2 * Copyright © 2010 Intel Corporation
3 * Copyright © 2014-2017 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
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.
33 * The goal of scheduling here is to pack pairs of operations together in a
34 * single QPU instruction.
37 #include "qpu/qpu_disasm.h"
38 #include "v3d_compiler.h"
39 #include "util/ralloc.h"
43 struct schedule_node_child
;
45 struct schedule_node
{
46 struct list_head link
;
48 struct schedule_node_child
*children
;
50 uint32_t child_array_size
;
51 uint32_t parent_count
;
53 /* Longest cycles + instruction_latency() of any parent of this node. */
54 uint32_t unblocked_time
;
57 * Minimum number of cycles from scheduling this instruction until the
58 * end of the program, based on the slowest dependency chain through
64 * cycles between this instruction being scheduled and when its result
70 struct schedule_node_child
{
71 struct schedule_node
*node
;
72 bool write_after_read
;
75 /* When walking the instructions in reverse, we need to swap before/after in
78 enum direction
{ F
, R
};
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
;
93 /* Estimated cycle when the current instruction would start. */
98 add_dep(struct schedule_state
*state
,
99 struct schedule_node
*before
,
100 struct schedule_node
*after
,
103 bool write_after_read
= !write
&& state
->dir
== R
;
105 if (!before
|| !after
)
108 assert(before
!= after
);
110 if (state
->dir
== R
) {
111 struct schedule_node
*t
= before
;
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
)) {
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
);
130 before
->children
[before
->child_count
].node
= after
;
131 before
->children
[before
->child_count
].write_after_read
=
133 before
->child_count
++;
134 after
->parent_count
++;
138 add_read_dep(struct schedule_state
*state
,
139 struct schedule_node
*before
,
140 struct schedule_node
*after
)
142 add_dep(state
, before
, after
, false);
146 add_write_dep(struct schedule_state
*state
,
147 struct schedule_node
**before
,
148 struct schedule_node
*after
)
150 add_dep(state
, *before
, after
, true);
155 qpu_inst_is_tlb(const struct v3d_qpu_instr
*inst
)
157 if (inst
->type
!= V3D_QPU_INSTR_TYPE_ALU
)
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
))
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
))
174 process_mux_deps(struct schedule_state
*state
, struct schedule_node
*n
,
175 enum v3d_qpu_mux mux
)
179 add_read_dep(state
, state
->last_rf
[n
->inst
->qpu
.raddr_a
], n
);
182 add_read_dep(state
, state
->last_rf
[n
->inst
->qpu
.raddr_b
], n
);
185 add_read_dep(state
, state
->last_r
[mux
- V3D_QPU_MUX_R0
], n
);
192 process_waddr_deps(struct schedule_state
*state
, struct schedule_node
*n
,
193 uint32_t waddr
, bool magic
)
196 add_write_dep(state
, &state
->last_rf
[waddr
], n
);
197 } else if (v3d_qpu_magic_waddr_is_tmu(waddr
)) {
198 /* XXX perf: For V3D 4.x, we could reorder TMU writes other
199 * than the TMUS/TMUD/TMUA to improve scheduling flexibility.
201 add_write_dep(state
, &state
->last_tmu_write
, n
);
203 case V3D_QPU_WADDR_TMUS
:
204 case V3D_QPU_WADDR_TMUSCM
:
205 case V3D_QPU_WADDR_TMUSF
:
206 case V3D_QPU_WADDR_TMUSLOD
:
207 add_write_dep(state
, &state
->last_tmu_config
, n
);
212 } else if (v3d_qpu_magic_waddr_is_sfu(waddr
)) {
213 /* Handled by v3d_qpu_writes_r4() check. */
216 case V3D_QPU_WADDR_R0
:
217 case V3D_QPU_WADDR_R1
:
218 case V3D_QPU_WADDR_R2
:
220 &state
->last_r
[waddr
- V3D_QPU_WADDR_R0
],
223 case V3D_QPU_WADDR_R3
:
224 case V3D_QPU_WADDR_R4
:
225 case V3D_QPU_WADDR_R5
:
226 /* Handled by v3d_qpu_writes_r*() checks below. */
229 case V3D_QPU_WADDR_VPM
:
230 case V3D_QPU_WADDR_VPMU
:
231 add_write_dep(state
, &state
->last_vpm
, n
);
234 case V3D_QPU_WADDR_TLB
:
235 case V3D_QPU_WADDR_TLBU
:
236 add_write_dep(state
, &state
->last_tlb
, n
);
239 case V3D_QPU_WADDR_NOP
:
243 fprintf(stderr
, "Unknown waddr %d\n", waddr
);
250 * Common code for dependencies that need to be tracked both forward and
253 * This is for things like "all reads of r4 have to happen between the r4
254 * writes that surround them".
257 calculate_deps(struct schedule_state
*state
, struct schedule_node
*n
)
259 const struct v3d_device_info
*devinfo
= state
->devinfo
;
260 struct qinst
*qinst
= n
->inst
;
261 struct v3d_qpu_instr
*inst
= &qinst
->qpu
;
262 /* If the input and output segments are shared, then all VPM reads to
263 * a location need to happen before all writes. We handle this by
264 * serializing all VPM operations for now.
