1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker // Template: src/x8-lut/avx2.c.in
3*4bdc9457SAndroid Build Coastguard Worker // Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2021 Google LLC
6*4bdc9457SAndroid Build Coastguard Worker //
7*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
8*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
9*4bdc9457SAndroid Build Coastguard Worker
10*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
11*4bdc9457SAndroid Build Coastguard Worker
12*4bdc9457SAndroid Build Coastguard Worker #include <immintrin.h>
13*4bdc9457SAndroid Build Coastguard Worker
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/intrinsics-polyfill.h>
15*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/lut.h>
16*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/common.h>
17*4bdc9457SAndroid Build Coastguard Worker
18*4bdc9457SAndroid Build Coastguard Worker
xnn_x8_lut_ukernel__avx2_x128(size_t n,const uint8_t * x,uint8_t * y,const uint8_t t[restrict XNN_MIN_ELEMENTS (256)])19*4bdc9457SAndroid Build Coastguard Worker void xnn_x8_lut_ukernel__avx2_x128(
20*4bdc9457SAndroid Build Coastguard Worker size_t n,
21*4bdc9457SAndroid Build Coastguard Worker const uint8_t* x,
22*4bdc9457SAndroid Build Coastguard Worker uint8_t* y,
23*4bdc9457SAndroid Build Coastguard Worker const uint8_t t[restrict XNN_MIN_ELEMENTS(256)])
24*4bdc9457SAndroid Build Coastguard Worker {
25*4bdc9457SAndroid Build Coastguard Worker assert(n != 0);
26*4bdc9457SAndroid Build Coastguard Worker assert(x != NULL);
27*4bdc9457SAndroid Build Coastguard Worker assert(y != NULL);
28*4bdc9457SAndroid Build Coastguard Worker
29*4bdc9457SAndroid Build Coastguard Worker const __m256i vt0 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) t));
30*4bdc9457SAndroid Build Coastguard Worker const __m256i vt1 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 16)));
31*4bdc9457SAndroid Build Coastguard Worker const __m256i vt2 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 32)));
32*4bdc9457SAndroid Build Coastguard Worker const __m256i vt3 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 48)));
33*4bdc9457SAndroid Build Coastguard Worker const __m256i vt4 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 64)));
34*4bdc9457SAndroid Build Coastguard Worker const __m256i vt5 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 80)));
35*4bdc9457SAndroid Build Coastguard Worker const __m256i vt6 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 96)));
36*4bdc9457SAndroid Build Coastguard Worker const __m256i vt7 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 112)));
37*4bdc9457SAndroid Build Coastguard Worker const __m256i vt8 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 128)));
38*4bdc9457SAndroid Build Coastguard Worker const __m256i vt9 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 144)));
39*4bdc9457SAndroid Build Coastguard Worker const __m256i vtA = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 160)));
40*4bdc9457SAndroid Build Coastguard Worker const __m256i vtB = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 176)));
41*4bdc9457SAndroid Build Coastguard Worker const __m256i vtC = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 192)));
42*4bdc9457SAndroid Build Coastguard Worker const __m256i vtD = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 208)));
43*4bdc9457SAndroid Build Coastguard Worker const __m256i vtE = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 224)));
44*4bdc9457SAndroid Build Coastguard Worker const __m256i vtF = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 240)));
45*4bdc9457SAndroid Build Coastguard Worker
46*4bdc9457SAndroid Build Coastguard Worker const __m256i vtable0 = vt0;
