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
|
/*
* Copyright © 2010 Intel Corporation
* Copyright © 2014-2017 Broadcom
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*/
/**
* @file
*
* The basic model of the list scheduler is to take a basic block, compute a
* DAG of the dependencies, and make a list of the DAG heads. Heuristically
* pick a DAG head, then put all the children that are now DAG heads into the
* list of things to schedule.
*
* The goal of scheduling here is to pack pairs of operations together in a
* single QPU instruction.
*/
#include "qpu/qpu_disasm.h"
#include "v3d_compiler.h"
#include "util/ralloc.h"
static bool debug;
struct schedule_node_child;
struct schedule_node {
struct list_head link;
struct qinst *inst;
struct schedule_node_child *children;
uint32_t child_count;
uint32_t child_array_size;
uint32_t parent_count;
/* Longest cycles + instruction_latency() of any parent of this node. */
uint32_t unblocked_time;
/**
* Minimum number of cycles from scheduling this instruction until the
* end of the program, based on the slowest dependency chain through
* the children.
*/
uint32_t delay;
/**
* cycles between this instruction being scheduled and when its result
* can be consumed.
*/
uint32_t latency;
};
struct schedule_node_child {
struct schedule_node *node;
bool write_after_read;
};
/* When walking the instructions in reverse, we need to swap before/after in
* add_dep().
*/
enum direction { F, R };
struct schedule_state {
const struct v3d_device_info *devinfo;
struct schedule_node *last_r[6];
struct schedule_node *last_rf[64];
struct schedule_node *last_sf;
struct schedule_node *last_vpm_read;
struct schedule_node *last_tmu_write;
struct schedule_node *last_tmu_config;
struct schedule_node *last_tlb;
struct schedule_node *last_vpm;
struct schedule_node *last_unif;
struct schedule_node *last_rtop;
enum direction dir;
/* Estimated cycle when the current instruction would start. */
uint32_t time;
};
static void
add_dep(struct schedule_state *state,
struct schedule_node *before,
struct schedule_node *after,
bool write)
{
bool write_after_read = !write && state->dir == R;
if (!before || !after)
return;
assert(before != after);
if (state->dir == R) {
struct schedule_node *t = before;
before = after;
after = t;
}
for (int i = 0; i < before->child_count; i++) {
if (before->children[i].node == after &&
(before->children[i].write_after_read == write_after_read)) {
return;
}
}
if (before->child_array_size <= before->child_count) {
before->child_array_size = MAX2(before->child_array_size * 2, 16);
before->children = reralloc(before, before->children,
struct schedule_node_child,
before->child_array_size);
}
before->children[before->child_count].node = after;
before->children[before->child_count].write_after_read =
write_after_read;
before->child_count++;
after->parent_count++;
}
static void
add_read_dep(struct schedule_state *state,
struct schedule_node *before,
struct schedule_node *after)
{
add_dep(state, before, after, false);
}
static void
add_write_dep(struct schedule_state *state,
struct schedule_node **before,
struct schedule_node *after)
{
add_dep(state, *before, after, true);
*before = after;
}
static bool
qpu_inst_is_tlb(const struct v3d_qpu_instr *inst)
{
if (inst->type != V3D_QPU_INSTR_TYPE_ALU)
return false;
if (inst->alu.add.magic_write &&
(inst->alu.add.waddr == V3D_QPU_WADDR_TLB ||
inst->alu.add.waddr == V3D_QPU_WADDR_TLBU))
return true;
if (inst->alu.mul.magic_write &&
(inst->alu.mul.waddr == V3D_QPU_WADDR_TLB ||
inst->alu.mul.waddr == V3D_QPU_WADDR_TLBU))
return true;
return false;
}
static void
process_mux_deps(struct schedule_state *state, struct schedule_node *n,
enum v3d_qpu_mux mux)
{
switch (mux) {
case V3D_QPU_MUX_A:
add_read_dep(state, state->last_rf[n->inst->qpu.raddr_a], n);
break;
case V3D_QPU_MUX_B:
add_read_dep(state, state->last_rf[n->inst->qpu.raddr_b], n);
break;
default:
add_read_dep(state, state->last_r[mux - V3D_QPU_MUX_R0], n);
break;
}
}
static void
process_waddr_deps(struct schedule_state *state, struct schedule_node *n,
uint32_t waddr, bool magic)
{
if (!magic) {
add_write_dep(state, &state->last_rf[waddr], n);
} else if (v3d_qpu_magic_waddr_is_tmu(waddr)) {
add_write_dep(state, &state->last_tmu_write, n);
switch (waddr) {
case V3D_QPU_WADDR_TMUS:
case V3D_QPU_WADDR_TMUSCM:
case V3D_QPU_WADDR_TMUSF:
case V3D_QPU_WADDR_TMUSLOD:
add_write_dep(state, &state->last_tmu_config, n);
break;
default:
break;
}
} else if (v3d_qpu_magic_waddr_is_sfu(waddr)) {
/* Handled by v3d_qpu_writes_r4() check. */
} else {
switch (waddr) {
case V3D_QPU_WADDR_R0:
case V3D_QPU_WADDR_R1:
case V3D_QPU_WADDR_R2:
add_write_dep(state,
&state->last_r[waddr - V3D_QPU_WADDR_R0],
n);
break;
case V3D_QPU_WADDR_R3:
case V3D_QPU_WADDR_R4:
case V3D_QPU_WADDR_R5:
/* Handled by v3d_qpu_writes_r*() checks below. */
break;
case V3D_QPU_WADDR_VPM:
case V3D_QPU_WADDR_VPMU:
add_write_dep(state, &state->last_vpm, n);
break;
case V3D_QPU_WADDR_TLB:
case V3D_QPU_WADDR_TLBU:
add_write_dep(state, &state->last_tlb, n);
break;
case V3D_QPU_WADDR_NOP:
break;
default:
fprintf(stderr, "Unknown waddr %d\n", waddr);
abort();
}
}
}
static void
process_cond_deps(struct schedule_state *state, struct schedule_node *n,
enum v3d_qpu_cond cond)
{
if (cond != V3D_QPU_COND_NONE)
add_read_dep(state, state->last_sf, n);
}
static void
process_pf_deps(struct schedule_state *state, struct schedule_node *n,
enum v3d_qpu_pf pf)
{
if (pf != V3D_QPU_PF_NONE)
add_write_dep(state, &state->last_sf, n);
}
static void
process_uf_deps(struct schedule_state *state, struct schedule_node *n,
enum v3d_qpu_uf uf)
{
if (uf != V3D_QPU_UF_NONE)
add_write_dep(state, &state->last_sf, n);
}
/**
* Common code for dependencies that need to be tracked both forward and
* backward.
