Commen=
t # 55
on bug 97879<=
/a>
from =
Timothee Besset
Rocket League uses an async I/O thread to stream in content in=
the background.
That thread works off of a queue and only does disk reads and decompression=
(it
doesn't issue any OpenGL calls for instance, that happens on a dedicated re=
nder
thread).
The CPU usage you are seeing out of Rocket League is this thread being busy
with decompression. This is generally normal and not an indication of a
problem.
However, when running with radeonsi, the async I/O thread is performing very
poorly. Much, much worse than with amdgpupro. To the point that the main th=
read
has to wait for async and causes a significant hitch (up to 4 seconds in so=
me
logs I captured).
I tend to agree that this problem is not related to shader compilation. The
bumps from shader compilation are barely showing up in profiles. But it's
definitely made a lot worse by running on radeonsi. On the same system with
amdgpupro those same hitches can happen but they are in the 100/200ms range.
At this point of the investigation, and due to scheduling constraints, I am
likely only going to come back to this once Timothy Arceri has a shader cac=
he
implementation available. I suspect this will not help significantly but le=
t's
hope I'm wrong there.
In the meantime there are some log files from the async thread compared bet=
ween
radeonsi and amdgpu-pro if someone feels like extracting some possibly usef=
ul
data:
https://www.dropbox.com/s/d73jk66uhfrhv8l/amdgpupro-=
compressedread-timings.log?dl=3D0
https://www.dropbox.com/s/1xfl3l24f5buzs0/radeonsi-co=
mpressedread-timings.log?dl=3D0