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
41 static bool debug;
42
43 struct schedule_node_child;
44
45 struct schedule_node {
46 struct list_head link;
47 struct qinst *inst;
48 struct schedule_node_child *children;
49 uint32_t child_count;
50 uint32_t child_array_size;
51 uint32_t parent_count;
52
53 /* Longest cycles + instruction_latency() of any parent of this node. */
54 uint32_t unblocked_time;
55
56 /**
57 * Minimum number of cycles from scheduling this instruction until the
58 * end of the program, based on the slowest dependency chain through
59 * the children.
60 */
61 uint32_t delay;
62
63 /**
64 * cycles between this instruction being scheduled and when its result
65 * can be consumed.
66 */
67 uint32_t latency;
68 };
69
70 struct schedule_node_child {
71 struct schedule_node *node;
72 bool write_after_read;
73 };
74
75 /* When walking the instructions in reverse, we need to swap before/after in
76 * add_dep().
77 */
78 enum direction { F, R };
79
80 struct schedule_state {
81 const struct v3d_device_info *devinfo;
82 struct schedule_node *last_r[6];
83 struct schedule_node *last_rf[64];
84 struct schedule_node *last_sf;
85 struct schedule_node *last_vpm_read;
86 struct schedule_node *last_tmu_write;
87 struct schedule_node *last_tmu_config;
88 struct schedule_node *last_tlb;
89 struct schedule_node *last_vpm;
90 struct schedule_node *last_unif;
91 struct schedule_node *last_rtop;
92 enum direction dir;
93 /* Estimated cycle when the current instruction would start. */
94 uint32_t time;
95 };
96
97 static void
add_dep(struct schedule_state * state,struct schedule_node * before,struct schedule_node * after,bool write)98 add_dep(struct schedule_state *state,
99 struct schedule_node *before,
100 struct schedule_node *after,
101 bool write)
102 {
103 bool write_after_read = !write && state->dir == R;
104
105 if (!before || !after)
106 return;
107
108 assert(before != after);
109
110 if (state->dir == R) {
111 struct schedule_node *t = before;
112 before = after;
113 after = t;
114 }
115
116 for (int i = 0; i < before->child_count; i++) {
117 if (before->children[i].node == after &&
118 (before->children[i].write_after_read == write_after_read)) {
119 return;
120 }
121 }
122
123 if (before->child_array_size <= before->child_count) {
124 before->child_array_size = MAX2(before->child_array_size * 2, 16);
125 before->children = reralloc(before, before->children,
126 struct schedule_node_child,
127 before->child_array_size);
128 }
129
130 before->children[before->child_count].node = after;
131 before->children[before->child_count].write_after_read =
132 write_after_read;
133 before->child_count++;
134 after->parent_count++;
135 }
136
137 static void
add_read_dep(struct schedule_state * state,struct schedule_node * before,struct schedule_node * after)138 add_read_dep(struct schedule_state *state,
139 struct schedule_node *before,
140 struct schedule_node *after)
141 {
142 add_dep(state, before, after, false);
143 }
144
145 static void
add_write_dep(struct schedule_state * state,struct schedule_node ** before,struct schedule_node * after)146 add_write_dep(struct schedule_state *state,
147 struct schedule_node **before,
148 struct schedule_node *after)
149 {
150 add_dep(state, *before, after, true);
151 *before = after;
152 }
153
154 static bool
qpu_inst_is_tlb(const struct v3d_qpu_instr * inst)155 qpu_inst_is_tlb(const struct v3d_qpu_instr *inst)
156 {
157 if (inst->type != V3D_QPU_INSTR_TYPE_ALU)
158 return false;
159
160 if (inst->alu.add.magic_write &&
161 (inst->alu.add.waddr == V3D_QPU_WADDR_TLB ||
162 inst->alu.add.waddr == V3D_QPU_WADDR_TLBU))
163 return true;
164
165 if (inst->alu.mul.magic_write &&
166 (inst->alu.mul.waddr == V3D_QPU_WADDR_TLB ||
167 inst->alu.mul.waddr == V3D_QPU_WADDR_TLBU))
168 return true;
169
170 return false;
171 }
172
173 static void
process_mux_deps(struct schedule_state * state,struct schedule_node * n,enum v3d_qpu_mux mux)174 process_mux_deps(struct schedule_state *state, struct schedule_node *n,
175 enum v3d_qpu_mux mux)
176 {
177 switch (mux) {
178 case V3D_QPU_MUX_A:
179 add_read_dep(state, state->last_rf[n->inst->qpu.raddr_a], n);
180 break;
181 case V3D_QPU_MUX_B:
182 add_read_dep(state, state->last_rf[n->inst->qpu.raddr_b], n);
183 break;
184 default:
185 add_read_dep(state, state->last_r[mux - V3D_QPU_MUX_R0], n);
186 break;
187 }
188 }
189
190
191 static void
process_waddr_deps(struct schedule_state * state,struct schedule_node * n,uint32_t waddr,bool magic)192 process_waddr_deps(struct schedule_state *state, struct schedule_node *n,
193 uint32_t waddr, bool magic)
194 {
195 if (!magic) {
196 add_write_dep(state, &state->last_rf[waddr], n);
197 } else if (v3d_qpu_magic_waddr_is_tmu(waddr)) {
198 add_write_dep(state, &state->last_tmu_write, n);
199 switch (waddr) {
200 case V3D_QPU_WADDR_TMUS:
201 case V3D_QPU_WADDR_TMUSCM:
202 case V3D_QPU_WADDR_TMUSF:
203 case V3D_QPU_WADDR_TMUSLOD:
204 add_write_dep(state, &state->last_tmu_config, n);
205 break;
206 default:
207 break;
208 }
209 } else if (v3d_qpu_magic_waddr_is_sfu(waddr)) {
210 /* Handled by v3d_qpu_writes_r4() check. */
211 } else {
212 switch (waddr) {
213 case V3D_QPU_WADDR_R0:
214 case V3D_QPU_WADDR_R1:
215 case V3D_QPU_WADDR_R2:
216 add_write_dep(state,
217 &state->last_r[waddr - V3D_QPU_WADDR_R0],
218 n);
219 break;
220 case V3D_QPU_WADDR_R3:
221 case V3D_QPU_WADDR_R4:
222 case V3D_QPU_WADDR_R5:
223 /* Handled by v3d_qpu_writes_r*() checks below. */
224 break;
225
226 case V3D_QPU_WADDR_VPM:
227 case V3D_QPU_WADDR_VPMU:
228 add_write_dep(state, &state->last_vpm, n);
229 break;
230
231 case V3D_QPU_WADDR_TLB:
232 case V3D_QPU_WADDR_TLBU:
233 add_write_dep(state, &state->last_tlb, n);
234 break;
235
236 case V3D_QPU_WADDR_NOP:
237 break;
238
239 default:
240 fprintf(stderr, "Unknown waddr %d\n", waddr);
241 abort();
242 }
243 }
244 }
245
246 static void
process_cond_deps(struct schedule_state * state,struct schedule_node * n,enum v3d_qpu_cond cond)247 process_cond_deps(struct schedule_state *state, struct schedule_node *n,
248 enum v3d_qpu_cond cond)
249 {
250 if (cond != V3D_QPU_COND_NONE)
251 add_read_dep(state, state->last_sf, n);
252 }
253
254 static void
process_pf_deps(struct schedule_state * state,struct schedule_node * n,enum v3d_qpu_pf pf)255 process_pf_deps(struct schedule_state *state, struct schedule_node *n,
256 enum v3d_qpu_pf pf)
257 {
258 if (pf != V3D_QPU_PF_NONE)
259 add_write_dep(state, &state->last_sf, n);
260 }
261
262 static void
process_uf_deps(struct schedule_state * state,struct schedule_node * n,enum v3d_qpu_uf uf)263 process_uf_deps(struct schedule_state *state, struct schedule_node *n,
264 enum v3d_qpu_uf uf)
265 {
266 if (uf != V3D_QPU_UF_NONE)
267 add_write_dep(state, &state->last_sf, n);
268 }
269
270 /**
271 * Common code for dependencies that need to be tracked both forward and
272 * backward.
