ppsspp
ppsspp copied to clipboard
hrydgard
Rewind ZSTD support
On top of #16958
Makes the rewind functionality much less of a memory hog, takes the memory requirement down to half or a third or so.
Has two modes, one where we first apply the old memcmp-based compression step, and one where it's replaced with xoring against the base state. The former seems to do surprisingly better, not sure why the xoring isn't as good.
It does seems like there are some potential improvements to the savestate format we could do. All the big memory blocks should come first, so that nothing that changes length is before those big blocks.
How is the performance speed wise, especially on less powerful devices?
I could imagine 1 being better than the default compression level (3 I think?) It could still use a big chunk of memory, though, at least a big chunk if you're using an armv7 phone with 1.5 GB of RAM or something. I know there were some stats that a lot of users still haven't moved to expensive powerful phones, despite how it might seem.
-[Unknown]
Performance seems fine on the devices I've tested on, but probably would be bad if your device was single-core, in which case you better stay away from the feature anyway... It is a bit more expensive on the battery though, since it spend up to half a second compressing a save state on a background thread on a pretty old device. But I don't really know if I want to add an option...
The idea was to get memory consumption down to a level where it's not much worse than for example cranking resolution up to 10x, in which case the "mem hog" label would need to be applied to such features too - or not at all, hopefully you shouldn't have to think about it.
It is not quite as good as I hoped yet though... heavy games are now at about 8-15MB / state, so about 300MB ram with 20 saves. Which is a drastic improvement and will be fine on most devices, but still.. it is a bit of a hog :P Though many games are more like 1-4 MB per state now, like Outrun. EDIT: Scratch that, most games shrink down a lot after a little while once the base state isn't from the menu or something...
We should probably move things around in save states a bit, so that things like memory will always be at the same offset, at least for the same game. Right now CoreTiming can throw wrenches in that in theory at least...
This question probably belongs to earlier commit, but still. How robust is this in face of non-monotonic wall clock? E.g. on my machine the clock is out of whack, and rebooting can make it jump backward, as (of course) can setting it to actual time manually. Even on otherwise fine machines, there's daylight saving time and all that. Timezone can also change (sometimes automatically).
Emulated PSP only takes system time on game boot, then it can freely move away from it by running emulation at different speed or getting back in time with save states(including rewind) or changing the system time itself.
Well, #16958 may have made it so rewind would stop working on clock skew for a little while, although it more matters if time_now_d() is non-monotonic. For example, QueryPerformanceCounter() is probably not affected by skew, but gettimeofday() isn't as safe. clock_gettime(CLOCK_MONOTONIC) would be better.
But this pull probably doesn't change anything.
-[Unknown]
If performance allows, it is desirable to add a finer time interval granularity, such as 0.1 seconds.This will result in a smoother rewinding experience.
@Josejeffslvh Don't post feature request in random unrelated pulls or issues please. Search for relevant existing issues, or create a new one.