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 struct schedule_node
*last_r
[6];
82 struct schedule_node
*last_rf
[64];
83 struct schedule_node
*last_sf
;
84 struct schedule_node
*last_vpm_read
;
85 struct schedule_node
*last_tmu_write
;
86 struct schedule_node
*last_tlb
;
87 struct schedule_node
*last_vpm
;
88 struct schedule_node
*last_unif
;
89 struct schedule_node
*last_rtop
;
91 /* Estimated cycle when the current instruction would start. */
96 add_dep(struct schedule_state
*state
,
97 struct schedule_node
*before
,
98 struct schedule_node
*after
,
101 bool write_after_read
= !write
&& state
->dir
== R
;
103 if (!before
|| !after
)
106 assert(before
!= after
);
108 if (state
->dir
== R
) {
109 struct schedule_node
*t
= before
;
114 for (int i
= 0; i
< before
->child_count
; i
++) {
115 if (before
->children
[i
].node
== after
&&
116 (before
->children
[i
].write_after_read
== write_after_read
)) {
121 if (before
->child_array_size
<= before
->child_count
) {
122 before
->child_array_size
= MAX2(before
->child_array_size
* 2, 16);
123 before
->children
= reralloc(before
, before
->children
,
124 struct schedule_node_child
,
125 before
->child_array_size
);
128 before
->children
[before
->child_count
].node
= after
;
129 before
->children
[before
->child_count
].write_after_read
=
131 before
->child_count
++;
132 after
->parent_count
++;
136 add_read_dep(struct schedule_state
*state
,
137 struct schedule_node
*before
,
138 struct schedule_node
*after
)
140 add_dep(state
, before
, after
, false);
144 add_write_dep(struct schedule_state
*state
,
145 struct schedule_node
**before
,
146 struct schedule_node
*after
)
148 add_dep(state
, *before
, after
, true);
153 qpu_inst_is_tlb(const struct v3d_qpu_instr
*inst
)
155 if (inst
->type
!= V3D_QPU_INSTR_TYPE_ALU
)
158 if (inst
->alu
.add
.magic_write
&&
159 (inst
->alu
.add
.waddr
== V3D_QPU_WADDR_TLB
||
160 inst
->alu
.add
.waddr
== V3D_QPU_WADDR_TLBU
))
163 if (inst
->alu
.mul
.magic_write
&&
164 (inst
->alu
.mul
.waddr
== V3D_QPU_WADDR_TLB
||
165 inst
->alu
.mul
.waddr
== V3D_QPU_WADDR_TLBU
))
172 process_mux_deps(struct schedule_state
*state
, struct schedule_node
*n
,
173 enum v3d_qpu_mux mux
)
177 add_read_dep(state
, state
->last_rf
[n
->inst
->qpu
.raddr_a
], n
);
180 add_read_dep(state
, state
->last_rf
[n
->inst
->qpu
.raddr_b
], n
);
183 add_read_dep(state
, state
->last_r
[mux
- V3D_QPU_MUX_R0
], n
);
190 process_waddr_deps(struct schedule_state
*state
, struct schedule_node
*n
,
191 uint32_t waddr
, bool magic
)
194 add_write_dep(state
, &state
->last_rf
[waddr
], n
);
195 } else if (v3d_qpu_magic_waddr_is_tmu(waddr
)) {
196 add_write_dep(state
, &state
->last_tmu_write
, n
);
197 } else if (v3d_qpu_magic_waddr_is_sfu(waddr
)) {
198 /* Handled by v3d_qpu_writes_r4() check. */
201 case V3D_QPU_WADDR_R0
:
202 case V3D_QPU_WADDR_R1
:
203 case V3D_QPU_WADDR_R2
:
204 case V3D_QPU_WADDR_R3
:
205 case V3D_QPU_WADDR_R4
:
206 case V3D_QPU_WADDR_R5
:
208 &state
->last_r
[waddr
- V3D_QPU_WADDR_R0
],
212 case V3D_QPU_WADDR_VPM
:
213 case V3D_QPU_WADDR_VPMU
:
214 add_write_dep(state
, &state
->last_vpm
, n
);
217 case V3D_QPU_WADDR_TLB
:
218 case V3D_QPU_WADDR_TLBU
:
219 add_write_dep(state
, &state
->last_tlb
, n
);
222 case V3D_QPU_WADDR_NOP
:
226 fprintf(stderr
, "Unknown waddr %d\n", waddr
);
233 process_cond_deps(struct schedule_state
*state
, struct schedule_node
*n
,
234 enum v3d_qpu_cond cond
)
236 if (cond
!= V3D_QPU_COND_NONE
)
237 add_read_dep(state
, state
->last_sf
, n
);
241 process_pf_deps(struct schedule_state
*state
, struct schedule_node
*n
,
244 if (pf
!= V3D_QPU_PF_NONE
)
245 add_write_dep(state
, &state
->last_sf
, n
);
249 process_uf_deps(struct schedule_state
*state
, struct schedule_node
*n
,
252 if (uf
!= V3D_QPU_UF_NONE
)
253 add_write_dep(state
, &state
->last_sf
, n
);
257 * Common code for dependencies that need to be tracked both forward and
260 * This is for things like "all reads of r4 have to happen between the r4
261 * writes that surround them".
