1 /*
2 * Copyright © 2010 Intel Corporation
3 * Copyright © 2014-2017 Broadcom
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
22 * IN THE SOFTWARE.
23 */
24
25 /**
26 * @file
27 *
28 * The basic model of the list scheduler is to take a basic block, compute a
29 * DAG of the dependencies, and make a list of the DAG heads. Heuristically
30 * pick a DAG head, then put all the children that are now DAG heads into the
31 * list of things to schedule.
32 *
33 * The goal of scheduling here is to pack pairs of operations together in a
34 * single QPU instruction.
35 */
36
37 #include "qpu/qpu_disasm.h"
38 #include "v3d_compiler.h"
39 #include "util/ralloc.h"
40 #include "util/dag.h"
41
42 static bool debug;
43
44 struct schedule_node_child;
45
46 struct schedule_node {
47 struct dag_node dag;
48 struct list_head link;
49 struct qinst *inst;
50
51 /* Longest cycles + instruction_latency() of any parent of this node. */
52 uint32_t unblocked_time;
53
54 /**
55 * Minimum number of cycles from scheduling this instruction until the
56 * end of the program, based on the slowest dependency chain through
57 * the children.
58 */
59 uint32_t delay;
60
61 /**
62 * cycles between this instruction being scheduled and when its result
63 * can be consumed.
64 */
65 uint32_t latency;
66 };
67
68 /* When walking the instructions in reverse, we need to swap before/after in
69 * add_dep().
70 */
71 enum direction { F, R };
72
73 struct schedule_state {
74 const struct v3d_device_info *devinfo;
75 struct dag *dag;
76 struct schedule_node *last_r[6];
77 struct schedule_node *last_rf[64];
78 struct schedule_node *last_sf;
79 struct schedule_node *last_vpm_read;
80 struct schedule_node *last_tmu_write;
81 struct schedule_node *last_tmu_config;
82 struct schedule_node *last_tlb;
83 struct schedule_node *last_vpm;
84 struct schedule_node *last_unif;
85 struct schedule_node *last_rtop;
86 enum direction dir;
87 /* Estimated cycle when the current instruction would start. */
88 uint32_t time;
89 };
90
91 static void
add_dep(struct schedule_state * state,struct schedule_node * before,struct schedule_node * after,bool write)92 add_dep(struct schedule_state *state,
93 struct schedule_node *before,
94 struct schedule_node *after,
95 bool write)
96 {
97 bool write_after_read = !write && state->dir == R;
98 void *edge_data = (void *)(uintptr_t)write_after_read;
99
100 if (!before || !after)
101 return;
102
103 assert(before != after);
104
105 if (state->dir == F)
106 dag_add_edge(&before->dag, &after->dag, edge_data);
107 else
108 dag_add_edge(&after->dag, &before->dag, edge_data);
109 }
110
111 static void
add_read_dep(struct schedule_state * state,struct schedule_node * before,struct schedule_node * after)112 add_read_dep(struct schedule_state *state,
113 struct schedule_node *before,
114 struct schedule_node *after)
115 {
116 add_dep(state, before, after, false);
117 }
118
119 static void
add_write_dep(struct schedule_state * state,struct schedule_node ** before,struct schedule_node * after)120 add_write_dep(struct schedule_state *state,
121 struct schedule_node **before,
122 struct schedule_node *after)
123 {
124 add_dep(state, *before, after, true);
125 *before = after;
126 }
127
128 static bool
qpu_inst_is_tlb(const struct v3d_qpu_instr * inst)129 qpu_inst_is_tlb(const struct v3d_qpu_instr *inst)
130 {
131 if (inst->sig.ldtlb || inst->sig.ldtlbu)
132 return true;
133
134 if (inst->type != V3D_QPU_INSTR_TYPE_ALU)
135 return false;
136
137 if (inst->alu.add.magic_write &&
138 (inst->alu.add.waddr == V3D_QPU_WADDR_TLB ||
139 inst->alu.add.waddr == V3D_QPU_WADDR_TLBU))
140 return true;
141
142 if (inst->alu.mul.magic_write &&
143 (inst->alu.mul.waddr == V3D_QPU_WADDR_TLB ||
144 inst->alu.mul.waddr == V3D_QPU_WADDR_TLBU))
145 return true;
146
147 return false;
148 }
149
150 static void
process_mux_deps(struct schedule_state * state,struct schedule_node * n,enum v3d_qpu_mux mux)151 process_mux_deps(struct schedule_state *state, struct schedule_node *n,
152 enum v3d_qpu_mux mux)
153 {
154 switch (mux) {
155 case V3D_QPU_MUX_A:
156 add_read_dep(state, state->last_rf[n->inst->qpu.raddr_a], n);
157 break;
158 case V3D_QPU_MUX_B:
159 if (!n->inst->qpu.sig.small_imm) {
160 add_read_dep(state,
161 state->last_rf[n->inst->qpu.raddr_b], n);
162 }
163 break;
164 default:
165 add_read_dep(state, state->last_r[mux - V3D_QPU_MUX_R0], n);
166 break;
167 }
168 }
169
170
171 static void
process_waddr_deps(struct schedule_state * state,struct schedule_node * n,uint32_t waddr,bool magic)172 process_waddr_deps(struct schedule_state *state, struct schedule_node *n,
173 uint32_t waddr, bool magic)
174 {
175 if (!magic) {
176 add_write_dep(state, &state->last_rf[waddr], n);
177 } else if (v3d_qpu_magic_waddr_is_tmu(waddr)) {
178 /* XXX perf: For V3D 4.x, we could reorder TMU writes other
179 * than the TMUS/TMUD/TMUA to improve scheduling flexibility.
180 */
181 add_write_dep(state, &state->last_tmu_write, n);
182 switch (waddr) {
183 case V3D_QPU_WADDR_TMUS:
184 case V3D_QPU_WADDR_TMUSCM:
185 case V3D_QPU_WADDR_TMUSF:
186 case V3D_QPU_WADDR_TMUSLOD:
187 add_write_dep(state, &state->last_tmu_config, n);
188 break;
189 default:
190 break;
191 }
192 } else if (v3d_qpu_magic_waddr_is_sfu(waddr)) {
193 /* Handled by v3d_qpu_writes_r4() check. */
194 } else {
195 switch (waddr) {
196 case V3D_QPU_WADDR_R0:
197 case V3D_QPU_WADDR_R1:
198 case V3D_QPU_WADDR_R2:
199 add_write_dep(state,
200 &state->last_r[waddr - V3D_QPU_WADDR_R0],
201 n);
202 break;
203 case V3D_QPU_WADDR_R3:
204 case V3D_QPU_WADDR_R4:
205 case V3D_QPU_WADDR_R5:
206 /* Handled by v3d_qpu_writes_r*() checks below. */
207 break;
208
209 case V3D_QPU_WADDR_VPM:
210 case V3D_QPU_WADDR_VPMU:
211 add_write_dep(state, &state->last_vpm, n);
212 break;
213
214 case V3D_QPU_WADDR_TLB:
215 case V3D_QPU_WADDR_TLBU:
216 add_write_dep(state, &state->last_tlb, n);
217 break;
218
219 case V3D_QPU_WADDR_SYNC:
220 case V3D_QPU_WADDR_SYNCB:
221 case V3D_QPU_WADDR_SYNCU:
222 /* For CS barrier(): Sync against any other memory
223 * accesses. There doesn't appear to be any need for
224 * barriers to affect ALU operations.