273 *
274 * This is for things like "all reads of r4 have to happen between the r4
275 * writes that surround them".
276 */
277 static void
calculate_deps(struct schedule_state * state,struct schedule_node * n)278 calculate_deps(struct schedule_state *state, struct schedule_node *n)
279 {
280 const struct v3d_device_info *devinfo = state->devinfo;
281 struct qinst *qinst = n->inst;
282 struct v3d_qpu_instr *inst = &qinst->qpu;
283
284 if (inst->type == V3D_QPU_INSTR_TYPE_BRANCH) {
285 if (inst->branch.cond != V3D_QPU_BRANCH_COND_ALWAYS)
286 add_read_dep(state, state->last_sf, n);
287
288 /* XXX: BDI */
289 /* XXX: BDU */
290 /* XXX: ub */
291 /* XXX: raddr_a */
292
293 add_write_dep(state, &state->last_unif, n);
294 return;
295 }
296
297 assert(inst->type == V3D_QPU_INSTR_TYPE_ALU);
298
299 /* XXX: LOAD_IMM */
300
301 if (v3d_qpu_add_op_num_src(inst->alu.add.op) > 0)
302 process_mux_deps(state, n, inst->alu.add.a);
303 if (v3d_qpu_add_op_num_src(inst->alu.add.op) > 1)
304 process_mux_deps(state, n, inst->alu.add.b);
305
306 if (v3d_qpu_mul_op_num_src(inst->alu.mul.op) > 0)
307 process_mux_deps(state, n, inst->alu.mul.a);
308 if (v3d_qpu_mul_op_num_src(inst->alu.mul.op) > 1)
309 process_mux_deps(state, n, inst->alu.mul.b);
310
311 switch (inst->alu.add.op) {
312 case V3D_QPU_A_VPMSETUP:
313 /* Could distinguish read/write by unpacking the uniform. */
314 add_write_dep(state, &state->last_vpm, n);
315 add_write_dep(state, &state->last_vpm_read, n);
316 break;
317
318 case V3D_QPU_A_STVPMV:
319 case V3D_QPU_A_STVPMD:
320 case V3D_QPU_A_STVPMP:
321 add_write_dep(state, &state->last_vpm, n);
322 break;
323
324 case V3D_QPU_A_VPMWT:
325 add_read_dep(state, state->last_vpm, n);
326 break;
327
328 case V3D_QPU_A_MSF:
329 add_read_dep(state, state->last_tlb, n);
330 break;
331
332 case V3D_QPU_A_SETMSF:
333 case V3D_QPU_A_SETREVF:
334 add_write_dep(state, &state->last_tlb, n);
335 break;
336
337 case V3D_QPU_A_FLAPUSH:
338 case V3D_QPU_A_FLBPUSH:
339 case V3D_QPU_A_VFLA:
340 case V3D_QPU_A_VFLNA:
341 case V3D_QPU_A_VFLB:
342 case V3D_QPU_A_VFLNB:
343 add_read_dep(state, state->last_sf, n);
344 break;
345
346 case V3D_QPU_A_FLBPOP:
347 add_write_dep(state, &state->last_sf, n);
348 break;
349
350 default:
351 break;
352 }
353
354 switch (inst->alu.mul.op) {
355 case V3D_QPU_M_MULTOP:
356 case V3D_QPU_M_UMUL24:
357 /* MULTOP sets rtop, and UMUL24 implicitly reads rtop and
358 * resets it to 0. We could possibly reorder umul24s relative
359 * to each other, but for now just keep all the MUL parts in
360 * order.
361 */
362 add_write_dep(state, &state->last_rtop, n);
363 break;
364 default:
365 break;
366 }
367
368 if (inst->alu.add.op != V3D_QPU_A_NOP) {
369 process_waddr_deps(state, n, inst->alu.add.waddr,
370 inst->alu.add.magic_write);
371 }
372 if (inst->alu.mul.op != V3D_QPU_M_NOP) {
373 process_waddr_deps(state, n, inst->alu.mul.waddr,
374 inst->alu.mul.magic_write);
375 }
376 if (v3d_qpu_sig_writes_address(devinfo, &inst->sig)) {
377 process_waddr_deps(state, n, inst->sig_addr,
378 inst->sig_magic);
379 }
380
381 if (v3d_qpu_writes_r3(devinfo, inst))
382 add_write_dep(state, &state->last_r[3], n);
383 if (v3d_qpu_writes_r4(devinfo, inst))
384 add_write_dep(state, &state->last_r[4], n);
385 if (v3d_qpu_writes_r5(devinfo, inst))
386 add_write_dep(state, &state->last_r[5], n);
387
388 if (inst->sig.thrsw) {
389 /* All accumulator contents and flags are undefined after the
390 * switch.
391 */
392 for (int i = 0; i < ARRAY_SIZE(state->last_r); i++)
393 add_write_dep(state, &state->last_r[i], n);
394 add_write_dep(state, &state->last_sf, n);
395
396 /* Scoreboard-locking operations have to stay after the last
397 * thread switch.
398 */
399 add_write_dep(state, &state->last_tlb, n);
400
401 add_write_dep(state, &state->last_tmu_write, n);
402 add_write_dep(state, &state->last_tmu_config, n);
403 }
404
405 if (inst->sig.ldtmu) {
406 /* TMU loads are coming from a FIFO, so ordering is important.
