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
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
|
.global RestoreRegisters_NEON
.global ReverseLine_NEON
.global ReverseLineUV_NEON
.global SaveRegisters_NEON
.global TransposeWx8_NEON
.global TransposeUVWx8_NEON
.type RestoreRegisters_NEON, function
.type ReverseLine_NEON, function
.type ReverseLineUV_NEON, function
.type SaveRegisters_NEON, function
.type TransposeWx8_NEON, function
.type TransposeUVWx8_NEON, function
@ void ReverseLine_NEON (const uint8* src, uint8* dst, int width)
@ r0 const uint8* src
@ r1 uint8* dst
@ r2 width
ReverseLine_NEON:
@ compute where to start writing destination
add r1, r2 @ dst + width
@ work on segments that are multiples of 16
lsrs r3, r2, #4
@ the output is written in two block. 8 bytes followed
@ by another 8. reading is done sequentially, from left to
@ right. writing is done from right to left in block sizes
@ r1, the destination pointer is incremented after writing
@ the first of the two blocks. need to subtract that 8 off
@ along with 16 to get the next location.
mov r3, #-24
beq Lline_residuals
@ back of destination by the size of the register that is
@ going to be reversed
sub r1, #16
@ the loop needs to run on blocks of 16. what will be left
@ over is either a negative number, the residuals that need
@ to be done, or 0. if this isn't subtracted off here the
@ loop will run one extra time.
sub r2, #16
Lsegments_of_16:
vld1.8 {q0}, [r0]! @ src += 16
@ reverse the bytes in the 64 bit segments. unable to reverse
@ the bytes in the entire 128 bits in one go.
vrev64.8 q0, q0
@ because of the inability to reverse the entire 128 bits
@ reverse the writing out of the two 64 bit segments.
vst1.8 {d1}, [r1]!
vst1.8 {d0}, [r1], r3 @ dst -= 16
subs r2, #16
bge Lsegments_of_16
@ add 16 back to the counter. if the result is 0 there is no
@ residuals so return
adds r2, #16
bxeq lr
add r1, #16
Lline_residuals:
mov r3, #-3
sub r1, #2
subs r2, #2
@ check for 16*n+1 scenarios where segments_of_2 should not
@ be run, but there is something left over.
blt Lsegment_of_1
@ do this in neon registers as per
@ http://blogs.arm.com/software-enablement/196-coding-for-neon-part-2-dealing-with-leftovers/
Lsegments_of_2:
vld2.8 {d0[0], d1[0]}, [r0]! @ src += 2
vst1.8 {d1[0]}, [r1]!
vst1.8 {d0[0]}, [r1], r3 @ dst -= 2
subs r2, #2
bge Lsegments_of_2
adds r2, #2
bxeq lr
Lsegment_of_1:
add r1, #1
vld1.8 {d0[0]}, [r0]
vst1.8 {d0[0]}, [r1]
bx lr
@ void TransposeWx8_NEON (const uint8* src, int src_stride,
@ uint8* dst, int dst_stride,
@ int w)
@ r0 const uint8* src
@ r1 int src_stride
@ r2 uint8* dst
@ r3 int dst_stride
@ stack int w
TransposeWx8_NEON:
push {r4,r8,r9,lr}
ldr r8, [sp, #16] @ width
@ loops are on blocks of 8. loop will stop when
@ counter gets to or below 0. starting the counter
@ at w-8 allow for this
sub r8, #8
@ handle 8x8 blocks. this should be the majority of the plane
Lloop_8x8:
mov r9, r0
vld1.8 {d0}, [r9], r1
vld1.8 {d1}, [r9], r1
vld1.8 {d2}, [r9], r1
vld1.8 {d3}, [r9], r1
vld1.8 {d4}, [r9], r1
vld1.8 {d5}, [r9], r1
vld1.8 {d6}, [r9], r1
vld1.8 {d7}, [r9]
vtrn.8 d1, d0
vtrn.8 d3, d2
vtrn.8 d5, d4
vtrn.8 d7, d6
vtrn.16 d1, d3
vtrn.16 d0, d2
vtrn.16 d5, d7
vtrn.16 d4, d6
vtrn.32 d1, d5
vtrn.32 d0, d4
vtrn.32 d3, d7
vtrn.32 d2, d6
vrev16.8 q0, q0
vrev16.8 q1, q1
vrev16.8 q2, q2
vrev16.8 q3, q3
mov r9, r2
vst1.8 {d1}, [r9], r3
vst1.8 {d0}, [r9], r3
vst1.8 {d3}, [r9], r3
vst1.8 {d2}, [r9], r3
vst1.8 {d5}, [r9], r3
vst1.8 {d4}, [r9], r3
vst1.8 {d7}, [r9], r3
vst1.8 {d6}, [r9]
add r0, #8 @ src += 8
add r2, r3, lsl #3 @ dst += 8 * dst_stride
subs r8, #8 @ w -= 8
bge Lloop_8x8
@ add 8 back to counter. if the result is 0 there are
@ no residuals.
