The Realtime Compositor crashes on undo after an operation like Dissolve
node.
The compositor evaluator stored a reference to the compositor node tree
assuming that it will always be valid. This is not guaranteed, however,
and changes to the node tree can invalidate that reference. So we get
the node tree from the context directly every time to fix the crash.
This patch implements the Streaks Glare node. Which is an approximation
to the existing implementation in the CPU compositor. The difference due
to the approximation is bearily visible in artificial test cases, but is
less visible in actual use cases. Since the difference is rather similar
to that we discussed in the Simple Star mode, the decision to allow that
difference would probably hold here.
For the future, we can look into approximating this further using a
closed form IIR recursive filter with parallel interconnection and
block-based parallelism. That's because the streak filter is already
very similar to the causal pass of a fourth order recursive filter,
just with exponential steps.
Differential Revision: https://developer.blender.org/D16789
Reviewed By: Clement Foucault
This patch implements the variable size mode of the blur node. This is
not identical to the CPU implementation, but is visually very close.
That's because of two things. First, the Extend Bounds option introduces
a 2px offset that doesn't make sense, which is likely a bug in the CPU
implementation. Second, the CPU implementation approximate the result
using three passes, the first two of which are separable morphological
operators applied on the size input. But this approximation does not
provide an advantage because the last pass is non-separable anyways. So
the GPU implementation does not attempt this approximation for more
accurate and faster results.
Differential Revision: https://developer.blender.org/D16762
Reviews By: Clement Foucault
Compile each static shader using shaderc to Spir-V binaries.
The main goal is to make sure that the GLSL created using ShaderCreateInfo and able to compile to Spir-V.
For the second stage a correct pipeline needs to be created and some shader would need more
adjustments (push constants size).
With this patch future changes to GLSL sources can already be checked against vulkan, without the
backend finished.
Mechanism has been tested using MacOS and MoltenVK. For other OS, we should finetune CMake
files to find the right location to shaderc.
```
************************************************************
*** Build Mon 12 Dec 2022 11:08:07 CET
************************************************************
Shader Test compilation result: 463 / 463 passed (skipped 118 for compatibility reasons)
OpenGL backend shader compilation succeeded.
Shader Test compilation result: 529 / 529 passed (skipped 52 for compatibility reasons)
Vulkan backend shader compilation succeeded.
```
Reviewed By: fclem
Maniphest Tasks: T102760
Differential Revision: https://developer.blender.org/D16610
This patch implements the Simple Star Glare node. This is only an approximation
of the existing implementation in the CPU compositor, an approximation that
removes the row-column dependency in the original algorithm, yielding an order
of magnitude faster computations. The difference due to the approximation is
readily visible in artificial test cases, but is less visible in actual use
cases, so it was agreed that this approximation is worthwhile.
For the future, we can look into approximating this further using a closed form
IIR recursive filter with parallel interconnection and block-based parallelism.
Which is expected to yield another order of magnitude faster computations.
The different passes can potentially be combined into a single shader with some
preprocessor tricks, but doing that complicated that code in a way that makes
it difficult to experiment with future optimizations, so I decided to leave it
as is for now.
Differential Revision: https://developer.blender.org/D16724
Reviewed By: Clement Foucault
This patch implements the Ghost Glare node. It is implemented using
direct convolution as opposed to a recursive one, which produces
slightly different results---more accurate ones, however, since the
ghosts are attenuated where it matters, the difference is barely
visible and is acceptable as far as I can tell.
A possible performance improvement is to implement all passes in a
single shader dispatch, where an array of all scales and color
modulators is computed recursively on the host then used in the shader
to add all ghosts, avoiding usage of global memory and unnecessary
copies. This optimization will be implemented separately.
Differential Revision: https://developer.blender.org/D16641
Reviewed By: Clement Foucault
Blender crashes when enabling Use Nodes after the viewport compositor is
already enabled.
This happens because the active viewer key is not yet initialized for the
node tree at this point, which eventually leads to a nullptr.
This patch fixes that by returning the root context in case the active
viewer key is not yet initialized.
Using output nodes inside node groups in compositor node trees doesn't
work for the realtime compositor.
Currently, the realtime compositor only considers top level output
nodes. That means if a user edits a node group and adds an output node
in the group, the output node outside of the node group will still be
used, which breaks the temporary viewers workflow where users debug
results inside a node group.
This patch fixes that by first considering the output nodes in the
active context, then consider the root context as a fallback. This is
mostly consistent with the CPU compositor, but the realtime compositor
allow viewing node group output nodes even if no output nodes exist at
the top level context.
Differential Revision: https://developer.blender.org/D16446
Reviewed By: Clement Foucault
When using two transformed compositor results, the transformation of one
of them is apparently in the local space of the other, while it should
be applied in the global space instead.
In order to realize a compositor result on a certain operation domain,
the domain of the result is projected on the operation domain and later
realized. This is done by multiplying by the inverse of the operation
domain. However, the order of multiplication was inverted, so the
transformation was applied in the local space of the operation domain.
This patch fixes that by inverting the order of multiplication in domain
realization.
This patch introduces the concept of a Cached Resource that can be
cached across compositor evaluations as well as used by multiple
operations in the same evaluation. Additionally, this patch implements a
new structure for the realtime compositor, the Static Cache Manager,
that manages all the cached resources and deletes them when they are no
longer needed.
This improves responsiveness while adjusting compositor node trees and
also conserves memory usage.