264 calculate_deps(struct schedule_state
*state
, struct schedule_node
*n
)
266 struct qinst
*qinst
= n
->inst
;
267 struct v3d_qpu_instr
*inst
= &qinst
->qpu
;
269 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
) {
270 if (inst
->branch
.cond
!= V3D_QPU_BRANCH_COND_ALWAYS
)
271 add_read_dep(state
, state
->last_sf
, n
);
278 add_write_dep(state
, &state
->last_unif
, n
);
282 assert(inst
->type
== V3D_QPU_INSTR_TYPE_ALU
);
286 if (v3d_qpu_add_op_num_src(inst
->alu
.add
.op
) > 0)
287 process_mux_deps(state
, n
, inst
->alu
.add
.a
);
288 if (v3d_qpu_add_op_num_src(inst
->alu
.add
.op
) > 1)
289 process_mux_deps(state
, n
, inst
->alu
.add
.b
);
291 if (v3d_qpu_mul_op_num_src(inst
->alu
.mul
.op
) > 0)
292 process_mux_deps(state
, n
, inst
->alu
.mul
.a
);
293 if (v3d_qpu_mul_op_num_src(inst
->alu
.mul
.op
) > 1)
294 process_mux_deps(state
, n
, inst
->alu
.mul
.b
);
296 switch (inst
->alu
.add
.op
) {
297 case V3D_QPU_A_VPMSETUP
:
298 /* Could distinguish read/write by unpacking the uniform. */
299 add_write_dep(state
, &state
->last_vpm
, n
);
300 add_write_dep(state
, &state
->last_vpm_read
, n
);
303 case V3D_QPU_A_STVPMV
:
304 case V3D_QPU_A_STVPMD
:
305 case V3D_QPU_A_STVPMP
:
306 add_write_dep(state
, &state
->last_vpm
, n
);
310 add_read_dep(state
, state
->last_tlb
, n
);
313 case V3D_QPU_A_SETMSF
:
314 case V3D_QPU_A_SETREVF
:
315 add_write_dep(state
, &state
->last_tlb
, n
);
318 case V3D_QPU_A_FLAPUSH
:
319 case V3D_QPU_A_FLBPUSH
:
321 case V3D_QPU_A_VFLNA
:
323 case V3D_QPU_A_VFLNB
:
324 add_read_dep(state
, state
->last_sf
, n
);
327 case V3D_QPU_A_FLBPOP
:
328 add_write_dep(state
, &state
->last_sf
, n
);
335 switch (inst
->alu
.mul
.op
) {
336 case V3D_QPU_M_MULTOP
:
337 case V3D_QPU_M_UMUL24
:
338 /* MULTOP sets rtop, and UMUL24 implicitly reads rtop and
339 * resets it to 0. We could possibly reorder umul24s relative
340 * to each other, but for now just keep all the MUL parts in
343 add_write_dep(state
, &state
->last_rtop
, n
);
349 if (inst
->alu
.add
.op
!= V3D_QPU_A_NOP
) {
350 process_waddr_deps(state
, n
, inst
->alu
.add
.waddr
,
351 inst
->alu
.add
.magic_write
);
353 if (inst
->alu
.mul
.op
!= V3D_QPU_M_NOP
) {
354 process_waddr_deps(state
, n
, inst
->alu
.mul
.waddr
,
355 inst
->alu
.mul
.magic_write
);
358 if (v3d_qpu_writes_r3(inst
))
359 add_write_dep(state
, &state
->last_r
[3], n
);
360 if (v3d_qpu_writes_r4(inst
))
361 add_write_dep(state
, &state
->last_r
[4], n
);
362 if (v3d_qpu_writes_r5(inst
))
363 add_write_dep(state
, &state
->last_r
[5], n
);
365 if (inst
->sig
.thrsw
) {
366 /* All accumulator contents and flags are undefined after the
369 for (int i
= 0; i
< ARRAY_SIZE(state
->last_r
); i
++)
370 add_write_dep(state
, &state
->last_r
[i
], n
);
371 add_write_dep(state
, &state
->last_sf
, n
);
373 /* Scoreboard-locking operations have to stay after the last
376 add_write_dep(state
, &state
->last_tlb
, n
);
378 add_write_dep(state
, &state
->last_tmu_write
, n
);
381 if (inst
->sig
.ldtmu
) {
382 /* TMU loads are coming from a FIFO, so ordering is important.
