Instead of replicating the RGSS Disposable interface in C++
and merely binding it, redefine the 'disposed' state as the
entire core object being deleted (and the binding object's
private pointer being null).
This makes the behavior more accurate in regard to RMXP.
It is now for example possible to subclass disposable classes
and access their 'dispose'/'disposed?' methods without
initializing the base class first (because the internal pointer
is simply null before initialization). Accessing any other
base methods will still raise an exception.
There are some quirks and irregular behavior in RMXP; eg.
most nullable bitmap attributes of disposable classes
(Sprite, Plane etc.) can still be queried afterwards, but
some cannot (Tilemap#tileset), and disposing certain
attributes crashes RMXP entirely (Tilemap#autotiles[n]).
mkxp tries to behave as close possible, but will be more
lenient some circumstances.
To the core, disposed bitmap attributes will look
identically to null, which slightly diverges from RMXP
(where they're treated as still existing, but aren't drawn).
The Disposable interface has been retained containing a
single signal, for the binding to inform core when
objects are disposed (so active attributes can be set to null).
This adds a new dependency with libfuildsynth. MIDI support
is built by default, but can be disabled if not desired.
All RTP songs should work well, but there are known problems
with other files (see README). Also, the pitch shift implementation
is somewhat poor and doesn't match RMXP (at least subjectively).
A soundfont is not included and must be provided by
the user themself.
This bug occured when starting playback of a stream, then immediately
stopping it, loading a different source, and starting playback again.
The real issue was that in stopStream(), the streaming thread had
not even queued anything yet, so it first decoded some data, then
started playing the source (which had already been stopped in the main
thread), and then finally saw the term request and stopped.
Instead stopping the source after the thread has definitely
terminated fixed the problem.
Very rarely rogue buffers would remain and play on loop on song
switch because we only ever cleared processed, not queued, buffers
from the source.
The correct way to completely clear a source's queue is to
simply attach a null buffer to it.
Previously, on creation, we would parse the entire map data,
translating it into and uploading vertices once, then rendering
the entire map on every draw (to keep the draw calls minimal).
This worked great for smaller and medium sized maps, but starting
with larger maps (200x200+) it doesn't scale as the GPUs vertex
processing/culling is overwhelmed by the amount of data each frame.
This rewrite instead changes the strategy to only processing and
uploading a small subregion of the map (the currently visible part)
and regenerating all buffers if this subregion changes. The amount
of data transferred is small enough that it can be done every frame
without causing lag.
The changes also have the convenient side effect that we no longer
require 32 bit indices in mkxp, easing the road to possible GLES2
support in the future.
RGSS allows the source rectangle in both `blt` and
`stretch_blt` to lie outside the source bitmap bounds
(treating the missing data as (0, 0, 0, 0)) and to be
inverted (in which case the blitted image is also inverted).
This commit only hanldes a corner case that
arises in the game "Last Scenario"; emulating the full
RGSS behavior is however desirable.
Previously, we would just stuff the entire tilemap vertex data
four times into the buffers, with only the autotile vertices
offset according to the animation frame. This meant we could
prepare the buffers once, and then just bind a different offset
for each animation frame without any shader changes, but it also
lead to a huge amount of data being duplicated (and blowing up
the buffer sizes).
The new method only requires one buffer, and instead animates by
recognizing vertices belonging to autotiles in a custom vertex
shader, which offsets them on the fly according to the animation
index.
With giant tilemaps, this method would turn out to be a little
less efficient, but considering the Tilemap is planned to be
rewritten to only hold the range of tiles visible on the screen
in its buffers, the on the fly offsetting will become neglient,
while at the same time the amount of data we have to send to the
GPU everytime the tilemap is updated is greatly reduced; so a
net win in the end.
Before, we would blindly rotate through the sources (like a
revolver through its chambers), which worked great if one
assumed all sounds to be relatively short and therefore oldest
use == most likely to be free, but breaks if there is one long
sound playing, which would be stopped and overtaken if we rotated
back to it even though there might be other free sources available.
Instead, keep an ascending priority list of sources with last
used == highest priorty that is iterated through for the first
free one, and only if none is found overtake the one with lowest
priority. This also ensures we're always able to play 'SE_SOURCES'
sounds at once independently of their length.
Fixes#37.
Performance can still be crudely measured by turning off
the framelimit and observing the FPS count. For everything
else, there's always callgrind / apitrace.
GL entrypoint resolution is now done manually. This has a couple
immediate benefits, such as not having to retrieve hundreds of
functions pointers that we'll never use. It's also nice to have
an exact overview of all the entrypoints used by mkxp.
This change allows mkxp to run fine with core contexts, not sure
how relevant that is going to be in the future.
What's noteworthy is that _all_ entrypoints, even the ones core
in 1.1 and guaranteed to be in every libGL, are resolved
dynamically.
This has the added benefit of not having to link directly against
libGL anymore, which also cleans up the output of `ldd` quite
a bit (SDL2 loads most system deps dynamically at runtime).
GL headers are still required at build time.
This bit was deprecated/removed in core GL.
