summaryrefslogtreecommitdiff
path: root/src/gallium/drivers/vc4/vc4_qpu_schedule.c
blob: 5f01f8c7093a58c0a4644580cfe493f03266be75 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
/*
 * Copyright © 2010 Intel Corporation
 * Copyright © 2014 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 vc4_qpu_schedule.c
 *
 * 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 "vc4_qir.h"
#include "vc4_qpu.h"
#include "util/ralloc.h"
#include "util/dag.h"

static bool debug;

struct schedule_node_child;

struct schedule_node {
        struct dag_node dag;
        struct list_head link;
        struct queued_qpu_inst *inst;

        /* 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;

        /**
         * Which uniform from uniform_data[] this instruction read, or -1 if
         * not reading a uniform.
         */
        int uniform;
};

/* When walking the instructions in reverse, we need to swap before/after in
 * add_dep().
 */
enum direction { F, R };

struct schedule_state {
        struct dag *dag;
        struct schedule_node *last_r[6];
        struct schedule_node *last_ra[32];
        struct schedule_node *last_rb[32];
        struct schedule_node *last_sf;
        struct schedule_node *last_vpm_read;
        struct schedule_node *last_tmu_write;
        struct schedule_node *last_tlb;
        struct schedule_node *last_vpm;
        struct schedule_node *last_uniforms_reset;
        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;
        void *edge_data = (void *)(uintptr_t)write_after_read;

        if (!before || !after)
                return;

        assert(before != after);

        if (state->dir == F)
                dag_add_edge(&before->dag, &after->dag, edge_data);
        else
                dag_add_edge(&after->dag, &before->dag, edge_data);
}

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_writes_r4(uint64_t inst)
{
        uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);

        switch(sig) {
        case QPU_SIG_COLOR_LOAD:
        case QPU_SIG_LOAD_TMU0:
        case QPU_SIG_LOAD_TMU1:
        case QPU_SIG_ALPHA_MASK_LOAD:
                return true;
        default:
                return false;
        }
}

static void
process_raddr_deps(struct schedule_state *state, struct schedule_node *n,
                   uint32_t raddr, bool is_a)
{
        switch (raddr) {
        case QPU_R_VARY:
                add_write_dep(state, &state->last_r[5], n);
                break;

        case QPU_R_VPM:
                add_write_dep(state, &state->last_vpm_read, n);
                break;

        case QPU_R_UNIF:
                add_read_dep(state, state->last_uniforms_reset, n);
                break;

        case QPU_R_NOP:
        case QPU_R_ELEM_QPU:
        case QPU_R_XY_PIXEL_COORD:
        case QPU_R_MS_REV_FLAGS:
                break;

        default:
                if (raddr < 32) {
                        if (is_a)
                                add_read_dep(state, state->last_ra[raddr], n);
                        else
                                add_read_dep(state, state->last_rb[raddr], n);
                } else {
                        fprintf(stderr, "unknown raddr %d\n", raddr);
                        abort();
                }
                break;
        }
}

static bool
is_tmu_write(uint32_t waddr)
{
        switch (waddr) {
        case QPU_W_TMU0_S:
        case QPU_W_TMU0_T:
        case QPU_W_TMU0_R:
        case QPU_W_TMU0_B:
        case QPU_W_TMU1_S:
        case QPU_W_TMU1_T:
        case QPU_W_TMU1_R:
        case QPU_W_TMU1_B:
                return true;
        default:
                return false;
        }
}

static bool
reads_uniform(uint64_t inst)
{
        if (QPU_GET_FIELD(inst, QPU_SIG) == QPU_SIG_LOAD_IMM)
                return false;

        return (QPU_GET_FIELD(inst, QPU_RADDR_A) == QPU_R_UNIF ||
                (QPU_GET_FIELD(inst, QPU_RADDR_B) == QPU_R_UNIF &&
                 QPU_GET_FIELD(inst, QPU_SIG) != QPU_SIG_SMALL_IMM) ||
                is_tmu_write(QPU_GET_FIELD(inst, QPU_WADDR_ADD)) ||
                is_tmu_write(QPU_GET_FIELD(inst, QPU_WADDR_MUL)));
}

static void
process_mux_deps(struct schedule_state *state, struct schedule_node *n,
                 uint32_t mux)
{
        if (mux != QPU_MUX_A && mux != QPU_MUX_B)
                add_read_dep(state, state->last_r[mux], n);
}


