_sub_epi8(vx0, voffset);
105 vx1 = _mm512_sub_epi8(vx1, voffset);
106 vx2 = _mm512_sub_epi8(vx2, voffset);
107 vy0 = _mm512_xor_si512(vy0, _mm512_shuffle_epi8(vtable6, vx0));
108 vy1 = _mm512_xor_si512(vy1, _mm512_shuffle_epi8(vtable6, vx1));
109 vy2 = _mm512_xor_si512(vy2, _mm512_shuffle_epi8(vtable6, vx2));
110 vx0 = _mm512_sub_epi8(vx0, voffset);
111 vx1 = _mm512_sub_epi8(vx1, voffset);
112 vx2 = _mm512_sub_epi8(vx2, voffset);
113 vy0 = _mm512_xor_si512(vy0, _mm512_shuffle_epi8(vtable7, vx0));
114 vy1 = _mm512_xor_si512(vy1, _mm512_shuffle_epi8(vtable7, vx1));
115 vy2 = _mm512_xor_si512(vy2, _mm512_shuffle_epi8(vtable7, vx2));
116 vx0 = _mm512_sub_epi8(vx0, voffset);
117 vx1 = _mm512_sub_epi8(vx1, voffset);
118 vx2 = _mm512_sub_epi8(vx2, voffset);
119 vy0 = _mm512_xor_si512(vy0, _mm512_shuffle_epi8(vtable8, vx0));
120 vy1 = _mm512_xor_si512(vy1, _mm512_shuffle_epi8(vtable8, vx1));
121 vy2 = _mm512_xor_si512(vy2, _mm512_shuffle_epi8(vtable8, vx2));
122
123 vx0 = _mm512_subs_epi8(vx0, voffset);
124 vx1 = _mm512_subs_epi8(vx1, voffset);
125 vx2 = _mm512_subs_epi8(vx2, voffset);
126 vy0 = _mm512_xor_si512(vy0, _mm512_shuffle_epi8(vtable9, vx0));
127 vy1 = _mm512_xor_si512(vy1, _mm512_shuffle_epi8(vtable9, vx1));
128 vy2 = _mm512_xor_si512(vy2, _mm512_shuffle_epi8(vtable9, vx2));
129 vx0 = _mm512_subs_epi8(vx0, voffset);
130 vx1 = _mm512_subs_epi8(vx1, voffset);
131 vx2 = _mm512_subs_epi8(vx2, voffset);
132 vy0 = _mm512_xor_si512(vy0, _mm512_shuffle_epi8(vtableA, vx0));
133 vy1 = _mm512_xor_si512(vy1, _mm512_shuffle_epi8(vtableA, vx1));
134 vy2 = _mm512_xor_si512(vy2, _mm512_shuffle_epi8(vtableA, vx2));
135 vx0 = _mm512_subs_epi8(vx0, voffset);
136 vx1 = _mm512_subs_epi8(vx1, voffset);
137 vx2 = _mm512_subs_epi8(vx2, voffset);
138 vy0 = _mm512_xor_si512(vy0, _mm512_shuffle_epi8(vtableB, vx0));
139 vy1 = _mm512_xor_si512(vy1, _mm512_shuffle_epi8(vtableB, vx1));
140 vy2 = _mm512_xor_si512(vy2, _mm512_shuffle_epi8(vtableB, vx2));
141 vx0 = _mm512_subs_epi8(vx0, voffset);
142 vx1 = _mm512_subs_epi8(vx1, voffset);
143 vx2 = _mm512_subs_epi8(vx2, voffset);
144 vy0 = _mm512_xor_si512(vy0, _mm512_shuffle_epi8(vtableC, vx0));
145 vy1 = _mm512_xor_si512(vy1, _mm512_shuffle_epi8(vtableC, vx1));
146 vy2 = _mm512_xor_si512(vy2, _mm512_shuffle_epi8(vtableC, vx2));
147 vx0 = _mm512_subs_epi8(vx0, voffset);
148 vx1 = _mm512_subs_epi8(vx1, voffset);
149 vx2 = _mm512_subs_epi8(vx2, voffset);
150 vy0 = _mm512_xor_si512(vy0, _mm512_shuffle_epi8(vtableD, vx0));
151 vy1 = _mm512_xor_si512(vy1, _mm512_shuffle_epi8(vtableD, vx1));
152 vy2 = _mm512_xor_si512(vy2, _mm512_shuffle_epi8(vtableD, vx2));
153 vx0 = _mm512_subs_epi8(vx0, voffset);
154 vx1 = _mm512_subs_epi8(vx1, voffset);
155 vx2 = _mm512_subs_epi8(vx2, voffset);
156 vy0 = _mm512_xor_si512(vy0, _mm512_shuffle_epi8(vtableE, vx0));
157 vy1 = _mm512_xor_si512(vy1, _mm512_shuffle_epi8(vtableE, vx1));
158 vy2 = _mm512_xor_si512(vy2, _mm512_shuffle_epi8(vtableE, vx2));
159 vx0 = _mm512_subs_epi8(vx0, voffset);
160 vx1 = _mm512_subs_epi8(vx1, voffset);
161 vx2 = _mm512_subs_epi8(vx2, voffset);
162 vy0 = _mm512_xor_si512(vy0, _mm512_shuffle_epi8(vtableF, vx0));
163 vy1 = _mm512_xor_si512(vy1, _mm512_shuffle_epi8(vtableF, vx1));
164 vy2 = _mm512_xor_si512(vy2, _mm512_shuffle_epi8(vtableF, vx2));
165
166 _mm512_storeu_si512(y, vy0);
167 _mm512_storeu_si512(y + 64, vy1);
168 _mm512_storeu_si512(y + 128, vy2);
169 y += 192;
170 }
171 for (; n >= 64 * sizeof(uint8_t); n -= 64 * sizeof(uint8_t)) {
172 __m512i vx = _mm512_loadu_si512(x);
173 x += 64;
174
175 __m512i vy = _mm512_shuffle_epi8(vtable0, vx);
176
177 vx = _mm512_sub_epi8(vx, voffset);
