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 add_write_dep(state
, &state
->last_tmu_write
, n
);
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
);
209 } else if (v3d_qpu_magic_waddr_is_sfu(waddr
)) {
210 /* Handled by v3d_qpu_writes_r4() check. */
213 case V3D_QPU_WADDR_R0
:
214 case V3D_QPU_WADDR_R1
:
215 case V3D_QPU_WADDR_R2
:
217 &state
->last_r
[waddr
- V3D_QPU_WADDR_R0
],
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. */
226 case V3D_QPU_WADDR_VPM
:
227 case V3D_QPU_WADDR_VPMU
:
228 add_write_dep(state
, &state
->last_vpm
, n
);
231 case V3D_QPU_WADDR_TLB
:
232 case V3D_QPU_WADDR_TLBU
:
233 add_write_dep(state
, &state
->last_tlb
, n
);
236 case V3D_QPU_WADDR_NOP
:
240 fprintf(stderr
, "Unknown waddr %d\n", waddr
);
247 process_cond_deps(struct schedule_state
*state
, struct schedule_node
*n
,
248 enum v3d_qpu_cond cond
)
250 if (cond
!= V3D_QPU_COND_NONE
)
251 add_read_dep(state
, state
->last_sf
, n
);
255 process_pf_deps(struct schedule_state
*state
, struct schedule_node
*n
,
258 if (pf
!= V3D_QPU_PF_NONE
)
259 add_write_dep(state
, &state
->last_sf
, n
);
263 process_uf_deps(struct schedule_state
*state
, struct schedule_node
*n
,
266 if (uf
!= V3D_QPU_UF_NONE
)
267 add_write_dep(state
, &state
->last_sf
, n
);
271 * Common code for dependencies that need to be tracked both forward and
274 * This is for things like "all reads of r4 have to happen between the r4
275 * writes that surround them".
278 calculate_deps(struct schedule_state
*state
, struct schedule_node
*n
)
280 const struct v3d_device_info
*devinfo
= state
->devinfo
;
281 struct qinst
*qinst
= n
->inst
;
282 struct v3d_qpu_instr
*inst
= &qinst
->qpu
;
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
);
293 add_write_dep(state
, &state
->last_unif
, n
);
297 assert(inst
->type
== V3D_QPU_INSTR_TYPE_ALU
);
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
);
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
);
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
);
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
);
324 case V3D_QPU_A_VPMWT
:
325 add_read_dep(state
, state
->last_vpm
, n
);
329 add_read_dep(state
, state
->last_tlb
, n
);
332 case V3D_QPU_A_SETMSF
:
333 case V3D_QPU_A_SETREVF
:
334 add_write_dep(state
, &state
->last_tlb
, n
);
337 case V3D_QPU_A_FLAPUSH
:
338 case V3D_QPU_A_FLBPUSH
:
340 case V3D_QPU_A_VFLNA
:
342 case V3D_QPU_A_VFLNB
:
343 add_read_dep(state
, state
->last_sf
, n
);
346 case V3D_QPU_A_FLPOP
:
347 add_write_dep(state
, &state
->last_sf
, n
);
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
362 add_write_dep(state
, &state
->last_rtop
, n
);
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
);
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
);
376 if (v3d_qpu_sig_writes_address(devinfo
, &inst
->sig
)) {
377 process_waddr_deps(state
, n
, inst
->sig_addr
,
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
);
388 if (inst
->sig
.thrsw
) {
389 /* All accumulator contents and flags are undefined after the
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
);
396 /* Scoreboard-locking operations have to stay after the last
399 add_write_dep(state
, &state
->last_tlb
, n
);
401 add_write_dep(state
, &state
->last_tmu_write
, n
);
402 add_write_dep(state
, &state
->last_tmu_config
, n
);
405 if (inst
->sig
.ldtmu
) {
406 /* TMU loads are coming from a FIFO, so ordering is important.