225 */
226 add_write_dep(state, &state->last_tmu_write, n);
227 break;
228
229 case V3D_QPU_WADDR_NOP:
230 break;
231
232 default:
233 fprintf(stderr, "Unknown waddr %d\n", waddr);
234 abort();
235 }
236 }
237 }
238
239 /**
240 * Common code for dependencies that need to be tracked both forward and
241 * backward.
242 *
243 * This is for things like "all reads of r4 have to happen between the r4
244 * writes that surround them".
245 */
246 static void
calculate_deps(struct schedule_state * state,struct schedule_node * n)247 calculate_deps(struct schedule_state *state, struct schedule_node *n)
248 {
249 const struct v3d_device_info *devinfo = state->devinfo;
250 struct qinst *qinst = n->inst;
251 struct v3d_qpu_instr *inst = &qinst->qpu;
252 /* If the input and output segments are shared, then all VPM reads to
253 * a location need to happen before all writes. We handle this by
254 * serializing all VPM operations for now.
255 */
256 bool separate_vpm_segment = false;
257
258 if (inst->type == V3D_QPU_INSTR_TYPE_BRANCH) {
259 if (inst->branch.cond != V3D_QPU_BRANCH_COND_ALWAYS)
260 add_read_dep(state, state->last_sf, n);
261
262 /* XXX: BDI */
263 /* XXX: BDU */
264 /* XXX: ub */
265 /* XXX: raddr_a */
266
267 add_write_dep(state, &state->last_unif, n);
268 return;
269 }
270
271 assert(inst->type == V3D_QPU_INSTR_TYPE_ALU);
272
273 /* XXX: LOAD_IMM */
274
275 if (v3d_qpu_add_op_num_src(inst->alu.add.op) > 0)
276 process_mux_deps(state, n, inst->alu.add.a);
277 if (v3d_qpu_add_op_num_src(inst->alu.add.op) > 1)
278 process_mux_deps(state, n, inst->alu.add.b);
279
280 if (v3d_qpu_mul_op_num_src(inst->alu.mul.op) > 0)
281 process_mux_deps(state, n, inst->alu.mul.a);
282 if (v3d_qpu_mul_op_num_src(inst->alu.mul.op) > 1)
283 process_mux_deps(state, n, inst->alu.mul.b);
284
285 switch (inst->alu.add.op) {
286 case V3D_QPU_A_VPMSETUP:
287 /* Could distinguish read/write by unpacking the uniform. */
288 add_write_dep(state, &state->last_vpm, n);
289 add_write_dep(state, &state->last_vpm_read, n);
290 break;
291
292 case V3D_QPU_A_STVPMV:
293 case V3D_QPU_A_STVPMD:
294 case V3D_QPU_A_STVPMP:
295 add_write_dep(state, &state->last_vpm, n);
296 break;
297
298 case V3D_QPU_A_LDVPMV_IN:
299 case V3D_QPU_A_LDVPMD_IN:
300 case V3D_QPU_A_LDVPMG_IN:
301 case V3D_QPU_A_LDVPMP:
302 if (!separate_vpm_segment)
303 add_write_dep(state, &state->last_vpm, n);
304 break;
305
306 case V3D_QPU_A_VPMWT:
307 add_read_dep(state, state->last_vpm, n);
308 break;
309
310 case V3D_QPU_A_MSF:
311 add_read_dep(state, state->last_tlb, n);
312 break;
313
314 case V3D_QPU_A_SETMSF:
315 case V3D_QPU_A_SETREVF:
316 add_write_dep(state, &state->last_tlb, n);
317 break;
318
319 default:
320 break;
321 }
322
323 switch (inst->alu.mul.op) {
324 case V3D_QPU_M_MULTOP:
325 case V3D_QPU_M_UMUL24:
326 /* MULTOP sets rtop, and UMUL24 implicitly reads rtop and
327 * resets it to 0. We could possibly reorder umul24s relative
328 * to each other, but for now just keep all the MUL parts in
329 * order.
330 */
331 add_write_dep(state, &state->last_rtop, n);
332 break;
333 default:
334 break;
335 }
336
337 if (inst->alu.add.op != V3D_QPU_A_NOP) {
338 process_waddr_deps(state, n, inst->alu.add.waddr,
339 inst->alu.add.magic_write);
340 }
341 if (inst->alu.mul.op != V3D_QPU_M_NOP) {
342 process_waddr_deps(state, n, inst->alu.mul.waddr,
343 inst->alu.mul.magic_write);
344 }
345 if (v3d_qpu_sig_writes_address(devinfo, &inst->sig)) {
346 process_waddr_deps(state, n, inst->sig_addr,
347 inst->sig_magic);
348 }
349
350 if (v3d_qpu_writes_r3(devinfo, inst))
351 add_write_dep(state, &state->last_r[3], n);
352 if (v3d_qpu_writes_r4(devinfo, inst))
353 add_write_dep(state, &state->last_r[4], n);
354 if (v3d_qpu_writes_r5(devinfo, inst))
355 add_write_dep(state, &state->last_r[5], n);
356
357 if (inst->sig.thrsw) {
358 /* All accumulator contents and flags are undefined after the
359 * switch.
360 */
361 for (int i = 0; i < ARRAY_SIZE(state->last_r); i++)
362 add_write_dep(state, &state->last_r[i], n);
363 add_write_dep(state, &state->last_sf, n);
364 add_write_dep(state, &state->last_rtop, n);
365
366 /* Scoreboard-locking operations have to stay after the last
367 * thread switch.
368 */
369 add_write_dep(state, &state->last_tlb, n);
370
371 add_write_dep(state, &state->last_tmu_write, n);
372 add_write_dep(state, &state->last_tmu_config, n);
373 }
374
375 if (v3d_qpu_waits_on_tmu(inst)) {
376 /* TMU loads are coming from a FIFO, so ordering is important.
377 */
378 add_write_dep(state, &state->last_tmu_write, n);
379 }
380
381 if (inst->sig.wrtmuc)
382 add_write_dep(state, &state->last_tmu_config, n);
383
384 if (inst->sig.ldtlb | inst->sig.ldtlbu)
385 add_write_dep(state, &state->last_tlb, n);
386
387 if (inst->sig.ldvpm) {
388 add_write_dep(state, &state->last_vpm_read, n);
389
390 /* At least for now, we're doing shared I/O segments, so queue
391 * all writes after all reads.