47*4bdc9457SAndroid Build Coastguard Worker const __m256i vtable1 = _mm256_xor_si256(vt0, vt1);
48*4bdc9457SAndroid Build Coastguard Worker const __m256i vtable2 = _mm256_xor_si256(vt1, vt2);
49*4bdc9457SAndroid Build Coastguard Worker const __m256i vtable3 = _mm256_xor_si256(vt2, vt3);
50*4bdc9457SAndroid Build Coastguard Worker const __m256i vtable4 = _mm256_xor_si256(vt3, vt4);
51*4bdc9457SAndroid Build Coastguard Worker const __m256i vtable5 = _mm256_xor_si256(vt4, vt5);
52*4bdc9457SAndroid Build Coastguard Worker const __m256i vtable6 = _mm256_xor_si256(vt5, vt6);
53*4bdc9457SAndroid Build Coastguard Worker const __m256i vtable7 = _mm256_xor_si256(vt6, vt7);
54*4bdc9457SAndroid Build Coastguard Worker const __m256i vtable8 = _mm256_xor_si256(_mm256_xor_si256(vt7, vt8), vtable0);
55*4bdc9457SAndroid Build Coastguard Worker const __m256i vtable9 = _mm256_xor_si256(_mm256_xor_si256(vt8, vt9), vtable1);
56*4bdc9457SAndroid Build Coastguard Worker const __m256i vtableA = _mm256_xor_si256(_mm256_xor_si256(vt9, vtA), vtable2);
57*4bdc9457SAndroid Build Coastguard Worker const __m256i vtableB = _mm256_xor_si256(_mm256_xor_si256(vtA, vtB), vtable3);
58*4bdc9457SAndroid Build Coastguard Worker const __m256i vtableC = _mm256_xor_si256(_mm256_xor_si256(vtB, vtC), vtable4);
59*4bdc9457SAndroid Build Coastguard Worker const __m256i vtableD = _mm256_xor_si256(_mm256_xor_si256(vtC, vtD), vtable5);
60*4bdc9457SAndroid Build Coastguard Worker const __m256i vtableE = _mm256_xor_si256(_mm256_xor_si256(vtD, vtE), vtable6);
61*4bdc9457SAndroid Build Coastguard Worker const __m256i vtableF = _mm256_xor_si256(_mm256_xor_si256(vtE, vtF), vtable7);
62*4bdc9457SAndroid Build Coastguard Worker
63*4bdc9457SAndroid Build Coastguard Worker const __m256i voffset = _mm256_set1_epi8(16);
64*4bdc9457SAndroid Build Coastguard Worker for (; n >= 128 * sizeof(uint8_t); n -= 128 * sizeof(uint8_t)) {
65*4bdc9457SAndroid Build Coastguard Worker __m256i vx0 = _mm256_loadu_si256((const __m256i*) x);
66*4bdc9457SAndroid Build Coastguard Worker __m256i vx1 = _mm256_loadu_si256((const __m256i*) (x + 32));
67*4bdc9457SAndroid Build Coastguard Worker __m256i vx2 = _mm256_loadu_si256((const __m256i*) (x + 64));
68*4bdc9457SAndroid Build Coastguard Worker __m256i vx3 = _mm256_loadu_si256((const __m256i*) (x + 96));
69*4bdc9457SAndroid Build Coastguard Worker x += 128;
70*4bdc9457SAndroid Build Coastguard Worker
71*4bdc9457SAndroid Build Coastguard Worker __m256i vy0 = _mm256_shuffle_epi8(vtable0, vx0);
72*4bdc9457SAndroid Build Coastguard Worker __m256i vy1 = _mm256_shuffle_epi8(vtable0, vx1);
73*4bdc9457SAndroid Build Coastguard Worker __m256i vy2 = _mm256_shuffle_epi8(vtable0, vx2);
74*4bdc9457SAndroid Build Coastguard Worker __m256i vy3 = _mm256_shuffle_epi8(vtable0, vx3);
75*4bdc9457SAndroid Build Coastguard Worker
76*4bdc9457SAndroid Build Coastguard Worker vx0 = _mm256_sub_epi8(vx0, voffset);
77*4bdc9457SAndroid Build Coastguard Worker vx1 = _mm256_sub_epi8(vx1, voffset);
78*4bdc9457SAndroid Build Coastguard Worker vx2 = _mm256_sub_epi8(vx2, voffset);
79*4bdc9457SAndroid Build Coastguard Worker vx3 = _mm256_sub_epi8(vx3, voffset);
80*4bdc9457SAndroid Build Coastguard Worker vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtable1, vx0));
81*4bdc9457SAndroid Build Coastguard Worker vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtable1, vx1));
82*4bdc9457SAndroid Build Coastguard Worker vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtable1, vx2));
83*4bdc9457SAndroid Build Coastguard Worker vy3 = _mm256_xor_si256(vy3, _mm256_shuffle_epi8(vtable1, vx3));
84*4bdc9457SAndroid Build Coastguard Worker vx0 = _mm256_sub_epi8(vx0, voffset);
85*4bdc9457SAndroid Build Coastguard Worker vx1 = _mm256_sub_epi8(vx1, voffset);
86*4bdc9457SAndroid Build Coastguard Worker vx2 = _mm256_sub_epi8(vx2, voffset);
