fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-02 20:58:04 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

aco: only consider cost of memory loads at waitcnt

Browse source 
We don't run this code before waitcnt insertion, so this isn't necessary.

This change improves accuracy in these two situations, because the waitcnt
insertion pass is more aware of divergent control flow:

v0 = valu
if (divergent) {
    v0 = vmem
} else {
    use(v0)
}

v0 = vmem
if (divergent) {
    wait vmcnt(0)
} else {
    wait vmcnt(0)
}
use(v0)

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38262>
This commit is contained in:
Rhys Perry 2025-11-03 15:23:14 +00:00 • committed by Marge Bot
parent bca5aab2be
commit 88b6b6db17
1 changed files with 1 additions and 8 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

9
src/amd/compiler/aco_statistics.cpp
View file

@ -441,14 +441,7 @@ BlockCycleEstimator::add(aco_ptr<Instruction>& instr)
mem_ops[i].push_back(cur_cycle + wait_info[i]);
}
/* This is inaccurate but shouldn't affect anything after waitcnt insertion.
* Before waitcnt insertion, this is necessary to consider memory operations.
*/
unsigned latency = 0;
for (unsigned i = 0; i < wait_type_num; i++)
latency = MAX2(latency, i == wait_type_vs ? 0 : wait_info[i]);
int32_t result_available = start + MAX2(perf.latency, (int32_t)latency);
int32_t result_available = start + perf.latency;
for (Definition& def : instr->definitions) {
int32_t* available = &reg_available[def.physReg().reg()];
for (unsigned i = 0; i < def.size(); i++)