392 */
393 if (!separate_vpm_segment)
394 add_write_dep(state, &state->last_vpm, n);
395 }
396
397 /* inst->sig.ldunif or sideband uniform read */
398 if (vir_has_uniform(qinst))
399 add_write_dep(state, &state->last_unif, n);
400
401 if (v3d_qpu_reads_flags(inst))
402 add_read_dep(state, state->last_sf, n);
403 if (v3d_qpu_writes_flags(inst))
404 add_write_dep(state, &state->last_sf, n);
405 }
406
407 static void
calculate_forward_deps(struct v3d_compile * c,struct dag * dag,struct list_head * schedule_list)408 calculate_forward_deps(struct v3d_compile *c, struct dag *dag,
409 struct list_head *schedule_list)
410 {
411 struct schedule_state state;
412
413 memset(&state, 0, sizeof(state));
414 state.dag = dag;
415 state.devinfo = c->devinfo;
416 state.dir = F;
417
418 list_for_each_entry(struct schedule_node, node, schedule_list, link)
419 calculate_deps(&state, node);
420 }
421
422 static void
calculate_reverse_deps(struct v3d_compile * c,struct dag * dag,struct list_head * schedule_list)423 calculate_reverse_deps(struct v3d_compile *c, struct dag *dag,
424 struct list_head *schedule_list)
425 {
426 struct schedule_state state;
427
428 memset(&state, 0, sizeof(state));
429 state.dag = dag;
430 state.devinfo = c->devinfo;
431 state.dir = R;
432
433 list_for_each_entry_rev(struct schedule_node, node, schedule_list,
434 link) {
435 calculate_deps(&state, (struct schedule_node *)node);
436 }
437 }
438
439 struct choose_scoreboard {
440 struct dag *dag;
441 int tick;
442 int last_magic_sfu_write_tick;
443 int last_stallable_sfu_reg;
444 int last_stallable_sfu_tick;
445 int last_ldvary_tick;
446 int last_uniforms_reset_tick;
447 int last_thrsw_tick;
448 bool tlb_locked;
449 };
450
451 static bool
mux_reads_too_soon(struct choose_scoreboard * scoreboard,const struct v3d_qpu_instr * inst,enum v3d_qpu_mux mux)452 mux_reads_too_soon(struct choose_scoreboard *scoreboard,
453 const struct v3d_qpu_instr *inst, enum v3d_qpu_mux mux)
454 {
455 switch (mux) {
456 case V3D_QPU_MUX_R4:
457 if (scoreboard->tick - scoreboard->last_magic_sfu_write_tick <= 2)
458 return true;
459 break;
460
461 case V3D_QPU_MUX_R5:
462 if (scoreboard->tick - scoreboard->last_ldvary_tick <= 1)
463 return true;
464 break;
465 default:
466 break;
467 }
468
469 return false;
470 }
471
472 static bool
reads_too_soon_after_write(struct choose_scoreboard * scoreboard,struct qinst * qinst)473 reads_too_soon_after_write(struct choose_scoreboard *scoreboard,
474 struct qinst *qinst)
475 {
476 const struct v3d_qpu_instr *inst = &qinst->qpu;
477
478 /* XXX: Branching off of raddr. */
479 if (inst->type == V3D_QPU_INSTR_TYPE_BRANCH)
480 return false;
481
482 assert(inst->type == V3D_QPU_INSTR_TYPE_ALU);
483
484 if (inst->alu.add.op != V3D_QPU_A_NOP) {
485 if (v3d_qpu_add_op_num_src(inst->alu.add.op) > 0 &&
486 mux_reads_too_soon(scoreboard, inst, inst->alu.add.a)) {
487 return true;
488 }
489 if (v3d_qpu_add_op_num_src(inst->alu.add.op) > 1 &&
490 mux_reads_too_soon(scoreboard, inst, inst->alu.add.b)) {
491 return true;
492 }
493 }
494
495 if (inst->alu.mul.op != V3D_QPU_M_NOP) {
496 if (v3d_qpu_mul_op_num_src(inst->alu.mul.op) > 0 &&
497 mux_reads_too_soon(scoreboard, inst, inst->alu.mul.a)) {
498 return true;
499 }
500 if (v3d_qpu_mul_op_num_src(inst->alu.mul.op) > 1 &&
501 mux_reads_too_soon(scoreboard, inst, inst->alu.mul.b)) {
502 return true;
503 }
504 }
505
506 /* XXX: imm */
507
508 return false;
509 }
510
511 static bool
writes_too_soon_after_write(const struct v3d_device_info * devinfo,struct choose_scoreboard * scoreboard,struct qinst * qinst)512 writes_too_soon_after_write(const struct v3d_device_info *devinfo,
513 struct choose_scoreboard *scoreboard,
514 struct qinst *qinst)
515 {
516 const struct v3d_qpu_instr *inst = &qinst->qpu;
517
518 /* Don't schedule any other r4 write too soon after an SFU write.
519 * This would normally be prevented by dependency tracking, but might
520 * occur if a dead SFU computation makes it to scheduling.
521 */
522 if (scoreboard->tick - scoreboard->last_magic_sfu_write_tick < 2 &&
523 v3d_qpu_writes_r4(devinfo, inst))
524 return true;
525
526 return false;
527 }
528
529 static bool
pixel_scoreboard_too_soon(struct choose_scoreboard * scoreboard,const struct v3d_qpu_instr * inst)530 pixel_scoreboard_too_soon(struct choose_scoreboard *scoreboard,
531 const struct v3d_qpu_instr *inst)
532 {
533 return (scoreboard->tick == 0 && qpu_inst_is_tlb(inst));
534 }
535
536 static bool
qpu_instruction_uses_rf(const struct v3d_qpu_instr * inst,uint32_t waddr)537 qpu_instruction_uses_rf(const struct v3d_qpu_instr *inst,
538 uint32_t waddr) {
539
540 if (inst->type != V3D_QPU_INSTR_TYPE_ALU)
541 return false;
542
543 if (v3d_qpu_uses_mux(inst, V3D_QPU_MUX_A) &&
544 inst->raddr_a == waddr)
545 return true;
546
547 if (v3d_qpu_uses_mux(inst, V3D_QPU_MUX_B) &&
548 !inst->sig.small_imm && (inst->raddr_b == waddr))
549 return true;
550
551 return false;
552 }
553
554 static bool
mux_read_stalls(struct choose_scoreboard * scoreboard,const struct v3d_qpu_instr * inst)555 mux_read_stalls(struct choose_scoreboard *scoreboard,
556 const struct v3d_qpu_instr *inst)
557 {
558 return scoreboard->tick == scoreboard->last_stallable_sfu_tick + 1 &&
559 qpu_instruction_uses_rf(inst,
560 scoreboard->last_stallable_sfu_reg);
561 }
562
563 /* We define a max schedule priority to allow negative priorities as result of
564 * substracting this max when an instruction stalls. So instructions that
565 * stall have lower priority than regular instructions. */
566 #define MAX_SCHEDULE_PRIORITY 16
567
568 static int
get_instruction_priority(const struct v3d_qpu_instr * inst)569 get_instruction_priority(const struct v3d_qpu_instr *inst)
570 {
571 uint32_t baseline_score;
572 uint32_t next_score = 0;
573
574 /* Schedule TLB operations as late as possible, to get more
575 * parallelism between shaders.
