2 * Copyright © 2010 Intel Corporation
3 * Copyright © 2014-2015 Broadcom
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:
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
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
26 * @file vc4_qir_schedule.c
28 * The basic model of the list scheduler is to take a basic block, compute a
29 * DAG of the dependencies from the bottom up, and make a list of the DAG
30 * heads. Heuristically pick a DAG head and schedule (remove) it, then put
31 * all the parents that are now DAG heads into the list of things to
34 * The goal of scheduling here, before register allocation and conversion to
35 * QPU instructions, is to reduce register pressure by reordering instructions
36 * to consume values when possible.
43 struct schedule_node
{
44 struct list_head link
;
47 struct schedule_node
**children
;
49 uint32_t child_array_size
;
50 uint32_t parent_count
;
52 /* Length of the longest (latency) chain from a DAG head to the this
57 /* Longest time + latency_between(parent, this) of any parent of this
60 uint32_t unblocked_time
;
63 struct schedule_state
{
64 /* List of struct schedule_node *. This starts out with all
65 * instructions, and after dependency updates it's trimmed to be just
68 struct list_head worklist
;
72 uint32_t *temp_writes
;
74 BITSET_WORD
*temp_live
;
77 /* When walking the instructions in reverse, we need to swap before/after in
80 enum direction
{ F
, R
};
83 * Marks a dependency between two intructions, that @after must appear after
86 * Our dependencies are tracked as a DAG. Since we're scheduling bottom-up,
87 * the latest instructions with nothing left to schedule are the DAG heads,
88 * and their inputs are their children.
91 add_dep(enum direction dir
,
92 struct schedule_node
*before
,
93 struct schedule_node
*after
)
95 if (!before
|| !after
)
98 assert(before
!= after
);
101 struct schedule_node
*t
= before
;
106 for (int i
= 0; i
< after
->child_count
; i
++) {
107 if (after
->children
[i
] == after
)
111 if (after
->child_array_size
<= after
->child_count
) {
112 after
->child_array_size
= MAX2(after
->child_array_size
* 2, 16);
113 after
->children
= reralloc(after
, after
->children
,
114 struct schedule_node
*,
115 after
->child_array_size
);
118 after
->children
[after
->child_count
] = before
;
119 after
->child_count
++;
120 before
->parent_count
++;
124 add_write_dep(enum direction dir
,
125 struct schedule_node
**before
,
126 struct schedule_node
*after
)
128 add_dep(dir
, *before
, after
);
132 struct schedule_setup_state
{
133 struct schedule_node
**last_temp_write
;
134 struct schedule_node
*last_sf
;
135 struct schedule_node
*last_vary_read
;
136 struct schedule_node
*last_vpm_read
;
137 struct schedule_node
*last_vpm_write
;
138 struct schedule_node
*last_tex_coord
;
139 struct schedule_node
*last_tex_result
;
140 struct schedule_node
*last_tlb
;
144 * Texture FIFO tracking. This is done top-to-bottom, and is used to
145 * track the QOP_TEX_RESULTs and add dependencies on previous ones
146 * when trying to submit texture coords with TFREQ full or new texture
147 * fetches with TXRCV full.
150 struct schedule_node
*node
;
153 int tfreq_count
; /**< Number of texture coords outstanding. */
154 int tfrcv_count
; /**< Number of texture results outstanding. */
159 block_until_tex_result(struct schedule_setup_state
*state
, struct schedule_node
*n
)
161 add_dep(state
->dir
, state
->tex_fifo
[0].node
, n
);
163 state
->tfreq_count
-= state
->tex_fifo
[0].coords
;
164 state
->tfrcv_count
--;
166 memmove(&state
->tex_fifo
[0],
168 state
->tex_fifo_pos
* sizeof(state
->tex_fifo
[0]));
169 state
->tex_fifo_pos
--;
173 * Common code for dependencies that need to be tracked both forward and
176 * This is for things like "all VPM reads have to happen in order."
