threers 0.0.7

Drop-in three.js replacement for Rust and WebAssembly (wgpu) — THREE.* API, scene graph, PBR, post-fx, mesh BVH, CSG, an OpenSCAD front end with a watertight exact-CSG kernel, headless render, and native media codecs
Documentation
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
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
//! A three-dimensional fluid: the same solver, a volume instead of a plane.
//!
//! ```no_run
//! use threers::fluid::{FluidConfig3d, FluidSim3d};
//! # fn demo(device: std::sync::Arc<wgpu::Device>, queue: std::sync::Arc<wgpu::Queue>,
//! #         camera: &dyn threers::cameras::Camera, frame: &wgpu::Texture) -> Result<(), String> {
//! let mut fluid = FluidSim3d::new(device, queue, FluidConfig3d::default())?;
//! fluid.splat([0.5, 0.2, 0.5], [0.0, 900.0, 0.0], [0.4, 0.2, 0.9]);
//!
//! // Per frame:
//! fluid.step(1.0 / 60.0);
//! fluid.render(camera, frame);
//! # Ok(()) }
//! ```
//!
//! # What changes between two dimensions and three
//!
//! Less than you would expect. Advection, divergence, the Jacobi sweep and the
//! projection are the same equations with a third term; the stencil grows from
//! five points to seven, so the pressure divisor is six rather than four.
//!
//! Two things genuinely change.
//!
//! **Vorticity is a vector.** In the plane the curl is a scalar — one number
//! saying how fast this point spins — and the confinement force is that
//! number times a quarter turn of the gradient. In space the curl is an axis,
//! and the quarter turn becomes the cross product it was always standing in
//! for: `f = ε (N × ω)`. That is the only formula in the solver whose *shape*
//! differs.
//!
//! **You cannot look at it.** A 2D fluid is already an image. A volume has to
//! be rendered, and since it has no surfaces that means marching rays through
//! it and accumulating emission against absorption — see
//! [`FluidSim3d::render`].
//!
//! # Why the volume lives in a 2D texture
//!
//! A render pass writes one attachment, and a 3D attachment is one z-slice, so
//! a 64³ volume would be sixty-four passes per stage and about eighteen
//! hundred a frame. Instead the slices are tiled across a 2D texture and every
//! stage is a single fullscreen pass, exactly as in two dimensions. It also
//! keeps what the rest of this module is built around: no compute stage, so it
//! runs on WebGL2, which has no 3D storage textures at all. The price is doing
//! trilinear filtering by hand — see `src/fluid/volume_shader.rs`.

use std::sync::Arc;

use crate::cameras::Camera;
use crate::renderer::fullscreen::{self, draw, draw_timed, DoubleFbo, Fbo, Load};
use crate::Vector3;

use super::passes::FluidFormats;
use super::PressureSolver;

/// Knobs for a 3D fluid.
#[derive(Clone, Copy, Debug, PartialEq)]
pub struct FluidConfig3d {
    /// Voxels along each axis of the solver's grid. Cubed, so this is the one
    /// number that decides everything: 64 is 262 144 voxels, 128 is two
    /// million.
    pub resolution: u32,
    /// Voxels along each axis of the dye. As in two dimensions it is usually
    /// finer than the solver's — the eye reads the ink, not the motion — but
    /// in three dimensions the cube law makes that expensive, so the default
    /// is only twice.
    pub dye_resolution: u32,
    pub density_dissipation: f32,
    pub velocity_dissipation: f32,
    /// How much of last frame's pressure to start this frame's solve from.
    pub pressure: f32,
    pub pressure_iterations: u32,
    /// How fast the fluid runs, as a multiple of real time. Applied by
    /// [`FluidSim3d::update`] and not by [`step`](FluidSim3d::step), for the
    /// reason the plane's [`time_scale`](super::FluidConfig::time_scale)
    /// gives.
    pub time_scale: f32,
    /// How the pressure equation is solved. The volume defaults to
    /// [`PressureSolver::Multigrid`] where the plane defaults to Jacobi, and
    /// the reason is the cube law — see [`QUALITY_LEVELS_3D`] and
    /// `the_v_cycle_beats_the_sweeps`.
    pub pressure_solver: PressureSolver,
    /// V-cycles a step, when the solver is the multigrid one.
    pub multigrid_cycles: u32,
    /// Smoothing sweeps either side of the recursion, per level.
    ///
    /// The textbook choice is two, and two is the default. It is here
    /// because it is most of the V-cycle's pass count — a 64³ solve with
    /// four levels spends sixteen of its forty-four passes on these — so it
    /// is the first place to look for work that is not buying anything.
    pub multigrid_smooth: u32,
    /// Stop coarsening at this many voxels a side.
    ///
    /// Each level costs a handful of render passes over a grid small enough
    /// that the arithmetic is free, which looks like pure overhead — and
    /// measured, it is not. Coarsening all the way to 4³ against stopping at
    /// 16³ is 10% of the solve for 23% of the residual
    /// (`the_v_cycle_needs_its_coarse_levels`), so the default goes all the
    /// way down. The knob is here because that balance is a property of the
    /// machine's per-pass cost, not of the fluid.
    pub multigrid_coarsest: u32,
    /// Keep the bind groups and uniform blocks between frames instead of
    /// rebuilding them.
    ///
    /// A step records about fifty passes whose descriptions are identical
    /// frame to frame, and recording them is most of what the CPU does here.
    /// Off is the way to measure that; there is no reason to run that way.
    pub cache_bindings: bool,
    /// Draw the frame rate over the picture.
    ///
    /// Off by default. The number is the volume's own render cadence,
    /// measured between calls to [`render`](FluidSim3d::render) — not the
    /// application's frame rate, which may be doing other things, and not
    /// the GPU's time in the solver, which is
    /// [`stage_ms`](FluidSim3d::stage_ms).
    pub show_fps: bool,
    /// Average the ray march over this many frames.
    ///
    /// One is off. Above that, each frame jitters its samples differently
    /// and blends into what came before with a weight of `1 / n`, so a march
    /// of 64 samples converges towards one of several hundred without ever
    /// costing more than 64 in a frame. That is the trade the big volumes
    /// need: past about 256³ the dye is finer than any affordable number of
    /// samples can resolve, and the picture is noise-limited rather than
    /// resolution-limited.
    ///
    /// It resets whenever the camera moves, so it buys nothing while you are
    /// dragging and everything the moment you stop. The fluid itself still
    /// moves, which it cannot reset on — a plume accumulated over eight
    /// frames is eight frames of motion blurred together, and that is the
    /// cost. Steady smoke takes it well; a splat landing does not.
    pub temporal_samples: u32,
    /// How far the history may stray from this frame, in standard
    /// deviations of its own neighbourhood, before it is rejected.
    ///
    /// Smaller rejects more. Measured, against a long march of the same
    /// moment — the first column is a still camera, where accumulating is
    /// supposed to help, and the second a rising plume, where it is supposed
    /// not to hurt:
    ///
    /// | clip | still | moving |
    /// |---|---|---|
    /// | 0.25 | 0.067 | 0.179 |
    /// | **0.5** | **0.024** | **0.235** |
    /// | 1.0 | 0.019 | 0.408 |
    /// | no clip | 0.019 | 1.233 |
    /// | no accumulation | 0.088 | 0.050 |
    ///
    /// Half a sigma is the knee: nearly all of the convergence, for a little
    /// over half the ghosting of one sigma. Below it the history is pinned
    /// so close to the local mean that accumulating stops buying anything.
    ///
    /// Only read when [`temporal_samples`](Self::temporal_samples) is above
    /// one.
    pub temporal_clip: f32,
    /// Step the ray march over the parts of the volume that hold nothing.
    ///
    /// Four fifths of a plume is empty, so this is most of the march. It
    /// costs two small reduction passes a frame to find the empty parts and
    /// changes the picture by a fraction of a level; set it false to measure
    /// either.
    pub empty_space_skip: bool,
    /// Pack the dye into `rg11b10ufloat` where the device allows it — half
    /// the memory and half the bandwidth of `rgba16float`, which on the
    /// largest field in the volume is the difference between one resolution
    /// level and the next. Set it false to measure what it costs; see
    /// [`FluidFormats::volume_dye`].
    pub compact_dye: bool,
    /// Vorticity confinement. The same knob as in two dimensions and the same
    /// effect: it is what decides whether this looks like smoke.
    pub curl: f32,
    /// Coefficient on the pressure gradient. 1.0 is the 2D module's default
    /// and over-relaxes by two, which compensates for an unconverged solve —
    /// see `FluidConfig::gradient_scale`.
    pub gradient_scale: f32,
    /// Radius of a splat, as a fraction of the volume.
    pub splat_radius: f32,
    /// How hard a drag pushes, for [`impulse_at_screen`](FluidSim3d::impulse_at_screen).
    ///
    /// The drag is already in world units at the depth it acts on, so this
    /// only has to turn "a fraction of the screen" into "voxels a second" —
    /// it is a feel knob, not a physical one.
    pub splat_force: f32,
    /// Buoyancy: how hard dye rises, in voxels per second squared. Smoke that
    /// does not rise reads as dust, and in three dimensions there is a clear
    /// up — which is the one force worth having that a 2D ink demo does not.
    pub buoyancy: f32,
    /// Which way buoyancy pushes. Up by default, because that is what
    /// buoyancy is — but a volume has no privileged axis, and the one case
    /// this exists for is a scene whose "up" is not the solver's: a plume
    /// along a tilted chimney, or a current in any direction at all.
    ///
    /// Normalised before use; a zero vector means no buoyancy.
    pub buoyancy_axis: [f32; 3],
    /// Ray-march steps through the volume. The cost of the picture, more or
    /// less entirely.
    pub march_steps: u32,
    /// How opaque the dye is.
    pub absorption: f32,
    /// How brightly it glows.
    pub emission: f32,
    pub exposure: f32,
    /// What is behind the volume when it is not `transparent`.
    pub background: [f32; 3],
    /// Keep the dye's own alpha, premultiplied, instead of compositing it
    /// onto [`background`](Self::background) — for drawing the volume over a
    /// scene rather than on its own.
    pub transparent: bool,
    /// The box the volume occupies, in world units.
    pub bounds_min: [f32; 3],
    pub bounds_max: [f32; 3],
    /// Seed for splat positions and hues, so a headless render repeats.
    pub seed: u64,
}

/// A named size for a volume: what to allocate and how hard to work it.
///
/// Separate from the 2D [`FluidQuality`](crate::fluid::FluidQuality) because
/// the numbers mean different things. In two dimensions doubling the dye grid
/// costs four times the memory and the simulation stays interactive; here it
/// costs eight, and the march has to take more steps to see the extra detail,
/// so a level that reads as one notch finer is nearer three times the frame.
#[derive(Clone, Copy, Debug)]
pub struct FluidQuality3d {
    pub name: &'static str,
    /// Velocity and pressure grid, in voxels a side.
    pub resolution: u32,
    /// Dye grid, in voxels a side.
    pub dye_resolution: u32,
    pub pressure_iterations: u32,
    /// Samples a ray takes crossing the volume.
    pub march_steps: u32,
}