266 bool separate_vpm_segment
= false;
268 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
) {
269 if (inst
->branch
.cond
!= V3D_QPU_BRANCH_COND_ALWAYS
)
270 add_read_dep(state
, state
->last_sf
, n
);
277 add_write_dep(state
, &state
->last_unif
, n
);
281 assert(inst
->type
== V3D_QPU_INSTR_TYPE_ALU
);
285 if (v3d_qpu_add_op_num_src(inst
->alu
.add
.op
) > 0)
286 process_mux_deps(state
, n
, inst
->alu
.add
.a
);
287 if (v3d_qpu_add_op_num_src(inst
->alu
.add
.op
) > 1)
288 process_mux_deps(state
, n
, inst
->alu
.add
.b
);
290 if (v3d_qpu_mul_op_num_src(inst
->alu
.mul
.op
) > 0)
291 process_mux_deps(state
, n
, inst
->alu
.mul
.a
);
292 if (v3d_qpu_mul_op_num_src(inst
->alu
.mul
.op
) > 1)
293 process_mux_deps(state
, n
, inst
->alu
.mul
.b
);
295 switch (inst
->alu
.add
.op
) {
296 case V3D_QPU_A_VPMSETUP
:
297 /* Could distinguish read/write by unpacking the uniform. */
298 add_write_dep(state
, &state
->last_vpm
, n
);
299 add_write_dep(state
, &state
->last_vpm_read
, n
);
302 case V3D_QPU_A_STVPMV
:
303 case V3D_QPU_A_STVPMD
:
304 case V3D_QPU_A_STVPMP
:
305 add_write_dep(state
, &state
->last_vpm
, n
);
308 case V3D_QPU_A_LDVPMV_IN
:
309 case V3D_QPU_A_LDVPMD_IN
:
310 case V3D_QPU_A_LDVPMG_IN
:
311 case V3D_QPU_A_LDVPMP
:
312 if (!separate_vpm_segment
)
313 add_write_dep(state
, &state
->last_vpm
, n
);
316 case V3D_QPU_A_VPMWT
:
317 add_read_dep(state
, state
->last_vpm
, n
);
321 add_read_dep(state
, state
->last_tlb
, n
);
324 case V3D_QPU_A_SETMSF
:
325 case V3D_QPU_A_SETREVF
:
326 add_write_dep(state
, &state
->last_tlb
, n
);
333 switch (inst
->alu
.mul
.op
) {
334 case V3D_QPU_M_MULTOP
:
335 case V3D_QPU_M_UMUL24
:
336 /* MULTOP sets rtop, and UMUL24 implicitly reads rtop and
337 * resets it to 0. We could possibly reorder umul24s relative
338 * to each other, but for now just keep all the MUL parts in
341 add_write_dep(state
, &state
->last_rtop
, n
);
347 if (inst
->alu
.add
.op
!= V3D_QPU_A_NOP
) {
348 process_waddr_deps(state
, n
, inst
->alu
.add
.waddr
,
349 inst
->alu
.add
.magic_write
);
351 if (inst
->alu
.mul
.op
!= V3D_QPU_M_NOP
) {
352 process_waddr_deps(state
, n
, inst
->alu
.mul
.waddr
,
353 inst
->alu
.mul
.magic_write
);
355 if (v3d_qpu_sig_writes_address(devinfo
, &inst
->sig
)) {
356 process_waddr_deps(state
, n
, inst
->sig_addr
,
360 if (v3d_qpu_writes_r3(devinfo
, inst
))
361 add_write_dep(state
, &state
->last_r
[3], n
);
362 if (v3d_qpu_writes_r4(devinfo
, inst
))
363 add_write_dep(state
, &state
->last_r
[4], n
);
364 if (v3d_qpu_writes_r5(devinfo
, inst
))
365 add_write_dep(state
, &state
->last_r
[5], n
);
367 if (inst
->sig
.thrsw
) {
368 /* All accumulator contents and flags are undefined after the
371 for (int i
= 0; i
< ARRAY_SIZE(state
->last_r
); i
++)
372 add_write_dep(state
, &state
->last_r
[i
], n
);
373 add_write_dep(state
, &state
->last_sf
, n
);
374 add_write_dep(state
, &state
->last_rtop
, n
);
376 /* Scoreboard-locking operations have to stay after the last
379 add_write_dep(state
, &state
->last_tlb
, n
);
381 add_write_dep(state
, &state
->last_tmu_write
, n
);
382 add_write_dep(state
, &state
->last_tmu_config
, n
);
385 if (v3d_qpu_waits_on_tmu(inst
)) {
386 /* TMU loads are coming from a FIFO, so ordering is important.
388 add_write_dep(state
, &state
->last_tmu_write
, n
);
391 if (inst
->sig
.wrtmuc
)
392 add_write_dep(state
, &state
->last_tmu_config
, n
);
394 if (inst
->sig
.ldtlb
| inst
->sig
.ldtlbu
)
395 add_read_dep(state
, state
->last_tlb
, n
);
397 if (inst
->sig
.ldvpm
) {
398 add_write_dep(state
, &state
->last_vpm_read
, n
);
400 /* At least for now, we're doing shared I/O segments, so queue
401 * all writes after all reads.
