Instead of using same matrix calculations 3 times, getMatrixCenter and getMatrix now both call getMatrixFast.
Additional function call might be slight cost in debug, in release compilers hopefully inline it away.
Also getMatrix_transposed now split into getMatrix_transposed and getMatrixFast_transposed to make it similar to getMatrix.
This also avoids a bunch of level 4 warnings in VS about function variables hiding class variables, which was why I started on this. Thought I also prefer having less code here.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6587 dfc29bdd-3216-0410-991c-e03cc46cb475
Double tessellation numbers for these 3 functions in your code to get same results as in Irrlicht 1.8
I noticed in last commit those all tessellated semi-circles instead of circles. I just documented it there, but it's annoying as that bug doesn't allow odd numbered tessellation. So for example one couldn't create a triangle pyramid.
I considered adding another parameter to describe what this number means, but... just too ugly. And it was always documented otherwise, so this could be considered a bug. Sorry if this changes the look of existing apps.
Also cleaned up code in CGeometryCreator::createCylinderMesh. No idea what the idea behind old code was. Aside from creating twice the tessellation numbers.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6581 dfc29bdd-3216-0410-991c-e03cc46cb475
CTriangleSelector constructor with IAnimatedMeshSceneNode parameter didn't initialize MeshBuffer.
Not used internally, so no big problem, but it got passed back to users in some cases
Replaced SCollisionTriangleRange, by CTriangleSelector::SBufferTriangleRange. Was a bad idea to use a struct which is for something else just because some member were the same (my fault, sorry).
Also started unifying the case of using ranges vs not using them.
Can maybe even get rid of the non-ranges case in future. Or unify the duplicated code for getting triangles at least.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6579 dfc29bdd-3216-0410-991c-e03cc46cb475
If we already know we start with an identity matrix (which is usually the case) then copy will do.
Also unifying code to create matrix of different CTriangleSelector::getTriangles functions
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6578 dfc29bdd-3216-0410-991c-e03cc46cb475
So far it only used the inverse of the node transformation to calculate the box used to check
Which gave wrong results as soon as one tried to pass an additional matrix transformation
Wasn't ever used internally or by examples in Irrlicht, so I guess no one ever noticed (also in some cases this still worked accidentally).
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6577 dfc29bdd-3216-0410-991c-e03cc46cb475
Until last summer we sorted by object origin to camera distance
Since then we used nearest transformed bbox-extent to camera.
I've now added an enum to allow switching those plus 2 new:
- none (so sorting based on scenegraph instead)
- object center to camera. Which I made the new default as it worked the best in my tests.
I already experimented with a few more ones like different sphere sizes (bbox radius, minimal inbound radius, maximal inbound radius) around center or origin to handle objects with different sizes, but that just gave worse results for all my test cases.
Likely algorithms we should still try:
- Collision point with bounding-box in line between camera and object center (sounds a bit slow, but maybe worth it)
- Distance to camera plane (instead of camera position). But needs additional parameter to distance functions first (maybe normalized view vector will do). That should be useful when working with planar objects.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6572 dfc29bdd-3216-0410-991c-e03cc46cb475
A bit annoying that it kinda has the opposite meaning for png and jpg compression for same parameter (png compression goes up, jpg goes down).
Also unfortunate we chose u32 instead of int here or we could at least use the usual zlib value range for png.
But I think it still won't mess up in many cases. Defaults (value 0) stay the same as before.
And setting a jpg range <= 10 is rarely done and even if so it just changes png sizes a bit now if people use writer for both.
People just have to be careful now when they override defaults for png and then also write jpg - but can't help it.
And it's too useful to not have this - this can massively change the write-speed for png's (like up to 3 times faster with no compression on my system).
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6570 dfc29bdd-3216-0410-991c-e03cc46cb475
Irrlicht generally avoided user pointers in the past, but after trying all kind of ugly workarounds - this is just easier and
not that much downsides really. Tiny speed costs due to additional SMaterial memory size and new comparison tests.
But allows to keep SMaterial alive and useful for a while longer without needing a complete rewrite and it can now be used for stuff like writing PBR shaders.
Using a new interface io::IUserData for this which also allows serialization the user data (that part is untested so far, sorry)
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6567 dfc29bdd-3216-0410-991c-e03cc46cb475
LensShift got added recently.