179 calculate_deps(struct schedule_setup_state
*state
, struct schedule_node
*n
)
181 struct qinst
*inst
= n
->inst
;
182 enum direction dir
= state
->dir
;
185 /* Add deps for temp registers and varyings accesses. Note that we
186 * ignore uniforms accesses, because qir_reorder_uniforms() happens
189 for (int i
= 0; i
< qir_get_op_nsrc(inst
->op
); i
++) {
190 switch (inst
->src
[i
].file
) {
193 state
->last_temp_write
[inst
->src
[i
].index
], n
);
197 add_write_dep(dir
, &state
->last_vary_read
, n
);
201 add_write_dep(dir
, &state
->last_vpm_read
, n
);
211 add_dep(dir
, state
->last_vary_read
, n
);
219 /* Texturing setup gets scheduled in order, because
220 * the uniforms referenced by them have to land in a
223 add_write_dep(dir
, &state
->last_tex_coord
, n
);
227 /* Results have to be fetched in order. */
228 add_write_dep(dir
, &state
->last_tex_result
, n
);
231 case QOP_TLB_COLOR_WRITE
:
232 case QOP_TLB_COLOR_READ
:
233 case QOP_TLB_Z_WRITE
:
234 case QOP_TLB_STENCIL_SETUP
:
236 add_write_dep(dir
, &state
->last_tlb
, n
);
239 case QOP_TLB_DISCARD_SETUP
:
240 add_write_dep(dir
, &state
->last_sf
, n
);
241 add_write_dep(dir
, &state
->last_tlb
, n
);
248 if (inst
->dst
.file
== QFILE_VPM
)
249 add_write_dep(dir
, &state
->last_vpm_write
, n
);
250 else if (inst
->dst
.file
== QFILE_TEMP
)
251 add_write_dep(dir
, &state
->last_temp_write
[inst
->dst
.index
], n
);
253 if (qir_depends_on_flags(inst
))
254 add_dep(dir
, state
->last_sf
, n
);
257 add_write_dep(dir
, &state
->last_sf
, n
);
261 calculate_forward_deps(struct vc4_compile
*c
, void *mem_ctx
,
262 struct list_head
*schedule_list
)
264 struct schedule_setup_state state
;
266 memset(&state
, 0, sizeof(state
));
267 state
.last_temp_write
= rzalloc_array(mem_ctx
, struct schedule_node
*,
271 list_for_each_entry(struct schedule_node
, n
, schedule_list
, link
) {
272 struct qinst
*inst
= n
->inst
;
274 calculate_deps(&state
, n
);
282 /* If the texture coordinate fifo is full,
283 * block this on the last QOP_TEX_RESULT.
285 if (state
.tfreq_count
== 8) {
286 block_until_tex_result(&state
, n
);
289 /* If the texture result fifo is full, block
290 * adding any more to it until the last
293 if (inst
->op
== QOP_TEX_S
||
294 inst
->op
== QOP_TEX_DIRECT
) {
295 if (state
.tfrcv_count
== 4)
296 block_until_tex_result(&state
, n
);
300 state
.tex_fifo
[state
.tex_fifo_pos
].coords
++;
305 /* Results have to be fetched after the
306 * coordinate setup. Note that we're assuming
307 * here that our input shader has the texture
308 * coord setup and result fetch in order,
309 * which is true initially but not of our
310 * instruction stream after this pass.
