1 // Auto-generated file. Do not edit!
2 // Template: src/x8-lut/ssse3.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 <tmmintrin.h>
13
14 #include <xnnpack/lut.h>
15 #include <xnnpack/common.h>
16
17
xnn_x8_lut_ukernel__ssse3_x16(size_t n,const uint8_t * x,uint8_t * y,const uint8_t t[restrict XNN_MIN_ELEMENTS (256)])18 void xnn_x8_lut_ukernel__ssse3_x16(
19 size_t n,
20 const uint8_t* x,
21 uint8_t* y,
22 const uint8_t t[restrict XNN_MIN_ELEMENTS(256)])
23 {
24 assert(n != 0);
25 assert(x != NULL);
26 assert(y != NULL);
27
28 const __m128i vt0 = _mm_load_si128((const __m128i*) t);
29 const __m128i vt1 = _mm_load_si128((const __m128i*) (t + 16));
30 const __m128i vt2 = _mm_load_si128((const __m128i*) (t + 32));
31 const __m128i vt3 = _mm_load_si128((const __m128i*) (t + 48));
32 const __m128i vt4 = _mm_load_si128((const __m128i*) (t + 64));
33 const __m128i vt5 = _mm_load_si128((const __m128i*) (t + 80));
34 const __m128i vt6 = _mm_load_si128((const __m128i*) (t + 96));
35 const __m128i vt7 = _mm_load_si128((const __m128i*) (t + 112));
36 const __m128i vt8 = _mm_load_si128((const __m128i*) (t + 128));
37 const __m128i vt9 = _mm_load_si128((const __m128i*) (t + 144));
38 const __m128i vtA = _mm_load_si128((const __m128i*) (t + 160));
39 const __m128i vtB = _mm_load_si128((const __m128i*) (t + 176));
40 const __m128i vtC = _mm_load_si128((const __m128i*) (t + 192));
41 const __m128i vtD = _mm_load_si128((const __m128i*) (t + 208));
42 const __m128i vtE = _mm_load_si128((const __m128i*) (t + 224));
43 const __m128i vtF = _mm_load_si128((const __m128i*) (t + 240));
44
45 const __m128i vtable0 = vt0;
46 const __m128i vtable1 = _mm_xor_si128(vt0, vt1);
47 const __m128i vtable2 = _mm_xor_si128(vt1, vt2);
48 const __m128i vtable3 = _mm_xor_si128(vt2, vt3);
49 const __m128i vtable4 = _mm_xor_si128(vt3, vt4);
50 const __m128i vtable5 = _mm_xor_si128(vt4, vt5);
51 const __m128i vtable6 = _mm_xor_si128(vt5, vt6);
52 const __m128i vtable7 = _mm_xor_si128(vt6, vt7);
53 const __m128i vtable8 = _mm_xor_si128(_mm_xor_si128(vt7, vt8), vtable0);
54 const __m128i vtable9 = _mm_xor_si128(_mm_xor_si128(vt8, vt9), vtable1);
55 const __m128i vtableA = _mm_xor_si128(_mm_xor_si128(vt9, vtA), vtable2);
56 const __m128i vtableB = _mm_xor_si128(_mm_xor_si128(vtA, vtB), vtable3);
57 const __m128i vtableC = _mm_xor_si128(_mm_xor_si128(vtB, vtC), vtable4);
58 const __m128i vtableD = _mm_xor_si128(_mm_xor_si128(vtC, vtD), vtable5);
59 const __m128i vtableE = _mm_xor_si128(_mm_xor_si128(vtD, vtE), vtable6);
60 const __m128i vtableF = _mm_xor_si128(_mm_xor_si128(vtE, vtF), vtable7);
61
62 const __m128i voffset = _mm_set1_epi8(16);
63 for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
64 __m128i vx = _mm_loadu_si128((const __m128i*) x);
65 x += 16;
66
67 __m128i vy = _mm_shuffle_epi8(vtable0, vx);
68
69 vx = _mm_sub_epi8(vx, voffset);
70 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable1, vx));
71 vx = _mm_sub_epi8(vx, voffset);
72 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable2, vx));
73 vx = _mm_sub_epi8(vx, voffset);
74 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable3, vx));
75 vx = _mm_sub_epi8(vx, voffset);
76 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable4, vx));
77 vx = _mm_sub_epi8(vx, voffset);
78 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable5, vx));
79 vx = _mm_sub_epi8(vx, voffset);
80 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable6, vx));
81 vx = _mm_sub_epi8(vx, voffset);
82 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable7, vx));
83 vx = _mm_sub_epi8(vx, voffset);
84 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable8, vx));
85
86 vx = _mm_subs_epi8(vx, voffset);
87 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable9, vx));
88 vx = _mm_subs_epi8(vx, voffset);
89 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableA, vx));
90 vx = _mm_subs_epi8(vx, voffset);
91 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableB, vx));
92 vx = _mm_subs_epi8(vx, voffset);
93 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableC, vx));
94 vx = _mm_subs_epi8(vx, voffset);
95 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableD, vx));
96 vx = _mm_subs_epi8(vx, voffset);
97 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableE, vx));
98 vx = _mm_subs_epi8(vx, voffset);
99 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableF, vx));
100
101 _mm_storeu_si128((__m128i*) y, vy);
102 y += 16;
103 }
104 if XNN_UNLIKELY(n != 0) {
105 __m128i vx = _mm_loadu_si128((const __m128i*) x);
106
107 __m128i vy = _mm_shuffle_epi8(vtable0, vx);
108
109 vx = _mm_sub_epi8(vx, voffset);
110 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable1, vx));
111 vx = _mm_sub_epi8(vx, voffset);
112 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable2, vx));
113 vx = _mm_sub_epi8(vx, voffset);
114 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable3, vx));
115 vx = _mm_sub_epi8(vx, voffset);
116 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable4, vx));
117 vx = _mm_sub_epi8(vx, voffset);
118 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable5, vx));
119 vx = _mm_sub_epi8(vx, voffset);
120 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable6, vx));
121 vx = _mm_sub_epi8(vx, voffset);
122 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable7, vx));
123 vx = _mm_sub_epi8(vx, voffset);
124 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable8, vx));
125
126 vx = _mm_subs_epi8(vx, voffset);
127 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable9, vx));
128 vx = _mm_subs_epi8(vx, voffset);
129 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableA, vx));
130 vx = _mm_subs_epi8(vx, voffset);
131 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableB, vx));
132 vx = _mm_subs_epi8(vx, voffset);
133 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableC, vx));
134 vx = _mm_subs_epi8(vx, voffset);
135 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableD, vx));
136 vx = _mm_subs_epi8(vx, voffset);
137 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableE, vx));
138 vx = _mm_subs_epi8(vx, voffset);
139 vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableF, vx));
140
141 if (n & (8 * sizeof(uint8_t))) {
142 _mm_storel_epi64((__m128i*) y, vy);
143 vy = _mm_unpackhi_epi64(vy, vy);
144 y += 8;
145 }
146 if (n & (4 * sizeof(uint8_t))) {
147 *((uint32_t*) y) = (uint32_t) _mm_cvtsi128_si32(vy);
148 vy = _mm_srli_epi64(vy, 32);
149 y += 4;
150 }
151 uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
152 if (n & (2 * sizeof(uint8_t))) {
153 *((uint16_t*) y) = (uint16_t) vy_lo;
154 vy_lo >>= 16;
155 y += 2;
156 }
157 if (n & (1 * sizeof(uint8_t))) {
158 *y = (uint8_t) vy_lo;
159 }
160 }
161 }
162