aboutsummaryrefslogtreecommitdiffhomepage
path: root/src/jumper/SkJumper_stages.cpp
blob: a200adffea58800fad9bce9cc7b5d02fca845ce6 (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
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
/*
 * Copyright 2017 Google Inc.
 *
 * Use of this source code is governed by a BSD-style license that can be
 * found in the LICENSE file.
 */

#include "SkJumper.h"
#include "SkJumper_misc.h"     // SI, unaligned_load(), bit_cast()
#include "SkJumper_vectors.h"  // F, I32, U32, U16, U8, cast(), expand()

// Our fundamental vector depth is our pixel stride.
static const size_t kStride = sizeof(F) / sizeof(float);

// A reminder:
// Code guarded by defined(JUMPER) can assume that it will be compiled by Clang
// and that F, I32, etc. are kStride-deep ext_vector_types of the appropriate type.
// Otherwise, F, I32, etc. just alias the basic scalar types (and so kStride == 1).

// You can use most constants in this file, but in a few rare exceptions we read from this struct.
using K = const SkJumper_constants;

// A little wrapper macro to name Stages differently depending on the instruction set.
// That lets us link together several options.
#if !defined(JUMPER)
    #define WRAP(name) sk_##name
#elif defined(__aarch64__)
    #define WRAP(name) sk_##name##_aarch64
#elif defined(__arm__)
    #define WRAP(name) sk_##name##_vfp4
#elif defined(__AVX2__)
    #define WRAP(name) sk_##name##_hsw
#elif defined(__AVX__)
    #define WRAP(name) sk_##name##_avx
#elif defined(__SSE4_1__)
    #define WRAP(name) sk_##name##_sse41
#elif defined(__SSE2__)
    #define WRAP(name) sk_##name##_sse2
#endif

// We're finally going to get to what a Stage function looks like!
//    tail == 0 ~~> work on a full kStride pixels
//    tail != 0 ~~> work on only the first tail pixels
// tail is always < kStride.
//
// We keep program the second argument, so that it's passed in rsi for load_and_inc().
using Stage = void(K* k, void** program, size_t x, size_t y, size_t tail, F,F,F,F, F,F,F,F);

#if defined(JUMPER) && defined(__AVX__)
    // We really want to make sure all paths go through this function's (implicit) vzeroupper.
    // If they don't, we'll experience severe slowdowns when we first use SSE instructions again.
    __attribute__((disable_tail_calls))
#endif
MAYBE_MSABI
extern "C" void WRAP(start_pipeline)(size_t x, size_t y, size_t limit, void** program, K* k) {
#if defined(JUMPER)
    F v;
#else
    F v{};
#endif
    auto start = (Stage*)load_and_inc(program);
    while (x + kStride <= limit) {
        start(k,program,x,y,0,    v,v,v,v, v,v,v,v);
        x += kStride;
    }
    if (size_t tail = limit - x) {
        start(k,program,x,y,tail, v,v,v,v, v,v,v,v);
    }
}

#define STAGE(name)                                                                   \
    SI void name##_k(K* k, LazyCtx ctx, size_t x, size_t y, size_t tail,              \
                     F& r, F& g, F& b, F& a, F& dr, F& dg, F& db, F& da);             \
    extern "C" void WRAP(name)(K* k, void** program, size_t x, size_t y, size_t tail, \
                               F r, F g, F b, F a, F dr, F dg, F db, F da) {          \
        LazyCtx ctx(program);                                                         \
        name##_k(k,ctx,x,y,tail, r,g,b,a, dr,dg,db,da);                               \
        auto next = (Stage*)load_and_inc(program);                                    \
        next(k,program,x,y,tail, r,g,b,a, dr,dg,db,da);                               \
    }                                                                                 \
    SI void name##_k(K* k, LazyCtx ctx, size_t x, size_t y, size_t tail,              \
                     F& r, F& g, F& b, F& a, F& dr, F& dg, F& db, F& da)


// just_return() is a simple no-op stage that only exists to end the chain,
// returning back up to start_pipeline(), and from there to the caller.
extern "C" void WRAP(just_return)(K*, void**, size_t,size_t,size_t, F,F,F,F, F,F,F,F) {}


// We could start defining normal Stages now.  But first, some helper functions.

// These load() and store() methods are tail-aware,
// but focus mainly on keeping the at-stride tail==0 case fast.

template <typename V, typename T>
SI V load(const T* src, size_t tail) {
#if defined(JUMPER)
    __builtin_assume(tail < kStride);
    if (__builtin_expect(tail, 0)) {
        V v{};  // Any inactive lanes are zeroed.
        switch (tail) {
            case 7: v[6] = src[6];
            case 6: v[5] = src[5];
            case 5: v[4] = src[4];
            case 4: memcpy(&v, src, 4*sizeof(T)); break;
            case 3: v[2] = src[2];
            case 2: memcpy(&v, src, 2*sizeof(T)); break;
            case 1: memcpy(&v, src, 1*sizeof(T)); break;
        }
        return v;
    }
#endif
    return unaligned_load<V>(src);
}

template <typename V, typename T>
SI void store(T* dst, V v, size_t tail) {
#if defined(JUMPER)
    __builtin_assume(tail < kStride);
    if (__builtin_expect(tail, 0)) {
        switch (tail) {
            case 7: dst[6] = v[6];
            case 6: dst[5] = v[5];
            case 5: dst[4] = v[4];
            case 4: memcpy(dst, &v, 4*sizeof(T)); break;
            case 3: dst[2] = v[2];
            case 2: memcpy(dst, &v, 2*sizeof(T)); break;
            case 1: memcpy(dst, &v, 1*sizeof(T)); break;
        }
        return;
    }
#endif
    unaligned_store(dst, v);
}

// AVX adds some mask loads and stores that make for shorter, faster code.
#if defined(JUMPER) && defined(__AVX__)
    SI U32 mask(size_t tail) {
        // We go a little out of our way to avoid needing large constant values here.

        // It's easiest to build the mask as 8 8-bit values, either 0x00 or 0xff.
        // Start fully on, then shift away lanes from the top until we've got our mask.
        uint64_t mask = 0xffffffffffffffff >> 8*(kStride-tail);

        // Sign-extend each mask lane to its full width, 0x00000000 or 0xffffffff.
        using S8  = int8_t  __attribute__((ext_vector_type(8)));
        using S32 = int32_t __attribute__((ext_vector_type(8)));
        return (U32)__builtin_convertvector(unaligned_load<S8>(&mask), S32);
    }

    template <>
    inline U32 load(const uint32_t* src, size_t tail) {
        __builtin_assume(tail < kStride);
        if (__builtin_expect(tail, 0)) {
            return (U32)_mm256_maskload_ps((const float*)src, mask(tail));
        }
        return unaligned_load<U32>(src);
    }

    template <>
    inline void store(uint32_t* dst, U32 v, size_t tail) {
        __builtin_assume(tail < kStride);
        if (__builtin_expect(tail, 0)) {
            return _mm256_maskstore_ps((float*)dst, mask(tail), (F)v);
        }
        unaligned_store(dst, v);
    }
#endif

