Just disable these tests on macOS for now as fixing seems hard, and we want to
be able to cross-compile and test x86_64 on Arm machines on the buildbot.
This is required for Cycles to report a meaningful error message when it fails to load a PTX module
created with a newer CUDA toolkit version than the driver supports.
Fix crash when kernel loading failed (T91879)
Ref T91879
This fixes a performance regression on Ampere cards, on specific scenes like
classroom. For cycles-x there is little difference, but this is still helpful
for LTS releases, and we need to upgrade at some point anyway.
Both before and after can have artifacts with some normal maps, but this seems to give
worse artifacts on average which are not worth the minor performance increase.
This reverts commit 21bc1a99ba.
Ref T88368, D10084
These seem to be causing some stability issues, and really are just not that
useful in practice. Compiling them is slow already, so it does not improve
the user experience much to show an AO preview if it's not nearly instant.
This is giving too bright pixel values, as the sample scaling and random number
sample are wrong. The proper fix for this is complicated. It will be solved in
Cycles X, for now we disable this combination.
Contributed by Intel. On some scenes like classroom with particular integrated
GPUs this speeds up rendering 1.97x. With other benchmarks and GPUs it's
between 0.99-1.14x.
This issue originates from a missing BVH packing for visibility data
when it is modified.
To fix this, this adds update flags to the managers to carry the
modified visibility information from the Objects' modified flag to the
GeometryManager.
Another set of flags is added to determine which data need to be packed:
geometry, vertices, or visibility. Those flags are then used when
packing the primivites.
Reviewed By: brecht
Maniphest Tasks: T87929
Differential Revision: https://developer.blender.org/D11219
Was causing calculation issues later on in the kernel.
This change catches the most obvious case: missing attribute. The old
code was trying to set tangent to 0, but because it was transformed as
a normal it got converted to non-finite value. This change makes it so
that no transform is involved and 0 is written directly to the SVM
stack.
To cover all cases it will require using safe_normalize() in this node
and in the normal transform function. This is more involved change from
performance point of view, would be nice to verify whether we really want
to go this route.
I've left asserts in the BSDF allocation functions. Don't have strong
connection to them, but think they are handy and are not different from
having an assert in the path radiance checks.
Differential Revision: https://developer.blender.org/D11235
It is possible that BSDF allocation will advance pointer in the
allocation "pool" but will return null pointer if the weight is
too small.
One artist-measurable issue this change fixes is random issues
with denoising: normal pass for denoising could have accessed
non-initialized normal of a closure.
Differential Revision: https://developer.blender.org/D11230
The render session is keeping track of the scene update, which includes
kernel loading time.
This fixes negative render times reported when CUDA kernels are compiled
at runtime.
A bit fragile logic, can be re-implemented using some user-counted
scope utility classes, so that only outer-most time skip is applied.
This is done to ensure building with newer OptiX SDK releases that add new struct fields gives
deterministic results (no uninitialized fields and therefore random data is passed to OptiX).
The problem was that you could getting write access to a grid from a
`const Volume *` without breaking const correctness. I encountered this
when working on support for volumes in the bounding box node. For
geometry nodes there is an important distinction between getting data
"for read" and "for write", with the former returning a `const` version
of the data.
Also, for volumes it was necessary to cast away const, since all of
the relevant functions in `volume.cc` didn't have const versions. This
patch adds `const` in these places, distinguising between "for read"
and "for write" versions of functions where necessary.
The downside is that loading and unloading in the global volume cache
needs const write-access to some member variables. I see that as an
inherent problem that comes up with caching that never has a beautiful
solution anyway.
Some of the const-ness could probably be propogated futher in EEVEE
code, but I'll leave that out, since there is another level of caching.
Differential Revision: https://developer.blender.org/D10916
For Cycles, when enabling the Persistent Data option, the full render data
will be preserved from frame-to-frame in animation renders and between
re-renders of the scene. This means that any modifier evaluation, BVH
building, OpenGL vertex buffer uploads, etc, can be done only once for
unchanged objects. This comes at an increased memory cost.
Previously there option was named Persistent Images and had a more limited
impact on render time and memory.
When using multiple view layers, only data from a single view layer is
preserved to keep memory usage somewhat under control. However objects
shared between view layers are preserved, and so this can speedup such
renders as well, even single frame renders.
For Eevee and Workbench this option is not available, however these engines
will now always reuse the depsgraph for animation and multiple view layers.
This can significantly speed up rendering.
These engines do not support sharing the depsgraph between re-renders, due
to technical issues regarding OpenGL contexts. Support for this could be added
if those are solved, see the code comments for details.
* Move out of Simplify panel, into Light Paths > Fast Global Illumination
* Add separate boolan setting to enable/disable it separate from Simplify
* Default AO bounces to 1
* Put ambient occlusion distance in this panel as well