87*4bdc9457SAndroid Build Coastguard Worker vx3 = _mm256_sub_epi8(vx3, voffset);
88*4bdc9457SAndroid Build Coastguard Worker vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtable2, vx0));
89*4bdc9457SAndroid Build Coastguard Worker vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtable2, vx1));
90*4bdc9457SAndroid Build Coastguard Worker vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtable2, vx2));
91*4bdc9457SAndroid Build Coastguard Worker vy3 = _mm256_xor_si256(vy3, _mm256_shuffle_epi8(vtable2, vx3));
92*4bdc9457SAndroid Build Coastguard Worker vx0 = _mm256_sub_epi8(vx0, voffset);
93*4bdc9457SAndroid Build Coastguard Worker vx1 = _mm256_sub_epi8(vx1, voffset);
94*4bdc9457SAndroid Build Coastguard Worker vx2 = _mm256_sub_epi8(vx2, voffset);
95*4bdc9457SAndroid Build Coastguard Worker vx3 = _mm256_sub_epi8(vx3, voffset);
96*4bdc9457SAndroid Build Coastguard Worker vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtable3, vx0));
97*4bdc9457SAndroid Build Coastguard Worker vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtable3, vx1));
98*4bdc9457SAndroid Build Coastguard Worker vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtable3, vx2));
99*4bdc9457SAndroid Build Coastguard Worker vy3 = _mm256_xor_si256(vy3, _mm256_shuffle_epi8(vtable3, vx3));
100*4bdc9457SAndroid Build Coastguard Worker vx0 = _mm256_sub_epi8(vx0, voffset);
101*4bdc9457SAndroid Build Coastguard Worker vx1 = _mm256_sub_epi8(vx1, voffset);
102*4bdc9457SAndroid Build Coastguard Worker vx2 = _mm256_sub_epi8(vx2, voffset);
103*4bdc9457SAndroid Build Coastguard Worker vx3 = _mm256_sub_epi8(vx3, voffset);
104*4bdc9457SAndroid Build Coastguard Worker vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtable4, vx0));
105*4bdc9457SAndroid Build Coastguard Worker vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtable4, vx1));
106*4bdc9457SAndroid Build Coastguard Worker vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtable4, vx2));
107*4bdc9457SAndroid Build Coastguard Worker vy3 = _mm256_xor_si256(vy3, _mm256_shuffle_epi8(vtable4, vx3));
108*4bdc9457SAndroid Build Coastguard Worker vx0 = _mm256_sub_epi8(vx0, voffset);
109*4bdc9457SAndroid Build Coastguard Worker vx1 = _mm256_sub_epi8(vx1, voffset);
110*4bdc9457SAndroid Build Coastguard Worker vx2 = _mm256_sub_epi8(vx2, voffset);
111*4bdc9457SAndroid Build Coastguard Worker vx3 = _mm256_sub_epi8(vx3, voffset);
112*4bdc9457SAndroid Build Coastguard Worker vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtable5, vx0));
113*4bdc9457SAndroid Build Coastguard Worker vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtable5, vx1));
114*4bdc9457SAndroid Build Coastguard Worker vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtable5, vx2));
115*4bdc9457SAndroid Build Coastguard Worker vy3 = _mm256_xor_si256(vy3, _mm256_shuffle_epi8(vtable5, vx3));
116*4bdc9457SAndroid Build Coastguard Worker vx0 = _mm256_sub_epi8(vx0, voffset);
117*4bdc9457SAndroid Build Coastguard Worker vx1 = _mm256_sub_epi8(vx1, voffset);
118*4bdc9457SAndroid Build Coastguard Worker vx2 = _mm256_sub_epi8(vx2, voffset);
119*4bdc9457SAndroid Build Coastguard Worker vx3 = _mm256_sub_epi8(vx3, voffset);
120*4bdc9457SAndroid Build Coastguard Worker vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtable6, vx0));
121*4bdc9457SAndroid Build Coastguard Worker vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtable6, vx1));
122*4bdc9457SAndroid Build Coastguard Worker vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtable6, vx2));
123*4bdc9457SAndroid Build Coastguard Worker vy3 = _mm256_xor_si256(vy3, _mm256_shuffle_epi8(vtable6, vx3));
124*4bdc9457SAndroid Build Coastguard Worker vx0 = _mm256_sub_epi8(vx0, voffset);
