fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-01-29 20:20:24 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 3

broadcom/compiler: don't spill in between multop and umul24

Browse source 
The multop instruction implicitly writes rtop which is not preserved
acrosss thread switches. We can spill the sources of the multop
(since these would happen before multop) and the destination of
umul24 (since that would happen after umul24).

Fixes some OpenCL tests when V3D_DEBUG=opt_compile_time is used to
choose a different compile configuration.

cc: mesa-stable

Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/29909>
(cherry picked from commit 38b7f411a1)
This commit is contained in:
Iago Toral Quiroga 2024-06-26 08:06:33 +02:00 • committed by Eric Engestrom
parent e1ff3ff373
commit 1bf3d6c8e6
2 changed files with 21 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
.pick_status.json
View file

@ -354,7 +354,7 @@
"description": "broadcom/compiler: don't spill in between multop and umul24",
"nominated": true,
"nomination_type": 0,
"resolution": 0,
"resolution": 1,
"main_sha": null,
"because_sha": null,
"notes": null

22
src/broadcom/compiler/vir_register_allocate.c
View file

@ -271,6 +271,7 @@ v3d_choose_spill_node(struct v3d_compile *c)
float block_scale = 1.0;
float spill_costs[c->num_temps];
bool in_tmu_operation = false;
bool rtop_hazard = false;
bool started_last_seg = false;
for (unsigned i = 0; i < c->num_temps; i++)
@ -279,6 +280,20 @@ v3d_choose_spill_node(struct v3d_compile *c)
/* XXX: Scale the cost up when inside of a loop. */
vir_for_each_block(block, c) {
vir_for_each_inst(inst, block) {
/* RTOP is not preserved across thread switches, so
* we can't spill in the middle of multop + umul24.
*/
bool is_multop = false;
bool is_umul24 = false;
if (inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU) {
if (inst->qpu.alu.mul.op == V3D_QPU_M_MULTOP) {
is_multop = true;
rtop_hazard = true;
} else if (inst->qpu.alu.mul.op == V3D_QPU_M_UMUL24) {
is_umul24 = true;
}
}
/* We can't insert new thread switches after
* starting output writes.
*/
@ -297,7 +312,7 @@ v3d_choose_spill_node(struct v3d_compile *c)
if (spill_type != SPILL_TYPE_TMU) {
spill_costs[temp] += block_scale;
} else if (!no_spilling) {
} else if (!no_spilling && (!rtop_hazard || is_multop)) {
float tmu_op_scale = in_tmu_operation ?
3.0 : 1.0;
spill_costs[temp] += (block_scale *
@ -315,7 +330,7 @@ v3d_choose_spill_node(struct v3d_compile *c)
if (spill_type != SPILL_TYPE_TMU) {
/* We just rematerialize it later */
} else if (!no_spilling) {
} else if (!no_spilling && (!rtop_hazard || is_umul24)) {
spill_costs[temp] += (block_scale *
tmu_scale);
} else {
@ -344,6 +359,9 @@ v3d_choose_spill_node(struct v3d_compile *c)
if (qinst_writes_tmu(c->devinfo, inst))
in_tmu_operation = true;
if (is_umul24)
rtop_hazard = false;
}
}