This is mostly the same approach as in GL1. I'm not quite sure if the
software rasterizer can work with all aspects and the like but I wasn't
able to crash it by trying several random resultions.
With this renamed cvars can be rewritten when config.cfg is first
loaded. Please note that once this was done older YQ2 versions can't
parse that config.cfg anymore.
gl_maxfps > 1000 breaks things, and cl_maxfps starts to behave weird
at >90, and while up to 125 or so you get the bugfeature of higher
jumping, beyond that things just get even buggier, at some point causing
bugs like #261
If too many of these sounds are started in one frame (for example if the
player shoots with the super shotgun into the power screen of a Brain)
things get too loud and OpenAL is forced to scale the volume of several
other sounds and the background music down. That leads to a noticable
and annoying drop in the overall volume.
Work around that by limiting the number of sounds started. 16 was
choosen by empirical testing.
This was so broken... Casting the type of an array to silence a
warning... It worked on x86, of course. But gave a SIGBUS on ARM.
Do it right, cast / copy the content of the array into another
array of the correct type. Yeah.
This fixes issue #231.
There're two possible problems with the calculation of the number of
sound buffers for Vorbis if OpenAL is in use:
* We assume that the (more or less) maximum number buffers is allocated
during map load. This is not correct if in a multiplayer game a lot of
custom models with custom sound connect at a later time.
* 64 buffers (about 3 seconds worth of music) may be too low in some
situations.
Work around this by recalculating the number of buffers if necessary.
We're now reserving about 256 (== 12 seconds) buffers.
This may fix issue #252.
turns out clock_get_time() uses mach_timespec_t which is very similar
to POSIX timespec_t so we're back to just one Sys_Microseconds() function
with an #ifdef __APPLE__ for the (relatively small) differences
Older versions of OS X don't implement clock_gettime() and no(?) version
seems to implement CLOCK_MONOTONIC. Work around this by implementing an
OS X specific variant of Sys_Microseconds() that relies on Mach APIs
provided by all OS X versions...
While at it alter the generic variant so that CLOCK_MONOTONIC is used
only if it's available. CLOCK_REALTIME as a fallback should be good
enough in most cases.
This is believed to fix issue #239.
We need to take in account that scaling the characters makes them
bigger, thus they need need to be places depending on the scale and not
at a precaclulated position. This should fix issue #247.
Returning 'microseconds / 1000ll' at the first call is wrong, the game
would thing that the first frame too way too much time. For some reason
this wirks in (my) Win10, but breaks on (my) Win7...
The original client used single precision mode on Windows and the
default mode on all other platforms. Most platform (at least OS X,
FreeBSD, NetBSD up to 6.0, OpenBSD and Solaris) set double precision
as default, Linux sets extended double precision... When playing a
network game there're several possibilities:
* Same precision on both sides: This one is okay, of course.
* single precision <-> double precision: This one is okay, too. I guess
this is because the code allows a small deviation between client and
server to work around imprecisions introduced be the network protocol.
* double precision <-> extended double precision: This one is okay,
likely for the same reasons given above.
* single precision <-> extended double precision: This one gives a lot
of misspredictions at client side.
All of these are more or less academic these days. Yamagi Quake II used
the platforms default mode for ages. And both gcc and clang default to
SSE2 math (with double precision as default on all platforms) when
compiling for amd64. So the only reasonable case is Linux/i386 on one
side and the original client or another source port on Windows/i386 at
the other side.
Work around this by forcing the x87 to double precision mode.
Miscframes are coupled to renderframes and are just checking for
renderer changes (very cheap) and advancing CD audio if implemented.
There's no reason not to that at every frame.
Until now the curtime variable was set at every call Sys_*seconds().
That's a little bit unfortunate because calls to that functions are
scattered around the code. Instead set it once every frame in
Qcommon_Frame().
The dedicated server runs at cl_maxfps frames per second. Een with very
large values one server frame can never be shorter than 1 milliseconds.
And the timing doesn't need to be very precise since the network
latency adds a lot of more jitter.
Yes, this duplicates some code. But it's at least 100 times more
readable to have two distinct functions for distinct purposes instead
of about 25 #ifdef.
This shouldn't have any noteable impact on timing (besides the machine
is way too slow for Quake II) and saves a lot of CPU cycles. 100% load
vs. 17% load on my desktop.
Having the server in an own timing zone seems to simplify things but
introduces slight timing discrepancies. The most visible effect is that
the game runs a little bit too fast, especially in the first cl_maxfps
frames.
Therefor: Remove timeframes, they're unnecessary. Track the time since
the last (client|server) frame instead and pass it to the client and
server when it's called.
