This actually works somewhat now, although viewport rendering is broken and any
kind of network error or connection failure will kill Blender.
* Experimental WITH_CYCLES_NETWORK cmake option
* Networked Device is shown as an option next to CPU and GPU Compute
* Various updates to work with the latest Cycles code
* Locks and thread safety for RPC calls and tiles
* Refactored pointer mapping code
* Fix error in CPU brand string retrieval code
This includes work by Doug Gale, Martijn Berger and Brecht Van Lommel.
Reviewers: brecht
Differential Revision: http://developer.blender.org/D36
This code can't actually be enabled for building and is incomplete, but it's
here because we know we want to support this at some point and there's not much
reason to have it in a separate branch if a simple #ifdef can disable it.
This is mostly work towards enabling the __KERNEL_SSE__ option to start using
SIMD operations for vector math operations. This 4.1 kernel performes about 8%
faster with that option but overall is still slower than without the option.
WITH_CYCLES_OPTIMIZED_KERNEL_SSE41 is the cmake flag for testing this kernel.
Alignment of int3, int4, float3, float4 to 16 bytes seems to give a slight 1-2%
speedup on tested systems with the current kernel already, so is enabled now.
This to avoids build conflicts with libc++ on FreeBSD, these __ prefixed values
are reserved for compilers. I apologize to anyone who has patches or branches
and has to go through the pain of merging this change, it may be easiest to do
these same replacements in your code and then apply/merge the patch.
Ref T37477.
* 32 bit GCC builds now have the SSE BVH optimizations turned off, but still
compile with SSE flags for better performance.
* White color when rendering on Windows seems to have been unrelated to SSE,
rather it was a graphics driver not supporting half float textures, added a
check for that now.
There is some sort of problem with the SSE2 code path, but I couldn't find
the cause, maybe a compiler bug due to the large amount of inlining? For
now I've disabled SSE2 optimizatons in 32 bit GCC builds.
except for curves, that's still missing from the OpenColorIO GLSL shader.
The pixels are stored in a half float texture, converterd from full float with
native GPU instructions and SIMD on the CPU, so it should be pretty quick.
Using a GLSL shader is useful for GPU render because it avoids a copy through
CPU memory.
New features:
* Bump mapping now works with SSS
* Texture Blur factor for SSS, see the documentation for details:
http://wiki.blender.org/index.php/Doc:2.6/Manual/Render/Cycles/Nodes/Shaders#Subsurface_Scattering
Work in progress for feedback:
Initial implementation of the "BSSRDF Importance Sampling" paper, which uses
a different importance sampling method. It gives better quality results in
many ways, with the availability of both Cubic and Gaussian falloff functions,
but also tends to be more noisy when using the progressive integrator and does
not give great results with some geometry. It works quite well for the
non-progressive integrator and is often less noisy there.
This code may still change a lot, so unless you're testing it may be best to
stick to the Compatible falloff function.
Skin test render and file that takes advantage of the gaussian falloff:
http://www.pasteall.org/pic/show.php?id=57661http://www.pasteall.org/pic/show.php?id=57662http://www.pasteall.org/blend/23501
* "Auto Detect" now again uses the umber of cores, instead number of cores + 1.
This was added before we had Tile rendering and benchmarks on several systems showed that there is no gain with this now. There might be some slight difference (0.5% or so) slower/faster depending on the scene, but this is negligible.
RGB color components gave non-grey results when you might no expect it.
What happens is that some of the color channels are zero in the direct light
pass because their channel is zero in the color pass. The direct light pass is
defined as lighting divided by the color pass, and we can't divide by zero. We
do a division after all samples are added together to ensure that multiplication
in the compositor gives the exact combined pass even with antialiasing, DoF, ..
Found a simple tweak here, instead of setting such channels to zero it will set
it to the average of other non-zero color channels, which makes the results look
like the expected grey.
* Reshuffle SSE #ifdefs to try to avoid compilation errors enabling SSE on 32 bit.