312 add_dep(state
.dir
, state
.last_tex_coord
, n
);
314 state
.tex_fifo
[state
.tex_fifo_pos
].node
= n
;
316 state
.tex_fifo_pos
++;
317 memset(&state
.tex_fifo
[state
.tex_fifo_pos
], 0,
318 sizeof(state
.tex_fifo
[0]));
321 assert(!qir_is_tex(inst
));
328 calculate_reverse_deps(struct vc4_compile
*c
, void *mem_ctx
,
329 struct list_head
*schedule_list
)
331 struct schedule_setup_state state
;
333 memset(&state
, 0, sizeof(state
));
335 state
.last_temp_write
= rzalloc_array(mem_ctx
, struct schedule_node
*,
338 list_for_each_entry_rev(struct schedule_node
, n
, schedule_list
, link
) {
339 calculate_deps(&state
, n
);
344 get_register_pressure_cost(struct schedule_state
*state
, struct qinst
*inst
)
348 if (inst
->dst
.file
== QFILE_TEMP
&&
349 state
->temp_writes
[inst
->dst
.index
] == 1)
352 for (int i
= 0; i
< qir_get_op_nsrc(inst
->op
); i
++) {
353 if (inst
->src
[i
].file
== QFILE_TEMP
&&
354 !BITSET_TEST(state
->temp_live
, inst
->src
[i
].index
)) {
363 locks_scoreboard(struct qinst
*inst
)
366 case QOP_TLB_Z_WRITE
:
367 case QOP_TLB_COLOR_WRITE
:
368 case QOP_TLB_COLOR_WRITE_MS
:
369 case QOP_TLB_COLOR_READ
:
376 static struct schedule_node
*
377 choose_instruction(struct schedule_state
*state
)
379 struct schedule_node
*chosen
= NULL
;
381 list_for_each_entry(struct schedule_node
, n
, &state
->worklist
, link
) {
387 /* Prefer scheduling things that lock the scoreboard, so that
388 * they appear late in the program and we get more parallelism
389 * between shaders on multiple QPUs hitting the same fragment.
391 if (locks_scoreboard(n
->inst
) &&
392 !locks_scoreboard(chosen
->inst
)) {
395 } else if (!locks_scoreboard(n
->inst
) &&
396 locks_scoreboard(chosen
->inst
)) {
400 /* If we would block on the previously chosen node, but would
401 * block less on this one, then then prefer it.
403 if (chosen
->unblocked_time
> state
->time
&&
404 n
->unblocked_time
< chosen
->unblocked_time
) {
407 } else if (n
->unblocked_time
> state
->time
&&
408 n
->unblocked_time
> chosen
->unblocked_time
) {
412 /* If we can definitely reduce register pressure, do so
415 int register_pressure_cost
=
416 get_register_pressure_cost(state
, n
->inst
);
417 int chosen_register_pressure_cost
=
418 get_register_pressure_cost(state
, chosen
->inst
);
420 if (register_pressure_cost
< chosen_register_pressure_cost
) {
423 } else if (register_pressure_cost
>
424 chosen_register_pressure_cost
) {
428 /* Otherwise, prefer instructions with the deepest chain to
429 * the end of the program. This avoids the problem of
430 * "everything generates a temp, nothing finishes freeing one,
431 * guess I'll just keep emitting varying mul/adds".
433 if (n
->delay
> chosen
->delay
) {
436 } else if (n
->delay
< chosen
->delay
) {
445 dump_state(struct vc4_compile
*c
, struct schedule_state
*state
)
448 list_for_each_entry(struct schedule_node
, n
, &state
->worklist
, link
) {
449 fprintf(stderr
, "%3d: ", i
++);
450 qir_dump_inst(c
, n
->inst
);
451 fprintf(stderr
, " (%d cost)\n",
452 get_register_pressure_cost(state
, n
->inst
));
454 for (int i
= 0; i
< n
->child_count
; i
++) {
455 struct schedule_node
*child
= n
->children
[i
];
456 fprintf(stderr
, " - ");
457 qir_dump_inst(c
, child
->inst
);
458 fprintf(stderr
, " (%d parents)\n", child
->parent_count
);
463 /* Estimate of how many instructions we should schedule between operations.
465 * These aren't in real cycle counts, because we're just estimating cycle
466 * times anyway. QIR instructions will get paired up when turned into QPU
467 * instructions, or extra NOP delays will have to be added due to register
468 * allocation choices.
471 latency_between(struct schedule_node
*before
, struct schedule_node
*after
)
473 if ((before
->inst
->op
== QOP_TEX_S
||
474 before
->inst
->op
== QOP_TEX_DIRECT
) &&
475 after
->inst
->op
== QOP_TEX_RESULT
)
481 /** Recursive computation of the delay member of a node. */
483 compute_delay(struct schedule_node
*n
)
485 if (!n
->child_count
) {
486 /* The color read needs to be scheduled late, to avoid locking
487 * the scoreboard early. This is our best tool for
488 * encouraging that. The other scoreboard locking ops will
489 * have this happen by default, since they are generally the
490 * DAG heads or close to them.