*
* This is for things like "all reads of r4 have to happen between the r4
* writes that surround them".
*/
static void
calculate_deps(struct schedule_state *state, struct schedule_node *n)
{
const struct v3d_device_info *devinfo = state->devinfo;
struct qinst *qinst = n->inst;
struct v3d_qpu_instr *inst = &qinst->qpu;
if (inst->type == V3D_QPU_INSTR_TYPE_BRANCH) {
if (inst->branch.cond != V3D_QPU_BRANCH_COND_ALWAYS)
add_read_dep(state, state->last_sf, n);
/* XXX: BDI */
/* XXX: BDU */
/* XXX: ub */
/* XXX: raddr_a */
add_write_dep(state, &state->last_unif, n);
return;
}
assert(inst->type == V3D_QPU_INSTR_TYPE_ALU);
/* XXX: LOAD_IMM */
if (v3d_qpu_add_op_num_src(inst->alu.add.op) > 0)
process_mux_deps(state, n, inst->alu.add.a);
if (v3d_qpu_add_op_num_src(inst->alu.add.op) > 1)
process_mux_deps(state, n, inst->alu.add.b);
if (v3d_qpu_mul_op_num_src(inst->alu.mul.op) > 0)
process_mux_deps(state, n, inst->alu.mul.a);
if (v3d_qpu_mul_op_num_src(inst->alu.mul.op) > 1)
process_mux_deps(state, n, inst->alu.mul.b);
switch (inst->alu.add.op) {
case V3D_QPU_A_VPMSETUP:
/* Could distinguish read/write by unpacking the uniform. */
add_write_dep(state, &state->last_vpm, n);
add_write_dep(state, &state->last_vpm_read, n);
break;
case V3D_QPU_A_STVPMV:
case V3D_QPU_A_STVPMD:
case V3D_QPU_A_STVPMP:
add_write_dep(state, &state->last_vpm, n);
break;
case V3D_QPU_A_VPMWT:
add_read_dep(state, state->last_vpm, n);
break;
case V3D_QPU_A_MSF:
add_read_dep(state, state->last_tlb, n);
break;
case V3D_QPU_A_SETMSF:
case V3D_QPU_A_SETREVF:
add_write_dep(state, &state->last_tlb, n);
break;
case V3D_QPU_A_FLAPUSH:
case V3D_QPU_A_FLBPUSH:
case V3D_QPU_A_VFLA:
case V3D_QPU_A_VFLNA:
case V3D_QPU_A_VFLB:
case V3D_QPU_A_VFLNB:
add_read_dep(state, state->last_sf, n);
break;
case V3D_QPU_A_FLPOP:
add_write_dep(state, &state->last_sf, n);
break;
default:
break;
}
switch (inst->alu.mul.op) {
case V3D_QPU_M_MULTOP:
case V3D_QPU_M_UMUL24:
/* MULTOP sets rtop, and UMUL24 implicitly reads rtop and
* resets it to 0. We could possibly reorder umul24s relative
* to each other, but for now just keep all the MUL parts in
* order.
*/
add_write_dep(state, &state->last_rtop, n);
break;
default:
break;
}
if (inst->alu.add.op != V3D_QPU_A_NOP) {
process_waddr_deps(state, n, inst->alu.add.waddr,
inst->alu.add.magic_write);
}
if (inst->alu.mul.op != V3D_QPU_M_NOP) {
process_waddr_deps(state, n, inst->alu.mul.waddr,
inst->alu.mul.magic_write);
}
if (v3d_qpu_sig_writes_address(devinfo, &inst->sig)) {
process_waddr_deps(state, n, inst->sig_addr,
inst->sig_magic);
}
if (v3d_qpu_writes_r3(devinfo, inst))
add_write_dep(state, &state->last_r[3], n);
if (v3d_qpu_writes_r4(devinfo, inst))
add_write_dep(state, &state->last_r[4], n);
if (v3d_qpu_writes_r5(devinfo, inst))
add_write_dep(state, &state->last_r[5], n);
if (inst->sig.thrsw) {
/* All accumulator contents and flags are undefined after the
* switch.
*/
for (int i = 0; i < ARRAY_SIZE(state->last_r); i++)
add_write_dep(state, &state->last_r[i], n);
add_write_dep(state, &state->last_sf, n);
/* Scoreboard-locking operations have to stay after the last
* thread switch.