SI F from_byte(U8 b) {
    return cast(expand(b)) * (1/255.0f);
}
SI void from_565(U16 _565, F* r, F* g, F* b) {
    U32 wide = expand(_565);
    *r = cast(wide & (31<<11)) * (1.0f / (31<<11));
    *g = cast(wide & (63<< 5)) * (1.0f / (63<< 5));
    *b = cast(wide & (31<< 0)) * (1.0f / (31<< 0));
}
SI void from_4444(U16 _4444, F* r, F* g, F* b, F* a) {
    U32 wide = expand(_4444);
    *r = cast(wide & (15<<12)) * (1.0f / (15<<12));
    *g = cast(wide & (15<< 8)) * (1.0f / (15<< 8));
    *b = cast(wide & (15<< 4)) * (1.0f / (15<< 4));
    *a = cast(wide & (15<< 0)) * (1.0f / (15<< 0));
}
SI void from_8888(U32 _8888, F* r, F* g, F* b, F* a) {
    *r = cast((_8888      ) & 0xff) * (1/255.0f);
    *g = cast((_8888 >>  8) & 0xff) * (1/255.0f);
    *b = cast((_8888 >> 16) & 0xff) * (1/255.0f);
    *a = cast((_8888 >> 24)       ) * (1/255.0f);
}

template <typename T>
SI U32 ix_and_ptr(T** ptr, const SkJumper_GatherCtx* ctx, F x, F y) {
    *ptr = (const T*)ctx->pixels;
    return trunc_(y)*ctx->stride + trunc_(x);
}

// Now finally, normal Stages!

STAGE(seed_shader) {
    // It's important for speed to explicitly cast(x) and cast(y),
    // which has the effect of splatting them to vectors before converting to floats.
    // On Intel this breaks a data dependency on previous loop iterations' registers.
    r = cast(x) + 0.5f + unaligned_load<F>(k->iota_F);
    g = cast(y) + 0.5f;
    b = 1.0f;
    a = 0;
    dr = dg = db = da = 0;
}

STAGE(dither) {
    auto rate = *(const float*)ctx;

    // Get [(x,y), (x+1,y), (x+2,y), ...] loaded up in integer vectors.
    U32 X = x + unaligned_load<U32>(k->iota_U32),
        Y = y;

    // We're doing 8x8 ordered dithering, see https://en.wikipedia.org/wiki/Ordered_dithering.
    // In this case n=8 and we're using the matrix that looks like 1/64 x [ 0 48 12 60 ... ].

    // We only need X and X^Y from here on, so it's easier to just think of that as "Y".
    Y ^= X;

    // We'll mix the bottom 3 bits of each of X and Y to make 6 bits,
    // for 2^6 == 64 == 8x8 matrix values.  If X=abc and Y=def, we make fcebda.
    U32 M = (Y & 1) << 5 | (X & 1) << 4
          | (Y & 2) << 2 | (X & 2) << 1
          | (Y & 4) >> 1 | (X & 4) >> 2;

    // Scale that dither to [0,1), then (-0.5,+0.5), here using 63/128 = 0.4921875 as 0.5-epsilon.
    // We want to make sure our dither is less than 0.5 in either direction to keep exact values
    // like 0 and 1 unchanged after rounding.
    F dither = cast(M) * (2/128.0f) - (63/128.0f);

    r += rate*dither;
    g += rate*dither;
    b += rate*dither;

    r = max(0, min(r, a));
    g = max(0, min(g, a));
    b = max(0, min(b, a));
}

// load 4 floats from memory, and splat them into r,g,b,a
STAGE(uniform_color) {
    auto rgba = (const float*)ctx;
    r = rgba[0];
    g = rgba[1];
    b = rgba[2];
    a = rgba[3];
}

// splats opaque-black into r,g,b,a
STAGE(black_color) {
    r = g = b = 0.0f;
    a = 1.0f;
}

STAGE(white_color) {
    r = g = b = a = 1.0f;
}

// load registers r,g,b,a from context (mirrors store_rgba)
STAGE(load_rgba) {
    auto ptr = (const float*)ctx;
    r = unaligned_load<F>(ptr + 0*kStride);
    g = unaligned_load<F>(ptr + 1*kStride);
    b = unaligned_load<F>(ptr + 2*kStride);
    a = unaligned_load<F>(ptr + 3*kStride);
}

// store registers r,g,b,a into context (mirrors load_rgba)
STAGE(store_rgba) {
    auto ptr = (float*)ctx;
    unaligned_store(ptr + 0*kStride, r);
    unaligned_store(ptr + 1*kStride, g);
    unaligned_store(ptr + 2*kStride, b);
    unaligned_store(ptr + 3*kStride, a);
}

// Most blend modes apply the same logic to each channel.
#define BLEND_MODE(name)                       \
    SI F name##_channel(F s, F d, F sa, F da); \
    STAGE(name) {                              \
        r = name##_channel(r,dr,a,da);         \
        g = name##_channel(g,dg,a,da);         \
        b = name##_channel(b,db,a,da);         \
        a = name##_channel(a,da,a,da);         \
    }                                          \
    SI F name##_channel(F s, F d, F sa, F da)

SI F inv(F x) { return 1.0f - x; }
SI F two(F x) { return x + x; }

BLEND_MODE(clear)    { return 0; }
BLEND_MODE(srcatop)  { return s*da + d*inv(sa); }
BLEND_MODE(dstatop)  { return d*sa + s*inv(da); }
BLEND_MODE(srcin)    { return s * da; }
BLEND_MODE(dstin)    { return d * sa; }
BLEND_MODE(srcout)   { return s * inv(da); }
BLEND_MODE(dstout)   { return d * inv(sa); }
BLEND_MODE(srcover)  { return mad(d, inv(sa), s); }
BLEND_MODE(dstover)  { return mad(s, inv(da), d); }

BLEND_MODE(modulate) { return s*d; }
BLEND_MODE(multiply) { return s*inv(da) + d*inv(sa) + s*d; }
BLEND_MODE(plus_)    { return s + d; }
BLEND_MODE(screen)   { return s + d - s*d; }
BLEND_MODE(xor_)     { return s*inv(da) + d*inv(sa); }
#undef BLEND_MODE

// Most other blend modes apply the same logic to colors, and srcover to alpha.
#define BLEND_MODE(name)                       \
    SI F name##_channel(F s, F d, F sa, F da); \
    STAGE(name) {                              \
        r = name##_channel(r,dr,a,da);         \
        g = name##_channel(g,dg,a,da);         \
        b = name##_channel(b,db,a,da);         \
        a = mad(da, inv(a), a);                \
    }                                          \
    SI F name##_channel(F s, F d, F sa, F da)

BLEND_MODE(darken)     { return s + d -     max(s*da, d*sa) ; }
BLEND_MODE(lighten)    { return s + d -     min(s*da, d*sa) ; }
BLEND_MODE(difference) { return s + d - two(min(s*da, d*sa)); }
BLEND_MODE(exclusion)  { return s + d - two(s*d); }