576 */
577 if (qpu_inst_is_tlb(inst))
578 return next_score;
579 next_score++;
580
581 /* Schedule texture read results collection late to hide latency. */
582 if (v3d_qpu_waits_on_tmu(inst))
583 return next_score;
584 next_score++;
585
586 /* Default score for things that aren't otherwise special. */
587 baseline_score = next_score;
588 next_score++;
589
590 /* Schedule texture read setup early to hide their latency better. */
591 if (v3d_qpu_writes_tmu(inst))
592 return next_score;
593 next_score++;
594
595 /* We should increase the maximum if we assert here */
596 assert(next_score < MAX_SCHEDULE_PRIORITY);
597
598 return baseline_score;
599 }
600
601 static bool
qpu_magic_waddr_is_periph(enum v3d_qpu_waddr waddr)602 qpu_magic_waddr_is_periph(enum v3d_qpu_waddr waddr)
603 {
604 return (v3d_qpu_magic_waddr_is_tmu(waddr) ||
605 v3d_qpu_magic_waddr_is_sfu(waddr) ||
606 v3d_qpu_magic_waddr_is_tlb(waddr) ||
607 v3d_qpu_magic_waddr_is_vpm(waddr) ||
608 v3d_qpu_magic_waddr_is_tsy(waddr));
609 }
610
611 static bool
qpu_accesses_peripheral(const struct v3d_qpu_instr * inst)612 qpu_accesses_peripheral(const struct v3d_qpu_instr *inst)
613 {
614 if (v3d_qpu_uses_vpm(inst))
615 return true;
616 if (v3d_qpu_uses_sfu(inst))
617 return true;
618
619 if (inst->type == V3D_QPU_INSTR_TYPE_ALU) {
620 if (inst->alu.add.op != V3D_QPU_A_NOP &&
621 inst->alu.add.magic_write &&
622 qpu_magic_waddr_is_periph(inst->alu.add.waddr)) {
623 return true;
624 }
625
626 if (inst->alu.add.op == V3D_QPU_A_TMUWT)
627 return true;
628
629 if (inst->alu.mul.op != V3D_QPU_M_NOP &&
630 inst->alu.mul.magic_write &&
631 qpu_magic_waddr_is_periph(inst->alu.mul.waddr)) {
632 return true;
633 }
634 }
635
636 return (inst->sig.ldvpm ||
637 inst->sig.ldtmu ||
638 inst->sig.ldtlb ||
639 inst->sig.ldtlbu ||
640 inst->sig.wrtmuc);
641 }
642
643 static bool
qpu_compatible_peripheral_access(const struct v3d_device_info * devinfo,const struct v3d_qpu_instr * a,const struct v3d_qpu_instr * b)644 qpu_compatible_peripheral_access(const struct v3d_device_info *devinfo,
645 const struct v3d_qpu_instr *a,
646 const struct v3d_qpu_instr *b)
647 {
648 const bool a_uses_peripheral = qpu_accesses_peripheral(a);
649 const bool b_uses_peripheral = qpu_accesses_peripheral(b);
650
651 /* We can always do one peripheral access per instruction. */
652 if (!a_uses_peripheral || !b_uses_peripheral)
653 return true;
654
655 if (devinfo->ver < 41)
656 return false;
657
658 /* V3D 4.1 and later allow TMU read along with a VPM read or write, and
659 * WRTMUC with a TMU magic register write (other than tmuc).
660 */
661 if ((a->sig.ldtmu && v3d_qpu_reads_or_writes_vpm(b)) ||
662 (b->sig.ldtmu && v3d_qpu_reads_or_writes_vpm(a))) {
663 return true;
664 }
665
666 if ((a->sig.wrtmuc && v3d_qpu_writes_tmu_not_tmuc(b)) ||
667 (b->sig.wrtmuc && v3d_qpu_writes_tmu_not_tmuc(a))) {
668 return true;
669 }
670
671 return false;
672 }
673
674 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)675 qpu_merge_inst(const struct v3d_device_info *devinfo,
676 struct v3d_qpu_instr *result,
677 const struct v3d_qpu_instr *a,
678 const struct v3d_qpu_instr *b)
679 {
680 if (a->type != V3D_QPU_INSTR_TYPE_ALU ||
681 b->type != V3D_QPU_INSTR_TYPE_ALU) {
682 return false;
683 }
684
685 if (!qpu_compatible_peripheral_access(devinfo, a, b))
686 return false;
687
688 struct v3d_qpu_instr merge = *a;
689
690 if (b->alu.add.op != V3D_QPU_A_NOP) {
691 if (a->alu.add.op != V3D_QPU_A_NOP)
692 return false;
693 merge.alu.add = b->alu.add;
694
695 merge.flags.ac = b->flags.ac;
696 merge.flags.apf = b->flags.apf;
697 merge.flags.auf = b->flags.auf;
698 }
699
700 if (b->alu.mul.op != V3D_QPU_M_NOP) {
701 if (a->alu.mul.op != V3D_QPU_M_NOP)
702 return false;
703 merge.alu.mul = b->alu.mul;
704
705 merge.flags.mc = b->flags.mc;
706 merge.flags.mpf = b->flags.mpf;
707 merge.flags.muf = b->flags.muf;
708 }
709
710 if (v3d_qpu_uses_mux(b, V3D_QPU_MUX_A)) {
711 if (v3d_qpu_uses_mux(a, V3D_QPU_MUX_A) &&
712 a->raddr_a != b->raddr_a) {
713 return false;
714 }
715 merge.raddr_a = b->raddr_a;
716 }
717
718 if (v3d_qpu_uses_mux(b, V3D_QPU_MUX_B)) {
719 if (v3d_qpu_uses_mux(a, V3D_QPU_MUX_B) &&
720 (a->raddr_b != b->raddr_b ||
721 a->sig.small_imm != b->sig.small_imm)) {
722 return false;
723 }
724 merge.raddr_b = b->raddr_b;
725 }
726
727 merge.sig.thrsw |= b->sig.thrsw;
728 merge.sig.ldunif |= b->sig.ldunif;
729 merge.sig.ldunifrf |= b->sig.ldunifrf;
730 merge.sig.ldunifa |= b->sig.ldunifa;
731 merge.sig.ldunifarf |= b->sig.ldunifarf;
732 merge.sig.ldtmu |= b->sig.ldtmu;
733 merge.sig.ldvary |= b->sig.ldvary;
734 merge.sig.ldvpm |= b->sig.ldvpm;
735 merge.sig.small_imm |= b->sig.small_imm;
736 merge.sig.ldtlb |= b->sig.ldtlb;
737 merge.sig.ldtlbu |= b->sig.ldtlbu;
738 merge.sig.ucb |= b->sig.ucb;
739 merge.sig.rotate |= b->sig.rotate;
740 merge.sig.wrtmuc |= b->sig.wrtmuc;
741
742 if (v3d_qpu_sig_writes_address(devinfo, &a->sig) &&
743 v3d_qpu_sig_writes_address(devinfo, &b->sig))
744 return false;
745 merge.sig_addr |= b->sig_addr;
746 merge.sig_magic |= b->sig_magic;
747
748 uint64_t packed;
749 bool ok = v3d_qpu_instr_pack(devinfo, &merge, &packed);
750
751 *result = merge;
752 /* No modifying the real instructions on failure. */
753 assert(ok || (a != result && b != result));
754
755 return ok;
756 }
757
758 static struct schedule_node *
choose_instruction_to_schedule(const struct v3d_device_info * devinfo,struct choose_scoreboard * scoreboard,struct schedule_node * prev_inst)759 choose_instruction_to_schedule(const struct v3d_device_info *devinfo,
760 struct choose_scoreboard *scoreboard,
761 struct schedule_node *prev_inst)
762 {
763 struct schedule_node *chosen = NULL;
764 int chosen_prio = 0;
765
766 /* Don't pair up anything with a thread switch signal -- emit_thrsw()
767 * will handle pairing it along with filling the delay slots.
768 */
769 if (prev_inst) {
770 if (prev_inst->inst->qpu.sig.thrsw)
771 return NULL;
772 }
773
774 list_for_each_entry(struct schedule_node, n, &scoreboard->dag->heads,
775 dag.link) {
776 const struct v3d_qpu_instr *inst = &n->inst->qpu;
777
778 /* Don't choose the branch instruction until it's the last one
779 * left. We'll move it up to fit its delay slots after we
780 * choose it.
781 */
782 if (inst->type == V3D_QPU_INSTR_TYPE_BRANCH &&
783 !list_is_singular(&scoreboard->dag->heads)) {
784 continue;
785 }
786
787 /* "An instruction must not read from a location in physical
788 * regfile A or B that was written to by the previous
789 * instruction."
790 */
791 if (reads_too_soon_after_write(scoreboard, n->inst))
792 continue;
793
794 if (writes_too_soon_after_write(devinfo, scoreboard, n->inst))
795 continue;
796
797 /* "A scoreboard wait must not occur in the first two
798 * instructions of a fragment shader. This is either the
799 * explicit Wait for Scoreboard signal or an implicit wait
800 * with the first tile-buffer read or write instruction."
