summaryrefslogtreecommitdiffstats
path: root/media/libjpeg/simd/arm/jdmrgext-neon.c
blob: 5b89bdb339424816d90e133f9e27baa76afe4e7e (plain)
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
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
/*
 * jdmrgext-neon.c - merged upsampling/color conversion (Arm Neon)
 *
 * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
 * Copyright (C) 2020, D. R. Commander.  All Rights Reserved.
 *
 * This software is provided 'as-is', without any express or implied
 * warranty.  In no event will the authors be held liable for any damages
 * arising from the use of this software.
 *
 * Permission is granted to anyone to use this software for any purpose,
 * including commercial applications, and to alter it and redistribute it
 * freely, subject to the following restrictions:
 *
 * 1. The origin of this software must not be misrepresented; you must not
 *    claim that you wrote the original software. If you use this software
 *    in a product, an acknowledgment in the product documentation would be
 *    appreciated but is not required.
 * 2. Altered source versions must be plainly marked as such, and must not be
 *    misrepresented as being the original software.
 * 3. This notice may not be removed or altered from any source distribution.
 */

/* This file is included by jdmerge-neon.c. */


/* These routines combine simple (non-fancy, i.e. non-smooth) h2v1 or h2v2
 * chroma upsampling and YCbCr -> RGB color conversion into a single function.
 *
 * As with the standalone functions, YCbCr -> RGB conversion is defined by the
 * following equations:
 *    R = Y                        + 1.40200 * (Cr - 128)
 *    G = Y - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128)
 *    B = Y + 1.77200 * (Cb - 128)
 *
 * Scaled integer constants are used to avoid floating-point arithmetic:
 *    0.3441467 = 11277 * 2^-15
 *    0.7141418 = 23401 * 2^-15
 *    1.4020386 = 22971 * 2^-14
 *    1.7720337 = 29033 * 2^-14
 * These constants are defined in jdmerge-neon.c.
 *
 * To ensure correct results, rounding is used when descaling.
 */

/* Notes on safe memory access for merged upsampling/YCbCr -> RGB conversion
 * routines:
 *
 * Input memory buffers can be safely overread up to the next multiple of
 * ALIGN_SIZE bytes, since they are always allocated by alloc_sarray() in
 * jmemmgr.c.
 *
 * The output buffer cannot safely be written beyond output_width, since
 * output_buf points to a possibly unpadded row in the decompressed image
 * buffer allocated by the calling program.
 */

/* Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical.
 */

void jsimd_h2v1_merged_upsample_neon(JDIMENSION output_width,
                                     JSAMPIMAGE input_buf,
                                     JDIMENSION in_row_group_ctr,
                                     JSAMPARRAY output_buf)
{
  JSAMPROW outptr;
  /* Pointers to Y, Cb, and Cr data */
  JSAMPROW inptr0, inptr1, inptr2;

  const int16x4_t consts = vld1_s16(jsimd_ycc_rgb_convert_neon_consts);
  const int16x8_t neg_128 = vdupq_n_s16(-128);

  inptr0 = input_buf[0][in_row_group_ctr];
  inptr1 = input_buf[1][in_row_group_ctr];
  inptr2 = input_buf[2][in_row_group_ctr];
  outptr = output_buf[0];

