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
;
283 /* If the input and output segments are shared, then all VPM reads to
284 * a location need to happen before all writes. We handle this by
285 * serializing all VPM operations for now.
287 bool separate_vpm_segment
= false;
289 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
) {
290 if (inst
->branch
.cond
!= V3D_QPU_BRANCH_COND_ALWAYS
)
291 add_read_dep(state
, state
->last_sf
, n
);
298 add_write_dep(state
, &state
->last_unif
, n
);
302 assert(inst
->type
== V3D_QPU_INSTR_TYPE_ALU
);
306 if (v3d_qpu_add_op_num_src(inst
->alu
.add
.op
) > 0)
307 process_mux_deps(state
, n
, inst
->alu
.add
.a
);
308 if (v3d_qpu_add_op_num_src(inst
->alu
.add
.op
) > 1)
309 process_mux_deps(state
, n
, inst
->alu
.add
.b
);
311 if (v3d_qpu_mul_op_num_src(inst
->alu
.mul
.op
) > 0)
312 process_mux_deps(state
, n
, inst
->alu
.mul
.a
);
313 if (v3d_qpu_mul_op_num_src(inst
->alu
.mul
.op
) > 1)
314 process_mux_deps(state
, n
, inst
->alu
.mul
.b
);
316 switch (inst
->alu
.add
.op
) {
317 case V3D_QPU_A_VPMSETUP
:
318 /* Could distinguish read/write by unpacking the uniform. */
319 add_write_dep(state
, &state
->last_vpm
, n
);
320 add_write_dep(state
, &state
->last_vpm_read
, n
);
323 case V3D_QPU_A_STVPMV
:
324 case V3D_QPU_A_STVPMD
:
325 case V3D_QPU_A_STVPMP
:
326 add_write_dep(state
, &state
->last_vpm
, n
);
329 case V3D_QPU_A_LDVPMV_IN
:
330 case V3D_QPU_A_LDVPMD_IN
:
331 case V3D_QPU_A_LDVPMG_IN
:
332 case V3D_QPU_A_LDVPMP
:
333 if (!separate_vpm_segment
)
334 add_write_dep(state
, &state
->last_vpm
, n
);
337 case V3D_QPU_A_VPMWT
:
338 add_read_dep(state
, state
->last_vpm
, n
);
342 add_read_dep(state
, state
->last_tlb
, n
);
345 case V3D_QPU_A_SETMSF
:
346 case V3D_QPU_A_SETREVF
:
347 add_write_dep(state
, &state
->last_tlb
, n
);
350 case V3D_QPU_A_FLAPUSH
:
351 case V3D_QPU_A_FLBPUSH
:
353 case V3D_QPU_A_VFLNA
:
355 case V3D_QPU_A_VFLNB
:
356 add_read_dep(state
, state
->last_sf
, n
);
359 case V3D_QPU_A_FLPOP
:
360 add_write_dep(state
, &state
->last_sf
, n
);
367 switch (inst
->alu
.mul
.op
) {
368 case V3D_QPU_M_MULTOP
:
369 case V3D_QPU_M_UMUL24
:
370 /* MULTOP sets rtop, and UMUL24 implicitly reads rtop and
371 * resets it to 0. We could possibly reorder umul24s relative
372 * to each other, but for now just keep all the MUL parts in
375 add_write_dep(state
, &state
->last_rtop
, n
);
381 if (inst
->alu
.add
.op
!= V3D_QPU_A_NOP
) {
382 process_waddr_deps(state
, n
, inst
->alu
.add
.waddr
,
383 inst
->alu
.add
.magic_write
);
385 if (inst
->alu
.mul
.op
!= V3D_QPU_M_NOP
) {
386 process_waddr_deps(state
, n
, inst
->alu
.mul
.waddr
,
387 inst
->alu
.mul
.magic_write
);
389 if (v3d_qpu_sig_writes_address(devinfo
, &inst
->sig
)) {
390 process_waddr_deps(state
, n
, inst
->sig_addr
,
394 if (v3d_qpu_writes_r3(devinfo
, inst
))
395 add_write_dep(state
, &state
->last_r
[3], n
);
396 if (v3d_qpu_writes_r4(devinfo
, inst
))
397 add_write_dep(state
, &state
->last_r
[4], n
);
398 if (v3d_qpu_writes_r5(devinfo
, inst
))
399 add_write_dep(state
, &state
->last_r
[5], n
);
401 if (inst
->sig
.thrsw
) {
402 /* All accumulator contents and flags are undefined after the
405 for (int i
= 0; i
< ARRAY_SIZE(state
->last_r
); i
++)
406 add_write_dep(state
, &state
->last_r
[i
], n
);
407 add_write_dep(state
, &state
->last_sf
, n
);
408 add_write_dep(state
, &state
->last_rtop
, n
);
410 /* Scoreboard-locking operations have to stay after the last
413 add_write_dep(state
, &state
->last_tlb
, n
);
415 add_write_dep(state
, &state
->last_tmu_write
, n
);
416 add_write_dep(state
, &state
->last_tmu_config
, n
);
419 if (v3d_qpu_waits_on_tmu(inst
)) {
420 /* TMU loads are coming from a FIFO, so ordering is important.