125*4bdc9457SAndroid Build Coastguard Worker vx1 = _mm256_sub_epi8(vx1, voffset);
126*4bdc9457SAndroid Build Coastguard Worker vx2 = _mm256_sub_epi8(vx2, voffset);
127*4bdc9457SAndroid Build Coastguard Worker vx3 = _mm256_sub_epi8(vx3, voffset);
128*4bdc9457SAndroid Build Coastguard Worker vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtable7, vx0));
129*4bdc9457SAndroid Build Coastguard Worker vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtable7, vx1));
130*4bdc9457SAndroid Build Coastguard Worker vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtable7, vx2));
131*4bdc9457SAndroid Build Coastguard Worker vy3 = _mm256_xor_si256(vy3, _mm256_shuffle_epi8(vtable7, vx3));
132*4bdc9457SAndroid Build Coastguard Worker vx0 = _mm256_sub_epi8(vx0, voffset);
133*4bdc9457SAndroid Build Coastguard Worker vx1 = _mm256_sub_epi8(vx1, voffset);
134*4bdc9457SAndroid Build Coastguard Worker vx2 = _mm256_sub_epi8(vx2, voffset);
135*4bdc9457SAndroid Build Coastguard Worker vx3 = _mm256_sub_epi8(vx3, voffset);
136*4bdc9457SAndroid Build Coastguard Worker vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtable8, vx0));
137*4bdc9457SAndroid Build Coastguard Worker vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtable8, vx1));
138*4bdc9457SAndroid Build Coastguard Worker vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtable8, vx2));
139*4bdc9457SAndroid Build Coastguard Worker vy3 = _mm256_xor_si256(vy3, _mm256_shuffle_epi8(vtable8, vx3));
140*4bdc9457SAndroid Build Coastguard Worker
141*4bdc9457SAndroid Build Coastguard Worker vx0 = _mm256_subs_epi8(vx0, voffset);
142*4bdc9457SAndroid Build Coastguard Worker vx1 = _mm256_subs_epi8(vx1, voffset);
143*4bdc9457SAndroid Build Coastguard Worker vx2 = _mm256_subs_epi8(vx2, voffset);
144*4bdc9457SAndroid Build Coastguard Worker vx3 = _mm256_subs_epi8(vx3, voffset);
145*4bdc9457SAndroid Build Coastguard Worker vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtable9, vx0));
146*4bdc9457SAndroid Build Coastguard Worker vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtable9, vx1));
147*4bdc9457SAndroid Build Coastguard Worker vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtable9, vx2));
148*4bdc9457SAndroid Build Coastguard Worker vy3 = _mm256_xor_si256(vy3, _mm256_shuffle_epi8(vtable9, vx3));
149*4bdc9457SAndroid Build Coastguard Worker vx0 = _mm256_subs_epi8(vx0, voffset);
150*4bdc9457SAndroid Build Coastguard Worker vx1 = _mm256_subs_epi8(vx1, voffset);
151*4bdc9457SAndroid Build Coastguard Worker vx2 = _mm256_subs_epi8(vx2, voffset);
152*4bdc9457SAndroid Build Coastguard Worker vx3 = _mm256_subs_epi8(vx3, voffset);
153*4bdc9457SAndroid Build Coastguard Worker vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtableA, vx0));
154*4bdc9457SAndroid Build Coastguard Worker vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtableA, vx1));
155*4bdc9457SAndroid Build Coastguard Worker vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtableA, vx2));
156*4bdc9457SAndroid Build Coastguard Worker vy3 = _mm256_xor_si256(vy3, _mm256_shuffle_epi8(vtableA, vx3));
157*4bdc9457SAndroid Build Coastguard Worker vx0 = _mm256_subs_epi8(vx0, voffset);
158*4bdc9457SAndroid Build Coastguard Worker vx1 = _mm256_subs_epi8(vx1, voffset);
159*4bdc9457SAndroid Build Coastguard Worker vx2 = _mm256_subs_epi8(vx2, voffset);
160*4bdc9457SAndroid Build Coastguard Worker vx3 = _mm256_subs_epi8(vx3, voffset);
161*4bdc9457SAndroid Build Coastguard Worker vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtableB, vx0));
162*4bdc9457SAndroid Build Coastguard Worker vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtableB, vx1));
163*4bdc9457SAndroid Build Coastguard Worker vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtableB, vx2));