403 if (!separate_vpm_segment
)
404 add_write_dep(state
, &state
->last_vpm
, n
);
407 /* inst->sig.ldunif or sideband uniform read */
408 if (qinst
->uniform
!= ~0)
409 add_write_dep(state
, &state
->last_unif
, n
);
411 if (v3d_qpu_reads_flags(inst
))
412 add_read_dep(state
, state
->last_sf
, n
);
413 if (v3d_qpu_writes_flags(inst
))
414 add_write_dep(state
, &state
->last_sf
, n
);
418 calculate_forward_deps(struct v3d_compile
*c
, struct list_head
*schedule_list
)
420 struct schedule_state state
;
422 memset(&state
, 0, sizeof(state
));
423 state
.devinfo
= c
->devinfo
;
426 list_for_each_entry(struct schedule_node
, node
, schedule_list
, link
)
427 calculate_deps(&state
, node
);
431 calculate_reverse_deps(struct v3d_compile
*c
, struct list_head
*schedule_list
)
433 struct list_head
*node
;
434 struct schedule_state state
;
436 memset(&state
, 0, sizeof(state
));
437 state
.devinfo
= c
->devinfo
;
440 for (node
= schedule_list
->prev
; schedule_list
!= node
; node
= node
->prev
) {
441 calculate_deps(&state
, (struct schedule_node
*)node
);
445 struct choose_scoreboard
{
447 int last_magic_sfu_write_tick
;
448 int last_ldvary_tick
;
449 int last_uniforms_reset_tick
;
455 mux_reads_too_soon(struct choose_scoreboard
*scoreboard
,
456 const struct v3d_qpu_instr
*inst
, enum v3d_qpu_mux mux
)
460 if (scoreboard
->tick
- scoreboard
->last_magic_sfu_write_tick
<= 2)
465 if (scoreboard
->tick
- scoreboard
->last_ldvary_tick
<= 1)
476 reads_too_soon_after_write(struct choose_scoreboard
*scoreboard
,
479 const struct v3d_qpu_instr
*inst
= &qinst
->qpu
;
481 /* XXX: Branching off of raddr. */
482 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
)
485 assert(inst
->type
== V3D_QPU_INSTR_TYPE_ALU
);
487 if (inst
->alu
.add
.op
!= V3D_QPU_A_NOP
) {
488 if (v3d_qpu_add_op_num_src(inst
->alu
.add
.op
) > 0 &&
489 mux_reads_too_soon(scoreboard
, inst
, inst
->alu
.add
.a
)) {
492 if (v3d_qpu_add_op_num_src(inst
->alu
.add
.op
) > 1 &&
493 mux_reads_too_soon(scoreboard
, inst
, inst
->alu
.add
.b
)) {
498 if (inst
->alu
.mul
.op
!= V3D_QPU_M_NOP
) {
499 if (v3d_qpu_mul_op_num_src(inst
->alu
.mul
.op
) > 0 &&
500 mux_reads_too_soon(scoreboard
, inst
, inst
->alu
.mul
.a
)) {
503 if (v3d_qpu_mul_op_num_src(inst
->alu
.mul
.op
) > 1 &&
504 mux_reads_too_soon(scoreboard
, inst
, inst
->alu
.mul
.b
)) {
515 writes_too_soon_after_write(const struct v3d_device_info
*devinfo
,
516 struct choose_scoreboard
*scoreboard
,
519 const struct v3d_qpu_instr
*inst
= &qinst
->qpu
;
521 /* Don't schedule any other r4 write too soon after an SFU write.
522 * This would normally be prevented by dependency tracking, but might
523 * occur if a dead SFU computation makes it to scheduling.
525 if (scoreboard
->tick
- scoreboard
->last_magic_sfu_write_tick
< 2 &&
526 v3d_qpu_writes_r4(devinfo
, inst
))
533 pixel_scoreboard_too_soon(struct choose_scoreboard
*scoreboard
,
534 const struct v3d_qpu_instr
*inst
)
536 return (scoreboard
->tick
== 0 && qpu_inst_is_tlb(inst
));
540 get_instruction_priority(const struct v3d_qpu_instr
*inst
)
542 uint32_t baseline_score
;
543 uint32_t next_score
= 0;
545 /* Schedule TLB operations as late as possible, to get more
546 * parallelism between shaders.
548 if (qpu_inst_is_tlb(inst
))
552 /* Schedule texture read results collection late to hide latency. */
553 if (v3d_qpu_waits_on_tmu(inst
))
557 /* XXX perf: We should schedule SFU ALU ops so that the reader is 2
558 * instructions after the producer if possible, not just 1.
561 /* Default score for things that aren't otherwise special. */
562 baseline_score
= next_score
;
565 /* Schedule texture read setup early to hide their latency better. */
566 if (v3d_qpu_writes_tmu(inst
))
570 return baseline_score
;
574 qpu_magic_waddr_is_periph(enum v3d_qpu_waddr waddr
)
576 return (v3d_qpu_magic_waddr_is_tmu(waddr
) ||
577 v3d_qpu_magic_waddr_is_sfu(waddr
) ||
578 v3d_qpu_magic_waddr_is_tlb(waddr
) ||
579 v3d_qpu_magic_waddr_is_vpm(waddr
) ||
580 v3d_qpu_magic_waddr_is_tsy(waddr
));
584 qpu_accesses_peripheral(const struct v3d_qpu_instr
*inst
)
586 if (v3d_qpu_uses_vpm(inst
))
588 if (v3d_qpu_uses_sfu(inst
))
591 if (inst
->type
== V3D_QPU_INSTR_TYPE_ALU
) {
592 if (inst
->alu
.add
.op
!= V3D_QPU_A_NOP
&&
593 inst
->alu
.add
.magic_write
&&
594 qpu_magic_waddr_is_periph(inst
->alu
.add
.waddr
)) {
598 if (inst
->alu
.add
.op
== V3D_QPU_A_TMUWT
)
601 if (inst
->alu
.mul
.op
!= V3D_QPU_M_NOP
&&
602 inst
->alu
.mul
.magic_write
&&
603 qpu_magic_waddr_is_periph(inst
->alu
.mul
.waddr
)) {
608 return (inst
->sig
.ldvpm
||
616 qpu_merge_inst(const struct v3d_device_info
*devinfo
,
617 struct v3d_qpu_instr
*result
,
618 const struct v3d_qpu_instr
*a
,
619 const struct v3d_qpu_instr
*b
)
621 if (a
->type
!= V3D_QPU_INSTR_TYPE_ALU
||
622 b
->type
!= V3D_QPU_INSTR_TYPE_ALU
) {
626 /* Can't do more than one peripheral access in an instruction.
628 * XXX: V3D 4.1 allows TMU read along with a VPM read or write, and
629 * WRTMUC with a TMU magic register write (other than tmuc).