static void
process_waddr_deps(struct schedule_state *state, struct schedule_node *n,
                   uint32_t waddr, bool is_add)
{
        uint64_t inst = n->inst->inst;
        bool is_a = is_add ^ ((inst & QPU_WS) != 0);

        if (waddr < 32) {
                if (is_a) {
                        add_write_dep(state, &state->last_ra[waddr], n);
                } else {
                        add_write_dep(state, &state->last_rb[waddr], n);
                }
        } else if (is_tmu_write(waddr)) {
                add_write_dep(state, &state->last_tmu_write, n);
                add_read_dep(state, state->last_uniforms_reset, n);
        } else if (qpu_waddr_is_tlb(waddr) ||
                   waddr == QPU_W_MS_FLAGS) {
                add_write_dep(state, &state->last_tlb, n);
        } else {
                switch (waddr) {
                case QPU_W_ACC0:
                case QPU_W_ACC1:
                case QPU_W_ACC2:
                case QPU_W_ACC3:
                case QPU_W_ACC5:
                        add_write_dep(state, &state->last_r[waddr - QPU_W_ACC0],
                                      n);
                        break;

                case QPU_W_VPM:
                        add_write_dep(state, &state->last_vpm, n);
                        break;

                case QPU_W_VPMVCD_SETUP:
                        if (is_a)
                                add_write_dep(state, &state->last_vpm_read, n);
                        else
                                add_write_dep(state, &state->last_vpm, n);
                        break;

                case QPU_W_SFU_RECIP:
                case QPU_W_SFU_RECIPSQRT:
                case QPU_W_SFU_EXP:
                case QPU_W_SFU_LOG:
                        add_write_dep(state, &state->last_r[4], n);
                        break;

                case QPU_W_TLB_STENCIL_SETUP:
                        /* This isn't a TLB operation that does things like
                         * implicitly lock the scoreboard, but it does have to
                         * appear before TLB_Z, and each of the TLB_STENCILs
                         * have to schedule in the same order relative to each
                         * other.
                         */
                        add_write_dep(state, &state->last_tlb, n);
                        break;

                case QPU_W_MS_FLAGS:
                        add_write_dep(state, &state->last_tlb, n);
                        break;

                case QPU_W_UNIFORMS_ADDRESS:
                        add_write_dep(state, &state->last_uniforms_reset, n);
                        break;

                case QPU_W_NOP:
                        break;

                default:
                        fprintf(stderr, "Unknown waddr %d\n", waddr);
                        abort();
                }
        }
}

static void
process_cond_deps(struct schedule_state *state, struct schedule_node *n,
                  uint32_t cond)
{
        switch (cond) {
        case QPU_COND_NEVER:
        case QPU_COND_ALWAYS:
                break;
        default:
                add_read_dep(state, state->last_sf, n);
                break;
        }
}

/**
 * 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)
{
        uint64_t inst = n->inst->inst;
        uint32_t add_op = QPU_GET_FIELD(inst, QPU_OP_ADD);
        uint32_t mul_op = QPU_GET_FIELD(inst, QPU_OP_MUL);
        uint32_t waddr_add = QPU_GET_FIELD(inst, QPU_WADDR_ADD);
        uint32_t waddr_mul = QPU_GET_FIELD(inst, QPU_WADDR_MUL);
        uint32_t raddr_a = QPU_GET_FIELD(inst, QPU_RADDR_A);
        uint32_t raddr_b = QPU_GET_FIELD(inst, QPU_RADDR_B);
        uint32_t add_a = QPU_GET_FIELD(inst, QPU_ADD_A);
        uint32_t add_b = QPU_GET_FIELD(inst, QPU_ADD_B);
        uint32_t mul_a = QPU_GET_FIELD(inst, QPU_MUL_A);
        uint32_t mul_b = QPU_GET_FIELD(inst, QPU_MUL_B);
        uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);

        if (sig != QPU_SIG_LOAD_IMM) {
                process_raddr_deps(state, n, raddr_a, true);
                if (sig != QPU_SIG_SMALL_IMM &&
                    sig != QPU_SIG_BRANCH)
                        process_raddr_deps(state, n, raddr_b, false);
        }

