In that case it can now fall back to CPU memory, at the cost of reduced
performance. For scenes that fit in GPU memory, this commit should not
cause any noticeable slowdowns.
We don't use all physical system RAM, since that can cause OS instability.
We leave at least half of system RAM or 4GB to other software, whichever
is smaller.
For image textures in host memory, performance was maybe 20-30% slower
in our tests (although this is highly hardware and scene dependent). Once
other type of data doesn't fit on the GPU, performance can be e.g. 10x
slower, and at that point it's probably better to just render on the CPU.
Differential Revision: https://developer.blender.org/D2056
Previously, the NLM kernels would be launched once per offset with one thread per pixel.
However, with the smaller tile sizes that are now feasible, there wasn't enough work to fully occupy GPUs which results in a significant slowdown.
Therefore, the kernels are now launched in a single call that handles all offsets at once.
This has two downsides: Memory accesses to accumulating buffers are now atomic, and more importantly, the temporary memory now has to be allocated for every shift at once, increasing the required memory.
On the other hand, of course, the smaller tiles significantly reduce the size of the memory.
The main bottleneck right now is the construction of the transformation - there is nothing to be parallelized there, one thread per pixel is the maximum.
I tried to parallelize the SVD implementation by storing the matrix in shared memory and launching one block per pixel, but that wasn't really going anywhere.
To make the new code somewhat readable, the handling of rectangular regions was cleaned up a bit and commented, it should be easier to understand what's going on now.
Also, some variables have been renamed to make the difference between buffer width and stride more apparent, in addition to some general style cleanup.
There was some changes about namespaces, which causes ambiguities.
Replaces using namespace with an explicit symbols we need. Is good idea to NOT
pull in the whole namespace anyway!
* Remove tex_* and pixels_* functions, replace by mem_*.
* Add MEM_TEXTURE and MEM_PIXELS as memory types recognized by devices.
* No longer create device_memory and call mem_* directly, always go
through device_only_memory, device_vector and device_pixels.
* 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.
One problem is that it was always using __mm_blendv_ps emulation even if the
instruction was supported. The other that the emulation function was wrong.
Thanks a lot to Ray Molenkamp for tracking this one down.
While unlikely to have had any serious effects because of limited use, the
previous implementation was not actually atomic due to a data race and
incorrectly coded CAS loop. We also had duplicates of this code in a few
places, it's now been moved to a single location with all other atomic
operations.
It is defined to & for CPU side compilation, and defined to an empty for any GPU
platform. The idea here is to use this macro instead of #ifdef block with bunch
of duplicated lines just to make it so CPU code is efficient.
Eventually we might switch to references on CUDA as well, but that would require
some intensive testing.
