Sleep for 1ms (which is always (a lot) more than 1ms) is just
randomly guessing and hoping you hit your deadline, give or take.
But given we can calculate when our next frame is happening, we
can just sleep for that exact amount. As these values are often
a bit larger, it is also more likely the OS can schedule us back
in close to our requested target. This means it is more likely we
hit our deadlines, which makes the FPS a lot more stable.
Before, every next frame was calculated from the current time.
If for some reason the current frame was drifting a bit, the
next would too, and the next more, etc etc. This meant we rarely
hit the targets we would like, like 33.33fps.
Instead, allow video-drivers to drift slightly, and schedule the
next frame based on the time the last should have happened. Only
if the drift gets too much, that deadlines are missed for longer
period of times, schedule the next frame based on the current
time.
This makes the FPS a lot smoother, as sleeps aren't as exact as
you might think.
During fast-forward, the game was drawing as fast as it could. This
means that the fast-forward was limited also by how fast we could
draw, something that people in general don't expect.
To give an extreme case, if you are fully zoomed out on a busy
map, fast-forward would be mostly limited because of the time it
takes to draw the screen.
By decoupling the draw-tick and game-tick, we can keep the pace
of the draw-tick the same while speeding up the game-tick. To use
the extreme case as example again, if you are fully zoomed out
now, the screen only redraws 33.33 times per second, fast-forwarding
or not. This means fast-forward is much more likely to go at the
same speed, no matter what you are looking at.
_realtime_tick was reset every time the diff was calculated. This
means if it would trigger, say, every N.9 milliseconds, it would
after two iterations already drift a millisecond. This adds up
pretty quick.
On all OSes we tested the std::chrono::steady_clock is of a high
enough resolution to do millisecond measurements, which is all we
need.
By accident, this fixes a Win32 driver bug, where we would never
hit our targets, as the resolution of the clock was too low to
do accurate millisecond measurements with (it was ~16ms resolution
instead).
In 10 years there was no active development on DOS. Although it
turned out to still work, the FPS was very bad. There is little
interest in the current community to look into this.
Further more, we like to switch to c++11 functions for threads,
which are not implemented by DJGPP, the only current compiler
for DOS.
Additionally, DOS is the only platform which does not support
networking. It is the reason we have tons of #ifdefs to support
disabling networking.
By removing DOS support, we can both use c++11 functions for threads,
and remove all the code related to disabling network. Sadly, this
means we have to see DOS go.
Of course, if you feel up for the task, simply revert this commit,
and implement stub c++11 functions for threads and stub functions
for networking. We are more than happy to accept such Pull Request.
Frame rate and various game loop/graphics timing measurements and graphs. Accessible via the Help menu, and can print some stats in the console via the fps command.
The benefit of making this function asynchronious ranges from 2%-25% (real time) during fast forward on dual core/hyperthreading-enabled CPUs; 8bpp improvements are, in my test cases, significantly smaller than 32bpp improvements.
On single core non-hyperthreading-enabled CPUs the extra locking/scheduling costs up to 1% extra realtime in fast forward. You can use -v sdl:no_threads to disable threading and undo this loss.
During normal non-fast-forwarded games the benefit/costs are negligable except when the gameloop takes more than about 90% of the time of a tick.
Note that allegro's performance does not improve with this system, likely due to their way of getting data to the video card. It is not implemented for the OS X/Windows video backends, unless (ofcourse) SDL is used there.
Funny is that the performance of the 32bpp(-anim) blitter is, at least in some test cases, significantly faster (more than 10%) than the 8bpp(-optimized) blitter when looking at real time in fast forward on a dual core CPU; it was slower.
The idea comes from a paper/report by Idar Borlaug and Knut Imar Hagen.
-Fix [FS#2864]: autopause and manual pausing conflict with eachother
-Fix: new game + pause on new game + autopause make the game not unpause on the first join