422 add_write_dep(state
, &state
->last_tmu_write
, n
);
425 if (inst
->sig
.wrtmuc
)
426 add_write_dep(state
, &state
->last_tmu_config
, n
);
428 if (inst
->sig
.ldtlb
| inst
->sig
.ldtlbu
)
429 add_read_dep(state
, state
->last_tlb
, n
);
431 if (inst
->sig
.ldvpm
) {
432 add_write_dep(state
, &state
->last_vpm_read
, n
);
434 /* At least for now, we're doing shared I/O segments, so queue
435 * all writes after all reads.
437 if (!separate_vpm_segment
)
438 add_write_dep(state
, &state
->last_vpm
, n
);
441 /* inst->sig.ldunif or sideband uniform read */
442 if (qinst
->uniform
!= ~0)
443 add_write_dep(state
, &state
->last_unif
, n
);
445 process_cond_deps(state
, n
, inst
->flags
.ac
);
446 process_cond_deps(state
, n
, inst
->flags
.mc
);
447 process_pf_deps(state
, n
, inst
->flags
.apf
);
448 process_pf_deps(state
, n
, inst
->flags
.mpf
);
449 process_uf_deps(state
, n
, inst
->flags
.auf
);
450 process_uf_deps(state
, n
, inst
->flags
.muf
);
454 calculate_forward_deps(struct v3d_compile
*c
, struct list_head
*schedule_list
)
456 struct schedule_state state
;
458 memset(&state
, 0, sizeof(state
));
459 state
.devinfo
= c
->devinfo
;
462 list_for_each_entry(struct schedule_node
, node
, schedule_list
, link
)
463 calculate_deps(&state
, node
);
467 calculate_reverse_deps(struct v3d_compile
*c
, struct list_head
*schedule_list
)
469 struct list_head
*node
;
470 struct schedule_state state
;
472 memset(&state
, 0, sizeof(state
));
473 state
.devinfo
= c
->devinfo
;
476 for (node
= schedule_list
->prev
; schedule_list
!= node
; node
= node
->prev
) {
477 calculate_deps(&state
, (struct schedule_node
*)node
);
481 struct choose_scoreboard
{
483 int last_magic_sfu_write_tick
;
484 int last_ldvary_tick
;
485 int last_uniforms_reset_tick
;
491 mux_reads_too_soon(struct choose_scoreboard
*scoreboard
,
492 const struct v3d_qpu_instr
*inst
, enum v3d_qpu_mux mux
)
496 if (scoreboard
->tick
- scoreboard
->last_magic_sfu_write_tick
<= 2)
501 if (scoreboard
->tick
- scoreboard
->last_ldvary_tick
<= 1)
512 reads_too_soon_after_write(struct choose_scoreboard
*scoreboard
,
515 const struct v3d_qpu_instr
*inst
= &qinst
->qpu
;
517 /* XXX: Branching off of raddr. */
518 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
)
521 assert(inst
->type
== V3D_QPU_INSTR_TYPE_ALU
);
523 if (inst
->alu
.add
.op
!= V3D_QPU_A_NOP
) {
524 if (v3d_qpu_add_op_num_src(inst
->alu
.add
.op
) > 0 &&
525 mux_reads_too_soon(scoreboard
, inst
, inst
->alu
.add
.a
)) {
528 if (v3d_qpu_add_op_num_src(inst
->alu
.add
.op
) > 1 &&
529 mux_reads_too_soon(scoreboard
, inst
, inst
->alu
.add
.b
)) {
534 if (inst
->alu
.mul
.op
!= V3D_QPU_M_NOP
) {
535 if (v3d_qpu_mul_op_num_src(inst
->alu
.mul
.op
) > 0 &&
536 mux_reads_too_soon(scoreboard
, inst
, inst
->alu
.mul
.a
)) {
539 if (v3d_qpu_mul_op_num_src(inst
->alu
.mul
.op
) > 1 &&
540 mux_reads_too_soon(scoreboard
, inst
, inst
->alu
.mul
.b
)) {
551 writes_too_soon_after_write(const struct v3d_device_info
*devinfo
,
552 struct choose_scoreboard
*scoreboard
,
555 const struct v3d_qpu_instr
*inst
= &qinst
->qpu
;
557 /* Don't schedule any other r4 write too soon after an SFU write.
558 * This would normally be prevented by dependency tracking, but might
559 * occur if a dead SFU computation makes it to scheduling.
561 if (scoreboard
->tick
- scoreboard
->last_magic_sfu_write_tick
< 2 &&
562 v3d_qpu_writes_r4(devinfo
, inst
))
569 pixel_scoreboard_too_soon(struct choose_scoreboard
*scoreboard
,
570 const struct v3d_qpu_instr
*inst
)
572 return (scoreboard
->tick
== 0 && qpu_inst_is_tlb(inst
));
576 get_instruction_priority(const struct v3d_qpu_instr
*inst
)
578 uint32_t baseline_score
;
579 uint32_t next_score
= 0;
581 /* Schedule TLB operations as late as possible, to get more
582 * parallelism between shaders.