*/
add_write_dep(state, &state->last_tlb, n);
add_write_dep(state, &state->last_tmu_write, n);
add_write_dep(state, &state->last_tmu_config, n);
}
if (v3d_qpu_waits_on_tmu(inst)) {
/* TMU loads are coming from a FIFO, so ordering is important.
*/
add_write_dep(state, &state->last_tmu_write, n);
}
if (inst->sig.wrtmuc)
add_write_dep(state, &state->last_tmu_config, n);
if (inst->sig.ldtlb | inst->sig.ldtlbu)
add_read_dep(state, state->last_tlb, n);
if (inst->sig.ldvpm)
add_write_dep(state, &state->last_vpm_read, n);
/* inst->sig.ldunif or sideband uniform read */
if (qinst->uniform != ~0)
add_write_dep(state, &state->last_unif, n);
process_cond_deps(state, n, inst->flags.ac);
process_cond_deps(state, n, inst->flags.mc);
process_pf_deps(state, n, inst->flags.apf);
process_pf_deps(state, n, inst->flags.mpf);
process_uf_deps(state, n, inst->flags.auf);
process_uf_deps(state, n, inst->flags.muf);
}
static void
calculate_forward_deps(struct v3d_compile *c, struct list_head *schedule_list)
{
struct schedule_state state;
memset(&state, 0, sizeof(state));
state.devinfo = c->devinfo;
state.dir = F;
list_for_each_entry(struct schedule_node, node, schedule_list, link)
calculate_deps(&state, node);
}
static void
calculate_reverse_deps(struct v3d_compile *c, struct list_head *schedule_list)
{
struct list_head *node;
struct schedule_state state;
memset(&state, 0, sizeof(state));
state.devinfo = c->devinfo;
state.dir = R;
for (node = schedule_list->prev; schedule_list != node; node = node->prev) {
calculate_deps(&state, (struct schedule_node *)node);
}
}
struct choose_scoreboard {
int tick;
int last_magic_sfu_write_tick;
int last_ldvary_tick;
int last_uniforms_reset_tick;
bool tlb_locked;
};
static bool
mux_reads_too_soon(struct choose_scoreboard *scoreboard,
const struct v3d_qpu_instr *inst, enum v3d_qpu_mux mux)
{
switch (mux) {
case V3D_QPU_MUX_R4:
if (scoreboard->tick - scoreboard->last_magic_sfu_write_tick <= 2)
return true;
break;
case V3D_QPU_MUX_R5:
if (scoreboard->tick - scoreboard->last_ldvary_tick <= 1)
return true;
break;
default:
break;
}
return false;
}
static bool
reads_too_soon_after_write(struct choose_scoreboard *scoreboard,
struct qinst *qinst)
{
const struct v3d_qpu_instr *inst = &qinst->qpu;
/* XXX: Branching off of raddr. */
if (inst->type == V3D_QPU_INSTR_TYPE_BRANCH)
return false;
assert(inst->type == V3D_QPU_INSTR_TYPE_ALU);
if (inst->alu.add.op != V3D_QPU_A_NOP) {
if (v3d_qpu_add_op_num_src(inst->alu.add.op) > 0 &&
mux_reads_too_soon(scoreboard, inst, inst->alu.add.a)) {
return true;
}
if (v3d_qpu_add_op_num_src(inst->alu.add.op) > 1 &&
mux_reads_too_soon(scoreboard, inst, inst->alu.add.b)) {
return true;
}
}
if (inst->alu.mul.op != V3D_QPU_M_NOP) {
if (v3d_qpu_mul_op_num_src(inst->alu.mul.op) > 0 &&
mux_reads_too_soon(scoreboard, inst, inst->alu.mul.a)) {
return true;
}
if (v3d_qpu_mul_op_num_src(inst->alu.mul.op) > 1 &&
mux_reads_too_soon(scoreboard, inst, inst->alu.mul.b)) {
return true;
}
}
/* XXX: imm */
return false;
}
static bool
writes_too_soon_after_write(const struct v3d_device_info *devinfo,
struct choose_scoreboard *scoreboard,
struct qinst *qinst)
{
const struct v3d_qpu_instr *inst = &qinst->qpu;
/* Don't schedule any other r4 write too soon after an SFU write.
* This would normally be prevented by dependency tracking, but might
* occur if a dead SFU computation makes it to scheduling.
*/
if (scoreboard->tick - scoreboard->last_magic_sfu_write_tick < 2 &&
v3d_qpu_writes_r4(devinfo, inst))
return true;
return false;
}
static bool
pixel_scoreboard_too_soon(struct choose_scoreboard *scoreboard,
const struct v3d_qpu_instr *inst)
{
return (scoreboard->tick == 0 && qpu_inst_is_tlb(inst));
}
static int
get_instruction_priority(const struct v3d_qpu_instr *inst)
{
uint32_t baseline_score;
uint32_t next_score = 0;
/* Schedule TLB operations as late as possible, to get more
* parallelism between shaders.