        if (add_op != QPU_A_NOP) {
                process_mux_deps(state, n, add_a);
                process_mux_deps(state, n, add_b);
        }
        if (mul_op != QPU_M_NOP) {
                process_mux_deps(state, n, mul_a);
                process_mux_deps(state, n, mul_b);
        }

        process_waddr_deps(state, n, waddr_add, true);
        process_waddr_deps(state, n, waddr_mul, false);
        if (qpu_writes_r4(inst))
                add_write_dep(state, &state->last_r[4], n);

        switch (sig) {
        case QPU_SIG_SW_BREAKPOINT:
        case QPU_SIG_NONE:
        case QPU_SIG_SMALL_IMM:
        case QPU_SIG_LOAD_IMM:
                break;

        case QPU_SIG_THREAD_SWITCH:
        case QPU_SIG_LAST_THREAD_SWITCH:
                /* 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);
                break;

        case QPU_SIG_LOAD_TMU0:
        case QPU_SIG_LOAD_TMU1:
                /* TMU loads are coming from a FIFO, so ordering is important.
                 */
                add_write_dep(state, &state->last_tmu_write, n);
                break;

        case QPU_SIG_COLOR_LOAD:
                add_read_dep(state, state->last_tlb, n);
                break;

        case QPU_SIG_BRANCH:
                add_read_dep(state, state->last_sf, n);
                break;

        case QPU_SIG_PROG_END:
        case QPU_SIG_WAIT_FOR_SCOREBOARD:
        case QPU_SIG_SCOREBOARD_UNLOCK:
        case QPU_SIG_COVERAGE_LOAD:
        case QPU_SIG_COLOR_LOAD_END:
        case QPU_SIG_ALPHA_MASK_LOAD:
                fprintf(stderr, "Unhandled signal bits %d\n", sig);
                abort();
        }

        process_cond_deps(state, n, QPU_GET_FIELD(inst, QPU_COND_ADD));
        process_cond_deps(state, n, QPU_GET_FIELD(inst, QPU_COND_MUL));
        if ((inst & QPU_SF) && sig != QPU_SIG_BRANCH)
                add_write_dep(state, &state->last_sf, n);
}

static void
calculate_forward_deps(struct vc4_compile *c, struct dag *dag,
                       struct list_head *schedule_list)
{
        struct schedule_state state;

        memset(&state, 0, sizeof(state));
        state.dag = dag;
        state.dir = F;

        list_for_each_entry(struct schedule_node, node, schedule_list, link)
                calculate_deps(&state, node);
}

static void
calculate_reverse_deps(struct vc4_compile *c, struct dag *dag,
                       struct list_head *schedule_list)
{
        struct schedule_state state;

        memset(&state, 0, sizeof(state));
        state.dag = dag;
        state.dir = R;

        list_for_each_entry_rev(struct schedule_node, node, schedule_list,
                                link) {
                calculate_deps(&state, (struct schedule_node *)node);
        }
}

struct choose_scoreboard {
        struct dag *dag;
        int tick;
        int last_sfu_write_tick;
        int last_uniforms_reset_tick;
        uint32_t last_waddr_a, last_waddr_b;
        bool tlb_locked;
};

static bool
reads_too_soon_after_write(struct choose_scoreboard *scoreboard, uint64_t inst)
{
        uint32_t raddr_a = QPU_GET_FIELD(inst, QPU_RADDR_A);
        uint32_t raddr_b = QPU_GET_FIELD(inst, QPU_RADDR_B);
        uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);

        /* Full immediate loads don't read any registers. */
        if (sig == QPU_SIG_LOAD_IMM)
                return false;

        uint32_t src_muxes[] = {
                QPU_GET_FIELD(inst, QPU_ADD_A),
                QPU_GET_FIELD(inst, QPU_ADD_B),
                QPU_GET_FIELD(inst, QPU_MUL_A),
                QPU_GET_FIELD(inst, QPU_MUL_B),
        };
        for (int i = 0; i < ARRAY_SIZE(src_muxes); i++) {
                if ((src_muxes[i] == QPU_MUX_A &&
                     raddr_a < 32 &&
                     scoreboard->last_waddr_a == raddr_a) ||
                    (src_muxes[i] == QPU_MUX_B &&
                     sig != QPU_SIG_SMALL_IMM &&
                     raddr_b < 32 &&
                     scoreboard->last_waddr_b == raddr_b)) {
                        return true;
                }