407 */
408 add_write_dep(state, &state->last_tmu_write, n);
409 }
410
411 if (inst->sig.wrtmuc)
412 add_write_dep(state, &state->last_tmu_config, n);
413
414 if (inst->sig.ldtlb | inst->sig.ldtlbu)
415 add_read_dep(state, state->last_tlb, n);
416
417 if (inst->sig.ldvpm)
418 add_write_dep(state, &state->last_vpm_read, n);
419
420 /* inst->sig.ldunif or sideband uniform read */
421 if (qinst->uniform != ~0)
422 add_write_dep(state, &state->last_unif, n);
423
424 process_cond_deps(state, n, inst->flags.ac);
425 process_cond_deps(state, n, inst->flags.mc);
426 process_pf_deps(state, n, inst->flags.apf);
427 process_pf_deps(state, n, inst->flags.mpf);
428 process_uf_deps(state, n, inst->flags.auf);
429 process_uf_deps(state, n, inst->flags.muf);
430 }
431
432 static void
calculate_forward_deps(struct v3d_compile * c,struct list_head * schedule_list)433 calculate_forward_deps(struct v3d_compile *c, struct list_head *schedule_list)
434 {
435 struct schedule_state state;
436
437 memset(&state, 0, sizeof(state));
438 state.devinfo = c->devinfo;
439 state.dir = F;
440
441 list_for_each_entry(struct schedule_node, node, schedule_list, link)
442 calculate_deps(&state, node);
443 }
444
445 static void
calculate_reverse_deps(struct v3d_compile * c,struct list_head * schedule_list)446 calculate_reverse_deps(struct v3d_compile *c, struct list_head *schedule_list)
447 {
448 struct list_head *node;
449 struct schedule_state state;
450
451 memset(&state, 0, sizeof(state));
452 state.devinfo = c->devinfo;
453 state.dir = R;
454
455 for (node = schedule_list->prev; schedule_list != node; node = node->prev) {
456 calculate_deps(&state, (struct schedule_node *)node);
457 }
458 }
459
460 struct choose_scoreboard {
461 int tick;
462 int last_sfu_write_tick;
463 int last_ldvary_tick;
464 int last_uniforms_reset_tick;
465 uint32_t last_waddr_add, last_waddr_mul;
466 bool tlb_locked;
467 };
468
469 static bool
mux_reads_too_soon(struct choose_scoreboard * scoreboard,const struct v3d_qpu_instr * inst,enum v3d_qpu_mux mux)470 mux_reads_too_soon(struct choose_scoreboard *scoreboard,
471 const struct v3d_qpu_instr *inst, enum v3d_qpu_mux mux)
472 {
473 switch (mux) {
474 case V3D_QPU_MUX_A:
475 if (scoreboard->last_waddr_add == inst->raddr_a ||
476 scoreboard->last_waddr_mul == inst->raddr_a) {
477 return true;
478 }
479 break;
480
481 case V3D_QPU_MUX_B:
482 if (scoreboard->last_waddr_add == inst->raddr_b ||
483 scoreboard->last_waddr_mul == inst->raddr_b) {
484 return true;
485 }
486 break;
487
488 case V3D_QPU_MUX_R4:
489 if (scoreboard->tick - scoreboard->last_sfu_write_tick <= 2)
490 return true;
491 break;
492
493 case V3D_QPU_MUX_R5:
494 if (scoreboard->tick - scoreboard->last_ldvary_tick <= 1)
495 return true;
496 break;
497 default:
498 break;
499 }
500
501 return false;
502 }
503
504 static bool
reads_too_soon_after_write(struct choose_scoreboard * scoreboard,struct qinst * qinst)505 reads_too_soon_after_write(struct choose_scoreboard *scoreboard,
506 struct qinst *qinst)
507 {
508 const struct v3d_qpu_instr *inst = &qinst->qpu;
509
510 /* XXX: Branching off of raddr. */
511 if (inst->type == V3D_QPU_INSTR_TYPE_BRANCH)
512 return false;
513
514 assert(inst->type == V3D_QPU_INSTR_TYPE_ALU);
515
516 if (inst->alu.add.op != V3D_QPU_A_NOP) {
517 if (v3d_qpu_add_op_num_src(inst->alu.add.op) > 0 &&
518 mux_reads_too_soon(scoreboard, inst, inst->alu.add.a)) {
519 return true;
520 }
521 if (v3d_qpu_add_op_num_src(inst->alu.add.op) > 1 &&
522 mux_reads_too_soon(scoreboard, inst, inst->alu.add.b)) {
523 return true;
524 }
525 }
526
527 if (inst->alu.mul.op != V3D_QPU_M_NOP) {
528 if (v3d_qpu_mul_op_num_src(inst->alu.mul.op) > 0 &&
529 mux_reads_too_soon(scoreboard, inst, inst->alu.mul.a)) {
530 return true;
531 }
532 if (v3d_qpu_mul_op_num_src(inst->alu.mul.op) > 1 &&
533 mux_reads_too_soon(scoreboard, inst, inst->alu.mul.b)) {
534 return true;
535 }
536 }
537
538 /* XXX: imm */
539
540 return false;
541 }
542
543 static bool
writes_too_soon_after_write(const struct v3d_device_info * devinfo,struct choose_scoreboard * scoreboard,struct qinst * qinst)544 writes_too_soon_after_write(const struct v3d_device_info *devinfo,
545 struct choose_scoreboard *scoreboard,
546 struct qinst *qinst)
547 {
548 const struct v3d_qpu_instr *inst = &qinst->qpu;
549
550 /* Don't schedule any other r4 write too soon after an SFU write.
551 * This would normally be prevented by dependency tracking, but might
552 * occur if a dead SFU computation makes it to scheduling.
553 */
554 if (scoreboard->tick - scoreboard->last_sfu_write_tick < 2 &&
555 v3d_qpu_writes_r4(devinfo, inst))
556 return true;
557
558 return false;
559 }
560
561 static bool
pixel_scoreboard_too_soon(struct choose_scoreboard * scoreboard,const struct v3d_qpu_instr * inst)562 pixel_scoreboard_too_soon(struct choose_scoreboard *scoreboard,
563 const struct v3d_qpu_instr *inst)
564 {
565 return (scoreboard->tick == 0 && qpu_inst_is_tlb(inst));
566 }
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 (inst->sig.ldtmu)
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 (inst->type == V3D_QPU_INSTR_TYPE_ALU &&
592 ((inst->alu.add.magic_write &&
593 v3d_qpu_magic_waddr_is_tmu(inst->alu.add.waddr)) ||
594 (inst->alu.mul.magic_write &&
595 v3d_qpu_magic_waddr_is_tmu(inst->alu.mul.waddr)))) {
596 return next_score;
597 }
598 next_score++;
599
600 return baseline_score;
601 }
602
603 static bool
qpu_magic_waddr_is_periph(enum v3d_qpu_waddr waddr)604 qpu_magic_waddr_is_periph(enum v3d_qpu_waddr waddr)
605 {
606 return (v3d_qpu_magic_waddr_is_tmu(waddr) ||
607 v3d_qpu_magic_waddr_is_sfu(waddr) ||
608 v3d_qpu_magic_waddr_is_tlb(waddr) ||
609 v3d_qpu_magic_waddr_is_vpm(waddr) ||
610 v3d_qpu_magic_waddr_is_tsy(waddr));
611 }
612
613 static bool
qpu_accesses_peripheral(const struct v3d_qpu_instr * inst)614 qpu_accesses_peripheral(const struct v3d_qpu_instr *inst)
615 {
616 if (v3d_qpu_uses_vpm(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.mul.op != V3D_QPU_M_NOP &&
627 inst->alu.mul.magic_write &&
628 qpu_magic_waddr_is_periph(inst->alu.mul.waddr)) {
629 return true;
630 }
631 }
632
633 return (inst->sig.ldvpm ||
634 inst->sig.ldtmu ||
635 inst->sig.ldtlb ||
636 inst->sig.ldtlbu ||
637 inst->sig.wrtmuc);
638 }
639
640 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)641 qpu_merge_inst(const struct v3d_device_info *devinfo,
642 struct v3d_qpu_instr *result,
643 const struct v3d_qpu_instr *a,
644 const struct v3d_qpu_instr *b)
645 {
646 if (a->type != V3D_QPU_INSTR_TYPE_ALU ||
647 b->type != V3D_QPU_INSTR_TYPE_ALU) {
648 return false;
649 }
650
651 /* Can't do more than one peripheral access in an instruction.