/// The volume's quality levels, coarsest first.
///
/// The names match the 2D levels so one dropdown can drive both, but the
/// numbers are the volume's own. Measured on an M4 Pro with
/// `examples/fluid3d_bench.rs`, marching into 1280x720, as step + march:
///
/// | level | solver | dye | memory | frame |
/// |---|---|---|---|---|
/// | low | 24³ | 48³ | 2 MB | 2.5 ms |
/// | medium | 40³ | 80³ | 10 MB | 3.0 ms |
/// | high | 64³ | 128³ | 41 MB | 3.2 ms |
/// | ultra | 96³ | 192³ | 139 MB | 3.9 ms |
/// | extreme | 128³ | 256³ | 324 MB | 5.2 ms |
///
/// Memory is the part to look at before `extreme`: three double-buffered
/// vector fields and a scalar pair, all at the cube of the side. The frame
/// grows far more slowly than the memory does, because half of it is the
/// march and the march is set by the pixels on the screen rather than by the
/// voxels behind them.
pub const QUALITY_LEVELS_3D: &[FluidQuality3d] = &[
    FluidQuality3d {
        name: "low",
        resolution: 24,
        dye_resolution: 48,
        pressure_iterations: 16,
        march_steps: 48,
    },
    FluidQuality3d {
        name: "medium",
        resolution: 40,
        dye_resolution: 80,
        pressure_iterations: 18,
        march_steps: 64,
    },
    // What `FluidConfig3d::default` gives.
    FluidQuality3d {
        name: "high",
        resolution: 64,
        dye_resolution: 128,
        pressure_iterations: 20,
        march_steps: 96,
    },
    FluidQuality3d {
        name: "ultra",
        resolution: 96,
        dye_resolution: 192,
        pressure_iterations: 24,
        march_steps: 128,
    },
    FluidQuality3d {
        name: "extreme",
        resolution: 128,
        dye_resolution: 256,
        pressure_iterations: 28,
        march_steps: 192,
    },
];

/// What a volume of this size will allocate, in bytes, before allocating it.
///
/// `formats` must be [resolved](FluidFormats::resolve) against the
/// configuration's `compact_dye` first — what the device *allows* and what
/// the volume will *use* are different, and the difference is a factor of two
/// on the largest field.
///
/// The atlas rounds a cube up to a rectangle, so this is what the textures
/// actually cost rather than `n³ × bytes`: a 100³ volume tiles ten across and
/// ten down and wastes nothing, a 101³ volume tiles eleven across and ten
/// down and wastes a tenth.
///
/// Mirrors [`FluidSim3d::memory_bytes`] — `the_memory_estimate_is_the_memory`
/// holds the two together — so it can be used to choose a resolution on a
/// device where getting it wrong means the app is killed rather than slow.
pub fn memory_for(resolution: u32, dye_resolution: u32, formats: FluidFormats) -> u64 {
    let texels = |grid: u32| {
        let a = Atlas::new(grid);
        let (w, h) = a.size();
        w as u64 * h as u64
    };
    let size = |f: wgpu::TextureFormat| crate::renderer::readback::bytes_per_texel(f) as u64;
    let (vec3, dye_fmt, scalar) = (
        size(formats.dye),
        size(formats.volume_dye),
        size(formats.scalar),
    );
    let sim = texels(resolution.max(2));
    let dye = texels(dye_resolution.max(2));

    // velocity ×2, pressure ×2, divergence, curl, residual.
    let mut total = sim * (2 * vec3 + 2 * scalar + scalar + vec3 + scalar);
    total += dye * 2 * dye_fmt;
    // The coarse chain: each level is an eighth of the one above, four fields
    // apiece.
    let mut n = resolution.max(2);
    let mut levels = 0;
    while n / 2 >= MG_COARSEST_3D && levels < 6 {
        n /= 2;
        levels += 1;
        total += texels(n) * 4 * scalar;
    }
    // The occupancy pyramid: half the dye's side, then a quarter.
    total += (texels(dye_resolution / 2) + texels(dye_resolution / 4)) * scalar;
    total
}

/// The largest volume that fits a memory budget, as `(solver, dye)`.
///
/// `ratio` is how many dye voxels to a solver voxel along each axis — two is
/// the default everywhere else here, because dye is what the eye reads and
/// motion is what the solver has to resolve.
///
/// For a device where the number matters. A desktop can be asked for 272³ and
/// told no; a phone that is asked for 272³ is killed by the watchdog, and the
/// honest move is to ask for what fits.
///
/// ```
/// use threers::fluid::{within_memory, FluidConfig3d, FluidFormats};
/// // A couple of hundred megabytes, which is a reasonable share of what a
/// // phone will let one app have. Resolved against the configuration, or the
/// // dye is predicted at half what it will cost.
/// let config = FluidConfig3d::default();
/// let formats = FluidFormats::default().resolve(config.compact_dye);
/// let (solver, dye) = within_memory(200 << 20, 2, formats);
/// assert!(solver >= 64 && dye == solver * 2);
/// ```
pub fn within_memory(budget: u64, ratio: u32, formats: FluidFormats) -> (u32, u32) {
    let ratio = ratio.max(1);
    // Eight at a time: the atlas pads a cube into a rectangle and the waste
    // is worst at awkward sizes, so walking in steps that tile well keeps the
    // answer from sitting just inside a cliff.
    let mut best = (8u32, 8 * ratio);
    for solver in (8..=512).step_by(8) {
        let dye = solver * ratio;
        if memory_for(solver, dye, formats) > budget {
            break;
        }
        best = (solver, dye);
    }
    best
}

/// Look a volume quality level up by name.
pub fn quality3d(name: &str) -> Option<&'static FluidQuality3d> {
    QUALITY_LEVELS_3D.iter().find(|q| q.name == name)
}

/// Every volume quality level's name, coarsest first.
pub fn quality3d_names() -> impl Iterator<Item = &'static str> {
    QUALITY_LEVELS_3D.iter().map(|q| q.name)
}

impl FluidConfig3d {
    /// Apply a [`FluidQuality3d`], leaving every other setting alone.
    ///
    /// ```no_run
    /// use threers::fluid::{quality3d, FluidConfig3d};
    /// let config = FluidConfig3d {
    ///     curl: 26.0,
    ///     ..FluidConfig3d::default().with_quality(quality3d("ultra").unwrap())
    /// };
    /// ```
    pub fn with_quality(mut self, quality: &FluidQuality3d) -> Self {
        self.resolution = quality.resolution;
        self.dye_resolution = quality.dye_resolution;
        self.pressure_iterations = quality.pressure_iterations;
        self.march_steps = quality.march_steps;
        self
    }
}

impl Default for FluidConfig3d {
    fn default() -> Self {
        Self {
            resolution: 64,
            dye_resolution: 128,
            density_dissipation: 0.4,
            velocity_dissipation: 0.2,
            // All of it, not the 0.8 the plane inherits from the original.
            // More carry-over converges better — 0.269 of a splat's
            // divergence left after eight frames against 0.281 — and the
            // drift the damping is for does not happen here: see
            // `the_pressure_carry_over_is_damped_against_drift`, which runs
            // nine hundred frames with and without a coarse grid to anchor
            // the constant mode and finds both bounded.
            pressure: 1.0,
            pressure_iterations: 20,
            time_scale: 1.0,
            pressure_solver: PressureSolver::Multigrid,
            multigrid_cycles: 1,
            multigrid_smooth: 2,
            multigrid_coarsest: 4,
            cache_bindings: true,
            show_fps: false,
            temporal_samples: 1,
            temporal_clip: 0.5,
            empty_space_skip: true,
            // Off, and measured: it halves the dye's memory, and on an M4 Pro
            // it also doubles the frame — sampling the packed format is the
            // cost, not the bandwidth it saves. See `FluidFormats::volume_dye`.
            compact_dye: false,
            curl: 12.0,
            gradient_scale: 1.0,
            splat_radius: 0.6,
            splat_force: 6.0,
            buoyancy: 0.0,
            buoyancy_axis: [0.0, 1.0, 0.0],
            march_steps: 96,
            absorption: 28.0,
            emission: 1.2,
            exposure: 1.0,
            background: [0.015, 0.015, 0.025],
            transparent: false,
            bounds_min: [-0.5, -0.5, -0.5],
            bounds_max: [0.5, 0.5, 0.5],
            seed: 0x5EED_F101_D1CE,
        }
    }
}

/// How a volume's z-slices are laid out across a 2D texture.
///
/// Square-ish, so the atlas is as close to square as the slice count allows —
/// a long thin texture wastes more of the device's maximum dimension than it
/// needs to and is likelier to hit it.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub struct Atlas {
    /// Voxels along each axis.
    pub grid: u32,
    /// Tiles across and down.
    pub tiles: (u32, u32),
}

impl Atlas {
    pub fn new(grid: u32) -> Self {
        let grid = grid.max(2);
        let across = (grid as f32).sqrt().ceil() as u32;
        let down = grid.div_ceil(across);
        Self {
            grid,
            tiles: (across, down),
        }
    }

    /// Size of the 2D texture holding it.
    pub fn size(&self) -> (u32, u32) {
        (self.grid * self.tiles.0, self.grid * self.tiles.1)
    }

    /// The largest grid whose atlas fits in `max` on both axes.
    pub fn largest_within(mut grid: u32, max: u32) -> Self {
        loop {
            let a = Self::new(grid);
            let (w, h) = a.size();
            if (w <= max && h <= max) || grid <= 8 {
                return a;
            }
            grid /= 2;
        }
    }
}

/// One field's atlas, mirroring `Grid` in the shader.
#[repr(C)]
#[derive(Clone, Copy, Debug, Default, bytemuck::Pod, bytemuck::Zeroable)]
struct GridUniform {
    size: [f32; 4],
    tiles: [f32; 4],
}

impl From<Atlas> for GridUniform {
    fn from(a: Atlas) -> Self {
        let n = a.grid as f32;
        let (tx, ty) = (a.tiles.0 as f32, a.tiles.1 as f32);
        Self {
            size: [n, n, n, 1.0 / n],
            tiles: [tx, ty, 1.0 / tx, 1.0 / ty],
        }
    }
}

/// The uniform block, mirroring `Vol` in `volume_shader.rs`.
#[repr(C)]
#[derive(Clone, Copy, Debug, Default, bytemuck::Pod, bytemuck::Zeroable)]
struct Vol {
    out_grid: GridUniform,
    a_grid: GridUniform,
    b_grid: GridUniform,
    texel: [f32; 4],
    point: [f32; 4],
    color: [f32; 4],
    params: [f32; 4],
    bounds_min: [f32; 4],
    bounds_max: [f32; 4],
    march: [f32; 4],
    background: [f32; 4],
    cam_pos: [f32; 4],
    inv_view_proj: [f32; 16],
    /// Last frame's view-projection, for the resolve to reproject through.
    prev_view_proj: [f32; 16],
}

const VOL_SIZE: u64 = std::mem::size_of::<Vol>() as u64;