584 if (qpu_inst_is_tlb(inst
))
588 /* Schedule texture read results collection late to hide latency. */
589 if (v3d_qpu_waits_on_tmu(inst
))
593 /* Default score for things that aren't otherwise special. */
594 baseline_score
= next_score
;
597 /* Schedule texture read setup early to hide their latency better. */
598 if (v3d_qpu_writes_tmu(inst
))
602 return baseline_score
;
606 qpu_magic_waddr_is_periph(enum v3d_qpu_waddr waddr
)
608 return (v3d_qpu_magic_waddr_is_tmu(waddr
) ||
609 v3d_qpu_magic_waddr_is_sfu(waddr
) ||
610 v3d_qpu_magic_waddr_is_tlb(waddr
) ||
611 v3d_qpu_magic_waddr_is_vpm(waddr
) ||
612 v3d_qpu_magic_waddr_is_tsy(waddr
));
616 qpu_accesses_peripheral(const struct v3d_qpu_instr
*inst
)
618 if (v3d_qpu_uses_vpm(inst
))
620 if (v3d_qpu_uses_sfu(inst
))
623 if (inst
->type
== V3D_QPU_INSTR_TYPE_ALU
) {
624 if (inst
->alu
.add
.op
!= V3D_QPU_A_NOP
&&
625 inst
->alu
.add
.magic_write
&&
626 qpu_magic_waddr_is_periph(inst
->alu
.add
.waddr
)) {
630 if (inst
->alu
.add
.op
== V3D_QPU_A_TMUWT
)
633 if (inst
->alu
.mul
.op
!= V3D_QPU_M_NOP
&&
634 inst
->alu
.mul
.magic_write
&&
635 qpu_magic_waddr_is_periph(inst
->alu
.mul
.waddr
)) {
640 return (inst
->sig
.ldvpm
||
648 qpu_merge_inst(const struct v3d_device_info
*devinfo
,
649 struct v3d_qpu_instr
*result
,
650 const struct v3d_qpu_instr
*a
,
651 const struct v3d_qpu_instr
*b
)
653 if (a
->type
!= V3D_QPU_INSTR_TYPE_ALU
||
654 b
->type
!= V3D_QPU_INSTR_TYPE_ALU
) {
658 /* Can't do more than one peripheral access in an instruction.
660 * XXX: V3D 4.1 allows TMU read along with a VPM read or write, and
661 * WRTMUC with a TMU magic register write (other than tmuc).
663 if (qpu_accesses_peripheral(a
) && qpu_accesses_peripheral(b
))
666 struct v3d_qpu_instr merge
= *a
;
668 if (b
->alu
.add
.op
!= V3D_QPU_A_NOP
) {
669 if (a
->alu
.add
.op
!= V3D_QPU_A_NOP
)
671 merge
.alu
.add
= b
->alu
.add
;
673 merge
.flags
.ac
= b
->flags
.ac
;
674 merge
.flags
.apf
= b
->flags
.apf
;
675 merge
.flags
.auf
= b
->flags
.auf
;
678 if (b
->alu
.mul
.op
!= V3D_QPU_M_NOP
) {
679 if (a
->alu
.mul
.op
!= V3D_QPU_M_NOP
)
681 merge
.alu
.mul
= b
->alu
.mul
;
683 merge
.flags
.mc
= b
->flags
.mc
;
684 merge
.flags
.mpf
= b
->flags
.mpf
;
685 merge
.flags
.muf
= b
->flags
.muf
;
688 if (v3d_qpu_uses_mux(b
, V3D_QPU_MUX_A
)) {
689 if (v3d_qpu_uses_mux(a
, V3D_QPU_MUX_A
) &&
690 a
->raddr_a
!= b
->raddr_a
) {
693 merge
.raddr_a
= b
->raddr_a
;
696 if (v3d_qpu_uses_mux(b
, V3D_QPU_MUX_B
)) {
697 if (v3d_qpu_uses_mux(a
, V3D_QPU_MUX_B
) &&
698 (a
->raddr_b
!= b
->raddr_b
||
699 a
->sig
.small_imm
!= b
->sig
.small_imm
)) {
702 merge
.raddr_b
= b
->raddr_b
;
705 merge
.sig
.thrsw
|= b
->sig
.thrsw
;
706 merge
.sig
.ldunif
|= b
->sig
.ldunif
;
707 merge
.sig
.ldunifrf
|= b
->sig
.ldunifrf
;
708 merge
.sig
.ldunifa
|= b
->sig
.ldunifa
;
709 merge
.sig
.ldunifarf
|= b
->sig
.ldunifarf
;
710 merge
.sig
.ldtmu
|= b
->sig
.ldtmu
;
711 merge
.sig
.ldvary
|= b
->sig
.ldvary
;
712 merge
.sig
.ldvpm
|= b
->sig
.ldvpm
;
713 merge
.sig
.small_imm
|= b
->sig
.small_imm
;
714 merge
.sig
.ldtlb
|= b
->sig
.ldtlb
;
715 merge
.sig
.ldtlbu
|= b
->sig
.ldtlbu
;
716 merge
.sig
.ucb
|= b
->sig
.ucb
;
717 merge
.sig
.rotate
|= b
->sig
.rotate
;
718 merge
.sig
.wrtmuc
|= b
->sig
.wrtmuc
;
720 if (v3d_qpu_sig_writes_address(devinfo
, &a
->sig
) &&
721 v3d_qpu_sig_writes_address(devinfo
, &b
->sig
))
723 merge
.sig_addr
|= b
->sig_addr
;
724 merge
.sig_magic
|= b
->sig_magic
;
727 bool ok
= v3d_qpu_instr_pack(devinfo
, &merge
, &packed
);
730 /* No modifying the real instructions on failure. */
731 assert(ok
|| (a
!= result
&& b
!= result
));
736 static struct schedule_node
*
737 choose_instruction_to_schedule(const struct v3d_device_info
*devinfo
,
738 struct choose_scoreboard
*scoreboard
,
739 struct list_head
*schedule_list
,
740 struct schedule_node
*prev_inst
)
742 struct schedule_node
*chosen
= NULL
;
745 /* Don't pair up anything with a thread switch signal -- emit_thrsw()
746 * will handle pairing it along with filling the delay slots.