408 add_write_dep(state
, &state
->last_tmu_write
, n
);
411 if (inst
->sig
.wrtmuc
)
412 add_write_dep(state
, &state
->last_tmu_config
, n
);
414 if (inst
->sig
.ldtlb
| inst
->sig
.ldtlbu
)
415 add_read_dep(state
, state
->last_tlb
, n
);
418 add_write_dep(state
, &state
->last_vpm_read
, n
);
420 /* inst->sig.ldunif or sideband uniform read */
421 if (qinst
->uniform
!= ~0)
422 add_write_dep(state
, &state
->last_unif
, n
);
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
);
433 calculate_forward_deps(struct v3d_compile
*c
, struct list_head
*schedule_list
)
435 struct schedule_state state
;
437 memset(&state
, 0, sizeof(state
));
438 state
.devinfo
= c
->devinfo
;
441 list_for_each_entry(struct schedule_node
, node
, schedule_list
, link
)
442 calculate_deps(&state
, node
);
446 calculate_reverse_deps(struct v3d_compile
*c
, struct list_head
*schedule_list
)
448 struct list_head
*node
;
449 struct schedule_state state
;
451 memset(&state
, 0, sizeof(state
));
452 state
.devinfo
= c
->devinfo
;
455 for (node
= schedule_list
->prev
; schedule_list
!= node
; node
= node
->prev
) {
456 calculate_deps(&state
, (struct schedule_node
*)node
);
460 struct choose_scoreboard
{
462 int last_magic_sfu_write_tick
;
463 int last_ldvary_tick
;
464 int last_uniforms_reset_tick
;
469 mux_reads_too_soon(struct choose_scoreboard
*scoreboard
,
470 const struct v3d_qpu_instr
*inst
, enum v3d_qpu_mux mux
)
474 if (scoreboard
->tick
- scoreboard
->last_magic_sfu_write_tick
<= 2)
479 if (scoreboard
->tick
- scoreboard
->last_ldvary_tick
<= 1)
490 reads_too_soon_after_write(struct choose_scoreboard
*scoreboard
,
493 const struct v3d_qpu_instr
*inst
= &qinst
->qpu
;
495 /* XXX: Branching off of raddr. */
496 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
)
499 assert(inst
->type
== V3D_QPU_INSTR_TYPE_ALU
);
501 if (inst
->alu
.add
.op
!= V3D_QPU_A_NOP
) {
502 if (v3d_qpu_add_op_num_src(inst
->alu
.add
.op
) > 0 &&
503 mux_reads_too_soon(scoreboard
, inst
, inst
->alu
.add
.a
)) {
506 if (v3d_qpu_add_op_num_src(inst
->alu
.add
.op
) > 1 &&
507 mux_reads_too_soon(scoreboard
, inst
, inst
->alu
.add
.b
)) {
512 if (inst
->alu
.mul
.op
!= V3D_QPU_M_NOP
) {
513 if (v3d_qpu_mul_op_num_src(inst
->alu
.mul
.op
) > 0 &&
514 mux_reads_too_soon(scoreboard
, inst
, inst
->alu
.mul
.a
)) {
517 if (v3d_qpu_mul_op_num_src(inst
->alu
.mul
.op
) > 1 &&
518 mux_reads_too_soon(scoreboard
, inst
, inst
->alu
.mul
.b
)) {
529 writes_too_soon_after_write(const struct v3d_device_info
*devinfo
,
530 struct choose_scoreboard
*scoreboard
,
533 const struct v3d_qpu_instr
*inst
= &qinst
->qpu
;
535 /* Don't schedule any other r4 write too soon after an SFU write.
536 * This would normally be prevented by dependency tracking, but might
537 * occur if a dead SFU computation makes it to scheduling.
539 if (scoreboard
->tick
- scoreboard
->last_magic_sfu_write_tick
< 2 &&
540 v3d_qpu_writes_r4(devinfo
, inst
))
547 pixel_scoreboard_too_soon(struct choose_scoreboard
*scoreboard
,
548 const struct v3d_qpu_instr
*inst
)
550 return (scoreboard
->tick
== 0 && qpu_inst_is_tlb(inst
));
554 get_instruction_priority(const struct v3d_qpu_instr
*inst
)
556 uint32_t baseline_score
;
557 uint32_t next_score
= 0;
559 /* Schedule TLB operations as late as possible, to get more
560 * parallelism between shaders.