384 add_write_dep(state
, &state
->last_tmu_write
, n
);
387 if (inst
->sig
.ldtlb
| inst
->sig
.ldtlbu
)
388 add_read_dep(state
, state
->last_tlb
, n
);
391 add_write_dep(state
, &state
->last_vpm_read
, n
);
393 /* inst->sig.ldunif or sideband uniform read */
394 if (qinst
->uniform
!= ~0)
395 add_write_dep(state
, &state
->last_unif
, n
);
397 process_cond_deps(state
, n
, inst
->flags
.ac
);
398 process_cond_deps(state
, n
, inst
->flags
.mc
);
399 process_pf_deps(state
, n
, inst
->flags
.apf
);
400 process_pf_deps(state
, n
, inst
->flags
.mpf
);
401 process_uf_deps(state
, n
, inst
->flags
.auf
);
402 process_uf_deps(state
, n
, inst
->flags
.muf
);
406 calculate_forward_deps(struct v3d_compile
*c
, struct list_head
*schedule_list
)
408 struct schedule_state state
;
410 memset(&state
, 0, sizeof(state
));
413 list_for_each_entry(struct schedule_node
, node
, schedule_list
, link
)
414 calculate_deps(&state
, node
);
418 calculate_reverse_deps(struct v3d_compile
*c
, struct list_head
*schedule_list
)
420 struct list_head
*node
;
421 struct schedule_state state
;
423 memset(&state
, 0, sizeof(state
));
426 for (node
= schedule_list
->prev
; schedule_list
!= node
; node
= node
->prev
) {
427 calculate_deps(&state
, (struct schedule_node
*)node
);
431 struct choose_scoreboard
{
433 int last_sfu_write_tick
;
434 int last_ldvary_tick
;
435 int last_uniforms_reset_tick
;
436 uint32_t last_waddr_add
, last_waddr_mul
;
441 mux_reads_too_soon(struct choose_scoreboard
*scoreboard
,
442 const struct v3d_qpu_instr
*inst
, enum v3d_qpu_mux mux
)
446 if (scoreboard
->last_waddr_add
== inst
->raddr_a
||
447 scoreboard
->last_waddr_mul
== inst
->raddr_a
) {
453 if (scoreboard
->last_waddr_add
== inst
->raddr_b
||
454 scoreboard
->last_waddr_mul
== inst
->raddr_b
) {
460 if (scoreboard
->tick
- scoreboard
->last_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(struct choose_scoreboard
*scoreboard
,
518 const struct v3d_qpu_instr
*inst
= &qinst
->qpu
;
520 /* Don't schedule any other r4 write too soon after an SFU write.
521 * This would normally be prevented by dependency tracking, but might
522 * occur if a dead SFU computation makes it to scheduling.
524 if (scoreboard
->tick
- scoreboard
->last_sfu_write_tick
< 2 &&
525 v3d_qpu_writes_r4(inst
))
532 pixel_scoreboard_too_soon(struct choose_scoreboard
*scoreboard
,
533 const struct v3d_qpu_instr
*inst
)
535 return (scoreboard
->tick
== 0 && qpu_inst_is_tlb(inst
));
539 get_instruction_priority(const struct v3d_qpu_instr
*inst
)
541 uint32_t baseline_score
;
542 uint32_t next_score
= 0;
544 /* Schedule TLB operations as late as possible, to get more
545 * parallelism between shaders.