BLEND_MODE(colorburn) {
    return if_then_else(d == da, d + s*inv(da),
           if_then_else(s ==  0, s + d*inv(sa),
                                 sa*(da - min(da, (da-d)*sa/s)) + s*inv(da) + d*inv(sa)));
}
BLEND_MODE(colordodge) {
    return if_then_else(d ==  0, d + s*inv(da),
           if_then_else(s == sa, s + d*inv(sa),
                                 sa*min(da, (d*sa)/(sa - s)) + s*inv(da) + d*inv(sa)));
}
BLEND_MODE(hardlight) {
    return s*inv(da) + d*inv(sa)
         + if_then_else(two(s) <= sa, two(s*d), sa*da - two((da-d)*(sa-s)));
}
BLEND_MODE(overlay) {
    return s*inv(da) + d*inv(sa)
         + if_then_else(two(d) <= da, two(s*d), sa*da - two((da-d)*(sa-s)));
}

BLEND_MODE(softlight) {
    F m  = if_then_else(da > 0, d / da, 0),
      s2 = two(s),
      m4 = two(two(m));

    // The logic forks three ways:
    //    1. dark src?
    //    2. light src, dark dst?
    //    3. light src, light dst?
    F darkSrc = d*(sa + (s2 - sa)*(1.0f - m)),     // Used in case 1.
      darkDst = (m4*m4 + m4)*(m - 1.0f) + 7.0f*m,  // Used in case 2.
      liteDst = rcp(rsqrt(m)) - m,                 // Used in case 3.
      liteSrc = d*sa + da*(s2 - sa) * if_then_else(two(two(d)) <= da, darkDst, liteDst); // 2 or 3?
    return s*inv(da) + d*inv(sa) + if_then_else(s2 <= sa, darkSrc, liteSrc);      // 1 or (2 or 3)?
}
#undef BLEND_MODE

// We're basing our implemenation of non-separable blend modes on
//   https://www.w3.org/TR/compositing-1/#blendingnonseparable.
// and
//   https://www.khronos.org/registry/OpenGL/specs/es/3.2/es_spec_3.2.pdf
// They're equivalent, but ES' math has been better simplified.
//
// Anything extra we add beyond that is to make the math work with premul inputs.

SI F max(F r, F g, F b) { return max(r, max(g, b)); }
SI F min(F r, F g, F b) { return min(r, min(g, b)); }

SI F sat(F r, F g, F b) { return max(r,g,b) - min(r,g,b); }
SI F lum(F r, F g, F b) { return r*0.30f + g*0.59f + b*0.11f; }

SI void set_sat(F* r, F* g, F* b, F s) {
    F mn  = min(*r,*g,*b),
      mx  = max(*r,*g,*b),
      sat = mx - mn;

    // Map min channel to 0, max channel to s, and scale the middle proportionally.
    auto scale = [=](F c) {
        return if_then_else(sat == 0, 0, (c - mn) * s / sat);
    };
    *r = scale(*r);
    *g = scale(*g);
    *b = scale(*b);
}
SI void set_lum(F* r, F* g, F* b, F l) {
    F diff = l - lum(*r, *g, *b);
    *r += diff;
    *g += diff;
    *b += diff;
}
SI void clip_color(F* r, F* g, F* b, F a) {
    F mn = min(*r, *g, *b),
      mx = max(*r, *g, *b),
      l  = lum(*r, *g, *b);

    auto clip = [=](F c) {
        c = if_then_else(mn >= 0, c, l + (c - l) * (    l) / (l - mn)   );
        c = if_then_else(mx >  a,    l + (c - l) * (a - l) / (mx - l), c);
        c = max(c, 0);  // Sometimes without this we may dip just a little negative.
        return c;
    };
    *r = clip(*r);
    *g = clip(*g);
    *b = clip(*b);
}

STAGE(hue) {
    F R = r*a,
      G = g*a,
      B = b*a;

    set_sat(&R, &G, &B, sat(dr,dg,db)*a);
    set_lum(&R, &G, &B, lum(dr,dg,db)*a);
    clip_color(&R,&G,&B, a*da);

    r = r*inv(da) + dr*inv(a) + R;
    g = g*inv(da) + dg*inv(a) + G;
    b = b*inv(da) + db*inv(a) + B;
    a = a + da - a*da;
}
STAGE(saturation) {
    F R = dr*a,
      G = dg*a,
      B = db*a;

    set_sat(&R, &G, &B, sat( r, g, b)*da);
    set_lum(&R, &G, &B, lum(dr,dg,db)* a);  // (This is not redundant.)
    clip_color(&R,&G,&B, a*da);

    r = r*inv(da) + dr*inv(a) + R;
    g = g*inv(da) + dg*inv(a) + G;
    b = b*inv(da) + db*inv(a) + B;
    a = a + da - a*da;
}
STAGE(color) {
    F R = r*da,
      G = g*da,
      B = b*da;

    set_lum(&R, &G, &B, lum(dr,dg,db)*a);
    clip_color(&R,&G,&B, a*da);

    r = r*inv(da) + dr*inv(a) + R;
    g = g*inv(da) + dg*inv(a) + G;
    b = b*inv(da) + db*inv(a) + B;
    a = a + da - a*da;
}
STAGE(luminosity) {
    F R = dr*a,
      G = dg*a,
      B = db*a;

    set_lum(&R, &G, &B, lum(r,g,b)*da);
    clip_color(&R,&G,&B, a*da);

    r = r*inv(da) + dr*inv(a) + R;
    g = g*inv(da) + dg*inv(a) + G;
    b = b*inv(da) + db*inv(a) + B;
    a = a + da - a*da;
}

STAGE(srcover_rgba_8888) {
    auto ptr = *(uint32_t**)ctx + x;

    U32 dst = load<U32>(ptr, tail);
    dr = cast((dst      ) & 0xff);
    dg = cast((dst >>  8) & 0xff);
    db = cast((dst >> 16) & 0xff);
    da = cast((dst >> 24)       );
    // {dr,dg,db,da} are in [0,255]
    // { r, g, b, a} are in [0,  1]

    r = mad(dr, inv(a), r*255.0f);
    g = mad(dg, inv(a), g*255.0f);
    b = mad(db, inv(a), b*255.0f);
    a = mad(da, inv(a), a*255.0f);
    // { r, g, b, a} are now in [0,255]

    dst = round(r, 1.0f)
        | round(g, 1.0f) <<  8
        | round(b, 1.0f) << 16
        | round(a, 1.0f) << 24;
    store(ptr, dst, tail);
}

STAGE(clamp_0) {
    r = max(r, 0);
    g = max(g, 0);
    b = max(b, 0);
    a = max(a, 0);
}

STAGE(clamp_1) {
    r = min(r, 1.0f);
    g = min(g, 1.0f);
    b = min(b, 1.0f);
    a = min(a, 1.0f);
}

STAGE(clamp_a) {
    a = min(a, 1.0f);
    r = min(r, a);
    g = min(g, a);
    b = min(b, a);
}

STAGE(clamp_a_dst) {
    da = min(da, 1.0f);
    dr = min(dr, da);
    dg = min(dg, da);
    db = min(db, da);
}

STAGE(set_rgb) {
    auto rgb = (const float*)ctx;
    r = rgb[0];
    g = rgb[1];
    b = rgb[2];
}
STAGE(swap_rb) {
    auto tmp = r;
    r = b;
    b = tmp;
}