562 if (qpu_inst_is_tlb(inst
))
566 /* Schedule texture read results collection late to hide latency. */
571 /* Default score for things that aren't otherwise special. */
572 baseline_score
= next_score
;
575 /* Schedule texture read setup early to hide their latency better. */
576 if (v3d_qpu_writes_tmu(inst
))
580 return baseline_score
;
584 qpu_magic_waddr_is_periph(enum v3d_qpu_waddr waddr
)
586 return (v3d_qpu_magic_waddr_is_tmu(waddr
) ||
587 v3d_qpu_magic_waddr_is_sfu(waddr
) ||
588 v3d_qpu_magic_waddr_is_tlb(waddr
) ||
589 v3d_qpu_magic_waddr_is_vpm(waddr
) ||
590 v3d_qpu_magic_waddr_is_tsy(waddr
));
594 qpu_accesses_peripheral(const struct v3d_qpu_instr
*inst
)
596 if (v3d_qpu_uses_vpm(inst
))
598 if (v3d_qpu_uses_sfu(inst
))
601 if (inst
->type
== V3D_QPU_INSTR_TYPE_ALU
) {
602 if (inst
->alu
.add
.op
!= V3D_QPU_A_NOP
&&
603 inst
->alu
.add
.magic_write
&&
604 qpu_magic_waddr_is_periph(inst
->alu
.add
.waddr
)) {
608 if (inst
->alu
.mul
.op
!= V3D_QPU_M_NOP
&&
609 inst
->alu
.mul
.magic_write
&&
610 qpu_magic_waddr_is_periph(inst
->alu
.mul
.waddr
)) {
615 return (inst
->sig
.ldvpm
||
623 qpu_merge_inst(const struct v3d_device_info
*devinfo
,
624 struct v3d_qpu_instr
*result
,
625 const struct v3d_qpu_instr
*a
,
626 const struct v3d_qpu_instr
*b
)
628 if (a
->type
!= V3D_QPU_INSTR_TYPE_ALU
||
629 b
->type
!= V3D_QPU_INSTR_TYPE_ALU
) {
633 /* Can't do more than one peripheral access in an instruction.
635 * XXX: V3D 4.1 allows TMU read along with a VPM read or write, and
636 * WRTMUC with a TMU magic register write (other than tmuc).
638 if (qpu_accesses_peripheral(a
) && qpu_accesses_peripheral(b
))
641 struct v3d_qpu_instr merge
= *a
;
643 if (b
->alu
.add
.op
!= V3D_QPU_A_NOP
) {
644 if (a
->alu
.add
.op
!= V3D_QPU_A_NOP
)
646 merge
.alu
.add
= b
->alu
.add
;
648 merge
.flags
.ac
= b
->flags
.ac
;
649 merge
.flags
.apf
= b
->flags
.apf
;
650 merge
.flags
.auf
= b
->flags
.auf
;
653 if (b
->alu
.mul
.op
!= V3D_QPU_M_NOP
) {
654 if (a
->alu
.mul
.op
!= V3D_QPU_M_NOP
)
656 merge
.alu
.mul
= b
->alu
.mul
;
658 merge
.flags
.mc
= b
->flags
.mc
;
659 merge
.flags
.mpf
= b
->flags
.mpf
;
660 merge
.flags
.muf
= b
->flags
.muf
;
663 if (v3d_qpu_uses_mux(b
, V3D_QPU_MUX_A
)) {
664 if (v3d_qpu_uses_mux(a
, V3D_QPU_MUX_A
) &&
665 a
->raddr_a
!= b
->raddr_a
) {
668 merge
.raddr_a
= b
->raddr_a
;
671 if (v3d_qpu_uses_mux(b
, V3D_QPU_MUX_B
)) {
672 if (v3d_qpu_uses_mux(a
, V3D_QPU_MUX_B
) &&
673 a
->raddr_b
!= b
->raddr_b
) {
676 merge
.raddr_b
= b
->raddr_b
;
679 merge
.sig
.thrsw
|= b
->sig
.thrsw
;
680 merge
.sig
.ldunif
|= b
->sig
.ldunif
;
681 merge
.sig
.ldunifrf
|= b
->sig
.ldunifrf
;
682 merge
.sig
.ldunifa
|= b
->sig
.ldunifa
;
683 merge
.sig
.ldunifarf
|= b
->sig
.ldunifarf
;
684 merge
.sig
.ldtmu
|= b
->sig
.ldtmu
;
685 merge
.sig
.ldvary
|= b
->sig
.ldvary
;
686 merge
.sig
.ldvpm
|= b
->sig
.ldvpm
;
687 merge
.sig
.small_imm
|= b
->sig
.small_imm
;
688 merge
.sig
.ldtlb
|= b
->sig
.ldtlb
;