/// A three-dimensional fluid and the textures it lives in.
pub struct FluidSim3d {
    device: Arc<wgpu::Device>,
    queue: Arc<wgpu::Queue>,
    config: FluidConfig3d,
    formats: FluidFormats,

    sim: Atlas,
    dye_atlas: Atlas,

    velocity: DoubleFbo,
    dye: DoubleFbo,
    pressure: DoubleFbo,
    divergence: Fbo,
    curl: Fbo,
    /// The top level's residual, and the coarse chain under it.
    residual: Fbo,
    mg: Vec<MgLevel3>,
    /// The dye's largest value per brick, at half the dye's side and then a
    /// quarter, for the march to skip by. See `fs_maxdown3`.
    occupancy: [(Atlas, Fbo); 2],
    /// Whether the dye has changed since the occupancy was last built. The
    /// reduction costs the same whatever the render is: rebuilding it per
    /// render would be most of the cost of a small one.
    occupancy_stale: std::cell::Cell<bool>,

    bgl: wgpu::BindGroupLayout,
    sampler: wgpu::Sampler,
    pipes: Pipes,
    out: fullscreen::FormatPipelines<wgpu::RenderPipeline>,
    /// The march into `current`, which is the ordinary march at the
    /// accumulation's format.
    accum_march: Option<wgpu::RenderPipeline>,
    /// The resolve, which writes the history and the caller's target at
    /// once — so it is keyed on the second of those.
    resolve: fullscreen::FormatPipelines<wgpu::RenderPipeline>,
    /// The frame-rate read-out, composited over the target.
    overlay: fullscreen::FormatPipelines<wgpu::RenderPipeline>,
    layout: wgpu::PipelineLayout,
    module: wgpu::ShaderModule,
    ubos: std::cell::RefCell<Vec<wgpu::Buffer>>,
    /// Next free uniform block this frame; reset per submit.
    next_ubo: std::cell::Cell<usize>,
    /// What was last written to each uniform block, so a block whose contents
    /// have not changed since the last frame is not written again.
    ubo_contents: std::cell::RefCell<Vec<Vol>>,
    /// GPU timestamps, when something asked for them. `None` until
    /// [`profile`](Self::profile) is called, and on an adapter with no
    /// `TIMESTAMP_QUERY`.
    profiler: Option<crate::renderer::gpu_timer::GpuTimer>,
    /// When the last render was, and the smoothed rate since — for the
    /// read-out, and for anyone who wants the number without drawing it.
    rate: super::read_out::Rate,
    /// The last sixty-four frame times, for the chart under the number.
    history: super::read_out::History,
    /// This frame's march, before it is mixed with the history: half float,
    /// at the size of the target it was last asked for.
    current: Option<Fbo>,
    /// How far along each ray this frame's march saw what it saw, for the
    /// resolve to reproject by.
    depth: Option<Fbo>,
    /// This frame's camera, and its inverse, kept so the resolve can use
    /// them without being handed the camera again.
    view_proj: [f32; 16],
    view_proj_inverse: [f32; 16],
    /// The history, ping-ponged — the resolve reads one and writes the other.
    accum: Option<DoubleFbo>,
    /// The view-projection the accumulation was started from. Any change
    /// means the pixels no longer line up and the history is worthless.
    accum_from: Option<[f32; 16]>,
    /// Frames since the accumulation was reset, which moves the jitter.
    accum_frame: u32,
    /// Render passes the last `step` recorded.
    ///
    /// A deterministic companion to [`stage_ms`](Self::stage_ms): at small
    /// grids the solve is bound by how many passes it is, not by what they
    /// compute, and a pass count does not need a quiet machine to measure.
    passes: std::cell::Cell<u32>,
    /// Which mark the next stage writes. Reset at the top of a frame.
    mark: std::cell::Cell<u32>,
    /// Bind groups, kept between frames.
    ///
    /// A step records about fifty passes and every one of them wants a bind
    /// group over (uniform block, two textures). Those triples repeat exactly
    /// frame after frame — the pass order is fixed, the uniform blocks are
    /// handed out in the same order, and the textures only change when the
    /// grids are rebuilt — so building them again each frame is work whose
    /// answer is already known. Keyed on the slot and on *which* textures,
    /// because the ping-pong means the same slot alternates between two.
    binds: std::cell::RefCell<std::collections::HashMap<(usize, u64, u64), wgpu::BindGroup>>,

    rng: super::Rng,
    paused: bool,
}

struct Pipes {
    fill: wgpu::RenderPipeline,
    buoyancy: wgpu::RenderPipeline,
    splat_velocity: wgpu::RenderPipeline,
    splat_dye: wgpu::RenderPipeline,
    advect_velocity: wgpu::RenderPipeline,
    advect_dye: wgpu::RenderPipeline,
    divergence: wgpu::RenderPipeline,
    curl: wgpu::RenderPipeline,
    vorticity: wgpu::RenderPipeline,
    pressure: wgpu::RenderPipeline,
    gradient: wgpu::RenderPipeline,
    maxdown: wgpu::RenderPipeline,
    residual: wgpu::RenderPipeline,
    restrict: wgpu::RenderPipeline,
    prolong: wgpu::RenderPipeline,
}

/// The stages a frame is reported in, in the order they are recorded.
///
/// A mark is written at the start of each one's first pass and a last mark at
/// the end of the frame, so consecutive differences are the stages — see
/// [`GpuTimer::with_marks`](crate::renderer::gpu_timer::GpuTimer::with_marks).
pub const STAGE_NAMES: [&str; 5] = ["curl+vort", "divergence", "solve", "project", "advect"];

/// Marks: one opening each stage, plus one that closes the last.
///
/// Every mark is written at the *beginning* of a pass. Writing one at each
/// end of a single pass is the obvious way to time it and it does not work
/// here — asked for both on the ray march, this GPU returned a pair in the
/// wrong order on most frames, and a plausible number on some. Marks across
/// passes are reliable. So the closing mark goes on whatever pass comes next
/// — the occupancy reduction, or the march if skipping is off — and the
/// stage it opens is not one of these. The march is timed by rendering in a
/// loop and watching the clock, which works because it is one pass of real
/// work rather than tens of microseconds.
const MARKS: u32 = STAGE_NAMES.len() as u32 + 1;

/// Half and quarter of the dye's side, holding its largest value per brick.
///
/// Two levels because the march wants a brick big enough to be worth a test
/// and small enough to hug the plume: a quarter of the side is four dye
/// voxels, which measured out as 85% empty on a settled plume. Half a
/// megabyte at the default resolution, against the dye's thirty-three.
fn occupancy_levels(
    device: &wgpu::Device,
    dye_grid: u32,
    format: wgpu::TextureFormat,
) -> [(Atlas, Fbo); 2] {
    let level = |n: u32| {
        let atlas = Atlas::new(n.max(2));
        let (w, h) = atlas.size();
        (atlas, Fbo::new(device, "fluid3d occupancy", w, h, format))
    };
    [level(dye_grid / 2), level(dye_grid / 4)]
}

/// Which format the dye volume lives in.
///
/// `rg11b10ufloat` is filterable everywhere but a render target only where
/// the device enabled `RG11B10UFLOAT_RENDERABLE`, so this asks the device
/// rather than assuming. Falling back costs memory, not correctness.
fn packed_dye(
    device: &wgpu::Device,
    config: &FluidConfig3d,
    formats: FluidFormats,
) -> wgpu::TextureFormat {
    if !config.compact_dye {
        return formats.resolve(false).volume_dye;
    }
    if device
        .features()
        .contains(wgpu::Features::RG11B10UFLOAT_RENDERABLE)
    {
        wgpu::TextureFormat::Rg11b10Ufloat
    } else {
        formats.dye
    }
}

/// A filled uniform block, and which slot of the pool it came from.
///
/// The slot travels with the buffer because the bind-group cache is keyed on
/// it, and deriving it from "the last one handed out" is only correct while
/// every block is bound before the next is filled.
struct Ubo {
    buffer: wgpu::Buffer,
    slot: usize,
}

/// Below this many voxels a side there is nothing left for a seven-point
/// stencil to average that clamp-to-edge is not already giving it.
const MG_COARSEST_3D: u32 = 4;

/// One coarse level of the volume's multigrid solver.
struct MgLevel3 {
    atlas: Atlas,
    /// The correction being solved for here, ping-ponged by the smoother.
    pressure: DoubleFbo,
    /// The restricted residual from the level above.
    rhs: Fbo,
    /// This level's own residual, on the way down.
    residual: Fbo,
}

/// Coarse levels for a grid of `grid` voxels a side, halving as far as it
/// goes.
///
/// They are cheap: a level is an eighth of the one above it, so the whole
/// chain below a 64³ grid costs about a seventh of one more 64³ field.
fn mg_levels3(
    device: &wgpu::Device,
    grid: u32,
    coarsest: u32,
    format: wgpu::TextureFormat,
) -> Vec<MgLevel3> {
    let mut levels = Vec::new();
    let mut n = grid;
    let coarsest = coarsest.max(MG_COARSEST_3D);
    while n / 2 >= coarsest && levels.len() < 6 {
        n /= 2;
        let atlas = Atlas::new(n);
        let (w, h) = atlas.size();
        levels.push(MgLevel3 {
            atlas,
            pressure: DoubleFbo::new(device, "fluid3d mg pressure", w, h, format),
            rhs: Fbo::new(device, "fluid3d mg rhs", w, h, format),
            residual: Fbo::new(device, "fluid3d mg residual", w, h, format),
        });
    }
    levels
}

impl FluidSim3d {
    pub fn new(
        device: Arc<wgpu::Device>,
        queue: Arc<wgpu::Queue>,
        config: FluidConfig3d,
    ) -> Result<Self, String> {
        Self::with_formats(device, queue, config, FluidFormats::default())
    }

    pub fn with_formats(
        device: Arc<wgpu::Device>,
        queue: Arc<wgpu::Queue>,
        config: FluidConfig3d,
        formats: FluidFormats,
    ) -> Result<Self, String> {
        if config.pressure_iterations == 0 {
            return Err("fluid3d: pressure_iterations must be at least 1".into());
        }
        let max = device.limits().max_texture_dimension_2d;
        let sim = Atlas::largest_within(config.resolution, max);
        let dye_atlas = Atlas::largest_within(config.dye_resolution, max);
        // The caller may have handed over the baseline formats, which cannot
        // know what this device will render to. Upgrade the dye here rather
        // than make every caller thread an adapter through.
        let formats = FluidFormats {
            volume_dye: packed_dye(&device, &config, formats),
            ..formats
        };

        let bgl = device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
            label: Some("threers fluid3d bgl"),
            entries: &[
                wgpu::BindGroupLayoutEntry {
                    binding: 0,
                    visibility: wgpu::ShaderStages::FRAGMENT,
                    ty: wgpu::BindingType::Buffer {
                        ty: wgpu::BufferBindingType::Uniform,
                        has_dynamic_offset: false,
                        min_binding_size: wgpu::BufferSize::new(VOL_SIZE),
                    },
                    count: None,
                },
                tex(1),
                tex(2),
                // A third, for the one pass that needs three: the resolve
                // reads this frame's march, the history, and the depth the
                // march saw. Every other pass binds the dye here and
                // ignores it — an entry the shader does not declare is
                // allowed, and one layout for everything is worth more than
                // a binding nobody reads.
                tex(4),
                wgpu::BindGroupLayoutEntry {
                    binding: 3,
                    visibility: wgpu::ShaderStages::FRAGMENT,
                    ty: wgpu::BindingType::Sampler(wgpu::SamplerBindingType::Filtering),
                    count: None,
                },
            ],
        });
        let layout = device.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
            label: Some("threers fluid3d layout"),
            bind_group_layouts: &[Some(&bgl)],
            immediate_size: 0,
        });
        let module = device.create_shader_module(wgpu::ShaderModuleDescriptor {
            label: Some("threers fluid3d"),
            source: wgpu::ShaderSource::Wgsl(
                format!(
                    "{}\n{}\n{}",
                    fullscreen::VS,
                    super::read_out::glyphs(),
                    super::volume_shader::SHADER
                )
                .into(),
            ),
        });