STAGE(move_src_dst) {
    dr = r;
    dg = g;
    db = b;
    da = a;
}
STAGE(move_dst_src) {
    r = dr;
    g = dg;
    b = db;
    a = da;
}

STAGE(premul) {
    r = r * a;
    g = g * a;
    b = b * a;
}
STAGE(unpremul) {
    auto scale = if_then_else(a == 0, 0, 1.0f / a);
    r *= scale;
    g *= scale;
    b *= scale;
}

SI F from_srgb(F s) {
    auto lo = s * (1/12.92f);
    auto hi = mad(s*s, mad(s, 0.3000f, 0.6975f), 0.0025f);
    return if_then_else(s < 0.055f, lo, hi);
}

STAGE(from_srgb) {
    r = from_srgb(r);
    g = from_srgb(g);
    b = from_srgb(b);
}
STAGE(from_srgb_dst) {
    dr = from_srgb(dr);
    dg = from_srgb(dg);
    db = from_srgb(db);
}
STAGE(to_srgb) {
    auto fn = [&](F l) {
        // We tweak c and d for each instruction set to make sure fn(1) is exactly 1.
    #if defined(JUMPER) && defined(__SSE2__)
        const float c = 1.130048394203f,
                    d = 0.141357362270f;
    #elif defined(JUMPER) && (defined(__aarch64__) || defined(__arm__))
        const float c = 1.129999995232f,
                    d = 0.141381442547f;
    #else
        const float c = 1.129999995232f,
                    d = 0.141377761960f;
    #endif
        F t = rsqrt(l);
        auto lo = l * 12.92f;
        auto hi = mad(t, mad(t, -0.0024542345f, 0.013832027f), c)
                * rcp(d + t);
        return if_then_else(l < 0.00465985f, lo, hi);
    };
    r = fn(r);
    g = fn(g);
    b = fn(b);
}

STAGE(rgb_to_hsl) {
    F mx = max(max(r,g), b),
      mn = min(min(r,g), b),
      d = mx - mn,
      d_rcp = 1.0f / d;

    F h = (1/6.0f) *
          if_then_else(mx == mn, 0,
          if_then_else(mx ==  r, (g-b)*d_rcp + if_then_else(g < b, 6.0f, 0),
          if_then_else(mx ==  g, (b-r)*d_rcp + 2.0f,
                                 (r-g)*d_rcp + 4.0f)));

    F l = (mx + mn) * 0.5f;
    F s = if_then_else(mx == mn, 0,
                       d / if_then_else(l > 0.5f, 2.0f-mx-mn, mx+mn));

    r = h;
    g = s;
    b = l;
}
STAGE(hsl_to_rgb) {
    F h = r,
      s = g,
      l = b;

    F q = l + if_then_else(l >= 0.5f, s - l*s, l*s),
      p = 2.0f*l - q;

    auto hue_to_rgb = [&](F t) {
        t = fract(t);

        F r = p;
        r = if_then_else(t >= 4/6.0f, r, p + (q-p)*(4.0f - 6.0f*t));
        r = if_then_else(t >= 3/6.0f, r, q);
        r = if_then_else(t >= 1/6.0f, r, p + (q-p)*(       6.0f*t));
        return r;
    };

    r = if_then_else(s == 0, l, hue_to_rgb(h + (1/3.0f)));
    g = if_then_else(s == 0, l, hue_to_rgb(h           ));
    b = if_then_else(s == 0, l, hue_to_rgb(h - (1/3.0f)));
}

STAGE(scale_1_float) {
    auto c = *(const float*)ctx;

    r = r * c;
    g = g * c;
    b = b * c;
    a = a * c;
}
STAGE(scale_u8) {
    auto ptr = *(const uint8_t**)ctx + x;

    auto scales = load<U8>(ptr, tail);
    auto c = from_byte(scales);

    r = r * c;
    g = g * c;
    b = b * c;
    a = a * c;
}

SI F lerp(F from, F to, F t) {
    return mad(to-from, t, from);
}

STAGE(lerp_1_float) {
    auto c = *(const float*)ctx;

    r = lerp(dr, r, c);
    g = lerp(dg, g, c);
    b = lerp(db, b, c);
    a = lerp(da, a, c);
}
STAGE(lerp_u8) {
    auto ptr = *(const uint8_t**)ctx + x;

    auto scales = load<U8>(ptr, tail);
    auto c = from_byte(scales);

    r = lerp(dr, r, c);
    g = lerp(dg, g, c);
    b = lerp(db, b, c);
    a = lerp(da, a, c);
}
STAGE(lerp_565) {
    auto ptr = *(const uint16_t**)ctx + x;

    F cr,cg,cb;
    from_565(load<U16>(ptr, tail), &cr, &cg, &cb);

    r = lerp(dr, r, cr);
    g = lerp(dg, g, cg);
    b = lerp(db, b, cb);
    a = max(lerp(da, a, cr), lerp(da, a, cg), lerp(da, a, cb));
}

STAGE(load_tables) {
    auto c = (const SkJumper_LoadTablesCtx*)ctx;

    auto px = load<U32>((const uint32_t*)c->src + x, tail);
    r = gather(c->r, (px      ) & 0xff);
    g = gather(c->g, (px >>  8) & 0xff);
    b = gather(c->b, (px >> 16) & 0xff);
    a = cast(        (px >> 24)) * (1/255.0f);
}
STAGE(load_tables_u16_be) {
    auto c = (const SkJumper_LoadTablesCtx*)ctx;
    auto ptr = (const uint16_t*)c->src + 4*x;

    U16 R,G,B,A;
    load4(ptr, tail, &R,&G,&B,&A);

    // c->src is big-endian, so & 0xff grabs the 8 most signficant bits.
    r = gather(c->r, expand(R) & 0xff);
    g = gather(c->g, expand(G) & 0xff);
    b = gather(c->b, expand(B) & 0xff);
    a = (1/65535.0f) * cast(expand(bswap(A)));
}
STAGE(load_tables_rgb_u16_be) {
    auto c = (const SkJumper_LoadTablesCtx*)ctx;
    auto ptr = (const uint16_t*)c->src + 3*x;

    U16 R,G,B;
    load3(ptr, tail, &R,&G,&B);

    // c->src is big-endian, so & 0xff grabs the 8 most signficant bits.
    r = gather(c->r, expand(R) & 0xff);
    g = gather(c->g, expand(G) & 0xff);
    b = gather(c->b, expand(B) & 0xff);
    a = 1.0f;
}

STAGE(byte_tables) {
    struct Tables { const uint8_t *r, *g, *b, *a; };
    auto tables = (const Tables*)ctx;

    r = from_byte(gather(tables->r, round(r, 255.0f)));
    g = from_byte(gather(tables->g, round(g, 255.0f)));
    b = from_byte(gather(tables->b, round(b, 255.0f)));
    a = from_byte(gather(tables->a, round(a, 255.0f)));
}

STAGE(byte_tables_rgb) {
    struct Tables { const uint8_t *r, *g, *b; int n; };
    auto tables = (const Tables*)ctx;

