As some people have already noticed, the "Add" menu for nodes is a bit messy since pynodes merge. The reason for this is that the order of nodes in submenus (categories) was previously defined by the order in which all nodes are registered (at the bottom of blenkernel/intern/node.c). For the dynamic registration of node types now possible this system of defining node order along with registration is no longer viable: while it would still sort of work for C nodes, it is completely meaningless for dynamic (python) nodes, which are basically registered automatically in whatever order modules and addons are loaded, with the added complexity of unloading and reloading.
To fix this problem and add a bunch of desirable features this commit replaces the C menu with a python implementation. The new menu does not rely on any particular order of types in the node registry, but instead uses a simple explicit list of all the available nodes, grouped by categories (in scripts/nodeitems_builtins.py).
There are a number of additional features that become possible with this implementation:
1) Node Toolbar can be populated!
The list of nodes is used to create 2 UI items for each node: 1 entry in a submenu of "Add" menu and 1 item in a node toolbar panel with basically the same functionality. Clicking a button in the toolbar will add a new node of this type, just like selecting an item in the menu. The toolbar has the advantage of having collapsible panels for each category, so users can decide if they don't need certain nodes categories and have the rest more easily accessible.
2) Each node item is a true operator call.
The old Add menu is a pretty old piece of C code which doesn't even use proper operator buttons. Now there is a generic node_add operator which can be used very flexibly for adding any of the available nodes.
3) Node Items support additional settings.
Each "NodeItem" consists of the basic node type plus an optional list of initial settings that shall be applied to a new instance. This gives additional flexibility for creating variants of the same node or for defining preferred initial settings. E.g. it has been requested to disable previews for all nodes except inputs, this would be simple change in the py code and much less intrusive than in C.
4) Node items can be generated with a function.
A callback can be used in any category instead of the fixed list, which generates a set of items based on the context (much like dynamic enum items in bpy.props). Originally this was implemented for group nodes, because these nodes only make sense when linked to a node tree from the library data. This principle could come in handy for a number of other nodes, e.g. Image nodes could provide a similar list of node variants based on images in the library - no need to first add node, then select an image.
WARNING: pynodes scripters will have to rework their "draw_add_menu" callback in node tree types, this has been removed now! It was already pretty redundant, since one can add draw functions to the Add menu just like for any other menu. In the future i'd like to improve the categories system further so scripters can use it for custom node systems too, for now just make a draw callback and attach it to the Add menu.
If you have two windows, each with different scene, the render output for a window
would go to the other, if it was already drawing a render for the other scene.
Now you can have renders draw correct in two windows for two scenes.
The feature "Keep Session" was also loading that session when you double-click
on a .blend to open it, or when a .blend file was on commandline.
Moved this feature to the main() in creator.c, so it can check on it properly, skipping the
kept session when a file was loaded.
* Overlay invalidation did not happen when changing brush, .either
through UI or through shortcuts
* Add initiliazation of curves before threaded overlay texture
evaluation or we may get memory leaks due to race conditions.
* Fix precision overflow issue with overlay previews,
* Expose alpha mask mapping to UI (still not functional but coming soon).
* More overlay refactoring:
Overlay now does minimal checking for texture refresh.
Instead, we now have invalidation flags to set an aspect of the brush
overlay as invalid. This is necessary because this way we will be able to
separate and preview different brush attributes on the overlays, using
different textures:
These attributes/aspects are:
Primary texture (main texture for sculpt, vertex, imapaint)
Secondary texture (mask/alpha texture for imapaint)
Cursor texture (cursor texture. It involves brush strength and curves)
Modified the relevant RNA property update functions and C update callback
functions to call the relevant cursor invalidation functions instead
of checking every frame for multiple properties.
Properties that affect this are:
Image changes, if image is used by current brush,
Texture slot changes, similarly
Curve changes,
Object mode change invalidates the cursor
Paint tool change invalidates the cursor.
These changes give slightly more invalidation cases than simply
comparing the relevant properties each frame, but these do not occur in
performance critical moments and it's a much more elegant system than
adding more variables to check per frame each time we add something on
the system.
code was modifying pointers in the Mesh which gave crashes with another thread
accessing the data at the same time. This could crash other threaded operations
like blender internal render or physics baking too but was less likely.
As a solution I've now changed the save code that it does not modify the mesh
data structure in place but rather a copy, as undo file saving should probably
be fully read-only regardless of how an improved threading architecture might
work.
Thanks to Sergey for tracking down the cause of this crash.
This option replaces previously added GPU limit
option, which became tricky to follow after GLSL
display space conversion.
There're 4 modes available:
- AUTO which will try to guess which mode is
best to use.
Currently It'll try using GLSL and if it fails,
will fallback to 2D textures.
Probably it'll make sense checking on whether
2D textures works well but currently such behavior
shall be sufficient.
Later we could make this method smarter (for example
don't try to use GLSL on certain GPU or so).
- GLSL will currently behave the same way as AUTO,
but it is intended to always try using GLSL
(unless it can not be used because of existing
limitation of dither and RGB curves).
- 2D Textures will use CPU-based color space conversion
and use OGL 2D Texture to display the image.
Image will be displayed in tiles, so there shall be
no big GPU memory consumption.
- DrawPixels will straightly fallback to glDrawPixels
without trying to use any fancy GPU stuff.
Hopefully this will also fix
#34943: Blender crashes when resizing the Compositing Screen Window
now, with a 'resize' routine for the engine we can/should also recreate some
buffers that are created only at init time (e.g., 2d filters, dome fbos, ...).
This bug was always present in Blender (since 2.49 at least).
Bugfix supported by NF-UBC Nereus Program as part of the development
of OceanViz/NereusViz
Doing linearization with GLSL was already faster, but even faster is to just read the
bytes instead of floats and convert those to linear, since byte => float is just a quick
256 entry table lookup. Also made it assign the bytes directly to the image buffer so
they do not need to be converted back from float to byte for file saving, and made sky
render write the background color with OpenGL instead of doing it on the CPU.
Allocate float buffer outside of image buffer,
so work-in-progress color space conversion doesn't
interfere with other parts of blender.
Covers most of cases -- since image buffer wouldn't
have partially-update float buffer all the rest
areas would be happy.
However, if there're places which updates float
buffer from byte buffer, it's still possible
some WIP color space conversion is displayed on
the screen.
But what a heck someone will do such a crappy
conversion anyway!