Also there was some change in PolygonOffsetDirection longer ago (although that one didn't cause tests to fail)
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6566 dfc29bdd-3216-0410-991c-e03cc46cb475
Returns now getMipMapsData instead which for now will always be 0 in this case (but might work one day)
Also removing comment about ITexture::lock() mipmapLevel parameter being broken. It got fixed in November 2019.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6563 dfc29bdd-3216-0410-991c-e03cc46cb475
We had 4 near identical functions, those now all call buildProjectionMatrixPerspectiveFov
They were a bit hard to check for errors otherwise.
Especially with the tiny confusing non-differences like one using (a-b) and other -(b-a)
Also new one uses matrix template parameter in case someone needs for example a high-precision matrix.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6530 dfc29bdd-3216-0410-991c-e03cc46cb475
Did no longer render nodes which were rendered in SceneManager + outside SceneManager if they didn't have the per-buffer-rendering enabled.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6491 dfc29bdd-3216-0410-991c-e03cc46cb475
SMaterialLayers are now identical when both have identity matrices.
Before it didn't consider them identical when one layer had set the identity matrix explicitly and the other didn't.
Generally didn't matter, just caused very rarely some extra state switches in the drivers. And just as rarely had a cheaper comparison. Just seems more correct this way.
operator= no longer releases texture memory which was allocated at one point.
Unless explicitly requested such memory is now always released later in the destructor.
This can avoid quite a few memory allocations/released in the driver. Usually not a noticeable performance difference on most platforms. But it can help avoid memory fragmentation.
We instead use an extra bool now to tell if the texture memory is used. So slight increase in SMaterialLayer and SMaterial size. But I did a quick performance test and this had no negative influence here, while it did improve speed in the case where it switched between material layers using/not using texture matrices a bit.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6488 dfc29bdd-3216-0410-991c-e03cc46cb475
Wasn't ever used by anything and not that well defined anyway.
So they all just passed it on to the drivers. And then sometimes the driver version was called and sometimes the IMaterialRendererServices version. So now everything just calls the driver - all places which need it have access to the driver anyway. Also made the driver version non-virtual for now. If someone actually really needs this for some reason I can add it back as virtual function directly in IVideoDriver. But I doubt it - the interface was hardly accessible until recently and originally only meant for internal stuff.
GLES version still to do, but checked them earlier and they also just do nothing with it.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6486 dfc29bdd-3216-0410-991c-e03cc46cb475
So far SceneManager always sorted Nodes per render stage.
Now we allow sorting per mesh-buffer per render stage by creating a new node for each mesh-buffer.
It's only supported for CMeshSceneNode so far.
This allows to enable better transparency sorting for meshes which have transparent buffers.
Previously those always got rendered in the order in which they got added and ignored mesh-buffer bounding-boxes, but just used the bbox of the full mesh. Now they can use the bbox for each meshbuffer which can sometimes avoid render errors.
Also depending on the scene this can be quite a bit faster because it can help reduce texture changes. We sort solid nodes per texture, but only per node. So nodes with several textures had a texture switch between rendering each meshbuffer. And those are rather expensive in Irrlicht right now (and we support no bindless textures yet...)
Lastly it's now also used to buffer the render-stage. Checking this twice (once in registering the node and once in render) constantly showed up in the profiler. Which was a bit surprising really, but anyway - now it's gone.
I tried to keep it working for all cases we had before (all kind of situations, like when people may want to call render() outside the SceneManager). But not (yet) supporting the case of changing the meshbuffers (adding or removing some) without calling setMesh() again. Reason is that this wasn't well supported before either (node materials never updated). So for now I just assume people will call setMesh() again when they change the mesh.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6483 dfc29bdd-3216-0410-991c-e03cc46cb475
New AbsPosUpdateBehavior which makes updateAbsolutePosition calls behave as if a node had no parent.
Allows for micro optimizations in cases where we have non-moving root node (all scenenodes are always added to the SceneManager which is generally not moved but it's transformation is still multiplied each frame for each node)
As a side-effect this also allows abusing the SceneManager to group objects without affecting transformations.
No real extra cost since I added ESNUA_TRANSFORM_POSITION already.
Thought turns out those matrix transformations are so fast that I also didn't noticed any difference in tests with > 20.000 nodes.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6481 dfc29bdd-3216-0410-991c-e03cc46cb475
Accessing IMaterialRendererServices outside of OnSetConstants can be useful and Irrlicht made this a bit too hard to access.
Also OnCreate allows actually for nicer code where initialization and update of shader constants are strictly separated (see changed example).
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6465 dfc29bdd-3216-0410-991c-e03cc46cb475
Basically now the same as GLSL material renderer already worked.