631 if (qpu_accesses_peripheral(a
) && qpu_accesses_peripheral(b
))
634 struct v3d_qpu_instr merge
= *a
;
636 if (b
->alu
.add
.op
!= V3D_QPU_A_NOP
) {
637 if (a
->alu
.add
.op
!= V3D_QPU_A_NOP
)
639 merge
.alu
.add
= b
->alu
.add
;
641 merge
.flags
.ac
= b
->flags
.ac
;
642 merge
.flags
.apf
= b
->flags
.apf
;
643 merge
.flags
.auf
= b
->flags
.auf
;
646 if (b
->alu
.mul
.op
!= V3D_QPU_M_NOP
) {
647 if (a
->alu
.mul
.op
!= V3D_QPU_M_NOP
)
649 merge
.alu
.mul
= b
->alu
.mul
;
651 merge
.flags
.mc
= b
->flags
.mc
;
652 merge
.flags
.mpf
= b
->flags
.mpf
;
653 merge
.flags
.muf
= b
->flags
.muf
;
656 if (v3d_qpu_uses_mux(b
, V3D_QPU_MUX_A
)) {
657 if (v3d_qpu_uses_mux(a
, V3D_QPU_MUX_A
) &&
658 a
->raddr_a
!= b
->raddr_a
) {
661 merge
.raddr_a
= b
->raddr_a
;
664 if (v3d_qpu_uses_mux(b
, V3D_QPU_MUX_B
)) {
665 if (v3d_qpu_uses_mux(a
, V3D_QPU_MUX_B
) &&
666 (a
->raddr_b
!= b
->raddr_b
||
667 a
->sig
.small_imm
!= b
->sig
.small_imm
)) {
670 merge
.raddr_b
= b
->raddr_b
;
673 merge
.sig
.thrsw
|= b
->sig
.thrsw
;
674 merge
.sig
.ldunif
|= b
->sig
.ldunif
;
675 merge
.sig
.ldunifrf
|= b
->sig
.ldunifrf
;
676 merge
.sig
.ldunifa
|= b
->sig
.ldunifa
;
677 merge
.sig
.ldunifarf
|= b
->sig
.ldunifarf
;
678 merge
.sig
.ldtmu
|= b
->sig
.ldtmu
;
679 merge
.sig
.ldvary
|= b
->sig
.ldvary
;
680 merge
.sig
.ldvpm
|= b
->sig
.ldvpm
;
681 merge
.sig
.small_imm
|= b
->sig
.small_imm
;
682 merge
.sig
.ldtlb
|= b
->sig
.ldtlb
;
683 merge
.sig
.ldtlbu
|= b
->sig
.ldtlbu
;
684 merge
.sig
.ucb
|= b
->sig
.ucb
;
685 merge
.sig
.rotate
|= b
->sig
.rotate
;
686 merge
.sig
.wrtmuc
|= b
->sig
.wrtmuc
;
688 if (v3d_qpu_sig_writes_address(devinfo
, &a
->sig
) &&
689 v3d_qpu_sig_writes_address(devinfo
, &b
->sig
))
691 merge
.sig_addr
|= b
->sig_addr
;
692 merge
.sig_magic
|= b
->sig_magic
;
695 bool ok
= v3d_qpu_instr_pack(devinfo
, &merge
, &packed
);
698 /* No modifying the real instructions on failure. */
699 assert(ok
|| (a
!= result
&& b
!= result
));
704 static struct schedule_node
*
705 choose_instruction_to_schedule(const struct v3d_device_info
*devinfo
,
706 struct choose_scoreboard
*scoreboard
,
707 struct list_head
*schedule_list
,
708 struct schedule_node
*prev_inst
)
710 struct schedule_node
*chosen
= NULL
;
713 /* Don't pair up anything with a thread switch signal -- emit_thrsw()
714 * will handle pairing it along with filling the delay slots.
717 if (prev_inst
->inst
->qpu
.sig
.thrsw
)
721 list_for_each_entry(struct schedule_node
, n
, schedule_list
, link
) {
722 const struct v3d_qpu_instr
*inst
= &n
->inst
->qpu
;
724 /* Don't choose the branch instruction until it's the last one
725 * left. We'll move it up to fit its delay slots after we
728 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
&&
729 !list_is_singular(schedule_list
)) {
733 /* "An instruction must not read from a location in physical
734 * regfile A or B that was written to by the previous
737 if (reads_too_soon_after_write(scoreboard
, n
->inst
))
740 if (writes_too_soon_after_write(devinfo
, scoreboard
, n
->inst
))
743 /* "A scoreboard wait must not occur in the first two
744 * instructions of a fragment shader. This is either the
745 * explicit Wait for Scoreboard signal or an implicit wait
746 * with the first tile-buffer read or write instruction."
748 if (pixel_scoreboard_too_soon(scoreboard
, inst
))
751 /* ldunif and ldvary both write r5, but ldunif does so a tick
752 * sooner. If the ldvary's r5 wasn't used, then ldunif might
753 * otherwise get scheduled so ldunif and ldvary try to update
754 * r5 in the same tick.
756 * XXX perf: To get good pipelining of a sequence of varying
757 * loads, we need to figure out how to pair the ldvary signal
758 * up to the instruction before the last r5 user in the
759 * previous ldvary sequence. Currently, it usually pairs with
762 if ((inst
->sig
.ldunif
|| inst
->sig
.ldunifa
) &&
763 scoreboard
->tick
== scoreboard
->last_ldvary_tick
+ 1) {
767 /* If we're trying to pair with another instruction, check
768 * that they're compatible.
771 /* Don't pair up a thread switch signal -- we'll
772 * handle pairing it when we pick it on its own.