    F scale = tables->n - 1;
    r = from_byte(gather(tables->r, round(r, scale)));
    g = from_byte(gather(tables->g, round(g, scale)));
    b = from_byte(gather(tables->b, round(b, scale)));
}

SI F table(F v, const SkJumper_TableCtx* ctx) {
    return gather(ctx->table, round(v, ctx->size - 1));
}
STAGE(table_r) { r = table(r, ctx); }
STAGE(table_g) { g = table(g, ctx); }
STAGE(table_b) { b = table(b, ctx); }
STAGE(table_a) { a = table(a, ctx); }

SI F parametric(F v, const SkJumper_ParametricTransferFunction* ctx) {
    F r = if_then_else(v <= ctx->D, mad(ctx->C, v, ctx->F)
                                  , approx_powf(mad(ctx->A, v, ctx->B), ctx->G) + ctx->E);
    return min(max(r, 0), 1.0f);  // Clamp to [0,1], with argument order mattering to handle NaN.
}
STAGE(parametric_r) { r = parametric(r, ctx); }
STAGE(parametric_g) { g = parametric(g, ctx); }
STAGE(parametric_b) { b = parametric(b, ctx); }
STAGE(parametric_a) { a = parametric(a, ctx); }

STAGE(lab_to_xyz) {
    F L = r * 100.0f,
      A = g * 255.0f - 128.0f,
      B = b * 255.0f - 128.0f;

    F Y = (L + 16.0f) * (1/116.0f),
      X = Y + A*(1/500.0f),
      Z = Y - B*(1/200.0f);

    X = if_then_else(X*X*X > 0.008856f, X*X*X, (X - (16/116.0f)) * (1/7.787f));
    Y = if_then_else(Y*Y*Y > 0.008856f, Y*Y*Y, (Y - (16/116.0f)) * (1/7.787f));
    Z = if_then_else(Z*Z*Z > 0.008856f, Z*Z*Z, (Z - (16/116.0f)) * (1/7.787f));

    // Adjust to D50 illuminant.
    r = X * 0.96422f;
    g = Y           ;
    b = Z * 0.82521f;
}

STAGE(load_a8) {
    auto ptr = *(const uint8_t**)ctx + x;

    r = g = b = 0.0f;
    a = from_byte(load<U8>(ptr, tail));
}
STAGE(load_a8_dst) {
    auto ptr = *(const uint8_t**)ctx + x;

    dr = dg = db = 0.0f;
    da = from_byte(load<U8>(ptr, tail));
}
STAGE(gather_a8) {
    const uint8_t* ptr;
    U32 ix = ix_and_ptr(&ptr, ctx, r,g);
    r = g = b = 0.0f;
    a = from_byte(gather(ptr, ix));
}
STAGE(store_a8) {
    auto ptr = *(uint8_t**)ctx + x;

    U8 packed = pack(pack(round(a, 255.0f)));
    store(ptr, packed, tail);
}

STAGE(load_g8) {
    auto ptr = *(const uint8_t**)ctx + x;

    r = g = b = from_byte(load<U8>(ptr, tail));
    a = 1.0f;
}
STAGE(load_g8_dst) {
    auto ptr = *(const uint8_t**)ctx + x;

    dr = dg = db = from_byte(load<U8>(ptr, tail));
    da = 1.0f;
}
STAGE(gather_g8) {
    const uint8_t* ptr;
    U32 ix = ix_and_ptr(&ptr, ctx, r,g);
    r = g = b = from_byte(gather(ptr, ix));
    a = 1.0f;
}

STAGE(gather_i8) {
    auto c = (const SkJumper_GatherCtx*)ctx;
    const uint8_t* ptr;
    U32 ix = ix_and_ptr(&ptr, ctx, r,g);
    ix = expand(gather(ptr, ix));
    from_8888(gather(c->ctable, ix), &r,&g,&b,&a);
}

STAGE(load_565) {
    auto ptr = *(const uint16_t**)ctx + x;

    from_565(load<U16>(ptr, tail), &r,&g,&b);
    a = 1.0f;
}
STAGE(load_565_dst) {
    auto ptr = *(const uint16_t**)ctx + x;

    from_565(load<U16>(ptr, tail), &dr,&dg,&db);
    da = 1.0f;
}
STAGE(gather_565) {
    const uint16_t* ptr;
    U32 ix = ix_and_ptr(&ptr, ctx, r,g);
    from_565(gather(ptr, ix), &r,&g,&b);
    a = 1.0f;
}
STAGE(store_565) {
    auto ptr = *(uint16_t**)ctx + x;

    U16 px = pack( round(r, 31.0f) << 11
                 | round(g, 63.0f) <<  5
                 | round(b, 31.0f)      );
    store(ptr, px, tail);
}

STAGE(load_4444) {
    auto ptr = *(const uint16_t**)ctx + x;
    from_4444(load<U16>(ptr, tail), &r,&g,&b,&a);
}
STAGE(load_4444_dst) {
    auto ptr = *(const uint16_t**)ctx + x;
    from_4444(load<U16>(ptr, tail), &dr,&dg,&db,&da);
}
STAGE(gather_4444) {
    const uint16_t* ptr;
    U32 ix = ix_and_ptr(&ptr, ctx, r,g);
    from_4444(gather(ptr, ix), &r,&g,&b,&a);
}
STAGE(store_4444) {
    auto ptr = *(uint16_t**)ctx + x;
    U16 px = pack( round(r, 15.0f) << 12
                 | round(g, 15.0f) <<  8
                 | round(b, 15.0f) <<  4
                 | round(a, 15.0f)      );
    store(ptr, px, tail);
}

STAGE(load_8888) {
    auto ptr = *(const uint32_t**)ctx + x;
    from_8888(load<U32>(ptr, tail), &r,&g,&b,&a);
}
STAGE(load_8888_dst) {
    auto ptr = *(const uint32_t**)ctx + x;
    from_8888(load<U32>(ptr, tail), &dr,&dg,&db,&da);
}
STAGE(gather_8888) {
    const uint32_t* ptr;
    U32 ix = ix_and_ptr(&ptr, ctx, r,g);
    from_8888(gather(ptr, ix), &r,&g,&b,&a);
}
STAGE(store_8888) {
    auto ptr = *(uint32_t**)ctx + x;

    U32 px = round(r, 255.0f)
           | round(g, 255.0f) <<  8
           | round(b, 255.0f) << 16
           | round(a, 255.0f) << 24;
    store(ptr, px, tail);
}

STAGE(load_bgra) {
    auto ptr = *(const uint32_t**)ctx + x;
    from_8888(load<U32>(ptr, tail), &b,&g,&r,&a);
}
STAGE(load_bgra_dst) {
    auto ptr = *(const uint32_t**)ctx + x;
    from_8888(load<U32>(ptr, tail), &db,&dg,&dr,&da);
}
STAGE(gather_bgra) {
    const uint32_t* ptr;
    U32 ix = ix_and_ptr(&ptr, ctx, r,g);
    from_8888(gather(ptr, ix), &b,&g,&r,&a);
}
STAGE(store_bgra) {
    auto ptr = *(uint32_t**)ctx + x;