547 if (qpu_inst_is_tlb(inst
))
551 /* Schedule texture read results collection late to hide latency. */
556 /* Default score for things that aren't otherwise special. */
557 baseline_score
= next_score
;
560 /* Schedule texture read setup early to hide their latency better. */
561 if (inst
->type
== V3D_QPU_INSTR_TYPE_ALU
&&
562 ((inst
->alu
.add
.magic_write
&&
563 v3d_qpu_magic_waddr_is_tmu(inst
->alu
.add
.waddr
)) ||
564 (inst
->alu
.mul
.magic_write
&&
565 v3d_qpu_magic_waddr_is_tmu(inst
->alu
.mul
.waddr
)))) {
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 (inst
->type
== V3D_QPU_INSTR_TYPE_ALU
) {
587 if (inst
->alu
.add
.op
!= V3D_QPU_A_NOP
&&
588 inst
->alu
.add
.magic_write
&&
589 qpu_magic_waddr_is_periph(inst
->alu
.add
.waddr
)) {
593 if (inst
->alu
.add
.op
== V3D_QPU_A_VPMSETUP
)
596 if (inst
->alu
.mul
.op
!= V3D_QPU_M_NOP
&&
597 inst
->alu
.mul
.magic_write
&&
598 qpu_magic_waddr_is_periph(inst
->alu
.mul
.waddr
)) {
603 return (inst
->sig
.ldvpm
||
610 qpu_merge_inst(const struct v3d_device_info
*devinfo
,
611 struct v3d_qpu_instr
*result
,
612 const struct v3d_qpu_instr
*a
,
613 const struct v3d_qpu_instr
*b
)
615 if (a
->type
!= V3D_QPU_INSTR_TYPE_ALU
||
616 b
->type
!= V3D_QPU_INSTR_TYPE_ALU
) {
620 /* Can't do more than one peripheral access in an instruction. */
621 if (qpu_accesses_peripheral(a
) && qpu_accesses_peripheral(b
))
624 struct v3d_qpu_instr merge
= *a
;
626 if (b
->alu
.add
.op
!= V3D_QPU_A_NOP
) {
627 if (a
->alu
.add
.op
!= V3D_QPU_A_NOP
)
629 merge
.alu
.add
= b
->alu
.add
;
631 merge
.flags
.ac
= b
->flags
.ac
;
632 merge
.flags
.apf
= b
->flags
.apf
;
633 merge
.flags
.auf
= b
->flags
.auf
;
636 if (b
->alu
.mul
.op
!= V3D_QPU_M_NOP
) {
637 if (a
->alu
.mul
.op
!= V3D_QPU_M_NOP
)
639 merge
.alu
.mul
= b
->alu
.mul
;
641 merge
.flags
.mc
= b
->flags
.mc
;
642 merge
.flags
.mpf
= b
->flags
.mpf
;
643 merge
.flags
.muf
= b
->flags
.muf
;
646 if (v3d_qpu_uses_mux(b
, V3D_QPU_MUX_A
)) {
647 if (v3d_qpu_uses_mux(a
, V3D_QPU_MUX_A
) &&
648 a
->raddr_a
!= b
->raddr_a
) {
651 merge
.raddr_a
= b
->raddr_a
;
654 if (v3d_qpu_uses_mux(b
, V3D_QPU_MUX_B
)) {
655 if (v3d_qpu_uses_mux(a
, V3D_QPU_MUX_B
) &&
656 a
->raddr_b
!= b
->raddr_b
) {
659 merge
.raddr_b
= b
->raddr_b
;
662 merge
.sig
.thrsw
|= b
->sig
.thrsw
;
663 merge
.sig
.ldunif
|= b
->sig
.ldunif
;
664 merge
.sig
.ldtmu
|= b
->sig
.ldtmu
;
665 merge
.sig
.ldvary
|= b
->sig
.ldvary
;
666 merge
.sig
.ldvpm
|= b
->sig
.ldvpm
;
667 merge
.sig
.small_imm
|= b
->sig
.small_imm
;
668 merge
.sig
.ldtlb
|= b
->sig
.ldtlb
;
669 merge
.sig
.ldtlbu
|= b
->sig
.ldtlbu
;
670 merge
.sig
.ucb
|= b
->sig
.ucb
;
671 merge
.sig
.rotate
|= b
->sig
.rotate
;
672 merge
.sig
.wrtmuc
|= b
->sig
.wrtmuc
;
675 bool ok
= v3d_qpu_instr_pack(devinfo
, &merge
, &packed
);
678 /* No modifying the real instructions on failure. */
679 assert(ok
|| (a
!= result
&& b
!= result
));
684 static struct schedule_node
*
685 choose_instruction_to_schedule(const struct v3d_device_info
*devinfo
,
686 struct choose_scoreboard
*scoreboard
,
687 struct list_head
*schedule_list
,
688 struct schedule_node
*prev_inst
)
690 struct schedule_node
*chosen
= NULL
;
693 /* Don't pair up anything with a thread switch signal -- emit_thrsw()
694 * will handle pairing it along with filling the delay slots.
697 if (prev_inst
->inst
->qpu
.sig
.thrsw
)
701 list_for_each_entry(struct schedule_node
, n
, schedule_list
, link
) {
702 const struct v3d_qpu_instr
*inst
= &n
->inst
->qpu
;
704 /* Don't choose the branch instruction until it's the last one
705 * left. We'll move it up to fit its delay slots after we
708 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
&&
709 !list_is_singular(schedule_list
)) {
713 /* "An instruction must not read from a location in physical
714 * regfile A or B that was written to by the previous
717 if (reads_too_soon_after_write(scoreboard
, n
->inst
))
720 if (writes_too_soon_after_write(scoreboard
, n
->inst
))
723 /* "A scoreboard wait must not occur in the first two
724 * instructions of a fragment shader. This is either the
725 * explicit Wait for Scoreboard signal or an implicit wait
726 * with the first tile-buffer read or write instruction."