777 if (prev_inst
->inst
->uniform
!= -1 &&
778 n
->inst
->uniform
!= -1)
781 /* Don't merge in something that will lock the TLB.
782 * Hopwefully what we have in inst will release some
783 * other instructions, allowing us to delay the
784 * TLB-locking instruction until later.
786 if (!scoreboard
->tlb_locked
&& qpu_inst_is_tlb(inst
))
789 struct v3d_qpu_instr merged_inst
;
790 if (!qpu_merge_inst(devinfo
, &merged_inst
,
791 &prev_inst
->inst
->qpu
, inst
)) {
796 int prio
= get_instruction_priority(inst
);
798 /* Found a valid instruction. If nothing better comes along,
807 if (prio
> chosen_prio
) {
810 } else if (prio
< chosen_prio
) {
814 if (n
->delay
> chosen
->delay
) {
817 } else if (n
->delay
< chosen
->delay
) {
826 update_scoreboard_for_magic_waddr(struct choose_scoreboard
*scoreboard
,
827 enum v3d_qpu_waddr waddr
)
829 if (v3d_qpu_magic_waddr_is_sfu(waddr
))
830 scoreboard
->last_magic_sfu_write_tick
= scoreboard
->tick
;
834 update_scoreboard_for_chosen(struct choose_scoreboard
*scoreboard
,
835 const struct v3d_qpu_instr
*inst
)
837 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
)
840 assert(inst
->type
== V3D_QPU_INSTR_TYPE_ALU
);
842 if (inst
->alu
.add
.op
!= V3D_QPU_A_NOP
) {
843 if (inst
->alu
.add
.magic_write
) {
844 update_scoreboard_for_magic_waddr(scoreboard
,
845 inst
->alu
.add
.waddr
);
849 if (inst
->alu
.mul
.op
!= V3D_QPU_M_NOP
) {
850 if (inst
->alu
.mul
.magic_write
) {
851 update_scoreboard_for_magic_waddr(scoreboard
,
852 inst
->alu
.mul
.waddr
);
856 if (inst
->sig
.ldvary
)
857 scoreboard
->last_ldvary_tick
= scoreboard
->tick
;
859 if (qpu_inst_is_tlb(inst
))
860 scoreboard
->tlb_locked
= true;
864 dump_state(const struct v3d_device_info
*devinfo
,
865 struct list_head
*schedule_list
)
867 list_for_each_entry(struct schedule_node
, n
, schedule_list
, link
) {
868 fprintf(stderr
, " t=%4d: ", n
->unblocked_time
);
869 v3d_qpu_dump(devinfo
, &n
->inst
->qpu
);
870 fprintf(stderr
, "\n");
872 for (int i
= 0; i
< n
->child_count
; i
++) {
873 struct schedule_node
*child
= n
->children
[i
].node
;
877 fprintf(stderr
, " - ");
878 v3d_qpu_dump(devinfo
, &child
->inst
->qpu
);
879 fprintf(stderr
, " (%d parents, %c)\n",
881 n
->children
[i
].write_after_read
? 'w' : 'r');
886 static uint32_t magic_waddr_latency(enum v3d_qpu_waddr waddr
,
887 const struct v3d_qpu_instr
*after
)
889 /* Apply some huge latency between texture fetch requests and getting
890 * their results back.
892 * FIXME: This is actually pretty bogus. If we do:
901 * we count that as worse than
910 * because we associate the first load_tmu0 with the *second* tmu0_s.
912 if (v3d_qpu_magic_waddr_is_tmu(waddr
) && v3d_qpu_waits_on_tmu(after
))
915 /* Assume that anything depending on us is consuming the SFU result. */
916 if (v3d_qpu_magic_waddr_is_sfu(waddr
))
923 instruction_latency(struct schedule_node
*before
, struct schedule_node
*after
)
925 const struct v3d_qpu_instr
*before_inst
= &before
->inst
->qpu
;
926 const struct v3d_qpu_instr
*after_inst
= &after
->inst
->qpu
;
927 uint32_t latency
= 1;
929 if (before_inst
->type
!= V3D_QPU_INSTR_TYPE_ALU
||
930 after_inst
->type
!= V3D_QPU_INSTR_TYPE_ALU
)
933 if (before_inst
->alu
.add
.magic_write
) {
934 latency
= MAX2(latency
,
935 magic_waddr_latency(before_inst
->alu
.add
.waddr
,
939 if (before_inst
->alu
.mul
.magic_write
) {
940 latency
= MAX2(latency
,
941 magic_waddr_latency(before_inst
->alu
.mul
.waddr
,
948 /** Recursive computation of the delay member of a node. */
950 compute_delay(struct schedule_node
*n
)
952 if (!n
->child_count
) {
955 for (int i
= 0; i
< n
->child_count
; i
++) {
956 if (!n
->children
[i
].node
->delay
)
957 compute_delay(n
->children
[i
].node
);
958 n
->delay
= MAX2(n
->delay
,
959 n
->children
[i
].node
->delay
+
960 instruction_latency(n
, n
->children
[i
].node
));
966 mark_instruction_scheduled(struct list_head
*schedule_list
,
968 struct schedule_node
*node
,
974 for (int i
= node
->child_count
- 1; i
>= 0; i
--) {
975 struct schedule_node
*child
=
976 node
->children
[i
].node
;
981 if (war_only
&& !node
->children
[i
].write_after_read
)
984 /* If the requirement is only that the node not appear before
985 * the last read of its destination, then it can be scheduled
986 * immediately after (or paired with!) the thing reading the
989 uint32_t latency
= 0;
991 latency
= instruction_latency(node
,
992 node
->children
[i
].node
);
995 child
->unblocked_time
= MAX2(child
->unblocked_time
,
997 child
->parent_count
--;
998 if (child
->parent_count
== 0)
999 list_add(&child
->link
, schedule_list
);
1001 node
->children
[i
].node
= NULL
;
1006 insert_scheduled_instruction(struct v3d_compile
*c
,
1007 struct qblock
*block
,
1008 struct choose_scoreboard
*scoreboard
,
1011 list_addtail(&inst
->link
, &block
->instructions
);
1013 update_scoreboard_for_chosen(scoreboard
, &inst
->qpu
);
1014 c
->qpu_inst_count
++;
1018 static struct qinst
*
1021 struct qreg undef
= { QFILE_NULL
, 0 };
1022 struct qinst
*qinst
= vir_add_inst(V3D_QPU_A_NOP
, undef
, undef
, undef
);
1028 emit_nop(struct v3d_compile
*c
, struct qblock
*block
,
1029 struct choose_scoreboard
*scoreboard
)
1031 insert_scheduled_instruction(c
, block
, scoreboard
, vir_nop());
1035 qpu_instruction_valid_in_thrend_slot(struct v3d_compile
*c
,
1036 const struct qinst
*qinst
, int slot
)
1038 const struct v3d_qpu_instr
*inst
= &qinst
->qpu
;
1040 /* Only TLB Z writes are prohibited in the last slot, but we don't
1041 * have those flagged so prohibit all TLB ops for now.
