Windows defaults to `wassapi`, which is a sensible choice. But WASAPI
only guarantees 32 bit float samples, anything else only works if the
driver or something supports it and YQ2 requires 16 bit samples. That
can be worked around by having SDL recode the audio, but I don't want
such a invasive change right before a release.
Another part of #1132.
Hardcoding default driver for some of the supported platform is a
remnant of SDL 1.2 and hasn't been necessary since SDL 2.0 a long
time ago. In fact it has a high properbility to break things, SDL
could easily end up with a non working driver.
When `s_sdldriver` is set to the newly introduced value `auto` the
driver is selected by SDL. In all other cases the string is the
driver name which SDL will be forced to.
This doesn't fix existing configs. Since the OpenAL sound backend has
been the default for nearly 15 years and we haven't received bug reports
for some other problem with the SDL sound backend in the past, I'm half
sure that there are next to users out there. These can reset the cvar
by hand if necessary.
Closes#1132.
Done for all platforms, CMakeLists included. Unneeded with glad.
Renderer fallback logic now includes gles1, just before soft:
custom -> gl3 -> gles3 -> gl1 -> gles1 -> soft.
Static arrays there have their dimensions swapped, they make more sense now.
Added important detail in gl1_lightmapcopies doc.
Also added doc for gl_polyblend, just because of its "popularity".
Available in both GL1 and GLES1. Keep multiple copies of "the same"
lightmap on video memory; they are actually different, because they're
used in different frames. This is a workaround for the usage of
glTexSubImage2D() for dynamic lighting, since modifying textures used
recently causes slowdown in embedded/mobile devices.
Controlled by gl1_lightmapcopies cvar; default in GL1 is `0`, while
in GLES1 is `1`.
Available only on GLES1, allows to use a "performance hint" to the
GPU to discard the contents of depth and stencil buffers after each
frame. Some hardware might want to reuse them, but Quake 2 doesn't
work that way.
Controlled by gl1_discardfb cvar, default `1`.
Variant of GL1, meant for embedded/mobile devices only.
Build it with "make with_gles1".
For Windows, you'll need MSYS2 and a decent ANGLE implementation
(probably not worth the trouble).
Building with CMake only works in Linux, so it has been commented out.
Changed "buffer" functions (vertex, single/mtex, color) to macros.
Macros eliminate the function call overhead, resulting in faster
execution times. Depending of the hardware, there might be a
performance increase of almost 10%.
This forced the buffer to be exposed as global.
Not worth doing this for 2D elements though.
glTexSubImage2D() calls are very slow, and are even slower when
the texture is big. Dynamic lighting changes are small compared
to the huge 512x512 size of the lightmap this option provided,
so it was detrimental to performance.
Original logic remains underneath if there's a need of a comeback.
Implemented a batching procedure, to try to group meshes in a buffer
and use a final GL call to draw them all in one step, instead of the
many GL draw calls existing today.
For now, only 2D textures are included, especifically console text
("conchars"), scrap and tiles. It's not worth doing this for
individual 2D elements (e.g. crosshair).
They were in this form:
glTexParameteri(GL_TEXTURE_2D,GL_TEXTURE_MAG/MIN_FILTER,GL_LINEAR)
Only appeared at startup, but we didn't need so many of them.
ref_gl renders use 0.0..1.0 float texture coordinates in gl commands
and skin ratio/resolution could be any.
ref_soft render uses absolute coordinates as result textures with
different to expected size are applied incorrectly.
Fix will not add support of scalled retextured skins, just fix case
when ratio is incorrect after scale down such or place a incorrect
pcx skin to the model directory.
The first try didn't take into account that an evil server could
override the filter list by sending a stuff command. Fix this by
hardcoding the filters for .dll, .dylib and .so. Make sure that the
filters are always applied, either when the download is requested
through the `download` command or because game data is missing.
This is just a poor mans fix, trying to rule out an obvious way to
inject code into the client.
`cl_nodownload_list` is a whitespace seperated list of strings, files
containing one of these strings in their name are never downloaded. Set
to `.dll .dylib .so` by default to prevent downloading libraries which
can be injected into client.
Closes#1114.
It will be used in a later commit. While at it change Q_strcasecmp() and
Q_strncasecmp() to take the arguments as const, matching the signatures
of the non-replacement versions.
Applied too many OpenGL calls for an effect that could not be told
apart from normal GL_MODULATE; explains its absence from Q3A code.
Also, removed calls to glPixelStorei when no dynamic lights present.
With "hand 2" and no "alias" models in player sight, sometimes only
lightmap was visible. Now forcing disabled multitexture when needed.
Another bugfix: "gl_showbbox 1" no longer produces OpenGL errors.
Controlled by gl1_biglightmaps cvar, works like new lightmaps.
