This patch unifies the Defocus node between the CPU and GPU compositors.
Both nodes now use a variable sized bokeh kernel which is always odd
sized for a center pixel guarantee. Further the CPU implementation now
properly handles half pixel offsets when doing interpolation, and always
sets the threshold to zero similar to the GPU implementation.
This patch unifies the CPU and GPU implementation for the Bokeh Image
node. The bokeh is now evaluated at the center of pixels and allows
different sizes for the output kernel.
The Sun Beams node is off by half a pixel. That's because we add a half
pixel offset to the initial coordinates, but then sample the texture
with the half pixel. To fix thus, use the texture_bilinear_extend
utility to sample the image, which takes care of the half pixel offset.
The Lens Distortion node is off by half a pixel because their normalized
coordinates were at the pixel corners as opposed to their centers, where
this patch changes the behavior to the latter.
The Plane Track and Corner Pin nodes are off by half a pixel because
their mask is computed at the pixel corners as opposed to their center,
where this patch changes the behavior to the latter.
The Directional Blur node is off by half a pixel because it transforms
the pixels at their corner as opposed to their center, where this patch
changes the behavior to the latter.
The Movie Distort node is off by half a pixel because it evaluates the
distortion at the corner of pixels as opposed to their center, where
this patch changes the behavior to the latter.
The compositor does not correctly free, wrap, or allocate images. That's
because the resetting mechanism didn't reset all members. This patch
rests all members and only retains the important ones.
This patch refactors the backdrop offset to be stored as a float instead
of an int and to be stored in the image runtime structure instead of the
image itself.
Pull Request: https://projects.blender.org/blender/blender/pulls/119877
In the tiled compositor ensure_delta() can be called from multiple threads,
but without any threading synchronization. This worked fine when the node
only supported absolute transform: multiple threads would do the same work
and assign delta to the same values.
With the addition of relative transform in #115947 a code which adjusts
previously calculated delta was added, leading to possible double-applying
relative transform.
The solution is to avoid multiple threads modifying the same data by using
a double-locked check.
This issue does not happen in 4.2 (main branch) because it switched to full
frame compositor, which works differently.
Pull Request: https://projects.blender.org/blender/blender/pulls/119883
This patch ports the GLSL SMAA library to the CPU compositor in order to
unify the anti-aliasing behavior between the CPU and GPU compositor.
Additionally, the SMAA texture generator was removed since it is now
unused.
Previously, we used an external C++ library for SMAA anti-aliasing,
which is itself a port of the GLSL SMAA library. However, the code
structure and results of the library were different, which made it quite
difficult to match results between CPU and GPU, hence the decision to
port the library ourselves.
The port was performed through a complete copy of the library to C++,
retaining the same function and variable names, even if they are
different from Blender's naming conversions. The necessary code changes
were done to make it work in C++, including manually doing swizzling
which changes the code structure a bit.
Even after porting the library, there were still major differences
between CPU and GPU, due to different arithmetic precision. To fix this
some of the bilinear samplers used in branches and selections were
carefully changed to use point samplers to avoid discontinuities around
branches, also resulting in a nice performance improvement. Some slight
differences still exist due to different bilinear interpolation, but
they shall be looked into later once we have a baseline implementation.
The new implementation is slower than the existing implementation, most
likely due to the liberal use of bilinear interpolation, since it is
quite cheap on GPUs and the code even does more work to use bilinear
interpolation to avoid multiple texture fetches, except this causes a
slow down on CPUs. Some of those were alleviated as mentioned in the
previous section, but we can probably look into optimizing it further.
Pull Request: https://projects.blender.org/blender/blender/pulls/119414
This patch unifies the sRGB to Linear color space conversion between the
CPU and GPU compositors. This is because CPU uses an optimized path that
produces values that are very slightly off. To fix this, for the GPU, we
do the conversion CPU side instead of doing it in a shader. Since images
are cached, the performance implications are not significant.
Another added benefit is that we no longer get differences due to the
order of alpha pre-multiplication and sRGB conversion, demonstrated in
#114305. And we no longer require any preprocessing of the images.
This patch adds some new utilities to the Image Buffer module to assign
float, byte, and compressed buffers along with their color spaces. It
also adds an ownership flag to compressed data. Those were added as a
way to facilitate the implementation.
Pull Request: https://projects.blender.org/blender/blender/pulls/118624
This patch adds a utility function for nearest interpolation with
clamped boundaries and normalized coordinates to the MemoryBuffer class.
Similar to the GLSL texture() function.
This patch adds clamped boundaries variants of the nearest interpolation
functions in the BLI module. The naming convention used by the bilinear
functions were followed.
Needed by #119414.