689 merge
.sig
.ldtlbu
|= b
->sig
.ldtlbu
;
690 merge
.sig
.ucb
|= b
->sig
.ucb
;
691 merge
.sig
.rotate
|= b
->sig
.rotate
;
692 merge
.sig
.wrtmuc
|= b
->sig
.wrtmuc
;
694 if (v3d_qpu_sig_writes_address(devinfo
, &a
->sig
) &&
695 v3d_qpu_sig_writes_address(devinfo
, &b
->sig
))
697 merge
.sig_addr
|= b
->sig_addr
;
698 merge
.sig_magic
|= b
->sig_magic
;
701 bool ok
= v3d_qpu_instr_pack(devinfo
, &merge
, &packed
);
704 /* No modifying the real instructions on failure. */
705 assert(ok
|| (a
!= result
&& b
!= result
));
710 static struct schedule_node
*
711 choose_instruction_to_schedule(const struct v3d_device_info
*devinfo
,
712 struct choose_scoreboard
*scoreboard
,
713 struct list_head
*schedule_list
,
714 struct schedule_node
*prev_inst
)
716 struct schedule_node
*chosen
= NULL
;
719 /* Don't pair up anything with a thread switch signal -- emit_thrsw()
720 * will handle pairing it along with filling the delay slots.
723 if (prev_inst
->inst
->qpu
.sig
.thrsw
)
727 list_for_each_entry(struct schedule_node
, n
, schedule_list
, link
) {
728 const struct v3d_qpu_instr
*inst
= &n
->inst
->qpu
;
730 /* Don't choose the branch instruction until it's the last one
731 * left. We'll move it up to fit its delay slots after we
734 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
&&
735 !list_is_singular(schedule_list
)) {
739 /* "An instruction must not read from a location in physical
740 * regfile A or B that was written to by the previous
743 if (reads_too_soon_after_write(scoreboard
, n
->inst
))
746 if (writes_too_soon_after_write(devinfo
, scoreboard
, n
->inst
))
749 /* "A scoreboard wait must not occur in the first two
750 * instructions of a fragment shader. This is either the
751 * explicit Wait for Scoreboard signal or an implicit wait
752 * with the first tile-buffer read or write instruction."
754 if (pixel_scoreboard_too_soon(scoreboard
, inst
))
757 /* ldunif and ldvary both write r5, but ldunif does so a tick
758 * sooner. If the ldvary's r5 wasn't used, then ldunif might
759 * otherwise get scheduled so ldunif and ldvary try to update
760 * r5 in the same tick.
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
) && after
->sig
.ldtmu
)
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
,
1095 struct qinst
*qinst
, int instructions_in_sequence
,
1098 for (int slot
= 0; slot
< instructions_in_sequence
; slot
++) {
1099 /* No scheduling SFU when the result would land in the other
1100 * thread. The simulator complains for safety, though it
1101 * would only occur for dead code in our case.
1104 qinst
->qpu
.type
== V3D_QPU_INSTR_TYPE_ALU
&&
1105 (v3d_qpu_magic_waddr_is_sfu(qinst
->qpu
.alu
.add
.waddr
) ||
1106 v3d_qpu_magic_waddr_is_sfu(qinst
->qpu
.alu
.mul
.waddr
))) {
1110 if (slot
> 0 && qinst
->qpu
.sig
.ldvary
)
1114 !qpu_instruction_valid_in_thrend_slot(c
, qinst
, slot
)) {
1118 /* Note that the list is circular, so we can only do this up
1119 * to instructions_in_sequence.
