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
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
|
/*
* Copyright (c) 2021-2023 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.
*/
#include <cstdint>
#include <cstddef>
#if defined(__aarch64__)
namespace arm_conv {
namespace pooling {
void a64_u8_nhwc_max_generic_depthfirst_impl(
const uint64_t,
const uint64_t n_valid_cells,
uint64_t n_channels,
const uint8_t *const *const inptrs,
uint8_t *outptr
)
{
__asm__ __volatile__(
"cmp %x[n_channels], #0x40\n"
"mov x27, #0x0\n"
"mov x26, #0x10\n" // cntb _, ALL, #1
"mov x24, #0x20\n" // cntb _, ALL, #2
"mov x23, #0x30\n" // cntb _, ALL, #3
"blt 7f\n"
"1:" // 4-vectors of channels
"lsr x25, %x[n_valid_cells], #0x2\n"
"movi v8.16b, #0x0\n"
"movi v7.16b, #0x0\n"
"mov x22, %x[inptrs]\n"
"movi v6.16b, #0x0\n"
"movi v5.16b, #0x0\n"
"cbz x25, 4f\n"
"ldp x21, x20, [x22, #0x0]\n"
"ldr q4, [x21, x27]\n"
"subs x25, x25, #0x1\n"
"ldr q3, [x20, x27]\n"
"ldr q2, [x21, x26]\n"
"ldr q1, [x20, x26]\n"
"ldr q0, [x21, x24]\n"
"ldr q31, [x20, x24]\n"
"ldr q30, [x21, x23]\n"
"ldr q29, [x20, x23]\n"
"ldp x21, x20, [x22, #0x10]\n"
"add x22, x22, #0x20\n"
"ldr q28, [x21, x27]\n"
"ldr q22, [x20, x27]\n"
"ldr q27, [x21, x26]\n"
"ldr q21, [x20, x26]\n"
"ldr q26, [x21, x24]\n"
"ldr q20, [x20, x24]\n"
"ldr q25, [x21, x23]\n"
"ldr q24, [x20, x23]\n"
"beq 3f\n"
"2:" // 4-vectors of channels: 4 inputs loop
"umax v23.16b, v4.16b, v3.16b\n"
"umax v19.16b, v28.16b, v22.16b\n"
"ldp x21, x20, [x22, #0x0]\n"
"ldr q4, [x21, x27]\n"
"ldr q3, [x20, x27]\n"
"umax v22.16b, v2.16b, v1.16b\n"
"ldr q2, [x21, x26]\n"
"umax v18.16b, v27.16b, v21.16b\n"
"ldr q1, [x20, x26]\n"
"umax v21.16b, v0.16b, v31.16b\n"
"ldr q0, [x21, x24]\n"
"umax v17.16b, v26.16b, v20.16b\n"
"ldr q31, [x20, x24]\n"
"umax v20.16b, v30.16b, v29.16b\n"
"ldr q30, [x21, x23]\n"
"umax v16.16b, v25.16b, v24.16b\n"
"ldr q29, [x20, x23]\n"
"umax v19.16b, v23.16b, v19.16b\n"
"umax v18.16b, v22.16b, v18.16b\n"
"ldp x21, x20, [x22, #0x10]\n"
"ldr q28, [x21, x27]\n"
"ldr q22, [x20, x27]\n"
"umax v17.16b, v21.16b, v17.16b\n"
"umax v16.16b, v20.16b, v16.16b\n"
"ldr q27, [x21, x26]\n"
"ldr q21, [x20, x26]\n"
"subs x25, x25, #0x1\n"
"umax v8.16b, v8.16b, v19.16b\n"
"ldr q26, [x21, x24]\n"
"ldr q20, [x20, x24]\n"
"umax v7.16b, v7.16b, v18.16b\n"
"umax v6.16b, v6.16b, v17.16b\n"
"ldr q25, [x21, x23]\n"
"ldr q24, [x20, x23]\n"
"umax v5.16b, v5.16b, v16.16b\n"
"add x22, x22, #0x20\n"
"bgt 2b\n"
"3:" // 4-vectors of channels: 4 inputs tail
"umax v23.16b, v4.16b, v3.16b\n"