801 */
802 if (pixel_scoreboard_too_soon(scoreboard, inst))
803 continue;
804
805 /* ldunif and ldvary both write r5, but ldunif does so a tick
806 * sooner. If the ldvary's r5 wasn't used, then ldunif might
807 * otherwise get scheduled so ldunif and ldvary try to update
808 * r5 in the same tick.
809 *
810 * XXX perf: To get good pipelining of a sequence of varying
811 * loads, we need to figure out how to pair the ldvary signal
812 * up to the instruction before the last r5 user in the
813 * previous ldvary sequence. Currently, it usually pairs with
814 * the last r5 user.
815 */
816 if ((inst->sig.ldunif || inst->sig.ldunifa) &&
817 scoreboard->tick == scoreboard->last_ldvary_tick + 1) {
818 continue;
819 }
820
821 /* If we're trying to pair with another instruction, check
822 * that they're compatible.
823 */
824 if (prev_inst) {
825 /* Don't pair up a thread switch signal -- we'll
826 * handle pairing it when we pick it on its own.
827 */
828 if (inst->sig.thrsw)
829 continue;
830
831 if (prev_inst->inst->uniform != -1 &&
832 n->inst->uniform != -1)
833 continue;
834
835 /* Don't merge in something that will lock the TLB.
836 * Hopwefully what we have in inst will release some
837 * other instructions, allowing us to delay the
838 * TLB-locking instruction until later.
839 */
840 if (!scoreboard->tlb_locked && qpu_inst_is_tlb(inst))
841 continue;
842
843 struct v3d_qpu_instr merged_inst;
844 if (!qpu_merge_inst(devinfo, &merged_inst,
845 &prev_inst->inst->qpu, inst)) {
846 continue;
847 }
848 }
849
850 int prio = get_instruction_priority(inst);
851
852 if (mux_read_stalls(scoreboard, inst)) {
853 /* Don't merge an instruction that stalls */
854 if (prev_inst)
855 continue;
856 else {
857 /* Any instruction that don't stall will have
858 * higher scheduling priority */
859 prio -= MAX_SCHEDULE_PRIORITY;
860 assert(prio < 0);
861 }
862 }
863
864 /* Found a valid instruction. If nothing better comes along,
865 * this one works.
866 */
867 if (!chosen) {
868 chosen = n;
869 chosen_prio = prio;
870 continue;
871 }
872
873 if (prio > chosen_prio) {
874 chosen = n;
875 chosen_prio = prio;
876 } else if (prio < chosen_prio) {
877 continue;
878 }
879
880 if (n->delay > chosen->delay) {
881 chosen = n;
882 chosen_prio = prio;
883 } else if (n->delay < chosen->delay) {
884 continue;
885 }
886 }
887
888 return chosen;
889 }
890
891 static void
update_scoreboard_for_magic_waddr(struct choose_scoreboard * scoreboard,enum v3d_qpu_waddr waddr)892 update_scoreboard_for_magic_waddr(struct choose_scoreboard *scoreboard,
893 enum v3d_qpu_waddr waddr)
894 {
895 if (v3d_qpu_magic_waddr_is_sfu(waddr))
896 scoreboard->last_magic_sfu_write_tick = scoreboard->tick;
897 }
898
899 static void
update_scoreboard_for_sfu_stall_waddr(struct choose_scoreboard * scoreboard,const struct v3d_qpu_instr * inst)900 update_scoreboard_for_sfu_stall_waddr(struct choose_scoreboard *scoreboard,
901 const struct v3d_qpu_instr *inst)
902 {
903 if (v3d_qpu_instr_is_sfu(inst)) {
904 scoreboard->last_stallable_sfu_reg = inst->alu.add.waddr;
905 scoreboard->last_stallable_sfu_tick = scoreboard->tick;
906 }
907 }
908
909 static void
update_scoreboard_for_chosen(struct choose_scoreboard * scoreboard,const struct v3d_qpu_instr * inst)910 update_scoreboard_for_chosen(struct choose_scoreboard *scoreboard,
911 const struct v3d_qpu_instr *inst)
912 {
913 if (inst->type == V3D_QPU_INSTR_TYPE_BRANCH)
914 return;
915
916 assert(inst->type == V3D_QPU_INSTR_TYPE_ALU);
917
918 if (inst->alu.add.op != V3D_QPU_A_NOP) {
919 if (inst->alu.add.magic_write) {
920 update_scoreboard_for_magic_waddr(scoreboard,
921 inst->alu.add.waddr);
922 } else {
923 update_scoreboard_for_sfu_stall_waddr(scoreboard,
924 inst);
925 }
926 }
927
928 if (inst->alu.mul.op != V3D_QPU_M_NOP) {
929 if (inst->alu.mul.magic_write) {
930 update_scoreboard_for_magic_waddr(scoreboard,
931 inst->alu.mul.waddr);
932 }
933 }
934
935 if (inst->sig.ldvary)
936 scoreboard->last_ldvary_tick = scoreboard->tick;
937
938 if (qpu_inst_is_tlb(inst))
939 scoreboard->tlb_locked = true;
940 }
941
942 static void
dump_state(const struct v3d_device_info * devinfo,struct dag * dag)943 dump_state(const struct v3d_device_info *devinfo, struct dag *dag)
944 {
945 list_for_each_entry(struct schedule_node, n, &dag->heads, dag.link) {
946 fprintf(stderr, " t=%4d: ", n->unblocked_time);
947 v3d_qpu_dump(devinfo, &n->inst->qpu);
948 fprintf(stderr, "\n");
949
950 util_dynarray_foreach(&n->dag.edges, struct dag_edge, edge) {
951 struct schedule_node *child =
952 (struct schedule_node *)edge->child;
953 if (!child)
954 continue;
955
956 fprintf(stderr, " - ");
957 v3d_qpu_dump(devinfo, &child->inst->qpu);
958 fprintf(stderr, " (%d parents, %c)\n",
959 child->dag.parent_count,
960 edge->data ? 'w' : 'r');
961 }
962 }
963 }
964
magic_waddr_latency(enum v3d_qpu_waddr waddr,const struct v3d_qpu_instr * after)965 static uint32_t magic_waddr_latency(enum v3d_qpu_waddr waddr,
966 const struct v3d_qpu_instr *after)
967 {
968 /* Apply some huge latency between texture fetch requests and getting
969 * their results back.
970 *
971 * FIXME: This is actually pretty bogus. If we do:
972 *
973 * mov tmu0_s, a
974 * <a bit of math>
975 * mov tmu0_s, b
976 * load_tmu0
977 * <more math>
978 * load_tmu0
979 *
980 * we count that as worse than
981 *
982 * mov tmu0_s, a
983 * mov tmu0_s, b
984 * <lots of math>
985 * load_tmu0
986 * <more math>
987 * load_tmu0
988 *
989 * because we associate the first load_tmu0 with the *second* tmu0_s.