Pull Request: https://projects.blender.org/blender/blender/pulls/119732
Previously, all information outside the render area did not get considered because transform operations such as translation and rotation might cause the output area to be smaller.
In this patch, the whole image gets flipped regardless of the output area, which makes the behavior consistent with the GPU compositor.
Pull Request: https://projects.blender.org/blender/blender/pulls/119278
Rendering with F12 causes some inconsistencies between canvases and areas of interest. This is due to some corrections of composite node canvas to fit to the render size, which are not necessary for the Viewer Node. This patch fixes the issue by considering the input image as a whole before translating.
Note: this is still consistent with immediate realization of transform nodes in GPU compositor, where nodes are to be evaluated from left to right.
Pull Request: https://projects.blender.org/blender/blender/pulls/119276
GPU compositor crashes when the Hue Correct and one of the Curves node
are used together. That's due to a utility function that goes by the
same name in both nodes. To fix this, rename one of them to be more
specific.
This patch implements the GPU Bloom glare for the CPU compositor, and
adds a new option for it, leaving the Fog Glow option unimplemented once
again for the GPU compositor.
Pull Request: https://projects.blender.org/blender/blender/pulls/119128
The Realtime Compositor crashes with intricate node setups. That's
due to hardware limitations where shaders can't have as many output
images as needed by the compositor shader operations. To fix this, we
recursively split shader operations until their output count fits the
hardware limitation.
Pull Request: https://projects.blender.org/blender/blender/pulls/119075
Compositor outputs are sometimes clipped. That's because consecutive
shader operations in the GPU compositor were assumed to have the same
domain. So make sure to reset the domain during compile unit resets.
Blender crashes when using the compositor debugger. That's because it
assumes the names of the operation classes will be demangled and will
proceed to slice them causing crashes. The standard makes no guarantees
about the form of the returned type_info::name, and it is implementation
defined.
While it is possible to demangle those names and return something that
is more human readable, it doesn't seem worth it as this is only used by
developers who should have no problem identifying mangled names. So we
just display the name as is.
Pull Request: https://projects.blender.org/blender/blender/pulls/118904
This patch unifies the anti-aliasing of plane deforms between the CPU
and GPU compositors. The CPU used a multi-sample approach, where the
mask was computed 8 times with a jitter, then averaged to get smooth
edges. The GPU relied on the anisotropic filtering with zero boundaries
to smooth the edges.
Furthermore, the CPU implementation ignored the anti-aliasing for the
deformed image and also relied anisotropic filtering like the GPU, so
its outputs were different.
To unify both implementation, we use the existing SMAA anti-aliasing
algorithm instead, and use the anti-aliased mask for the image output as
well. This affects both the Corner Pin and Plane Deform nodes.
A consequence of this change for the Plane Deform node is that motion
blur will appear to have less samples, that's because it was sampled
8-times more in the previous implementation. But users can just increase
the samples in the node to account for that.
Pull Request: https://projects.blender.org/blender/blender/pulls/118853
This patch unifies the variable size blur between the CPU and GPU
compositor. The difference is due to how weights are computed and used.
The CPU computed a nested array of weights for every possible size, that
is, from size 1 to the base size. Then, it assumed the kernel was
separable and reconstructed a 2D kernel by selecting two 1D weights
array and multiplying them for every pixel of the blur window.
The GPU on the other hand computes a single quadrant of the 2D weights
kernel and sampled it directly in the blur window. We favor the GPU
implementation since it makes no assumptions about the separability of
the weights kernel and since the CPU has no performance advantage even
with the assumption in place.
Pull Request: https://projects.blender.org/blender/blender/pulls/118834
Currently, the CPU compositor smoothes its input size in variable size
mode. It is unclear why this is the case, but it seems the logic is that
sharp transitions in the size input are undesirable. Alternatively, this
is similar to the morphological blurring step in the Defocus node. But
it does not use standard weights and it is not morphological in nature
at all.
This patch removes the smoothing step and uses the original size
provided by the user. Looking at resources online, it seems users almost
always expect the size inputs to be used directly, so there is no reason
for force smooth their inputs.
Pull Request: https://projects.blender.org/blender/blender/pulls/118757
New ("fullframe") CPU compositor backend is being used now, and all the code
related to "tiled" CPU compositor is just never used anymore. The new backend
is faster, uses less memory, better matches GPU compositor, etc.
TL;DR: 20 thousand lines of code gone.
This commit:
- Removes various bits and pieces related to "tiled" compositor (execution
groups, one-pixel-at-a-time node processing, read/write buffer operations
related to node execution groups).
- "GPU" (OpenCL) execution device, that was only used by several nodes of
the tiled compositor.
- With that, remove CLEW external library too, since nothing within Blender
uses OpenCL directly anymore.
Pull Request: https://projects.blender.org/blender/blender/pulls/118819