* Remove CUDA kernel launch size exception on Mac, is not needed.
* Make OSL file compilation quiet like c/cpp files.
* 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.
On the BMW scene, this gives roughly a 10% speedup overall with clang/gcc, and 30%
speedup with visual studio (2008). It turns out visual studio was optimizing the
existing code quite poorly compared to pretty good autovectorization by clang/gcc,
but hand written SSE code also gives a smaller speed boost there.
This code isn't enabled when using the hair minimum width feature yet, need to
make that work with the SSE code still.
* Added a node to convert wavelength (in nanometers, from 380nm to 780nm) to RGB values. This can be useful to match real world colors easier.
* Code cleanup:
** Moved color functions (xyz and hsv) into dedicated utility files.
** Remove svm_lerp(), use interp() instead.
Documentation:
http://wiki.blender.org/index.php/Doc:2.6/Manual/Render/Cycles/Nodes/More#Wavelength
Example render:
http://www.pasteall.org/pic/show.php?id=53202
This is part of my GSoC 2013. (revisions 57322, 57326, 57335 and 57367 from soc-2013-dingto).
instead of sobol. So far one doesn't seem to be consistently better or worse than
the other for the same number of samples but more testing is needed.
The random number generator itself is slower than sobol for most number of samples,
except 16, 64, 256, .. because they can be computed faster. This can probably be
optimized, but we can do that when/if this actually turns out to be useful.
Paper this implementation is based on:
http://graphics.pixar.com/library/MultiJitteredSampling/
Also includes some refactoring of RNG code, fixing a Sobol correlation issue with
the first BSDF and < 16 samples, skipping some unneeded RNG calls and using a
simpler unit square to unit disk function.
* Revert r57203 (len() renaming)
There seems to be a problem with nVidia OpenCL after this and I haven't figured out the real cause yet.
Better to selectively enable native length() later, after figuring out what's wrong.
This fixes [#35612].
* Rename some math functions:
len -> length
len_squared -> length_squared
normalize_len -> normalize_length
* This way OpenCL uses its inbuilt length() function, rather than our own. The other two functions have been renamed for consistency.
* Tested CPU, CUDA and OpenCL compile, should be no functional changes.
the second time, as for example Intel CPU startup time is 9 seconds.
* Adds an cache for contexts and programs for each platform and device pair,
which also ensure now no two threads try to compile and write the binary cache
file at the same time.
* Change clFinish to clFlush so we don't block until the result is done, instead
it will block at the moment we copy back memory.
* Fix error in Cycles time_sleep implementation, does not affect any active code
though.
* Adds some (disabled) debugging code in the task scheduler.
Patch #35559 by Doug Gale.
* Support using devices from all OpenCL platforms, so that you can use e.g. both
Intel and NVidia OpenCL implementations if you have them installed.
* Fix compile error due to missing fmodf after recent math node change.
* Enable advanced shading for Intel OpenCL.
* CYCLES_OPENCL_DEBUG environment variable for generating debug symbols so you
can debug with gdb. This crashes the compiler with Intel OpenCL on Linux though.
To make this work the preprocessed kernel source code is written out, as gdb
needs this.
* Show OpenCL compiler warnings even if the build succeeded.
* Some small fixes to initialize cdDevice to NULL, add missing NULL check when
creating buffer and add missing space at end of build options for Apple OpenCL.
* Fix crash with multi device + opencl, now e.g. CPU + GPU render should work.
I did a few tweaks to the code and also:
* Fix viewport render failing sometimes with Apple CPU OpenCL, was not taking
workgroup size limits into account properly.
* Add compile error when advanced shading in the Blender binary and OpenCL kernel
are not in sync.
for Apple OpenCL on OS X 10.8 and simple AO render.
Also environment variable CYCLES_OPENCL_TEST can now be set to CPU, GPU,
ACCELERATOR, DEFAULT or ALL values to test particuler devices.