  int cols_remaining = output_width;
  for (; cols_remaining >= 16; cols_remaining -= 16) {
    /* De-interleave Y component values into two separate vectors, one
     * containing the component values with even-numbered indices and one
     * containing the component values with odd-numbered indices.
     */
    uint8x8x2_t y = vld2_u8(inptr0);
    uint8x8_t cb = vld1_u8(inptr1);
    uint8x8_t cr = vld1_u8(inptr2);
    /* Subtract 128 from Cb and Cr. */
    int16x8_t cr_128 =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
    int16x8_t cb_128 =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
    /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
    int32x4_t g_sub_y_l = vmull_lane_s16(vget_low_s16(cb_128), consts, 0);
    int32x4_t g_sub_y_h = vmull_lane_s16(vget_high_s16(cb_128), consts, 0);
    g_sub_y_l = vmlsl_lane_s16(g_sub_y_l, vget_low_s16(cr_128), consts, 1);
    g_sub_y_h = vmlsl_lane_s16(g_sub_y_h, vget_high_s16(cr_128), consts, 1);
    /* Descale G components: shift right 15, round, and narrow to 16-bit. */
    int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
                                     vrshrn_n_s32(g_sub_y_h, 15));
    /* Compute R-Y: 1.40200 * (Cr - 128) */
    int16x8_t r_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128, 1), consts, 2);
    /* Compute B-Y: 1.77200 * (Cb - 128) */
    int16x8_t b_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128, 1), consts, 3);
    /* Add the chroma-derived values (G-Y, R-Y, and B-Y) to both the "even" and
     * "odd" Y component values.  This effectively upsamples the chroma
     * components horizontally.
     */
    int16x8_t g_even =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
                                     y.val[0]));
    int16x8_t r_even =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
                                     y.val[0]));
    int16x8_t b_even =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
                                     y.val[0]));
    int16x8_t g_odd =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
                                     y.val[1]));
    int16x8_t r_odd =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
                                     y.val[1]));
    int16x8_t b_odd =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
                                     y.val[1]));
    /* Convert each component to unsigned and narrow, clamping to [0-255].
     * Re-interleave the "even" and "odd" component values.
     */
    uint8x8x2_t r = vzip_u8(vqmovun_s16(r_even), vqmovun_s16(r_odd));
    uint8x8x2_t g = vzip_u8(vqmovun_s16(g_even), vqmovun_s16(g_odd));
    uint8x8x2_t b = vzip_u8(vqmovun_s16(b_even), vqmovun_s16(b_odd));

#ifdef RGB_ALPHA
    uint8x16x4_t rgba;
    rgba.val[RGB_RED] = vcombine_u8(r.val[0], r.val[1]);
    rgba.val[RGB_GREEN] = vcombine_u8(g.val[0], g.val[1]);
    rgba.val[RGB_BLUE] = vcombine_u8(b.val[0], b.val[1]);
    /* Set alpha channel to opaque (0xFF). */
    rgba.val[RGB_ALPHA] = vdupq_n_u8(0xFF);
    /* Store RGBA pixel data to memory. */
    vst4q_u8(outptr, rgba);
#else
    uint8x16x3_t rgb;
    rgb.val[RGB_RED] = vcombine_u8(r.val[0], r.val[1]);
    rgb.val[RGB_GREEN] = vcombine_u8(g.val[0], g.val[1]);
    rgb.val[RGB_BLUE] = vcombine_u8(b.val[0], b.val[1]);
    /* Store RGB pixel data to memory. */
    vst3q_u8(outptr, rgb);
#endif

    /* Increment pointers. */
    inptr0 += 16;
    inptr1 += 8;
    inptr2 += 8;
    outptr += (RGB_PIXELSIZE * 16);
  }

  if (cols_remaining > 0) {
    /* De-interleave Y component values into two separate vectors, one
     * containing the component values with even-numbered indices and one
     * containing the component values with odd-numbered indices.
     */
    uint8x8x2_t y = vld2_u8(inptr0);
    uint8x8_t cb = vld1_u8(inptr1);
    uint8x8_t cr = vld1_u8(inptr2);
    /* Subtract 128 from Cb and Cr. */
    int16x8_t cr_128 =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
    int16x8_t cb_128 =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
    /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
    int32x4_t g_sub_y_l = vmull_lane_s16(vget_low_s16(cb_128), consts, 0);
    int32x4_t g_sub_y_h = vmull_lane_s16(vget_high_s16(cb_128), consts, 0);
    g_sub_y_l = vmlsl_lane_s16(g_sub_y_l, vget_low_s16(cr_128), consts, 1);
    g_sub_y_h = vmlsl_lane_s16(g_sub_y_h, vget_high_s16(cr_128), consts, 1);
    /* Descale G components: shift right 15, round, and narrow to 16-bit. */
    int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
                                     vrshrn_n_s32(g_sub_y_h, 15));
    /* Compute R-Y: 1.40200 * (Cr - 128) */
    int16x8_t r_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128, 1), consts, 2);
    /* Compute B-Y: 1.77200 * (Cb - 128) */
    int16x8_t b_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128, 1), consts, 3);
    /* Add the chroma-derived values (G-Y, R-Y, and B-Y) to both the "even" and
     * "odd" Y component values.  This effectively upsamples the chroma
     * components horizontally.
     */
    int16x8_t g_even =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
                                     y.val[0]));
    int16x8_t r_even =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
                                     y.val[0]));
    int16x8_t b_even =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
                                     y.val[0]));
    int16x8_t g_odd =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
                                     y.val[1]));
    int16x8_t r_odd =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
                                     y.val[1]));
    int16x8_t b_odd =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
                                     y.val[1]));
    /* Convert each component to unsigned and narrow, clamping to [0-255].
     * Re-interleave the "even" and "odd" component values.
     */
    uint8x8x2_t r = vzip_u8(vqmovun_s16(r_even), vqmovun_s16(r_odd));
    uint8x8x2_t g = vzip_u8(vqmovun_s16(g_even), vqmovun_s16(g_odd));
    uint8x8x2_t b = vzip_u8(vqmovun_s16(b_even), vqmovun_s16(b_odd));