652 *
653 * XXX: V3D 4.1 allows TMU read along with a VPM read or write, and
654 * WRTMUC with a TMU magic register write (other than tmuc).
655 */
656 if (qpu_accesses_peripheral(a) && qpu_accesses_peripheral(b))
657 return false;
658
659 struct v3d_qpu_instr merge = *a;
660
661 if (b->alu.add.op != V3D_QPU_A_NOP) {
662 if (a->alu.add.op != V3D_QPU_A_NOP)
663 return false;
664 merge.alu.add = b->alu.add;
665
666 merge.flags.ac = b->flags.ac;
667 merge.flags.apf = b->flags.apf;
668 merge.flags.auf = b->flags.auf;
669 }
670
671 if (b->alu.mul.op != V3D_QPU_M_NOP) {
672 if (a->alu.mul.op != V3D_QPU_M_NOP)
673 return false;
674 merge.alu.mul = b->alu.mul;
675
676 merge.flags.mc = b->flags.mc;
677 merge.flags.mpf = b->flags.mpf;
678 merge.flags.muf = b->flags.muf;
679 }
680
681 if (v3d_qpu_uses_mux(b, V3D_QPU_MUX_A)) {
682 if (v3d_qpu_uses_mux(a, V3D_QPU_MUX_A) &&
683 a->raddr_a != b->raddr_a) {
684 return false;
685 }
686 merge.raddr_a = b->raddr_a;
687 }
688
689 if (v3d_qpu_uses_mux(b, V3D_QPU_MUX_B)) {
690 if (v3d_qpu_uses_mux(a, V3D_QPU_MUX_B) &&
691 a->raddr_b != b->raddr_b) {
692 return false;
693 }
694 merge.raddr_b = b->raddr_b;
695 }
696
697 merge.sig.thrsw |= b->sig.thrsw;
698 merge.sig.ldunif |= b->sig.ldunif;
699 merge.sig.ldunifrf |= b->sig.ldunifrf;
700 merge.sig.ldunifa |= b->sig.ldunifa;
701 merge.sig.ldunifarf |= b->sig.ldunifarf;
702 merge.sig.ldtmu |= b->sig.ldtmu;
703 merge.sig.ldvary |= b->sig.ldvary;
704 merge.sig.ldvpm |= b->sig.ldvpm;
705 merge.sig.small_imm |= b->sig.small_imm;
706 merge.sig.ldtlb |= b->sig.ldtlb;
707 merge.sig.ldtlbu |= b->sig.ldtlbu;
708 merge.sig.ucb |= b->sig.ucb;
709 merge.sig.rotate |= b->sig.rotate;
710 merge.sig.wrtmuc |= b->sig.wrtmuc;
711
712 if (v3d_qpu_sig_writes_address(devinfo, &a->sig) &&
713 v3d_qpu_sig_writes_address(devinfo, &b->sig))
714 return false;
715 merge.sig_addr |= b->sig_addr;
716 merge.sig_magic |= b->sig_magic;
717
718 uint64_t packed;
719 bool ok = v3d_qpu_instr_pack(devinfo, &merge, &packed);
720
721 *result = merge;
722 /* No modifying the real instructions on failure. */
723 assert(ok || (a != result && b != result));
724
725 return ok;
726 }
727
728 static struct schedule_node *
choose_instruction_to_schedule(const struct v3d_device_info * devinfo,struct choose_scoreboard * scoreboard,struct list_head * schedule_list,struct schedule_node * prev_inst)729 choose_instruction_to_schedule(const struct v3d_device_info *devinfo,
730 struct choose_scoreboard *scoreboard,
731 struct list_head *schedule_list,
732 struct schedule_node *prev_inst)
733 {
734 struct schedule_node *chosen = NULL;
735 int chosen_prio = 0;
736
737 /* Don't pair up anything with a thread switch signal -- emit_thrsw()
738 * will handle pairing it along with filling the delay slots.
739 */
740 if (prev_inst) {
741 if (prev_inst->inst->qpu.sig.thrsw)
742 return NULL;
743 }
744
745 list_for_each_entry(struct schedule_node, n, schedule_list, link) {
746 const struct v3d_qpu_instr *inst = &n->inst->qpu;
747
748 /* Don't choose the branch instruction until it's the last one
749 * left. We'll move it up to fit its delay slots after we
750 * choose it.
751 */
752 if (inst->type == V3D_QPU_INSTR_TYPE_BRANCH &&
753 !list_is_singular(schedule_list)) {
754 continue;
755 }
756
757 /* "An instruction must not read from a location in physical
758 * regfile A or B that was written to by the previous
759 * instruction."
760 */
761 if (reads_too_soon_after_write(scoreboard, n->inst))
762 continue;
763
764 if (writes_too_soon_after_write(devinfo, scoreboard, n->inst))
765 continue;
766
767 /* "A scoreboard wait must not occur in the first two
768 * instructions of a fragment shader. This is either the
769 * explicit Wait for Scoreboard signal or an implicit wait
770 * with the first tile-buffer read or write instruction."
771 */
772 if (pixel_scoreboard_too_soon(scoreboard, inst))
773 continue;
774
775 /* ldunif and ldvary both write r5, but ldunif does so a tick
776 * sooner. If the ldvary's r5 wasn't used, then ldunif might
777 * otherwise get scheduled so ldunif and ldvary try to update
778 * r5 in the same tick.
779 */
780 if ((inst->sig.ldunif || inst->sig.ldunifa) &&
781 scoreboard->tick == scoreboard->last_ldvary_tick + 1) {
782 continue;
783 }
784
785 /* If we're trying to pair with another instruction, check
786 * that they're compatible.