        // Three channels of velocity and of curl, so four: there is no
        // *signed* three-channel float format. The dye is unsigned and gets
        // the packed one.
        let vec_fmt = formats.dye;
        let dye_fmt = formats.volume_dye;
        let scalar = formats.scalar;
        let p = |fs: &str, format| build(&device, &layout, &module, fs, format);
        let pipes = Pipes {
            fill: p("fs_fill3", vec_fmt),
            buoyancy: p("fs_buoyancy3", vec_fmt),
            splat_velocity: p("fs_splat3", vec_fmt),
            splat_dye: p("fs_splat3", dye_fmt),
            advect_velocity: p("fs_advect3_self", vec_fmt),
            advect_dye: p("fs_advect3", dye_fmt),
            divergence: p("fs_divergence3", scalar),
            curl: p("fs_curl3", vec_fmt),
            vorticity: p("fs_vorticity3", vec_fmt),
            pressure: p("fs_pressure3", scalar),
            gradient: p("fs_gradient3", vec_fmt),
            maxdown: p("fs_maxdown3", scalar),
            residual: p("fs_residual3", scalar),
            restrict: p("fs_restrict3", scalar),
            prolong: p("fs_prolong3", scalar),
        };

        let (sw, sh) = sim.size();
        let (dw, dh) = dye_atlas.size();
        Ok(Self {
            velocity: DoubleFbo::new(&device, "fluid3d velocity", sw, sh, vec_fmt),
            dye: DoubleFbo::new(&device, "fluid3d dye", dw, dh, dye_fmt),
            pressure: DoubleFbo::new(&device, "fluid3d pressure", sw, sh, scalar),
            divergence: Fbo::new(&device, "fluid3d divergence", sw, sh, scalar),
            curl: Fbo::new(&device, "fluid3d curl", sw, sh, vec_fmt),
            residual: Fbo::new(&device, "fluid3d residual", sw, sh, scalar),
            mg: mg_levels3(&device, sim.grid, config.multigrid_coarsest, scalar),
            occupancy: occupancy_levels(&device, dye_atlas.grid, scalar),
            occupancy_stale: std::cell::Cell::new(true),
            sampler: fullscreen::linear_sampler(&device),
            ubos: std::cell::RefCell::new((0..48).map(|_| ubo(&device)).collect()),
            next_ubo: std::cell::Cell::new(0),
            profiler: None,
            rate: Default::default(),
            history: Default::default(),
            current: None,
            depth: None,
            view_proj: [0.0; 16],
            view_proj_inverse: [0.0; 16],
            accum: None,
            accum_from: None,
            accum_frame: 0,
            passes: std::cell::Cell::new(0),
            mark: std::cell::Cell::new(0),
            ubo_contents: std::cell::RefCell::new(Vec::new()),
            binds: std::cell::RefCell::new(std::collections::HashMap::new()),
            rng: super::Rng::new(config.seed),
            device,
            queue,
            config,
            formats,
            sim,
            dye_atlas,
            bgl,
            pipes,
            out: fullscreen::FormatPipelines::new(),
            accum_march: None,
            resolve: fullscreen::FormatPipelines::new(),
            overlay: fullscreen::FormatPipelines::new(),
            layout,
            module,
            paused: false,
        })
    }

    pub fn config(&self) -> &FluidConfig3d {
        &self.config
    }

    /// Replace the configuration, rebuilding the grids if the resolution
    /// changed.
    ///
    /// A resolution change throws the fluid away — see
    /// [`set_resolution`](Self::set_resolution) for what that costs and why
    /// there is nothing to carry over.
    pub fn set_config(&mut self, config: FluidConfig3d) {
        let (solver, dye) = (config.resolution, config.dye_resolution);
        // `compact_dye` is a texture format, so changing it means new
        // textures — which `set_resolution` only does when a size changed.
        let refmt = packed_dye(&self.device, &config, self.formats) != self.formats.volume_dye;
        let relevel = config.multigrid_coarsest != self.config.multigrid_coarsest;
        self.config = config;
        if refmt || relevel {
            self.formats.volume_dye = packed_dye(&self.device, &self.config, self.formats);
            self.rebuild_fields();
        }
        self.set_resolution(solver, dye);
        // Anything cached names a uniform block or a texture whose contents
        // the new configuration may have changed.
        self.ubo_contents.borrow_mut().clear();
    }

    /// Rebuild the volume at a new size, in voxels a side.
    ///
    /// The fields are dropped and reallocated empty: an interpolated volume
    /// would be eight textures of resampling, and the thing it would preserve
    /// — a plume mid-flight — is a second of simulation. The splat colours
    /// and the camera survive because they are not in here.
    ///
    /// Both are requests. A volume costs the cube of its side, and the atlas
    /// that holds it is a 2D texture with a device limit on it, so the
    /// numbers that come back from [`grids`](Self::grids) are halved until
    /// they fit. Nothing is allocated if neither changed, so this is cheap to
    /// call from a slider's event handler.
    pub fn set_resolution(&mut self, solver: u32, dye: u32) {
        let max = self.device.limits().max_texture_dimension_2d;
        let sim = Atlas::largest_within(solver.max(2), max);
        let dye_atlas = Atlas::largest_within(dye.max(2), max);
        self.config.resolution = solver;
        self.config.dye_resolution = dye;
        if sim.grid == self.sim.grid && dye_atlas.grid == self.dye_atlas.grid {
            return;
        }
        self.sim = sim;
        self.dye_atlas = dye_atlas;
        self.rebuild_fields();
    }

    /// Allocate every field at the current grids and formats, and empty it.
    fn rebuild_fields(&mut self) {
        let (sw, sh) = self.sim.size();
        let (dw, dh) = self.dye_atlas.size();
        let vec_fmt = self.formats.dye;
        let scalar = self.formats.scalar;
        self.velocity = DoubleFbo::new(&self.device, "fluid3d velocity", sw, sh, vec_fmt);
        self.dye = DoubleFbo::new(&self.device, "fluid3d dye", dw, dh, self.formats.volume_dye);
        self.pressure = DoubleFbo::new(&self.device, "fluid3d pressure", sw, sh, scalar);
        self.divergence = Fbo::new(&self.device, "fluid3d divergence", sw, sh, scalar);
        self.curl = Fbo::new(&self.device, "fluid3d curl", sw, sh, vec_fmt);
        self.residual = Fbo::new(&self.device, "fluid3d residual", sw, sh, scalar);
        self.mg = mg_levels3(
            &self.device,
            self.sim.grid,
            self.config.multigrid_coarsest,
            scalar,
        );
        self.occupancy = occupancy_levels(&self.device, self.dye_atlas.grid, scalar);
        self.occupancy_stale.set(true);
        // Every cached bind group names a texture that no longer exists.
        self.binds.borrow_mut().clear();
        // A fresh texture is undefined until something writes it, and every
        // stage here reads before it writes.
        self.clear_fields();
    }

    /// Zero every field, so a rebuilt volume starts from nothing rather than
    /// from whatever the allocator handed over.
    fn clear_fields(&mut self) {
        let mut enc = self.encoder("fluid3d clear");
        for target in [
            &self.velocity.read.view,
            &self.velocity.write.view,
            &self.pressure.read.view,
            &self.pressure.write.view,
            &self.divergence.view,
            &self.curl.view,
            &self.dye.read.view,
            &self.dye.write.view,
        ] {
            // A render pass that loads nothing and draws nothing: the clear
            // is the whole operation, and it costs one tile flush rather than
            // a fullscreen triangle.
            fullscreen::begin(&mut enc, target, Load::Clear);
        }
        self.queue.submit([enc.finish()]);
    }

    /// How many coarse levels the multigrid solver is using, and how many
    /// render passes a step records. The second is what the GPU time in the
    /// solve is mostly made of — see `fluid3d_bench`.
    pub fn solver_levels(&self) -> usize {
        self.mg.len()
    }

    /// The solver and dye grids, in voxels a side, after the device's texture
    /// limit has had its say.
    pub fn grids(&self) -> (u32, u32) {
        (self.sim.grid, self.dye_atlas.grid)
    }

    /// Bytes of GPU memory the fields occupy.
    pub fn memory_bytes(&self) -> u64 {
        let bytes = |f: &Fbo| {
            crate::renderer::readback::bytes_per_texel(f.texture.format()) as u64
                * f.width as u64
                * f.height as u64
        };
        let double = |f: &DoubleFbo| bytes(&f.read) + bytes(&f.write);
        double(&self.velocity)
            + double(&self.dye)
            + double(&self.pressure)
            + bytes(&self.divergence)
            + bytes(&self.curl)
            + bytes(&self.residual)
            + self
                .mg
                .iter()
                .map(|l| double(&l.pressure) + bytes(&l.rhs) + bytes(&l.residual))
                .sum::<u64>()
            + self.occupancy.iter().map(|(_, f)| bytes(f)).sum::<u64>()
    }

    pub fn formats(&self) -> FluidFormats {
        self.formats
    }

    /// Renders a second, smoothed, measured between calls to
    /// [`render`](Self::render). Zero until the second one.
    pub fn fps(&self) -> f32 {
        self.rate.fps()
    }

    /// Ask the GPU how long each stage of a frame takes.
    ///
    /// Wall-clock timing round `step` measures the CPU recording the work and
    /// almost none of the work; on a machine with anything else running it
    /// stops meaning anything at all. These come off the GPU's own clock.
    ///
    /// Returns false where the adapter has no `TIMESTAMP_QUERY`. The figures
    /// are a frame or two behind — reading them back without stalling the
    /// pipeline is the whole trick — so they are a statistic, not something
    /// to make a per-frame decision on.
    pub fn profile(&mut self) -> bool {
        if self.profiler.is_none() {
            self.profiler =
                crate::renderer::gpu_timer::GpuTimer::with_marks(&self.device, &self.queue, MARKS);
        }
        self.profiler.is_some()
    }

    /// Milliseconds the GPU spent in each of [`STAGE_NAMES`], from the most
    /// recent completed readback. All zero until one lands.
    ///
    /// The ray march is not among them: every mark is written at the start
    /// of a pass, so the last pass of a frame has nothing after it to close
    /// its stage. Time the march by rendering in a loop and watching the
    /// clock — it is one pass of real work, so the wall is good enough.
    pub fn stage_ms(&self) -> [f32; STAGE_NAMES.len()] {
        let mut out = [0.0; STAGE_NAMES.len()];
        if let Some(p) = &self.profiler {
            for (i, slot) in out.iter_mut().enumerate() {
                *slot = p.span_ms(i);
            }
        }
        out
    }

    /// Render passes the last [`step`](Self::step) recorded.
    ///
    /// Zero before the first one. The splats and the render are not counted:
    /// this is the solver's own cost, which is what the number is for.
    pub fn passes_per_step(&self) -> u32 {
        self.passes.get()
    }

    /// Record one pass, counting it.
    ///
    /// Every draw in a step goes through here so that the count cannot drift
    /// from the thing it counts.
    fn pass(
        &self,
        enc: &mut wgpu::CommandEncoder,
        target: &wgpu::TextureView,
        load: Load,
        pipeline: &wgpu::RenderPipeline,
        bind: &wgpu::BindGroup,
        marks: Option<wgpu::RenderPassTimestampWrites<'_>>,
    ) {
        self.passes.set(self.passes.get() + 1);
        draw_timed(enc, target, load, pipeline, bind, marks);
    }