#ifdef RGB_ALPHA
    uint8x8x4_t rgba_h;
    rgba_h.val[RGB_RED] = r.val[1];
    rgba_h.val[RGB_GREEN] = g.val[1];
    rgba_h.val[RGB_BLUE] = b.val[1];
    /* Set alpha channel to opaque (0xFF). */
    rgba_h.val[RGB_ALPHA] = vdup_n_u8(0xFF);
    uint8x8x4_t rgba_l;
    rgba_l.val[RGB_RED] = r.val[0];
    rgba_l.val[RGB_GREEN] = g.val[0];
    rgba_l.val[RGB_BLUE] = b.val[0];
    /* Set alpha channel to opaque (0xFF). */
    rgba_l.val[RGB_ALPHA] = vdup_n_u8(0xFF);
    /* Store RGBA pixel data to memory. */
    switch (cols_remaining) {
    case 15:
      vst4_lane_u8(outptr + 14 * RGB_PIXELSIZE, rgba_h, 6);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 14:
      vst4_lane_u8(outptr + 13 * RGB_PIXELSIZE, rgba_h, 5);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 13:
      vst4_lane_u8(outptr + 12 * RGB_PIXELSIZE, rgba_h, 4);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 12:
      vst4_lane_u8(outptr + 11 * RGB_PIXELSIZE, rgba_h, 3);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 11:
      vst4_lane_u8(outptr + 10 * RGB_PIXELSIZE, rgba_h, 2);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 10:
      vst4_lane_u8(outptr + 9 * RGB_PIXELSIZE, rgba_h, 1);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 9:
      vst4_lane_u8(outptr + 8 * RGB_PIXELSIZE, rgba_h, 0);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 8:
      vst4_u8(outptr, rgba_l);
      break;
    case 7:
      vst4_lane_u8(outptr + 6 * RGB_PIXELSIZE, rgba_l, 6);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 6:
      vst4_lane_u8(outptr + 5 * RGB_PIXELSIZE, rgba_l, 5);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 5:
      vst4_lane_u8(outptr + 4 * RGB_PIXELSIZE, rgba_l, 4);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 4:
      vst4_lane_u8(outptr + 3 * RGB_PIXELSIZE, rgba_l, 3);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 3:
      vst4_lane_u8(outptr + 2 * RGB_PIXELSIZE, rgba_l, 2);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 2:
      vst4_lane_u8(outptr + RGB_PIXELSIZE, rgba_l, 1);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 1:
      vst4_lane_u8(outptr, rgba_l, 0);
      FALLTHROUGH               /*FALLTHROUGH*/
    default:
      break;
    }
#else
    uint8x8x3_t rgb_h;
    rgb_h.val[RGB_RED] = r.val[1];
    rgb_h.val[RGB_GREEN] = g.val[1];
    rgb_h.val[RGB_BLUE] = b.val[1];
    uint8x8x3_t rgb_l;
    rgb_l.val[RGB_RED] = r.val[0];
    rgb_l.val[RGB_GREEN] = g.val[0];
    rgb_l.val[RGB_BLUE] = b.val[0];
    /* Store RGB pixel data to memory. */
    switch (cols_remaining) {
    case 15:
      vst3_lane_u8(outptr + 14 * RGB_PIXELSIZE, rgb_h, 6);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 14:
      vst3_lane_u8(outptr + 13 * RGB_PIXELSIZE, rgb_h, 5);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 13:
      vst3_lane_u8(outptr + 12 * RGB_PIXELSIZE, rgb_h, 4);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 12:
      vst3_lane_u8(outptr + 11 * RGB_PIXELSIZE, rgb_h, 3);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 11:
      vst3_lane_u8(outptr + 10 * RGB_PIXELSIZE, rgb_h, 2);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 10:
      vst3_lane_u8(outptr + 9 * RGB_PIXELSIZE, rgb_h, 1);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 9:
      vst3_lane_u8(outptr + 8 * RGB_PIXELSIZE, rgb_h, 0);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 8:
      vst3_u8(outptr, rgb_l);
      break;
    case 7:
      vst3_lane_u8(outptr + 6 * RGB_PIXELSIZE, rgb_l, 6);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 6:
      vst3_lane_u8(outptr + 5 * RGB_PIXELSIZE, rgb_l, 5);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 5:
      vst3_lane_u8(outptr + 4 * RGB_PIXELSIZE, rgb_l, 4);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 4:
      vst3_lane_u8(outptr + 3 * RGB_PIXELSIZE, rgb_l, 3);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 3:
      vst3_lane_u8(outptr + 2 * RGB_PIXELSIZE, rgb_l, 2);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 2:
      vst3_lane_u8(outptr + RGB_PIXELSIZE, rgb_l, 1);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 1:
      vst3_lane_u8(outptr, rgb_l, 0);
      FALLTHROUGH               /*FALLTHROUGH*/
    default:
      break;
    }
#endif
  }
}


/* Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical.
 *
 * See comments above for details regarding color conversion and safe memory
 * access.
 */

void jsimd_h2v2_merged_upsample_neon(JDIMENSION output_width,
                                     JSAMPIMAGE input_buf,
                                     JDIMENSION in_row_group_ctr,
                                     JSAMPARRAY output_buf)
{
  JSAMPROW outptr0, outptr1;
  /* Pointers to Y (both rows), Cb, and Cr data */
  JSAMPROW inptr0_0, inptr0_1, inptr1, inptr2;

  const int16x4_t consts = vld1_s16(jsimd_ycc_rgb_convert_neon_consts);
  const int16x8_t neg_128 = vdupq_n_s16(-128);

  inptr0_0 = input_buf[0][in_row_group_ctr * 2];
  inptr0_1 = input_buf[0][in_row_group_ctr * 2 + 1];
  inptr1 = input_buf[1][in_row_group_ctr];
  inptr2 = input_buf[2][in_row_group_ctr];
  outptr0 = output_buf[0];
  outptr1 = output_buf[1];

  int cols_remaining = output_width;
  for (; cols_remaining >= 16; cols_remaining -= 16) {
    /* For each row, de-interleave Y component values into two separate
     * vectors, one containing the component values with even-numbered indices
     * and one containing the component values with odd-numbered indices.
     */
    uint8x8x2_t y0 = vld2_u8(inptr0_0);
    uint8x8x2_t y1 = vld2_u8(inptr0_1);
    uint8x8_t cb = vld1_u8(inptr1);
    uint8x8_t cr = vld1_u8(inptr2);
    /* Subtract 128 from Cb and Cr. */
    int16x8_t cr_128 =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
    int16x8_t cb_128 =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
    /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
    int32x4_t g_sub_y_l = vmull_lane_s16(vget_low_s16(cb_128), consts, 0);
    int32x4_t g_sub_y_h = vmull_lane_s16(vget_high_s16(cb_128), consts, 0);
    g_sub_y_l = vmlsl_lane_s16(g_sub_y_l, vget_low_s16(cr_128), consts, 1);
    g_sub_y_h = vmlsl_lane_s16(g_sub_y_h, vget_high_s16(cr_128), consts, 1);
    /* Descale G components: shift right 15, round, and narrow to 16-bit. */
    int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
                                     vrshrn_n_s32(g_sub_y_h, 15));
    /* Compute R-Y: 1.40200 * (Cr - 128) */
    int16x8_t r_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128, 1), consts, 2);
    /* Compute B-Y: 1.77200 * (Cb - 128) */
    int16x8_t b_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128, 1), consts, 3);
    /* For each row, add the chroma-derived values (G-Y, R-Y, and B-Y) to both
     * the "even" and "odd" Y component values.  This effectively upsamples the
     * chroma components both horizontally and vertically.
     */
    int16x8_t g0_even =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
                                     y0.val[0]));
    int16x8_t r0_even =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
                                     y0.val[0]));
    int16x8_t b0_even =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
                                     y0.val[0]));
    int16x8_t g0_odd =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
                                     y0.val[1]));
    int16x8_t r0_odd =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
                                     y0.val[1]));
    int16x8_t b0_odd =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
                                     y0.val[1]));
    int16x8_t g1_even =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
                                     y1.val[0]));
    int16x8_t r1_even =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
                                     y1.val[0]));
    int16x8_t b1_even =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
                                     y1.val[0]));
    int16x8_t g1_odd =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
                                     y1.val[1]));
    int16x8_t r1_odd =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
                                     y1.val[1]));
    int16x8_t b1_odd =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
                                     y1.val[1]));
    /* Convert each component to unsigned and narrow, clamping to [0-255].
     * Re-interleave the "even" and "odd" component values.
     */
    uint8x8x2_t r0 = vzip_u8(vqmovun_s16(r0_even), vqmovun_s16(r0_odd));
    uint8x8x2_t r1 = vzip_u8(vqmovun_s16(r1_even), vqmovun_s16(r1_odd));
    uint8x8x2_t g0 = vzip_u8(vqmovun_s16(g0_even), vqmovun_s16(g0_odd));
    uint8x8x2_t g1 = vzip_u8(vqmovun_s16(g1_even), vqmovun_s16(g1_odd));
    uint8x8x2_t b0 = vzip_u8(vqmovun_s16(b0_even), vqmovun_s16(b0_odd));
    uint8x8x2_t b1 = vzip_u8(vqmovun_s16(b1_even), vqmovun_s16(b1_odd));