This allows us to implement the global timing without an artificial
brake slowing the game unnecessary down. This is only partial working,
more changes and fixes are coming.
This is a no-op for now. We need this to get a much higher precision
when calculating the frame times. This changes the fixedtime cvar from
milli- to microseconds.
This is the same as the client does for it's realtime. It looks at least
somewhat more correct since it pevents rounding errors. And things are
simplified a litte bit since the server timing is now independent of the
global timing.
The old framecounter had two problems:
* It measured only the time of the current render frame, not the total
time spend between the last and the current render frame. Therefor the
calculated value was too high.
* It was based upon milliseconds and rather inaccurate.
This new frame counter solves both problems. The total time spend
between two render frames is measured and the measurement done in
microseconds.
There're three modes:
* cl_drawfps 1 displayes the average frame rate calculated over the last
60 frames.
* cl_drawfps 2 displays a nice string with minimal framerate, maximum
framerate and average framerate. All three values are calculated over
the last 60 frames.
* cl_drawfps 3 is the same as number 2 but with a second line showing the
raw values.
TODO:
* Discuss if cl_drawfps should be renamed to cl_showfps. All other
status displays are named cl_show*.
While at it remove several unsused drawing functions.
This is the same as the well known Sys_Milliseconds() but like the name
suggests with microsecond precision. To be used in the upcoming new
framecounter.
For some fucking reason, if you set an unsupported
SDL_GL_MULTISAMPLESAMPLES value on Windows (at least Win10 with Intel GPU
drivers, there 16 is unsupported), creating the Window and OpenGL context
will succeed, but you'll get Microsofts stupid GDI OpenGL software
implementation that only supports OpenGL 1.1.
Before these fixes, the GL3 renderer would just crash and the GL1 renderer
would fail to load, which caused the game to run in the background:
No Window, no Input, but sound was playing..
Now this problem should be handled properly and if initialization fails,
the rendering backend will be considered not working, and it will
try the gl1 backend next, and if that also fails it'll give up and exit
the game.
Until now the video menu enforced:
* fov set to 90 and horplus set to 1
* fov set to something other than 90 and horplus to 0
If the user hat configured another configuration through the console the
menu would reset it, even if only unrelated changes are applied. With
this change horplus is ignored by the menu and only fov is altered. The
rationale behind this is that most users want horplus enabled and all
others can disable it through the console.
This is believed to fix issue #225.
While here reimplement the same hack for baseq2/players, lost somewhere
on the way. This is just another searchpath f*ckup. For some reasons
paks have a higher priority than plain directories. We do not want that
for the maps.lst and players/ since id Software decided to put updated
versions of them directly into baseq2/...
This closes issue #217.
SDL_WINDOW_FULLSCREEN changes the display resolution if the requested
resolution is different to the actual resultion. SDL_WINDOW_FULLSCREEN_
DESKTOP doesn't do that, it places a smaller or bigger render area
somewhere inside the fullscreen area. This is somewhat nicer with modern
high resolution flatscreens.
This commit changes vid_fullscreen 1 from SDL_WINDOW_FULLSCREEN to
SDL_WINDOW_FULLSCREEN_DESKTOP. Additional vid_fullscreen 2 is
implemented, it uses SDL_WINDOW_FULLSCREEN to create the fullscreen
area.
TL;DR: Use vid_fullscreen 1 to keep the current resolution or use
vid_fullscreen 2 to switch the resolution.
Implementation details: The whole fullscreen stuff is a horrible mess.
Like generations of hackers before me I'm not desperated enough to clean
it up. GLimp_InitGraphics() is modified to take the fullscreen mode as
an integer and not as a boolean. That's a change to the renderer API.
In GLimp_InitGraphics() the needed SDL fullscreen mode flag is
determined once at the top and just used further down below. That saves
dome SDL1 <-> SDL2 compatibility cruft. IsFullscreen() was modified to
return the actual fullscreen mode and not just if fullscreen is enabled.
Several platforms - OpenBSD being a prominent example - don't provide a
way to get the executable path. Don't abort, just return the current
dir ./ executable dir. This is just a work around, of course. The user
needs to supply a script that calls ./quake2 in the correct directory.
The big problem with the old implementation was that stdout.txt and
stderr.txt on Windows became available when nearly all the low level
initialization was already done. Regardless if the client was in
normal or in portable mode.
Solve this by scanning the command line for the string '-portable'. If
it's not found, stdout and stderr are redirected as early as possible.
If found the global variable (*sigh*) is_portable is set to true. It's
evaluated later on to set the cvar 'portable', which in turn is used
be the filesystem to decide if the home directory should be added to
the search path.
Maybe we should remove the cvar and stick to the global variable.