*/
if (qpu_inst_is_tlb(inst))
return next_score;
next_score++;
/* Schedule texture read results collection late to hide latency. */
if (v3d_qpu_waits_on_tmu(inst))
return next_score;
next_score++;
/* Default score for things that aren't otherwise special. */
baseline_score = next_score;
next_score++;
/* Schedule texture read setup early to hide their latency better. */
if (v3d_qpu_writes_tmu(inst))
return next_score;
next_score++;
return baseline_score;
}
static bool
qpu_magic_waddr_is_periph(enum v3d_qpu_waddr waddr)
{
return (v3d_qpu_magic_waddr_is_tmu(waddr) ||
v3d_qpu_magic_waddr_is_sfu(waddr) ||
v3d_qpu_magic_waddr_is_tlb(waddr) ||
v3d_qpu_magic_waddr_is_vpm(waddr) ||
v3d_qpu_magic_waddr_is_tsy(waddr));
}
static bool
qpu_accesses_peripheral(const struct v3d_qpu_instr *inst)
{
if (v3d_qpu_uses_vpm(inst))
return true;
if (v3d_qpu_uses_sfu(inst))
return true;
if (inst->type == V3D_QPU_INSTR_TYPE_ALU) {
if (inst->alu.add.op != V3D_QPU_A_NOP &&
inst->alu.add.magic_write &&
qpu_magic_waddr_is_periph(inst->alu.add.waddr)) {
return true;
}
if (inst->alu.add.op == V3D_QPU_A_TMUWT)
return true;
if (inst->alu.mul.op != V3D_QPU_M_NOP &&
inst->alu.mul.magic_write &&
qpu_magic_waddr_is_periph(inst->alu.mul.waddr)) {
return true;
}
}
return (inst->sig.ldvpm ||
inst->sig.ldtmu ||
inst->sig.ldtlb ||
inst->sig.ldtlbu ||
inst->sig.wrtmuc);
}
static bool
qpu_merge_inst(const struct v3d_device_info *devinfo,
struct v3d_qpu_instr *result,
const struct v3d_qpu_instr *a,
const struct v3d_qpu_instr *b)
{
if (a->type != V3D_QPU_INSTR_TYPE_ALU ||
b->type != V3D_QPU_INSTR_TYPE_ALU) {
return false;
}
/* Can't do more than one peripheral access in an instruction.
*
* XXX: V3D 4.1 allows TMU read along with a VPM read or write, and
* WRTMUC with a TMU magic register write (other than tmuc).
*/
if (qpu_accesses_peripheral(a) && qpu_accesses_peripheral(b))
return false;
struct v3d_qpu_instr merge = *a;
if (b->alu.add.op != V3D_QPU_A_NOP) {
if (a->alu.add.op != V3D_QPU_A_NOP)
return false;
merge.alu.add = b->alu.add;
merge.flags.ac = b->flags.ac;
merge.flags.apf = b->flags.apf;
merge.flags.auf = b->flags.auf;
}
if (b->alu.mul.op != V3D_QPU_M_NOP) {
if (a->alu.mul.op != V3D_QPU_M_NOP)
return false;
merge.alu.mul = b->alu.mul;
merge.flags.mc = b->flags.mc;
merge.flags.mpf = b->flags.mpf;
merge.flags.muf = b->flags.muf;
}
if (v3d_qpu_uses_mux(b, V3D_QPU_MUX_A)) {
if (v3d_qpu_uses_mux(a, V3D_QPU_MUX_A) &&
a->raddr_a != b->raddr_a) {
return false;
}
merge.raddr_a = b->raddr_a;
}
if (v3d_qpu_uses_mux(b, V3D_QPU_MUX_B)) {
if (v3d_qpu_uses_mux(a, V3D_QPU_MUX_B) &&
(a->raddr_b != b->raddr_b ||
a->sig.small_imm != b->sig.small_imm)) {
return false;
}
merge.raddr_b = b->raddr_b;
}
merge.sig.thrsw |= b->sig.thrsw;
merge.sig.ldunif |= b->sig.ldunif;
merge.sig.ldunifrf |= b->sig.ldunifrf;
merge.sig.ldunifa |= b->sig.ldunifa;
merge.sig.ldunifarf |= b->sig.ldunifarf;
merge.sig.ldtmu |= b->sig.ldtmu;
merge.sig.ldvary |= b->sig.ldvary;
merge.sig.ldvpm |= b->sig.ldvpm;
merge.sig.small_imm |= b->sig.small_imm;
merge.sig.ldtlb |= b->sig.ldtlb;
merge.sig.ldtlbu |= b->sig.ldtlbu;
merge.sig.ucb |= b->sig.ucb;
merge.sig.rotate |= b->sig.rotate;
merge.sig.wrtmuc |= b->sig.wrtmuc;
if (v3d_qpu_sig_writes_address(devinfo, &a->sig) &&
v3d_qpu_sig_writes_address(devinfo, &b->sig))
return false;
merge.sig_addr |= b->sig_addr;
merge.sig_magic |= b->sig_magic;
uint64_t packed;
bool ok = v3d_qpu_instr_pack(devinfo, &merge, &packed);
*result = merge;
/* No modifying the real instructions on failure. */
assert(ok || (a != result && b != result));
return ok;
}
static struct schedule_node *
choose_instruction_to_schedule(const struct v3d_device_info *devinfo,
struct choose_scoreboard *scoreboard,
struct list_head *schedule_list,
struct schedule_node *prev_inst)
{
struct schedule_node *chosen = NULL;
int chosen_prio = 0;
/* Don't pair up anything with a thread switch signal -- emit_thrsw()
* will handle pairing it along with filling the delay slots.
*/
if (prev_inst) {
if (prev_inst->inst->qpu.sig.thrsw)
return NULL;
}
list_for_each_entry(struct schedule_node, n, schedule_list, link) {
const struct v3d_qpu_instr *inst = &n->inst->qpu;
/* Don't choose the branch instruction until it's the last one
* left. We'll move it up to fit its delay slots after we
* choose it.
*/
if (inst->type == V3D_QPU_INSTR_TYPE_BRANCH &&
!list_is_singular(schedule_list)) {
continue;
}
/* "An instruction must not read from a location in physical
* regfile A or B that was written to by the previous
* instruction."