728 if (pixel_scoreboard_too_soon(scoreboard
, inst
))
731 /* ldunif and ldvary both write r5, but ldunif does so a tick
732 * sooner. If the ldvary's r5 wasn't used, then ldunif might
733 * otherwise get scheduled so ldunif and ldvary try to update
734 * r5 in the same tick.
736 if (inst
->sig
.ldunif
&&
737 scoreboard
->tick
== scoreboard
->last_ldvary_tick
+ 1) {
741 /* If we're trying to pair with another instruction, check
742 * that they're compatible.
745 /* Don't pair up a thread switch signal -- we'll
746 * handle pairing it when we pick it on its own.
751 if (prev_inst
->inst
->uniform
!= -1 &&
752 n
->inst
->uniform
!= -1)
755 /* Don't merge in something that will lock the TLB.
756 * Hopwefully what we have in inst will release some
757 * other instructions, allowing us to delay the
758 * TLB-locking instruction until later.
760 if (!scoreboard
->tlb_locked
&& qpu_inst_is_tlb(inst
))
763 struct v3d_qpu_instr merged_inst
;
764 if (!qpu_merge_inst(devinfo
, &merged_inst
,
765 &prev_inst
->inst
->qpu
, inst
)) {
770 int prio
= get_instruction_priority(inst
);
772 /* Found a valid instruction. If nothing better comes along,
781 if (prio
> chosen_prio
) {
784 } else if (prio
< chosen_prio
) {
788 if (n
->delay
> chosen
->delay
) {
791 } else if (n
->delay
< chosen
->delay
) {
800 update_scoreboard_for_magic_waddr(struct choose_scoreboard
*scoreboard
,
801 enum v3d_qpu_waddr waddr
)
803 if (v3d_qpu_magic_waddr_is_sfu(waddr
))
804 scoreboard
->last_sfu_write_tick
= scoreboard
->tick
;
808 update_scoreboard_for_chosen(struct choose_scoreboard
*scoreboard
,
809 const struct v3d_qpu_instr
*inst
)
811 scoreboard
->last_waddr_add
= ~0;
812 scoreboard
->last_waddr_mul
= ~0;
814 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
)
817 assert(inst
->type
== V3D_QPU_INSTR_TYPE_ALU
);
819 if (inst
->alu
.add
.op
!= V3D_QPU_A_NOP
) {
820 if (inst
->alu
.add
.magic_write
) {
821 update_scoreboard_for_magic_waddr(scoreboard
,
822 inst
->alu
.add
.waddr
);
824 scoreboard
->last_waddr_add
= inst
->alu
.add
.waddr
;
828 if (inst
->alu
.mul
.op
!= V3D_QPU_M_NOP
) {
829 if (inst
->alu
.mul
.magic_write
) {
830 update_scoreboard_for_magic_waddr(scoreboard
,
831 inst
->alu
.mul
.waddr
);
833 scoreboard
->last_waddr_mul
= inst
->alu
.mul
.waddr
;
837 if (inst
->sig
.ldvary
)
838 scoreboard
->last_ldvary_tick
= scoreboard
->tick
;
840 if (qpu_inst_is_tlb(inst
))
841 scoreboard
->tlb_locked
= true;
845 dump_state(const struct v3d_device_info
*devinfo
,
846 struct list_head
*schedule_list
)
848 list_for_each_entry(struct schedule_node
, n
, schedule_list
, link
) {
849 fprintf(stderr
, " t=%4d: ", n
->unblocked_time
);
850 v3d_qpu_dump(devinfo
, &n
->inst
->qpu
);
851 fprintf(stderr
, "\n");
853 for (int i
= 0; i
< n
->child_count
; i
++) {
854 struct schedule_node
*child
= n
->children
[i
].node
;
858 fprintf(stderr
, " - ");
859 v3d_qpu_dump(devinfo
, &child
->inst
->qpu
);
860 fprintf(stderr
, " (%d parents, %c)\n",
862 n
->children
[i
].write_after_read
? 'w' : 'r');
867 static uint32_t magic_waddr_latency(enum v3d_qpu_waddr waddr
,
868 const struct v3d_qpu_instr
*after
)
870 /* Apply some huge latency between texture fetch requests and getting
871 * their results back.
873 * FIXME: This is actually pretty bogus. If we do:
882 * we count that as worse than
891 * because we associate the first load_tmu0 with the *second* tmu0_s.