164*4bdc9457SAndroid Build Coastguard Worker vy3 = _mm256_xor_si256(vy3, _mm256_shuffle_epi8(vtableB, vx3));
165*4bdc9457SAndroid Build Coastguard Worker vx0 = _mm256_subs_epi8(vx0, voffset);
166*4bdc9457SAndroid Build Coastguard Worker vx1 = _mm256_subs_epi8(vx1, voffset);
167*4bdc9457SAndroid Build Coastguard Worker vx2 = _mm256_subs_epi8(vx2, voffset);
168*4bdc9457SAndroid Build Coastguard Worker vx3 = _mm256_subs_epi8(vx3, voffset);
169*4bdc9457SAndroid Build Coastguard Worker vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtableC, vx0));
170*4bdc9457SAndroid Build Coastguard Worker vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtableC, vx1));
171*4bdc9457SAndroid Build Coastguard Worker vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtableC, vx2));
172*4bdc9457SAndroid Build Coastguard Worker vy3 = _mm256_xor_si256(vy3, _mm256_shuffle_epi8(vtableC, vx3));
173*4bdc9457SAndroid Build Coastguard Worker vx0 = _mm256_subs_epi8(vx0, voffset);
174*4bdc9457SAndroid Build Coastguard Worker vx1 = _mm256_subs_epi8(vx1, voffset);
175*4bdc9457SAndroid Build Coastguard Worker vx2 = _mm256_subs_epi8(vx2, voffset);
176*4bdc9457SAndroid Build Coastguard Worker vx3 = _mm256_subs_epi8(vx3, voffset);
177*4bdc9457SAndroid Build Coastguard Worker vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtableD, vx0));
178*4bdc9457SAndroid Build Coastguard Worker vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtableD, vx1));
179*4bdc9457SAndroid Build Coastguard Worker vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtableD, vx2));
180*4bdc9457SAndroid Build Coastguard Worker vy3 = _mm256_xor_si256(vy3, _mm256_shuffle_epi8(vtableD, vx3));
181*4bdc9457SAndroid Build Coastguard Worker vx0 = _mm256_subs_epi8(vx0, voffset);
182*4bdc9457SAndroid Build Coastguard Worker vx1 = _mm256_subs_epi8(vx1, voffset);
183*4bdc9457SAndroid Build Coastguard Worker vx2 = _mm256_subs_epi8(vx2, voffset);
184*4bdc9457SAndroid Build Coastguard Worker vx3 = _mm256_subs_epi8(vx3, voffset);
185*4bdc9457SAndroid Build Coastguard Worker vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtableE, vx0));
186*4bdc9457SAndroid Build Coastguard Worker vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtableE, vx1));
187*4bdc9457SAndroid Build Coastguard Worker vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtableE, vx2));
188*4bdc9457SAndroid Build Coastguard Worker vy3 = _mm256_xor_si256(vy3, _mm256_shuffle_epi8(vtableE, vx3));
189*4bdc9457SAndroid Build Coastguard Worker vx0 = _mm256_subs_epi8(vx0, voffset);
190*4bdc9457SAndroid Build Coastguard Worker vx1 = _mm256_subs_epi8(vx1, voffset);
191*4bdc9457SAndroid Build Coastguard Worker vx2 = _mm256_subs_epi8(vx2, voffset);
192*4bdc9457SAndroid Build Coastguard Worker vx3 = _mm256_subs_epi8(vx3, voffset);
193*4bdc9457SAndroid Build Coastguard Worker vy0 = _mm256_xor_si256(vy0, _mm256_shuffle_epi8(vtableF, vx0));
194*4bdc9457SAndroid Build Coastguard Worker vy1 = _mm256_xor_si256(vy1, _mm256_shuffle_epi8(vtableF, vx1));
195*4bdc9457SAndroid Build Coastguard Worker vy2 = _mm256_xor_si256(vy2, _mm256_shuffle_epi8(vtableF, vx2));
196*4bdc9457SAndroid Build Coastguard Worker vy3 = _mm256_xor_si256(vy3, _mm256_shuffle_epi8(vtableF, vx3));
197*4bdc9457SAndroid Build Coastguard Worker
198*4bdc9457SAndroid Build Coastguard Worker _mm256_storeu_si256((__m256i*) y, vy0);
199*4bdc9457SAndroid Build Coastguard Worker _mm256_storeu_si256((__m256i*) (y + 32), vy1);
200*4bdc9457SAndroid Build Coastguard Worker _mm256_storeu_si256((__m256i*) (y + 64), vy2);
201*4bdc9457SAndroid Build Coastguard Worker _mm256_storeu_si256((__m256i*) (y + 96), vy3);
202*4bdc9457SAndroid Build Coastguard Worker y += 128;
203*4bdc9457SAndroid Build Coastguard Worker }
204*4bdc9457SAndroid Build Coastguard Worker for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
