Discovered with valgrind. Previously, some memory that later gets freed using
free(...) (in StdBuf::Clear) was allocated with new char[...] rather than
malloc, in which case delete[] should be used to free them.
The red color channel calculation could overflow into the sign bit,
which is undefined behavior. At least one compiler takes advantage of
this and assumes it cannot happen, resulting in incorrect results.
BltAlphaAdd looks similar, but does in fact not have this bug because it
shifts the color channel far enough that multiplication can't overflow.
This is required for glDrawElements() with a core profile. Furthermore, for
meshes that have a fixed face ordering (non-transparent meshes at 100%
completion), this puts more static data on the GPU that we don't have to
upload every frame.
For meshes with non-fixed face ordering, we still have to upload faces every
frame, since the ordering might change if the mesh rotates. We could still
improve this by figuring out if the order actually changed after the sort
step, and only update to the GPU if it has. In many cases it hasn't, so there
is some potential for more optimization there. But that's for later.
Since mesh instances can now be attached to attached definitions, attached
definitions need to denumerate pointers recursively to make sure that the
reference to the mesh instance is restored correctly.
Changing the mesh material to something that does not use alpha blending
restores the originial face order. This caused a crash for meshes with
Completion below 1, since all original faces were copied, but the buffer
in the mesh instance was only large enough to hold the number of faces of
the incompleted mesh. To fix this, simply run the standard
select-faces-for-a-given-completion-value procedure.
Replace the hardcoded VAI_ constants with that. The VAI constants are now
moved to C4DrawGL as C4SSA_ constants, similar to the C4SSU ones. This allows
to introduce other attributes to replace vertex positions, normals, colors
and texture coordinates with attributes in later commits. This, in turn, is
needed because the built-in attributes are no longer available in the OpenGL
core profile.
Also, while at it, cleanup C4Shader a bit. Use std::vector<> instead of
maintaining an own array of uniform names. Delete the vertex and fragment
objects after the full shader program has been linked. Make sure that
C4Shader::Init keeps the old program in place if the new one cannot be
compiled or linked.
The only use of C4RTF in its final moments was parsing out plain text
from RTF files anyway, so why even go to all the trouble instead of just
storing plain text in the beginning?
Instead, compute the projection, modelview and normal matrices explictly
and upload them as shader uniforms. This is one step towards using the
OpenGL 3 core profile.
The C++ standard library comes with perfectly fine implementations of
these functions, so there's no point in reimplementing them just for the
hell of it.
Using memset to initialize non-POD types doesn't work. Or rather, it may
work right now, but will fail when somebody adds a member that relies on
its constructor doing something (like for example any STL container).
Either way it's undefined behavior and needs to go. Furthermore, using
it to reinitialize an object also prevents any dtors from doing their
work when needed.
A new helper function InplaceReconstruct will take an object of nothrow-
default-constructible type, and call the dtor to properly clean up
before placement-new reconstructing the object in the same location.
This is still bad design, but unfortunately removing the Default/Clear
functions from every object currently using them is a herculean task.
This implements the proposal made in the forum for "shiny" materials -
material can now determine the angle at which the most light is reflected.
Shiny materials might set this lower to approximate a "reflection" effect,
and increase the "spottiness" at the same time. To compensate for the
lack of brightness without light, "emittance" can be used.
Not sure this is the most elegant way to model this - the "proper" way
here would be to have emittance, shading and specular as three separate
light parameters instead of molding one into the other and using the third
to compensate.
Furthermore, this reorganises shaders in a major way: We reduce the
number of shader files down to three, pushing a number of possible
configurations into preprocessor. I believe this should be easier to
understand, which for the moment trumps theoretical extensibility
benefits.
We already require support for std::unique_ptr, which itself requires
support for rvalue references. As such, we know we can use rvalue
references, and thus don't have to keep carrying dead code around.
In comparison to the old system, this is a downgrade - instead of being
able to set a full color mapping by gamma ramp, we now get just a value
per colour channel.
