Without this, some code will enable (and set other parameters for)
whatever random texture unit the previous stage had selected. This can
break any fixed-function rendering, since either it will not have
enabled the correct texture unit, or its correct setup will be
interfered with by another texture unit that shouldn't be enabled. (This
doesn't affect ARB/GLSL shader rendering, since they don't care about
the enabled state, but the GUI (and some debug things) are still FFP in
the default ARB mode.)
Fixes#2833 (in which the fancy water shader selected texture units 5/6,
which then got enabled, and remained enabled forever).
This was SVN commit r16372.
This avoids vswprintf failures when printing non-ASCII char* strings
from CLogger into the console.
Also convert ScriptInterface::ToString to return UTF-8, to avoid some
utf8_from_wstring calls.
Also remove some unused and redundant CConsole functions.
This was SVN commit r16333.
This fixes the problem where passing a non-ASCII string to
debug_printf(L"%hs", s) caused vswprintf_s to fail on Linux (because it
doesn't know what encoding the char* is meant to have). Now debug
messages will remain as UTF-8 until they reach the OS.
Fixes#3021.
This was SVN commit r16332.
Some meshes (e.g. pers_gardens_struct with gentangents enabled) have
>32K vertexes, and won't fit in a 2MB VBO. 0ef6c7555e had reduced them
from 4MB to 2MB. So just make the VBOs bigger again, and try not to
worry about the waste of memory.
Fixes#3026, #3042.
This was SVN commit r16309.
Some drivers (at least the Intel drivers on Windows) are slow at
incrementally updating a VBO with hundreds of calls to glBufferSubData
every frame. Performance is significantly better if you use
glBufferData(NULL) to tell it to discard all the previous contents, and
then re-upload all the data at once.
Update CVertexBuffer so that GL_DYNAMIC_DRAW/GL_STREAM_DRAW buffers are
handled with the new mechanism. This requires the caller to hold onto
the backing store so it can be re-uploaded when necessary, and needs a
bit more signalling to indicate exactly what needs uploading.
I see an improvement from roughly 60 to 75 fps on Intel HD Graphics
3000, Windows, 1024x768, Siwa Oasis.
This was SVN commit r16241.
This reduces the total number of different vertex sizes in the system,
allowing more data to share a single CVertexBuffer, therefore reducing
the amount of wasted space in each CVertexBuffer and reducing VRAM
usage.
This was SVN commit r16230.
Nobody ever looks at the data, and it takes up a lot of server disk
space. The feature should probably be restored in the future once we've
fixed all the obvious performance problems, since a wider range of data
lets us search for system-dependent issues (e.g. some GL drivers are
particularly slow with particular features), but it's not worthwhile
now.
This was SVN commit r16213.
The implementation on Windows was gross. The only user was CLogger,
which no longer uses it.
Also fix vswprintf_s to handle truncated output correctly (by returning
"") on Linux, now that CLogger is no longer relying on the buggy
behaviour.
This was SVN commit r16190.
Everything is char* now, so we don't need to mess around with different
string types.
Done with:
ag -ls 'LOG(MESSAGE|MESSAGERENDER|WARNING|ERROR)' source | xargs perl
-pi -e'1 while
s/(LOG(MESSAGE|MESSAGERENDER|WARNING|ERROR).*)%[hl]s/$1%s/g'
This was SVN commit r16187.
This saves the hassle of writing utf8_from_wstring(path.string()) in
places like log messages, and can be extended to better handle
non-ISO-8859-1 paths on Linux.
This was SVN commit r16185.
Done with:
ag -ls 'LOG(MESSAGE|MESSAGERENDER|WARNING|ERROR)' source | xargs sed
-i 's/LOG\(MESSAGE\|MESSAGERENDER\|WARNING\|ERROR\)(L/LOG\1(/g'
This was SVN commit r16183.
sys_vswprintf relies on platform-specific printf implementations, which
vary widely between platforms (in handling of truncation, return values,
use of %s/%S/%hs/%ls for mixing char and wchar_t strings, etc) and are
therefore a pain.
Use cppformat's fmt::sprintf instead, which has very similar syntax to
sprintf but is more C++ish and is portable.