205*4bdc9457SAndroid Build Coastguard Worker __m128i vx = _mm_loadu_si128((const __m128i*) x);
206*4bdc9457SAndroid Build Coastguard Worker x += 16;
207*4bdc9457SAndroid Build Coastguard Worker
208*4bdc9457SAndroid Build Coastguard Worker __m128i vy = _mm_shuffle_epi8(_mm256_castsi256_si128(vtable0), vx);
209*4bdc9457SAndroid Build Coastguard Worker
210*4bdc9457SAndroid Build Coastguard Worker vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
211*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable1), vx));
212*4bdc9457SAndroid Build Coastguard Worker vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
213*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable2), vx));
214*4bdc9457SAndroid Build Coastguard Worker vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
215*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable3), vx));
216*4bdc9457SAndroid Build Coastguard Worker vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
217*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable4), vx));
218*4bdc9457SAndroid Build Coastguard Worker vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
219*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable5), vx));
220*4bdc9457SAndroid Build Coastguard Worker vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
221*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable6), vx));
222*4bdc9457SAndroid Build Coastguard Worker vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
223*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable7), vx));
224*4bdc9457SAndroid Build Coastguard Worker vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
225*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable8), vx));
226*4bdc9457SAndroid Build Coastguard Worker
227*4bdc9457SAndroid Build Coastguard Worker vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
228*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable9), vx));
229*4bdc9457SAndroid Build Coastguard Worker vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
230*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableA), vx));
231*4bdc9457SAndroid Build Coastguard Worker vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
232*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableB), vx));
233*4bdc9457SAndroid Build Coastguard Worker vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
234*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableC), vx));
235*4bdc9457SAndroid Build Coastguard Worker vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
236*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableD), vx));
237*4bdc9457SAndroid Build Coastguard Worker vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
238*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableE), vx));
239*4bdc9457SAndroid Build Coastguard Worker vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
240*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableF), vx));
241*4bdc9457SAndroid Build Coastguard Worker
242*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_si128((__m128i*) y, vy);
243*4bdc9457SAndroid Build Coastguard Worker y += 16;
244*4bdc9457SAndroid Build Coastguard Worker }
245*4bdc9457SAndroid Build Coastguard Worker if XNN_UNLIKELY(n != 0) {
246*4bdc9457SAndroid Build Coastguard Worker __m128i vx = _mm_loadu_si128((const __m128i*) x);
247*4bdc9457SAndroid Build Coastguard Worker
248*4bdc9457SAndroid Build Coastguard Worker __m128i vy = _mm_shuffle_epi8(_mm256_castsi256_si128(vtable0), vx);
249*4bdc9457SAndroid Build Coastguard Worker
250*4bdc9457SAndroid Build Coastguard Worker vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
251*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable1), vx));