There was only one place where this was used (flash tiles
in Tilemap), and since the full shader rewrite, it was
effectively a no-op anyway (flash shader doesn't sample texture).
Previously, any font names requested by RGSS would be translated
directly to filenames by lowercasing and replacing spaces with
underscores (and finally doing some extension substitution).
To make this whole thing work smoother as well as get closer to
how font discovery is done in VX, we now scan the "Fonts/" folder
at startup and index all present font assets by their family name;
now, if an "Open Sans" font is present in "Fonts/", it will be
used regardless of filename.
Font assets with "Regular" style are preferred, but in their
absence, mkxp will make use of any other style it can find for
the respective family. This is not the exact same behavior as
VX, but it should cover 95% of use cases.
Previously, one could substitute fonts via filenames, ie. to
substitute "Arial" with "Open Sans", one would just rename
"OpenSans.ttf" to "arial.ttf" and put it in "Fonts/". With the
above change, this is no longer possible. As an alternative, one
can now explicitly specify font family substitutions via mkxp.conf;
eg. for the above case, one would add
fontSub=Arial>Open Sans
to the configuration file. Multiple such rules can be specified.
In the process, I also added the ability to provide
'Font.(default_)name' with an array of font families to search
for the first existing one instead of a plain string.
This makes the behavior closer to RMXP; however, it doesn't
work 100% the same: when a reference to the 'Font.name' array is
held and additional strings are added to it without re-assignig
the array to 'Font.name', those will be ignored.
This initial implementation emulates the way RMVX splits
the sprite into "chunks" of about 8 pixels, which it then
scrolls left/right on a vertical sine wave. It even
replicates the weird behavior when wave_amp < 0, namely
"shrinking" the src_rect horizontally.
As with bush_opacity, this effect in combination with
rotation will render differently from RMVX.
This gets rid of the "batch/flush" semantics for #set_pixel
and instead just directly uploads the pixel color to the
texture, circumventing the float conversion entirely.
Also makes a lot of code simpler in many places as calling
'flush()' is no longer required for bitmaps.
Any options that are not arrays (ie. RTP paths) specified
as command line options will override entries in mkxp.conf.
The syntax is: --<option>=<value>
Because Windows has case insensitive paths, this should
be turned on (which it is by default) for maximum
compatibility. Can be turned off as an optimization
(this will speed up startup a little depending on the
number of game assets).
nearly all of the previous required extensions are CORE in OpenGL 2.0
the remaining ones need to have fallback checks for ARB vs EXT vs APPLE
variants..
This looks like a pretty major change, but in reality,
80% of it is just renames of types and corresponding
methods.
The config parsing code has been completely replaced
with a boost::program_options based version. This
means that the config file format slightly changed
(checkout the updated README).
I still expect there to be bugs / unforseen events.
Those should be fixed in follow up commits.
Also, finally reverted back to using pkg-config to
locate and link libruby. Yay for less hacks!
An exception is made of TexPool, which will need a
bit more testing before transitioning to std containers.
Also replace 'int' with 'size_t' where it is used only
as an array index.
When using something like Valgrind that will run
mkxp 20 times slower than normal, frame skip will
make the redraw loop completely grind to a halt.
Set 'frameSkip' to false in the config to
avert this.
A previous commit prevented the MeWatch from
starting a BGM stream that was in stopped state.
However, when a new BGM is loaded while the ME
is still playing, the BGM stream will be stopped
even though we want it to start after the ME
finishes.
Also add some comments trying to explain members
of 'AudioStream' a bit better.
We didn't account for the spec dictating that
scissor test does affect FBO blit operations,
which resulted in corrupted output if more than
one viewport with an active effect (tone, color)
was created.
The reason I never caught this before must be
that the fglrx-legacy driver is actually bugged
in this aspect and ignores the scissor on blit.
The general rule I'm aiming for is to <> include
system wide / installed paths / generally everything
that's outside the git managed source tree (this means
mruby paths too!), and "" include everything else,
ie. local mkxp headers.
The only current exception are the mri headers, which
all have './' at their front as to not clash with
system wide ruby headers. I'm leaving them be for now
until I can come up with a better general solution.
With this we now link to libvorbis/ogg directly.
When this is enabled, one can theoretically also
build SDL_sound without ogg support, although I
doubt it makes much of a difference.
Adittionally, count frames instead of samples
for playback offset calculation.
We will not be using librubberband for in place
pitch shifting. RMXP "shifts" PCM based audio
by just playing it back slower/faster, which
OpenAL takes care of for us. A native midi backend
will be able to effortlessly pitch shift by
multiplying note pitches, should we ever get one.
I have been chasing this ghost for way too long.
Making this an option makes no sense. It ought to
be the default behavior, as RMXP pitch shifts PCM
based audio files the exact same way.
This reverts commit ac35d4214e.
This is a major change in the Audio module that comes with
many changes throughout the codebase and dependency list.
The main gist is that we're finally nuking the last pieces
of SFML from the project. sfml-audio brought with itself
unneeded and big dependencies (libsndfile, libvorbisenc)
while at the same time limiting the amount of audio formats
mkxp can support (eg. we now get mp3 for free, and wma/midi
can be implemented by extending SDL_sound directly).