#ifdef RGB_ALPHA
    uint8x16x4_t rgba0, rgba1;
    rgba0.val[RGB_RED] = vcombine_u8(r0.val[0], r0.val[1]);
    rgba1.val[RGB_RED] = vcombine_u8(r1.val[0], r1.val[1]);
    rgba0.val[RGB_GREEN] = vcombine_u8(g0.val[0], g0.val[1]);
    rgba1.val[RGB_GREEN] = vcombine_u8(g1.val[0], g1.val[1]);
    rgba0.val[RGB_BLUE] = vcombine_u8(b0.val[0], b0.val[1]);
    rgba1.val[RGB_BLUE] = vcombine_u8(b1.val[0], b1.val[1]);
    /* Set alpha channel to opaque (0xFF). */
    rgba0.val[RGB_ALPHA] = vdupq_n_u8(0xFF);
    rgba1.val[RGB_ALPHA] = vdupq_n_u8(0xFF);
    /* Store RGBA pixel data to memory. */
    vst4q_u8(outptr0, rgba0);
    vst4q_u8(outptr1, rgba1);
#else
    uint8x16x3_t rgb0, rgb1;
    rgb0.val[RGB_RED] = vcombine_u8(r0.val[0], r0.val[1]);
    rgb1.val[RGB_RED] = vcombine_u8(r1.val[0], r1.val[1]);
    rgb0.val[RGB_GREEN] = vcombine_u8(g0.val[0], g0.val[1]);
    rgb1.val[RGB_GREEN] = vcombine_u8(g1.val[0], g1.val[1]);
    rgb0.val[RGB_BLUE] = vcombine_u8(b0.val[0], b0.val[1]);
    rgb1.val[RGB_BLUE] = vcombine_u8(b1.val[0], b1.val[1]);
    /* Store RGB pixel data to memory. */
    vst3q_u8(outptr0, rgb0);
    vst3q_u8(outptr1, rgb1);
#endif

    /* Increment pointers. */
    inptr0_0 += 16;
    inptr0_1 += 16;
    inptr1 += 8;
    inptr2 += 8;
    outptr0 += (RGB_PIXELSIZE * 16);
    outptr1 += (RGB_PIXELSIZE * 16);
  }

