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_x64(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_x64(
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 >= 64 * sizeof(uint8_t); n -= 64 * sizeof(uint8_t)) {
65 __m512i vx = _mm512_loadu_si512(x);
66 x += 64;
67
68 __m512i vy = _mm512_shuffle_epi8(vtable0, vx);
69
70 vx = _mm512_sub_epi8(vx, voffset);
71 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable1, vx));
72 vx = _mm512_sub_epi8(vx, voffset);
73 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable2, vx));
74 vx = _mm512_sub_epi8(vx, voffset);
75 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable3, vx));
76 vx = _mm512_sub_epi8(vx, voffset);
77 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable4, vx));
78 vx = _mm512_sub_epi8(vx, voffset);
79 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable5, vx));
80 vx = _mm512_sub_epi8(vx, voffset);
81 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable6, vx));
82 vx = _mm512_sub_epi8(vx, voffset);
83 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable7, vx));
84 vx = _mm512_sub_epi8(vx, voffset);
85 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable8, vx));
86
87 vx = _mm512_subs_epi8(vx, voffset);
88 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable9, vx));
89 vx = _mm512_subs_epi8(vx, voffset);
90 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableA, vx));
91 vx = _mm512_subs_epi8(vx, voffset);
92 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableB, vx));
93 vx = _mm512_subs_epi8(vx, voffset);
94 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableC, vx));
95 vx = _mm512_subs_epi8(vx, voffset);
96 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableD, vx));
97 vx = _mm512_subs_epi8(vx, voffset);
98 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableE, vx));
99 vx = _mm512_subs_epi8(vx, voffset);
100 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableF, vx));
101
102 _mm512_storeu_si512(y, vy);
103 y += 64;
104 }
105 if XNN_UNLIKELY(n != 0) {
106 assert(n < 64);
107 const __mmask64 vmask = _cvtu64_mask64((uint64_t) ((UINT64_C(1) << n) - UINT64_C(1)));
108
109 __m512i vx = _mm512_maskz_loadu_epi8(vmask, x);
110
111 __m512i vy = _mm512_shuffle_epi8(vtable0, vx);
112
113 vx = _mm512_sub_epi8(vx, voffset);
114 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable1, vx));
115 vx = _mm512_sub_epi8(vx, voffset);
116 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable2, vx));
117 vx = _mm512_sub_epi8(vx, voffset);
118 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable3, vx));
119 vx = _mm512_sub_epi8(vx, voffset);
120 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable4, vx));
121 vx = _mm512_sub_epi8(vx, voffset);
122 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable5, vx));
123 vx = _mm512_sub_epi8(vx, voffset);
124 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable6, vx));
125 vx = _mm512_sub_epi8(vx, voffset);
126 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable7, vx));
127 vx = _mm512_sub_epi8(vx, voffset);
128 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable8, vx));
129
130 vx = _mm512_subs_epi8(vx, voffset);
131 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtable9, vx));
132 vx = _mm512_subs_epi8(vx, voffset);
133 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableA, vx));
134 vx = _mm512_subs_epi8(vx, voffset);
135 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableB, vx));
136 vx = _mm512_subs_epi8(vx, voffset);
137 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableC, vx));
138 vx = _mm512_subs_epi8(vx, voffset);
139 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableD, vx));
140 vx = _mm512_subs_epi8(vx, voffset);
141 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableE, vx));
142 vx = _mm512_subs_epi8(vx, voffset);
143 vy = _mm512_xor_si512(vy, _mm512_shuffle_epi8(vtableF, vx));
144
145 _mm512_mask_storeu_epi8(y, vmask, vy);
146 }
147 }
148