749 if (prev_inst
->inst
->qpu
.sig
.thrsw
)
753 list_for_each_entry(struct schedule_node
, n
, schedule_list
, link
) {
754 const struct v3d_qpu_instr
*inst
= &n
->inst
->qpu
;
756 /* Don't choose the branch instruction until it's the last one
757 * left. We'll move it up to fit its delay slots after we
760 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
&&
761 !list_is_singular(schedule_list
)) {
765 /* "An instruction must not read from a location in physical
766 * regfile A or B that was written to by the previous
769 if (reads_too_soon_after_write(scoreboard
, n
->inst
))
772 if (writes_too_soon_after_write(devinfo
, scoreboard
, n
->inst
))
775 /* "A scoreboard wait must not occur in the first two
776 * instructions of a fragment shader. This is either the
777 * explicit Wait for Scoreboard signal or an implicit wait
778 * with the first tile-buffer read or write instruction."
780 if (pixel_scoreboard_too_soon(scoreboard
, inst
))
783 /* ldunif and ldvary both write r5, but ldunif does so a tick
784 * sooner. If the ldvary's r5 wasn't used, then ldunif might
785 * otherwise get scheduled so ldunif and ldvary try to update
786 * r5 in the same tick.
788 if ((inst
->sig
.ldunif
|| inst
->sig
.ldunifa
) &&
789 scoreboard
->tick
== scoreboard
->last_ldvary_tick
+ 1) {
793 /* If we're trying to pair with another instruction, check
794 * that they're compatible.
797 /* Don't pair up a thread switch signal -- we'll
798 * handle pairing it when we pick it on its own.
803 if (prev_inst
->inst
->uniform
!= -1 &&
804 n
->inst
->uniform
!= -1)
807 /* Don't merge in something that will lock the TLB.
808 * Hopwefully what we have in inst will release some
809 * other instructions, allowing us to delay the
810 * TLB-locking instruction until later.
812 if (!scoreboard
->tlb_locked
&& qpu_inst_is_tlb(inst
))
815 struct v3d_qpu_instr merged_inst
;
816 if (!qpu_merge_inst(devinfo
, &merged_inst
,
817 &prev_inst
->inst
->qpu
, inst
)) {
822 int prio
= get_instruction_priority(inst
);
824 /* Found a valid instruction. If nothing better comes along,
833 if (prio
> chosen_prio
) {
836 } else if (prio
< chosen_prio
) {
840 if (n
->delay
> chosen
->delay
) {
843 } else if (n
->delay
< chosen
->delay
) {
852 update_scoreboard_for_magic_waddr(struct choose_scoreboard
*scoreboard
,
853 enum v3d_qpu_waddr waddr
)
855 if (v3d_qpu_magic_waddr_is_sfu(waddr
))
856 scoreboard
->last_magic_sfu_write_tick
= scoreboard
->tick
;
860 update_scoreboard_for_chosen(struct choose_scoreboard
*scoreboard
,
861 const struct v3d_qpu_instr
*inst
)
863 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
)
866 assert(inst
->type
== V3D_QPU_INSTR_TYPE_ALU
);
868 if (inst
->alu
.add
.op
!= V3D_QPU_A_NOP
) {
869 if (inst
->alu
.add
.magic_write
) {
870 update_scoreboard_for_magic_waddr(scoreboard
,
871 inst
->alu
.add
.waddr
);
875 if (inst
->alu
.mul
.op
!= V3D_QPU_M_NOP
) {
876 if (inst
->alu
.mul
.magic_write
) {
877 update_scoreboard_for_magic_waddr(scoreboard
,
878 inst
->alu
.mul
.waddr
);
882 if (inst
->sig
.ldvary
)
883 scoreboard
->last_ldvary_tick
= scoreboard
->tick
;
885 if (qpu_inst_is_tlb(inst
))
886 scoreboard
->tlb_locked
= true;
890 dump_state(const struct v3d_device_info
*devinfo
,
891 struct list_head
*schedule_list
)
893 list_for_each_entry(struct schedule_node
, n
, schedule_list
, link
) {
894 fprintf(stderr
, " t=%4d: ", n
->unblocked_time
);
895 v3d_qpu_dump(devinfo
, &n
->inst
->qpu
);
896 fprintf(stderr
, "\n");
898 for (int i
= 0; i
< n
->child_count
; i
++) {
899 struct schedule_node
*child
= n
->children
[i
].node
;
903 fprintf(stderr
, " - ");
904 v3d_qpu_dump(devinfo
, &child
->inst
->qpu
);
905 fprintf(stderr
, " (%d parents, %c)\n",
907 n
->children
[i
].write_after_read
? 'w' : 'r');
912 static uint32_t magic_waddr_latency(enum v3d_qpu_waddr waddr
,
913 const struct v3d_qpu_instr
*after
)
915 /* Apply some huge latency between texture fetch requests and getting
916 * their results back.