Before it was using IMaterialRendererServices from CD3D9Driver, but there had been several problems with that:
- The d3d9 driver called functions through the CD3D9MaterialRenderer interface, but CD3D9HLSLMaterialRenderer is not (or maybe no longer?) derived from that class. Reason it still worked was accidental luck - the same functions had been in the same order and due to casts the compiler wasn't noticing it was calling the functions of an unrelated class.
- It was making calls to set shader constants depend on the currently set material in the driver. Which was not necessary and just prevents we can use the IMaterialRendererServices interface without setting the material first (how I found the bug).
Still some problems left for now:
- There's 2 ways to call shader constants. One seems to be only used by hi-level shaders which links constants to the shader. The the other only by low-level shaders which uses global (shader independent) registers.
So maybe this should be 2 interfaces? But not certain, glsl material renderer seems to prevent setting the global registers, but maybe those could be used additionally? I've still allowed it for now in HLSL, just in case it had it's uses.
- setBasicRenderStates probably should not be in IMaterialRendererServices. I'm not seeing any case where this isn't just passed on to the driver. And all classes using it have access to the driver unless I missed one. So probably can just avoid that additional indirection and kick that out of the IMaterialRendererServices interface.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6464 dfc29bdd-3216-0410-991c-e03cc46cb475
S3DVertexTangents also documents that it's passed as TEXCOORD1 and TEXCOORD2, but easy to miss, so added some comments to the hlsl shader.
Also not all materials work well as base materials.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6452 dfc29bdd-3216-0410-991c-e03cc46cb475
Bitfields for PolygonOffsetDirection, ZWriteEnable and BlendOperation were chosen too small.
As we have pre c++11 code and therefore didn't use unsigned qualifiers for enums they were generally signed (up to compiler in theory, but I think they all choose signed).
Which means the bitfield also had a sign.
So for example setting PolygonOffsetDirection to EPO_FRONT set it to -1 instead of 1.
Which then would fail with comparison checks (PolygonOffsetDirection == EPO_FRONT would be false).
We kind of got lucky that we usually not checked for the last enum inside Irrlicht, so it worked to due being the "else" case.
Or in the ZWriteEnable case the last one was identical to the default return value so it also worked accidentally.
But obviously still wrong and user code could be messed up.
While at it I also re-ordered SMaterial variable so most bitfield variables are close together again to give compiler at least a chance to use packing. Thought at least in my quick debug compile test it didn't seem to use any packing (but maybe on other compilers).
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6440 dfc29bdd-3216-0410-991c-e03cc46cb475
- Only the getRotationDegrees without parameter is allowed to try fixing scale.
My fault when I added a new function which takes scale parameter, that one is
not allowed to be changed.
On the up-side - we know have for the first time an option which works in cases only
scale and rotation had been used and the user still has the correct scale.
Before any solution for that was broken
- getRotationDegrees fixes 2 places which caused wrong results due to floating point inaccuracies
New test for that got added
- Document the current restrains and problems of getRotationDegrees and getScale some more.
- Improve docs for other matrix4 functions.
- Add some comments about further improvements (I'll try if I find time)
Note: Irrlicht still assumes in at least 2 places (getting bone animations and Collada loader) that matrix
decomposing works. Which it doesn't yet for matrices which switch handedness (or have further transformations like skewing axes)
The bone animation is mostly fine for now with recent workaround (but that might cause other problems as it may be used too often), haven't checked Collada yet in detail.
TL/DR: This improves things with getRotationDegrees, but does not yet fix all troubles.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6439 dfc29bdd-3216-0410-991c-e03cc46cb475
ISceneNode::setUpdateAbsolutePosBehavior can now control what ISceneNode::updateAbsolutePosition really does.
Having only the position and not the rotation/scale of a child node affected by the parent transformation was previously impossible inside the scene-graph. So people always had to break the scene-graph and code it themselves.
Old behaviour is default.
Extra check for new variable has a small cost, thought new behaviour can actually be faster when it's used.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6432 dfc29bdd-3216-0410-991c-e03cc46cb475
Previously search functions only worked when called with the same type as the array elements had.
Which forced users sometimes to create dummy objects to be able to search for elements by another type.
linear_search and linear_reverse_search now work with any type for which <T>::operator== is implemented.
Similar binary_search now works when <T>::operator< is implemented in both directions (T < E and E < T).
Note: It might be possible to further improve binary_search so only one operator< is needed (I think STL managed that somehow).
So if someone likes a challenge - have a go at it! :-)
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6398 dfc29bdd-3216-0410-991c-e03cc46cb475
Nothing good could come out of putting negative values into any of those functions.