    /// Take the timestamp writes that open stage `i`, once per frame.
    ///
    /// The first pass of a stage asks for this and the rest do not, so a
    /// stage whose first pass is conditional — buoyancy, confinement — still
    /// opens on whatever pass actually runs first.
    fn take_mark(&self, stage: u32) -> Option<u32> {
        // Strictly in order. A configuration that skips a stage — no
        // confinement, so no curl — would otherwise shift every later mark
        // into the wrong slot and report five plausible wrong numbers; this
        // way the frame is not reported at all and the last good one stands.
        if self.profiler.is_none() || self.mark.get() != stage {
            return None;
        }
        self.mark.set(stage + 1);
        Some(stage)
    }

    /// The timestamp writes for a mark index, if there is a profiler.
    ///
    /// Separate from [`take_mark`](Self::take_mark) because the result
    /// borrows the query set, and the ray march has to claim its mark before
    /// taking a mutable borrow to compile its pipeline.
    fn writes(&self, mark: Option<u32>) -> Option<wgpu::RenderPassTimestampWrites<'_>> {
        match (mark, &self.profiler) {
            (Some(i), Some(p)) => Some(p.mark(i)),
            _ => None,
        }
    }

    fn open(&self, stage: u32) -> Option<wgpu::RenderPassTimestampWrites<'_>> {
        let mark = self.take_mark(stage);
        self.writes(mark)
    }

    pub fn paused(&self) -> bool {
        self.paused
    }

    pub fn set_paused(&mut self, paused: bool) {
        self.paused = paused;
    }

    /// The dye volume, as the atlas it is stored in.
    pub fn dye_texture(&self) -> &wgpu::Texture {
        &self.dye.read.texture
    }

    /// How the dye volume is packed into that texture.
    pub fn dye_atlas(&self) -> Atlas {
        self.dye_atlas
    }

    // -- input ------------------------------------------------------------

    /// A ball of dye and a push, at `at` in the unit cube.
    pub fn splat(&mut self, at: [f32; 3], velocity: [f32; 3], color: [f32; 3]) {
        self.splat_batch(&[(at, velocity, color)]);
    }

    /// `count` splats at random places, as the 2D module's burst.
    /// A push into the volume, where a point on the screen looks into it.
    ///
    /// `uv` is in 0..1 with the origin at the top left — a pointer position
    /// over the canvas — and `drag` is how far it moved in the same units,
    /// zero for a click that did not move.
    ///
    /// # Where the impulse lands
    ///
    /// A screen point is a ray, not a place: everything along it is under
    /// the pointer. Something has to pick the depth, and there is no right
    /// answer without reading the volume back, which would mean a stall.
    ///
    /// So it lands on the plane through the volume's centre, facing the
    /// camera. Poking the middle of the box is what a pointer over the
    /// middle of the box looks like it should do, it is stable as the camera
    /// orbits — the plane turns with it — and it cannot land behind the
    /// viewer or outside the box's depth range. A ray that misses the box
    /// entirely does nothing.
    ///
    /// The push is the drag, in world units at that same depth, so dragging
    /// across the screen pushes across the screen however the camera is
    /// turned. `strength` scales it into voxels a second.
    pub fn impulse_at_screen(
        &mut self,
        camera: &dyn Camera,
        uv: [f32; 2],
        drag: [f32; 2],
        strength: f32,
        colour: [f32; 3],
    ) {
        let inv = camera
            .projection_matrix()
            .multiply(&camera.view_matrix())
            .invert();
        // uv, top-left origin, to a world ray.
        let ray = |uv: [f32; 2]| -> (Vector3, Vector3) {
            let ndc =
                |z: f32| Vector3::new(uv[0] * 2.0 - 1.0, 1.0 - uv[1] * 2.0, z).apply_matrix4(&inv);
            let near = ndc(0.0);
            let far = ndc(1.0);
            (near, (far - near).normalize())
        };
        let c = self.config;
        let centre = Vector3::new(
            0.5 * (c.bounds_min[0] + c.bounds_max[0]),
            0.5 * (c.bounds_min[1] + c.bounds_max[1]),
            0.5 * (c.bounds_min[2] + c.bounds_max[2]),
        );
        let eye = camera.position();
        let normal = (centre - eye).normalize();
        // Where a ray meets the plane through the centre facing the camera.
        let on_plane = |uv: [f32; 2]| -> Option<Vector3> {
            let (origin, dir) = ray(uv);
            let denom = dir.dot(normal);
            if denom.abs() < 1e-6 {
                return None;
            }
            let t = (centre - origin).dot(normal) / denom;
            if t <= 0.0 {
                return None;
            }
            Some(origin + dir * t)
        };
        let Some(hit) = on_plane(uv) else { return };
        let size = Vector3::new(
            c.bounds_max[0] - c.bounds_min[0],
            c.bounds_max[1] - c.bounds_min[1],
            c.bounds_max[2] - c.bounds_min[2],
        );
        let at = [
            (hit.x - c.bounds_min[0]) / size.x,
            (hit.y - c.bounds_min[1]) / size.y,
            (hit.z - c.bounds_min[2]) / size.z,
        ];
        if at.iter().any(|v| !(0.0..=1.0).contains(v)) {
            return;
        }
        // The drag, as a world direction at the same depth, in the solver's
        // voxels a second.
        let moved = on_plane([uv[0] + drag[0], uv[1] + drag[1]])
            .map(|to| to - hit)
            .unwrap_or(Vector3::ZERO);
        let scale = strength * self.sim.grid as f32;
        self.splat(
            at,
            [
                moved.x / size.x * scale,
                moved.y / size.y * scale,
                moved.z / size.z * scale,
            ],
            colour,
        );
    }

    pub fn multiple_splats(&mut self, count: u32) {
        let mut batch = Vec::with_capacity(count as usize);
        for _ in 0..count {
            let hue = self.rng.next_f32();
            let c = super::hsv_to_rgb(hue, 1.0, 1.0);
            let at = [
                0.2 + 0.6 * self.rng.next_f32(),
                0.2 + 0.6 * self.rng.next_f32(),
                0.2 + 0.6 * self.rng.next_f32(),
            ];
            // A direction, uniformly over the sphere, times a speed.
            //
            // Not three independent components: that is uniform over a
            // *cube*, which is denser towards its corners, so the pushes
            // would favour the diagonals. And the y term used to be
            // `260 * rand` rather than `rand - 0.5`, which is to say it was
            // never negative — every random splat went up, which is what
            // this looked like from outside.
            let z = 2.0 * self.rng.next_f32() - 1.0;
            let theta = std::f32::consts::TAU * self.rng.next_f32();
            let r = (1.0 - z * z).max(0.0).sqrt();
            let speed = 120.0 + 180.0 * self.rng.next_f32();
            let v = [r * theta.cos() * speed, z * speed, r * theta.sin() * speed];
            batch.push((at, v, [c[0] * 1.5, c[1] * 1.5, c[2] * 1.5]));
        }
        self.splat_batch(&batch);
    }

    fn splat_batch(&mut self, splats: &[([f32; 3], [f32; 3], [f32; 3])]) {
        // A pass each, unlike the 2D module's batch: in three dimensions the
        // pass is over an atlas rather than an image, and the arithmetic per
        // texel is a single exponential either way, so batching would save
        // only the submits — and a 3D burst is a handful of blobs, not
        // twenty-five.
        let radius = self.config.splat_radius / 100.0;
        for (at, vel, color) in splats {
            let mut enc = self.encoder("fluid3d splat");
            self.next_ubo.set(0);
            for dye in [false, true] {
                let field = if dye { &self.dye } else { &self.velocity };
                let atlas = if dye { self.dye_atlas } else { self.sim };
                let value = if dye { *color } else { *vel };
                let u = self.params(
                    atlas,
                    Vol {
                        point: [at[0], at[1], at[2], radius],
                        color: [value[0], value[1], value[2], 1.0],
                        ..Default::default()
                    },
                );
                let bg = self.bind(&u, &field.read, None);
                let pipe = if dye {
                    &self.pipes.splat_dye
                } else {
                    &self.pipes.splat_velocity
                };
                draw(&mut enc, &field.write.view, Load::Clear, pipe, &bg);
            }
            self.queue.submit([enc.finish()]);
            self.velocity.swap();
            self.dye.swap();
            self.occupancy_stale.set(true);
        }
    }

    /// Set the whole velocity field to one vector, in voxels per second.
    pub fn set_uniform_velocity(&mut self, v: [f32; 3]) {
        let mut enc = self.encoder("fluid3d fill");
        self.next_ubo.set(0);
        let u = self.params(
            self.sim,
            Vol {
                color: [v[0], v[1], v[2], 0.0],
                ..Default::default()
            },
        );
        let bg = self.bind(&u, &self.velocity.read, None);
        draw(
            &mut enc,
            &self.velocity.write.view,
            Load::Clear,
            &self.pipes.fill,
            &bg,
        );
        self.queue.submit([enc.finish()]);
        self.velocity.swap();
    }

    // -- the frame --------------------------------------------------------

    pub fn update(&mut self, dt: f32) {
        let dt = dt * self.config.time_scale.max(0.0);
        if !self.paused && dt > 0.0 {
            self.step(dt);
        }
    }

    /// One step of the solver.
    pub fn step(&mut self, dt: f32) {
        let c = self.config;
        let mut enc = self.encoder("fluid3d step");
        self.next_ubo.set(0);
        self.mark.set(0);
        self.passes.set(0);

        // Buoyancy, before anything else: dye is hot, hot fluid rises. In two
        // dimensions this is a stylistic choice; in three there is an up.
        let axis = {
            let a = c.buoyancy_axis;
            let len = (a[0] * a[0] + a[1] * a[1] + a[2] * a[2]).sqrt();
            if len > 1e-6 {
                [a[0] / len, a[1] / len, a[2] / len]
            } else {
                [0.0; 3]
            }
        };
        if c.buoyancy != 0.0 && axis != [0.0; 3] {
            let u = self.params_mixed(
                self.sim,
                self.sim,
                self.dye_atlas,
                Vol {
                    texel: [0.0, 0.0, 0.0, dt],
                    params: [0.0, c.buoyancy, 0.0, 0.0],
                    // `color` is unused by this pass, so the axis rides in
                    // it rather than widening the block for three floats.
                    color: [axis[0], axis[1], axis[2], 0.0],
                    ..Default::default()
                },
            );
            let bg = self.bind(&u, &self.velocity.read, Some(&self.dye.read));
            self.pass(
                &mut enc,
                &self.velocity.write.view,
                Load::Clear,
                &self.pipes.buoyancy,
                &bg,
                self.open(0),
            );
            self.velocity.swap();
        }

        // 1. Curl — a vector field in three dimensions. Two passes with a
        // seven-point stencil each, and nothing downstream reads the curl, so
        // with confinement off they are skipped rather than multiplied by
        // zero.
        if c.curl != 0.0 {
            let u = self.params(self.sim, Vol::default());
            let bg = self.bind(&u, &self.velocity.read, None);
            self.pass(
                &mut enc,
                &self.curl.view,
                Load::Clear,
                &self.pipes.curl,
                &bg,
                self.open(0),
            );

            // 2. Vorticity confinement, from it.
            let u = self.params(
                self.sim,
                Vol {
                    texel: [0.0, 0.0, 0.0, dt],
                    params: [0.0, c.curl, 0.0, 0.0],
                    ..Default::default()
                },
            );
            let bg = self.bind(&u, &self.velocity.read, Some(&self.curl));
            self.pass(
                &mut enc,
                &self.velocity.write.view,
                Load::Clear,
                &self.pipes.vorticity,
                &bg,
                None,
            );
            self.velocity.swap();
        }