Max size is 384x384; a bigger size is not justified.
Added a transparent border so colors from one picture don't bleed
into the next.
Controlled by new cvar, gl1_biglightmaps.
Size is now up to 512x512, for a max quantity of 8 lightmaps.
Should reduce rebinding of them, and reduce the number of calls
to glTexSubImage2D(), although it may increase data transfer on
each call.
Only happens in Windows. A cvar was being queried just after
all cvars were freed, in Qcommon_Shutdown().
GL_GENERATE_MIPMAP definition relocated.
scrap_uploads counter deleted, it wasn't being read anywhere.
Using MH's solution, which is keeping all lightmaps in memory
to modify and upload them as a batch when possible.
lightmap_buffer is now an array; index 0 is used as the legacy
lightmap buffer (when no mtex), and the rest of the indexes are
to store the different lightmaps (only when using mtex).
Unlike the old, buggy one, this implementation follows the
texture chain, just like the standard execution path. It also
avoids doing the lightmap chains, since it has already done it
in the second TMU; there's no duplicated work for lightmaps.
No errors appear in the lava on the "boss1" map either.
It's still slow when having an overdraw of dynamic lights.
Further work is needed.
While it is based on the old, buggy multitexture implementation
(removed in 68a12d4e & 8fe8f832), rework will be done to make it
function as intended this time.
For this commit, just init and function declarations.
Using GL 1.4 declarations in code, although 'point parameters' will
use legacy extensions if they're the only ones found.
Moved stuff out of local.h; better to have definitions in one place.
Extracted from <glext.h>.
Somehow, R_TextureAnimation() returns different results for the same
surface when called consecutively. We force it to be called once, so
the chains in R_DrawTextureChains() are getting respected now.
...and alter the render finding and fallback sequences to include GLES3
as the first fallback. This should have been done years ago, when GL3
became the clear better option above GL1.
SDL major versions must not be combined in one process, otherwise bad
things will happen. In the best case the game crashes, in the worst case
strange bugs will occure. To prevent that:
* Add a new field `framework_version` to the renderer API and use it to
pass the SDL major version from the renderer to client. Don't load the
renderer if it and the client were build with different SDL major
versions.
* Bump the renderer API version to 7. This could have been implemented
by assuming `framework_version == 0` (not filled by renderer) as SDL
2, but let's keep things clean and bump the version.
While here fix a long standing bug with printing the error when not
loading a renderer lib. The message must be generated before shutting
down the renderer, otherwise the API version will alsways be 0. The
struct is zeroed at renderer shutdown.
Changes are:
* SDL_SetWindowGrab() was renamed to SDL_SetWindowMouseGrab().
* SDL doesn't give precalculated battery states anymore. Use the same
scale as SDL 3.1.0, 20% is low and 1% is empty.
With SDL3 the high dpi support is in much better shape than with SDL2.
And for Wayland the experience on high dpi displays is generally better
when the application is high dpi aware.
Reimplement `SDL_BackendInit()` and `SDL_BackendShutdown()` for SDL 3.
Hide now unnecessary locking calls behind !USE_SDL3. To minimize code
duplication `SDL_Callback()` is hidden behind a wrapper function.
SDL_INIT_VIDEO includes SDL_INIT_EVENTS when initialized through
SDL_Init(), but not when shutdown to SDL_Quit(). Handle the shutdown
in the input subsystem. This closes some memory leaks in SDL.
With SDL 2 the refreshrate handling was rather limited by SDL 2 only
supporting integer refreshrates, making it impossible to represent the
common 59.95hz and other non-integer refreshrates. Quake II works around
this by assuming a refreshrate of 2hz higher than returned by SDL.
SDL 3 finally supports float refreshrates. Refactor the internal API to
also use floats.
This should be a no-op since the 'refreshrate * 1.02' logic stayes in
place for the time being. This can be reevaluated at a later time for
the SDL 3 build, but needs testing.
As a nice side effect fix the `vid_displayrefreshrate` cvar. It was
broken in both SDL 2 and 3 since it was always overwritten with the
actual display refresh rate.
In SDL 2 the first display was 0 and the error code -1. In SDL 3 this
changed to 1 as the first display and 0 for the error code. While at
it implement error handling and fallbacks for all cases.
SDL 3 has a new approach to fullscreen handling, which is way better
than the old fiddling with several window flags. In SDL 3:
* The SDL_FULLSCREEN_FLAG puts the window in fullscreen window mode.
* For exclusive fullscreen a fullscreen mode must be requested and set.
* Applying the mode to the window is asynchronous, the mode change may
occure at any later time. The window can be synchronized to force the
mode change.
This is - of course - totaly untested, since the client doesn't build
yet with SDL 3. There will likely be bugs and shortcomings.
And the code is still ugly, it can be refactored somewhat more.