And they are used a lot in image loaders which often can be tricked into passing large enough values to make those functions be called with negative numbers.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6389 dfc29bdd-3216-0410-991c-e03cc46cb475
It's not 100% identical to what we had as I only copied up to MATERIAL_MAX_TEXTURES_USED before, but I think that was rather premature optimiziation.
(the cost for the extra textures is really in other places, don't think the copy here ever matters).
Was a bit based on Mintest patch, but they messed it up: 4931b34625
(because they didn't notice the difference between MATERIAL_MAX_TEXTURES_USED and MATERIAL_MAX_TEXTURES).
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6363 dfc29bdd-3216-0410-991c-e03cc46cb475
Previously vertex buffer did some invalid casts to references which could cause it to copy whatever was in memory into the vertex-arrays.
Generally it worked up to S3DVertex - but switch from another type to S3DVertex2TCoords or S3DVertexTangents caused it to be filled with whatever was in memory behind it.
Setter functions in IVertexBuffer have now overloads for all known vertex types.
Also adding const version of IVertexBuffer::getData.
And some warnings in comments about using the array functions (if even we mess it up...)
Also de-deprecate IIndexBuffer::pointer() again. I don't like having 2 functions for same stuff, but in the end it doesn't really hurt keeping it around.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6360 dfc29bdd-3216-0410-991c-e03cc46cb475
Thanks @sfan5 for patch for Mintest: 4931b34625
Forum: https://irrlicht.sourceforge.io/forum/viewtopic.php?f=2&t=52819
Note - patch is not applied exaclty yet. SMaterial and matrix4 still missing (might be applied later).
Also patch was based on older Irrlicht and we removed a few already earlier this year (around r6280).
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6355 dfc29bdd-3216-0410-991c-e03cc46cb475
CMeshManipulator::createMeshCopy creates new meshes which have copies of the actual meshbuffers instead of copying everything into SMeshBuffers (which didn't support 32 bit or any of the other special features).
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6335 dfc29bdd-3216-0410-991c-e03cc46cb475
- IMeshLoader::setPreferredIndexType and getPreferredIndexType allow setting hints for the loaders if users prefer 16 or 32 bit meshbuffers. Loaders are free to ignore those hints (all but .obj will do that for now).
- obj meshloader loads now 32-bit buffers when setPreferredIndexType is set to EIT_32BIT.
NOTE: It's 16 bit meshes use now also an IDynamicMeshbuffer instead of an SMeshBuffer.
That will break the code of people who accessed meshbuffer before by casting to SMeshBuffer*
And might even be somewhat slower (lot's of virtual functions...), but shouldn't really matter and can maybe be a bit improved.
Sorry about that, I considered keeping SMeshBuffer for 16-bit (still considering it), but it would add some overhead in code and I don't think it's worth that. If there are any complains I'll maybe consider it again.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6333 dfc29bdd-3216-0410-991c-e03cc46cb475
Avoid creating dummy arrays when we work with non array images.
Just a minor speed improvement.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6329 dfc29bdd-3216-0410-991c-e03cc46cb475
Textures involved had all power of two sizes anyway, so no real difference in this case, just a bit cleaner.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6326 dfc29bdd-3216-0410-991c-e03cc46cb475
Drawing the screenshot itself before taking final result to see if the problem is in the render or the screenshot
Note: Test-images for software-drivers show some other problems there which this test wasn't about.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6322 dfc29bdd-3216-0410-991c-e03cc46cb475
Tests seem to work fine without -Xext and -Xcursor so kicking those out for now.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6321 dfc29bdd-3216-0410-991c-e03cc46cb475
That was a bad case of premature optimization.
Multiplication is indeed faster, but when working with floats this can introduce some rather unexpected inaccuracies.
Like x/x suddenly no longer being 1.0 (something guaranteed by division).
If someone really needs this back, then please add some new function which makes it clear we don't just have a typical division here.
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6298 dfc29bdd-3216-0410-991c-e03cc46cb475
Invisible difference, probably some mip-mapping pixel.
Getting rid of some unnecessary global in example 14
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6297 dfc29bdd-3216-0410-991c-e03cc46cb475
They all just implemented the same the default functions do.
This causes now warnings with newer gcc -Wdeprecated settings (otherwise they would have had to implement always both, but makes no sense as they did nothing special).
git-svn-id: svn://svn.code.sf.net/p/irrlicht/code/trunk@6280 dfc29bdd-3216-0410-991c-e03cc46cb475