Upside is that we do not need to play around with the global gamma ramps
any more, which was arguably the wrong way to do it.
This commit will likely break everything that has been using gamma so far.
Instead of doing the transformation when drawing a mesh. This allows making
the OpenGL normal matrix more consistent, since it does not include the
Ogre-To-Clonk transformation, and so that the transformation does not need
to be inverted in the shader.
As a side effect, all Attach transformations were updated, since before
they were specified in the OGRE reference frame, not the Clonk reference
frame.
Several rendering changes have resulted in a non-rendering build that
failed to build from source. Dummy out all of these functions to make it
work again.
On success, flock() returns 0, so !flock() is true then. Testing for
!flock() to detect a locking error is therefore wrong. This got the log
file closed immediately after opening it on my GNU/Linux machine,
resulting in an always empty OpenClonkShader.log.
Deserializing proplists tries to read a boolean variable from the
representation in order to be backwards compatible to old (<6.0)
savegames. Newer savegames do not write this boolean, so StdCompiler
(correctly) throws an exception. This exception uses line numbers, which
StdCompiler used to obtain by scanning through the entire buffer. Having
to scan through the entire buffer for every serialized proplist is
ridiculously slow (and the data isn't used anyway in this instance).
Since line information is still valuable outside of proplist
deserialization, replace the repeated SGetLine calls with an on-demand
cache.
This reduces load times of a 14 MB Game.txt from slightly over 3 minutes
to 7 seconds on my computer.
As long as we're not actually using a different shader for meshes
without bones, we need to upload an identity matrix so there's defined
data in the bone slot.
Instead of transforming all vertices on the CPU every time an animation
progresses, we now only recalculate the skeleton, leaving the heavy
lifting for the GPU. This also means we no longer have to push all
vertices onto the bus every frame, because the mesh isn't changing and
can therefore be stored in a GL_STATIC_DRAW VBO when it's first loaded.
The downside of this approach is that there's only a limited number of
uniforms and vertex attributes we can pass to the shader. At the moment
these limits are a maximum of 128 bones per skeleton, and no vertex can
be influenced by more than 8 bones at once. So far this is no problem,
as the most complex skeleton in the base game uses less than 64 bones
and no more than 6 bone weights per vertex.
While we're still not doing skinning on the GPU, copying the vertex data
to a VBO immediately after updating the animation allows us to re-use
that data for unanimated meshes. It also allows us to store unanimated
data on the GPU, instead of transferring it over the bus for each frame.
In applications targeting the GUI subsystem, stdout and stderr aren't
valid file handles, and trying to write to them sets the system error
code to ERROR_INVALID_HANDLE.
If vsnprintf fails for other reasons than too small a buffer, it returns
a negative value. Comparing that with a size_t value promotes the
negative value to unsigned, which makes for a very large result, almost
guaranteed to be larger than the buffer size.
We're trying to ensure that no caller passes a char* to ssprintf, but a
char array instead. This is way easier done by using template parameter
deduction than a specialized class template.
While none of the mismatches were having a side-effect, this silences a
number of -Wreorder warnings which were drowning out potentially
important diagnostics.
This should improve cache coherency by having all surface tiles adjacent
instead of strewn across the heap. This will also remove an indirection
in the common case of only using one tile.
Appending and including skeletons works now. Documentation updated.
Skins such as appendto.Clonk.Farmer.skeleton are handled as if the skin name were not included: appendto.Clonk.skeleton
This also changes a bit the structure of how meshes are reloaded in general,
in particular mesh materials are only reloaded once and not once per
C4DefGraphics instance. (Fix crash on reload with included skeletons, cherry picked from commit 2f69aa4850ab7a4b621e5f30cbc537d40f32c0df)
Fix linking of C4Script standalone tool (cherry picked from commit b03a332b73463b378c94e4e92b66d45b923b3b9c)
Lukas Werling