1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
|
/*
* Copyright (c) 2019-2020 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*/
#ifdef __aarch64__
template<>
void interleave_block<8, 1, VLType::None, false>(
float * &out_ptr, const bfloat16 * const * in, size_t width, size_t height,
size_t row_offset, bool
)
{
__asm__ __volatile__(
"movi v29.8h, #0x0\n"
"ldr x27, [%x[in], #0x0]\n"
"cmp %x[height], #0x8\n"
"ldr x26, [%x[in], #0x8]\n"
"add x27, x27, %x[row_offset], LSL #1\n"
"ldr x25, [%x[in], #0x10]\n"
"ldr x24, [%x[in], #0x18]\n"
"add x26, x26, %x[row_offset], LSL #1\n"
"ldr x23, [%x[in], #0x20]\n"
"add x25, x25, %x[row_offset], LSL #1\n"
"ldr x22, [%x[in], #0x28]\n"
"ldr x21, [%x[in], #0x30]\n"
"add x24, x24, %x[row_offset], LSL #1\n"
"ldr x20, [%x[in], #0x38]\n"
"add x23, x23, %x[row_offset], LSL #1\n"
"add x22, x22, %x[row_offset], LSL #1\n"
"add x21, x21, %x[row_offset], LSL #1\n"
"add x20, x20, %x[row_offset], LSL #1\n"
"beq 1f\n"
"mov x20, x27\n"
"cmp %x[height], #0x2\n"
"csel x26, x26, x27, GE\n"
"csel x25, x25, x27, GT\n"
"cmp %x[height], #0x4\n"
"csel x24, x24, x27, GE\n"
"csel x23, x23, x27, GT\n"
"cmp %x[height], #0x6\n"
"csel x22, x22, x27, GE\n"
"csel x21, x21, x27, GT\n"
"1:" // no_pointer_adj
"prfm pldl1keep, [x27, #0x0]\n"
"cmp %x[width], #0x4\n"
"prfm pldl1keep, [x26, #0x0]\n"
"prfm pldl1keep, [x25, #0x0]\n"
"prfm pldl1keep, [x24, #0x0]\n"
"prfm pldl1keep, [x23, #0x0]\n"
"prfm pldl1keep, [x22, #0x0]\n"
"prfm pldl1keep, [x21, #0x0]\n"
"prfm pldl1keep, [x20, #0x0]\n"
"prfm pldl1keep, [x27, #0x40]\n"
"prfm pldl1keep, [x26, #0x40]\n"
"prfm pldl1keep, [x25, #0x40]\n"
"prfm pldl1keep, [x24, #0x40]\n"
"prfm pldl1keep, [x23, #0x40]\n"
"prfm pldl1keep, [x22, #0x40]\n"
"prfm pldl1keep, [x21, #0x40]\n"
"prfm pldl1keep, [x20, #0x40]\n"
"blt 3f\n"
"2:" // Main loop head
"ldr d28, [x27], #0x8\n"
"zip1 v28.8h, v29.8h, v28.8h\n"
"prfm pldl1keep, [x27, #0x70]\n"
"ldr d27, [x26], #0x8\n"
"zip1 v27.8h, v29.8h, v27.8h\n"
"prfm pldl1keep, [x26, #0x70]\n"
"ldr d26, [x25], #0x8\n"
"zip1 v26.8h, v29.8h, v26.8h\n"
"prfm pldl1keep, [x25, #0x70]\n"
"ldr d25, [x24], #0x8\n"
"zip1 v20.4s, v28.4s, v26.4s\n"
"prfm pldl1keep, [x24, #0x70]\n"
"zip1 v25.8h, v29.8h, v25.8h\n"
"ldr d24, [x23], #0x8\n"
"zip1 v19.4s, v27.4s, v25.4s\n"
"prfm pldl1keep, [x23, #0x70]\n"
"zip1 v24.8h, v29.8h, v24.8h\n"
"ldr d23, [x22], #0x8\n"
"zip1 v16.4s, v20.4s, v19.4s\n"
"prfm pldl1keep, [x22, #0x70]\n"
"zip1 v23.8h, v29.8h, v23.8h\n"
"ldr d22, [x21], #0x8\n"
"zip2 v19.4s, v20.4s, v19.4s\n"
"prfm pldl1keep, [x21, #0x70]\n"
"zip1 v22.8h, v29.8h, v22.8h\n"
"ldr d21, [x20], #0x8\n"
"zip1 v18.4s, v24.4s, v22.4s\n"
"prfm pldl1keep, [x20, #0x70]\n"
"zip1 v21.8h, v29.8h, v21.8h\n"
"str q16, [%x[out_ptr], #0x0]\n"
"zip1 v17.4s, v23.4s, v21.4s\n"
"subs %x[width], %x[width], #0x4\n"