    U32 px = round(b, 255.0f)
           | round(g, 255.0f) <<  8
           | round(r, 255.0f) << 16
           | round(a, 255.0f) << 24;
    store(ptr, px, tail);
}

STAGE(load_f16) {
    auto ptr = *(const uint64_t**)ctx + x;

    U16 R,G,B,A;
    load4((const uint16_t*)ptr,tail, &R,&G,&B,&A);
    r = from_half(R);
    g = from_half(G);
    b = from_half(B);
    a = from_half(A);
}
STAGE(load_f16_dst) {
    auto ptr = *(const uint64_t**)ctx + x;

    U16 R,G,B,A;
    load4((const uint16_t*)ptr,tail, &R,&G,&B,&A);
    dr = from_half(R);
    dg = from_half(G);
    db = from_half(B);
    da = from_half(A);
}
STAGE(gather_f16) {
    const uint64_t* ptr;
    U32 ix = ix_and_ptr(&ptr, ctx, r,g);
    auto px = gather(ptr, ix);

    U16 R,G,B,A;
    load4((const uint16_t*)&px,0, &R,&G,&B,&A);
    r = from_half(R);
    g = from_half(G);
    b = from_half(B);
    a = from_half(A);
}
STAGE(store_f16) {
    auto ptr = *(uint64_t**)ctx + x;
    store4((uint16_t*)ptr,tail, to_half(r)
                              , to_half(g)
                              , to_half(b)
                              , to_half(a));
}

STAGE(load_u16_be) {
    auto ptr = *(const uint16_t**)ctx + 4*x;

    U16 R,G,B,A;
    load4(ptr,tail, &R,&G,&B,&A);

    r = (1/65535.0f) * cast(expand(bswap(R)));
    g = (1/65535.0f) * cast(expand(bswap(G)));
    b = (1/65535.0f) * cast(expand(bswap(B)));
    a = (1/65535.0f) * cast(expand(bswap(A)));
}
STAGE(load_rgb_u16_be) {
    auto ptr = *(const uint16_t**)ctx + 3*x;

    U16 R,G,B;
    load3(ptr,tail, &R,&G,&B);

    r = (1/65535.0f) * cast(expand(bswap(R)));
    g = (1/65535.0f) * cast(expand(bswap(G)));
    b = (1/65535.0f) * cast(expand(bswap(B)));
    a = 1.0f;
}
STAGE(store_u16_be) {
    auto ptr = *(uint16_t**)ctx + 4*x;

    U16 R = bswap(pack(round(r, 65535.0f))),
        G = bswap(pack(round(g, 65535.0f))),
        B = bswap(pack(round(b, 65535.0f))),
        A = bswap(pack(round(a, 65535.0f)));

    store4(ptr,tail, R,G,B,A);
}

STAGE(load_f32) {
    auto ptr = *(const float**)ctx + 4*x;
    load4(ptr,tail, &r,&g,&b,&a);
}
STAGE(load_f32_dst) {
    auto ptr = *(const float**)ctx + 4*x;
    load4(ptr,tail, &dr,&dg,&db,&da);
}
STAGE(store_f32) {
    auto ptr = *(float**)ctx + 4*x;
    store4(ptr,tail, r,g,b,a);
}

SI F ulp_before(F f) {
    U32 bits = -1 + unaligned_load<U32>(&f);
    return unaligned_load<F>(&bits);
}

SI F exclusive_clamp(F v, const SkJumper_TileCtx* ctx) {
    v = max(0,v);
    return min(v, ulp_before(ctx->scale));
}
SI F exclusive_repeat(F v, const SkJumper_TileCtx* ctx) {
    v = v - floor_(v*ctx->invScale)*ctx->scale;
    return min(v, ulp_before(ctx->scale));
}
SI F exclusive_mirror(F v, const SkJumper_TileCtx* ctx) {
    auto limit = ctx->scale;
    auto invLimit = ctx->invScale;
    v = abs_( (v-limit) - (limit+limit)*floor_((v-limit)*(invLimit*0.5f)) - limit );
    return min(v, ulp_before(limit));
}
// Clamp x or y to [0,limit) == [0,limit - 1 ulp] (think, sampling from images).
STAGE(clamp_x)  { r = exclusive_clamp (r, (const SkJumper_TileCtx*)ctx); }
STAGE(clamp_y)  { g = exclusive_clamp (g, (const SkJumper_TileCtx*)ctx); }
STAGE(repeat_x) { r = exclusive_repeat(r, (const SkJumper_TileCtx*)ctx); }
STAGE(repeat_y) { g = exclusive_repeat(g, (const SkJumper_TileCtx*)ctx); }
STAGE(mirror_x) { r = exclusive_mirror(r, (const SkJumper_TileCtx*)ctx); }
STAGE(mirror_y) { g = exclusive_mirror(g, (const SkJumper_TileCtx*)ctx); }

// Clamp x to [0,1], both sides exclusive (think, gradients).
STAGE( clamp_x_1) { r = min(max(0, r), 1.0f); }
STAGE(repeat_x_1) { r = r - floor_(r); }
STAGE(mirror_x_1) { r = abs_( (r-1.0f) - two(floor_((r-1.0f)*0.5f)) - 1.0f ); }

STAGE(luminance_to_alpha) {
    a = r*0.2126f + g*0.7152f + b*0.0722f;
    r = g = b = 0;
}

STAGE(matrix_translate) {
    auto m = (const float*)ctx;

    r += m[0];
    g += m[1];
}
STAGE(matrix_scale_translate) {
    auto m = (const float*)ctx;

    r = mad(r,m[2], m[0]);
    g = mad(g,m[3], m[1]);
}
STAGE(matrix_2x3) {
    auto m = (const float*)ctx;

    auto R = mad(r,m[0], mad(g,m[2], m[4])),
         G = mad(r,m[1], mad(g,m[3], m[5]));
    r = R;
    g = G;
}
STAGE(matrix_3x4) {
    auto m = (const float*)ctx;

    auto R = mad(r,m[0], mad(g,m[3], mad(b,m[6], m[ 9]))),
         G = mad(r,m[1], mad(g,m[4], mad(b,m[7], m[10]))),
         B = mad(r,m[2], mad(g,m[5], mad(b,m[8], m[11])));
    r = R;
    g = G;
    b = B;
}
STAGE(matrix_4x5) {
    auto m = (const float*)ctx;

    auto R = mad(r,m[0], mad(g,m[4], mad(b,m[ 8], mad(a,m[12], m[16])))),
         G = mad(r,m[1], mad(g,m[5], mad(b,m[ 9], mad(a,m[13], m[17])))),
         B = mad(r,m[2], mad(g,m[6], mad(b,m[10], mad(a,m[14], m[18])))),
         A = mad(r,m[3], mad(g,m[7], mad(b,m[11], mad(a,m[15], m[19]))));
    r = R;
    g = G;
    b = B;
    a = A;
}
STAGE(matrix_4x3) {
    auto m = (const float*)ctx;
    auto X = r,
         Y = g;

    r = mad(X, m[0], mad(Y, m[4], m[ 8]));
    g = mad(X, m[1], mad(Y, m[5], m[ 9]));
    b = mad(X, m[2], mad(Y, m[6], m[10]));
    a = mad(X, m[3], mad(Y, m[7], m[11]));
}
STAGE(matrix_perspective) {
    // N.B. Unlike the other matrix_ stages, this matrix is row-major.
    auto m = (const float*)ctx;

    auto R = mad(r,m[0], mad(g,m[1], m[2])),
         G = mad(r,m[3], mad(g,m[4], m[5])),
         Z = mad(r,m[6], mad(g,m[7], m[8]));
    r = R * rcp(Z);
    g = G * rcp(Z);
}