While at it change the maximum path length for qconsole.log from
MAX_QPATH to MAX_OSPATH. At least on my Linux laptop MAX_QPATH is
too short.
This commit is still untested on Windows!
A new linked list fs_rawPath with nodes of type fsRawPath_t is added.
The new function FS_BuildRawPath() fills it at filesystem initialization
with the raw search path directories. Later FS_BuildGenericSearchPath()
and FS_BuildGameSpecificSearchPath() use it to derive the actual search
directories.
Remove all functions that are no longer used:
* FS_AddGameDirectory()
* FS_SetGameDir()
* FS_AddHomeAsGameDirectory()
* FS_AddBinaryDirAsGameDirectory()
While at it try remove as much global variables from filesystem.c as
possible. Also fix a small, longstandig bug: The download code should
treat .zip and .pk3 files as pak files and not as normal directories.
Refactor FS_SetGamedir() into FS_BuildGameSpecificSearchPath(). The new
function removes the specialized part of the search path if necessary
and create a new specialized part based upon the given directory. It
uses the FS_AddDirToSearchpath() function added in the last commit.
This moves the code used to add a directory and it's paks to the search
path into one well defined function FS_AddDirToSearchPath(). Also create
a new function FS_BuildGenericSearchPath() that builds the generic part
of the global search path. This obsoletes several other, specialized
functions. They'll be removed in a later commit.
This prevents Windows from scaling our (fullscreen) window to crap if
the whole desktop is scaled and we're rendering more than 1080p. This is
believed to fix#208.
if that cvar is set to 1, particles aren't rendered as nice circles, but
as squares, like in the software renderer or in Quake1.
Also documented it in cvarlist.md and fixed some typos there
The model shadows are rendered after all entities are rendered.
This fixes them making entity brushes below them translucent (#194)
The model rendering code used lots of global variables, many of them
totally superfluous (esp. currententity, currentmodel).
I refactored the code to use less global variables (this was at least
partly needed to render the shadows later).
So this looks like lots of changes, but many of them are just using
"entity" instead of "currententity" or "model" instead of "currentmodel"
Like GL1 gl_shadows + gl_stencilshadows: no shadow volumes, but looks
ok apart from standing over edges
The gl_stencilshadows cvar isn't used in GL3, it always uses the stencil
buffer if available (and if gl_shadows != 0)
This still needs performance optimizations: Like the GL1 impl it takes
lots of draw calls per model, it could be done with one per model like
when rendering the actual model.
there have been complaints that those things look too bright, so let
people configure their intensity independently of the general intensity
used for levels, monsters etc.
fixes#189
When applied to y SURF_FLOWING textures are scrolled into the wrong
direction. I guess that in GL1 the offset is also applied to x.
This fixes issue #186.
Sometimes cinematics are skipped after the first frame even if the
player didn't press any key. I'm unable to reliable reproduce that,
so my educated guess is that one or more events are still waiting in
SDLs event queue.
For example, during intermission IN_Update() is not called for 5
seconds, key presses by impatient players are just added to the queue
and not processed. The first event is used to skip leave the
intermission, the second event skips the cinematic...
Fix this by implementing a new function IN_FlushQueue() to flush SDLs
event queue and calling it when starting cinematic playback. Yes, this
is just another layer violation. :(
When the client is paused (either explicit or by entering the menu or
console) the cinematic is paused, too. Therefor no more sound samples
are generated and added to the playback queue, the existing samples are
played over and over again. Until now these samples weren't hearable,
because OpenAL marked them as processed and AL_StreamUpdate() removed
them from OpenALs playback queues. This changed in the previous commit,
now the stay in OpenALs queue and are hearable.
Fix this by calling AL_UnqueueRawSamples() when the menu or console is
entered during cinematic playback.
Newer openal-soft versions changed the way how the processed buffers are
counted when in AL_STOPPED state. Previously only processed buffers were
counted, now all buffers are. Change our unqueue logic to match this new
behavior.
This was debugged and fixed @xorw, I'm just committing the patch. This
closes issue #185.
For some reasons setting the MSAA fails at window creation and not at
GL context creation. And of course SDL is unable to detect before, that
the requested number of MSAA samples is invalid... Implement a work
around: Fall back to gl_msaa_samples == 0 if the window cannot be
created.
Resurrect support for render / refresher loadable libraries and use them to implement an experimental OpenGL 3.2 renderer. Please note that the new renderer interface is somewhat different from the original one, old render libraries will NOT work!
there's a target_secret (with targetname "t117"), but no one triggers
it - that's why the help computer shows four secrets, but you can only
get three of them.