                if (src_muxes[i] == QPU_MUX_R4) {
                        if (scoreboard->tick -
                            scoreboard->last_sfu_write_tick <= 2) {
                                return true;
                        }
                }
        }

        if (sig == QPU_SIG_SMALL_IMM &&
            QPU_GET_FIELD(inst, QPU_SMALL_IMM) >= QPU_SMALL_IMM_MUL_ROT) {
                uint32_t mux_a = QPU_GET_FIELD(inst, QPU_MUL_A);
                uint32_t mux_b = QPU_GET_FIELD(inst, QPU_MUL_B);

                if (scoreboard->last_waddr_a == mux_a + QPU_W_ACC0 ||
                    scoreboard->last_waddr_a == mux_b + QPU_W_ACC0 ||
                    scoreboard->last_waddr_b == mux_a + QPU_W_ACC0 ||
                    scoreboard->last_waddr_b == mux_b + QPU_W_ACC0) {
                        return true;
                }
        }

        if (reads_uniform(inst) &&
            scoreboard->tick - scoreboard->last_uniforms_reset_tick <= 2) {
                return true;
        }

        return false;
}

static bool
pixel_scoreboard_too_soon(struct choose_scoreboard *scoreboard, uint64_t inst)
{
        return (scoreboard->tick < 2 && qpu_inst_is_tlb(inst));
}

static int
get_instruction_priority(uint64_t inst)
{
        uint32_t waddr_add = QPU_GET_FIELD(inst, QPU_WADDR_ADD);
        uint32_t waddr_mul = QPU_GET_FIELD(inst, QPU_WADDR_MUL);
        uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
        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 (sig == QPU_SIG_LOAD_TMU0 || sig == QPU_SIG_LOAD_TMU1)
                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 (is_tmu_write(waddr_add) || is_tmu_write(waddr_mul))
                return next_score;
        next_score++;

        return baseline_score;
}

static struct schedule_node *
choose_instruction_to_schedule(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) {
                uint32_t prev_sig = QPU_GET_FIELD(prev_inst->inst->inst,
                                                  QPU_SIG);
                if (prev_sig == QPU_SIG_THREAD_SWITCH ||
                    prev_sig == QPU_SIG_LAST_THREAD_SWITCH) {
                        return NULL;
                }
        }

        list_for_each_entry(struct schedule_node, n, &scoreboard->dag->heads,
                            dag.link) {
                uint64_t inst = n->inst->inst;
                uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);

                /* Don't choose the branch instruction until it's the last one
                 * left.  XXX: We could potentially choose it before it's the
                 * last one, if the remaining instructions fit in the delay
                 * slots.
                 */
                if (sig == QPU_SIG_BRANCH &&
                    !list_is_singular(&scoreboard->dag->heads)) {
                        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, 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;

                /* 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 (sig == QPU_SIG_THREAD_SWITCH ||
                            sig == QPU_SIG_LAST_THREAD_SWITCH) {
                                continue;
                        }

                        if (prev_inst->uniform != -1 && n->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;

                        inst = qpu_merge_inst(prev_inst->inst->inst, inst);
                        if (!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_chosen(struct choose_scoreboard *scoreboard,
                             uint64_t inst)
{
        uint32_t waddr_add = QPU_GET_FIELD(inst, QPU_WADDR_ADD);
        uint32_t waddr_mul = QPU_GET_FIELD(inst, QPU_WADDR_MUL);

        if (!(inst & QPU_WS)) {
                scoreboard->last_waddr_a = waddr_add;
                scoreboard->last_waddr_b = waddr_mul;
        } else {
                scoreboard->last_waddr_b = waddr_add;
                scoreboard->last_waddr_a = waddr_mul;
        }

        if ((waddr_add >= QPU_W_SFU_RECIP && waddr_add <= QPU_W_SFU_LOG) ||
            (waddr_mul >= QPU_W_SFU_RECIP && waddr_mul <= QPU_W_SFU_LOG)) {
                scoreboard->last_sfu_write_tick = scoreboard->tick;
        }

        if (waddr_add == QPU_W_UNIFORMS_ADDRESS ||
            waddr_mul == QPU_W_UNIFORMS_ADDRESS) {
                scoreboard->last_uniforms_reset_tick = scoreboard->tick;
        }