adds r8, #8
beq Ldone
@ some residual, so between 1 and 7 lines left to transpose
cmp r8, #2
blt Lblock_1x8
cmp r8, #4
blt Lblock_2x8
Lblock_4x8:
mov r9, r0
vld1.32 {d0[0]}, [r9], r1
vld1.32 {d0[1]}, [r9], r1
vld1.32 {d1[0]}, [r9], r1
vld1.32 {d1[1]}, [r9], r1
vld1.32 {d2[0]}, [r9], r1
vld1.32 {d2[1]}, [r9], r1
vld1.32 {d3[0]}, [r9], r1
vld1.32 {d3[1]}, [r9]
mov r9, r2
adr r12, vtbl_4x4_transpose
vld1.8 {q3}, [r12]
vtbl.8 d4, {d0, d1}, d6
vtbl.8 d5, {d0, d1}, d7
vtbl.8 d0, {d2, d3}, d6
vtbl.8 d1, {d2, d3}, d7
@ TODO: rework shuffle above to write
@ out with 4 instead of 8 writes
vst1.32 {d4[0]}, [r9], r3
vst1.32 {d4[1]}, [r9], r3
vst1.32 {d5[0]}, [r9], r3
vst1.32 {d5[1]}, [r9]
add r9, r2, #4
vst1.32 {d0[0]}, [r9], r3
vst1.32 {d0[1]}, [r9], r3
vst1.32 {d1[0]}, [r9], r3
vst1.32 {d1[1]}, [r9]
add r0, #4 @ src += 4
add r2, r3, lsl #2 @ dst += 4 * dst_stride
subs r8, #4 @ w -= 4
beq Ldone
@ some residual, check to see if it includes a 2x8 block,
@ or less
cmp r8, #2
blt Lblock_1x8
Lblock_2x8:
mov r9, r0
vld1.16 {d0[0]}, [r9], r1
vld1.16 {d1[0]}, [r9], r1
vld1.16 {d0[1]}, [r9], r1
vld1.16 {d1[1]}, [r9], r1
vld1.16 {d0[2]}, [r9], r1
vld1.16 {d1[2]}, [r9], r1
vld1.16 {d0[3]}, [r9], r1
vld1.16 {d1[3]}, [r9]
vtrn.8 d0, d1
mov r9, r2
vst1.64 {d0}, [r9], r3
vst1.64 {d1}, [r9]
add r0, #2 @ src += 2
add r2, r3, lsl #1 @ dst += 2 * dst_stride
subs r8, #2 @ w -= 2
beq Ldone
Lblock_1x8:
vld1.8 {d0[0]}, [r0], r1
vld1.8 {d0[1]}, [r0], r1
vld1.8 {d0[2]}, [r0], r1
vld1.8 {d0[3]}, [r0], r1
vld1.8 {d0[4]}, [r0], r1
vld1.8 {d0[5]}, [r0], r1
vld1.8 {d0[6]}, [r0], r1
vld1.8 {d0[7]}, [r0]
vst1.64 {d0}, [r2]
Ldone:
pop {r4,r8,r9,pc}
vtbl_4x4_transpose:
.byte 0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15
@ void SaveRegisters_NEON (unsigned long long store)
@ r0 unsigned long long store
SaveRegisters_NEON:
vst1.i64 {d8, d9, d10, d11}, [r0]!
vst1.i64 {d12, d13, d14, d15}, [r0]!
bx lr
@ void RestoreRegisters_NEON (unsigned long long store)
@ r0 unsigned long long store
RestoreRegisters_NEON:
vld1.i64 {d8, d9, d10, d11}, [r0]!
vld1.i64 {d12, d13, d14, d15}, [r0]!
bx lr
@ void ReverseLineUV_NEON (const uint8* src,
@ uint8* dst_a,
@ uint8* dst_b,
@ int width)
@ r0 const uint8* src
@ r1 uint8* dst_a
@ r2 uint8* dst_b
@ r3 width
ReverseLineUV_NEON:
@ compute where to start writing destination
add r1, r1, r3 @ dst_a + width
add r2, r2, r3 @ dst_b + width
@ work on input segments that are multiples of 16, but
@ width that has been passed is output segments, half
@ the size of input.