178 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable1, vx));
179 vx = _mm512_sub_epi8(vx, voffset);
180 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable2, vx));
181 vx = _mm512_sub_epi8(vx, voffset);
182 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable3, vx));
183 vx = _mm512_sub_epi8(vx, voffset);
184 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable4, vx));
185 vx = _mm512_sub_epi8(vx, voffset);
186 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable5, vx));
187 vx = _mm512_sub_epi8(vx, voffset);
188 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable6, vx));
189 vx = _mm512_sub_epi8(vx, voffset);
190 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable7, vx));
191 vx = _mm512_sub_epi8(vx, voffset);
192 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable8, vx));
193
194 vx = _mm512_subs_epi8(vx, voffset);
195 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable9, vx));
196 vx = _mm512_subs_epi8(vx, voffset);
197 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableA, vx));
198 vx = _mm512_subs_epi8(vx, voffset);
199 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableB, vx));
200 vx = _mm512_subs_epi8(vx, voffset);
201 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableC, vx));
202 vx = _mm512_subs_epi8(vx, voffset);
203 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableD, vx));
204 vx = _mm512_subs_epi8(vx, voffset);
205 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableE, vx));
206 vx = _mm512_subs_epi8(vx, voffset);
207 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableF, vx));
208
209 _mm512_storeu_si512(y, vy);
210 y += 64;
211 }
212 if XNN_UNLIKELY(n != 0) {
213 assert(n < 64);
214 const __mmask64 vmask = _cvtu64_mask64((uint64_t) ((UINT64_C(1) << n) - UINT64_C(1)));
215
216 __m512i vx = _mm512_maskz_loadu_epi8(vmask, x);
217
218 __m512i vy = _mm512_shuffle_epi8(vtable0, vx);
219
220 vx = _mm512_sub_epi8(vx, voffset);
221 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable1, vx));
222 vx = _mm512_sub_epi8(vx, voffset);
223 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable2, vx));
224 vx = _mm512_sub_epi8(vx, voffset);
225 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable3, vx));
226 vx = _mm512_sub_epi8(vx, voffset);
227 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable4, vx));
228 vx = _mm512_sub_epi8(vx, voffset);
229 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable5, vx));
230 vx = _mm512_sub_epi8(vx, voffset);
231 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable6, vx));
232 vx = _mm512_sub_epi8(vx, voffset);
233 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable7, vx));
234 vx = _mm512_sub_epi8(vx, voffset);
235 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable8, vx));
236
237 vx = _mm512_subs_epi8(vx, voffset);
238 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable9, vx));
239 vx = _mm512_subs_epi8(vx, voffset);
240 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableA, vx));
241 vx = _mm512_subs_epi8(vx, voffset);
242 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableB, vx));
243 vx = _mm512_subs_epi8(vx, voffset);
244 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableC, vx));
245 vx = _mm512_subs_epi8(vx, voffset);
246 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableD, vx));
247 vx = _mm512_subs_epi8(vx, voffset);
248 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableE, vx));
249 vx = _mm512_subs_epi8(vx, voffset);
250 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableF, vx));
251
252 _mm512_mask_storeu_epi8(y, vmask, vy);
253 }
254 }
255