918 * FIXME: This is actually pretty bogus. If we do:
927 * we count that as worse than
936 * because we associate the first load_tmu0 with the *second* tmu0_s.
938 if (v3d_qpu_magic_waddr_is_tmu(waddr
) && v3d_qpu_waits_on_tmu(after
))
941 /* Assume that anything depending on us is consuming the SFU result. */
942 if (v3d_qpu_magic_waddr_is_sfu(waddr
))
949 instruction_latency(struct schedule_node
*before
, struct schedule_node
*after
)
951 const struct v3d_qpu_instr
*before_inst
= &before
->inst
->qpu
;
952 const struct v3d_qpu_instr
*after_inst
= &after
->inst
->qpu
;
953 uint32_t latency
= 1;
955 if (before_inst
->type
!= V3D_QPU_INSTR_TYPE_ALU
||
956 after_inst
->type
!= V3D_QPU_INSTR_TYPE_ALU
)
959 if (before_inst
->alu
.add
.magic_write
) {
960 latency
= MAX2(latency
,
961 magic_waddr_latency(before_inst
->alu
.add
.waddr
,
965 if (before_inst
->alu
.mul
.magic_write
) {
966 latency
= MAX2(latency
,
967 magic_waddr_latency(before_inst
->alu
.mul
.waddr
,
974 /** Recursive computation of the delay member of a node. */
976 compute_delay(struct schedule_node
*n
)
978 if (!n
->child_count
) {
981 for (int i
= 0; i
< n
->child_count
; i
++) {
982 if (!n
->children
[i
].node
->delay
)
983 compute_delay(n
->children
[i
].node
);
984 n
->delay
= MAX2(n
->delay
,
985 n
->children
[i
].node
->delay
+
986 instruction_latency(n
, n
->children
[i
].node
));
992 mark_instruction_scheduled(struct list_head
*schedule_list
,
994 struct schedule_node
*node
,
1000 for (int i
= node
->child_count
- 1; i
>= 0; i
--) {
1001 struct schedule_node
*child
=
1002 node
->children
[i
].node
;
1007 if (war_only
&& !node
->children
[i
].write_after_read
)
1010 /* If the requirement is only that the node not appear before
1011 * the last read of its destination, then it can be scheduled
1012 * immediately after (or paired with!) the thing reading the
1015 uint32_t latency
= 0;
1017 latency
= instruction_latency(node
,
1018 node
->children
[i
].node
);
1021 child
->unblocked_time
= MAX2(child
->unblocked_time
,
1023 child
->parent_count
--;
1024 if (child
->parent_count
== 0)
1025 list_add(&child
->link
, schedule_list
);
1027 node
->children
[i
].node
= NULL
;
1032 insert_scheduled_instruction(struct v3d_compile
*c
,
1033 struct qblock
*block
,
1034 struct choose_scoreboard
*scoreboard
,
1037 list_addtail(&inst
->link
, &block
->instructions
);
1039 update_scoreboard_for_chosen(scoreboard
, &inst
->qpu
);
1040 c
->qpu_inst_count
++;
1044 static struct qinst
*
1047 struct qreg undef
= { QFILE_NULL
, 0 };
1048 struct qinst
*qinst
= vir_add_inst(V3D_QPU_A_NOP
, undef
, undef
, undef
);
1054 emit_nop(struct v3d_compile
*c
, struct qblock
*block
,
1055 struct choose_scoreboard
*scoreboard
)
1057 insert_scheduled_instruction(c
, block
, scoreboard
, vir_nop());
1061 qpu_instruction_valid_in_thrend_slot(struct v3d_compile
*c
,
1062 const struct qinst
*qinst
, int slot
)
1064 const struct v3d_qpu_instr
*inst
= &qinst
->qpu
;
1066 /* Only TLB Z writes are prohibited in the last slot, but we don't
1067 * have those flagged so prohibit all TLB ops for now.