787 */
788 if (prev_inst) {
789 /* Don't pair up a thread switch signal -- we'll
790 * handle pairing it when we pick it on its own.
791 */
792 if (inst->sig.thrsw)
793 continue;
794
795 if (prev_inst->inst->uniform != -1 &&
796 n->inst->uniform != -1)
797 continue;
798
799 /* Don't merge in something that will lock the TLB.
800 * Hopwefully what we have in inst will release some
801 * other instructions, allowing us to delay the
802 * TLB-locking instruction until later.
803 */
804 if (!scoreboard->tlb_locked && qpu_inst_is_tlb(inst))
805 continue;
806
807 struct v3d_qpu_instr merged_inst;
808 if (!qpu_merge_inst(devinfo, &merged_inst,
809 &prev_inst->inst->qpu, inst)) {
810 continue;
811 }
812 }
813
814 int prio = get_instruction_priority(inst);
815
816 /* Found a valid instruction. If nothing better comes along,
817 * this one works.
818 */
819 if (!chosen) {
820 chosen = n;
821 chosen_prio = prio;
822 continue;
823 }
824
825 if (prio > chosen_prio) {
826 chosen = n;
827 chosen_prio = prio;
828 } else if (prio < chosen_prio) {
829 continue;
830 }
831
832 if (n->delay > chosen->delay) {
833 chosen = n;
834 chosen_prio = prio;
835 } else if (n->delay < chosen->delay) {
836 continue;
837 }
838 }
839
840 return chosen;
841 }
842
843 static void
update_scoreboard_for_magic_waddr(struct choose_scoreboard * scoreboard,enum v3d_qpu_waddr waddr)844 update_scoreboard_for_magic_waddr(struct choose_scoreboard *scoreboard,
845 enum v3d_qpu_waddr waddr)
846 {
847 if (v3d_qpu_magic_waddr_is_sfu(waddr))
848 scoreboard->last_sfu_write_tick = scoreboard->tick;
849 }
850
851 static void
update_scoreboard_for_chosen(struct choose_scoreboard * scoreboard,const struct v3d_qpu_instr * inst)852 update_scoreboard_for_chosen(struct choose_scoreboard *scoreboard,
853 const struct v3d_qpu_instr *inst)
854 {
855 scoreboard->last_waddr_add = ~0;
856 scoreboard->last_waddr_mul = ~0;
857
858 if (inst->type == V3D_QPU_INSTR_TYPE_BRANCH)
859 return;
860
861 assert(inst->type == V3D_QPU_INSTR_TYPE_ALU);
862
863 if (inst->alu.add.op != V3D_QPU_A_NOP) {
864 if (inst->alu.add.magic_write) {
865 update_scoreboard_for_magic_waddr(scoreboard,
866 inst->alu.add.waddr);
867 } else {
868 scoreboard->last_waddr_add = inst->alu.add.waddr;
869 }
870 }
871
872 if (inst->alu.mul.op != V3D_QPU_M_NOP) {
873 if (inst->alu.mul.magic_write) {
874 update_scoreboard_for_magic_waddr(scoreboard,
875 inst->alu.mul.waddr);
876 } else {
877 scoreboard->last_waddr_mul = inst->alu.mul.waddr;
878 }
879 }
880
881 if (inst->sig.ldvary)
882 scoreboard->last_ldvary_tick = scoreboard->tick;
883
884 if (qpu_inst_is_tlb(inst))
885 scoreboard->tlb_locked = true;
886 }
887
888 static void
dump_state(const struct v3d_device_info * devinfo,struct list_head * schedule_list)889 dump_state(const struct v3d_device_info *devinfo,
890 struct list_head *schedule_list)
891 {
892 list_for_each_entry(struct schedule_node, n, schedule_list, link) {
893 fprintf(stderr, " t=%4d: ", n->unblocked_time);
894 v3d_qpu_dump(devinfo, &n->inst->qpu);
895 fprintf(stderr, "\n");
896
897 for (int i = 0; i < n->child_count; i++) {
898 struct schedule_node *child = n->children[i].node;
899 if (!child)
900 continue;
901
902 fprintf(stderr, " - ");
903 v3d_qpu_dump(devinfo, &child->inst->qpu);
904 fprintf(stderr, " (%d parents, %c)\n",
905 child->parent_count,
906 n->children[i].write_after_read ? 'w' : 'r');
907 }
908 }
909 }
910
magic_waddr_latency(enum v3d_qpu_waddr waddr,const struct v3d_qpu_instr * after)911 static uint32_t magic_waddr_latency(enum v3d_qpu_waddr waddr,
912 const struct v3d_qpu_instr *after)
913 {
914 /* Apply some huge latency between texture fetch requests and getting
915 * their results back.
916 *
917 * FIXME: This is actually pretty bogus. If we do:
918 *
919 * mov tmu0_s, a
920 * <a bit of math>
921 * mov tmu0_s, b
922 * load_tmu0
923 * <more math>
924 * load_tmu0
925 *
926 * we count that as worse than
927 *
928 * mov tmu0_s, a
929 * mov tmu0_s, b
930 * <lots of math>
931 * load_tmu0
932 * <more math>
933 * load_tmu0
934 *
935 * because we associate the first load_tmu0 with the *second* tmu0_s.
936 */
937 if (v3d_qpu_magic_waddr_is_tmu(waddr) && after->sig.ldtmu)
938 return 100;
939
940 /* Assume that anything depending on us is consuming the SFU result. */
941 if (v3d_qpu_magic_waddr_is_sfu(waddr))
942 return 3;
943
944 return 1;
945 }
946
947 static uint32_t
instruction_latency(struct schedule_node * before,struct schedule_node * after)948 instruction_latency(struct schedule_node *before, struct schedule_node *after)
949 {
950 const struct v3d_qpu_instr *before_inst = &before->inst->qpu;
951 const struct v3d_qpu_instr *after_inst = &after->inst->qpu;
952 uint32_t latency = 1;
953
954 if (before_inst->type != V3D_QPU_INSTR_TYPE_ALU ||
955 after_inst->type != V3D_QPU_INSTR_TYPE_ALU)
956 return latency;
957
958 if (before_inst->alu.add.magic_write) {
959 latency = MAX2(latency,
960 magic_waddr_latency(before_inst->alu.add.waddr,
961 after_inst));
962 }
963
964 if (before_inst->alu.mul.magic_write) {
965 latency = MAX2(latency,
966 magic_waddr_latency(before_inst->alu.mul.waddr,
967 after_inst));
968 }
969
970 return latency;
971 }
972
973 /** Recursive computation of the delay member of a node. */
974 static void
compute_delay(struct schedule_node * n)975 compute_delay(struct schedule_node *n)
976 {
977 if (!n->child_count) {
978 n->delay = 1;
979 } else {
980 for (int i = 0; i < n->child_count; i++) {
981 if (!n->children[i].node->delay)
982 compute_delay(n->children[i].node);
983 n->delay = MAX2(n->delay,
984 n->children[i].node->delay +
985 instruction_latency(n, n->children[i].node));
986 }
987 }
988 }
989
990 static void
mark_instruction_scheduled(struct list_head * schedule_list,uint32_t time,struct schedule_node * node,bool war_only)991 mark_instruction_scheduled(struct list_head *schedule_list,
992 uint32_t time,
993 struct schedule_node *node,
994 bool war_only)
995 {
996 if (!node)
997 return;
998
999 for (int i = node->child_count - 1; i >= 0; i--) {
1000 struct schedule_node *child =
1001 node->children[i].node;
1002
1003 if (!child)
1004 continue;
1005
1006 if (war_only && !node->children[i].write_after_read)
1007 continue;
1008
1009 /* If the requirement is only that the node not appear before
1010 * the last read of its destination, then it can be scheduled
1011 * immediately after (or paired with!) the thing reading the
1012 * destination.
