Second try after the revert in 8a28339 due to an unexpected regression.
- Rigidly animated models (e.g. the glTF frog node) were not working correctly,
since cloning the mesh ignored the transformation matrices.
Note that scaling the mesh needs to occur *after* transforming the vertices.
- Visual scale did not apply to skinned models,
as resetting the animation overwrote scaled vertex data with static positions & normals.
For backwards compatibility, we now apply a 10x scale to static, non-glTF models.
We now do scale static meshes, as the bug that caused meshes not to be scaled was limited to skeletally animated meshes,
hence we ought not to reproduce it for skinned meshes that do not take advantage of skeletal animations (e.g. current MTG doors).
However, glTF models (e.g. Wuzzy's eyeballs) up until recently were always affected due to technical reasons
(using skeletal animation for rigid animation).
Thus, to preserve behavior, we:
1. Do not apply 10x scale to glTF models.
2. Apply 10x scale to obj models.
3. Apply 10x scale to static x or b3d models, but not to animated ones.
See also: #16141
We don't have a C++ expert on hand, but taking a pointer to one member
and expecting to access another by an offset is very fishy:
- for one, there could theoretically be padding
- the compiler might assume that we are only writing to that first member
The new code has shown to be free for constant parameter values.
Non-constant ones cause the assembly to have branches (why?), but we don't
use that much.
Textures created through a render target are flipped along the X axis.
For the same reason, screenshots must be flipped back as well ('IVideoDriver::createScreenShot').
This commit implements the same flipping concept in two places:
1. Batch rendering of images (mostly used by fonts/text)
2. In-world rendering of such textures