1069 if (slot
== 2 && qpu_inst_is_tlb(inst
))
1072 if (slot
> 0 && qinst
->uniform
!= ~0)
1075 if (v3d_qpu_uses_vpm(inst
))
1078 if (inst
->sig
.ldvary
)
1081 if (inst
->type
== V3D_QPU_INSTR_TYPE_ALU
) {
1082 /* GFXH-1625: TMUWT not allowed in the final instruction. */
1083 if (slot
== 2 && inst
->alu
.add
.op
== V3D_QPU_A_TMUWT
)
1086 /* No writing physical registers at the end. */
1087 if (!inst
->alu
.add
.magic_write
||
1088 !inst
->alu
.mul
.magic_write
) {
1092 if (c
->devinfo
->ver
< 40 && inst
->alu
.add
.op
== V3D_QPU_A_SETMSF
)
1095 /* RF0-2 might be overwritten during the delay slots by
1096 * fragment shader setup.
1098 if (inst
->raddr_a
< 3 &&
1099 (inst
->alu
.add
.a
== V3D_QPU_MUX_A
||
1100 inst
->alu
.add
.b
== V3D_QPU_MUX_A
||
1101 inst
->alu
.mul
.a
== V3D_QPU_MUX_A
||
1102 inst
->alu
.mul
.b
== V3D_QPU_MUX_A
)) {
1106 if (inst
->raddr_b
< 3 &&
1107 !inst
->sig
.small_imm
&&
1108 (inst
->alu
.add
.a
== V3D_QPU_MUX_B
||
1109 inst
->alu
.add
.b
== V3D_QPU_MUX_B
||
1110 inst
->alu
.mul
.a
== V3D_QPU_MUX_B
||
1111 inst
->alu
.mul
.b
== V3D_QPU_MUX_B
)) {
1120 valid_thrsw_sequence(struct v3d_compile
*c
, struct choose_scoreboard
*scoreboard
,
1121 struct qinst
*qinst
, int instructions_in_sequence
,
1124 /* No emitting our thrsw while the previous thrsw hasn't happened yet. */
1125 if (scoreboard
->last_thrsw_tick
+ 3 >
1126 scoreboard
->tick
- instructions_in_sequence
) {
1130 for (int slot
= 0; slot
< instructions_in_sequence
; slot
++) {
1131 /* No scheduling SFU when the result would land in the other
1132 * thread. The simulator complains for safety, though it
1133 * would only occur for dead code in our case.
1136 qinst
->qpu
.type
== V3D_QPU_INSTR_TYPE_ALU
&&
1137 (v3d_qpu_magic_waddr_is_sfu(qinst
->qpu
.alu
.add
.waddr
) ||
1138 v3d_qpu_magic_waddr_is_sfu(qinst
->qpu
.alu
.mul
.waddr
))) {
1142 if (slot
> 0 && qinst
->qpu
.sig
.ldvary
)
1146 !qpu_instruction_valid_in_thrend_slot(c
, qinst
, slot
)) {
1150 /* Note that the list is circular, so we can only do this up
1151 * to instructions_in_sequence.
1153 qinst
= (struct qinst
*)qinst
->link
.next
;
1160 * Emits a THRSW signal in the stream, trying to move it up to pair with
1161 * another instruction.
1164 emit_thrsw(struct v3d_compile
*c
,
1165 struct qblock
*block
,
1166 struct choose_scoreboard
*scoreboard
,
1172 /* There should be nothing in a thrsw inst being scheduled other than
1175 assert(inst
->qpu
.type
== V3D_QPU_INSTR_TYPE_ALU
);
1176 assert(inst
->qpu
.alu
.add
.op
== V3D_QPU_A_NOP
);
1177 assert(inst
->qpu
.alu
.mul
.op
== V3D_QPU_M_NOP
);
1179 /* Find how far back into previous instructions we can put the THRSW. */
1180 int slots_filled
= 0;
1181 struct qinst
*merge_inst
= NULL
;
1182 vir_for_each_inst_rev(prev_inst
, block
) {
1183 struct v3d_qpu_sig sig
= prev_inst
->qpu
.sig
;
1185 uint32_t packed_sig
;
1187 if (!v3d_qpu_sig_pack(c
->devinfo
, &sig
, &packed_sig
))
1190 if (!valid_thrsw_sequence(c
, scoreboard
,
1191 prev_inst
, slots_filled
+ 1,
1196 merge_inst
= prev_inst
;
1197 if (++slots_filled
== 3)
1201 bool needs_free
= false;
1203 merge_inst
->qpu
.sig
.thrsw
= true;
1205 scoreboard
->last_thrsw_tick
= scoreboard
->tick
- slots_filled
;
1207 scoreboard
->last_thrsw_tick
= scoreboard
->tick
;
1208 insert_scheduled_instruction(c
, block
, scoreboard
, inst
);
1214 /* Insert any extra delay slot NOPs we need. */
1215 for (int i
= 0; i
< 3 - slots_filled
; i
++) {
1216 emit_nop(c
, block
, scoreboard
);
1220 /* If we're emitting the last THRSW (other than program end), then
1221 * signal that to the HW by emitting two THRSWs in a row.
1223 if (inst
->is_last_thrsw
) {
1224 struct qinst
*second_inst
=
1225 (struct qinst
*)merge_inst
->link
.next
;
1226 second_inst
->qpu
.sig
.thrsw
= true;
1229 /* If we put our THRSW into another instruction, free up the
1230 * instruction that didn't end up scheduled into the list.