990 */
991 if (v3d_qpu_magic_waddr_is_tmu(waddr) && v3d_qpu_waits_on_tmu(after))
992 return 100;
993
994 /* Assume that anything depending on us is consuming the SFU result. */
995 if (v3d_qpu_magic_waddr_is_sfu(waddr))
996 return 3;
997
998 return 1;
999 }
1000
1001 static uint32_t
instruction_latency(struct schedule_node * before,struct schedule_node * after)1002 instruction_latency(struct schedule_node *before, struct schedule_node *after)
1003 {
1004 const struct v3d_qpu_instr *before_inst = &before->inst->qpu;
1005 const struct v3d_qpu_instr *after_inst = &after->inst->qpu;
1006 uint32_t latency = 1;
1007
1008 if (before_inst->type != V3D_QPU_INSTR_TYPE_ALU ||
1009 after_inst->type != V3D_QPU_INSTR_TYPE_ALU)
1010 return latency;
1011
1012 if (before_inst->alu.add.magic_write) {
1013 latency = MAX2(latency,
1014 magic_waddr_latency(before_inst->alu.add.waddr,
1015 after_inst));
1016 }
1017
1018 if (before_inst->alu.mul.magic_write) {
1019 latency = MAX2(latency,
1020 magic_waddr_latency(before_inst->alu.mul.waddr,
1021 after_inst));
1022 }
1023
1024 if (v3d_qpu_instr_is_sfu(before_inst))
1025 return 2;
1026
1027 return latency;
1028 }
1029
1030 /** Recursive computation of the delay member of a node. */
1031 static void
compute_delay(struct dag_node * node,void * state)1032 compute_delay(struct dag_node *node, void *state)
1033 {
1034 struct schedule_node *n = (struct schedule_node *)node;
1035
1036 n->delay = 1;
1037
1038 util_dynarray_foreach(&n->dag.edges, struct dag_edge, edge) {
1039 struct schedule_node *child =
1040 (struct schedule_node *)edge->child;
1041
1042 n->delay = MAX2(n->delay, (child->delay +
1043 instruction_latency(n, child)));
1044 }
1045 }
1046
1047 /* Removes a DAG head, but removing only the WAR edges. (dag_prune_head()
1048 * should be called on it later to finish pruning the other edges).
1049 */
1050 static void
pre_remove_head(struct dag * dag,struct schedule_node * n)1051 pre_remove_head(struct dag *dag, struct schedule_node *n)
1052 {
1053 list_delinit(&n->dag.link);
1054
1055 util_dynarray_foreach(&n->dag.edges, struct dag_edge, edge) {
1056 if (edge->data)
1057 dag_remove_edge(dag, edge);
1058 }
1059 }
1060
1061 static void
mark_instruction_scheduled(struct dag * dag,uint32_t time,struct schedule_node * node)1062 mark_instruction_scheduled(struct dag *dag,
1063 uint32_t time,
1064 struct schedule_node *node)
1065 {
1066 if (!node)
1067 return;
1068
1069 util_dynarray_foreach(&node->dag.edges, struct dag_edge, edge) {
1070 struct schedule_node *child =
1071 (struct schedule_node *)edge->child;
1072
1073 if (!child)
1074 continue;
1075
1076 uint32_t latency = instruction_latency(node, child);
1077
1078 child->unblocked_time = MAX2(child->unblocked_time,
1079 time + latency);
1080 }
1081 dag_prune_head(dag, &node->dag);
1082 }
1083
1084 static void
insert_scheduled_instruction(struct v3d_compile * c,struct qblock * block,struct choose_scoreboard * scoreboard,struct qinst * inst)1085 insert_scheduled_instruction(struct v3d_compile *c,
1086 struct qblock *block,
1087 struct choose_scoreboard *scoreboard,
1088 struct qinst *inst)
1089 {
1090 list_addtail(&inst->link, &block->instructions);
1091
1092 update_scoreboard_for_chosen(scoreboard, &inst->qpu);
1093 c->qpu_inst_count++;
1094 scoreboard->tick++;
1095 }
1096
1097 static struct qinst *
vir_nop()1098 vir_nop()
1099 {
1100 struct qreg undef = vir_nop_reg();
1101 struct qinst *qinst = vir_add_inst(V3D_QPU_A_NOP, undef, undef, undef);
1102
1103 return qinst;
1104 }
1105
1106 static void
emit_nop(struct v3d_compile * c,struct qblock * block,struct choose_scoreboard * scoreboard)1107 emit_nop(struct v3d_compile *c, struct qblock *block,
1108 struct choose_scoreboard *scoreboard)
1109 {
1110 insert_scheduled_instruction(c, block, scoreboard, vir_nop());
1111 }
1112
1113 static bool
qpu_instruction_valid_in_thrend_slot(struct v3d_compile * c,const struct qinst * qinst,int slot)1114 qpu_instruction_valid_in_thrend_slot(struct v3d_compile *c,
1115 const struct qinst *qinst, int slot)
1116 {
1117 const struct v3d_qpu_instr *inst = &qinst->qpu;
1118
1119 /* Only TLB Z writes are prohibited in the last slot, but we don't
1120 * have those flagged so prohibit all TLB ops for now.
1121 */
1122 if (slot == 2 && qpu_inst_is_tlb(inst))
1123 return false;
1124
1125 if (slot > 0 && qinst->uniform != ~0)
1126 return false;
1127
1128 if (v3d_qpu_uses_vpm(inst))
1129 return false;
1130
1131 if (inst->sig.ldvary)
1132 return false;
1133
1134 if (inst->type == V3D_QPU_INSTR_TYPE_ALU) {
1135 /* GFXH-1625: TMUWT not allowed in the final instruction. */
1136 if (slot == 2 && inst->alu.add.op == V3D_QPU_A_TMUWT)
1137 return false;
1138
1139 /* No writing physical registers at the end. */
1140 if (!inst->alu.add.magic_write ||
1141 !inst->alu.mul.magic_write) {
1142 return false;
1143 }
1144
1145 if (c->devinfo->ver < 40 && inst->alu.add.op == V3D_QPU_A_SETMSF)
1146 return false;
1147
1148 /* RF0-2 might be overwritten during the delay slots by
1149 * fragment shader setup.
1150 */
1151 if (inst->raddr_a < 3 &&
1152 (inst->alu.add.a == V3D_QPU_MUX_A ||
1153 inst->alu.add.b == V3D_QPU_MUX_A ||
1154 inst->alu.mul.a == V3D_QPU_MUX_A ||
1155 inst->alu.mul.b == V3D_QPU_MUX_A)) {
1156 return false;
1157 }
1158
1159 if (inst->raddr_b < 3 &&
1160 !inst->sig.small_imm &&
1161 (inst->alu.add.a == V3D_QPU_MUX_B ||
1162 inst->alu.add.b == V3D_QPU_MUX_B ||
1163 inst->alu.mul.a == V3D_QPU_MUX_B ||
1164 inst->alu.mul.b == V3D_QPU_MUX_B)) {
1165 return false;
1166 }
1167 }
1168
1169 return true;
1170 }
1171
1172 static bool
valid_thrsw_sequence(struct v3d_compile * c,struct choose_scoreboard * scoreboard,struct qinst * qinst,int instructions_in_sequence,bool is_thrend)1173 valid_thrsw_sequence(struct v3d_compile *c, struct choose_scoreboard *scoreboard,
1174 struct qinst *qinst, int instructions_in_sequence,
1175 bool is_thrend)
1176 {
1177 /* No emitting our thrsw while the previous thrsw hasn't happened yet. */
1178 if (scoreboard->last_thrsw_tick + 3 >
1179 scoreboard->tick - instructions_in_sequence) {
1180 return false;
1181 }
1182
1183 for (int slot = 0; slot < instructions_in_sequence; slot++) {
1184 /* No scheduling SFU when the result would land in the other
1185 * thread. The simulator complains for safety, though it
1186 * would only occur for dead code in our case.