1043 if (slot
== 2 && qpu_inst_is_tlb(inst
))
1046 if (slot
> 0 && qinst
->uniform
!= ~0)
1049 if (v3d_qpu_uses_vpm(inst
))
1052 if (inst
->sig
.ldvary
)
1055 if (inst
->type
== V3D_QPU_INSTR_TYPE_ALU
) {
1056 /* GFXH-1625: TMUWT not allowed in the final instruction. */
1057 if (slot
== 2 && inst
->alu
.add
.op
== V3D_QPU_A_TMUWT
)
1060 /* No writing physical registers at the end. */
1061 if (!inst
->alu
.add
.magic_write
||
1062 !inst
->alu
.mul
.magic_write
) {
1066 if (c
->devinfo
->ver
< 40 && inst
->alu
.add
.op
== V3D_QPU_A_SETMSF
)
1069 /* RF0-2 might be overwritten during the delay slots by
1070 * fragment shader setup.
1072 if (inst
->raddr_a
< 3 &&
1073 (inst
->alu
.add
.a
== V3D_QPU_MUX_A
||
1074 inst
->alu
.add
.b
== V3D_QPU_MUX_A
||
1075 inst
->alu
.mul
.a
== V3D_QPU_MUX_A
||
1076 inst
->alu
.mul
.b
== V3D_QPU_MUX_A
)) {
1080 if (inst
->raddr_b
< 3 &&
1081 !inst
->sig
.small_imm
&&
1082 (inst
->alu
.add
.a
== V3D_QPU_MUX_B
||
1083 inst
->alu
.add
.b
== V3D_QPU_MUX_B
||
1084 inst
->alu
.mul
.a
== V3D_QPU_MUX_B
||
1085 inst
->alu
.mul
.b
== V3D_QPU_MUX_B
)) {
1094 valid_thrsw_sequence(struct v3d_compile
*c
, struct choose_scoreboard
*scoreboard
,
1095 struct qinst
*qinst
, int instructions_in_sequence
,
1098 /* No emitting our thrsw while the previous thrsw hasn't happened yet. */
1099 if (scoreboard
->last_thrsw_tick
+ 3 >
1100 scoreboard
->tick
- instructions_in_sequence
) {
1104 for (int slot
= 0; slot
< instructions_in_sequence
; slot
++) {
1105 /* No scheduling SFU when the result would land in the other
1106 * thread. The simulator complains for safety, though it
1107 * would only occur for dead code in our case.
1110 qinst
->qpu
.type
== V3D_QPU_INSTR_TYPE_ALU
&&
1111 (v3d_qpu_magic_waddr_is_sfu(qinst
->qpu
.alu
.add
.waddr
) ||
1112 v3d_qpu_magic_waddr_is_sfu(qinst
->qpu
.alu
.mul
.waddr
))) {
1116 if (slot
> 0 && qinst
->qpu
.sig
.ldvary
)
1120 !qpu_instruction_valid_in_thrend_slot(c
, qinst
, slot
)) {
1124 /* Note that the list is circular, so we can only do this up
1125 * to instructions_in_sequence.
1127 qinst
= (struct qinst
*)qinst
->link
.next
;
1134 * Emits a THRSW signal in the stream, trying to move it up to pair with
1135 * another instruction.
1138 emit_thrsw(struct v3d_compile
*c
,
1139 struct qblock
*block
,
1140 struct choose_scoreboard
*scoreboard
,
1146 /* There should be nothing in a thrsw inst being scheduled other than
1149 assert(inst
->qpu
.type
== V3D_QPU_INSTR_TYPE_ALU
);
1150 assert(inst
->qpu
.alu
.add
.op
== V3D_QPU_A_NOP
);
1151 assert(inst
->qpu
.alu
.mul
.op
== V3D_QPU_M_NOP
);
1153 /* Find how far back into previous instructions we can put the THRSW. */
1154 int slots_filled
= 0;
1155 struct qinst
*merge_inst
= NULL
;
1156 vir_for_each_inst_rev(prev_inst
, block
) {
1157 struct v3d_qpu_sig sig
= prev_inst
->qpu
.sig
;
1159 uint32_t packed_sig
;
1161 if (!v3d_qpu_sig_pack(c
->devinfo
, &sig
, &packed_sig
))
1164 if (!valid_thrsw_sequence(c
, scoreboard
,
1165 prev_inst
, slots_filled
+ 1,
1170 merge_inst
= prev_inst
;
1171 if (++slots_filled
== 3)
1175 bool needs_free
= false;
1177 merge_inst
->qpu
.sig
.thrsw
= true;
1179 scoreboard
->last_thrsw_tick
= scoreboard
->tick
- slots_filled
;
1181 scoreboard
->last_thrsw_tick
= scoreboard
->tick
;
1182 insert_scheduled_instruction(c
, block
, scoreboard
, inst
);
1188 /* Insert any extra delay slot NOPs we need. */
1189 for (int i
= 0; i
< 3 - slots_filled
; i
++) {
1190 emit_nop(c
, block
, scoreboard
);
1194 /* If we're emitting the last THRSW (other than program end), then
1195 * signal that to the HW by emitting two THRSWs in a row.