1239 schedule_instructions(struct v3d_compile
*c
,
1240 struct choose_scoreboard
*scoreboard
,
1241 struct qblock
*block
,
1242 struct list_head
*schedule_list
,
1243 enum quniform_contents
*orig_uniform_contents
,
1244 uint32_t *orig_uniform_data
,
1245 uint32_t *next_uniform
)
1247 const struct v3d_device_info
*devinfo
= c
->devinfo
;
1251 fprintf(stderr
, "initial deps:\n");
1252 dump_state(devinfo
, schedule_list
);
1253 fprintf(stderr
, "\n");
1256 /* Remove non-DAG heads from the list. */
1257 list_for_each_entry_safe(struct schedule_node
, n
, schedule_list
, link
) {
1258 if (n
->parent_count
!= 0)
1262 while (!list_empty(schedule_list
)) {
1263 struct schedule_node
*chosen
=
1264 choose_instruction_to_schedule(devinfo
,
1268 struct schedule_node
*merge
= NULL
;
1270 /* If there are no valid instructions to schedule, drop a NOP
1273 struct qinst
*qinst
= chosen
? chosen
->inst
: vir_nop();
1274 struct v3d_qpu_instr
*inst
= &qinst
->qpu
;
1277 fprintf(stderr
, "t=%4d: current list:\n",
1279 dump_state(devinfo
, schedule_list
);
1280 fprintf(stderr
, "t=%4d: chose: ", time
);
1281 v3d_qpu_dump(devinfo
, inst
);
1282 fprintf(stderr
, "\n");
1285 /* We can't mark_instruction_scheduled() the chosen inst until
1286 * we're done identifying instructions to merge, so put the
1287 * merged instructions on a list for a moment.
1289 struct list_head merged_list
;
1290 list_inithead(&merged_list
);
1292 /* Schedule this instruction onto the QPU list. Also try to
1293 * find an instruction to pair with it.
1296 time
= MAX2(chosen
->unblocked_time
, time
);
1297 list_del(&chosen
->link
);
1298 mark_instruction_scheduled(schedule_list
, time
,
1302 choose_instruction_to_schedule(devinfo
,
1306 time
= MAX2(merge
->unblocked_time
, time
);
1307 list_del(&merge
->link
);
1308 list_addtail(&merge
->link
, &merged_list
);
1309 (void)qpu_merge_inst(devinfo
, inst
,
1310 inst
, &merge
->inst
->qpu
);
1311 if (merge
->inst
->uniform
!= -1) {
1312 chosen
->inst
->uniform
=
1313 merge
->inst
->uniform
;
1317 fprintf(stderr
, "t=%4d: merging: ",
1319 v3d_qpu_dump(devinfo
, &merge
->inst
->qpu
);
1320 fprintf(stderr
, "\n");
1321 fprintf(stderr
, " result: ");
1322 v3d_qpu_dump(devinfo
, inst
);
1323 fprintf(stderr
, "\n");
1328 /* Update the uniform index for the rewritten location --
1329 * branch target updating will still need to change
1330 * c->uniform_data[] using this index.
1332 if (qinst
->uniform
!= -1) {
1333 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
)
1334 block
->branch_uniform
= *next_uniform
;
1336 c
->uniform_data
[*next_uniform
] =
1337 orig_uniform_data
[qinst
->uniform
];
1338 c
->uniform_contents
[*next_uniform
] =
1339 orig_uniform_contents
[qinst
->uniform
];
1340 qinst
->uniform
= *next_uniform
;
1345 fprintf(stderr
, "\n");
1348 /* Now that we've scheduled a new instruction, some of its
1349 * children can be promoted to the list of instructions ready to
1350 * be scheduled. Update the children's unblocked time for this
1351 * DAG edge as we do so.
1353 mark_instruction_scheduled(schedule_list
, time
, chosen
, false);
1354 list_for_each_entry(struct schedule_node
, merge
, &merged_list
,
1356 mark_instruction_scheduled(schedule_list
, time
, merge
,
1359 /* The merged VIR instruction doesn't get re-added to the
1360 * block, so free it now.
1365 if (inst
->sig
.thrsw
) {
1366 time
+= emit_thrsw(c
, block
, scoreboard
, qinst
, false);
1368 insert_scheduled_instruction(c
, block
,
1371 if (inst
->type
== V3D_QPU_INSTR_TYPE_BRANCH
) {
1372 block
->branch_qpu_ip
= c
->qpu_inst_count
- 1;
1373 /* Fill the delay slots.