        if (qpu_inst_is_tlb(inst))
                scoreboard->tlb_locked = true;
}

static void
dump_state(struct dag *dag)
{
        list_for_each_entry(struct schedule_node, n, &dag->heads, dag.link) {
                fprintf(stderr, "         t=%4d: ", n->unblocked_time);
                vc4_qpu_disasm(&n->inst->inst, 1);
                fprintf(stderr, "\n");

                util_dynarray_foreach(&n->dag.edges, struct dag_edge, edge) {
                        struct schedule_node *child =
                                (struct schedule_node *)edge->child;
                        if (!child)
                                continue;

                        fprintf(stderr, "                 - ");
                        vc4_qpu_disasm(&child->inst->inst, 1);
                        fprintf(stderr, " (%d parents, %c)\n",
                                child->dag.parent_count,
                                edge->data ? 'w' : 'r');
                }
        }
}

static uint32_t waddr_latency(uint32_t waddr, uint64_t after)
{
        if (waddr < 32)
                return 2;

        /* 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 (waddr == QPU_W_TMU0_S) {
                if (QPU_GET_FIELD(after, QPU_SIG) == QPU_SIG_LOAD_TMU0)
                        return 100;
        }
        if (waddr == QPU_W_TMU1_S) {
                if (QPU_GET_FIELD(after, QPU_SIG) == QPU_SIG_LOAD_TMU1)
                        return 100;
        }

        switch(waddr) {
        case QPU_W_SFU_RECIP:
        case QPU_W_SFU_RECIPSQRT:
        case QPU_W_SFU_EXP:
        case QPU_W_SFU_LOG:
                return 3;
        default:
                return 1;
        }
}

static uint32_t
instruction_latency(struct schedule_node *before, struct schedule_node *after)
{
        uint64_t before_inst = before->inst->inst;
        uint64_t after_inst = after->inst->inst;

        return MAX2(waddr_latency(QPU_GET_FIELD(before_inst, QPU_WADDR_ADD),
                                  after_inst),
                    waddr_latency(QPU_GET_FIELD(before_inst, QPU_WADDR_MUL),
                                  after_inst));
}

/** Recursive computation of the delay member of a node. */
static void
compute_delay(struct dag_node *node, void *state)
{
        struct schedule_node *n = (struct schedule_node *)node;

        n->delay = 1;

        util_dynarray_foreach(&n->dag.edges, struct dag_edge, edge) {
                struct schedule_node *child =
                        (struct schedule_node *)edge->child;
                n->delay = MAX2(n->delay, (child->delay +
                                           instruction_latency(n, child)));
        }
}

/* Removes a DAG head, but removing only the WAR edges. (dag_prune_head()
 * should be called on it later to finish pruning the other edges).
 */
static void
pre_remove_head(struct dag *dag, struct schedule_node *n)
{
        list_delinit(&n->dag.link);

        util_dynarray_foreach(&n->dag.edges, struct dag_edge, edge) {
                if (edge->data)
                        dag_remove_edge(dag, edge);
        }
}

static void
mark_instruction_scheduled(struct dag *dag,
                           uint32_t time,
                           struct schedule_node *node)
{
        if (!node)
                return;

        util_dynarray_foreach(&node->dag.edges, struct dag_edge, edge) {
                struct schedule_node *child =
                        (struct schedule_node *)edge->child;

                if (!child)
                        continue;

                uint32_t latency = instruction_latency(node, child);

                child->unblocked_time = MAX2(child->unblocked_time,
                                             time + latency);
        }
        dag_prune_head(dag, &node->dag);
}

/**
 * Emits a THRSW/LTHRSW signal in the stream, trying to move it up to pair
 * with another instruction.
 */
static void
emit_thrsw(struct vc4_compile *c,
           struct choose_scoreboard *scoreboard,
           uint64_t inst)
{
        uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);

        /* There should be nothing in a thrsw inst being scheduled other than
         * the signal bits.
         */
        assert(QPU_GET_FIELD(inst, QPU_OP_ADD) == QPU_A_NOP);
        assert(QPU_GET_FIELD(inst, QPU_OP_MUL) == QPU_M_NOP);