1013 */
1014 uint32_t latency = 0;
1015 if (!war_only) {
1016 latency = instruction_latency(node,
1017 node->children[i].node);
1018 }
1019
1020 child->unblocked_time = MAX2(child->unblocked_time,
1021 time + latency);
1022 child->parent_count--;
1023 if (child->parent_count == 0)
1024 list_add(&child->link, schedule_list);
1025
1026 node->children[i].node = NULL;
1027 }
1028 }
1029
1030 static void
insert_scheduled_instruction(struct v3d_compile * c,struct qblock * block,struct choose_scoreboard * scoreboard,struct qinst * inst)1031 insert_scheduled_instruction(struct v3d_compile *c,
1032 struct qblock *block,
1033 struct choose_scoreboard *scoreboard,
1034 struct qinst *inst)
1035 {
1036 list_addtail(&inst->link, &block->instructions);
1037
1038 update_scoreboard_for_chosen(scoreboard, &inst->qpu);
1039 c->qpu_inst_count++;
1040 scoreboard->tick++;
1041 }
1042
1043 static struct qinst *
vir_nop()1044 vir_nop()
1045 {
1046 struct qreg undef = { QFILE_NULL, 0 };
1047 struct qinst *qinst = vir_add_inst(V3D_QPU_A_NOP, undef, undef, undef);
1048
1049 return qinst;
1050 }
1051
1052 static void
emit_nop(struct v3d_compile * c,struct qblock * block,struct choose_scoreboard * scoreboard)1053 emit_nop(struct v3d_compile *c, struct qblock *block,
1054 struct choose_scoreboard *scoreboard)
1055 {
1056 insert_scheduled_instruction(c, block, scoreboard, vir_nop());
1057 }
1058
1059 static bool
qpu_instruction_valid_in_thrend_slot(struct v3d_compile * c,const struct qinst * qinst,int slot)1060 qpu_instruction_valid_in_thrend_slot(struct v3d_compile *c,
1061 const struct qinst *qinst, int slot)
1062 {
1063 const struct v3d_qpu_instr *inst = &qinst->qpu;
1064
1065 /* Only TLB Z writes are prohibited in the last slot, but we don't
1066 * have those flagged so prohibit all TLB ops for now.
1067 */
1068 if (slot == 2 && qpu_inst_is_tlb(inst))
1069 return false;
1070
1071 if (slot > 0 && qinst->uniform != ~0)
1072 return false;
1073
1074 if (v3d_qpu_uses_vpm(inst))
1075 return false;
1076
1077 if (inst->sig.ldvary)
1078 return false;
1079
1080 if (inst->type == V3D_QPU_INSTR_TYPE_ALU) {
1081 /* No writing physical registers at the end. */
1082 if (!inst->alu.add.magic_write ||
1083 !inst->alu.mul.magic_write) {
1084 return false;
1085 }
1086
1087 if (c->devinfo->ver < 40 && inst->alu.add.op == V3D_QPU_A_SETMSF)
1088 return false;
1089
1090 /* RF0-2 might be overwritten during the delay slots by
1091 * fragment shader setup.
1092 */
1093 if (inst->raddr_a < 3 &&
1094 (inst->alu.add.a == V3D_QPU_MUX_A ||
1095 inst->alu.add.b == V3D_QPU_MUX_A ||
1096 inst->alu.mul.a == V3D_QPU_MUX_A ||
1097 inst->alu.mul.b == V3D_QPU_MUX_A)) {
1098 return false;
1099 }
1100
1101 if (inst->raddr_b < 3 &&
1102 !inst->sig.small_imm &&
1103 (inst->alu.add.a == V3D_QPU_MUX_B ||
1104 inst->alu.add.b == V3D_QPU_MUX_B ||
1105 inst->alu.mul.a == V3D_QPU_MUX_B ||
1106 inst->alu.mul.b == V3D_QPU_MUX_B)) {
1107 return false;
1108 }
1109 }
1110
1111 return true;
1112 }
1113
1114 static bool
valid_thrsw_sequence(struct v3d_compile * c,struct qinst * qinst,int instructions_in_sequence,bool is_thrend)1115 valid_thrsw_sequence(struct v3d_compile *c,
1116 struct qinst *qinst, int instructions_in_sequence,
1117 bool is_thrend)
1118 {
1119 for (int slot = 0; slot < instructions_in_sequence; slot++) {
1120 /* No scheduling SFU when the result would land in the other
1121 * thread. The simulator complains for safety, though it
1122 * would only occur for dead code in our case.
1123 */
1124 if (slot > 0 &&
1125 qinst->qpu.type == V3D_QPU_INSTR_TYPE_ALU &&
1126 (v3d_qpu_magic_waddr_is_sfu(qinst->qpu.alu.add.waddr) ||
1127 v3d_qpu_magic_waddr_is_sfu(qinst->qpu.alu.mul.waddr))) {
1128 return false;
1129 }
1130
1131 if (slot > 0 && qinst->qpu.sig.ldvary)
1132 return false;
1133
1134 if (is_thrend &&
1135 !qpu_instruction_valid_in_thrend_slot(c, qinst, slot)) {
1136 return false;
1137 }
1138
1139 /* Note that the list is circular, so we can only do this up
1140 * to instructions_in_sequence.
1141 */
1142 qinst = (struct qinst *)qinst->link.next;
1143 }
1144
1145 return true;
1146 }
1147
1148 /**
1149 * Emits a THRSW signal in the stream, trying to move it up to pair with
1150 * another instruction.
1151 */
1152 static int
emit_thrsw(struct v3d_compile * c,struct qblock * block,struct choose_scoreboard * scoreboard,struct qinst * inst,bool is_thrend)1153 emit_thrsw(struct v3d_compile *c,
1154 struct qblock *block,
1155 struct choose_scoreboard *scoreboard,
1156 struct qinst *inst,
1157 bool is_thrend)
1158 {
1159 int time = 0;
1160
1161 /* There should be nothing in a thrsw inst being scheduled other than
1162 * the signal bits.