lsrs r12, r3, #3
beq Lline_residuals_di
@ the output is written in to two blocks.
mov r12, #-8
@ back of destination by the size of the register that is
@ going to be reversed
sub r1, r1, #8
sub r2, r2, #8
@ the loop needs to run on blocks of 8. what will be left
@ over is either a negative number, the residuals that need
@ to be done, or 0. if this isn't subtracted off here the
@ loop will run one extra time.
sub r3, r3, #8
Lsegments_of_8_di:
vld2.8 {d0, d1}, [r0]! @ src += 16
@ reverse the bytes in the 64 bit segments
vrev64.8 q0, q0
vst1.8 {d0}, [r1], r12 @ dst_a -= 8
vst1.8 {d1}, [r2], r12 @ dst_b -= 8
subs r3, r3, #8
bge Lsegments_of_8_di
@ add 8 back to the counter. if the result is 0 there is no
@ residuals so return
adds r3, r3, #8
bxeq lr
add r1, r1, #8
add r2, r2, #8
Lline_residuals_di:
mov r12, #-1
sub r1, r1, #1
sub r2, r2, #1
@ do this in neon registers as per
@ http://blogs.arm.com/software-enablement/196-coding-for-neon-part-2-dealing-with-leftovers/
Lsegments_of_1:
vld2.8 {d0[0], d1[0]}, [r0]! @ src += 2
vst1.8 {d0[0]}, [r1], r12 @ dst_a -= 1
vst1.8 {d1[0]}, [r2], r12 @ dst_b -= 1
subs r3, r3, #1
bgt Lsegments_of_1
bx lr
@ void TransposeUVWx8_NEON (const uint8* src, int src_stride,
@ uint8* dst_a, int dst_stride_a,
@ uint8* dst_b, int dst_stride_b,
@ int width)
@ r0 const uint8* src
@ r1 int src_stride
@ r2 uint8* dst_a
@ r3 int dst_stride_a
@ stack uint8* dst_b
@ stack int dst_stride_b
@ stack int width
TransposeUVWx8_NEON:
push {r4-r9,lr}
ldr r4, [sp, #28] @ dst_b
ldr r5, [sp, #32] @ dst_stride_b
ldr r8, [sp, #36] @ width
@ loops are on blocks of 8. loop will stop when
@ counter gets to or below 0. starting the counter
@ at w-8 allow for this
sub r8, #8
@ handle 8x8 blocks. this should be the majority of the plane
Lloop_8x8_di:
mov r9, r0
vld2.8 {d0, d1}, [r9], r1
vld2.8 {d2, d3}, [r9], r1
vld2.8 {d4, d5}, [r9], r1
vld2.8 {d6, d7}, [r9], r1
vld2.8 {d8, d9}, [r9], r1
vld2.8 {d10, d11}, [r9], r1
vld2.8 {d12, d13}, [r9], r1
vld2.8 {d14, d15}, [r9]
vtrn.8 q1, q0
vtrn.8 q3, q2
vtrn.8 q5, q4
vtrn.8 q7, q6
vtrn.16 q1, q3
vtrn.16 q0, q2
vtrn.16 q5, q7
vtrn.16 q4, q6
vtrn.32 q1, q5
vtrn.32 q0, q4
vtrn.32 q3, q7
vtrn.32 q2, q6
vrev16.8 q0, q0
vrev16.8 q1, q1
vrev16.8 q2, q2
vrev16.8 q3, q3
vrev16.8 q4, q4
vrev16.8 q5, q5
vrev16.8 q6, q6
vrev16.8 q7, q7
mov r9, r2
vst1.8 {d2}, [r9], r3
vst1.8 {d0}, [r9], r3
vst1.8 {d6}, [r9], r3
vst1.8 {d4}, [r9], r3
vst1.8 {d10}, [r9], r3
vst1.8 {d8}, [r9], r3
vst1.8 {d14}, [r9], r3
vst1.8 {d12}, [r9]
mov r9, r4
vst1.8 {d3}, [r9], r5
vst1.8 {d1}, [r9], r5
vst1.8 {d7}, [r9], r5
vst1.8 {d5}, [r9], r5
vst1.8 {d11}, [r9], r5
vst1.8 {d9}, [r9], r5
vst1.8 {d15}, [r9], r5
vst1.8 {d13}, [r9]
add r0, #8*2 @ src += 8*2
add r2, r3, lsl #3 @ dst_a += 8 * dst_stride_a
add r4, r5, lsl #3 @ dst_b += 8 * dst_stride_b
subs r8, #8 @ w -= 8
bge Lloop_8x8_di
@ add 8 back to counter. if the result is 0 there are
@ no residuals.
