This change affects CUDA GPUs not connected to a display or connected to a
display but supporting compute preemption so that the display does not
freeze. I couldn't find an official list, but compute preemption seems to be
only supported with GTX 1070+ and Linux (not GTX 1060- or Windows).
This helps improve small tile rendering performance further if there are
sufficient samples x number of pixels in a single tile to keep the GPU busy.
Best guess is that cuInit() somehow interferes with the AMD graphics driver
on Windows, and switching the initialization order to do OpenCL first seems
to solve the issue.
* Use common TextureInfo struct for all devices, except CUDA fermi.
* Move image sampling code to kernels/*/kernel_*_image.h files.
* Use arrays for data textures on Fermi too, so device_vector<Struct> works.
Two issues here:
- Checking table size to be non-zero is not a proper way to go here. This is
because we first resize the table and then fill it in. So it was possible that
non-initialized table was used.
Trickery with using temporary memory and then doing table.swap() might work,
but we can not guarantee that table size will be set after the data pointer.
- Mutex guard was useless, because every thread was using own mutex. Need to
make mutex guard static so all threads are using same mutex.
The issue was caused by light sample being evaluated to nan at some point.
This is root of the cause which is to be fixed, but is very hard to trace down
especially via ssh (the issue only happens on AVX2 release build). Will give it
a closer look when back to my AVX2 machine.
For until then this is a good check to have anyway, it corresponds to what's
happening in regular radiance sum.
The work size is still very conservative, and this doesn't help for progressive
refine. For that we will need to render multiple tiles at the same time. But this
should already help for denoising renders that require too much memory with big
tiles, and just generally soften the performance dropoff with small tiles.
Differential Revision: https://developer.blender.org/D2856
This was originally done with the first sample in the kernel for better
performance, but it doesn't work anymore with atomics. Any benefit was
very minor anyway, too small to measure it seems.
This removes a bunch of code that is no longer needed, and running
"make update" will now automatically download the new libraries.
Differential Revision: https://developer.blender.org/D2861
This is done by storing only a subset of PathRadiance, and by storing
direct light immediately in the main PathRadiance. Saves about 10% of
CUDA stack memory, and simplifies subsurface indirect ray code.
One crucial thing here: OpenVDB shoudl be compiled WITHOUT
OPENVDB_ENABLE_3_ABI_COMPATIBLE flag. This is how OpenVDB's Makefile is
configured and it's not really possible to detect this for a compiled library.
If we ever want to support that option, we need to add extra CMake argument and
use old version 3 API everywhere.
It has been deprecated since at least macOS 10.9 and fully removed in 10.12.
I am unsure if we should remove it only in 2.8. But you cannot build blender with it supported when using a modern xcode version anyway so I would tend towards just removing it also for 2.79 if that ever happens.
Reviewers: mont29, dfelinto, juicyfruit, brecht
Reviewed By: mont29, brecht
Subscribers: Blendify, brecht
Maniphest Tasks: T52807
Differential Revision: https://developer.blender.org/D2333
For the first bounce we now give each BSDF or BSSRDF a minimum sample weight,
which helps reduce noise for a typical case where you have a glossy BSDF with
a small weight due to Fresnel, but not necessarily small contribution relative
to a diffuse or transmission BSDF below.
We can probably find a better heuristic that also enables this on further
bounces, for example when looking through a perfect mirror, but I wasn't able
to find a robust one so far.
Similar to what we did for area lights previously, this should help
preserve stratification when using multiple BSDFs in theory. Improvements
are not easily noticeable in practice though, because the number of BSDFs
is usually low. Still nice to eliminate one sampling dimension.
Previously the Sobol pattern suffered from some correlation issues that
made the outline of objects like a smoke domain visible. This helps
simplify the code and also makes some other optimizations possible.
Now we replace O(N^2) computational complexity with O(N) extra memory penalty.
Memory is much cheaper than CPU time. Keep in mind, memory penalty is like
4 megabytes per 1M vertices.
