fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-09 06:48:06 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

broadcom/compiler: update ldvary thread switch delay slot restriction for v7.x

Browse source 
In V3D 7.x we don't have accumulators which would not survive a thread
switch, so the only restriction is that ldvary can't be placed in the
second delay slot of a thread switch.

shader-db results for UnrealEngine4 shaders:

total instructions in shared programs: 446458 -> 446401 (-0.01%)
instructions in affected programs: 13492 -> 13435 (-0.42%)
helped: 58
HURT: 3
Instructions are helped.

total nops in shared programs: 19571 -> 19541 (-0.15%)
nops in affected programs: 161 -> 131 (-18.63%)
helped: 30
HURT: 0
Nops are helped.

Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25450>
This commit is contained in:
Iago Toral Quiroga 2021-11-23 10:04:49 +01:00 • committed by Marge Bot
parent 526c1889e5
commit 5e9b405aa7
2 changed files with 33 additions and 10 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

33
src/broadcom/compiler/qpu_schedule.c
View file

@ -1491,11 +1491,20 @@ retry:
* ldvary now if the follow-up fixup would place
* it in the delay slots of a thrsw, which is not
* allowed and would prevent the fixup from being
* successful.
* successful. In V3D 7.x we can allow this to happen
* as long as it is not the last delay slot.
*/
if (inst->sig.ldvary &&
scoreboard->last_thrsw_tick + 2 >= scoreboard->tick - 1) {
continue;
if (inst->sig.ldvary) {
if (c->devinfo->ver <= 42 &&
scoreboard->last_thrsw_tick + 2 >=
scoreboard->tick - 1) {
continue;
}
if (c->devinfo->ver >= 71 &&
scoreboard->last_thrsw_tick + 2 ==
scoreboard->tick - 1) {
continue;
}
}
/* We can emit a new tmu lookup with a previous ldtmu
@ -2020,8 +2029,12 @@ qpu_inst_before_thrsw_valid_in_delay_slot(struct v3d_compile *c,
if (slot > 0 && v3d_qpu_instr_is_legacy_sfu(&qinst->qpu))
return false;
if (slot > 0 && qinst->qpu.sig.ldvary)
return false;
if (qinst->qpu.sig.ldvary) {
if (c->devinfo->ver <= 42 && slot > 0)
return false;
if (c->devinfo->ver >= 71 && slot == 2)
return false;
}
/* unifa and the following 3 instructions can't overlap a
* thread switch/end. The docs further clarify that this means
@ -2618,9 +2631,13 @@ fixup_pipelined_ldvary(struct v3d_compile *c,
/* We can't put an ldvary in the delay slots of a thrsw. We should've
* prevented this when pairing up the ldvary with another instruction
* and flagging it for a fixup.
* and flagging it for a fixup. In V3D 7.x this is limited only to the
* second delay slot.
*/
assert(scoreboard->last_thrsw_tick + 2 < scoreboard->tick - 1);
assert((devinfo->ver <= 42 &&
scoreboard->last_thrsw_tick + 2 < scoreboard->tick - 1) ||
(devinfo->ver >= 71 &&
scoreboard->last_thrsw_tick + 2 != scoreboard->tick - 1));
/* Move the ldvary to the previous instruction and remove it from the
* current one.

10
src/broadcom/compiler/qpu_validate.c
View file

@ -215,8 +215,14 @@ qpu_validate_inst(struct v3d_qpu_validate_state *state, struct qinst *qinst)
"SFU write started during THRSW delay slots ");
}
if (inst->sig.ldvary)
fail_instr(state, "LDVARY during THRSW delay slots");
if (inst->sig.ldvary) {
if (devinfo->ver <= 42)
fail_instr(state, "LDVARY during THRSW delay slots");
if (devinfo->ver >= 71 &&
state->ip - state->last_thrsw_ip == 2) {
fail_instr(state, "LDVARY in 2nd THRSW delay slot");
}
}
}
(void)qpu_magic_waddr_matches; /* XXX */