1121 qinst
= (struct qinst
*)qinst
->link
.next
;
1128 * Emits a THRSW signal in the stream, trying to move it up to pair with
1129 * another instruction.
1132 emit_thrsw(struct v3d_compile
*c
,
1133 struct qblock
*block
,
1134 struct choose_scoreboard
*scoreboard
,
1140 /* There should be nothing in a thrsw inst being scheduled other than
1143 assert(inst
->qpu
.type
== V3D_QPU_INSTR_TYPE_ALU
);
1144 assert(inst
->qpu
.alu
.add
.op
== V3D_QPU_A_NOP
);
1145 assert(inst
->qpu
.alu
.mul
.op
== V3D_QPU_M_NOP
);
1147 /* Find how far back into previous instructions we can put the THRSW. */
1148 int slots_filled
= 0;
1149 struct qinst
*merge_inst
= NULL
;
1150 vir_for_each_inst_rev(prev_inst
, block
) {
1151 struct v3d_qpu_sig sig
= prev_inst
->qpu
.sig
;
1153 uint32_t packed_sig
;
1155 if (!v3d_qpu_sig_pack(c
->devinfo
, &sig
, &packed_sig
))
1158 if (!valid_thrsw_sequence(c
, prev_inst
, slots_filled
+ 1,
1163 merge_inst
= prev_inst
;
1164 if (++slots_filled
== 3)
1168 bool needs_free
= false;
1170 merge_inst
->qpu
.sig
.thrsw
= true;
1173 insert_scheduled_instruction(c
, block
, scoreboard
, inst
);
1179 /* Insert any extra delay slot NOPs we need. */
1180 for (int i
= 0; i
< 3 - slots_filled
; i
++) {
1181 emit_nop(c
, block
, scoreboard
);
1185 /* If we're emitting the last THRSW (other than program end), then
1186 * signal that to the HW by emitting two THRSWs in a row.
1188 if (inst
->is_last_thrsw
) {
1189 struct qinst
*second_inst
=
1190 (struct qinst
*)merge_inst
->link
.next
;
1191 second_inst
->qpu
.sig
.thrsw
= true;
1194 /* If we put our THRSW into another instruction, free up the
1195 * instruction that didn't end up scheduled into the list.
1204 schedule_instructions(struct v3d_compile
*c
,
1205 struct choose_scoreboard
*scoreboard
,
1206 struct qblock
*block
,
1207 struct list_head
*schedule_list
,
1208 enum quniform_contents
*orig_uniform_contents
,
1209 uint32_t *orig_uniform_data
,
1210 uint32_t *next_uniform
)
1212 const struct v3d_device_info
*devinfo
= c
->devinfo
;
1216 fprintf(stderr
, "initial deps:\n");
1217 dump_state(devinfo
, schedule_list
);
1218 fprintf(stderr
, "\n");
1221 /* Remove non-DAG heads from the list. */
1222 list_for_each_entry_safe(struct schedule_node
, n
, schedule_list
, link
) {
1223 if (n
->parent_count
!= 0)
1227 while (!list_empty(schedule_list
)) {
1228 struct schedule_node
*chosen
=
1229 choose_instruction_to_schedule(devinfo
,
1233 struct schedule_node
*merge
= NULL
;
1235 /* If there are no valid instructions to schedule, drop a NOP
1238 struct qinst
*qinst
= chosen
? chosen
->inst
: vir_nop();
1239 struct v3d_qpu_instr
*inst
= &qinst
->qpu
;
1242 fprintf(stderr
, "t=%4d: current list:\n",
1244 dump_state(devinfo
, schedule_list
);
1245 fprintf(stderr
, "t=%4d: chose: ", time
);
1246 v3d_qpu_dump(devinfo
, inst
);
1247 fprintf(stderr
, "\n");
1250 /* We can't mark_instruction_scheduled() the chosen inst until
1251 * we're done identifying instructions to merge, so put the
1252 * merged instructions on a list for a moment.
1254 struct list_head merged_list
;
1255 list_inithead(&merged_list
);
1257 /* Schedule this instruction onto the QPU list. Also try to
1258 * find an instruction to pair with it.