*/
if (reads_too_soon_after_write(scoreboard, n->inst))
continue;
if (writes_too_soon_after_write(devinfo, scoreboard, n->inst))
continue;
/* "A scoreboard wait must not occur in the first two
* instructions of a fragment shader. This is either the
* explicit Wait for Scoreboard signal or an implicit wait
* with the first tile-buffer read or write instruction."
*/
if (pixel_scoreboard_too_soon(scoreboard, inst))
continue;
/* ldunif and ldvary both write r5, but ldunif does so a tick
* sooner. If the ldvary's r5 wasn't used, then ldunif might
* otherwise get scheduled so ldunif and ldvary try to update
* r5 in the same tick.
*/
if ((inst->sig.ldunif || inst->sig.ldunifa) &&
scoreboard->tick == scoreboard->last_ldvary_tick + 1) {
continue;
}
/* If we're trying to pair with another instruction, check
* that they're compatible.
*/
if (prev_inst) {
/* Don't pair up a thread switch signal -- we'll
* handle pairing it when we pick it on its own.
*/
if (inst->sig.thrsw)
continue;
if (prev_inst->inst->uniform != -1 &&
n->inst->uniform != -1)
continue;
/* Don't merge in something that will lock the TLB.
* Hopwefully what we have in inst will release some
* other instructions, allowing us to delay the
* TLB-locking instruction until later.
*/
if (!scoreboard->tlb_locked && qpu_inst_is_tlb(inst))
continue;
struct v3d_qpu_instr merged_inst;
if (!qpu_merge_inst(devinfo, &merged_inst,
&prev_inst->inst->qpu, inst)) {
continue;
}
}
int prio = get_instruction_priority(inst);
/* Found a valid instruction. If nothing better comes along,
* this one works.
*/
if (!chosen) {
chosen = n;
chosen_prio = prio;
continue;
}
if (prio > chosen_prio) {
chosen = n;
chosen_prio = prio;
} else if (prio < chosen_prio) {
continue;
}
if (n->delay > chosen->delay) {
chosen = n;
chosen_prio = prio;
} else if (n->delay < chosen->delay) {
continue;
}
}
return chosen;
}
static void
update_scoreboard_for_magic_waddr(struct choose_scoreboard *scoreboard,
enum v3d_qpu_waddr waddr)
{
if (v3d_qpu_magic_waddr_is_sfu(waddr))
scoreboard->last_magic_sfu_write_tick = scoreboard->tick;
}
static void
update_scoreboard_for_chosen(struct choose_scoreboard *scoreboard,
const struct v3d_qpu_instr *inst)
{
if (inst->type == V3D_QPU_INSTR_TYPE_BRANCH)
return;
assert(inst->type == V3D_QPU_INSTR_TYPE_ALU);
if (inst->alu.add.op != V3D_QPU_A_NOP) {
if (inst->alu.add.magic_write) {
update_scoreboard_for_magic_waddr(scoreboard,
inst->alu.add.waddr);
}
}
if (inst->alu.mul.op != V3D_QPU_M_NOP) {
if (inst->alu.mul.magic_write) {
update_scoreboard_for_magic_waddr(scoreboard,
inst->alu.mul.waddr);
}
}
if (inst->sig.ldvary)
scoreboard->last_ldvary_tick = scoreboard->tick;
if (qpu_inst_is_tlb(inst))
scoreboard->tlb_locked = true;
}
static void
dump_state(const struct v3d_device_info *devinfo,
struct list_head *schedule_list)
{
list_for_each_entry(struct schedule_node, n, schedule_list, link) {
fprintf(stderr, " t=%4d: ", n->unblocked_time);
v3d_qpu_dump(devinfo, &n->inst->qpu);
fprintf(stderr, "\n");
for (int i = 0; i < n->child_count; i++) {
struct schedule_node *child = n->children[i].node;
if (!child)
continue;
fprintf(stderr, " - ");
v3d_qpu_dump(devinfo, &child->inst->qpu);
fprintf(stderr, " (%d parents, %c)\n",
child->parent_count,
n->children[i].write_after_read ? 'w' : 'r');
}
}
}
static uint32_t magic_waddr_latency(enum v3d_qpu_waddr waddr,
const struct v3d_qpu_instr *after)
{
/* Apply some huge latency between texture fetch requests and getting
* their results back.
*
* FIXME: This is actually pretty bogus. If we do:
*
* mov tmu0_s, a
* <a bit of math>
* mov tmu0_s, b
* load_tmu0
* <more math>
* load_tmu0
*
* we count that as worse than
*
* mov tmu0_s, a
* mov tmu0_s, b
* <lots of math>
* load_tmu0
* <more math>
* load_tmu0
*
* because we associate the first load_tmu0 with the *second* tmu0_s.