        /* Try to find an earlier scheduled instruction that we can merge the
         * thrsw into.
         */
        int thrsw_ip = c->qpu_inst_count;
        for (int i = 1; i <= MIN2(c->qpu_inst_count, 3); i++) {
                uint64_t prev_instr = c->qpu_insts[c->qpu_inst_count - i];
                uint32_t prev_sig = QPU_GET_FIELD(prev_instr, QPU_SIG);

                if (prev_sig == QPU_SIG_NONE)
                        thrsw_ip = c->qpu_inst_count - i;
        }

        if (thrsw_ip != c->qpu_inst_count) {
                /* Merge the thrsw into the existing instruction. */
                c->qpu_insts[thrsw_ip] =
                        QPU_UPDATE_FIELD(c->qpu_insts[thrsw_ip], sig, QPU_SIG);
        } else {
                qpu_serialize_one_inst(c, inst);
                update_scoreboard_for_chosen(scoreboard, inst);
        }

        /* Fill the delay slots. */
        while (c->qpu_inst_count < thrsw_ip + 3) {
                update_scoreboard_for_chosen(scoreboard, qpu_NOP());
                qpu_serialize_one_inst(c, qpu_NOP());
        }
}

static uint32_t
schedule_instructions(struct vc4_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)
{
        uint32_t time = 0;

        while (!list_is_empty(&scoreboard->dag->heads)) {
                struct schedule_node *chosen =
                        choose_instruction_to_schedule(scoreboard,
                                                       schedule_list,
                                                       NULL);
                struct schedule_node *merge = NULL;

                /* If there are no valid instructions to schedule, drop a NOP
                 * in.
                 */
                uint64_t inst = chosen ? chosen->inst->inst : qpu_NOP();

                if (debug) {
                        fprintf(stderr, "t=%4d: current list:\n",
                                time);
                        dump_state(scoreboard->dag);
                        fprintf(stderr, "t=%4d: chose: ", time);
                        vc4_qpu_disasm(&inst, 1);
                        fprintf(stderr, "\n");
                }

                /* 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);
                        pre_remove_head(scoreboard->dag, chosen);
                        if (chosen->uniform != -1) {
                                c->uniform_data[*next_uniform] =
                                        orig_uniform_data[chosen->uniform];
                                c->uniform_contents[*next_uniform] =
                                        orig_uniform_contents[chosen->uniform];
                                (*next_uniform)++;
                        }

                        merge = choose_instruction_to_schedule(scoreboard,
                                                               schedule_list,
                                                               chosen);
                        if (merge) {
                                time = MAX2(merge->unblocked_time, time);
                                inst = qpu_merge_inst(inst, merge->inst->inst);
                                assert(inst != 0);
                                if (merge->uniform != -1) {
                                        c->uniform_data[*next_uniform] =
                                                orig_uniform_data[merge->uniform];
                                        c->uniform_contents[*next_uniform] =
                                                orig_uniform_contents[merge->uniform];
                                        (*next_uniform)++;
                                }

                                if (debug) {
                                        fprintf(stderr, "t=%4d: merging: ",
                                                time);
                                        vc4_qpu_disasm(&merge->inst->inst, 1);
                                        fprintf(stderr, "\n");
                                        fprintf(stderr, "            resulting in: ");
                                        vc4_qpu_disasm(&inst, 1);
                                        fprintf(stderr, "\n");
                                }
                        }
                }

                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(scoreboard->dag, time, chosen);
                mark_instruction_scheduled(scoreboard->dag, time, merge);

                if (QPU_GET_FIELD(inst, QPU_SIG) == QPU_SIG_THREAD_SWITCH ||
                    QPU_GET_FIELD(inst, QPU_SIG) == QPU_SIG_LAST_THREAD_SWITCH) {
                        emit_thrsw(c, scoreboard, inst);
                } else {
                        qpu_serialize_one_inst(c, inst);
                        update_scoreboard_for_chosen(scoreboard, inst);
                }

                scoreboard->tick++;
                time++;

                if (QPU_GET_FIELD(inst, QPU_SIG) == QPU_SIG_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)
                        */
                        inst = qpu_NOP();
                        update_scoreboard_for_chosen(scoreboard, inst);
                        qpu_serialize_one_inst(c, inst);
                        qpu_serialize_one_inst(c, inst);
                        qpu_serialize_one_inst(c, inst);
                }
        }

        return time;
}

static uint32_t
qpu_schedule_instructions_block(struct vc4_compile *c,
                                struct choose_scoreboard *scoreboard,
                                struct qblock *block,
                                enum quniform_contents *orig_uniform_contents,
                                uint32_t *orig_uniform_data,
                                uint32_t *next_uniform)
{
        scoreboard->dag = dag_create(NULL);
        struct list_head setup_list;

        list_inithead(&setup_list);