"umax v19.16b, v28.16b, v22.16b\n"
"umax v22.16b, v2.16b, v1.16b\n"
"umax v18.16b, v27.16b, v21.16b\n"
"umax v21.16b, v0.16b, v31.16b\n"
"umax v17.16b, v26.16b, v20.16b\n"
"umax v20.16b, v30.16b, v29.16b\n"
"umax v16.16b, v25.16b, v24.16b\n"
"umax v19.16b, v23.16b, v19.16b\n"
"umax v18.16b, v22.16b, v18.16b\n"
"umax v17.16b, v21.16b, v17.16b\n"
"umax v16.16b, v20.16b, v16.16b\n"
"umax v8.16b, v8.16b, v19.16b\n"
"umax v7.16b, v7.16b, v18.16b\n"
"umax v6.16b, v6.16b, v17.16b\n"
"umax v5.16b, v5.16b, v16.16b\n"
"4:" // 4-vectors of channels: After loop
"ands x21, %x[n_valid_cells], #0x3\n"
"beq 6f\n"
"5:" // 4-vectors of channels: Single input loop
"ldr x20, [x22], #0x8\n"
"ldr q16, [x20, x27]\n"
"subs x21, x21, #0x1\n"
"umax v8.16b, v8.16b, v16.16b\n"
"ldr q17, [x20, x26]\n"
"ldr q16, [x20, x24]\n"
"umax v7.16b, v7.16b, v17.16b\n"
"umax v6.16b, v6.16b, v16.16b\n"
"ldr q16, [x20, x23]\n"
"umax v5.16b, v5.16b, v16.16b\n"
"bgt 5b\n"
"6:" // 4-vectors of channels: Single input loop: End
"sub %x[n_channels], %x[n_channels], #0x40\n"
"cmp %x[n_channels], #0x40\n"
"str q8, [%x[outptr], x27]\n"
"str q7, [%x[outptr], x26]\n"
"add x27, x27, #0x40\n"
"add x26, x26, #0x40\n"
"str q6, [%x[outptr], x24]\n"
"add x24, x24, #0x40\n"
"str q5, [%x[outptr], x23]\n"
"add x23, x23, #0x40\n"
"bge 1b\n"
"cbz %x[n_channels], 43f\n"
"7:" // Single vector of channels
"cmp %x[n_channels], #0x10\n"
"blt 14f\n"
"8:" // Single vector of channels: Loop
"lsr x25, %x[n_valid_cells], #0x2\n"
"movi v8.16b, #0x0\n"
"mov x22, %x[inptrs]\n"
"cbz x25, 11f\n"
"ldp x21, x20, [x22, #0x0]\n"
"ldr q4, [x21, x27]\n"
"subs x25, x25, #0x1\n"
"ldr q3, [x20, x27]\n"
"ldp x21, x20, [x22, #0x10]\n"
"add x22, x22, #0x20\n"
"ldr q28, [x21, x27]\n"
"ldr q22, [x20, x27]\n"
"beq 10f\n"
"9:" // Single vector of channels: Loop: 4 inputs loop
"umax v17.16b, v4.16b, v3.16b\n"
"umax v16.16b, v28.16b, v22.16b\n"
"ldp x21, x20, [x22, #0x0]\n"
"ldr q4, [x21, x27]\n"
"ldr q3, [x20, x27]\n"
"umax v16.16b, v17.16b, v16.16b\n"
"ldp x21, x20, [x22, #0x10]\n"
"subs x25, x25, #0x1\n"
"ldr q28, [x21, x27]\n"
"ldr q22, [x20, x27]\n"
"umax v8.16b, v8.16b, v16.16b\n"
"add x22, x22, #0x20\n"
"bgt 9b\n"
"10:" // Single vector of channels: Loop: 4 inputs tail
"umax v17.16b, v4.16b, v3.16b\n"
"umax v16.16b, v28.16b, v22.16b\n"
"umax v16.16b, v17.16b, v16.16b\n"
"umax v8.16b, v8.16b, v16.16b\n"
"11:" // Single vector of channels: Loop: After loop
"ands x21, %x[n_valid_cells], #0x3\n"
"beq 13f\n"
"12:" // Single vector of channels: Loop: Single input loop
"ldr x20, [x22], #0x8\n"
