xref: /aosp_15_r20/external/XNNPACK/src/x8-lut/gen/lut-ssse3-x16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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/common.h>
15 #include <xnnpack/lut.h>
16 #include <xnnpack/unaligned.h>
17 
18 
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)])19 void xnn_x8_lut_ukernel__ssse3_x16(
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 __m128i vt0 = _mm_load_si128((const __m128i*) t);
30   const __m128i vt1 = _mm_load_si128((const __m128i*) (t + 16));
31   const __m128i vt2 = _mm_load_si128((const __m128i*) (t + 32));
32   const __m128i vt3 = _mm_load_si128((const __m128i*) (t + 48));
33   const __m128i vt4 = _mm_load_si128((const __m128i*) (t + 64));
34   const __m128i vt5 = _mm_load_si128((const __m128i*) (t + 80));
35   const __m128i vt6 = _mm_load_si128((const __m128i*) (t + 96));
36   const __m128i vt7 = _mm_load_si128((const __m128i*) (t + 112));
37   const __m128i vt8 = _mm_load_si128((const __m128i*) (t + 128));
38   const __m128i vt9 = _mm_load_si128((const __m128i*) (t + 144));
39   const __m128i vtA = _mm_load_si128((const __m128i*) (t + 160));
40   const __m128i vtB = _mm_load_si128((const __m128i*) (t + 176));
41   const __m128i vtC = _mm_load_si128((const __m128i*) (t + 192));
42   const __m128i vtD = _mm_load_si128((const __m128i*) (t + 208));
43   const __m128i vtE = _mm_load_si128((const __m128i*) (t + 224));
44   const __m128i vtF = _mm_load_si128((const __m128i*) (t + 240));
45 
46   const __m128i vtable0 = vt0;
47   const __m128i vtable1 = _mm_xor_si128(vt0, vt1);
48   const __m128i vtable2 = _mm_xor_si128(vt1, vt2);
49   const __m128i vtable3 = _mm_xor_si128(vt2, vt3);
50   const __m128i vtable4 = _mm_xor_si128(vt3, vt4);
51   const __m128i vtable5 = _mm_xor_si128(vt4, vt5);
52   const __m128i vtable6 = _mm_xor_si128(vt5, vt6);
53   const __m128i vtable7 = _mm_xor_si128(vt6, vt7);
54   const __m128i vtable8 = _mm_xor_si128(_mm_xor_si128(vt7, vt8), vtable0);
55   const __m128i vtable9 = _mm_xor_si128(_mm_xor_si128(vt8, vt9), vtable1);
56   const __m128i vtableA = _mm_xor_si128(_mm_xor_si128(vt9, vtA), vtable2);
57   const __m128i vtableB = _mm_xor_si128(_mm_xor_si128(vtA, vtB), vtable3);
58   const __m128i vtableC = _mm_xor_si128(_mm_xor_si128(vtB, vtC), vtable4);
59   const __m128i vtableD = _mm_xor_si128(_mm_xor_si128(vtC, vtD), vtable5);
60   const __m128i vtableE = _mm_xor_si128(_mm_xor_si128(vtD, vtE), vtable6);
61   const __m128i vtableF = _mm_xor_si128(_mm_xor_si128(vtE, vtF), vtable7);
62 
63   const __m128i voffset = _mm_set1_epi8(16);
64   for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
65     __m128i vx = _mm_loadu_si128((const __m128i*) x);
66     x += 16;
67 
68     __m128i vy = _mm_shuffle_epi8(vtable0, vx);
69 
70     vx = _mm_sub_epi8(vx, voffset);
71     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable1, vx));
72     vx = _mm_sub_epi8(vx, voffset);
73     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable2, vx));
74     vx = _mm_sub_epi8(vx, voffset);
75     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable3, vx));
76     vx = _mm_sub_epi8(vx, voffset);
77     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable4, vx));
78     vx = _mm_sub_epi8(vx, voffset);
79     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable5, vx));
80     vx = _mm_sub_epi8(vx, voffset);
81     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable6, vx));
82     vx = _mm_sub_epi8(vx, voffset);
83     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable7, vx));
84     vx = _mm_sub_epi8(vx, voffset);
85     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable8, vx));
86 
87     vx = _mm_subs_epi8(vx, voffset);
88     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable9, vx));
89     vx = _mm_subs_epi8(vx, voffset);
90     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableA, vx));
91     vx = _mm_subs_epi8(vx, voffset);
92     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableB, vx));
93     vx = _mm_subs_epi8(vx, voffset);
94     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableC, vx));
95     vx = _mm_subs_epi8(vx, voffset);
96     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableD, vx));
97     vx = _mm_subs_epi8(vx, voffset);
98     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableE, vx));
99     vx = _mm_subs_epi8(vx, voffset);
100     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableF, vx));
101 
102     _mm_storeu_si128((__m128i*) y, vy);
103     y += 16;
104   }
105   if XNN_UNLIKELY(n != 0) {
106     __m128i vx = _mm_loadu_si128((const __m128i*) x);
107 
108     __m128i vy = _mm_shuffle_epi8(vtable0, vx);
109 
110     vx = _mm_sub_epi8(vx, voffset);
111     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable1, vx));
112     vx = _mm_sub_epi8(vx, voffset);
113     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable2, vx));
114     vx = _mm_sub_epi8(vx, voffset);
115     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable3, vx));
116     vx = _mm_sub_epi8(vx, voffset);
117     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable4, vx));
118     vx = _mm_sub_epi8(vx, voffset);
119     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable5, vx));
120     vx = _mm_sub_epi8(vx, voffset);
121     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable6, vx));
122     vx = _mm_sub_epi8(vx, voffset);
123     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable7, vx));
124     vx = _mm_sub_epi8(vx, voffset);
125     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable8, vx));
126 
127     vx = _mm_subs_epi8(vx, voffset);
128     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable9, vx));
129     vx = _mm_subs_epi8(vx, voffset);
130     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableA, vx));
131     vx = _mm_subs_epi8(vx, voffset);
132     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableB, vx));
133     vx = _mm_subs_epi8(vx, voffset);
134     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableC, vx));
135     vx = _mm_subs_epi8(vx, voffset);
136     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableD, vx));
137     vx = _mm_subs_epi8(vx, voffset);
138     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableE, vx));
139     vx = _mm_subs_epi8(vx, voffset);
140     vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtableF, vx));
141 
142     if (n & (8 * sizeof(uint8_t))) {
143       _mm_storel_epi64((__m128i*) y, vy);
144       vy = _mm_unpackhi_epi64(vy, vy);
145       y += 8;
146     }
147     if (n & (4 * sizeof(uint8_t))) {
148       unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy));
149       vy = _mm_srli_epi64(vy, 32);
150       y += 4;
151     }
152     uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
153     if (n & (2 * sizeof(uint8_t))) {
154       unaligned_store_u16(y, (uint16_t) vy_lo);
155       vy_lo >>= 16;
156       y += 2;
157     }
158     if (n & (1 * sizeof(uint8_t))) {
159       *y = (uint8_t) vy_lo;
160     }
161   }
162 }
163