        // 3. Divergence.
        let u = self.params(self.sim, Vol::default());
        let bg = self.bind(&u, &self.velocity.read, None);
        self.pass(
            &mut enc,
            &self.divergence.view,
            Load::Clear,
            &self.pipes.divergence,
            &bg,
            self.open(1),
        );

        // 4. The solve, starting from a fraction of last frame's answer —
        // which the first sweep applies to what it reads rather than a pass
        // of its own multiplying a whole grid by a constant.
        let open = self.take_mark(2);
        match c.pressure_solver {
            PressureSolver::Jacobi => {
                self.record_jacobi(&mut enc, c.pressure_iterations, c.pressure, open)
            }
            PressureSolver::Multigrid => self.record_multigrid(&mut enc, open),
        }

        // 6. Subtract the gradient.
        let u = self.params(
            self.sim,
            Vol {
                params: [0.0, 0.0, c.gradient_scale, 0.0],
                ..Default::default()
            },
        );
        let bg = self.bind(&u, &self.pressure.read, Some(&self.velocity.read));
        self.pass(
            &mut enc,
            &self.velocity.write.view,
            Load::Clear,
            &self.pipes.gradient,
            &bg,
            self.open(3),
        );
        self.velocity.swap();

        // 7. The field carries itself.
        let u = self.params(
            self.sim,
            Vol {
                texel: [0.0, 0.0, 0.0, dt],
                params: [c.velocity_dissipation, 0.0, 0.0, 0.0],
                ..Default::default()
            },
        );
        let bg = self.bind(&u, &self.velocity.read, Some(&self.velocity.read));
        self.pass(
            &mut enc,
            &self.velocity.write.view,
            Load::Clear,
            &self.pipes.advect_velocity,
            &bg,
            self.open(4),
        );
        self.velocity.swap();

        // 8. …and it carries the dye. The velocity is in *solver* voxels a
        // second, and the advection steps in the sampled field's own voxels,
        // so the dye pass needs the solver's texel size rather than its own.
        let u = self.params_mixed(
            self.dye_atlas,
            self.sim,
            self.dye_atlas,
            Vol {
                texel: [0.0, 0.0, 0.0, dt],
                params: [c.density_dissipation, 0.0, 0.0, 0.0],
                ..Default::default()
            },
        );
        let bg = self.bind(&u, &self.velocity.read, Some(&self.dye.read));
        self.pass(
            &mut enc,
            &self.dye.write.view,
            Load::Clear,
            &self.pipes.advect_dye,
            &bg,
            None,
        );
        self.dye.swap();
        self.occupancy_stale.set(true);

        self.queue.submit([enc.finish()]);
    }

    /// `count` Jacobi sweeps on the top-level pressure, ping-ponging.
    ///
    /// No relaxation weight, unlike the textbook multigrid smoother. The
    /// argument is the one the 2D module sets out at `MG_WEIGHT` and it
    /// carries over unchanged: damping is there to kill the checkerboard
    /// mode, this divergence is a central difference over two voxels, and
    /// that operator has a zero at exactly that frequency — so there is no
    /// checkerboard in the right-hand side to kill, and damping would only
    /// slow the modes that do exist.
    fn record_jacobi(
        &mut self,
        enc: &mut wgpu::CommandEncoder,
        count: u32,
        first_scale: f32,
        open: Option<u32>,
    ) {
        if count == 0 {
            return;
        }
        // Two uniform blocks, because the first sweep may scale what it
        // reads and the rest never do — see `fs_pressure3`.
        let scaled = self.params(
            self.sim,
            Vol {
                params: [0.0, 0.0, first_scale, 0.0],
                ..Default::default()
            },
        );
        let plain = self.params(
            self.sim,
            Vol {
                params: [0.0, 0.0, 1.0, 0.0],
                ..Default::default()
            },
        );
        let groups = [
            self.bind(&scaled, &self.pressure.read, Some(&self.divergence)),
            self.bind(&scaled, &self.pressure.write, Some(&self.divergence)),
            self.bind(&plain, &self.pressure.read, Some(&self.divergence)),
            self.bind(&plain, &self.pressure.write, Some(&self.divergence)),
        ];
        let mut marks = self.writes(open);
        for i in 0..count {
            let even = i % 2 == 0;
            let bg = &groups[usize::from(i > 0) * 2 + usize::from(!even)];
            let target = if even {
                &self.pressure.write.view
            } else {
                &self.pressure.read.view
            };
            self.pass(
                enc,
                target,
                Load::Clear,
                &self.pipes.pressure,
                bg,
                marks.take(),
            );
        }
        if count % 2 == 1 {
            self.pressure.swap();
        }
    }

    /// A multigrid V-cycle: down to the coarsest grid, then back up.
    ///
    /// The shape is the textbook one, and the same one the 2D module uses. On
    /// the way down each level smooths a little, measures what it failed to
    /// satisfy, and hands that residual to a grid half the size as a fresh
    /// problem starting from zero. At the bottom — four to eight voxels
    /// across — a few sweeps *are* a solve. On the way up each correction is
    /// interpolated into the level above and smoothed again, to take the
    /// interpolation's own roughness back out.
    ///
    /// It matters more here than in two dimensions. A sweep moves information
    /// one voxel either way, and a volume is wider in the sense that counts:
    /// the pressure from a puff falls off as 1/r rather than log r, and six
    /// neighbours share the correction instead of four.
    fn record_multigrid(&mut self, enc: &mut wgpu::CommandEncoder, open: Option<u32>) {
        if self.mg.is_empty() {
            let (count, scale) = (self.config.pressure_iterations, self.config.pressure);
            self.record_jacobi(enc, count, scale, open);
            return;
        }
        let nu = self.config.multigrid_smooth.max(1);
        // The bottom grid gets enough sweeps to *be* a solve, and no more.
        // Information crosses a grid one voxel a sweep, so twice the side is
        // the whole domain twice over — on a 4³ grid that is eight sweeps,
        // and the twenty the 2D module spends there would be twenty render
        // passes over sixty-four voxels. The arithmetic is free at that size;
        // the passes are not.
        let coarsest = self.mg.last().map(|l| l.atlas.grid).unwrap_or(4);
        let coarse_sweeps = (2 * coarsest).clamp(nu * 2, self.config.pressure_iterations.max(8));
        let mut open = open;
        for cycle in 0..self.config.multigrid_cycles.max(1) {
            // --- down -------------------------------------------------
            // Only the very first sweep carries the carry-over scale; after
            // that the pressure is this frame's and is read as it stands.
            let scale = if cycle == 0 {
                self.config.pressure
            } else {
                1.0
            };
            self.record_jacobi(enc, nu, scale, open.take());
            for level in 0..self.mg.len() {
                self.record_residual3(enc, level);
                self.record_restrict3(enc, level);
                let sweeps = if level + 1 == self.mg.len() {
                    coarse_sweeps
                } else {
                    nu
                };
                // The coarse problem is for a *correction*, so it starts at
                // zero. A scale of zero says so, and the sweep that used to
                // follow a clear-only pass now *is* the clear: it reads
                // nothing and writes `-rhs / 6`.
                self.record_smooth3(enc, level, sweeps, 0.0);
            }
            // --- up ---------------------------------------------------
            for level in (0..self.mg.len()).rev() {
                self.record_prolong3(enc, level);
                if level == 0 {
                    self.record_jacobi(enc, nu, 1.0, None);
                } else {
                    self.record_smooth3(enc, level - 1, nu, 1.0);
                }
            }
        }
    }

    /// Smooth coarse `level`'s correction against its own right-hand side.
    fn record_smooth3(
        &mut self,
        enc: &mut wgpu::CommandEncoder,
        level: usize,
        sweeps: u32,
        first_scale: f32,
    ) {
        if sweeps == 0 {
            return;
        }
        let atlas = self.mg[level].atlas;
        let scaled = self.params(
            atlas,
            Vol {
                params: [0.0, 0.0, first_scale, 0.0],
                ..Default::default()
            },
        );
        let plain = self.params(
            atlas,
            Vol {
                params: [0.0, 0.0, 1.0, 0.0],
                ..Default::default()
            },
        );
        let l = &self.mg[level];
        let groups = [
            self.bind(&scaled, &l.pressure.read, Some(&l.rhs)),
            self.bind(&scaled, &l.pressure.write, Some(&l.rhs)),
            self.bind(&plain, &l.pressure.read, Some(&l.rhs)),
            self.bind(&plain, &l.pressure.write, Some(&l.rhs)),
        ];
        for i in 0..sweeps {
            let even = i % 2 == 0;
            let bg = &groups[usize::from(i > 0) * 2 + usize::from(!even)];
            let target = if even {
                &l.pressure.write.view
            } else {
                &l.pressure.read.view
            };
            self.pass(enc, target, Load::Clear, &self.pipes.pressure, bg, None);
        }
        if sweeps % 2 == 1 {
            self.mg[level].pressure.swap();
        }
    }

    /// What the level above `level` fails to satisfy, at its own resolution.
    ///
    /// At its own resolution and not straight into the coarse texture: that
    /// would be a point sample, which is the one thing a restriction must not
    /// be — it aliases exactly the high-frequency error the smoother was
    /// there to remove.
    fn record_residual3(&mut self, enc: &mut wgpu::CommandEncoder, level: usize) {
        let (atlas, pressure, rhs, target) = if level == 0 {
            (
                self.sim,
                &self.pressure.read,
                &self.divergence,
                &self.residual,
            )
        } else {
            let l = &self.mg[level - 1];
            (l.atlas, &l.pressure.read, &l.rhs, &l.residual)
        };
        let u = self.params(atlas, Vol::default());
        let bg = self.bind(&u, pressure, Some(rhs));
        self.pass(
            enc,
            &target.view,
            Load::Clear,
            &self.pipes.residual,
            &bg,
            None,
        );
    }

    /// Average `level`'s residual onto the half-size grid below it.
    fn record_restrict3(&mut self, enc: &mut wgpu::CommandEncoder, level: usize) {
        let (fine_atlas, residual) = if level == 0 {
            (self.sim, &self.residual)
        } else {
            let l = &self.mg[level - 1];
            (l.atlas, &l.residual)
        };
        let coarse = self.mg[level].atlas;
        let u = self.params_mixed(coarse, fine_atlas, fine_atlas, Vol::default());
        let bg = self.bind(&u, residual, None);
        self.pass(
            enc,
            &self.mg[level].rhs.view,
            Load::Clear,
            &self.pipes.restrict,
            &bg,
            None,
        );
    }

    /// Interpolate `level`'s correction into the level above and add it.
    fn record_prolong3(&mut self, enc: &mut wgpu::CommandEncoder, level: usize) {
        let coarse_atlas = self.mg[level].atlas;
        let (fine_atlas, fine_read, fine_write) = if level == 0 {
            (self.sim, &self.pressure.read, &self.pressure.write)
        } else {
            let l = &self.mg[level - 1];
            (l.atlas, &l.pressure.read, &l.pressure.write)
        };
        let u = self.params_mixed(fine_atlas, coarse_atlas, fine_atlas, Vol::default());
        let bg = self.bind(&u, &self.mg[level].pressure.read, Some(fine_read));
        self.pass(
            enc,
            &fine_write.view,
            Load::Clear,
            &self.pipes.prolong,
            &bg,
            None,
        );
        if level == 0 {
            self.pressure.swap();
        } else {
            self.mg[level - 1].pressure.swap();
        }
    }