"ldr q16, [x20, x27]\n"
"subs x21, x21, #0x1\n"
"umax v8.16b, v8.16b, v16.16b\n"
"bgt 12b\n"
"13:" // Single vector of channels: Loop: Single input loop: End
"sub %x[n_channels], %x[n_channels], #0x10\n"
"cmp %x[n_channels], #0x10\n"
"str q8, [%x[outptr], x27]\n"
"add x27, x27, #0x10\n"
"bge 8b\n"
"cbz %x[n_channels], 43f\n"
"14:" // Oddments
"lsr x25, %x[n_valid_cells], #0x2\n"
"add %x[outptr], %x[outptr], x27\n"
"movi v8.16b, #0x0\n"
"mov x24, %x[inptrs]\n"
"cbz x25, 24f\n"
"15:" // Oddments: 4 inputs loop
"ldp x23, x22, [x24, #0x0]\n"
"ldp x21, x20, [x24, #0x10]\n"
"add x24, x24, #0x20\n"
"add x23, x23, x27\n"
"add x22, x22, x27\n"
"add x21, x21, x27\n"
"movi v4.16b, #0x0\n"
"movi v3.16b, #0x0\n"
"add x20, x20, x27\n"
"movi v28.16b, #0x0\n"
"movi v22.16b, #0x0\n"
"tbz %x[n_channels], #3, 19f\n"
"ldr d4, [x23], #0x8\n"
"ldr d3, [x22], #0x8\n"
"ldr d28, [x21], #0x8\n"
"ldr d22, [x20], #0x8\n"
"tbz %x[n_channels], #2, 17f\n"
"ld1 { v4.s }[2], [x23], #0x4\n"
"ld1 { v3.s }[2], [x22], #0x4\n"
"ld1 { v28.s }[2], [x21], #0x4\n"
"ld1 { v22.s }[2], [x20], #0x4\n"
"tbz %x[n_channels], #1, 16f\n"
"ld1 { v4.h }[6], [x23], #0x2\n"
"ld1 { v3.h }[6], [x22], #0x2\n"
"ld1 { v28.h }[6], [x21], #0x2\n"
"ld1 { v22.h }[6], [x20], #0x2\n"
"tbz %x[n_channels], #0, 23f\n"
"ld1 { v4.b }[14], [x23], #0x1\n"
"ld1 { v3.b }[14], [x22], #0x1\n"
"ld1 { v28.b }[14], [x21], #0x1\n"
"ld1 { v22.b }[14], [x20], #0x1\n"
"b 23f\n"
"16:" // Oddments: 4 inputs loop: Load: Bit 3: Bit 2: Bit 1: Unset
"tbz %x[n_channels], #0, 23f\n"
"ld1 { v4.b }[12], [x23], #0x1\n"
"ld1 { v3.b }[12], [x22], #0x1\n"
"ld1 { v28.b }[12], [x21], #0x1\n"
"ld1 { v22.b }[12], [x20], #0x1\n"
"b 23f\n"
"17:" // Oddments: 4 inputs loop: Load: Bit 3: Bit 2: Unset
"tbz %x[n_channels], #1, 18f\n"
"ld1 { v4.h }[4], [x23], #0x2\n"
"ld1 { v3.h }[4], [x22], #0x2\n"
"ld1 { v28.h }[4], [x21], #0x2\n"
"ld1 { v22.h }[4], [x20], #0x2\n"
"tbz %x[n_channels], #0, 23f\n"
"ld1 { v4.b }[10], [x23], #0x1\n"
"ld1 { v3.b }[10], [x22], #0x1\n"
"ld1 { v28.b }[10], [x21], #0x1\n"
"ld1 { v22.b }[10], [x20], #0x1\n"
"b 23f\n"
"18:" // Oddments: 4 inputs loop: Load: Bit 3: Bit 2: Unset: Bit 1: Unset
"tbz %x[n_channels], #0, 23f\n"
"ld1 { v4.b }[8], [x23], #0x1\n"
"ld1 { v3.b }[8], [x22], #0x1\n"
"ld1 { v28.b }[8], [x21], #0x1\n"
"ld1 { v22.b }[8], [x20], #0x1\n"
"b 23f\n"
"19:" // Oddments: 4 inputs loop: Load: Bit 3: Unset
"tbz %x[n_channels], #2, 21f\n"
"ldr s4, [x23], #0x4\n"
"ldr s3, [x22], #0x4\n"
"ldr s28, [x21], #0x4\n"
"ldr s22, [x20], #0x4\n"
"tbz %x[n_channels], #1, 20f\n"