893 if (v3d_qpu_magic_waddr_is_tmu(waddr
) && after
->sig
.ldtmu
)
896 /* Assume that anything depending on us is consuming the SFU result. */
897 if (v3d_qpu_magic_waddr_is_sfu(waddr
))
904 instruction_latency(struct schedule_node
*before
, struct schedule_node
*after
)
906 const struct v3d_qpu_instr
*before_inst
= &before
->inst
->qpu
;
907 const struct v3d_qpu_instr
*after_inst
= &after
->inst
->qpu
;
908 uint32_t latency
= 1;
910 if (before_inst
->type
!= V3D_QPU_INSTR_TYPE_ALU
||
911 after_inst
->type
!= V3D_QPU_INSTR_TYPE_ALU
)
914 if (before_inst
->alu
.add
.magic_write
) {
915 latency
= MAX2(latency
,
916 magic_waddr_latency(before_inst
->alu
.add
.waddr
,
920 if (before_inst
->alu
.mul
.magic_write
) {
921 latency
= MAX2(latency
,
922 magic_waddr_latency(before_inst
->alu
.mul
.waddr
,
929 /** Recursive computation of the delay member of a node. */
931 compute_delay(struct schedule_node
*n
)
933 if (!n
->child_count
) {
936 for (int i
= 0; i
< n
->child_count
; i
++) {
937 if (!n
->children
[i
].node
->delay
)
938 compute_delay(n
->children
[i
].node
);
939 n
->delay
= MAX2(n
->delay
,
940 n
->children
[i
].node
->delay
+
941 instruction_latency(n
, n
->children
[i
].node
));
947 mark_instruction_scheduled(struct list_head
*schedule_list
,
949 struct schedule_node
*node
,
955 for (int i
= node
->child_count
- 1; i
>= 0; i
--) {
956 struct schedule_node
*child
=
957 node
->children
[i
].node
;
962 if (war_only
&& !node
->children
[i
].write_after_read
)
965 /* If the requirement is only that the node not appear before
966 * the last read of its destination, then it can be scheduled
967 * immediately after (or paired with!) the thing reading the
970 uint32_t latency
= 0;
972 latency
= instruction_latency(node
,
973 node
->children
[i
].node
);
976 child
->unblocked_time
= MAX2(child
->unblocked_time
,
978 child
->parent_count
--;
979 if (child
->parent_count
== 0)
980 list_add(&child
->link
, schedule_list
);
982 node
->children
[i
].node
= NULL
;
986 static struct qinst
*
989 struct qreg undef
= { QFILE_NULL
, 0 };
990 struct qinst
*qinst
= vir_add_inst(V3D_QPU_A_NOP
, undef
, undef
, undef
);
996 static struct qinst
*
997 nop_after(struct qinst
*inst
)
999 struct qinst
*q
= vir_nop();
1001 list_add(&q
->link
, &inst
->link
);
1007 * Emits a THRSW/LTHRSW signal in the stream, trying to move it up to pair
1008 * with another instruction.
1011 emit_thrsw(struct v3d_compile
*c
,
1012 struct choose_scoreboard
*scoreboard
,
1013 const struct v3d_qpu_instr
*inst
)
1015 /* There should be nothing in a thrsw inst being scheduled other than
1018 assert(inst
->type
== V3D_QPU_INSTR_TYPE_ALU
);
1019 assert(inst
->alu
.add
.op
== V3D_QPU_A_NOP
);
1020 assert(inst
->alu
.mul
.op
== V3D_QPU_M_NOP
);
1022 /* Try to find an earlier scheduled instruction that we can merge the
1025 int thrsw_ip
= c
->qpu_inst_count
;
1026 for (int i
= 1; i
<= MIN2(c
->qpu_inst_count
, 3); i
++) {
1027 uint64_t prev_instr
= c
->qpu_insts
[c
->qpu_inst_count
- i
];
1028 uint32_t prev_sig
= QPU_GET_FIELD(prev_instr
, QPU_SIG
);
1030 if (prev_sig
== QPU_SIG_NONE
)
1031 thrsw_ip
= c
->qpu_inst_count
- i
;
1034 if (thrsw_ip
!= c
->qpu_inst_count
) {
1035 /* Merge the thrsw into the existing instruction. */
1036 c
->qpu_insts
[thrsw_ip
] =
1037 QPU_UPDATE_FIELD(c
->qpu_insts
[thrsw_ip
], sig
, QPU_SIG
);
1039 qpu_serialize_one_inst(c
, inst
);
1040 update_scoreboard_for_chosen(scoreboard
, inst
);
1043 /* Fill the delay slots. */
1044 while (c
->qpu_inst_count
< thrsw_ip
+ 3) {
1045 update_scoreboard_for_chosen(scoreboard
, v3d_qpu_nop());
1046 qpu_serialize_one_inst(c
, v3d_qpu_nop());
1052 schedule_instructions(struct v3d_compile
*c
,
1053 struct choose_scoreboard
*scoreboard
,
1054 struct qblock
*block
,
1055 struct list_head
*schedule_list
,
1056 enum quniform_contents
*orig_uniform_contents
,
1057 uint32_t *orig_uniform_data
,
1058 uint32_t *next_uniform
)
1060 const struct v3d_device_info
*devinfo
= c
->devinfo
;
1064 fprintf(stderr
, "initial deps:\n");
1065 dump_state(devinfo
, schedule_list
);
1066 fprintf(stderr
, "\n");
1069 /* Remove non-DAG heads from the list. */
1070 list_for_each_entry_safe(struct schedule_node
, n
, schedule_list
, link
) {
1071 if (n
->parent_count
!= 0)
1075 while (!list_empty(schedule_list
)) {
1076 struct schedule_node
*chosen
=
1077 choose_instruction_to_schedule(devinfo
,
1081 struct schedule_node
*merge
= NULL
;
1083 /* If there are no valid instructions to schedule, drop a NOP
1086 struct qinst
*qinst
= chosen
? chosen
->inst
: vir_nop();
1087 struct v3d_qpu_instr
*inst
= &qinst
->qpu
;
1090 fprintf(stderr
, "t=%4d: current list:\n",
1092 dump_state(devinfo
, schedule_list
);
1093 fprintf(stderr
, "t=%4d: chose: ", time
);
1094 v3d_qpu_dump(devinfo
, inst
);
1095 fprintf(stderr
, "\n");
1098 /* Schedule this instruction onto the QPU list. Also try to
1099 * find an instruction to pair with it.
