Since the kernel split work we're now having quite a few of new files, majority
of which are related on the kernel entry points. Keeping those files in the
root kernel folder will eventually make it really hard to follow which files are
actual implementation of Cycles kernel.
Those files are now moved to kernel/kernels/<device_type>. This way adding extra
entry points will be less noisy. It is also nice to have all device-specific
files grouped together.
Another change is in the way how split kernel invokes logic. Previously all the
logic was implemented directly in the .cl files, which makes it a bit tricky to
re-use the logic across other devices. Since we'll likely be looking into doing
same split work for CUDA devices eventually it makes sense to move logic from
.cl files to header files. Those files are stored in kernel/split. This does not
mean the header files will not give error messages when tried to be included
from other devices and their arguments will likely be changed, but having such
separation is a good start anyway.
There should be no functional changes.
Reviewers: juicyfruit, dingto
Differential Revision: https://developer.blender.org/D1314
Now we build 2 .cubins per architecture (e.g. kernel_sm_21.cubin, kernel_experimental_sm_21.cubin).
The experimental kernel can be used by switching to the Experimental Feature Set: http://wiki.blender.org/index.php/Doc:2.6/Manual/Render/Cycles/Experimental_Features
This enables Subsurface Scattering and Correlated Multi Jitter Sampling on GPU, while keeping the stability and performance of the regular kernel.
Differential Revision: https://developer.blender.org/D762
Patch by Sergey and myself.
Developer / Builder Note:
CUDA Toolkit 6.5 is highly recommended for this, also note that building the experimental kernel requires a lot of system memory (~7-8GB).
CMake had this --fast-math flag but scons not, makes a big difference on some
files. Slightly slower rendering might still happen though, but it should not
be this much.
This also updates the configurations to build kernels for compute capability
5.0 cards, when using and older CUDA toolkit version this will be skipped.
Also includes tweaks to improve performance with this version:
* Increase max registers on sm_30, sm_35 and sm_50
* No longer use texture storage on sm_30
This makes it easier to have per kernel number of registers. Also, all the
tunable parameters for this are now in kernel.cu, rather than spread over cmake,
scons and device_cuda.cpp.
This fixes the ptaxs "ACCESS_VIOLATION" error and should allow our Linux and Windows build bots to compile again.
Unfortunately this comes with a performance penalty on sm_2x cards, so this is only a workaround for now. Branched Path is still globally disabled on GPU.
* Remove support for CUDA Toolkit 4.x, only Toolkit 5.0 and above are supported now.
* Remove support for sm_1x cards (< Fermi) for good. We didn't officially support those cards for a few releases already, now remove some special code that was still there.
* Add CUDA compiler version detection to cmake/scons/runtime
* Remove noinline in kernel_shader.h and reenable --use_fast_math if CUDA 5.x
is used, these were workarounds for CUDA 4.2 bugs
* Change max number of registers to 32 for sm 2.x (based on performance tests
from Martijn Berger and confirmed here), and also for NVidia OpenCL.
Overall it seems that with these changes and the latest CUDA 5.0 download, that
performance is as good as or better than the 2.67b release with the scenes and
graphics cards I tested.
* Added option "WITH_BF_CYCLES_CUDA_THREADED_COMPILE" for the people who have much RAM (8 or more) and can compile several kernels at the same time. If enabled, it uses the general BF_NUMJOBS flag.
* The option is off per default.
* Compile all of cycles with -ffast-math again
* Add scons compilation of cuda binaries, tested on mac/linux.
* Add UI option for supported/experimental features, to make it
more clear what is supported, opencl/subdivision is experimental.
* Remove cycles xml exporter, was just for testing.