1375 * We should fill these with actual instructions,
1376 * instead, but that will probably need to be done
1377 * after this, once we know what the leading
1378 * instructions of the successors are (so we can
1379 * handle A/B register file write latency)
1381 for (int i
= 0; i
< 3; i
++)
1382 emit_nop(c
, block
, scoreboard
);
1391 qpu_schedule_instructions_block(struct v3d_compile
*c
,
1392 struct choose_scoreboard
*scoreboard
,
1393 struct qblock
*block
,
1394 enum quniform_contents
*orig_uniform_contents
,
1395 uint32_t *orig_uniform_data
,
1396 uint32_t *next_uniform
)
1398 void *mem_ctx
= ralloc_context(NULL
);
1399 struct list_head schedule_list
;
1401 list_inithead(&schedule_list
);
1403 /* Wrap each instruction in a scheduler structure. */
1404 while (!list_empty(&block
->instructions
)) {
1405 struct qinst
*qinst
= (struct qinst
*)block
->instructions
.next
;
1406 struct schedule_node
*n
=
1407 rzalloc(mem_ctx
, struct schedule_node
);
1411 list_del(&qinst
->link
);
1412 list_addtail(&n
->link
, &schedule_list
);
1415 calculate_forward_deps(c
, &schedule_list
);
1416 calculate_reverse_deps(c
, &schedule_list
);
1418 list_for_each_entry(struct schedule_node
, n
, &schedule_list
, link
) {
1422 uint32_t cycles
= schedule_instructions(c
, scoreboard
, block
,
1424 orig_uniform_contents
,
1428 ralloc_free(mem_ctx
);
1434 qpu_set_branch_targets(struct v3d_compile
*c
)
1436 vir_for_each_block(block
, c
) {
1437 /* The end block of the program has no branch. */
1438 if (!block
->successors
[0])
1441 /* If there was no branch instruction, then the successor
1442 * block must follow immediately after this one.
1444 if (block
->branch_qpu_ip
== ~0) {
1445 assert(block
->end_qpu_ip
+ 1 ==
1446 block
->successors
[0]->start_qpu_ip
);
1450 /* Walk back through the delay slots to find the branch
1453 struct list_head
*entry
= block
->instructions
.prev
;
1454 for (int i
= 0; i
< 3; i
++)
1455 entry
= entry
->prev
;
1456 struct qinst
*branch
= container_of(entry
, branch
, link
);
1457 assert(branch
->qpu
.type
== V3D_QPU_INSTR_TYPE_BRANCH
);
1459 /* Make sure that the if-we-don't-jump
1460 * successor was scheduled just after the
1463 assert(!block
->successors
[1] ||
1464 block
->successors
[1]->start_qpu_ip
==
1465 block
->branch_qpu_ip
+ 4);
1467 branch
->qpu
.branch
.offset
=
1468 ((block
->successors
[0]->start_qpu_ip
-
1469 (block
->branch_qpu_ip
+ 4)) *
1472 /* Set up the relative offset to jump in the
1475 * Use a temporary here, because
1476 * uniform_data[inst->uniform] may be shared
1477 * between multiple instructions.
1479 assert(c
->uniform_contents
[branch
->uniform
] == QUNIFORM_CONSTANT
);
1480 c
->uniform_data
[branch
->uniform
] =
1481 (block
->successors
[0]->start_uniform
-
1482 (block
->branch_uniform
+ 1)) * 4;
1487 v3d_qpu_schedule_instructions(struct v3d_compile
*c
)
1489 const struct v3d_device_info
*devinfo
= c
->devinfo
;
1490 struct qblock
*end_block
= list_last_entry(&c
->blocks
,
1491 struct qblock
, link
);
1493 /* We reorder the uniforms as we schedule instructions, so save the
1494 * old data off and replace it.
1496 uint32_t *uniform_data
= c
->uniform_data
;
1497 enum quniform_contents
*uniform_contents
= c
->uniform_contents
;
1498 c
->uniform_contents
= ralloc_array(c
, enum quniform_contents
,
1500 c
->uniform_data
= ralloc_array(c
, uint32_t, c
->num_uniforms
);
1501 c
->uniform_array_size
= c
->num_uniforms
;
1502 uint32_t next_uniform
= 0;
1504 struct choose_scoreboard scoreboard
;
1505 memset(&scoreboard
, 0, sizeof(scoreboard
));
1506 scoreboard
.last_ldvary_tick
= -10;
1507 scoreboard
.last_magic_sfu_write_tick
= -10;
1508 scoreboard
.last_uniforms_reset_tick
= -10;
1509 scoreboard
.last_thrsw_tick
= -10;
1512 fprintf(stderr
, "Pre-schedule instructions\n");
1513 vir_for_each_block(block
, c
) {
1514 fprintf(stderr
, "BLOCK %d\n", block
->index
);
1515 list_for_each_entry(struct qinst
, qinst
,
1516 &block
->instructions
, link
) {
1517 v3d_qpu_dump(devinfo
, &qinst
->qpu
);
1518 fprintf(stderr
, "\n");
1521 fprintf(stderr
, "\n");
1524 uint32_t cycles
= 0;
1525 vir_for_each_block(block
, c
) {
1526 block
->start_qpu_ip
= c
->qpu_inst_count
;
1527 block
->branch_qpu_ip
= ~0;
1528 block
->start_uniform
= next_uniform
;
1530 cycles
+= qpu_schedule_instructions_block(c
,
1537 block
->end_qpu_ip
= c
->qpu_inst_count
- 1;
1540 /* Emit the program-end THRSW instruction. */;
1541 struct qinst
*thrsw
= vir_nop();
1542 thrsw
->qpu
.sig
.thrsw
= true;
1543 emit_thrsw(c
, end_block
, &scoreboard
, thrsw
, true);
1545 qpu_set_branch_targets(c
);
1547 assert(next_uniform
== c
->num_uniforms
);