Instead of 95, we can use 145 images now. This only affects Kepler and above (sm30, sm_35 and sm_50).
This can be increased further if needed, but let's first test if this does not come with a performance impact.
Originally developed during my GSoC 2013.
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
Otherwise devices used for display will lock up the UI too much. This means
you might still get 100% CPU for the display device, but for others CPU usage
should be low still.
The check to see if a device is used for display may not be entirely reliable,
it checks if there is a watchdog timeout on the device, but I'm not entirely
sure that always exists for display devices or is disabled for non-display
devices, though some tools like cuda-gdb seem to make the same assumption.
Ref T39559
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.
Issue was caused by the wrong usage of OCIO GLSL binding API. To make it
work properly on pre-GLSL-1.3 drivers shader is to be enabled after the
texture is binded to the opengl context. Otherwise it wouldn't know the
proper texture size.
This is actually a regression in 2.70 and to be ported to 'a'.
All textures are sampled bi-linear currently with the exception of OSL there texture sampling is fixed and set to smart bi-cubic.
This patch adds user control to this setting.
Added:
- bits to DNA / RNA in the form of an enum for supporting multiple interpolations types
- changes to the image texture node drawing code ( add enum)
- to ImageManager (this needs to know to allocate second texture when interpolation type is different)
- to node compiler (pass on interpolation type)
- to device tex_alloc this also needs to get the concept of multiple interpolation types
- implementation for doing non interpolated lookup for cuda and cpu
- implementation where we pass this along to osl ( this makes OSL also do linear untill I add smartcubic to the interface / DNA/ RNA)
Reviewers: brecht, dingto
Reviewed By: brecht
CC: dingto, venomgfx
Differential Revision: https://developer.blender.org/D317
This switches api usage for cuda towards using more of the Async calls.
Updating only once every second is sufficiently cheap that I don't think it is worth doing it less often.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D262
Cycles GPU Performance Regression
From my testing this (what i should have done in the first place) reduces the regression a lot.
Lets hope it is enough or we have to go back to busy waiting.
This is my first stab at this and is based on this IRC converstation:
<mib2berlin> brecht: this is meaning as reminder only, I know you have other things to do > http://openvidia.sourceforge.net/index.php/Optimization_Notes#avoiding_busy_waits
<brecht> mib2berlin: thanks, bookmarked
only tested on Ubuntu 14.04 / cuda 5.0 but ill do some more testing tomorrow.
Also unsure about the placement and the lifetime of the stream and the event. But creating / deleting these seems to incur a non trivial cost.
Reviewers: brecht
Reviewed By: brecht
CC: mib2berlin, dingto
Differential Revision: https://developer.blender.org/D262
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
* 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.
use arrays instead of textures for general storage on this card (image textures
are still stored as texture). Textures were found to be faster on older cards,
but the limits on 1D texture size have not increased along with the memory size,
which meant that the full 6 GB could not be used.
The performance actually seems to be slightly better with arrays in some tests
on Titan. For older cards there seems to be a bit of a mix, some are better and
others not. We may change those to use arrays too, but more testing is needed,
only Titan and Tesla K20 (sm_35) is changed for now.
The fact that arrays are faster is a bit surprising, as others found textures
to be faster on Kepler. However even if they were, the memory limitation is
more important to solve anyway.
https://research.nvidia.com/publication/understanding-efficiency-ray-traversal-gpus-kepler-and-fermi-addendum
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.
* GPU kernel can now be compiled without __NON_PROGRESSIVE__ again, was broken after my last commit. Also add a check for have_error(), in case the GPU kernel comes without Non-Progressive, to avoid a crash.
* Don't compile progressive kernel twice on CPU, if __NON_PROGRESSIVE__ would be disabled there.
* Non-Progressive integrator is now available on the GPU (CUDA, sm_20 and above).
Implementation details:
* kernel_path_trace() has been split up into two functions:
kernel_path_trace_non_progressive() and kernel_path_trace_progressive().
* We compile two CUDA kernel entry functions (in kernel.cu) for the two integrators, they are still inside one .cubin file but due to the kernel separation there should be no performance problem. I tested with the BMW file on my Geforce 540M and the render times were the same for 100 samples (1.57 min in my case).
This is part of my GSoC project, SVN merge of r59032 + manual merge of UI changes for this from my branch.
* 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.
* Deprecate computing capability 1.3 (sm_13)
This commit disables auto build of sm_13 CUDA platform, which means that starting with Blender 2.67, we don't support sm_13 devices anymore. It has become difficult to support that and it was already feature incomplete (no render-passes, AO, Multi Closure etc).
It's still possible to manually enable sm_13 for own tests, but building might break in the future.
* CUDA: Make it more clear that sm_12 and below is not supported.
* OpenCL: __KERNEL_SHADING__ was declared twice for nvidia opencl device.
* Some reshuffle of defines in kernel_types.h. No functional changes.
precompiled cubins instead,
Logic here is following now:
- If there're precompiled cubins, assume CUDA compute is available,
otherwise
- If cuda toolkit found, assume CUDA compute is available
- In all other cases CUDA compute is not available
For windows there're still check for only precompiled binaries,
no runtime compilation is allowed.
Ended up with such decision after discussion with Brecht. The thing
is, if we'll support runtime compilation on windows we'll end up
having lots of reports about different aspects of something doesn't
work (you need particular toolkit version, msvc installed, environment
variables set properly and so) and giving feedback on such reports
will waste time.
This commit adds memory usage information while rendering.
It reports memory used by device, meaning:
- For CPU it'll report real memory consumption
- For GPU rendering it'll report GPU memory consumption, but it'll
also mean the same memory is used from host side.
This information displays information about memory requested by Cycles,
not memory really allocated on a device. Real memory usage might be
higher because of memory fragmentation or optimistic memory allocator.
There's really nothing we can do against this.
Also in contrast with blender internal's render cycles memory usage
does not include memory used by scene, only memory needed by cycles
itself will be displayed. So don't freak out if memory usage reported
by cycles would be much lower than blender internal's.
This commit also adds RenderEngine.update_memory_stats callback which
is used to tell memory consumption from external engine to blender.
This information is used to generate information line after rendering
is finished.
Just makes progressive refine :)
This means the whole image would be refined gradually using as much
threads as it's set in performance settings. Having enough tiles is
required to have this option working as it's expected.
Technically it's implemented by repeatedly computing next sample for
all the tiles before switching to next sample.
This works around 7-12% slower than regular tile-based rendering, so
use this option only if you really need it.
This commit also fixes progressive update of image when Save Buffers
option is enabled.
And one more thing this commit fixes is handling display buffer with
Save Buffers option enabled. If this option is enabled image buffer
wouldn't have neither byte nor float buffer until image is fully
rendered which could backfire in missing image while rendering in
cases color management cache became full.
This issue solved by allocating byte buffer for image buffer from
tile update callback.
Patch was reviewed by Brecht. He also made some minor edits to
original version to patch. Thanks, man!