    /// Reduce the dye to its largest value per brick, twice.
    ///
    /// Recorded at render rather than at the end of the step, because a splat
    /// lands outside the step: building it there would let a fresh splat fall
    /// into a brick still marked empty, and the march would step straight
    /// over it.
    fn record_occupancy(&self, enc: &mut wgpu::CommandEncoder, opening: Option<u32>) {
        let mut source = (self.dye_atlas, &self.dye.read);
        let mut marks = self.writes(opening);
        for level in 0..self.occupancy.len() {
            let (atlas, target) = &self.occupancy[level];
            let u = self.params_mixed(*atlas, source.0, source.0, Vol::default());
            let bg = self.bind(&u, source.1, None);
            self.pass(
                enc,
                &target.view,
                Load::Clear,
                &self.pipes.maxdown,
                &bg,
                marks.take(),
            );
            source = (*atlas, target);
        }
    }

    /// March the dye volume from `camera` into `target`.
    ///
    /// Emission and absorption, front to back, stopping a ray once what is
    /// left cannot change the pixel. The volume occupies
    /// [`bounds_min`](FluidConfig3d::bounds_min)`..`[`bounds_max`](FluidConfig3d::bounds_max)
    /// in world space, so it composites with anything else drawn with the same
    /// camera — though it writes alpha rather than reading depth, so it goes
    /// over a scene rather than into it.
    pub fn render(&mut self, camera: &dyn Camera, target: &wgpu::Texture) {
        let view = target.create_view(&wgpu::TextureViewDescriptor::default());
        let size = (target.width(), target.height());
        self.render_to_view(camera, &view, target.format(), size);
    }

    /// March into a view of `size` pixels.
    ///
    /// The size is not decoration: a `TextureView` does not carry its
    /// dimensions, and [`temporal_samples`](FluidConfig3d::temporal_samples)
    /// needs a buffer that matches the target exactly.
    pub fn render_to_view(
        &mut self,
        camera: &dyn Camera,
        target: &wgpu::TextureView,
        format: wgpu::TextureFormat,
        size: (u32, u32),
    ) {
        let c = self.config;
        self.next_ubo.set(0);
        let accumulating = self.prepare_accum(size);
        let view_proj = camera.projection_matrix().multiply(&camera.view_matrix());
        let inv = view_proj.invert();
        self.view_proj = view_proj.elements;
        self.view_proj_inverse = inv.elements;
        let pos = camera.position();
        // A brick below this cannot change a pixel: the march's own early-out
        // gives up at a thousandth of transmittance. Zero turns skipping off,
        // and the occupancy volume stays bound either way so that what the
        // uniform says about `b` is always what is in `b`.
        let skip_floor = if c.empty_space_skip { 0.0008 } else { 0.0 };
        let u = self.params_mixed(
            self.dye_atlas,
            self.dye_atlas,
            self.occupancy[1].0,
            Vol {
                params: [skip_floor, self.accum_frame as f32, 0.0, 0.0],
                bounds_min: [c.bounds_min[0], c.bounds_min[1], c.bounds_min[2], 0.0],
                bounds_max: [c.bounds_max[0], c.bounds_max[1], c.bounds_max[2], 0.0],
                march: [
                    c.march_steps.max(1) as f32,
                    c.absorption,
                    c.emission,
                    c.exposure,
                ],
                background: [
                    c.background[0],
                    c.background[1],
                    c.background[2],
                    if c.transparent { 1.0 } else { 0.0 },
                ],
                cam_pos: [pos.x, pos.y, pos.z, 1.0],
                inv_view_proj: inv.elements,
                prev_view_proj: self.accum_from.unwrap_or(view_proj.elements),
                ..Default::default()
            },
        );
        let bg = self.bind(&u, &self.dye.read, Some(&self.occupancy[1].1));
        let mut enc = self.encoder("fluid3d render");
        // Here rather than at the end of the step, because the march is the
        // only thing that reads it: a simulation stepped a hundred times and
        // drawn once would otherwise reduce the dye a hundred times and use
        // the answer once. Two passes a step, which is five percent of one.
        //
        // It carries the mark that closes the last stage when it runs; the
        // march carries it when it does not.
        if c.empty_space_skip && self.occupancy_stale.get() {
            let closing = self.take_mark(STAGE_NAMES.len() as u32);
            self.record_occupancy(&mut enc, closing);
            self.occupancy_stale.set(false);
        }
        // Carries the mark that closes the last stage, if the occupancy
        // reduction did not — it is the next pass either way. Taken before
        // the pipelines, which borrow `self` mutably to compile one on first
        // sight of a format.
        let mark = self.take_mark(STAGE_NAMES.len() as u32);
        if accumulating {
            self.record_accumulated_march(&mut enc, target, format, &bg, mark);
        } else {
            let (device, layout, module) = (&self.device, &self.layout, &self.module);
            let pipeline = self.out.get_or_build(format, || {
                build(device, layout, module, "fs_raymarch", format)
            });
            let marks = match (mark, &self.profiler) {
                (Some(i), Some(p)) => Some(p.mark(i)),
                _ => None,
            };
            draw_timed(&mut enc, target, Load::Clear, pipeline, &bg, marks);
        }
        // The march is the last pass of the frame, so it also carries the
        // mark that closes it — and the resolve only happens when every
        // stage opened, so a frame that skipped one reports nothing rather
        // than six wrong numbers.
        let complete = self.mark.get() == MARKS;
        if let Some(p) = &self.profiler {
            if complete {
                p.resolve(&mut enc);
            }
        }
        if c.show_fps {
            self.record_overlay(&mut enc, target, format);
        }
        self.queue.submit([enc.finish()]);
        if let Some(p) = &mut self.profiler {
            p.poll(&self.device);
        }
        if accumulating {
            self.accum_frame += 1;
            self.accum_from = Some(view_proj.elements);
        }
        self.rate.tick();
    }

    /// Composite the read-out over the target.
    fn record_overlay(
        &mut self,
        enc: &mut wgpu::CommandEncoder,
        target: &wgpu::TextureView,
        format: wgpu::TextureFormat,
    ) {
        // The target the chart is drawn against: what the display is
        // actually delivering, rounded to a sensible refresh, so the line
        // means "keeping up" rather than an arbitrary sixty.
        let target_ms = self.rate.target_ms();
        // Twice the current frame time, so a steady run sits halfway up
        // however fast it is going.
        let scale_ms = self.rate.scale_ms();
        let u = self.params(
            self.dye_atlas,
            Vol {
                params: [self.rate.fps(), scale_ms, target_ms, 0.0],
                ..Default::default()
            },
        );
        let history = {
            let (device, queue) = (self.device.clone(), self.queue.clone());
            self.history
                .view(&device, &queue, self.rate.frames())
                .clone()
        };
        let bind = self.device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("fluid3d read-out"),
            layout: &self.bgl,
            entries: &[
                wgpu::BindGroupEntry {
                    binding: 0,
                    resource: u.buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 1,
                    resource: wgpu::BindingResource::TextureView(&history),
                },
                wgpu::BindGroupEntry {
                    binding: 2,
                    resource: wgpu::BindingResource::TextureView(&history),
                },
                wgpu::BindGroupEntry {
                    binding: 3,
                    resource: wgpu::BindingResource::Sampler(&self.sampler),
                },
                wgpu::BindGroupEntry {
                    binding: 4,
                    resource: wgpu::BindingResource::TextureView(&history),
                },
            ],
        });
        let (device, layout, module) = (&self.device, &self.layout, &self.module);
        let pipeline = self.overlay.get_or_build(format, || {
            build_blended(
                device,
                layout,
                module,
                "fs_overlay3",
                format,
                wgpu::BlendState::ALPHA_BLENDING,
            )
        });
        // Keeps what is already there: the read-out is lit pixels over a
        // picture, and every other pixel is discarded.
        draw(enc, target, Load::Keep, pipeline, &bind);
    }

    /// Make sure there is an accumulation buffer of the right size, and
    /// decide whether this frame continues the average or starts it.
    ///
    /// Returns false when accumulation is off, which is the only path that
    /// touches no extra memory.
    fn prepare_accum(&mut self, size: (u32, u32)) -> bool {
        if self.config.temporal_samples <= 1 {
            self.accum = None;
            self.current = None;
            self.depth = None;
            self.accum_from = None;
            return false;
        }
        let (w, h) = (size.0.max(1), size.1.max(1));
        let resized = !matches!(&self.current, Some(f) if f.width == w && f.height == h);
        if resized {
            let fmt = self.formats.dye;
            self.current = Some(Fbo::new(&self.device, "fluid3d march", w, h, fmt));
            self.accum = Some(DoubleFbo::new(&self.device, "fluid3d history", w, h, fmt));
            // One channel, and it holds a distance in world units rather
            // than a normalised depth, so it wants the range a float gives.
            self.depth = Some(Fbo::new(
                &self.device,
                "fluid3d march depth",
                w,
                h,
                self.formats.scalar,
            ));
        }
        // Any camera movement and the pixels no longer line up, so the
        // history is not an average of this view, it is a smear of the last
        // one. Nothing here reprojects: a volume has no single depth to
        // reproject *by*, and guessing one puts the plume's near face and
        // its far face in different places.
        // Not reset on movement any more: the resolve reprojects through
        // `accum_from`, so a moved camera still has a usable history. Only
        // a resize invalidates it outright, because then the pixels are a
        // different set.
        if resized {
            self.accum_from = None;
            self.accum_frame = 0;
        }
        true
    }

    /// March into the accumulation, then put it on the target.
    ///
    /// Two passes where there was one, and the second is a copy — which is
    /// the price of the first being able to blend with what it wrote last
    /// time. `set_blend_constant` carries the weight, so one pipeline serves
    /// every sample count and the frame after a reset can set it to 1 and
    /// simply overwrite.
    fn record_accumulated_march(
        &mut self,
        enc: &mut wgpu::CommandEncoder,
        target: &wgpu::TextureView,
        format: wgpu::TextureFormat,
        bg: &wgpu::BindGroup,
        mark: Option<u32>,
    ) {
        let weight = if self.accum_frame == 0 {
            1.0
        } else {
            1.0 / self.config.temporal_samples.max(2) as f32
        };
        let accum_format = self.formats.dye;
        let depth_format = self.formats.scalar;
        if self.accum_march.is_none() {
            self.accum_march = Some(build_targets(
                &self.device,
                &self.layout,
                &self.module,
                "fs_raymarch_depth",
                &[Some(accum_format), Some(depth_format)],
            ));
        }
        {
            let current = self.current.as_ref().expect("prepare_accum made one");
            let depth = self.depth.as_ref().expect("prepare_accum made one");
            let marks = match (mark, &self.profiler) {
                (Some(i), Some(p)) => Some(p.mark(i)),
                _ => None,
            };
            fn attach(view: &wgpu::TextureView) -> wgpu::RenderPassColorAttachment<'_> {
                wgpu::RenderPassColorAttachment {
                    view,
                    depth_slice: None,
                    resolve_target: None,
                    ops: wgpu::Operations {
                        load: wgpu::LoadOp::Clear(wgpu::Color::TRANSPARENT),
                        store: wgpu::StoreOp::Store,
                    },
                }
            }
            let mut pass = enc.begin_render_pass(&wgpu::RenderPassDescriptor {
                label: Some("fluid3d march"),
                color_attachments: &[Some(attach(&current.view)), Some(attach(&depth.view))],
                depth_stencil_attachment: None,
                timestamp_writes: marks,
                occlusion_query_set: None,
                multiview_mask: None,
            });
            pass.set_pipeline(self.accum_march.as_ref().expect("just built"));
            pass.set_bind_group(0, bg, &[]);
            pass.draw(0..3, 0..1);
        }