  if (cols_remaining > 0) {
    /* For each row, de-interleave Y component values into two separate
     * vectors, one containing the component values with even-numbered indices
     * and one containing the component values with odd-numbered indices.
     */
    uint8x8x2_t y0 = vld2_u8(inptr0_0);
    uint8x8x2_t y1 = vld2_u8(inptr0_1);
    uint8x8_t cb = vld1_u8(inptr1);
    uint8x8_t cr = vld1_u8(inptr2);
    /* Subtract 128 from Cb and Cr. */
    int16x8_t cr_128 =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cr));
    int16x8_t cb_128 =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(neg_128), cb));
    /* Compute G-Y: - 0.34414 * (Cb - 128) - 0.71414 * (Cr - 128) */
    int32x4_t g_sub_y_l = vmull_lane_s16(vget_low_s16(cb_128), consts, 0);
    int32x4_t g_sub_y_h = vmull_lane_s16(vget_high_s16(cb_128), consts, 0);
    g_sub_y_l = vmlsl_lane_s16(g_sub_y_l, vget_low_s16(cr_128), consts, 1);
    g_sub_y_h = vmlsl_lane_s16(g_sub_y_h, vget_high_s16(cr_128), consts, 1);
    /* Descale G components: shift right 15, round, and narrow to 16-bit. */
    int16x8_t g_sub_y = vcombine_s16(vrshrn_n_s32(g_sub_y_l, 15),
                                     vrshrn_n_s32(g_sub_y_h, 15));
    /* Compute R-Y: 1.40200 * (Cr - 128) */
    int16x8_t r_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cr_128, 1), consts, 2);
    /* Compute B-Y: 1.77200 * (Cb - 128) */
    int16x8_t b_sub_y = vqrdmulhq_lane_s16(vshlq_n_s16(cb_128, 1), consts, 3);
    /* For each row, add the chroma-derived values (G-Y, R-Y, and B-Y) to both
     * the "even" and "odd" Y component values.  This effectively upsamples the
     * chroma components both horizontally and vertically.
     */
    int16x8_t g0_even =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
                                     y0.val[0]));
    int16x8_t r0_even =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
                                     y0.val[0]));
    int16x8_t b0_even =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
                                     y0.val[0]));
    int16x8_t g0_odd =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
                                     y0.val[1]));
    int16x8_t r0_odd =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
                                     y0.val[1]));
    int16x8_t b0_odd =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
                                     y0.val[1]));
    int16x8_t g1_even =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
                                     y1.val[0]));
    int16x8_t r1_even =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
                                     y1.val[0]));
    int16x8_t b1_even =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
                                     y1.val[0]));
    int16x8_t g1_odd =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(g_sub_y),
                                     y1.val[1]));
    int16x8_t r1_odd =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(r_sub_y),
                                     y1.val[1]));
    int16x8_t b1_odd =
      vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(b_sub_y),
                                     y1.val[1]));
    /* Convert each component to unsigned and narrow, clamping to [0-255].
     * Re-interleave the "even" and "odd" component values.
     */
    uint8x8x2_t r0 = vzip_u8(vqmovun_s16(r0_even), vqmovun_s16(r0_odd));
    uint8x8x2_t r1 = vzip_u8(vqmovun_s16(r1_even), vqmovun_s16(r1_odd));
    uint8x8x2_t g0 = vzip_u8(vqmovun_s16(g0_even), vqmovun_s16(g0_odd));
    uint8x8x2_t g1 = vzip_u8(vqmovun_s16(g1_even), vqmovun_s16(g1_odd));
    uint8x8x2_t b0 = vzip_u8(vqmovun_s16(b0_even), vqmovun_s16(b0_odd));
    uint8x8x2_t b1 = vzip_u8(vqmovun_s16(b1_even), vqmovun_s16(b1_odd));

#ifdef RGB_ALPHA
    uint8x8x4_t rgba0_h, rgba1_h;
    rgba0_h.val[RGB_RED] = r0.val[1];
    rgba1_h.val[RGB_RED] = r1.val[1];
    rgba0_h.val[RGB_GREEN] = g0.val[1];
    rgba1_h.val[RGB_GREEN] = g1.val[1];
    rgba0_h.val[RGB_BLUE] = b0.val[1];
    rgba1_h.val[RGB_BLUE] = b1.val[1];
    /* Set alpha channel to opaque (0xFF). */
    rgba0_h.val[RGB_ALPHA] = vdup_n_u8(0xFF);
    rgba1_h.val[RGB_ALPHA] = vdup_n_u8(0xFF);