1163 */
1164 assert(inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU);
1165 assert(inst->qpu.alu.add.op == V3D_QPU_A_NOP);
1166 assert(inst->qpu.alu.mul.op == V3D_QPU_M_NOP);
1167
1168 /* Find how far back into previous instructions we can put the THRSW. */
1169 int slots_filled = 0;
1170 struct qinst *merge_inst = NULL;
1171 vir_for_each_inst_rev(prev_inst, block) {
1172 struct v3d_qpu_sig sig = prev_inst->qpu.sig;
1173 sig.thrsw = true;
1174 uint32_t packed_sig;
1175
1176 if (!v3d_qpu_sig_pack(c->devinfo, &sig, &packed_sig))
1177 break;
1178
1179 if (!valid_thrsw_sequence(c, prev_inst, slots_filled + 1,
1180 is_thrend)) {
1181 break;
1182 }
1183
1184 merge_inst = prev_inst;
1185 if (++slots_filled == 3)
1186 break;
1187 }
1188
1189 bool needs_free = false;
1190 if (merge_inst) {
1191 merge_inst->qpu.sig.thrsw = true;
1192 needs_free = true;
1193 } else {
1194 insert_scheduled_instruction(c, block, scoreboard, inst);
1195 time++;
1196 slots_filled++;
1197 merge_inst = inst;
1198 }
1199
1200 /* Insert any extra delay slot NOPs we need. */
1201 for (int i = 0; i < 3 - slots_filled; i++) {
1202 emit_nop(c, block, scoreboard);
1203 time++;
1204 }
1205
1206 /* If we're emitting the last THRSW (other than program end), then
1207 * signal that to the HW by emitting two THRSWs in a row.
1208 */
1209 if (inst->is_last_thrsw) {
1210 struct qinst *second_inst =
1211 (struct qinst *)merge_inst->link.next;
1212 second_inst->qpu.sig.thrsw = true;
1213 }
1214
1215 /* If we put our THRSW into another instruction, free up the
1216 * instruction that didn't end up scheduled into the list.
1217 */
1218 if (needs_free)
1219 free(inst);
1220
1221 return time;
1222 }
1223
1224 static uint32_t
schedule_instructions(struct v3d_compile * c,struct choose_scoreboard * scoreboard,struct qblock * block,struct list_head * schedule_list,enum quniform_contents * orig_uniform_contents,uint32_t * orig_uniform_data,uint32_t * next_uniform)1225 schedule_instructions(struct v3d_compile *c,
1226 struct choose_scoreboard *scoreboard,
1227 struct qblock *block,
1228 struct list_head *schedule_list,
1229 enum quniform_contents *orig_uniform_contents,
1230 uint32_t *orig_uniform_data,
1231 uint32_t *next_uniform)
1232 {
1233 const struct v3d_device_info *devinfo = c->devinfo;
1234 uint32_t time = 0;
1235
1236 if (debug) {
1237 fprintf(stderr, "initial deps:\n");
1238 dump_state(devinfo, schedule_list);
1239 fprintf(stderr, "\n");
1240 }
1241
1242 /* Remove non-DAG heads from the list. */
1243 list_for_each_entry_safe(struct schedule_node, n, schedule_list, link) {
1244 if (n->parent_count != 0)
1245 list_del(&n->link);
1246 }
1247
1248 while (!list_empty(schedule_list)) {
1249 struct schedule_node *chosen =
1250 choose_instruction_to_schedule(devinfo,
1251 scoreboard,
1252 schedule_list,
1253 NULL);
1254 struct schedule_node *merge = NULL;
1255
1256 /* If there are no valid instructions to schedule, drop a NOP
1257 * in.
1258 */
1259 struct qinst *qinst = chosen ? chosen->inst : vir_nop();
1260 struct v3d_qpu_instr *inst = &qinst->qpu;
1261
1262 if (debug) {
1263 fprintf(stderr, "t=%4d: current list:\n",
1264 time);
1265 dump_state(devinfo, schedule_list);
1266 fprintf(stderr, "t=%4d: chose: ", time);
1267 v3d_qpu_dump(devinfo, inst);
1268 fprintf(stderr, "\n");
1269 }
1270
1271 /* Schedule this instruction onto the QPU list. Also try to
1272 * find an instruction to pair with it.
1273 */
1274 if (chosen) {
1275 time = MAX2(chosen->unblocked_time, time);
1276 list_del(&chosen->link);
1277 mark_instruction_scheduled(schedule_list, time,
1278 chosen, true);
1279
1280 merge = choose_instruction_to_schedule(devinfo,
1281 scoreboard,
1282 schedule_list,
1283 chosen);
1284 if (merge) {
1285 time = MAX2(merge->unblocked_time, time);
1286 list_del(&merge->link);
1287 (void)qpu_merge_inst(devinfo, inst,
1288 inst, &merge->inst->qpu);
1289 if (merge->inst->uniform != -1) {
1290 chosen->inst->uniform =
1291 merge->inst->uniform;
1292 }
1293
1294 if (debug) {
1295 fprintf(stderr, "t=%4d: merging: ",
1296 time);
1297 v3d_qpu_dump(devinfo, &merge->inst->qpu);
1298 fprintf(stderr, "\n");
1299 fprintf(stderr, " result: ");
1300 v3d_qpu_dump(devinfo, inst);
1301 fprintf(stderr, "\n");
1302 }
1303 }
1304 }
1305
1306 /* Update the uniform index for the rewritten location --
1307 * branch target updating will still need to change
1308 * c->uniform_data[] using this index.
1309 */
1310 if (qinst->uniform != -1) {
1311 if (inst->type == V3D_QPU_INSTR_TYPE_BRANCH)
1312 block->branch_uniform = *next_uniform;
1313
1314 c->uniform_data[*next_uniform] =
1315 orig_uniform_data[qinst->uniform];
1316 c->uniform_contents[*next_uniform] =
1317 orig_uniform_contents[qinst->uniform];
1318 qinst->uniform = *next_uniform;
1319 (*next_uniform)++;
1320 }
1321
1322 if (debug) {
1323 fprintf(stderr, "\n");
1324 }
1325
1326 /* Now that we've scheduled a new instruction, some of its
1327 * children can be promoted to the list of instructions ready to
1328 * be scheduled. Update the children's unblocked time for this
1329 * DAG edge as we do so.
1330 */
1331 mark_instruction_scheduled(schedule_list, time, chosen, false);
1332
1333 if (merge) {
1334 mark_instruction_scheduled(schedule_list, time, merge,
1335 false);
1336
1337 /* The merged VIR instruction doesn't get re-added to the
1338 * block, so free it now.
