This is the root cause of broken updates
on local lights.
The same local frustum was used for all of the
tilemap (up to 6) of a light. This made the
`intersect(frustum, box)` call buggy which
would return true only if the object would
intersect with the first tilemap of the light.
This lead to improper updates when the object
would hit this path.
Fix#122533
Basically this tries to make the API to stop and kill jobs more explicit &
consistent, so intent is expressed clearly & behavior as expected.
- Remove use of the job start callback address as identifier for the job.
6887dea786 already removed this pattern from the jobs system internals, this
commit also removes it from the API.
- Make stop & kill API and implementation consistent. E.g. don't stop/kill jobs
by either owner **or** type/callback in one function, and by owner (if
provided) **and** type/callback in another. Causes some small behavior
changes, documented inline.
- Use the same job type and API for all preview render jobs (change by Brecht).
There doesn't seem to be a need for the separated types, in fact the
separation might have caused some issues earlier (and added code complexity).
- Add/improve function documentation.
This does actually have subtle behavior changes that are known, see PR, but
they were investigated carefully and seem like implementing wanted behavior.
Co-authored-by: Brecht Van Lommel <brecht@blender.org>
Pull Request: https://projects.blender.org/blender/blender/pulls/123086
This was caused by the tangent basis computation that
had a threshold that was too noticeable.
Increasing the threshold makes the artifact
unoticeable.
Fix#122949
This was caused by some pixels having no good neighbors
to get the irradiance from. The weighted sum would
have huge precision issues and the values would blow-out.
Adding an epsilon weight tailored to the report file
to fix this issue.
Fix#123488
Update the batch specializations compilation to allow using it in an
async way.
The implementation has 2 main limitations:
- Only one batch at a time can be processed, extra batches will be
added to a queue.
- Binding a specialization variant that is still being compiled will fail.
Pull Request: https://projects.blender.org/blender/blender/pulls/123015
EEVEE stores light probes using octahedral mapping. Compared to the previous
cubemap storage octahedral has less pixels. The 64x64 is becoming useless
and can be removed. This PR also enables generating light probe maps upto 4k.
Some issues were found: the offset of the sphere inside the atlas
was always set to mipmap level 0 offset. This was hidden because of the texture
wrapping. Also the offset was substracted from the local texture
coordinate when calculating the direction of the pixel. Might be that due
to the incorrect offset (mipmap level 0), the latter issue was never detected.
Pull Request: https://projects.blender.org/blender/blender/pulls/123074
EEVEE stores light probes using octahedral mapping. Compared to the previous
cubemap storage octahedral has less pixels. The 64x64 is becoming useless
and can be removed. This PR also enables generating light probe maps upto 4k.
Some issues were found: the offset of the sphere inside the atlas
was always set to mipmap level 0 offset. This was hidden because of the texture
wrapping. Also the offset was substracted from the local texture
coordinate when calculating the direction of the pixel. Might be that due
to the incorrect offset (mipmap level 0), the latter issue was never detected.
Pull Request: https://projects.blender.org/blender/blender/pulls/123074
The offset was only 0.5 which centered the gather
sample exactly on the first pixel of the quad.
With floating point arithmetic differences on Nvidia
this lead to the wrong set of texture pixel being
fetched by gather.
Using the coordinate at the center of the quad fixes
the issue.
Fix#123262
Add a `.data<T>()` method that retrieves a mutable span. This is useful
more and more as we change to filling in vertex buffer data arrays
directly, and compared to raw pointers it's safer too because of asserts
in debug builds.
Pull Request: https://projects.blender.org/blender/blender/pulls/123338
This was a missing block of the TAA implementation.
TAA jitter and reprojection have a tedency to soften
the texture. Add a 1.5 bias to make them a bit sharper.
Note that this is a bit different than the usual TAA
blurring. In final render we don't do reprojection
so it is only because the texture filter (box filter
from the LOD) is applied at the same time than our pixel
filter (blackmann-harris). It is less noticeable than
the normal TAA blur, but still blurs ~2px instead of
1.5px.
This implement the holdout flag by switching to
the holdout case in the shader. This has a few benefits:
- Doesn't recompile the shaders.
- Makes the object infos mandatory (already the case in
practice)
- Handle transparent materials properly, keeping the
transparency working.
Fix#123284
Pull Request: https://projects.blender.org/blender/blender/pulls/123315
This happened in the following render test:
`render/light/all_light_types_in_volume.blend`
Unfortunately it seems non-deterministic.
To fix this, we change the heuristic to jump
out of the shadow update loop. Also introduce
a upper bound to the number of iteration.
On top of this, add a flush for every loop
to avoid huge command buffer submission.
Thanks @pragma37 for the fix.
At high light count, this missing barriers would
produce invalid, non-unique `prefix_sum` indices.
This then resulted in some slots inside `out_light_buf`
never written to, leaving undefined data inside them.
If the buffer was cleared to zero, these undefined light
slots would be interpreted as sun lights and the
shadow setup compute pass would critically fail because
of out of bound memory.
Fix#123195
EEVEE Film accumulation workaround for Metal/Intel iGPUs.
On Metal the Intel iGPUs do not support image read write
on array textures. However this limitation doesn't show
any artifacts when using the compute shader.
This PR is a work around that uses the film_comp shader
to process the film samples, but uses a separate film_copy_frag
shader to read the result and copy them to the frame buffer.
I deliberately didn't include the fix to the film_frag shader
as that would change the read/write resources and could lead
to performance issues for other platforms. Writable resources
are typically slower compared to read only resources.
Some code needed to be duplicated (and not added to `*_lib.glsl`)
as compilers would still raise compilation errors due to imageStore/Load
on incompatible resource access.
The Metal/Intel iGPU is also marked to have limited support as
raytracing and probes still produces big artifacts.
This workaround can be tested on any platform just by setting
`use_compute_ = true` in `Film::sync`
Related to #122361
Pull Request: https://projects.blender.org/blender/blender/pulls/123330
