This new implementation does all downsampling in a single compute shader
dispatch, removing a lot of complexity from the previous recursive
downsampling.
This is heavilly inspired by the Single-Pass-Downsampler from GPUOpen:
https://github.com/GPUOpen-Effects/FidelityFX-SPD
However I do not implement all the optimization bits as they require
vulkan (GL_KHR_shader_subgroup) and is not as versatile (it is only
for HiZ).
Timers inside renderdoc report ~0.4ms of saving on a 2048*1024 render for
the whole downsampling. Note that the previous implementation only
processed 6 mips where the new one processes 8 mips.
```
EEVEE ~1.0ms
EEVEE-Next ~0.6ms
```
Padding has been bumped to be of 128px for processing 8 mips.
A new debug option has been added (debug value 2) to validate the HiZ.
Compared to the previous implementation this has a limit of 65536 lights
per scene. Lights exceeding this limit will be ignored.
This also introduce fine grained GPU light culling, making rendering
many lights in a scene more efficient as long they don't overlap much.
Compatible light panels have been unhidden.
Note: This commit does not include surface evaluation, only light culling.
This is a port of the previous implementation but using compute
shaders instead of using the raster pipeline for every steps.
Only the scatter passes is kept as a raster pass for obvious performance
reasons.
Many steps have been rewritten to take advantage of LDS which allows faster
and simpler downsampling and filtering for some passes.
A new stabilize phase has been separated from another setup pass in order
to improve it in the future with better stabilization.
The scatter pass shaders and pipeline also changed. We now use indirect
drawcall to draw quads using triangle strips primitives. This reduces
fragment shader invocation count & overdraw compared to a bounding
triangle. This also reduces the amount of vertex shader invocation
drastically to the bare minimum instead of having always 3 verts per
4 pixels (for each ground).
The new implementation leverage compute shaders to reduce the
number of passes and complexity.
The max blur amount is now detected automatically, replacing the property
in the render panel by a simple checkbox.
The dilation algorithm has also been rewritten from scratch into a 1 pass
algorithm that does the dilation more efficiently and more precisely.
Some differences with the old implementation can be observed in areas with
complex motion.
The display depth is used to composite Gpencil and Overlays. For it to
be stable we bias it using the dFdx gradient functions. This makes
overlays like edit mode not flicker.
The previous approach to save the 1st center sample does not work anymore
since we jitter the projection matrix in a looping pattern when scene
is updated. So the center depth is only (almost) valid 1/8th of the times.
The biasing technique, even if not perfect, does the job of being stable.
This has a few cons:
- it makes the geometry below the ground plane unlike workbench engine.
- it makes overlays render over geometry at larger depth discontinuities.
This modules handles renderpasses allocation and filling. Also handles
blitting to viewport framebuffer and render result reading.
Changes against the old implementation:
- the filling of the renderpasses happens all at once requiring
only 1 geometry pass.
- The filtering is optimized with weights precomputed on CPU and
reuse of neighboor pixels.
- Only one accumulation buffer for renderpasses (no ping-pong).
- Accumulation happens in one pass for every passes using a single
dispatch or fullscreen triangle pass.
TAA and history reprojection is not yet implemented.
AOVs support is present but with a 16 AOV limit for now.
Cryptomatte is not yet implemented.
This module allow tracking of object and geometry data accross time.
This commit adds no user visible changes.
It work in both viewport (*) and render mode, gives correct motion
for any camera projection type and is compatible with displacement (**).
It is a huge improvement upon the old EEVEE velocity which was only used
for motion blur and only available in render.
It is also an improvement for speed as the animated objects do not need to
be rendered a 3rd time. The code is also much cleaner: no GPUVertBuf
duplication, no GPUBatch amendment, no special cases for different geometry
types, no DRWShadingGroup per object, no double buffering of velocity.
The module is still work in progress as the final output may still be
flawed.
(*): Viewport support is already working but there might be some cases where
mapping will fail. For instance if topology changes but not vertex count.
(**): Displacement does not contribute to motion vectors. Surfaces using
displacement will have the same motion vectors as if they were not displaced.
This commit introduce back support for all geometry types and all nodetree support.
Only the forward shading pipeline is implemented for now.
Vertex Displacement is automatically enabled for now.
Lighting & Shading is placeholder.
Related Task: T93220
# Conflicts:
# source/blender/draw/engines/eevee_next/eevee_engine.cc
# source/blender/gpu/CMakeLists.txt