"zip2 v20.4s, v28.4s, v26.4s\n"
"cmp %x[width], #0x4\n"
"zip1 v16.4s, v18.4s, v17.4s\n"
"str q16, [%x[out_ptr], #0x10]\n"
"zip2 v16.4s, v18.4s, v17.4s\n"
"str q19, [%x[out_ptr], #0x20]\n"
"zip2 v19.4s, v27.4s, v25.4s\n"
"str q16, [%x[out_ptr], #0x30]\n"
"zip1 v16.4s, v20.4s, v19.4s\n"
"str q16, [%x[out_ptr], #0x40]\n"
"zip2 v18.4s, v24.4s, v22.4s\n"
"zip2 v17.4s, v23.4s, v21.4s\n"
"zip1 v16.4s, v18.4s, v17.4s\n"
"str q16, [%x[out_ptr], #0x50]\n"
"zip2 v16.4s, v20.4s, v19.4s\n"
"str q16, [%x[out_ptr], #0x60]\n"
"zip2 v16.4s, v18.4s, v17.4s\n"
"str q16, [%x[out_ptr], #0x70]\n"
"add %x[out_ptr], %x[out_ptr], #0x80\n"
"bge 2b\n"
"3:" // Main loop skip
"cbz %x[width], 6f\n"
"tbz %x[width], #1, 4f\n"
"ldr s28, [x27], #0x4\n"
"ldr s27, [x26], #0x4\n"
"ldr s26, [x25], #0x4\n"
"ldr s25, [x24], #0x4\n"
"ldr s24, [x23], #0x4\n"
"ldr s23, [x22], #0x4\n"
"ldr s22, [x21], #0x4\n"
"ldr s21, [x20], #0x4\n"
"mov x19, #0x2\n"
"tbz %x[width], #0, 5f\n"
"ld1 { v28.h }[2], [x27]\n"
"ld1 { v27.h }[2], [x26]\n"
"ld1 { v26.h }[2], [x25]\n"
"ld1 { v25.h }[2], [x24]\n"
"ld1 { v24.h }[2], [x23]\n"
"ld1 { v23.h }[2], [x22]\n"
"ld1 { v22.h }[2], [x21]\n"
"ld1 { v21.h }[2], [x20]\n"
"mov x19, #0x3\n"
"b 5f\n"
"4:" // odd_loads_1_0
"ldr h28, [x27, #0x0]\n"
"ldr h27, [x26, #0x0]\n"
"ldr h26, [x25, #0x0]\n"
"ldr h25, [x24, #0x0]\n"
"ldr h24, [x23, #0x0]\n"
"ldr h23, [x22, #0x0]\n"
"ldr h22, [x21, #0x0]\n"
"ldr h21, [x20, #0x0]\n"
"mov x19, #0x1\n"
"5:" // Odd load end
"zip1 v28.8h, v29.8h, v28.8h\n"
"subs x19, x19, #0x1\n"
"zip1 v27.8h, v29.8h, v27.8h\n"
"zip1 v26.8h, v29.8h, v26.8h\n"
"zip1 v25.8h, v29.8h, v25.8h\n"
"zip1 v24.8h, v29.8h, v24.8h\n"
"zip1 v23.8h, v29.8h, v23.8h\n"
"zip1 v22.8h, v29.8h, v22.8h\n"
"zip1 v21.8h, v29.8h, v21.8h\n"
"zip1 v20.4s, v28.4s, v26.4s\n"
"zip1 v19.4s, v27.4s, v25.4s\n"
"zip1 v16.4s, v20.4s, v19.4s\n"
"str q16, [%x[out_ptr], #0x0]\n"
"zip1 v18.4s, v24.4s, v22.4s\n"
"zip1 v17.4s, v23.4s, v21.4s\n"
"zip1 v16.4s, v18.4s, v17.4s\n"
"str q16, [%x[out_ptr], #0x10]\n"
"add %x[out_ptr], %x[out_ptr], #0x20\n"
"beq 6f\n"
"zip2 v19.4s, v20.4s, v19.4s\n"
"zip2 v16.4s, v18.4s, v17.4s\n"
"str q19, [%x[out_ptr], #0x0]\n"
"str q16, [%x[out_ptr], #0x10]\n"
"subs x19, x19, #0x1\n"
"add %x[out_ptr], %x[out_ptr], #0x20\n"
"beq 6f\n"
"zip2 v20.4s, v28.4s, v26.4s\n"
"zip2 v19.4s, v27.4s, v25.4s\n"
"zip1 v16.4s, v20.4s, v19.4s\n"
"str q16, [%x[out_ptr], #0x0]\n"
"zip2 v18.4s, v24.4s, v22.4s\n"
"zip2 v17.4s, v23.4s, v21.4s\n"
"zip1 v16.4s, v18.4s, v17.4s\n"
"str q16, [%x[out_ptr], #0x10]\n"
"add %x[out_ptr], %x[out_ptr], #0x20\n"
"6:" // Odds skip
: [out_ptr] "+r" (out_ptr), [width] "+r" (width)
: [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset)
: "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27"
);
}
#endif // __aarch64__
|