adds r8, #8
beq Ldone_di
@ some residual, so between 1 and 7 lines left to transpose
cmp r8, #2
blt Lblock_1x8_di
cmp r8, #4
blt Lblock_2x8_di
@ TODO(frkoenig) : clean this up
Lblock_4x8_di:
mov r9, r0
vld1.64 {d0}, [r9], r1
vld1.64 {d1}, [r9], r1
vld1.64 {d2}, [r9], r1
vld1.64 {d3}, [r9], r1
vld1.64 {d4}, [r9], r1
vld1.64 {d5}, [r9], r1
vld1.64 {d6}, [r9], r1
vld1.64 {d7}, [r9]
adr r12, vtbl_4x4_transpose_di
vld1.8 {q7}, [r12]
vtrn.8 q0, q1
vtrn.8 q2, q3
vtbl.8 d8, {d0, d1}, d14
vtbl.8 d9, {d0, d1}, d15
vtbl.8 d10, {d2, d3}, d14
vtbl.8 d11, {d2, d3}, d15
vtbl.8 d12, {d4, d5}, d14
vtbl.8 d13, {d4, d5}, d15
vtbl.8 d0, {d6, d7}, d14
vtbl.8 d1, {d6, d7}, d15
mov r9, r2
vst1.32 {d8[0]}, [r9], r3
vst1.32 {d8[1]}, [r9], r3
vst1.32 {d9[0]}, [r9], r3
vst1.32 {d9[1]}, [r9], r3
add r9, r2, #4
vst1.32 {d12[0]}, [r9], r3
vst1.32 {d12[1]}, [r9], r3
vst1.32 {d13[0]}, [r9], r3
vst1.32 {d13[1]}, [r9]
mov r9, r4
vst1.32 {d10[0]}, [r9], r5
vst1.32 {d10[1]}, [r9], r5
vst1.32 {d11[0]}, [r9], r5
vst1.32 {d11[1]}, [r9], r5
add r9, r4, #4
vst1.32 {d0[0]}, [r9], r5
vst1.32 {d0[1]}, [r9], r5
vst1.32 {d1[0]}, [r9], r5
vst1.32 {d1[1]}, [r9]
add r0, #4*2 @ src += 4 * 2
add r2, r3, lsl #2 @ dst_a += 4 * dst_stride_a
add r4, r5, lsl #2 @ dst_b += 4 * dst_stride_b
subs r8, #4 @ w -= 4
beq Ldone_di
@ some residual, check to see if it includes a 2x8 block,
@ or less
cmp r8, #2
blt Lblock_1x8_di
Lblock_2x8_di:
mov r9, r0
vld2.16 {d0[0], d2[0]}, [r9], r1
vld2.16 {d1[0], d3[0]}, [r9], r1
vld2.16 {d0[1], d2[1]}, [r9], r1
vld2.16 {d1[1], d3[1]}, [r9], r1
vld2.16 {d0[2], d2[2]}, [r9], r1
vld2.16 {d1[2], d3[2]}, [r9], r1
vld2.16 {d0[3], d2[3]}, [r9], r1
vld2.16 {d1[3], d3[3]}, [r9]
vtrn.8 d0, d1
vtrn.8 d2, d3
mov r9, r2
vst1.64 {d0}, [r9], r3
vst1.64 {d2}, [r9]
mov r9, r4
vst1.64 {d1}, [r9], r5
vst1.64 {d3}, [r9]
add r0, #2*2 @ src += 2 * 2
add r2, r3, lsl #1 @ dst_a += 2 * dst_stride_a
add r4, r5, lsl #1 @ dst_a += 2 * dst_stride_a
subs r8, #2 @ w -= 2
beq Ldone_di
Lblock_1x8_di:
vld2.8 {d0[0], d1[0]}, [r0], r1
vld2.8 {d0[1], d1[1]}, [r0], r1
vld2.8 {d0[2], d1[2]}, [r0], r1
vld2.8 {d0[3], d1[3]}, [r0], r1
vld2.8 {d0[4], d1[4]}, [r0], r1
vld2.8 {d0[5], d1[5]}, [r0], r1
vld2.8 {d0[6], d1[6]}, [r0], r1
vld2.8 {d0[7], d1[7]}, [r0]
vst1.64 {d0}, [r2]
vst1.64 {d1}, [r4]
Ldone_di:
pop {r4-r9, pc}
vtbl_4x4_transpose_di:
.byte 0, 8, 1, 9, 2, 10, 3, 11, 4, 12, 5, 13, 6, 14, 7, 15
|