Old code was working quite unreliable in combination with fast math
flag, especially when compiling with Clang. It seems we were hitting
result of the following bug submitted to Clang [1].
Basically, it was happening so that (int)sqrtf(64) was 7 when Cycles
is built with Clang but was correct 8 when built with GCC.
This commit works this around. Annoying, but don't see other way to
keep sampling pattern the same for Clang and GCC.
[1] https://bugs.llvm.org//show_bug.cgi?id=24063
There were two cases where correlation issues were obvious:
- File from T38710 was giving issues in 2.78a again
- File from T50116 was having totally different shadow between
sample 1 and sample 32.
Use some more simplified version of CMJ hash which seems to give
nice randomized value which solves the correlation.
This commit will break all unit test files, but it's a bug fix
so perhaps OK to commit this.
This also fixes T41143: Sobol gives nonuniform noise
Proper science paper about hash function is coming.
Reviewers: brecht
Reviewed By: brecht
Subscribers: lukasstockner97
Differential Revision: https://developer.blender.org/D2385
This inconsistency drove me totally crazy, it's really confusing
when it's inconsistent especially when you work on both Cycles and
Blender sides.
Shouldn;t cause merge PITA, it's whitespace changes only, Git should
be able to merge it nicely.
Issue was caused by the precision issues which made sdivm by 1 under
it's actual value. We can try to do some eps magic, but from the tests
on laptop and desktop doing integer division is not slower than using
floats here.
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.
* 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.
and sm_30 cards, so hopefully it should all work now.
Also includes some warnings fixes related to nvcc compiler arguments, should make
no difference otherwise.
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.