1197 if (inst
->is_last_thrsw
) {
1198 struct qinst
*second_inst
=
1199 (struct qinst
*)merge_inst
->link
.next
;
1200 second_inst
->qpu
.sig
.thrsw
= true;
1203 /* If we put our THRSW into another instruction, free up the
1204 * instruction that didn't end up scheduled into the list.
1213 schedule_instructions(struct v3d_compile
*c
,
1214 struct choose_scoreboard
*scoreboard
,
1215 struct qblock
*block
,
1216 struct list_head
*schedule_list
,
1217 enum quniform_contents
*orig_uniform_contents
,
1218 uint32_t *orig_uniform_data
,
1219 uint32_t *next_uniform
)
1221 const struct v3d_device_info
*devinfo
= c
->devinfo
;
1225 fprintf(stderr
, "initial deps:\n");
1226 dump_state(devinfo
, schedule_list
);
1227 fprintf(stderr
, "\n");
1230 /* Remove non-DAG heads from the list. */
1231 list_for_each_entry_safe(struct schedule_node
, n
, schedule_list
, link
) {
1232 if (n
->parent_count
!= 0)
1236 while (!list_empty(schedule_list
)) {
1237 struct schedule_node
*chosen
=
1238 choose_instruction_to_schedule(devinfo
,
1242 struct schedule_node
*merge
= NULL
;
1244 /* If there are no valid instructions to schedule, drop a NOP
1247 struct qinst
*qinst
= chosen
? chosen
->inst
: vir_nop();
1248 struct v3d_qpu_instr
*inst
= &qinst
->qpu
;
1251 fprintf(stderr
, "t=%4d: current list:\n",
1253 dump_state(devinfo
, schedule_list
);
1254 fprintf(stderr
, "t=%4d: chose: ", time
);
1255 v3d_qpu_dump(devinfo
, inst
);
1256 fprintf(stderr
, "\n");
1259 /* We can't mark_instruction_scheduled() the chosen inst until
1260 * we're done identifying instructions to merge, so put the
1261 * merged instructions on a list for a moment.
1263 struct list_head merged_list
;
1264 list_inithead(&merged_list
);
1266 /* Schedule this instruction onto the QPU list. Also try to
1267 * find an instruction to pair with it.
1270 time
= MAX2(chosen
->unblocked_time
, time
);
1271 list_del(&chosen
->link
);
1272 mark_instruction_scheduled(schedule_list
, time
,
1276 choose_instruction_to_schedule(devinfo
,
1280 time
= MAX2(merge
->unblocked_time
, time
);
1281 list_del(&merge
->link
);
1282 list_addtail(&merge
->link
, &merged_list
);
1283 (void)qpu_merge_inst(devinfo
, inst
,
1284 inst
, &merge
->inst
->qpu
);
1285 if (merge
->inst
->uniform
!= -1) {
1286 chosen
->inst
->uniform
=
1287 merge
->inst
->uniform
;
1291 fprintf(stderr
, "t=%4d: merging: ",
1293 v3d_qpu_dump(devinfo
, &merge
->inst
->qpu
);
1294 fprintf(stderr
, "\n");
1295 fprintf(stderr
, " result: ");
1296 v3d_qpu_dump(devinfo
, inst
);
1297 fprintf(stderr
, "\n");
1302 /* Update the uniform index for the rewritten location --
1303 * branch target updating will still need to change
1304 * c->uniform_data[] using this index.
1306 if (qinst
->uniform
!= -1) {
1307 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
)
1308 block
->branch_uniform
= *next_uniform
;
1310 c
->uniform_data
[*next_uniform
] =
1311 orig_uniform_data
[qinst
->uniform
];
1312 c
->uniform_contents
[*next_uniform
] =
1313 orig_uniform_contents
[qinst
->uniform
];
1314 qinst
->uniform
= *next_uniform
;
1319 fprintf(stderr
, "\n");
1322 /* Now that we've scheduled a new instruction, some of its
1323 * children can be promoted to the list of instructions ready to
1324 * be scheduled. Update the children's unblocked time for this
1325 * DAG edge as we do so.
1327 mark_instruction_scheduled(schedule_list
, time
, chosen
, false);
1328 list_for_each_entry(struct schedule_node
, merge
, &merged_list
,
1330 mark_instruction_scheduled(schedule_list
, time
, merge
,
1333 /* The merged VIR instruction doesn't get re-added to the
1334 * block, so free it now.
1339 if (inst
->sig
.thrsw
) {
1340 time
+= emit_thrsw(c
, block
, scoreboard
, qinst
, false);
1342 insert_scheduled_instruction(c
, block
,
1345 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
) {
1346 block
->branch_qpu_ip
= c
->qpu_inst_count
- 1;
1347 /* Fill the delay slots.