SI void gradient_lookup(const SkJumper_GradientCtx* c, U32 idx, F t,
                        F* r, F* g, F* b, F* a) {
    F fr, br, fg, bg, fb, bb, fa, ba;
#if defined(JUMPER) && defined(__AVX2__)
    if (c->stopCount <=8) {
        fr = _mm256_permutevar8x32_ps(_mm256_loadu_ps(c->fs[0]), idx);
        br = _mm256_permutevar8x32_ps(_mm256_loadu_ps(c->bs[0]), idx);
        fg = _mm256_permutevar8x32_ps(_mm256_loadu_ps(c->fs[1]), idx);
        bg = _mm256_permutevar8x32_ps(_mm256_loadu_ps(c->bs[1]), idx);
        fb = _mm256_permutevar8x32_ps(_mm256_loadu_ps(c->fs[2]), idx);
        bb = _mm256_permutevar8x32_ps(_mm256_loadu_ps(c->bs[2]), idx);
        fa = _mm256_permutevar8x32_ps(_mm256_loadu_ps(c->fs[3]), idx);
        ba = _mm256_permutevar8x32_ps(_mm256_loadu_ps(c->bs[3]), idx);
    } else
#endif
    {
        fr = gather(c->fs[0], idx);
        br = gather(c->bs[0], idx);
        fg = gather(c->fs[1], idx);
        bg = gather(c->bs[1], idx);
        fb = gather(c->fs[2], idx);
        bb = gather(c->bs[2], idx);
        fa = gather(c->fs[3], idx);
        ba = gather(c->bs[3], idx);
    }

    *r = mad(t, fr, br);
    *g = mad(t, fg, bg);
    *b = mad(t, fb, bb);
    *a = mad(t, fa, ba);
}

STAGE(evenly_spaced_gradient) {
    auto c = (const SkJumper_GradientCtx*)ctx;
    auto t = r;
    auto idx = trunc_(t * (c->stopCount-1));
    gradient_lookup(c, idx, t, &r, &g, &b, &a);
}

STAGE(gauss_a_to_rgba) {
    // x = 1 - x;
    // exp(-x * x * 4) - 0.018f;
    // ... now approximate with quartic
    //
    const float c4 = -2.26661229133605957031f;
    const float c3 = 2.89795351028442382812f;
    const float c2 = 0.21345567703247070312f;
    const float c1 = 0.15489584207534790039f;
    const float c0 = 0.00030726194381713867f;
    a = mad(a, mad(a, mad(a, mad(a, c4, c3), c2), c1), c0);
    r = a;
    g = a;
    b = a;
}

STAGE(gradient) {
    auto c = (const SkJumper_GradientCtx*)ctx;
    auto t = r;
    U32 idx = 0;

    // N.B. The loop starts at 1 because idx 0 is the color to use before the first stop.
    for (size_t i = 1; i < c->stopCount; i++) {
        idx += if_then_else(t >= c->ts[i], U32(1), U32(0));
    }

    gradient_lookup(c, idx, t, &r, &g, &b, &a);
}

STAGE(evenly_spaced_2_stop_gradient) {
    struct Ctx { float f[4], b[4]; };
    auto c = (const Ctx*)ctx;

    auto t = r;
    r = mad(t, c->f[0], c->b[0]);
    g = mad(t, c->f[1], c->b[1]);
    b = mad(t, c->f[2], c->b[2]);
    a = mad(t, c->f[3], c->b[3]);
}

STAGE(xy_to_unit_angle) {
    F X = r,
      Y = g;
    F xabs = abs_(X),
      yabs = abs_(Y);

    F slope = min(xabs, yabs)/max(xabs, yabs);
    F s = slope * slope;

    // Use a 7th degree polynomial to approximate atan.
    // This was generated using sollya.gforge.inria.fr.
    // A float optimized polynomial was generated using the following command.
    // P1 = fpminimax((1/(2*Pi))*atan(x),[|1,3,5,7|],[|24...|],[2^(-40),1],relative);
    F phi = slope
             * (0.15912117063999176025390625f     + s
             * (-5.185396969318389892578125e-2f   + s
             * (2.476101927459239959716796875e-2f + s
             * (-7.0547382347285747528076171875e-3f))));

    phi = if_then_else(xabs < yabs, 1.0f/4.0f - phi, phi);
    phi = if_then_else(X < 0.0f   , 1.0f/2.0f - phi, phi);
    phi = if_then_else(Y < 0.0f   , 1.0f - phi     , phi);
    phi = if_then_else(phi != phi , 0              , phi);  // Check for NaN.
    r = phi;
}

STAGE(xy_to_radius) {
    F X2 = r * r,
      Y2 = g * g;
    r = sqrt_(X2 + Y2);
}

SI F solve_2pt_conical_quadratic(const SkJumper_2PtConicalCtx* c, F x, F y, F (*select)(F, F)) {
    // At this point, (x, y) is mapped into a synthetic gradient space with
    // the start circle centerd on (0, 0), and the end circle centered on (1, 0)
    // (see the stage setup).
    //
    // We're searching along X-axis for x', such that
    //
    //   1) r(x') is a linear interpolation between r0 and r1
    //   2) (x, y) is on the circle C(x', 0, r(x'))
    //
    // Solving this system boils down to a quadratic equation with coefficients
    //
    //   a = 1 - (r1 - r0)^2             <- constant, precomputed in ctx->fCoeffA)
    //
    //   b = -2 * (x + (r1 - r0) * r0)
    //
    //   c = x^2 + y^2 - r0^2
    //
    // Since the start/end circle centers are the extremes of the [0, 1] interval
    // on the X axis, the solution (x') is exactly the t we are looking for.

    const F coeffA = c->fCoeffA,
            coeffB = -2 * (x + c->fDR*c->fR0),
            coeffC = x*x + y*y - c->fR0*c->fR0;

    const F disc      = mad(coeffB, coeffB, -4 * coeffA * coeffC);
    const F sqrt_disc = sqrt_(disc);

    const F invCoeffA = c->fInvCoeffA;
    return select((-coeffB + sqrt_disc) * (invCoeffA * 0.5f),
                  (-coeffB - sqrt_disc) * (invCoeffA * 0.5f));
}

STAGE(xy_to_2pt_conical_quadratic_max) {
    r = solve_2pt_conical_quadratic(ctx, r, g, max);
}

STAGE(xy_to_2pt_conical_quadratic_min) {
    r = solve_2pt_conical_quadratic(ctx, r, g, min);
}