*/
if (v3d_qpu_magic_waddr_is_tmu(waddr) && v3d_qpu_waits_on_tmu(after))
return 100;
/* Assume that anything depending on us is consuming the SFU result. */
if (v3d_qpu_magic_waddr_is_sfu(waddr))
return 3;
return 1;
}
static uint32_t
instruction_latency(struct schedule_node *before, struct schedule_node *after)
{
const struct v3d_qpu_instr *before_inst = &before->inst->qpu;
const struct v3d_qpu_instr *after_inst = &after->inst->qpu;
uint32_t latency = 1;
if (before_inst->type != V3D_QPU_INSTR_TYPE_ALU ||
after_inst->type != V3D_QPU_INSTR_TYPE_ALU)
return latency;
if (before_inst->alu.add.magic_write) {
latency = MAX2(latency,
magic_waddr_latency(before_inst->alu.add.waddr,
after_inst));
}
if (before_inst->alu.mul.magic_write) {
latency = MAX2(latency,
magic_waddr_latency(before_inst->alu.mul.waddr,
after_inst));
}
return latency;
}
/** Recursive computation of the delay member of a node. */
static void
compute_delay(struct schedule_node *n)
{
if (!n->child_count) {
n->delay = 1;
} else {
for (int i = 0; i < n->child_count; i++) {
if (!n->children[i].node->delay)
compute_delay(n->children[i].node);
n->delay = MAX2(n->delay,
n->children[i].node->delay +
instruction_latency(n, n->children[i].node));
}
}
}
static void
mark_instruction_scheduled(struct list_head *schedule_list,
uint32_t time,
struct schedule_node *node,
bool war_only)
{
if (!node)
return;
for (int i = node->child_count - 1; i >= 0; i--) {
struct schedule_node *child =
node->children[i].node;
if (!child)
continue;
if (war_only && !node->children[i].write_after_read)
continue;
/* If the requirement is only that the node not appear before
* the last read of its destination, then it can be scheduled
* immediately after (or paired with!) the thing reading the
* destination.
*/
uint32_t latency = 0;
if (!war_only) {
latency = instruction_latency(node,
node->children[i].node);
}
child->unblocked_time = MAX2(child->unblocked_time,
time + latency);
child->parent_count--;
if (child->parent_count == 0)
list_add(&child->link, schedule_list);
node->children[i].node = NULL;
}
}
static void
insert_scheduled_instruction(struct v3d_compile *c,
struct qblock *block,
struct choose_scoreboard *scoreboard,
struct qinst *inst)
{
list_addtail(&inst->link, &block->instructions);
update_scoreboard_for_chosen(scoreboard, &inst->qpu);
c->qpu_inst_count++;
scoreboard->tick++;
}
static struct qinst *
vir_nop()
{
struct qreg undef = { QFILE_NULL, 0 };
struct qinst *qinst = vir_add_inst(V3D_QPU_A_NOP, undef, undef, undef);
return qinst;
}
static void
emit_nop(struct v3d_compile *c, struct qblock *block,
struct choose_scoreboard *scoreboard)
{
insert_scheduled_instruction(c, block, scoreboard, vir_nop());
}
static bool
qpu_instruction_valid_in_thrend_slot(struct v3d_compile *c,
const struct qinst *qinst, int slot)
{
const struct v3d_qpu_instr *inst = &qinst->qpu;
/* Only TLB Z writes are prohibited in the last slot, but we don't
* have those flagged so prohibit all TLB ops for now.
*/
if (slot == 2 && qpu_inst_is_tlb(inst))
return false;
if (slot > 0 && qinst->uniform != ~0)
return false;
if (v3d_qpu_uses_vpm(inst))
return false;
if (inst->sig.ldvary)
return false;
if (inst->type == V3D_QPU_INSTR_TYPE_ALU) {
/* GFXH-1625: TMUWT not allowed in the final instruction. */
if (slot == 2 && inst->alu.add.op == V3D_QPU_A_TMUWT)
return false;
/* No writing physical registers at the end. */
if (!inst->alu.add.magic_write ||
!inst->alu.mul.magic_write) {
return false;
}
if (c->devinfo->ver < 40 && inst->alu.add.op == V3D_QPU_A_SETMSF)
return false;
/* RF0-2 might be overwritten during the delay slots by
* fragment shader setup.
*/
if (inst->raddr_a < 3 &&
(inst->alu.add.a == V3D_QPU_MUX_A ||
inst->alu.add.b == V3D_QPU_MUX_A ||
inst->alu.mul.a == V3D_QPU_MUX_A ||
inst->alu.mul.b == V3D_QPU_MUX_A)) {
return false;
}
if (inst->raddr_b < 3 &&
!inst->sig.small_imm &&
(inst->alu.add.a == V3D_QPU_MUX_B ||
inst->alu.add.b == V3D_QPU_MUX_B ||
inst->alu.mul.a == V3D_QPU_MUX_B ||
inst->alu.mul.b == V3D_QPU_MUX_B)) {
return false;
}
}
return true;
}
static bool
valid_thrsw_sequence(struct v3d_compile *c,
struct qinst *qinst, int instructions_in_sequence,
bool is_thrend)
{
for (int slot = 0; slot < instructions_in_sequence; slot++) {
/* No scheduling SFU when the result would land in the other
* thread. The simulator complains for safety, though it
* would only occur for dead code in our case.
*/
if (slot > 0 &&
qinst->qpu.type == V3D_QPU_INSTR_TYPE_ALU &&
(v3d_qpu_magic_waddr_is_sfu(qinst->qpu.alu.add.waddr) ||
v3d_qpu_magic_waddr_is_sfu(qinst->qpu.alu.mul.waddr))) {
return false;
}
if (slot > 0 && qinst->qpu.sig.ldvary)
return false;
if (is_thrend &&
!qpu_instruction_valid_in_thrend_slot(c, qinst, slot)) {
return false;
}
/* Note that the list is circular, so we can only do this up
* to instructions_in_sequence.
*/
qinst = (struct qinst *)qinst->link.next;
}
return true;
}
/**
* Emits a THRSW signal in the stream, trying to move it up to pair with
* another instruction.