1261 time
= MAX2(chosen
->unblocked_time
, time
);
1262 list_del(&chosen
->link
);
1263 mark_instruction_scheduled(schedule_list
, time
,
1267 choose_instruction_to_schedule(devinfo
,
1271 time
= MAX2(merge
->unblocked_time
, time
);
1272 list_del(&merge
->link
);
1273 list_addtail(&merge
->link
, &merged_list
);
1274 (void)qpu_merge_inst(devinfo
, inst
,
1275 inst
, &merge
->inst
->qpu
);
1276 if (merge
->inst
->uniform
!= -1) {
1277 chosen
->inst
->uniform
=
1278 merge
->inst
->uniform
;
1282 fprintf(stderr
, "t=%4d: merging: ",
1284 v3d_qpu_dump(devinfo
, &merge
->inst
->qpu
);
1285 fprintf(stderr
, "\n");
1286 fprintf(stderr
, " result: ");
1287 v3d_qpu_dump(devinfo
, inst
);
1288 fprintf(stderr
, "\n");
1293 /* Update the uniform index for the rewritten location --
1294 * branch target updating will still need to change
1295 * c->uniform_data[] using this index.
1297 if (qinst
->uniform
!= -1) {
1298 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
)
1299 block
->branch_uniform
= *next_uniform
;
1301 c
->uniform_data
[*next_uniform
] =
1302 orig_uniform_data
[qinst
->uniform
];
1303 c
->uniform_contents
[*next_uniform
] =
1304 orig_uniform_contents
[qinst
->uniform
];
1305 qinst
->uniform
= *next_uniform
;
1310 fprintf(stderr
, "\n");
1313 /* Now that we've scheduled a new instruction, some of its
1314 * children can be promoted to the list of instructions ready to
1315 * be scheduled. Update the children's unblocked time for this
1316 * DAG edge as we do so.
1318 mark_instruction_scheduled(schedule_list
, time
, chosen
, false);
1319 list_for_each_entry(struct schedule_node
, merge
, &merged_list
,
1321 mark_instruction_scheduled(schedule_list
, time
, merge
,
1324 /* The merged VIR instruction doesn't get re-added to the
1325 * block, so free it now.
1330 if (inst
->sig
.thrsw
) {
1331 time
+= emit_thrsw(c
, block
, scoreboard
, qinst
, false);
1333 insert_scheduled_instruction(c
, block
,
1336 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
) {
1337 block
->branch_qpu_ip
= c
->qpu_inst_count
- 1;
1338 /* Fill the delay slots.
1340 * We should fill these with actual instructions,
1341 * instead, but that will probably need to be done
1342 * after this, once we know what the leading
1343 * instructions of the successors are (so we can
1344 * handle A/B register file write latency)
1346 for (int i
= 0; i
< 3; i
++)
1347 emit_nop(c
, block
, scoreboard
);
1356 qpu_schedule_instructions_block(struct v3d_compile
*c
,
1357 struct choose_scoreboard
*scoreboard
,
1358 struct qblock
*block
,
1359 enum quniform_contents
*orig_uniform_contents
,
1360 uint32_t *orig_uniform_data
,
1361 uint32_t *next_uniform
)
1363 void *mem_ctx
= ralloc_context(NULL
);
1364 struct list_head schedule_list
;
1366 list_inithead(&schedule_list
);
1368 /* Wrap each instruction in a scheduler structure. */
1369 while (!list_empty(&block
->instructions
)) {
1370 struct qinst
*qinst
= (struct qinst
*)block
->instructions
.next
;
1371 struct schedule_node
*n
=
1372 rzalloc(mem_ctx
, struct schedule_node
);
1376 list_del(&qinst
->link
);
1377 list_addtail(&n
->link
, &schedule_list
);
1380 calculate_forward_deps(c
, &schedule_list
);
1381 calculate_reverse_deps(c
, &schedule_list
);
1383 list_for_each_entry(struct schedule_node
, n
, &schedule_list
, link
) {
1387 uint32_t cycles
= schedule_instructions(c
, scoreboard
, block
,
1389 orig_uniform_contents
,
1393 ralloc_free(mem_ctx
);
1399 qpu_set_branch_targets(struct v3d_compile
*c
)
1401 vir_for_each_block(block
, c
) {
1402 /* The end block of the program has no branch. */
1403 if (!block
->successors
[0])
1406 /* If there was no branch instruction, then the successor
1407 * block must follow immediately after this one.