1339 */
1340 free(merge->inst);
1341 }
1342
1343 if (inst->sig.thrsw) {
1344 time += emit_thrsw(c, block, scoreboard, qinst, false);
1345 } else {
1346 insert_scheduled_instruction(c, block,
1347 scoreboard, qinst);
1348
1349 if (inst->type == V3D_QPU_INSTR_TYPE_BRANCH) {
1350 block->branch_qpu_ip = c->qpu_inst_count - 1;
1351 /* Fill the delay slots.
1352 *
1353 * We should fill these with actual instructions,
1354 * instead, but that will probably need to be done
1355 * after this, once we know what the leading
1356 * instructions of the successors are (so we can
1357 * handle A/B register file write latency)
1358 */
1359 for (int i = 0; i < 3; i++)
1360 emit_nop(c, block, scoreboard);
1361 }
1362 }
1363 }
1364
1365 return time;
1366 }
1367
1368 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)1369 qpu_schedule_instructions_block(struct v3d_compile *c,
1370 struct choose_scoreboard *scoreboard,
1371 struct qblock *block,
1372 enum quniform_contents *orig_uniform_contents,
1373 uint32_t *orig_uniform_data,
1374 uint32_t *next_uniform)
1375 {
1376 void *mem_ctx = ralloc_context(NULL);
1377 struct list_head schedule_list;
1378
1379 list_inithead(&schedule_list);
1380
1381 /* Wrap each instruction in a scheduler structure. */
1382 while (!list_empty(&block->instructions)) {
1383 struct qinst *qinst = (struct qinst *)block->instructions.next;
1384 struct schedule_node *n =
1385 rzalloc(mem_ctx, struct schedule_node);
1386
1387 n->inst = qinst;
1388
1389 list_del(&qinst->link);
1390 list_addtail(&n->link, &schedule_list);
1391 }
1392
1393 calculate_forward_deps(c, &schedule_list);
1394 calculate_reverse_deps(c, &schedule_list);
1395
1396 list_for_each_entry(struct schedule_node, n, &schedule_list, link) {
1397 compute_delay(n);
1398 }
1399
1400 uint32_t cycles = schedule_instructions(c, scoreboard, block,
1401 &schedule_list,
1402 orig_uniform_contents,
1403 orig_uniform_data,
1404 next_uniform);
1405
1406 ralloc_free(mem_ctx);
1407
1408 return cycles;
1409 }
1410
1411 static void
qpu_set_branch_targets(struct v3d_compile * c)1412 qpu_set_branch_targets(struct v3d_compile *c)
1413 {
1414 vir_for_each_block(block, c) {
1415 /* The end block of the program has no branch. */
1416 if (!block->successors[0])
1417 continue;
1418
1419 /* If there was no branch instruction, then the successor
1420 * block must follow immediately after this one.
1421 */
1422 if (block->branch_qpu_ip == ~0) {
1423 assert(block->end_qpu_ip + 1 ==
1424 block->successors[0]->start_qpu_ip);
1425 continue;
1426 }
1427
1428 /* Walk back through the delay slots to find the branch
1429 * instr.
1430 */
1431 struct list_head *entry = block->instructions.prev;
1432 for (int i = 0; i < 3; i++)
1433 entry = entry->prev;
1434 struct qinst *branch = container_of(entry, branch, link);
1435 assert(branch->qpu.type == V3D_QPU_INSTR_TYPE_BRANCH);
1436
1437 /* Make sure that the if-we-don't-jump
1438 * successor was scheduled just after the
1439 * delay slots.
1440 */
1441 assert(!block->successors[1] ||
1442 block->successors[1]->start_qpu_ip ==
1443 block->branch_qpu_ip + 4);
1444
1445 branch->qpu.branch.offset =
1446 ((block->successors[0]->start_qpu_ip -
1447 (block->branch_qpu_ip + 4)) *
1448 sizeof(uint64_t));
1449
1450 /* Set up the relative offset to jump in the
1451 * uniform stream.
1452 *
1453 * Use a temporary here, because
1454 * uniform_data[inst->uniform] may be shared
1455 * between multiple instructions.
1456 */
1457 assert(c->uniform_contents[branch->uniform] == QUNIFORM_CONSTANT);
1458 c->uniform_data[branch->uniform] =
1459 (block->successors[0]->start_uniform -
1460 (block->branch_uniform + 1)) * 4;
1461 }
1462 }
1463
1464 uint32_t
v3d_qpu_schedule_instructions(struct v3d_compile * c)1465 v3d_qpu_schedule_instructions(struct v3d_compile *c)
1466 {
1467 const struct v3d_device_info *devinfo = c->devinfo;
1468 struct qblock *end_block = list_last_entry(&c->blocks,
1469 struct qblock, link);
1470
1471 /* We reorder the uniforms as we schedule instructions, so save the
1472 * old data off and replace it.
1473 */
1474 uint32_t *uniform_data = c->uniform_data;
1475 enum quniform_contents *uniform_contents = c->uniform_contents;
1476 c->uniform_contents = ralloc_array(c, enum quniform_contents,
1477 c->num_uniforms);
1478 c->uniform_data = ralloc_array(c, uint32_t, c->num_uniforms);
1479 c->uniform_array_size = c->num_uniforms;
1480 uint32_t next_uniform = 0;
1481
1482 struct choose_scoreboard scoreboard;
1483 memset(&scoreboard, 0, sizeof(scoreboard));
1484 scoreboard.last_waddr_add = ~0;
1485 scoreboard.last_waddr_mul = ~0;
1486 scoreboard.last_ldvary_tick = -10;
1487 scoreboard.last_sfu_write_tick = -10;
1488 scoreboard.last_uniforms_reset_tick = -10;
1489
1490 if (debug) {
1491 fprintf(stderr, "Pre-schedule instructions\n");
1492 vir_for_each_block(block, c) {
1493 fprintf(stderr, "BLOCK %d\n", block->index);
1494 list_for_each_entry(struct qinst, qinst,
1495 &block->instructions, link) {
1496 v3d_qpu_dump(devinfo, &qinst->qpu);
1497 fprintf(stderr, "\n");
1498 }
1499 }
1500 fprintf(stderr, "\n");
1501 }
1502
1503 uint32_t cycles = 0;
1504 vir_for_each_block(block, c) {
1505 block->start_qpu_ip = c->qpu_inst_count;
1506 block->branch_qpu_ip = ~0;
1507 block->start_uniform = next_uniform;
1508
1509 cycles += qpu_schedule_instructions_block(c,
1510 &scoreboard,
1511 block,
1512 uniform_contents,
1513 uniform_data,
1514 &next_uniform);
1515
1516 block->end_qpu_ip = c->qpu_inst_count - 1;
1517 }
1518
1519 /* Emit the program-end THRSW instruction. */;
1520 struct qinst *thrsw = vir_nop();
1521 thrsw->qpu.sig.thrsw = true;
1522 emit_thrsw(c, end_block, &scoreboard, thrsw, true);
1523
1524 qpu_set_branch_targets(c);
1525
1526 assert(next_uniform == c->num_uniforms);
1527
1528 return cycles;
1529 }
1530