"ld1 { v4.h }[2], [x23], #0x2\n"
"ld1 { v3.h }[2], [x22], #0x2\n"
"ld1 { v28.h }[2], [x21], #0x2\n"
"ld1 { v22.h }[2], [x20], #0x2\n"
"tbz %x[n_channels], #0, 23f\n"
"ld1 { v4.b }[6], [x23], #0x1\n"
"ld1 { v3.b }[6], [x22], #0x1\n"
"ld1 { v28.b }[6], [x21], #0x1\n"
"ld1 { v22.b }[6], [x20], #0x1\n"
"b 23f\n"
"20:" // Oddments: 4 inputs loop: Load: Bit 3: Unset: Bit 2: Bit 1: Unset
"tbz %x[n_channels], #0, 23f\n"
"ld1 { v4.b }[4], [x23], #0x1\n"
"ld1 { v3.b }[4], [x22], #0x1\n"
"ld1 { v28.b }[4], [x21], #0x1\n"
"ld1 { v22.b }[4], [x20], #0x1\n"
"b 23f\n"
"21:" // Oddments: 4 inputs loop: Load: Bit 3: Unset: Bit 2: Unset
"tbz %x[n_channels], #1, 22f\n"
"ldr h4, [x23], #0x2\n"
"ldr h3, [x22], #0x2\n"
"ldr h28, [x21], #0x2\n"
"ldr h22, [x20], #0x2\n"
"tbz %x[n_channels], #0, 23f\n"
"ld1 { v4.b }[2], [x23], #0x1\n"
"ld1 { v3.b }[2], [x22], #0x1\n"
"ld1 { v28.b }[2], [x21], #0x1\n"
"ld1 { v22.b }[2], [x20], #0x1\n"
"b 23f\n"
"22:" // Oddments: 4 inputs loop: Load: Bit 3: Unset: Bit 2: Unset: Bit 1: Unset
"tbz %x[n_channels], #0, 23f\n"
"ldr b4, [x23], #0x1\n"
"ldr b3, [x22], #0x1\n"
"ldr b28, [x21], #0x1\n"
"ldr b22, [x20], #0x1\n"
"23:" // Oddments: 4 inputs loop: Load: Bit 3: End
"umax v17.16b, v4.16b, v3.16b\n"
"umax v16.16b, v28.16b, v22.16b\n"
"subs x25, x25, #0x1\n"
"umax v16.16b, v17.16b, v16.16b\n"
"umax v8.16b, v8.16b, v16.16b\n"
"bgt 15b\n"
"24:" // Oddments: After loop
"ands x21, %x[n_valid_cells], #0x3\n"
"beq 34f\n"
"25:" // Oddments: Single input loop
"ldr x23, [x24], #0x8\n"
"add x23, x23, x27\n"
"movi v4.16b, #0x0\n"
"tbz %x[n_channels], #3, 29f\n"
"ldr d4, [x23], #0x8\n"
"tbz %x[n_channels], #2, 27f\n"
"ld1 { v4.s }[2], [x23], #0x4\n"
"tbz %x[n_channels], #1, 26f\n"
"ld1 { v4.h }[6], [x23], #0x2\n"
"tbz %x[n_channels], #0, 33f\n"
"ld1 { v4.b }[14], [x23], #0x1\n"
"b 33f\n"
"26:" // Oddments: Single input loop: Load: Bit 3: Bit 2: Bit 1: Unset
"tbz %x[n_channels], #0, 33f\n"
"ld1 { v4.b }[12], [x23], #0x1\n"
"b 33f\n"
"27:" // Oddments: Single input loop: Load: Bit 3: Bit 2: Unset
"tbz %x[n_channels], #1, 28f\n"
"ld1 { v4.h }[4], [x23], #0x2\n"
"tbz %x[n_channels], #0, 33f\n"
"ld1 { v4.b }[10], [x23], #0x1\n"
"b 33f\n"
"28:" // Oddments: Single input loop: Load: Bit 3: Bit 2: Unset: Bit 1: Unset
"tbz %x[n_channels], #0, 33f\n"
"ld1 { v4.b }[8], [x23], #0x1\n"
"b 33f\n"
"29:" // Oddments: Single input loop: Load: Bit 3: Unset
"tbz %x[n_channels], #2, 31f\n"
"ldr s4, [x23], #0x4\n"
"tbz %x[n_channels], #1, 30f\n"
"ld1 { v4.h }[2], [x23], #0x2\n"
"tbz %x[n_channels], #0, 33f\n"