Differential Revision: https://developer.blender.org/D16357
Reviewed By: Clement Foucault
This patch moves the GLSL shaders and their infos to the compositor
module as decided by the EEVEE & Viewport module. This is a non
functional change.
Differential Revision: https://developer.blender.org/D16360
Reviewed By: Clement Foucault
Currently, the realtime compositor treat vector types as 3D vectors,
which is true for most operations. However, some operations deal with
vector types as 4D vectors that encode two 2D vectors or one 3D vector
with the last component ignored. So this patch expands vector types to
include a fourth component that is only sometimes used.
Since we already stored vectors in RGBA textures, the necessary changes
are straightforward and are mostly concerned with adjusting the Result
class.
Differential Revision: https://developer.blender.org/D16359
Reviewed By: Clement Foucault
This patch implements the normalize node for the realtime compositor.
Differential Revision: https://developer.blender.org/D16279
Reviewed By: Clement Foucault
This patch implements the tone map node for the realtime compositor
based on the two papers:
Reinhard, Erik, et al. "Photographic tone reproduction for digital
images." Proceedings of the 29th annual conference on Computer graphics
and interactive techniques. 2002.
Reinhard, Erik, and Kate Devlin. "Dynamic range reduction inspired by
photoreceptor physiology." IEEE transactions on visualization and
computer graphics 11.1 (2005): 13-24.
The original implementation should be revisited later due to apparent
incompatibilities with the reference papers, which makes the operation
less useful.
Differential Revision: https://developer.blender.org/D16306
Reviewed By: Clement Foucault
Currently, the scale node always changes the interpolation of its result
to bilinear. This was done because the scale node does not have an
interpolation option, unlike the Transform node, so a default of
bilinear was assumed. This turned out to be problematic, because in the
pixelation use cases, a nearest interpolation is typically preferred by
the user.
This patch changes the default interpolation of input nodes to bilinear,
makes the scale node keep the interpolation of the input it receives,
and makes the pixelate node changes the interpolation to nearest. In
effect, for non-pixelation use cases, the default bilinear interpolation
will be used, and for pixelation use cases, the nearest interpolation
will be used unless explicitly specified using a node that sets the
interpolation.
The parallel reduction file didn't include its own header, which can
yield "no previous declaration" warnings. This patch includes the header
to fix the warning.
This patch implements generic parallel reduction for the realtime
compositor and implements the Levels operation as an example. This patch
also introduces the notion of a "Compositor Algorithm", which is a
reusable operation that can be used to construct other operations.
Differential Revision: https://developer.blender.org/D16184
Reviewed By: Clement Foucault
This is the conventional way of dealing with unused arguments in C++,
since it works on all compilers.
Regex find and replace: `UNUSED\((\w+)\)` -> `/*$1*/`
This patch adds support for the skip realization option of the input
descriptor. Inputs that request skip realization will not be realized on
the operation domain of the operation and will not contribute to its
computation, and consequently, they can't be a domain input.
An example is the bokeh input of the Bokeh Blur node, which is actually
a resource that is decoupled from the rest of the inputs and should not
affect or be affected by their domain.
Differential Revision: https://developer.blender.org/D15767
Reviewed By: Clement Foucault
Push the const usage a bit further for compositor nodes, so that they
are more explicit about not modifying original nodes from the editor.
Differential Revision: https://developer.blender.org/D15822
The purpose of `NodeTreeRef` was to speed up various queries on a read-only
`bNodeTree`. Not that we have runtime data in nodes and sockets, we can also
store the result of some queries there. This has some benefits:
* No need for a read-only separate node tree data structure which increased
complexity.
* Makes it easier to reuse cached queries in more parts of Blender that can
benefit from it.
A downside is that we loose some type safety that we got by having different
types for input and output sockets, as well as internal and non-internal links.
This patch also refactors `DerivedNodeTree` so that it does not use
`NodeTreeRef` anymore, but uses `bNodeTree` directly instead.
To provide a convenient API (that is also close to what `NodeTreeRef` has), a
new approach is implemented: `bNodeTree`, `bNode`, `bNodeSocket` and `bNodeLink`
now have C++ methods declared in `DNA_node_types.h` which are implemented in
`BKE_node_runtime.hh`. To make this work, `makesdna` now skips c++ sections when
parsing dna header files.
No user visible changes are expected.
Differential Revision: https://developer.blender.org/D15491
The realtime compositor crashes when some nodes are unlinked.
This happens for GPU material nodes if it was compiled into its own
shader operation. Since it is unlinked, the shader operation will have
no inputs, a case that the current code didn't consider.
This patch fixes this by skipping code generation for inputs if no
inputs exist for the shader operation.
Fix a number of warnings reported by Clang Tidy in the realtime
compositor's code.
Differential Revision: https://developer.blender.org/D15654
Reviewed By: Clement Foucault
This patch adds a stub implementation for all unsupported nodes. The
inputs are passed through to the outputs where it make sense, while
other outputs will be allocated a single zero value.
This seems to be preferred by users as opposed to stopping execution and
displaying an error message.
Differential Revision: https://developer.blender.org/D15464
Reviewed By: Clement Foucault
This patch adds the core realtime compositor evaluator as well as a
compositor draw engine powered by the evaluator that operates in the
viewport. The realtime compositor is a new GPU accelerated compositor
that will be used to power the viewport compositor imminently as well as
the existing compositor in the future.
This patch only adds the evaluator and engine as an experimental
feature, the implementation of the nodes themselves will be committed
separately.
See T99210.
Differential Revision: https://developer.blender.org/D15206
Reviewed By: Clement Foucault
