forked from KolibriOS/kolibrios
Fixed Rate Pig - a fixed logic frame rate demo
----------------------------------------------
This SDL programming example - a simple
platform game - demonstrates the use of
a fixed virtual logic frame rate
together with interpolation, for smooth
and accurate game logic that is
independent of the rendering frame rate.
The example also demonstrates sprite
animation and partial display updating
techniques, suitable for games and
applications that need high frame rates
but can do without updating the whole
screen every frame.
Fixed Logic Frame Rate
----------------------
Having a fixed logic frame rate means that the game
logic (that is, what defines the gameplay in terms
of object behavior and user input handling) runs a
fixed number of times per unit of time. This makes
it possible to use "frame count" as a unit of time.
More interestingly, since the logic frame rate
can be set at any sufficient value (say, 20 Hz for
a slow turn based game, or 100 Hz for fast action)
the logic code will run exactly once per frame.
Thus, there is no need to take delta times in
account, solving equations, making calculations on
velocity, acceleration, jerk and stuff like that.
You can just deal with hardcoded "step" values and
simple tests.
Perhaps most importantly, you can *still* rely
on the game behaving *exactly* the same way,
regardless of the rendering frame rate or other
system dependent parameters - something that is
virtually impossible with delta times, since you
cannot have infinite accuracy in the calculations.
Virtual Logic Frame Rate
------------------------
By "virtual", I mean that the actual frame rate is
not necessarily stable at the nominal value at all
times. Rather, the *average* logic frame rate is
kept at the nominal value by means of controlling
the number of logic frames processed for each
rendered frame.
That is, if the rendering frame rate is lower
than the nominal logic frame rate, the engine will
run the game logic several times before rendering
each frame. Thus, the game logic may actually be
running at tens of kHz for a few frames at a time,
but this doesn't matter, as long as the game logic
code relies entirely on logic time.
So, do not try to read time using SDL_GetTicks()
or similar in the game logic code! Instead, just
count logic frames, like we did back in the C64 and
Amiga days, where video frames were actually a
reliable time unit. It really works!
Resampling Distortion
---------------------
Now, there is one problem with fixed logic frame
rates: Resampling distortion. (The same phenomena
that cause poor audio engines to squeal and feep
when playing back waveforms at certain pitches.)
The object coordinates generated by the game
logic engine can be thought of as streams of values
describing signals (in electrical engineering/DSP
terms) with a fixed sample rate. Each coordinate
value is one stream.
Since the logic frame rate is fixed, and the
game logic runs an integer number of times per
rendered frame, what we get is a "nearest point"
resampling from the logic frame rate to the
rendering frame rate. That's not very nice, since
only the last set of coordinates after each run of
logic frames is actually used - the rest are thrown
away!
What's maybe even worse, especially if the logic
frame rate is low, is that you get new coordinates
only every now and then, when the rendering frame
rate is higher than the logic frame rate.
Getting Smooth Animation
------------------------
So, what do we do? Well, given my hint above, the
answer is probably obvious: interpolation! We just
need to replace the basic "nearest sample" method
with something better.
Resampling is a science and an art in the audio
field, and countless papers have been written on
the subject, most of which are probably totally
incomprehensible for anyone who hasn't got a degree
in maths.
However, our requirements for the resampling can
be kept reasonably low by keeping the logic frame
rate reatively high (ie in the same order of
magnitude as the expected rendering frame rate) -
and we generally want to do that anyway, to reduce
the game's control latency.
Chosing An Interpolator
-----------------------
Since the rendering frame rate can vary constantly
in unpredictable ways, we will have to recalculate
the input/output ratio of the resampling filter for
every rendered frame.
However, using a polynomial interpolator (as
opposed to a FIR resampling filter), we can get
away without actually doing anything special. We
just feed the interpolator the coordinates and the
desired fractional frame time, and get the
coordinates calculated.
DSP people will complain that a polynomial
resampler (that is, without a brickwall filter, or
oversampling + bandlimited downsampling) doesn't
really solve the whole problem. Right, it doesn't
remove frequencies above Nyqvist of the rendering
frame rate, so those can cause aliasing distortion.
But let's consider this:
Do we actually *have* significant amounts of
energy at such frequencies in the data from the
game logic? Most probably not! You would have to
have objects bounce around or oscillate at insane
speed to get anywhere near Nyqvist of (that is, 50%
of) any reasonable (ie playable) rendering frame
rate. In fact, we can probably assume that we're
dealing with signals in the range 0..10 Hz. Not
even the transients caused by abrupt changes in
speed and direction will cause visible side
effects.