"ld1 { v4.b }[6], [x23], #0x1\n"
"b 33f\n"
"30:" // Oddments: Single input loop: Load: Bit 3: Unset: Bit 2: Bit 1: Unset
"tbz %x[n_channels], #0, 33f\n"
"ld1 { v4.b }[4], [x23], #0x1\n"
"b 33f\n"
"31:" // Oddments: Single input loop: Load: Bit 3: Unset: Bit 2: Unset
"tbz %x[n_channels], #1, 32f\n"
"ldr h4, [x23], #0x2\n"
"tbz %x[n_channels], #0, 33f\n"
"ld1 { v4.b }[2], [x23], #0x1\n"
"b 33f\n"
"32:" // Oddments: Single input loop: Load: Bit 3: Unset: Bit 2: Unset: Bit 1: Unset
"tbz %x[n_channels], #0, 33f\n"
"ldr b4, [x23], #0x1\n"
"33:" // Oddments: Single input loop: Load: Bit 3: End
"subs x21, x21, #0x1\n"
"umax v8.16b, v8.16b, v4.16b\n"
"bgt 25b\n"
"34:" // Oddments: Single input loop: End
"tbz %x[n_channels], #3, 38f\n"
"st1 { v8.d }[0], [%x[outptr]], #0x8\n"
"tbz %x[n_channels], #2, 36f\n"
"st1 { v8.s }[2], [%x[outptr]], #0x4\n"
"tbz %x[n_channels], #1, 35f\n"
"st1 { v8.h }[6], [%x[outptr]], #0x2\n"
"tbz %x[n_channels], #0, 42f\n"
"st1 { v8.b }[14], [%x[outptr]], #0x1\n"
"b 42f\n"
"35:" // Oddments: Store: Bit 3: Bit 2: Bit 1: Unset
"tbz %x[n_channels], #0, 42f\n"
"st1 { v8.b }[12], [%x[outptr]], #0x1\n"
"b 42f\n"
"36:" // Oddments: Store: Bit 3: Bit 2: Unset
"tbz %x[n_channels], #1, 37f\n"
"st1 { v8.h }[4], [%x[outptr]], #0x2\n"
"tbz %x[n_channels], #0, 42f\n"
"st1 { v8.b }[10], [%x[outptr]], #0x1\n"
"b 42f\n"
"37:" // Oddments: Store: Bit 3: Bit 2: Unset: Bit 1: Unset
"tbz %x[n_channels], #0, 42f\n"
"st1 { v8.b }[8], [%x[outptr]], #0x1\n"
"b 42f\n"
"38:" // Oddments: Store: Bit 3: Unset
"tbz %x[n_channels], #2, 40f\n"
"st1 { v8.s }[0], [%x[outptr]], #0x4\n"
"tbz %x[n_channels], #1, 39f\n"
"st1 { v8.h }[2], [%x[outptr]], #0x2\n"
"tbz %x[n_channels], #0, 42f\n"
"st1 { v8.b }[6], [%x[outptr]], #0x1\n"
"b 42f\n"
"39:" // Oddments: Store: Bit 3: Unset: Bit 2: Bit 1: Unset
"tbz %x[n_channels], #0, 42f\n"
"st1 { v8.b }[4], [%x[outptr]], #0x1\n"
"b 42f\n"
"40:" // Oddments: Store: Bit 3: Unset: Bit 2: Unset
"tbz %x[n_channels], #1, 41f\n"
"st1 { v8.h }[0], [%x[outptr]], #0x2\n"
"tbz %x[n_channels], #0, 42f\n"
"st1 { v8.b }[2], [%x[outptr]], #0x1\n"
"b 42f\n"
"41:" // Oddments: Store: Bit 3: Unset: Bit 2: Unset: Bit 1: Unset
"tbz %x[n_channels], #0, 42f\n"
"st1 { v8.b }[0], [%x[outptr]], #0x1\n"
"42:" // Oddments: Store: Bit 3: End
"43:" // End
: [n_channels] "+&r" (n_channels), [outptr] "+&r" (outptr)
: [inptrs] "r" (inptrs), [n_valid_cells] "r" (n_valid_cells)
: "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27"
);
}
} // namespace pooling
} // namespace arm_conv
#endif // defined(__aarch64__)
|