    uint8x8x4_t rgba0_l, rgba1_l;
    rgba0_l.val[RGB_RED] = r0.val[0];
    rgba1_l.val[RGB_RED] = r1.val[0];
    rgba0_l.val[RGB_GREEN] = g0.val[0];
    rgba1_l.val[RGB_GREEN] = g1.val[0];
    rgba0_l.val[RGB_BLUE] = b0.val[0];
    rgba1_l.val[RGB_BLUE] = b1.val[0];
    /* Set alpha channel to opaque (0xFF). */
    rgba0_l.val[RGB_ALPHA] = vdup_n_u8(0xFF);
    rgba1_l.val[RGB_ALPHA] = vdup_n_u8(0xFF);
    /* Store RGBA pixel data to memory. */
    switch (cols_remaining) {
    case 15:
      vst4_lane_u8(outptr0 + 14 * RGB_PIXELSIZE, rgba0_h, 6);
      vst4_lane_u8(outptr1 + 14 * RGB_PIXELSIZE, rgba1_h, 6);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 14:
      vst4_lane_u8(outptr0 + 13 * RGB_PIXELSIZE, rgba0_h, 5);
      vst4_lane_u8(outptr1 + 13 * RGB_PIXELSIZE, rgba1_h, 5);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 13:
      vst4_lane_u8(outptr0 + 12 * RGB_PIXELSIZE, rgba0_h, 4);
      vst4_lane_u8(outptr1 + 12 * RGB_PIXELSIZE, rgba1_h, 4);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 12:
      vst4_lane_u8(outptr0 + 11 * RGB_PIXELSIZE, rgba0_h, 3);
      vst4_lane_u8(outptr1 + 11 * RGB_PIXELSIZE, rgba1_h, 3);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 11:
      vst4_lane_u8(outptr0 + 10 * RGB_PIXELSIZE, rgba0_h, 2);
      vst4_lane_u8(outptr1 + 10 * RGB_PIXELSIZE, rgba1_h, 2);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 10:
      vst4_lane_u8(outptr0 + 9 * RGB_PIXELSIZE, rgba0_h, 1);
      vst4_lane_u8(outptr1 + 9 * RGB_PIXELSIZE, rgba1_h, 1);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 9:
      vst4_lane_u8(outptr0 + 8 * RGB_PIXELSIZE, rgba0_h, 0);
      vst4_lane_u8(outptr1 + 8 * RGB_PIXELSIZE, rgba1_h, 0);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 8:
      vst4_u8(outptr0, rgba0_l);
      vst4_u8(outptr1, rgba1_l);
      break;
    case 7:
      vst4_lane_u8(outptr0 + 6 * RGB_PIXELSIZE, rgba0_l, 6);
      vst4_lane_u8(outptr1 + 6 * RGB_PIXELSIZE, rgba1_l, 6);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 6:
      vst4_lane_u8(outptr0 + 5 * RGB_PIXELSIZE, rgba0_l, 5);
      vst4_lane_u8(outptr1 + 5 * RGB_PIXELSIZE, rgba1_l, 5);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 5:
      vst4_lane_u8(outptr0 + 4 * RGB_PIXELSIZE, rgba0_l, 4);
      vst4_lane_u8(outptr1 + 4 * RGB_PIXELSIZE, rgba1_l, 4);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 4:
      vst4_lane_u8(outptr0 + 3 * RGB_PIXELSIZE, rgba0_l, 3);
      vst4_lane_u8(outptr1 + 3 * RGB_PIXELSIZE, rgba1_l, 3);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 3:
      vst4_lane_u8(outptr0 + 2 * RGB_PIXELSIZE, rgba0_l, 2);
      vst4_lane_u8(outptr1 + 2 * RGB_PIXELSIZE, rgba1_l, 2);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 2:
      vst4_lane_u8(outptr0 + 1 * RGB_PIXELSIZE, rgba0_l, 1);
      vst4_lane_u8(outptr1 + 1 * RGB_PIXELSIZE, rgba1_l, 1);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 1:
      vst4_lane_u8(outptr0, rgba0_l, 0);
      vst4_lane_u8(outptr1, rgba1_l, 0);
      FALLTHROUGH               /*FALLTHROUGH*/
    default:
      break;
    }
#else
    uint8x8x3_t rgb0_h, rgb1_h;
    rgb0_h.val[RGB_RED] = r0.val[1];
    rgb1_h.val[RGB_RED] = r1.val[1];
    rgb0_h.val[RGB_GREEN] = g0.val[1];
    rgb1_h.val[RGB_GREEN] = g1.val[1];
    rgb0_h.val[RGB_BLUE] = b0.val[1];
    rgb1_h.val[RGB_BLUE] = b1.val[1];