1187 */
1188 if (slot > 0 &&
1189 qinst->qpu.type == V3D_QPU_INSTR_TYPE_ALU &&
1190 (v3d_qpu_magic_waddr_is_sfu(qinst->qpu.alu.add.waddr) ||
1191 v3d_qpu_magic_waddr_is_sfu(qinst->qpu.alu.mul.waddr))) {
1192 return false;
1193 }
1194
1195 if (slot > 0 && qinst->qpu.sig.ldvary)
1196 return false;
1197
1198 if (is_thrend &&
1199 !qpu_instruction_valid_in_thrend_slot(c, qinst, slot)) {
1200 return false;
1201 }
1202
1203 /* Note that the list is circular, so we can only do this up
1204 * to instructions_in_sequence.
1205 */
1206 qinst = (struct qinst *)qinst->link.next;
1207 }
1208
1209 return true;
1210 }
1211
1212 /**
1213 * Emits a THRSW signal in the stream, trying to move it up to pair with
1214 * another instruction.
1215 */
1216 static int
emit_thrsw(struct v3d_compile * c,struct qblock * block,struct choose_scoreboard * scoreboard,struct qinst * inst,bool is_thrend)1217 emit_thrsw(struct v3d_compile *c,
1218 struct qblock *block,
1219 struct choose_scoreboard *scoreboard,
1220 struct qinst *inst,
1221 bool is_thrend)
1222 {
1223 int time = 0;
1224
1225 /* There should be nothing in a thrsw inst being scheduled other than
1226 * the signal bits.
1227 */
1228 assert(inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU);
1229 assert(inst->qpu.alu.add.op == V3D_QPU_A_NOP);
1230 assert(inst->qpu.alu.mul.op == V3D_QPU_M_NOP);
1231
1232 /* Find how far back into previous instructions we can put the THRSW. */
1233 int slots_filled = 0;
1234 struct qinst *merge_inst = NULL;
1235 vir_for_each_inst_rev(prev_inst, block) {
1236 struct v3d_qpu_sig sig = prev_inst->qpu.sig;
1237 sig.thrsw = true;
1238 uint32_t packed_sig;
1239
1240 if (!v3d_qpu_sig_pack(c->devinfo, &sig, &packed_sig))
1241 break;
1242
1243 if (!valid_thrsw_sequence(c, scoreboard,
1244 prev_inst, slots_filled + 1,
1245 is_thrend)) {
1246 break;
1247 }
1248
1249 merge_inst = prev_inst;
1250 if (++slots_filled == 3)
1251 break;
1252 }
1253
1254 bool needs_free = false;
1255 if (merge_inst) {
1256 merge_inst->qpu.sig.thrsw = true;
1257 needs_free = true;
1258 scoreboard->last_thrsw_tick = scoreboard->tick - slots_filled;
1259 } else {
1260 scoreboard->last_thrsw_tick = scoreboard->tick;
1261 insert_scheduled_instruction(c, block, scoreboard, inst);
1262 time++;
1263 slots_filled++;
1264 merge_inst = inst;
1265 }
1266
1267 /* Insert any extra delay slot NOPs we need. */
1268 for (int i = 0; i < 3 - slots_filled; i++) {
1269 emit_nop(c, block, scoreboard);
1270 time++;
1271 }
1272
1273 /* If we're emitting the last THRSW (other than program end), then
1274 * signal that to the HW by emitting two THRSWs in a row.
1275 */
1276 if (inst->is_last_thrsw) {
1277 struct qinst *second_inst =
1278 (struct qinst *)merge_inst->link.next;
1279 second_inst->qpu.sig.thrsw = true;
1280 }
1281
1282 /* If we put our THRSW into another instruction, free up the
1283 * instruction that didn't end up scheduled into the list.
1284 */
1285 if (needs_free)
1286 free(inst);
1287
1288 return time;
1289 }
1290
1291 static uint32_t
schedule_instructions(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)1292 schedule_instructions(struct v3d_compile *c,
1293 struct choose_scoreboard *scoreboard,
1294 struct qblock *block,
1295 enum quniform_contents *orig_uniform_contents,
1296 uint32_t *orig_uniform_data,
1297 uint32_t *next_uniform)
1298 {
1299 const struct v3d_device_info *devinfo = c->devinfo;
1300 uint32_t time = 0;
1301
1302 while (!list_is_empty(&scoreboard->dag->heads)) {
1303 struct schedule_node *chosen =
1304 choose_instruction_to_schedule(devinfo,
1305 scoreboard,
1306 NULL);
1307 struct schedule_node *merge = NULL;
1308
1309 /* If there are no valid instructions to schedule, drop a NOP
1310 * in.
1311 */
1312 struct qinst *qinst = chosen ? chosen->inst : vir_nop();
1313 struct v3d_qpu_instr *inst = &qinst->qpu;
1314
1315 if (debug) {
1316 fprintf(stderr, "t=%4d: current list:\n",
1317 time);
1318 dump_state(devinfo, scoreboard->dag);
1319 fprintf(stderr, "t=%4d: chose: ", time);
1320 v3d_qpu_dump(devinfo, inst);
1321 fprintf(stderr, "\n");
1322 }
1323
1324 /* We can't mark_instruction_scheduled() the chosen inst until
1325 * we're done identifying instructions to merge, so put the
1326 * merged instructions on a list for a moment.
1327 */
1328 struct list_head merged_list;
1329 list_inithead(&merged_list);
1330
1331 /* Schedule this instruction onto the QPU list. Also try to
1332 * find an instruction to pair with it.
1333 */
1334 if (chosen) {
1335 time = MAX2(chosen->unblocked_time, time);
1336 pre_remove_head(scoreboard->dag, chosen);
1337
1338 while ((merge =
1339 choose_instruction_to_schedule(devinfo,
1340 scoreboard,
1341 chosen))) {
1342 time = MAX2(merge->unblocked_time, time);
1343 pre_remove_head(scoreboard->dag, chosen);
1344 list_addtail(&merge->link, &merged_list);
1345 (void)qpu_merge_inst(devinfo, inst,
1346 inst, &merge->inst->qpu);
1347 if (merge->inst->uniform != -1) {
1348 chosen->inst->uniform =
1349 merge->inst->uniform;
1350 }
1351
1352 if (debug) {
1353 fprintf(stderr, "t=%4d: merging: ",
1354 time);
1355 v3d_qpu_dump(devinfo, &merge->inst->qpu);
1356 fprintf(stderr, "\n");
1357 fprintf(stderr, " result: ");
1358 v3d_qpu_dump(devinfo, inst);
1359 fprintf(stderr, "\n");
1360 }
1361 }
1362 if (mux_read_stalls(scoreboard, inst))
1363 c->qpu_inst_stalled_count++;
1364 }
1365
1366 /* Update the uniform index for the rewritten location --
1367 * branch target updating will still need to change
1368 * c->uniform_data[] using this index.
1369 */
1370 if (qinst->uniform != -1) {
1371 if (inst->type == V3D_QPU_INSTR_TYPE_BRANCH)
1372 block->branch_uniform = *next_uniform;
1373
1374 c->uniform_data[*next_uniform] =
1375 orig_uniform_data[qinst->uniform];
1376 c->uniform_contents[*next_uniform] =
1377 orig_uniform_contents[qinst->uniform];
1378 qinst->uniform = *next_uniform;
1379 (*next_uniform)++;
1380 }
1381
1382 if (debug) {
1383 fprintf(stderr, "\n");
1384 }
1385
1386 /* Now that we've scheduled a new instruction, some of its
1387 * children can be promoted to the list of instructions ready to
1388 * be scheduled. Update the children's unblocked time for this
1389 * DAG edge as we do so.