So, in this programming example, I'm just using
a simple linear interpolator. No filters, no
oversampling or anything like that. As simple as it
gets, but still an incredible improvement over
"nearest sample" resampling. You can enable/disable
interpolation with the F1 key when running the
example.
Rendering Sprites
-----------------
In order to cover another animation related FAQ,
this example includes "smart" partial updates of
the screen. Only areas that are affected by moving
and/or animated sprites are updated.
To keep things simple and not annoyingly non-
deterministic, updates are done by removing all
sprites, updating their positions and animation
frames, and then rendering all sprites. This is
done every frame, and includes all sprites, whether
they move or not.
So, why not update only the sprites that
actually moved? That would allow for cheap but
powerful animated "backgrounds" and the like.
Well, the problem is that sprites can overlap,
and when they do, they start dragging each other
into the update loop, leading to recursion and
potentially circular dependencies. A non-recursive
two-pass (mark + render) algorithm is probably a
better idea than actual recursion. It's quite
doable and neat, if the updates are restricted by
clipping - but I'll leave that for another example.
Pretty much all sprites in Fixed Rate Pig move all
the time, so there's nothing to gain by using a
smarter algorithm.
Efficient Software Rendering
----------------------------
To make it a bit more interesting, I also added
alpha blending for sprite anti-aliasing and effects.
Most 2D graphics APIs and drivers (and as a result,
most SDL backends) lack h/w acceleration of alpha
blended blits, which means the CPU has to perform
the blending. That's relatively expensive, but
SDL's software blitters are pretty fast, and it
turns out *that's* usually not a problem.
However, there is one problem: Alpha blending
requires that data is read from the target surface,
modified, and then written back. Unfortunately,
modern video cards handle CPU reads from VRAM very
poorly. The bandwidth for CPU reads - even on the
latest monster AGP 8x card - is on par with that of
an old hard drive. (I'm not kidding!)
This is why I wanted to demonstrate how to avoid
this problem, by rendering into a s/w back buffer
instead of the h/w display surface. If you're on a
system that supports hardware display surfaces, you
can see the difference by hitting F2 in the game,
to enable/disable rendering directly into VRAM.
Indeed, SDL can set that up for you, but *only*
if you ask for a single buffered display - and we
do NOT want that! Single buffered displays cannot
sync animation with the retrace, and as a result,
we end up hogging the CPU (since we never block,
but just pump out new frames) and still getting
unsmooth animation.
Accidentally, this approach of using a s/w back
buffer for rendering mixes very well with partial
update strategies, so it fits right in.
Smart Dirty Rectangle Management
--------------------------------
The most complicated part of this implementation
is keeping track of the exact areas of the screen
that need updating. Just maintaining one rectangle
per sprite would not be sufficient. A moving sprite
has to be removed, animated and then re-rendered.
That's two rectangles that need to be pushed to the
screen; one to remove the old sprite image, and one
for the new position.
On a double buffered display, it gets even worse,
as the rendering is done into two alternating
buffers. When we update a buffer, the old sprites
in it are actually *two* frames old - not one.
I've chosen to implement a "smart" rectangle
merging algorithm that can deal with all of this
with a minimum of support from higher levels. The
algorithm merges rectangles in order to minimize
overdraw and rectangle count when blitting to and
updating the screen. See the file dirtyrects.txt for
details. You can (sort of) see what's going on by
hitting F3 in the game. Here's what's going on:
1. All sprites are removed from the rendering
buffer. The required information is found
in the variables that store the results of
the interpolation.
2. The dirtyrect table for the display surface
is swapped into a work dirtyrect table. The
display surface dirtyrect table is cleared.
3. New graphic coordinates are calculated, and
all sprites are rendered into the rendering
buffer. The bounding rectangles are fed
into the display surface dirtyrect table.
4. The dirtyrect table compiled in step 3 is
merged into the work dirtyrect table. The
result covers all areas that need to be
updated to remove old sprites and make the
new ones visible.
5. The dirtyrect table compiled in step 4 is
used to blit from the rendering buffer to
the display surface.
On a double buffered display, there is one
dirtyrect table for each display page, and there
is (obviously) a page flip operation after step 5,
but other than that, the algorithm is the same.
Command Line Options
--------------------
-f Fullscreen
-s Single buffer
<n> Depth = <n> bits
//David Olofson <david@olofson.net>