Now works with games that deliberately avoid invalidating TMEM because
they know textures are too large to fit:
* Sonic Riders
* Metal Arms: Glitch in the System
* Godzilla: Destroy All Monsters Melee
* NHL Slapshot
* Tak and the Power of Juju
* Night at the Museum: Battle of the Smithsonian
* 428: Fūsa Sareta Shibuya de
When I added the software FMA path in 2c38d64 and made us use
it when determinism is enabled, I was assuming that either the
performance impact of software FMA wouldn't be too large or CPUs
that were too old to have FMA instructions were too slow to run
Dolphin well anyway. This was wrong. To give an example, the
netplay performance went from 60 FPS to 30 FPS in one case.
This change makes netplay clients negotiate whether FMA should
be used. If all clients use an x64 CPU that supports FMA, or
AArch64, then FMA is enabled, and otherwise FMA is disabled.
In other words, we sacrifice accuracy if needed to avoid massive
slowdown, but not otherwise. When not using netplay, whether to
enable FMA is simply based on whether the host CPU supports it.
The only remaining case where the software FMA path gets used
under normal circumstances is when an input recording is created
on a CPU with FMA support and then played back on a CPU without.
This is not an especially common scenario (though it can happen),
and TASers are generally less picky about performance and more
picky about accuracy than other users anyway.
With this change, FMA desyncs are avoided between AArch64 and
modern x64 CPUs (unlike before 2c38d64), but we do get FMA
desyncs between AArch64 and old x64 CPUs (like before 2c38d64).
This desync can be avoided by adding a non-FMA path to JitArm64 as
an option, which I will wait with for another pull request so that
we can get the performance regression fixed as quickly as possible.
https://bugs.dolphin-emu.org/issues/12542
The STL has everything we need nowadays.
I have tried to not alter any behavior or semantics with this
change wherever possible. In particular, WriteLow and WriteHigh
in CommandProcessor retain the ability to accidentally undo
another thread's write to the upper half or lower half
respectively. If that should be fixed, it should be done in a
separate commit for clarity. One thing did change: The places
where we were using += on a volatile variable (not an atomic
operation) are now using fetch_add (actually an atomic operation).
Tested with single core and dual core on x86-64 and AArch64.