1390 */
1391 mark_instruction_scheduled(scoreboard->dag, time, chosen);
1392 list_for_each_entry(struct schedule_node, merge, &merged_list,
1393 link) {
1394 mark_instruction_scheduled(scoreboard->dag, time, merge);
1395
1396 /* The merged VIR instruction doesn't get re-added to the
1397 * block, so free it now.
1398 */
1399 free(merge->inst);
1400 }
1401
1402 if (inst->sig.thrsw) {
1403 time += emit_thrsw(c, block, scoreboard, qinst, false);
1404 } else {
1405 insert_scheduled_instruction(c, block,
1406 scoreboard, qinst);
1407
1408 if (inst->type == V3D_QPU_INSTR_TYPE_BRANCH) {
1409 block->branch_qpu_ip = c->qpu_inst_count - 1;
1410 /* Fill the delay slots.
1411 *
1412 * We should fill these with actual instructions,
1413 * instead, but that will probably need to be done
1414 * after this, once we know what the leading
1415 * instructions of the successors are (so we can
1416 * handle A/B register file write latency)
1417 */
1418 for (int i = 0; i < 3; i++)
1419 emit_nop(c, block, scoreboard);
1420 }
1421 }
1422 }
1423
1424 return time;
1425 }
1426
1427 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)1428 qpu_schedule_instructions_block(struct v3d_compile *c,
1429 struct choose_scoreboard *scoreboard,
1430 struct qblock *block,
1431 enum quniform_contents *orig_uniform_contents,
1432 uint32_t *orig_uniform_data,
1433 uint32_t *next_uniform)
1434 {
1435 void *mem_ctx = ralloc_context(NULL);
1436 scoreboard->dag = dag_create(mem_ctx);
1437 struct list_head setup_list;
1438
1439 list_inithead(&setup_list);
1440
1441 /* Wrap each instruction in a scheduler structure. */
1442 while (!list_is_empty(&block->instructions)) {
1443 struct qinst *qinst = (struct qinst *)block->instructions.next;
1444 struct schedule_node *n =
1445 rzalloc(mem_ctx, struct schedule_node);
1446
1447 dag_init_node(scoreboard->dag, &n->dag);
1448 n->inst = qinst;
1449
1450 list_del(&qinst->link);
1451 list_addtail(&n->link, &setup_list);
1452 }
1453
1454 calculate_forward_deps(c, scoreboard->dag, &setup_list);
1455 calculate_reverse_deps(c, scoreboard->dag, &setup_list);
1456
1457 dag_traverse_bottom_up(scoreboard->dag, compute_delay, NULL);
1458
1459 uint32_t cycles = schedule_instructions(c, scoreboard, block,
1460 orig_uniform_contents,
1461 orig_uniform_data,
1462 next_uniform);
1463
1464 ralloc_free(mem_ctx);
1465 scoreboard->dag = NULL;
1466
1467 return cycles;
1468 }
1469
1470 static void
qpu_set_branch_targets(struct v3d_compile * c)1471 qpu_set_branch_targets(struct v3d_compile *c)
1472 {
1473 vir_for_each_block(block, c) {
1474 /* The end block of the program has no branch. */
1475 if (!block->successors[0])
1476 continue;
1477
1478 /* If there was no branch instruction, then the successor
1479 * block must follow immediately after this one.
1480 */
1481 if (block->branch_qpu_ip == ~0) {
1482 assert(block->end_qpu_ip + 1 ==
1483 block->successors[0]->start_qpu_ip);
1484 continue;
1485 }
1486
1487 /* Walk back through the delay slots to find the branch
1488 * instr.
1489 */
1490 struct list_head *entry = block->instructions.prev;
1491 for (int i = 0; i < 3; i++)
1492 entry = entry->prev;
1493 struct qinst *branch = container_of(entry, branch, link);
1494 assert(branch->qpu.type == V3D_QPU_INSTR_TYPE_BRANCH);
1495
1496 /* Make sure that the if-we-don't-jump
1497 * successor was scheduled just after the
1498 * delay slots.
1499 */
1500 assert(!block->successors[1] ||
1501 block->successors[1]->start_qpu_ip ==
1502 block->branch_qpu_ip + 4);
1503
1504 branch->qpu.branch.offset =
1505 ((block->successors[0]->start_qpu_ip -
1506 (block->branch_qpu_ip + 4)) *
1507 sizeof(uint64_t));
1508
1509 /* Set up the relative offset to jump in the
1510 * uniform stream.
1511 *
1512 * Use a temporary here, because
1513 * uniform_data[inst->uniform] may be shared
1514 * between multiple instructions.
1515 */
1516 assert(c->uniform_contents[branch->uniform] == QUNIFORM_CONSTANT);
1517 c->uniform_data[branch->uniform] =
1518 (block->successors[0]->start_uniform -
1519 (block->branch_uniform + 1)) * 4;
1520 }
1521 }
1522
1523 uint32_t
v3d_qpu_schedule_instructions(struct v3d_compile * c)1524 v3d_qpu_schedule_instructions(struct v3d_compile *c)
1525 {
1526 const struct v3d_device_info *devinfo = c->devinfo;
1527 struct qblock *end_block = list_last_entry(&c->blocks,
1528 struct qblock, link);
1529
1530 /* We reorder the uniforms as we schedule instructions, so save the
1531 * old data off and replace it.
1532 */
1533 uint32_t *uniform_data = c->uniform_data;
1534 enum quniform_contents *uniform_contents = c->uniform_contents;
1535 c->uniform_contents = ralloc_array(c, enum quniform_contents,
1536 c->num_uniforms);
1537 c->uniform_data = ralloc_array(c, uint32_t, c->num_uniforms);
1538 c->uniform_array_size = c->num_uniforms;
1539 uint32_t next_uniform = 0;
1540
1541 struct choose_scoreboard scoreboard;
1542 memset(&scoreboard, 0, sizeof(scoreboard));
1543 scoreboard.last_ldvary_tick = -10;
1544 scoreboard.last_magic_sfu_write_tick = -10;
1545 scoreboard.last_uniforms_reset_tick = -10;
1546 scoreboard.last_thrsw_tick = -10;
1547 scoreboard.last_stallable_sfu_tick = -10;
1548
1549 if (debug) {
1550 fprintf(stderr, "Pre-schedule instructions\n");
1551 vir_for_each_block(block, c) {
1552 fprintf(stderr, "BLOCK %d\n", block->index);
1553 list_for_each_entry(struct qinst, qinst,
1554 &block->instructions, link) {
1555 v3d_qpu_dump(devinfo, &qinst->qpu);
1556 fprintf(stderr, "\n");
1557 }
1558 }
1559 fprintf(stderr, "\n");
1560 }
1561
1562 uint32_t cycles = 0;
1563 vir_for_each_block(block, c) {
1564 block->start_qpu_ip = c->qpu_inst_count;
1565 block->branch_qpu_ip = ~0;
1566 block->start_uniform = next_uniform;
1567
1568 cycles += qpu_schedule_instructions_block(c,
1569 &scoreboard,
1570 block,
1571 uniform_contents,
1572 uniform_data,
1573 &next_uniform);
1574
1575 block->end_qpu_ip = c->qpu_inst_count - 1;
1576 }
1577
1578 /* Emit the program-end THRSW instruction. */;
1579 struct qinst *thrsw = vir_nop();
1580 thrsw->qpu.sig.thrsw = true;
1581 emit_thrsw(c, end_block, &scoreboard, thrsw, true);
1582
1583 qpu_set_branch_targets(c);
1584
1585 assert(next_uniform == c->num_uniforms);
1586
1587 return cycles;
1588 }
1589