*/
static int
emit_thrsw(struct v3d_compile *c,
struct qblock *block,
struct choose_scoreboard *scoreboard,
struct qinst *inst,
bool is_thrend)
{
int time = 0;
/* There should be nothing in a thrsw inst being scheduled other than
* the signal bits.
*/
assert(inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU);
assert(inst->qpu.alu.add.op == V3D_QPU_A_NOP);
assert(inst->qpu.alu.mul.op == V3D_QPU_M_NOP);
/* Find how far back into previous instructions we can put the THRSW. */
int slots_filled = 0;
struct qinst *merge_inst = NULL;
vir_for_each_inst_rev(prev_inst, block) {
struct v3d_qpu_sig sig = prev_inst->qpu.sig;
sig.thrsw = true;
uint32_t packed_sig;
if (!v3d_qpu_sig_pack(c->devinfo, &sig, &packed_sig))
break;
if (!valid_thrsw_sequence(c, prev_inst, slots_filled + 1,
is_thrend)) {
break;
}
merge_inst = prev_inst;
if (++slots_filled == 3)
break;
}
bool needs_free = false;
if (merge_inst) {
merge_inst->qpu.sig.thrsw = true;
needs_free = true;
} else {
insert_scheduled_instruction(c, block, scoreboard, inst);
time++;
slots_filled++;
merge_inst = inst;
}
/* Insert any extra delay slot NOPs we need. */
for (int i = 0; i < 3 - slots_filled; i++) {
emit_nop(c, block, scoreboard);
time++;
}
/* If we're emitting the last THRSW (other than program end), then
* signal that to the HW by emitting two THRSWs in a row.
*/
if (inst->is_last_thrsw) {
struct qinst *second_inst =
(struct qinst *)merge_inst->link.next;
second_inst->qpu.sig.thrsw = true;
}
/* If we put our THRSW into another instruction, free up the
* instruction that didn't end up scheduled into the list.
*/
if (needs_free)
free(inst);
return time;
}
static uint32_t
schedule_instructions(struct v3d_compile *c,
struct choose_scoreboard *scoreboard,
struct qblock *block,
struct list_head *schedule_list,
enum quniform_contents *orig_uniform_contents,
uint32_t *orig_uniform_data,
uint32_t *next_uniform)
{
const struct v3d_device_info *devinfo = c->devinfo;
uint32_t time = 0;
if (debug) {
fprintf(stderr, "initial deps:\n");
dump_state(devinfo, schedule_list);
fprintf(stderr, "\n");
}
/* Remove non-DAG heads from the list. */
list_for_each_entry_safe(struct schedule_node, n, schedule_list, link) {
if (n->parent_count != 0)
list_del(&n->link);
}
while (!list_empty(schedule_list)) {
struct schedule_node *chosen =
choose_instruction_to_schedule(devinfo,
scoreboard,
schedule_list,
NULL);
struct schedule_node *merge = NULL;
/* If there are no valid instructions to schedule, drop a NOP
* in.
*/
struct qinst *qinst = chosen ? chosen->inst : vir_nop();
struct v3d_qpu_instr *inst = &qinst->qpu;
if (debug) {
fprintf(stderr, "t=%4d: current list:\n",
time);
dump_state(devinfo, schedule_list);
fprintf(stderr, "t=%4d: chose: ", time);
v3d_qpu_dump(devinfo, inst);
fprintf(stderr, "\n");
}
/* We can't mark_instruction_scheduled() the chosen inst until
* we're done identifying instructions to merge, so put the
* merged instructions on a list for a moment.
*/
struct list_head merged_list;
list_inithead(&merged_list);
/* Schedule this instruction onto the QPU list. Also try to
* find an instruction to pair with it.
*/
if (chosen) {
time = MAX2(chosen->unblocked_time, time);
list_del(&chosen->link);
mark_instruction_scheduled(schedule_list, time,
chosen, true);
while ((merge =
choose_instruction_to_schedule(devinfo,
scoreboard,
schedule_list,
chosen))) {
time = MAX2(merge->unblocked_time, time);
list_del(&merge->link);
list_addtail(&merge->link, &merged_list);
(void)qpu_merge_inst(devinfo, inst,
inst, &merge->inst->qpu);
if (merge->inst->uniform != -1) {
chosen->inst->uniform =
merge->inst->uniform;
}
if (debug) {
fprintf(stderr, "t=%4d: merging: ",
time);
v3d_qpu_dump(devinfo, &merge->inst->qpu);
fprintf(stderr, "\n");
fprintf(stderr, " result: ");
v3d_qpu_dump(devinfo, inst);
fprintf(stderr, "\n");
}
}
}
/* Update the uniform index for the rewritten location --
* branch target updating will still need to change
* c->uniform_data[] using this index.
*/
if (qinst->uniform != -1) {
if (inst->type == V3D_QPU_INSTR_TYPE_BRANCH)
block->branch_uniform = *next_uniform;
c->uniform_data[*next_uniform] =
orig_uniform_data[qinst->uniform];
c->uniform_contents[*next_uniform] =
orig_uniform_contents[qinst->uniform];
qinst->uniform = *next_uniform;
(*next_uniform)++;
}
if (debug) {
fprintf(stderr, "\n");
}
/* Now that we've scheduled a new instruction, some of its
* children can be promoted to the list of instructions ready to
* be scheduled. Update the children's unblocked time for this
* DAG edge as we do so.
*/
mark_instruction_scheduled(schedule_list, time, chosen, false);
list_for_each_entry(struct schedule_node, merge, &merged_list,
link) {
mark_instruction_scheduled(schedule_list, time, merge,
false);
/* The merged VIR instruction doesn't get re-added to the
* block, so free it now.