1409 if (block
->branch_qpu_ip
== ~0) {
1410 assert(block
->end_qpu_ip
+ 1 ==
1411 block
->successors
[0]->start_qpu_ip
);
1415 /* Walk back through the delay slots to find the branch
1418 struct list_head
*entry
= block
->instructions
.prev
;
1419 for (int i
= 0; i
< 3; i
++)
1420 entry
= entry
->prev
;
1421 struct qinst
*branch
= container_of(entry
, branch
, link
);
1422 assert(branch
->qpu
.type
== V3D_QPU_INSTR_TYPE_BRANCH
);
1424 /* Make sure that the if-we-don't-jump
1425 * successor was scheduled just after the
1428 assert(!block
->successors
[1] ||
1429 block
->successors
[1]->start_qpu_ip
==
1430 block
->branch_qpu_ip
+ 4);
1432 branch
->qpu
.branch
.offset
=
1433 ((block
->successors
[0]->start_qpu_ip
-
1434 (block
->branch_qpu_ip
+ 4)) *
1437 /* Set up the relative offset to jump in the
1440 * Use a temporary here, because
1441 * uniform_data[inst->uniform] may be shared
1442 * between multiple instructions.
1444 assert(c
->uniform_contents
[branch
->uniform
] == QUNIFORM_CONSTANT
);
1445 c
->uniform_data
[branch
->uniform
] =
1446 (block
->successors
[0]->start_uniform
-
1447 (block
->branch_uniform
+ 1)) * 4;
1452 v3d_qpu_schedule_instructions(struct v3d_compile
*c
)
1454 const struct v3d_device_info
*devinfo
= c
->devinfo
;
1455 struct qblock
*end_block
= list_last_entry(&c
->blocks
,
1456 struct qblock
, link
);
1458 /* We reorder the uniforms as we schedule instructions, so save the
1459 * old data off and replace it.
1461 uint32_t *uniform_data
= c
->uniform_data
;
1462 enum quniform_contents
*uniform_contents
= c
->uniform_contents
;
1463 c
->uniform_contents
= ralloc_array(c
, enum quniform_contents
,
1465 c
->uniform_data
= ralloc_array(c
, uint32_t, c
->num_uniforms
);
1466 c
->uniform_array_size
= c
->num_uniforms
;
1467 uint32_t next_uniform
= 0;
1469 struct choose_scoreboard scoreboard
;
1470 memset(&scoreboard
, 0, sizeof(scoreboard
));
1471 scoreboard
.last_ldvary_tick
= -10;
1472 scoreboard
.last_magic_sfu_write_tick
= -10;
1473 scoreboard
.last_uniforms_reset_tick
= -10;
1476 fprintf(stderr
, "Pre-schedule instructions\n");
1477 vir_for_each_block(block
, c
) {
1478 fprintf(stderr
, "BLOCK %d\n", block
->index
);
1479 list_for_each_entry(struct qinst
, qinst
,
1480 &block
->instructions
, link
) {
1481 v3d_qpu_dump(devinfo
, &qinst
->qpu
);
1482 fprintf(stderr
, "\n");
1485 fprintf(stderr
, "\n");
1488 uint32_t cycles
= 0;
1489 vir_for_each_block(block
, c
) {
1490 block
->start_qpu_ip
= c
->qpu_inst_count
;
1491 block
->branch_qpu_ip
= ~0;
1492 block
->start_uniform
= next_uniform
;
1494 cycles
+= qpu_schedule_instructions_block(c
,
1501 block
->end_qpu_ip
= c
->qpu_inst_count
- 1;
1504 /* Emit the program-end THRSW instruction. */;
1505 struct qinst
*thrsw
= vir_nop();
1506 thrsw
->qpu
.sig
.thrsw
= true;
1507 emit_thrsw(c
, end_block
, &scoreboard
, thrsw
, true);
1509 qpu_set_branch_targets(c
);
1511 assert(next_uniform
== c
->num_uniforms
);