Also, wchar_t is stupid, so use char* strings (which are expected to be
UTF-8) in CLogger. This creates a bit of a pain with changing all
callers to convert to char* strings, but that's their fault for not
using UTF-8 already.
Refs #3011.
This was SVN commit r16182.
Throwing exception on NULL is a bit extreme, and unhelpful when it
happens in rarely-tested error paths. Printing "(null)" is safer and
provides compatibility with glibc sprintf.
This was SVN commit r16180.
These cause a lot of type-safety trouble - unsupported types passed into
fmt::sprintf (like CStr or enums) will be accepted at compile time, but
trigger an exception at runtime. Remove them, so we'll get either an
implicit conversion to a supported type, or a compile-time error.
This was SVN commit r16179.
This needs to use the model-view matrix, not model-view-projection (the
transform uniform), else the axes won't be unit vectors and the particle
sizes will be wrong. But GLES doesn't have the pre-defined matrices, so
pass it in explicitly.
This was SVN commit r16165.
Do some intersection tests on the CPU so that the silhouette render
passes only have to draw models/patches that might actually contribute
to silhouettes, saving the CPU and GPU cost of rendering more objects
than necessary.
This was SVN commit r15483.
Previously we had a single culling frustum based on the main camera, and
any object outside the frustum would never get rendered, even if it
should actually contribute to shadows or reflections/refractions. This
caused ugly pop-in effects in the shadows and reflections while
scrolling.
Extend the renderer to support multiple cull groups, each with a
separate frustum and with separate lists of submitted objects, so that
shadows and reflections will render the correctly culled sets of
objects.
Update the shadow map generation to compute the (hopefully) correct
bounds and matrices for this new scheme.
Include terrain patches in the shadow bounds, so hills can cast shadows
correctly.
Remove the code that tried to render objects slightly outside the camera
frustum in order to reduce the pop-in effect, since that was a
workaround for the lack of a proper fix.
Remove the model/patch filtering code, which was used to cull objects
that were in the normal camera frustum but should be excluded from
reflections/refractions, since that's redundant now too.
Inline DistanceToPlane to save a few hundred usecs per frame inside
CCmpUnitRenderer::RenderSubmit.
Fixes#504, #579.
This was SVN commit r15445.
Allow components to individually subscribe/unsubscribe to messages,
instead of statically subscribing the entire component type. Use this
for most Interpolate/RenderSubmit messages, to avoid the performance
cost of passing those messages to a large number of components that will
just ignore them anyway.
On Azure Coast this reduces total time per frame by about 30% on a
CPU-bound system.
This was SVN commit r15400.
Once 'delete' is called on an object, that object no longer exists, and
accessing its member variables is undefined behaviour. GCC 4.9's
optimiser recognises this, and eliminates any writes to member variables
inside the destructor, since it knows they cannot legally be read later.
BoundaryTagManager relied on ~FreedBlock resetting its memory to 0, so
this optimisation broke it. Replace the placement new/delete with plain
non-magic Setup/Reset functions, to avoid the optimisation.
Fixes#2481.
This was SVN commit r15334.
I guess changes to the map loading sequence caused the TerrainChanged
message to be sent before the map was switched from square to circular
instead of after. The pathfinder didn't notice the switch, so it
continued treating the map as if it were square, allowing units to walk
into the permanent map-corner SOD and vanish, and allowing territories
to expand into the SOD.
Tell the pathfinder explicitly when the map shape changes, so it can
discard its cached data correctly.
This was SVN commit r15277.
Instead of each CCmpVisualActor rendering itself individually, collect
all the units in a single CCmpUnitRenderer. This avoids the overhead of
doing Interpolate/RenderSubmit calls every frame for every object in the
world. It also allows more efficient culling.
CCmpUnitRenderer knows the positions of each object at the start and end
of each turn, and computes the bounding sphere of the object along that
path. That allows quick culling without recomputing the precise
interpolated transform every frame. (In the future it could be improved
much more.)
Clarify and clean up the sending of PositionChanged messages, and add
new InterpolatedPositionChanged.