1102 time
= MAX2(chosen
->unblocked_time
, time
);
1103 list_del(&chosen
->link
);
1104 mark_instruction_scheduled(schedule_list
, time
,
1107 merge
= choose_instruction_to_schedule(devinfo
,
1112 time
= MAX2(merge
->unblocked_time
, time
);
1113 list_del(&merge
->link
);
1114 (void)qpu_merge_inst(devinfo
, inst
,
1115 inst
, &merge
->inst
->qpu
);
1116 if (merge
->inst
->uniform
!= -1) {
1117 chosen
->inst
->uniform
=
1118 merge
->inst
->uniform
;
1122 fprintf(stderr
, "t=%4d: merging: ",
1124 v3d_qpu_dump(devinfo
, &merge
->inst
->qpu
);
1125 fprintf(stderr
, "\n");
1126 fprintf(stderr
, " result: ");
1127 v3d_qpu_dump(devinfo
, inst
);
1128 fprintf(stderr
, "\n");
1133 /* Update the uniform index for the rewritten location --
1134 * branch target updating will still need to change
1135 * c->uniform_data[] using this index.
1137 if (qinst
->uniform
!= -1) {
1138 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
)
1139 block
->branch_uniform
= *next_uniform
;
1141 c
->uniform_data
[*next_uniform
] =
1142 orig_uniform_data
[qinst
->uniform
];
1143 c
->uniform_contents
[*next_uniform
] =
1144 orig_uniform_contents
[qinst
->uniform
];
1145 qinst
->uniform
= *next_uniform
;
1150 fprintf(stderr
, "\n");
1153 /* Now that we've scheduled a new instruction, some of its
1154 * children can be promoted to the list of instructions ready to
1155 * be scheduled. Update the children's unblocked time for this
1156 * DAG edge as we do so.
1158 mark_instruction_scheduled(schedule_list
, time
, chosen
, false);
1161 mark_instruction_scheduled(schedule_list
, time
, merge
,
1164 /* The merged VIR instruction doesn't get re-added to the
1165 * block, so free it now.
1170 if (0 && inst
->sig
.thrsw
) {
1171 /* XXX emit_thrsw(c, scoreboard, qinst); */
1173 c
->qpu_inst_count
++;
1174 list_addtail(&qinst
->link
, &block
->instructions
);
1175 update_scoreboard_for_chosen(scoreboard
, inst
);
1181 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
||
1182 inst
->sig
.thrsw
/* XXX */) {
1183 block
->branch_qpu_ip
= c
->qpu_inst_count
- 1;
1184 /* Fill the delay slots.
1186 * We should fill these with actual instructions,
1187 * instead, but that will probably need to be done
1188 * after this, once we know what the leading
1189 * instructions of the successors are (so we can
1190 * handle A/B register file write latency)
1192 /* XXX: scoreboard */
1193 int slots
= (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
?