*/
free(merge->inst);
}
if (inst->sig.thrsw) {
time += emit_thrsw(c, block, scoreboard, qinst, false);
} else {
insert_scheduled_instruction(c, block,
scoreboard, qinst);
if (inst->type == V3D_QPU_INSTR_TYPE_BRANCH) {
block->branch_qpu_ip = c->qpu_inst_count - 1;
/* Fill the delay slots.
*
* We should fill these with actual instructions,
* instead, but that will probably need to be done
* after this, once we know what the leading
* instructions of the successors are (so we can
* handle A/B register file write latency)
*/
for (int i = 0; i < 3; i++)
emit_nop(c, block, scoreboard);
}
}
}
return time;
}
static uint32_t
qpu_schedule_instructions_block(struct v3d_compile *c,
struct choose_scoreboard *scoreboard,
struct qblock *block,
enum quniform_contents *orig_uniform_contents,
uint32_t *orig_uniform_data,
uint32_t *next_uniform)
{
void *mem_ctx = ralloc_context(NULL);
struct list_head schedule_list;
list_inithead(&schedule_list);
/* Wrap each instruction in a scheduler structure. */
while (!list_empty(&block->instructions)) {
struct qinst *qinst = (struct qinst *)block->instructions.next;
struct schedule_node *n =
rzalloc(mem_ctx, struct schedule_node);
n->inst = qinst;
list_del(&qinst->link);
list_addtail(&n->link, &schedule_list);
}
calculate_forward_deps(c, &schedule_list);
calculate_reverse_deps(c, &schedule_list);
list_for_each_entry(struct schedule_node, n, &schedule_list, link) {
compute_delay(n);
}
uint32_t cycles = schedule_instructions(c, scoreboard, block,
&schedule_list,
orig_uniform_contents,
orig_uniform_data,
next_uniform);
ralloc_free(mem_ctx);
return cycles;
}
static void
qpu_set_branch_targets(struct v3d_compile *c)
{
vir_for_each_block(block, c) {
/* The end block of the program has no branch. */
if (!block->successors[0])
continue;
/* If there was no branch instruction, then the successor
* block must follow immediately after this one.
*/
if (block->branch_qpu_ip == ~0) {
assert(block->end_qpu_ip + 1 ==
block->successors[0]->start_qpu_ip);
continue;
}
/* Walk back through the delay slots to find the branch
* instr.
*/
struct list_head *entry = block->instructions.prev;
for (int i = 0; i < 3; i++)
entry = entry->prev;
struct qinst *branch = container_of(entry, branch, link);
assert(branch->qpu.type == V3D_QPU_INSTR_TYPE_BRANCH);
/* Make sure that the if-we-don't-jump
* successor was scheduled just after the
* delay slots.
*/
assert(!block->successors[1] ||
block->successors[1]->start_qpu_ip ==
block->branch_qpu_ip + 4);
branch->qpu.branch.offset =
((block->successors[0]->start_qpu_ip -
(block->branch_qpu_ip + 4)) *
sizeof(uint64_t));
/* Set up the relative offset to jump in the
* uniform stream.
*
* Use a temporary here, because
* uniform_data[inst->uniform] may be shared
* between multiple instructions.
*/
assert(c->uniform_contents[branch->uniform] == QUNIFORM_CONSTANT);
c->uniform_data[branch->uniform] =
(block->successors[0]->start_uniform -
(block->branch_uniform + 1)) * 4;
}
}
uint32_t
v3d_qpu_schedule_instructions(struct v3d_compile *c)
{
const struct v3d_device_info *devinfo = c->devinfo;
struct qblock *end_block = list_last_entry(&c->blocks,
struct qblock, link);
/* We reorder the uniforms as we schedule instructions, so save the
* old data off and replace it.
*/
uint32_t *uniform_data = c->uniform_data;
enum quniform_contents *uniform_contents = c->uniform_contents;
c->uniform_contents = ralloc_array(c, enum quniform_contents,
c->num_uniforms);
c->uniform_data = ralloc_array(c, uint32_t, c->num_uniforms);
c->uniform_array_size = c->num_uniforms;
uint32_t next_uniform = 0;
struct choose_scoreboard scoreboard;
memset(&scoreboard, 0, sizeof(scoreboard));
scoreboard.last_ldvary_tick = -10;
scoreboard.last_magic_sfu_write_tick = -10;
scoreboard.last_uniforms_reset_tick = -10;
if (debug) {
fprintf(stderr, "Pre-schedule instructions\n");
vir_for_each_block(block, c) {
fprintf(stderr, "BLOCK %d\n", block->index);
list_for_each_entry(struct qinst, qinst,
&block->instructions, link) {
v3d_qpu_dump(devinfo, &qinst->qpu);
fprintf(stderr, "\n");
}
}
fprintf(stderr, "\n");
}
uint32_t cycles = 0;
vir_for_each_block(block, c) {
block->start_qpu_ip = c->qpu_inst_count;
block->branch_qpu_ip = ~0;
block->start_uniform = next_uniform;
cycles += qpu_schedule_instructions_block(c,
&scoreboard,
block,
uniform_contents,
uniform_data,
&next_uniform);
block->end_qpu_ip = c->qpu_inst_count - 1;
}
/* Emit the program-end THRSW instruction. */;
struct qinst *thrsw = vir_nop();
thrsw->qpu.sig.thrsw = true;
emit_thrsw(c, end_block, &scoreboard, thrsw, true);
qpu_set_branch_targets(c);
assert(next_uniform == c->num_uniforms);
return cycles;
}
|