Now when you open the door in front of the hidden secret armor
(by shooting it), it'll trigger the target_secret and you can get all
four secrets.
fixes#182
The internal order of the items is determined by Menu_AddItem() and
not the y position. Without this change the cursor didn't jump from
item to item, but from the mode list box to the aspect list box,
skipping the brightness slider.
- Bump vid_gamma to 1.2 in both GL1 and GL3. A default value of 1.0 is
too dark.
- Lower gl3_overbrightbits to 1.3, the previous value of 1.5 was too
bright. This can be seen in later units, for example on mine1 some
textures blended into white.
- Lower gl3_particle_size to 40. A value of 60 may be okay, but with
gl3_particle_fade_factor 1.2 the particles take up too much screen
estate in close range combat.
With this changes GL3 looks (at least for me) nearly the same as GL1
rendered through the removed multitexturing path.
Without capping the brightness entity models may fade into pure white
which looks ugly. This can be seen when several flyer fire blaster
bolds onto the player or when multiple barrels are exploding. This
change was suggest by @DanielGibson, I'm just the messenger.
I really don't see why this constraint was ever necessary. It leads to
one line of pixels not rendered either at the bottom or the right edge
of the screen. In GL1 for whatever reasons this line is just black, in
GL3 garbage is drawn.
In SCR_DrawFieldScaled() the HUD scale factor wasn't taken into account
when calculating the screen area affected by the change. Therefor wrong
coordinates were passed to SCR_AddDirtyPoint() and a part of the changed
area wasn't marked dirty, leading to render artifacts.
This bug was present since HUD scaling was first introduced.
the flash should only be drawn in the part of the window where actual
3D rendering happens, not in the borders added if viewsize < 100
(and apparently also for with 1 pixel width if the resolution is odd).
* gl3_particle_size: in GL3 the particles should be a bit bigger because
the particles fade out towards the edge, so I put it in a seperate
CVar
* gl3_intensity: in GL3 the intensity can have any floating point value,
in GL1 only integers, so it gets its own CVar
* gl3_overbrightbits: gl_overbrightbits had to be 1, 2 or 4, in GL3 it
can have any floating point value.
Changed the particle scaling a bit so they look bigger.
One problem was that GL3_Shutdown() called several functions that use
that gl* function pointers - not a good idea if InitContext() failed
and the function pointers are all NULL. So check for that.
Similarly in GL3/R_ShutdownWindow() calling glClear() etc.
Another problem was that R_SetMode() would, if R_SetMode_impl() failed,
try again with a "safe" resolution (640x480 unless we had another
working resolution before) - which is bad if we're already using that
"safe" resolution because then GLimp_InitGraphics() would check mode
and fullscreen and decide it hasn't changed and do nothing and return
true, which would make SetMode() believe everything is fine and
afterwards all hell breaks loose.
This makes the fragment shader faster by skipping lights that haven't
marked this surface in GL3_MarkLights()
This seems to improve performance at least slightly everywhere, but
it really helps *a lot* on integrated intel GPUs like the one on their
Sandy Bridge, Ivy Bride and Haswell CPUs (those are the ones we tested).
adding dot(surfaceNormal, lightToPixelOnSurfaceNormal) to the equation,
should be Phong-y now? Looks good at least.
The Windows AMD legacy driver needed its usual manual padding..
OSX was totally weird.. There were no errors or warnings from OpenGL
at all, but the dynamic lights were just not visible.
After (too long) debugging the shader I figured out that
dynLights[i].lightIntensity was always 1, and thus
'dynLights[i].lightIntensity - distLightToPos - 64' was negative and set
to 0 with max(0, ...).
I still have no idea why that happens, but removing lightIntensity from
the struct, making lightColor a vec4 and using .a for intensity works...
Dynamic lights on normal world brushes work, on brush-based entities
probably not yet properly. For this I need the model matrix in the
shader to transform vertex positions and normals to worldspace
(they already are for world brushes, but not entities that might rotate
and move etc).
Furthermore, while they dynamic lights look nice and smooth they might
need some fine tuning in the shader..
For this to work there are two bigger changes:
* the vertex data for brushes (gl3_3D_vtx_t) now also contains the
vertex normal
- glpoly_t contains array of gl3_3D_vtx_t instead of 7 floats
* 3D shaders now have in vec3 normal, bound to GL3_ATTRIB_NORMAL
* There's a new UBO for light data: uniLights, containing an array of
up to 32 dynamic lights, with data copied from gl3_newrefdef.dlights
..with Radeon 6950 using AMDs legacy driver.
For uploading UBOs it turned out that glBufferData() is faster,
sometimes a lot faster, with several drivers, especially Intel/OSX.