1195 for (int i
= 0; i
< slots
; i
++) {
1196 struct qinst
*nop
= vir_nop();
1197 list_addtail(&nop
->link
, &block
->instructions
);
1199 update_scoreboard_for_chosen(scoreboard
,
1201 c
->qpu_inst_count
++;
1212 qpu_schedule_instructions_block(struct v3d_compile
*c
,
1213 struct choose_scoreboard
*scoreboard
,
1214 struct qblock
*block
,
1215 enum quniform_contents
*orig_uniform_contents
,
1216 uint32_t *orig_uniform_data
,
1217 uint32_t *next_uniform
)
1219 void *mem_ctx
= ralloc_context(NULL
);
1220 struct list_head schedule_list
;
1222 list_inithead(&schedule_list
);
1224 /* Wrap each instruction in a scheduler structure. */
1225 while (!list_empty(&block
->instructions
)) {
1226 struct qinst
*qinst
= (struct qinst
*)block
->instructions
.next
;
1227 struct schedule_node
*n
=
1228 rzalloc(mem_ctx
, struct schedule_node
);
1232 list_del(&qinst
->link
);
1233 list_addtail(&n
->link
, &schedule_list
);
1236 calculate_forward_deps(c
, &schedule_list
);
1237 calculate_reverse_deps(c
, &schedule_list
);
1239 list_for_each_entry(struct schedule_node
, n
, &schedule_list
, link
) {
1243 uint32_t cycles
= schedule_instructions(c
, scoreboard
, block
,
1245 orig_uniform_contents
,
1249 ralloc_free(mem_ctx
);
1255 qpu_set_branch_targets(struct v3d_compile
*c
)
1257 vir_for_each_block(block
, c
) {
1258 /* The end block of the program has no branch. */
1259 if (!block
->successors
[0])
1262 /* If there was no branch instruction, then the successor
1263 * block must follow immediately after this one.
1265 if (block
->branch_qpu_ip
== ~0) {
1266 assert(block
->end_qpu_ip
+ 1 ==
1267 block
->successors
[0]->start_qpu_ip
);
1271 /* Walk back through the delay slots to find the branch
1274 struct list_head
*entry
= block
->instructions
.prev
;
1275 for (int i
= 0; i
< 3; i
++)
1276 entry
= entry
->prev
;
1277 struct qinst
*branch
= container_of(entry
, branch
, link
);
1278 assert(branch
->qpu
.type
== V3D_QPU_INSTR_TYPE_BRANCH
);
1280 /* Make sure that the if-we-don't-jump
1281 * successor was scheduled just after the
1284 assert(!block
->successors
[1] ||
1285 block
->successors
[1]->start_qpu_ip
==
1286 block
->branch_qpu_ip
+ 4);
1288 branch
->qpu
.branch
.offset
=
1289 ((block
->successors
[0]->start_qpu_ip
-
1290 (block
->branch_qpu_ip
+ 4)) *
1293 /* Set up the relative offset to jump in the
1296 * Use a temporary here, because
1297 * uniform_data[inst->uniform] may be shared
1298 * between multiple instructions.
1300 assert(c
->uniform_contents
[branch
->uniform
] == QUNIFORM_CONSTANT
);
1301 c
->uniform_data
[branch
->uniform
] =
1302 (block
->successors
[0]->start_uniform
-
1303 (block
->branch_uniform
+ 1)) * 4;
1308 v3d_qpu_schedule_instructions(struct v3d_compile
*c
)
1310 const struct v3d_device_info
*devinfo
= c
->devinfo
;
1312 /* We reorder the uniforms as we schedule instructions, so save the
1313 * old data off and replace it.
1315 uint32_t *uniform_data
= c
->uniform_data
;
1316 enum quniform_contents
*uniform_contents
= c
->uniform_contents
;
1317 c
->uniform_contents
= ralloc_array(c
, enum quniform_contents
,
1319 c
->uniform_data
= ralloc_array(c
, uint32_t, c
->num_uniforms
);
1320 c
->uniform_array_size
= c
->num_uniforms
;
1321 uint32_t next_uniform
= 0;
1323 struct choose_scoreboard scoreboard
;
1324 memset(&scoreboard
, 0, sizeof(scoreboard
));
1325 scoreboard
.last_waddr_add
= ~0;
1326 scoreboard
.last_waddr_mul
= ~0;
1327 scoreboard
.last_ldvary_tick
= -10;
1328 scoreboard
.last_sfu_write_tick
= -10;
1329 scoreboard
.last_uniforms_reset_tick
= -10;
1332 fprintf(stderr
, "Pre-schedule instructions\n");
1333 vir_for_each_block(block
, c
) {
1334 fprintf(stderr
, "BLOCK %d\n", block
->index
);
1335 list_for_each_entry(struct qinst
, qinst
,
1336 &block
->instructions
, link
) {
1337 v3d_qpu_dump(devinfo
, &qinst
->qpu
);
1338 fprintf(stderr
, "\n");
1341 fprintf(stderr
, "\n");
1344 uint32_t cycles
= 0;
1345 vir_for_each_block(block
, c
) {
1346 block
->start_qpu_ip
= c
->qpu_inst_count
;
1347 block
->branch_qpu_ip
= ~0;
1348 block
->start_uniform
= next_uniform
;
1350 cycles
+= qpu_schedule_instructions_block(c
,
1357 block
->end_qpu_ip
= c
->qpu_inst_count
- 1;
1360 qpu_set_branch_targets(c
);
1362 assert(next_uniform
== c
->num_uniforms
);