492 if (n
->inst
->op
== QOP_TLB_COLOR_READ
)
497 for (int i
= 0; i
< n
->child_count
; i
++) {
498 if (!n
->children
[i
]->delay
)
499 compute_delay(n
->children
[i
]);
500 n
->delay
= MAX2(n
->delay
,
501 n
->children
[i
]->delay
+
502 latency_between(n
, n
->children
[i
]));
508 schedule_instructions(struct vc4_compile
*c
, struct schedule_state
*state
)
511 fprintf(stderr
, "initial deps:\n");
512 dump_state(c
, state
);
515 /* Remove non-DAG heads from the list. */
516 list_for_each_entry_safe(struct schedule_node
, n
,
517 &state
->worklist
, link
) {
518 if (n
->parent_count
!= 0)
523 while (!list_empty(&state
->worklist
)) {
524 struct schedule_node
*chosen
= choose_instruction(state
);
525 struct qinst
*inst
= chosen
->inst
;
528 fprintf(stderr
, "current list:\n");
529 dump_state(c
, state
);
530 fprintf(stderr
, "chose: ");
531 qir_dump_inst(c
, inst
);
532 fprintf(stderr
, " (%d cost)\n",
533 get_register_pressure_cost(state
, inst
));
536 state
->time
= MAX2(state
->time
, chosen
->unblocked_time
);
538 /* Schedule this instruction back onto the QIR list. */
539 list_del(&chosen
->link
);
540 list_add(&inst
->link
, &c
->instructions
);
542 /* Now that we've scheduled a new instruction, some of its
543 * children can be promoted to the list of instructions ready to
544 * be scheduled. Update the children's unblocked time for this
545 * DAG edge as we do so.
547 for (int i
= chosen
->child_count
- 1; i
>= 0; i
--) {
548 struct schedule_node
*child
= chosen
->children
[i
];
550 child
->unblocked_time
= MAX2(child
->unblocked_time
,
552 latency_between(chosen
,
554 child
->parent_count
--;
555 if (child
->parent_count
== 0)
556 list_add(&child
->link
, &state
->worklist
);
559 /* Update our tracking of register pressure. */
560 for (int i
= 0; i
< qir_get_op_nsrc(inst
->op
); i
++) {
561 if (inst
->src
[i
].file
== QFILE_TEMP
)
562 BITSET_SET(state
->temp_live
, inst
->src
[i
].index
);
564 if (inst
->dst
.file
== QFILE_TEMP
) {
565 state
->temp_writes
[inst
->dst
.index
]--;
566 if (state
->temp_writes
[inst
->dst
.index
] == 0)
567 BITSET_CLEAR(state
->temp_live
, inst
->dst
.index
);
575 qir_schedule_instructions(struct vc4_compile
*c
)
577 void *mem_ctx
= ralloc_context(NULL
);
578 struct schedule_state state
= { 0 };
581 fprintf(stderr
, "Pre-schedule instructions\n");
585 state
.temp_writes
= rzalloc_array(mem_ctx
, uint32_t, c
->num_temps
);
586 state
.temp_live
= rzalloc_array(mem_ctx
, BITSET_WORD
,
587 BITSET_WORDS(c
->num_temps
));
588 list_inithead(&state
.worklist
);
590 /* Wrap each instruction in a scheduler structure. */
591 list_for_each_entry_safe(struct qinst
, inst
, &c
->instructions
, link
) {
592 struct schedule_node
*n
= rzalloc(mem_ctx
, struct schedule_node
);
595 list_del(&inst
->link
);
596 list_addtail(&n
->link
, &state
.worklist
);
598 if (inst
->dst
.file
== QFILE_TEMP
)
599 state
.temp_writes
[inst
->dst
.index
]++;
602 /* Dependencies tracked top-to-bottom. */
603 calculate_forward_deps(c
, mem_ctx
, &state
.worklist
);
604 /* Dependencies tracked bottom-to-top. */
605 calculate_reverse_deps(c
, mem_ctx
, &state
.worklist
);
607 list_for_each_entry(struct schedule_node
, n
, &state
.worklist
, link
)
610 schedule_instructions(c
, &state
);
613 fprintf(stderr
, "Post-schedule instructions\n");
617 ralloc_free(mem_ctx
);