STAGE(xy_to_2pt_conical_linear) {
    auto* c = (const SkJumper_2PtConicalCtx*)ctx;

    const F coeffB = -2 * (r + c->fDR*c->fR0),
            coeffC = r*r + g*g - c->fR0*c->fR0;

    r = -coeffC / coeffB;
}

STAGE(mask_2pt_conical_degenerates) {
    auto* c = (SkJumper_2PtConicalCtx*)ctx;

    // Compute and save a mask for degenerate values.
    U32 mask = 0xffffffff;

    // TODO: mtklein kindly volunteered to revisit this at some point.
#if defined(JUMPER)
    // Vector comparisons set all bits, so we can use something like this.
    mask = mask & (mad(r, c->fDR, c->fR0) >= 0);  // R(t) >= 0
    mask = mask & (r == r);                       // t != NaN
#else
    // The portable version is more involved, 'cause we only get one bit back.
    mask = mask & if_then_else(mad(r, c->fDR, c->fR0) >= 0, U32(0xffffffff), U32(0)); // R(t) >= 0
    mask = mask & if_then_else(r == r,                      U32(0xffffffff), U32(0)); // t != NaN
#endif

    unaligned_store(&c->fMask, mask);
}

STAGE(apply_vector_mask) {
    const U32 mask = unaligned_load<U32>((const uint32_t*)ctx);
    r = bit_cast<F>(bit_cast<U32>(r) & mask);
    g = bit_cast<F>(bit_cast<U32>(g) & mask);
    b = bit_cast<F>(bit_cast<U32>(b) & mask);
    a = bit_cast<F>(bit_cast<U32>(a) & mask);
}

STAGE(save_xy) {
    auto c = (SkJumper_SamplerCtx*)ctx;

    // Whether bilinear or bicubic, all sample points are at the same fractional offset (fx,fy).
    // They're either the 4 corners of a logical 1x1 pixel or the 16 corners of a 3x3 grid
    // surrounding (x,y) at (0.5,0.5) off-center.
    F fx = fract(r + 0.5f),
      fy = fract(g + 0.5f);

    // Samplers will need to load x and fx, or y and fy.
    unaligned_store(c->x,  r);
    unaligned_store(c->y,  g);
    unaligned_store(c->fx, fx);
    unaligned_store(c->fy, fy);
}

STAGE(accumulate) {
    auto c = (const SkJumper_SamplerCtx*)ctx;

    // Bilinear and bicubic filters are both separable, so we produce independent contributions
    // from x and y, multiplying them together here to get each pixel's total scale factor.
    auto scale = unaligned_load<F>(c->scalex)
               * unaligned_load<F>(c->scaley);
    dr = mad(scale, r, dr);
    dg = mad(scale, g, dg);
    db = mad(scale, b, db);
    da = mad(scale, a, da);
}

// In bilinear interpolation, the 4 pixels at +/- 0.5 offsets from the sample pixel center
// are combined in direct proportion to their area overlapping that logical query pixel.
// At positive offsets, the x-axis contribution to that rectangle is fx, or (1-fx) at negative x.
// The y-axis is symmetric.

template <int kScale>
SI void bilinear_x(SkJumper_SamplerCtx* ctx, F* x) {
    *x = unaligned_load<F>(ctx->x) + (kScale * 0.5f);
    F fx = unaligned_load<F>(ctx->fx);

    F scalex;
    if (kScale == -1) { scalex = 1.0f - fx; }
    if (kScale == +1) { scalex =        fx; }
    unaligned_store(ctx->scalex, scalex);
}
template <int kScale>
SI void bilinear_y(SkJumper_SamplerCtx* ctx, F* y) {
    *y = unaligned_load<F>(ctx->y) + (kScale * 0.5f);
    F fy = unaligned_load<F>(ctx->fy);

    F scaley;
    if (kScale == -1) { scaley = 1.0f - fy; }
    if (kScale == +1) { scaley =        fy; }
    unaligned_store(ctx->scaley, scaley);
}

STAGE(bilinear_nx) { bilinear_x<-1>(ctx, &r); }
STAGE(bilinear_px) { bilinear_x<+1>(ctx, &r); }
STAGE(bilinear_ny) { bilinear_y<-1>(ctx, &g); }
STAGE(bilinear_py) { bilinear_y<+1>(ctx, &g); }


// In bicubic interpolation, the 16 pixels and +/- 0.5 and +/- 1.5 offsets from the sample
// pixel center are combined with a non-uniform cubic filter, with higher values near the center.
//
// We break this function into two parts, one for near 0.5 offsets and one for far 1.5 offsets.
// See GrCubicEffect for details of this particular filter.

SI F bicubic_near(F t) {
    // 1/18 + 9/18t + 27/18t^2 - 21/18t^3 == t ( t ( -21/18t + 27/18) + 9/18) + 1/18
    return mad(t, mad(t, mad((-21/18.0f), t, (27/18.0f)), (9/18.0f)), (1/18.0f));
}
SI F bicubic_far(F t) {
    // 0/18 + 0/18*t - 6/18t^2 + 7/18t^3 == t^2 (7/18t - 6/18)
    return (t*t)*mad((7/18.0f), t, (-6/18.0f));
}

template <int kScale>
SI void bicubic_x(SkJumper_SamplerCtx* ctx, F* x) {
    *x = unaligned_load<F>(ctx->x) + (kScale * 0.5f);
    F fx = unaligned_load<F>(ctx->fx);

    F scalex;
    if (kScale == -3) { scalex = bicubic_far (1.0f - fx); }
    if (kScale == -1) { scalex = bicubic_near(1.0f - fx); }
    if (kScale == +1) { scalex = bicubic_near(       fx); }
    if (kScale == +3) { scalex = bicubic_far (       fx); }
    unaligned_store(ctx->scalex, scalex);
}
template <int kScale>
SI void bicubic_y(SkJumper_SamplerCtx* ctx, F* y) {
    *y = unaligned_load<F>(ctx->y) + (kScale * 0.5f);
    F fy = unaligned_load<F>(ctx->fy);

    F scaley;
    if (kScale == -3) { scaley = bicubic_far (1.0f - fy); }
    if (kScale == -1) { scaley = bicubic_near(1.0f - fy); }
    if (kScale == +1) { scaley = bicubic_near(       fy); }
    if (kScale == +3) { scaley = bicubic_far (       fy); }
    unaligned_store(ctx->scaley, scaley);
}

STAGE(bicubic_n3x) { bicubic_x<-3>(ctx, &r); }
STAGE(bicubic_n1x) { bicubic_x<-1>(ctx, &r); }
STAGE(bicubic_p1x) { bicubic_x<+1>(ctx, &r); }
STAGE(bicubic_p3x) { bicubic_x<+3>(ctx, &r); }

STAGE(bicubic_n3y) { bicubic_y<-3>(ctx, &g); }
STAGE(bicubic_n1y) { bicubic_y<-1>(ctx, &g); }
STAGE(bicubic_p1y) { bicubic_y<+1>(ctx, &g); }
STAGE(bicubic_p3y) { bicubic_y<+3>(ctx, &g); }

STAGE(callback) {
    auto c = (SkJumper_CallbackCtx*)ctx;
    store4(c->rgba,0, r,g,b,a);
    c->fn(c, tail ? tail : kStride);
    load4(c->read_from,0, &r,&g,&b,&a);
}