252*4bdc9457SAndroid Build Coastguard Worker vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
253*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable2), vx));
254*4bdc9457SAndroid Build Coastguard Worker vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
255*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable3), vx));
256*4bdc9457SAndroid Build Coastguard Worker vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
257*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable4), vx));
258*4bdc9457SAndroid Build Coastguard Worker vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
259*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable5), vx));
260*4bdc9457SAndroid Build Coastguard Worker vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
261*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable6), vx));
262*4bdc9457SAndroid Build Coastguard Worker vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
263*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable7), vx));
264*4bdc9457SAndroid Build Coastguard Worker vx = _mm_sub_epi8(vx, _mm256_castsi256_si128(voffset));
265*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable8), vx));
266*4bdc9457SAndroid Build Coastguard Worker
267*4bdc9457SAndroid Build Coastguard Worker vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
268*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtable9), vx));
269*4bdc9457SAndroid Build Coastguard Worker vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
270*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableA), vx));
271*4bdc9457SAndroid Build Coastguard Worker vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
272*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableB), vx));
273*4bdc9457SAndroid Build Coastguard Worker vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
274*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableC), vx));
275*4bdc9457SAndroid Build Coastguard Worker vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
276*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableD), vx));
277*4bdc9457SAndroid Build Coastguard Worker vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
278*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableE), vx));
279*4bdc9457SAndroid Build Coastguard Worker vx = _mm_subs_epi8(vx, _mm256_castsi256_si128(voffset));
280*4bdc9457SAndroid Build Coastguard Worker vy = _mm_xor_si128(vy, _mm_shuffle_epi8(_mm256_castsi256_si128(vtableF), vx));
281*4bdc9457SAndroid Build Coastguard Worker
282*4bdc9457SAndroid Build Coastguard Worker if (n & (8 * sizeof(uint8_t))) {
283*4bdc9457SAndroid Build Coastguard Worker _mm_storel_epi64((__m128i*) y, vy);
284*4bdc9457SAndroid Build Coastguard Worker vy = _mm_unpackhi_epi64(vy, vy);
285*4bdc9457SAndroid Build Coastguard Worker y += 8;
286*4bdc9457SAndroid Build Coastguard Worker }
287*4bdc9457SAndroid Build Coastguard Worker if (n & (4 * sizeof(uint8_t))) {
288*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_si32(y, vy);
289*4bdc9457SAndroid Build Coastguard Worker vy = _mm_srli_epi64(vy, 32);
290*4bdc9457SAndroid Build Coastguard Worker y += 4;
291*4bdc9457SAndroid Build Coastguard Worker }
292*4bdc9457SAndroid Build Coastguard Worker if (n & (2 * sizeof(uint8_t))) {
293*4bdc9457SAndroid Build Coastguard Worker _mm_storeu_si16(y, vy);
294*4bdc9457SAndroid Build Coastguard Worker vy = _mm_srli_epi32(vy, 16);
295*4bdc9457SAndroid Build Coastguard Worker y += 2;
296*4bdc9457SAndroid Build Coastguard Worker }
297*4bdc9457SAndroid Build Coastguard Worker if (n & (1 * sizeof(uint8_t))) {
298*4bdc9457SAndroid Build Coastguard Worker *y = (uint8_t) _mm_extract_epi8(vy, 0);
299*4bdc9457SAndroid Build Coastguard Worker }
300*4bdc9457SAndroid Build Coastguard Worker }
301*4bdc9457SAndroid Build Coastguard Worker }
302