The increased control gained by interfacing with OpenAL directly
will also allow for easy integration of a dedicated audio
stretcher (librubberband), as well as enable us to implement
looped ogg vorbis (via the 'LOOPSTART'/'LOOPLENGTH' tags),
as required by RGSS2, in the future.
The FileSystem class has had its SFML parts removed.
Aditionally, audio file extensions to be supplemented are
now automatically detected based on how SDL_sound was
built (ie. if no mp3 support was built, mkxp won't try
to look for *.mp3 files). The final used extension
can be optionally returned by 'openRead' calls so
SDL_sound and SDL2_image can immediately choose the
right decoder.
The OpenAL context is created and destroyed in main.cpp
along side the GL context.
Finally got around to nuking that ugly pile of shit that was
previously there for PhysFS file enumeration because filepath
cache generation with unencrypted game files + archive + RTP
has started taking around 6 seconds. Thank $DEITY.
We now actively track how far behind / in front of an
ideal timestep we are during each frame, and try to
catch up / delay approximate this timing.
Therefore we use more precise timers and sleep functions
(nanosleep if available). We also delay **before** the
final buffer swap so the frame displays at more consistent
points in time.
Not only should this provide a somewhat more consistent
looking map scrolling at lower frame rates, it also
guarantees that we don't fall out of sync eg. with the
Audio during longer cutscenes.
'Graphics.frameReset()' now finally has a function, in
that it resets the ideal timestep approximation, which I
beliefe was also its job in the original RMXP engine.
I'm not sure how well this will work when the frame rate
is set to the monitor refresh rate and vsync is turned on.
Very likely unnecessary frame skips will occur here and there
due to imprecise timers. In the future we should probably
check if the frame rate is equal to or higher than the
monitor rate, and disable frame skip accordingly.
These changes currently break the F2 FPS display (it shows
a value that's slightly too high).
Releasing a Tilemap atlas into the pool on every map switch
will blow out tons of smaller textures for very little gain,
as atlas textures are already pretty much impossible to
recycle anywhere but in new Tilemaps.
This reverts commit 34d4103111.
Turns out we need at least GLSL 1.50, for which we'd
have to throw our OpenGL 2.0 compatibility in the water.
Nope, not yet.
Using "SDL2/SDL_xxx.h" instead of "SDL_xxx.h" caused
the include paths provided by pkg-config to be ignored,
and headers from a standard include path to be used instead.
If consecutive scanrows in the scene list have no foreign
elements in between them, we batch them up and draw them
in one glDrawElements() call.
This should reduce the Tilemap induced draw calls on
average by at least 50 percent.
This should be almost as fast as reading unencrypted
files from disk now. I also don't see any possible further
optimizations, so this is probably as fast as it gets.
This implementation is also heaps better than the old
one as it doesn't use a (differently sized) aux texture,
meaning the Bitmap discards its old texture and aquires
one of same size, making reuse through the TexPool a
lot more likely. It also saves on the aux texture blits
and binding switches.
As the setup / resource acquisition far outweighs the
actual rendering cost, operation time is relatively
constant no matter how many divisions are used.
This should make graphics.cpp somewhat easier to navigate/read.
GL_EXT_timer_query is also made optional, and if it's not present
dummy functions will be called instead.
The 'tainted' area of a Bitmap describes what parts are no
longer in a 'cleared' state. When we blit to a fully cleared
are of a Bitmap at full opacity, we can completely disregard
the existing pixels in the operation, meaning we can skip any
blending calculations and just blit / upload straight to the
texture. This greatly speeds up text message rendering.
In the process, pixman has become a new dependency for mkxp,
but the results of this optimization are well worth it!
The atlas packing algorithm has been reworked to pack autotiles
and tileset very efficiently into a texture, splitting the tileset
in multiple ways and eliminating the previous duplication of image
data in the atlas across "frames". Animation, which these frames
were designed for, is now done via duplicated buffer frames,
ie. each animation frame has its own VBO and IBO data. This was
not done to save on VRAM (hardly less memory is used), but to
make place for the new atlas layout.
Thanks to this new layout, even with a max texture size of 2048,
one can use tilesets with up to 15000 height. Of course, such
a tileset couldn't be stored in a regular Bitmap to begin with,
which is why I also introduced a hack called "mega surfaces":
software surfaces stored in RAM and wrapped inside a Bitmap,
whose sole purpose is to be passed to a Tilemap as tilesets.
Various other minor changes and fixes are included.
The drawing is now completely shader based, which makes away
with all usage of the depracted matrix stack. This also allows
us to do things like simple translations and texture coordinate
translation directly instead of doing everything indirectly
through matrices.
Fixed vertex attributes ('vertexPointer()' etc) are also
replaced with user defined attribute arrays.
A difference to RMXP is that negative height values
will result in exceptions too.
Also change Bitmap constructors to not allocate Private
struct before potential exceptions.
Jonas Kulla
cremno
Edward Rudd