1349 * We should fill these with actual instructions,
1350 * instead, but that will probably need to be done
1351 * after this, once we know what the leading
1352 * instructions of the successors are (so we can
1353 * handle A/B register file write latency)
1355 for (int i
= 0; i
< 3; i
++)
1356 emit_nop(c
, block
, scoreboard
);
1365 qpu_schedule_instructions_block(struct v3d_compile
*c
,
1366 struct choose_scoreboard
*scoreboard
,
1367 struct qblock
*block
,
1368 enum quniform_contents
*orig_uniform_contents
,
1369 uint32_t *orig_uniform_data
,
1370 uint32_t *next_uniform
)
1372 void *mem_ctx
= ralloc_context(NULL
);
1373 struct list_head schedule_list
;
1375 list_inithead(&schedule_list
);
1377 /* Wrap each instruction in a scheduler structure. */
1378 while (!list_empty(&block
->instructions
)) {
1379 struct qinst
*qinst
= (struct qinst
*)block
->instructions
.next
;
1380 struct schedule_node
*n
=
1381 rzalloc(mem_ctx
, struct schedule_node
);
1385 list_del(&qinst
->link
);
1386 list_addtail(&n
->link
, &schedule_list
);
1389 calculate_forward_deps(c
, &schedule_list
);
1390 calculate_reverse_deps(c
, &schedule_list
);
1392 list_for_each_entry(struct schedule_node
, n
, &schedule_list
, link
) {
1396 uint32_t cycles
= schedule_instructions(c
, scoreboard
, block
,
1398 orig_uniform_contents
,
1402 ralloc_free(mem_ctx
);
1408 qpu_set_branch_targets(struct v3d_compile
*c
)
1410 vir_for_each_block(block
, c
) {
1411 /* The end block of the program has no branch. */
1412 if (!block
->successors
[0])
1415 /* If there was no branch instruction, then the successor
1416 * block must follow immediately after this one.
1418 if (block
->branch_qpu_ip
== ~0) {
1419 assert(block
->end_qpu_ip
+ 1 ==
1420 block
->successors
[0]->start_qpu_ip
);
1424 /* Walk back through the delay slots to find the branch
1427 struct list_head
*entry
= block
->instructions
.prev
;
1428 for (int i
= 0; i
< 3; i
++)
1429 entry
= entry
->prev
;
1430 struct qinst
*branch
= container_of(entry
, branch
, link
);
1431 assert(branch
->qpu
.type
== V3D_QPU_INSTR_TYPE_BRANCH
);
1433 /* Make sure that the if-we-don't-jump
1434 * successor was scheduled just after the
1437 assert(!block
->successors
[1] ||
1438 block
->successors
[1]->start_qpu_ip
==
1439 block
->branch_qpu_ip
+ 4);
1441 branch
->qpu
.branch
.offset
=
1442 ((block
->successors
[0]->start_qpu_ip
-
1443 (block
->branch_qpu_ip
+ 4)) *
1446 /* Set up the relative offset to jump in the
1449 * Use a temporary here, because
1450 * uniform_data[inst->uniform] may be shared
1451 * between multiple instructions.
1453 assert(c
->uniform_contents
[branch
->uniform
] == QUNIFORM_CONSTANT
);
1454 c
->uniform_data
[branch
->uniform
] =
1455 (block
->successors
[0]->start_uniform
-
1456 (block
->branch_uniform
+ 1)) * 4;
1461 v3d_qpu_schedule_instructions(struct v3d_compile
*c
)
1463 const struct v3d_device_info
*devinfo
= c
->devinfo
;
1464 struct qblock
*end_block
= list_last_entry(&c
->blocks
,
1465 struct qblock
, link
);
1467 /* We reorder the uniforms as we schedule instructions, so save the
1468 * old data off and replace it.
1470 uint32_t *uniform_data
= c
->uniform_data
;
1471 enum quniform_contents
*uniform_contents
= c
->uniform_contents
;
1472 c
->uniform_contents
= ralloc_array(c
, enum quniform_contents
,
1474 c
->uniform_data
= ralloc_array(c
, uint32_t, c
->num_uniforms
);
1475 c
->uniform_array_size
= c
->num_uniforms
;
1476 uint32_t next_uniform
= 0;
1478 struct choose_scoreboard scoreboard
;
1479 memset(&scoreboard
, 0, sizeof(scoreboard
));
1480 scoreboard
.last_ldvary_tick
= -10;
1481 scoreboard
.last_magic_sfu_write_tick
= -10;
1482 scoreboard
.last_uniforms_reset_tick
= -10;
1483 scoreboard
.last_thrsw_tick
= -10;
1486 fprintf(stderr
, "Pre-schedule instructions\n");
1487 vir_for_each_block(block
, c
) {
1488 fprintf(stderr
, "BLOCK %d\n", block
->index
);
1489 list_for_each_entry(struct qinst
, qinst
,
1490 &block
->instructions
, link
) {
1491 v3d_qpu_dump(devinfo
, &qinst
->qpu
);
1492 fprintf(stderr
, "\n");
1495 fprintf(stderr
, "\n");
1498 uint32_t cycles
= 0;
1499 vir_for_each_block(block
, c
) {
1500 block
->start_qpu_ip
= c
->qpu_inst_count
;
1501 block
->branch_qpu_ip
= ~0;
1502 block
->start_uniform
= next_uniform
;
1504 cycles
+= qpu_schedule_instructions_block(c
,
1511 block
->end_qpu_ip
= c
->qpu_inst_count
- 1;
1514 /* Emit the program-end THRSW instruction. */;
1515 struct qinst
*thrsw
= vir_nop();
1516 thrsw
->qpu
.sig
.thrsw
= true;
1517 emit_thrsw(c
, end_block
, &scoreboard
, thrsw
, true);
1519 qpu_set_branch_targets(c
);
1521 assert(next_uniform
== c
->num_uniforms
);