        /* Wrap each instruction in a scheduler structure. */
        uint32_t next_sched_uniform = *next_uniform;
        while (!list_is_empty(&block->qpu_inst_list)) {
                struct queued_qpu_inst *inst =
                        (struct queued_qpu_inst *)block->qpu_inst_list.next;
                struct schedule_node *n = rzalloc(scoreboard->dag,
                                                  struct schedule_node);

                dag_init_node(scoreboard->dag, &n->dag);
                n->inst = inst;

                if (reads_uniform(inst->inst)) {
                        n->uniform = next_sched_uniform++;
                } else {
                        n->uniform = -1;
                }
                list_del(&inst->link);
                list_addtail(&n->link, &setup_list);
        }

        calculate_forward_deps(c, scoreboard->dag, &setup_list);
        calculate_reverse_deps(c, scoreboard->dag, &setup_list);

        dag_traverse_bottom_up(scoreboard->dag, compute_delay, NULL);

        uint32_t cycles = schedule_instructions(c, scoreboard, block,
                                                &setup_list,
                                                orig_uniform_contents,
                                                orig_uniform_data,
                                                next_uniform);

        ralloc_free(scoreboard->dag);
        scoreboard->dag = NULL;

        return cycles;
}

static void
qpu_set_branch_targets(struct vc4_compile *c)
{
        qir_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;
                }

                /* Set the branch target for the block that doesn't follow
                 * immediately after ours.
                 */
                uint64_t *branch_inst = &c->qpu_insts[block->branch_qpu_ip];
                assert(QPU_GET_FIELD(*branch_inst, QPU_SIG) == QPU_SIG_BRANCH);
                assert(QPU_GET_FIELD(*branch_inst, QPU_BRANCH_TARGET) == 0);

                uint32_t branch_target =
                        (block->successors[0]->start_qpu_ip -
                         (block->branch_qpu_ip + 4)) * sizeof(uint64_t);
                *branch_inst = (*branch_inst |
                                QPU_SET_FIELD(branch_target, QPU_BRANCH_TARGET));

                /* Make sure that the if-we-don't-jump successor was scheduled
                 * just after the delay slots.
                 */
                if (block->successors[1]) {
                        assert(block->successors[1]->start_qpu_ip ==
                               block->branch_qpu_ip + 4);
                }
        }
}

uint32_t
qpu_schedule_instructions(struct vc4_compile *c)
{
        /* 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_waddr_a = ~0;
        scoreboard.last_waddr_b = ~0;
        scoreboard.last_sfu_write_tick = -10;
        scoreboard.last_uniforms_reset_tick = -10;

        if (debug) {
                fprintf(stderr, "Pre-schedule instructions\n");
                qir_for_each_block(block, c) {
                        fprintf(stderr, "BLOCK %d\n", block->index);
                        list_for_each_entry(struct queued_qpu_inst, q,
                                            &block->qpu_inst_list, link) {
                                vc4_qpu_disasm(&q->inst, 1);
                                fprintf(stderr, "\n");
                        }
                }
                fprintf(stderr, "\n");
        }

        uint32_t cycles = 0;
        qir_for_each_block(block, c) {
                block->start_qpu_ip = c->qpu_inst_count;
                block->branch_qpu_ip = ~0;

                cycles += qpu_schedule_instructions_block(c,
                                                          &scoreboard,
                                                          block,
                                                          uniform_contents,
                                                          uniform_data,
                                                          &next_uniform);

                block->end_qpu_ip = c->qpu_inst_count - 1;
        }

        qpu_set_branch_targets(c);

        assert(next_uniform == c->num_uniforms);

        if (debug) {
                fprintf(stderr, "Post-schedule instructions\n");
                vc4_qpu_disasm(c->qpu_insts, c->qpu_inst_count);
                fprintf(stderr, "\n");
        }

        return cycles;
}