    uint8x8x3_t rgb0_l, rgb1_l;
    rgb0_l.val[RGB_RED] = r0.val[0];
    rgb1_l.val[RGB_RED] = r1.val[0];
    rgb0_l.val[RGB_GREEN] = g0.val[0];
    rgb1_l.val[RGB_GREEN] = g1.val[0];
    rgb0_l.val[RGB_BLUE] = b0.val[0];
    rgb1_l.val[RGB_BLUE] = b1.val[0];
    /* Store RGB pixel data to memory. */
    switch (cols_remaining) {
    case 15:
      vst3_lane_u8(outptr0 + 14 * RGB_PIXELSIZE, rgb0_h, 6);
      vst3_lane_u8(outptr1 + 14 * RGB_PIXELSIZE, rgb1_h, 6);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 14:
      vst3_lane_u8(outptr0 + 13 * RGB_PIXELSIZE, rgb0_h, 5);
      vst3_lane_u8(outptr1 + 13 * RGB_PIXELSIZE, rgb1_h, 5);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 13:
      vst3_lane_u8(outptr0 + 12 * RGB_PIXELSIZE, rgb0_h, 4);
      vst3_lane_u8(outptr1 + 12 * RGB_PIXELSIZE, rgb1_h, 4);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 12:
      vst3_lane_u8(outptr0 + 11 * RGB_PIXELSIZE, rgb0_h, 3);
      vst3_lane_u8(outptr1 + 11 * RGB_PIXELSIZE, rgb1_h, 3);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 11:
      vst3_lane_u8(outptr0 + 10 * RGB_PIXELSIZE, rgb0_h, 2);
      vst3_lane_u8(outptr1 + 10 * RGB_PIXELSIZE, rgb1_h, 2);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 10:
      vst3_lane_u8(outptr0 + 9 * RGB_PIXELSIZE, rgb0_h, 1);
      vst3_lane_u8(outptr1 + 9 * RGB_PIXELSIZE, rgb1_h, 1);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 9:
      vst3_lane_u8(outptr0 + 8 * RGB_PIXELSIZE, rgb0_h, 0);
      vst3_lane_u8(outptr1 + 8 * RGB_PIXELSIZE, rgb1_h, 0);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 8:
      vst3_u8(outptr0, rgb0_l);
      vst3_u8(outptr1, rgb1_l);
      break;
    case 7:
      vst3_lane_u8(outptr0 + 6 * RGB_PIXELSIZE, rgb0_l, 6);
      vst3_lane_u8(outptr1 + 6 * RGB_PIXELSIZE, rgb1_l, 6);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 6:
      vst3_lane_u8(outptr0 + 5 * RGB_PIXELSIZE, rgb0_l, 5);
      vst3_lane_u8(outptr1 + 5 * RGB_PIXELSIZE, rgb1_l, 5);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 5:
      vst3_lane_u8(outptr0 + 4 * RGB_PIXELSIZE, rgb0_l, 4);
      vst3_lane_u8(outptr1 + 4 * RGB_PIXELSIZE, rgb1_l, 4);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 4:
      vst3_lane_u8(outptr0 + 3 * RGB_PIXELSIZE, rgb0_l, 3);
      vst3_lane_u8(outptr1 + 3 * RGB_PIXELSIZE, rgb1_l, 3);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 3:
      vst3_lane_u8(outptr0 + 2 * RGB_PIXELSIZE, rgb0_l, 2);
      vst3_lane_u8(outptr1 + 2 * RGB_PIXELSIZE, rgb1_l, 2);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 2:
      vst3_lane_u8(outptr0 + 1 * RGB_PIXELSIZE, rgb0_l, 1);
      vst3_lane_u8(outptr1 + 1 * RGB_PIXELSIZE, rgb1_l, 1);
      FALLTHROUGH               /*FALLTHROUGH*/
    case 1:
      vst3_lane_u8(outptr0, rgb0_l, 0);
      vst3_lane_u8(outptr1, rgb1_l, 0);
      FALLTHROUGH               /*FALLTHROUGH*/
    default:
      break;
    }
#endif
  }
}