        // The resolve, writing the history and the caller's target together.
        // Two attachments rather than two passes: the value is the same and
        // only the format differs, so letting the hardware convert it twice
        // is cheaper than reading it back to convert it once.
        let u = self.params(
            self.dye_atlas,
            Vol {
                params: [0.0, 0.0, weight, self.config.temporal_clip],
                // The resolve unprojects this frame's pixel and projects it
                // through last frame's camera, so it needs both.
                inv_view_proj: self.view_proj_inverse,
                prev_view_proj: self.accum_from.unwrap_or(self.view_proj),
                ..Default::default()
            },
        );
        let current = self.current.take().expect("prepare_accum made one");
        let depth = self.depth.take().expect("prepare_accum made one");
        let mut accum = self.accum.take().expect("prepare_accum made one");
        let bind = self.bind3(&u, &current, &accum.read, &depth);
        {
            let (device, layout, module) = (&self.device, &self.layout, &self.module);
            let pipeline = self.resolve.get_or_build(format, || {
                build_targets(
                    device,
                    layout,
                    module,
                    "fs_resolve3",
                    &[Some(accum_format), Some(format)],
                )
            });
            let mut pass = enc.begin_render_pass(&wgpu::RenderPassDescriptor {
                label: Some("fluid3d resolve"),
                color_attachments: &[
                    Some(wgpu::RenderPassColorAttachment {
                        view: &accum.write.view,
                        depth_slice: None,
                        resolve_target: None,
                        ops: wgpu::Operations {
                            load: wgpu::LoadOp::Clear(wgpu::Color::TRANSPARENT),
                            store: wgpu::StoreOp::Store,
                        },
                    }),
                    Some(wgpu::RenderPassColorAttachment {
                        view: target,
                        depth_slice: None,
                        resolve_target: None,
                        ops: wgpu::Operations {
                            load: wgpu::LoadOp::Clear(wgpu::Color::TRANSPARENT),
                            store: wgpu::StoreOp::Store,
                        },
                    }),
                ],
                depth_stencil_attachment: None,
                timestamp_writes: None,
                occlusion_query_set: None,
                multiview_mask: None,
            });
            pass.set_pipeline(pipeline);
            pass.set_bind_group(0, &bind, &[]);
            pass.draw(0..3, 0..1);
        }
        accum.swap();
        self.current = Some(current);
        self.depth = Some(depth);
        self.accum = Some(accum);
    }

    /// Render into a fresh `Rgba8Unorm` texture and read it back. Native only.
    #[cfg(not(target_arch = "wasm32"))]
    pub fn capture_rgba(&mut self, camera: &dyn Camera, width: u32, height: u32) -> Vec<u8> {
        let target = self.device.create_texture(&wgpu::TextureDescriptor {
            label: Some("fluid3d capture"),
            size: wgpu::Extent3d {
                width: width.max(1),
                height: height.max(1),
                depth_or_array_layers: 1,
            },
            mip_level_count: 1,
            sample_count: 1,
            dimension: wgpu::TextureDimension::D2,
            format: wgpu::TextureFormat::Rgba8Unorm,
            usage: wgpu::TextureUsages::RENDER_ATTACHMENT | wgpu::TextureUsages::COPY_SRC,
            view_formats: &[],
        });
        self.render(camera, &target);
        crate::renderer::readback::texture_bytes(&self.device, &self.queue, &target)
    }

    // -- helpers ----------------------------------------------------------

    fn encoder(&self, label: &str) -> wgpu::CommandEncoder {
        self.device
            .create_command_encoder(&wgpu::CommandEncoderDescriptor { label: Some(label) })
    }

    /// Fill in the atlas geometry and take the next free uniform block. Every
    /// field on the same grid, which is every pass but the two that mix the
    /// velocity with the dye.
    fn params(&self, atlas: Atlas, rest: Vol) -> Ubo {
        self.params_mixed(atlas, atlas, atlas, rest)
    }

    /// As [`Self::params`], with each texture's own grid given separately.
    ///
    /// Sampling one field with another's tiling is the bug this exists to
    /// prevent, and it is a quiet one: the dye advection read the velocity
    /// atlas through the dye's tile layout and still produced a plausible
    /// plume.
    fn params_mixed(&self, out: Atlas, a: Atlas, b: Atlas, rest: Vol) -> Ubo {
        let v = Vol {
            out_grid: out.into(),
            a_grid: a.into(),
            b_grid: b.into(),
            ..rest
        };
        // One buffer a pass: `write_buffer` orders against submits, not
        // against draws, so two passes sharing a block in one submit both see
        // whatever was written last. The pool grows rather than wrapping —
        // wrapping is that bug with a cap on it, and a V-cycle needs five
        // blocks a level.
        let i = self.next_ubo.get();
        self.next_ubo.set(i + 1);
        let buffer = {
            let mut pool = self.ubos.borrow_mut();
            while pool.len() <= i {
                pool.push(ubo(&self.device));
            }
            pool[i].clone()
        };
        // Most of these blocks are the same every frame — a grid description
        // and a dt that has not changed — and the pass order is fixed, so
        // slot `i` sees the same contents it saw last frame. Comparing two
        // hundred and forty bytes is cheaper than staging them.
        if !self.config.cache_bindings {
            self.queue.write_buffer(&buffer, 0, bytemuck::bytes_of(&v));
            return Ubo { buffer, slot: i };
        }
        let mut seen = self.ubo_contents.borrow_mut();
        while seen.len() <= i {
            // A block nothing has written yet cannot match: `march.x` is a
            // step count and is never zero in a real one.
            seen.push(Vol::default());
        }
        if bytemuck::bytes_of(&seen[i]) != bytemuck::bytes_of(&v) {
            seen[i] = v;
            drop(seen);
            self.queue.write_buffer(&buffer, 0, bytemuck::bytes_of(&v));
        }
        Ubo { buffer, slot: i }
    }

    fn bind(&self, params: &Ubo, a: &Fbo, b: Option<&Fbo>) -> wgpu::BindGroup {
        if !self.config.cache_bindings {
            return self.build_bind(&params.buffer, a, b);
        }
        // The *uniform's own* slot, not whichever was handed out last. Those
        // are the same number only while every block is bound immediately
        // after it is filled — and the moment a pass wants two blocks before
        // it binds either, keying on the latter silently hands the second
        // pass the first one's bind group. Which is a solver that stops
        // converging and a sweep count that stops mattering.
        let key = (params.slot, a.id, b.map(|f| f.id).unwrap_or(a.id));
        if let Some(hit) = self.binds.borrow().get(&key) {
            return hit.clone();
        }
        let group = self.build_bind(&params.buffer, a, b);
        self.binds.borrow_mut().insert(key, group.clone());
        group
    }

    fn build_bind(&self, params: &wgpu::Buffer, a: &Fbo, b: Option<&Fbo>) -> wgpu::BindGroup {
        self.build_bind3(params, a, b, None)
    }

    fn build_bind3(
        &self,
        params: &wgpu::Buffer,
        a: &Fbo,
        b: Option<&Fbo>,
        c: Option<&Fbo>,
    ) -> wgpu::BindGroup {
        self.device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("fluid3d bind"),
            layout: &self.bgl,
            entries: &[
                wgpu::BindGroupEntry {
                    binding: 0,
                    resource: params.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 1,
                    resource: wgpu::BindingResource::TextureView(&a.view),
                },
                wgpu::BindGroupEntry {
                    binding: 2,
                    resource: wgpu::BindingResource::TextureView(
                        b.map(|f| &f.view).unwrap_or(&a.view),
                    ),
                },
                wgpu::BindGroupEntry {
                    binding: 3,
                    resource: wgpu::BindingResource::Sampler(&self.sampler),
                },
                wgpu::BindGroupEntry {
                    binding: 4,
                    resource: wgpu::BindingResource::TextureView(
                        c.map(|f| &f.view).unwrap_or(&a.view),
                    ),
                },
            ],
        })
    }

    /// [`bind`](Self::bind), with a third texture. Not cached: the one pass
    /// that uses it runs once a frame, and the cache is keyed on two.
    fn bind3(&self, params: &Ubo, a: &Fbo, b: &Fbo, c: &Fbo) -> wgpu::BindGroup {
        self.build_bind3(&params.buffer, a, Some(b), Some(c))
    }
}

fn tex(binding: u32) -> wgpu::BindGroupLayoutEntry {
    wgpu::BindGroupLayoutEntry {
        binding,
        visibility: wgpu::ShaderStages::FRAGMENT,
        ty: wgpu::BindingType::Texture {
            sample_type: wgpu::TextureSampleType::Float { filterable: true },
            view_dimension: wgpu::TextureViewDimension::D2,
            multisampled: false,
        },
        count: None,
    }
}

fn ubo(device: &wgpu::Device) -> wgpu::Buffer {
    device.create_buffer(&wgpu::BufferDescriptor {
        label: Some("fluid3d params"),
        size: VOL_SIZE,
        usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
        mapped_at_creation: false,
    })
}

fn build(
    device: &wgpu::Device,
    layout: &wgpu::PipelineLayout,
    module: &wgpu::ShaderModule,
    fs: &str,
    format: wgpu::TextureFormat,
) -> wgpu::RenderPipeline {
    build_targets(device, layout, module, fs, &[Some(format)])
}

/// [`build`], for a stage that writes more than one attachment — the resolve,
/// which puts the same value into the history and the caller's target and
/// lets the hardware convert it to each format.
fn build_targets(
    device: &wgpu::Device,
    layout: &wgpu::PipelineLayout,
    module: &wgpu::ShaderModule,
    fs: &str,
    formats: &[Option<wgpu::TextureFormat>],
) -> wgpu::RenderPipeline {
    let targets: Vec<_> = formats
        .iter()
        .map(|f| {
            f.map(|format| wgpu::ColorTargetState {
                format,
                blend: None,
                write_mask: wgpu::ColorWrites::ALL,
            })
        })
        .collect();
    build_with(device, layout, module, fs, &targets)
}

/// [`build`], for a stage that composites over what is already there.
fn build_blended(
    device: &wgpu::Device,
    layout: &wgpu::PipelineLayout,
    module: &wgpu::ShaderModule,
    fs: &str,
    format: wgpu::TextureFormat,
    blend: wgpu::BlendState,
) -> wgpu::RenderPipeline {
    build_with(
        device,
        layout,
        module,
        fs,
        &[Some(wgpu::ColorTargetState {
            format,
            blend: Some(blend),
            write_mask: wgpu::ColorWrites::ALL,
        })],
    )
}

fn build_with(
    device: &wgpu::Device,
    layout: &wgpu::PipelineLayout,
    module: &wgpu::ShaderModule,
    fs: &str,
    targets: &[Option<wgpu::ColorTargetState>],
) -> wgpu::RenderPipeline {
    device.create_render_pipeline(&wgpu::RenderPipelineDescriptor {
        label: Some(fs),
        layout: Some(layout),
        vertex: wgpu::VertexState {
            module,
            entry_point: Some(fullscreen::VS_ENTRY),
            buffers: &[],
            compilation_options: Default::default(),
        },
        fragment: Some(wgpu::FragmentState {
            module,
            entry_point: Some(fs),
            targets,
            compilation_options: Default::default(),
        }),
        primitive: wgpu::PrimitiveState::default(),
        depth_stencil: None,
        multisample: wgpu::MultisampleState::default(),
        multiview_mask: None,
        cache: None,
    })
}

#[cfg(test)]
#[path = "volume_tests.rs"]
mod tests;