Remove the forceFloating parameter from GetInterpolatedTransform, since
it doesn't fit the new design. Replace it with a (non-synchronised) flag
in CCmpPosition.
Move construction progress from CCmpVisualActor to CCmpPosition, so that
it consistently affects all position/transform computation.
Refs #2337.
This was SVN commit r15265.
Deserialize SYSTEM_ENTITY before any other entities. This makes it safe
for Deserialize() methods to access system components (mirroring how
Init() can already access system components).
Add a Deserialized message, sent after all entities have been
deserialized, to help with some other sequencing problems.
This was SVN commit r15264.
This allows component Init functions to make use of the terrain
immediately, instead of delaying some computations until the first
frame.
This was SVN commit r15262.
BoundingSphere is similar to a bounding box, but more spherical.
SOverlaySphere is useful for visualising BoundingSphere (it's quite
inefficient and should only be used for debug functionality).
Normalise the camera frustum clipping planes, so that IsSphereVisible
gives the correct result.
This was SVN commit r15261.
For unknown reasons, sometimes glTexSubImage2D acts as if
GL_UNPACK_ALIGNMENT = 2 or 4, instead of 1. This causes the odd-sized
array uploaded for the LOS texture to be interpreted incorrectly, and
the LOS texture gets rendered very incorrectly.
Pad the array to a multiple of 4 in all cases, so that
GL_UNPACK_ALIGNMENT shouldn't affect it.
Hopefully fixes#2594.
This was SVN commit r15216.
Fix EncryptPassword being called with the wrong argument order, and
encrypting username instead of password. (This will break all existing
lobby accounts.)
Fix EncryptPassword not using all of salt_base.
This was SVN commit r14123.
OglTex had a size_t added, which pushed it over HDATA_USER_SIZE.
Use u32 instead to minimise the difference between platforms.
This was SVN commit r14121.
SDL queues up all the input events received in a frame. When the hotkey
system saw a key up/down event, it immediately updated its
HotkeyIsPressed state and then pushed a hotkey event onto the end of the
queue.
If the initial queue was e.g. [key-down shift, key-press Z, key-up
shift], the hotkey event triggered by Z would be processed after the
key-up shift had updated the HotkeyIsPressed state, so the handler of
the Z hotkey would not think the shift hotkey was pressed.
If the initial queue was e.g. [key-press Z, mouse-click], the hotkey
triggered by Z would be processed after the mouse-click event, so it
could apply to the wrong building selection.
Fix by pushing the hotkey events onto a special queue that gets
processed before any subsequent SDL input events.
Also update the HotkeyIsPressed status when the HOTKEYDOWN/HOTKEYUP
events are processed, not when they are generated, to guarantee they are
consistent with the DOWN/UP events.
Fixes#1869.
This was SVN commit r14057.
Add optional clipping rectangle to CTextRenderer. Strings that are
printed
outside the vertical extent of the clipping rectangle will be
immediately
skipped. This greatly reduces the cost of large scrollable text boxes.
This was SVN commit r14019.
When CTextRenderer is given multiple strings with no differences other
than position, render them in a single glDrawElements call to reduce
driver overhead.
Also avoid some unnecessary copies of std::wstrings.
This helps performance a bit with large GUI text boxes.
This was SVN commit r14018.
Replace unifont with CFont and CFontManager, since the h_mgr interface
was
needlessly inconvenient.
Load the font textures through CTextureManager, to support dynamic
reloading (e.g. when resetting GL state - see #741).
Add CFontMetrics as a convenient wrapper for code that just wants to
measure text.
Fixes#1117.
This was SVN commit r14016.
When textures.xml specifies format="alpha", the input is expected to be
an
8-bit greyscale PNG, and the output will be an 8-bit uncompressed
alpha-only DDS.
Add format override to CTextureProperties, to select between e.g.
GL_ALPHA
and GL_LUMINANCE for 8-bit textures.
This is needed so fonts can use the new texture system.
This was SVN commit r14015.
This fixes an error with autostart and smoothlos=true, where the system
defines were being accessed before the first rendered frame.
This was SVN commit r13923.
Ykkrosh