[Bug 47767] Airport Mania First Flight - Very long loading times and awful performance

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=47767

Paul Gofman <gofmanp at gmail.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |gofmanp at gmail.com

--- Comment #1 from Paul Gofman <gofmanp at gmail.com> ---
I have tested the game trial.

The immediate workaround to get the speed to normal is to disable CSMT (set
HKCU\\Software\\Wine\\Direct3D\\csmt to 0).

The other workaround is to set process affinity for AirportMania.RWG process
after the game is started to default, like 'taskset -p 0xffffffff <pid>', where
pid is process id of 'AirportMania.RWG' process which game creates. This will
help if there is more than one CPU core present.

Yet another workaround is set WINED3D_CS_SPIN_COUNT to 0 in
dlls/wined3d/wined3d_private.h.

The game sets process affinity mask to 1 for the process. Here are related
calls (all that relates to process 'AirportMania.RWG', which is started from
game launcher):

---
002c:Call KERNEL32.GetProcessAffinityMask(ffffffff,0032fad8,0032fac0)
ret=00477314
002c:Ret  KERNEL32.GetProcessAffinityMask() retval=00000001 ret=00477314
002c:Call KERNEL32.SetProcessAffinityMask(ffffffff,00000001) ret=0047744d
002c:Ret  KERNEL32.SetProcessAffinityMask() retval=00000001 ret=0047744d
002c:Call KERNEL32.SetThreadAffinityMask(00000114,0000003e) ret=00477480
002c:Ret  KERNEL32.SetThreadAffinityMask() retval=00000000 ret=00477480
---

I have no idea what the game is trying to achieve by its inventive thread
affinity masks, but I verified that process affinity mask end up being 1 both
under Wine and Windows 7, so that process is bound to a single logical CPU.

Having done that, the game constantly locks and unlocks ddraw7 surfaces, that's
the main load it creates. The most of the CPU spent in wined3d_cs_run() in CS
thread, which is spinning waiting for task to be queued. Mapping of surfaces
has to go through map queue, and also spins waiting while the CS thread does
the actual mapping. Those spins create enough load for an effective single core
execution to leave a very little time for anything else. Removing spin locking
in wined3d_cs_run() (going to wined3d_cs_wait_event() every time) gets normal
performance (while this is not a general improvement to CS thread locking logic
of course).

-- 
Do not reply to this email, post in Bugzilla using the
above URL to reply.
You are receiving this mail because:
You are watching all bug changes.