This is a preparation work for the followup commit which wil l move
remaining parts of Woop intersection logic to an utility file.
Doing it as a separate commit to keep changes more atomic and easier
to bisect when/if needed.
There are following benefits:
- Modifying intersection algorithm will not cause so much re-compilation.
- It works around header dependency hell and allows us to use vectorization
types much easier in there.
Single program generally compiles kernels faster (2-3 times), loads faster,
takes less drive space (2-3 times), and reduces the number of cached kernels.
This does a few things at once:
- Refactors host side split kernel logic into a new device
agnostic class `DeviceSplitKernel`.
- Removes tile splitting, a new work pool implementation takes its place and
allows as many threads as will fit in memory regardless of tile size, which
can give performance gains.
- Refactors split state buffers into one buffer, as well as reduces the
number of arguments passed to kernels. Means there's less code to deal
with overall.
- Moves kernel logic out of OpenCL kernel files so they can later be used by
other device types.
- Replaced OpenCL specific APIs with new generic versions
- Tiles can now be seen updating during rendering
Please NEVER EVER use such a statement, it's only causing HUGE
issues. What is even worse: it's not always possible to immediately
see that the hell is coming from such a statement.
There is still some statements in the existing code, will leave
those for a later cleanup.
The Progress system in Cycles had two limitations so far:
- It just counted tiles, but ignored their size. For example, when rendering a 600x500 image with 512x512 tiles, the right 88x500 tile would count for 50% of the progress, although it only covers 15% of the image.
- Scene update time was incorrectly counted as rendering time - therefore, the remaining time started very long and gradually decreased.
This patch fixes both problems:
First of all, the Progress now has a function to ignore time spans, and that is used to ignore scene update time.
The larger change is the tile size: Instead of counting samples per tile, so that the final value is num_samples*num_tiles, the code now counts every sample for every pixel, so that the final value is num_samples*num_pixels.
Along with that, some unused variables were removed from the Progress and Session classes.
Reviewers: brecht, sergey, #cycles
Subscribers: brecht, candreacchio, sergey
Differential Revision: https://developer.blender.org/D2214
They are defined for MSVC but seems to be missing in GCC and CLang-3.8.
Maybe some further tweaks to policy when to define those functions is
needed, but should be fine for now.
I can no longer reproduce crash with neither of the files where
the crash was originally visible. This is something where other
changes (light threshold, sampling) had an effect and made code
to work as it is supposed to. Could have been optimizator issue
or something like that.
Let's see if we hit same issue again.
Main intention is to give some quick way to control scene's memory
usage by clamping textures which are too big. This is really handy
on the early production stages when you first create really nice
looking hi-res textures and only when it all works and approved
start investing time on optimizing your scene.
This is a new option in Scene Simplify panel and it acts as
following: when texture size is bigger than the given value it'll
be scaled down by half for until it fits into given limit.
There are various possible improvements, such as:
- Use threaded scaling using our own task manager.
This is actually one of the main reasons why image resize is
manually-implemented instead of using OIIO's resize. Other
reason here is that API seems limited to construct 3D texture
description easily.
- Vectorization of uchar4/float4/half4 textures.
- Use something smarter than box filter.
Was playing with some other filters, but not sure they are
really better: they kind of causes more fuzzy edges.
Even with such a TODOs in the code the option is already quite
useful.
Reviewers: brecht
Reviewed By: brecht
Subscribers: jtheninja, Blendify, gregzaal, venomgfx
Differential Revision: https://developer.blender.org/D2362
This allows to save a memory copy, which will be particularly useful for network rendering.
Reviewers: sergey, brecht, dingto, juicyfruit, maiself
Differential Revision: https://developer.blender.org/D2323
In scenes with many lights, some of them might have a very small contribution to some pixels, but the shadow rays are traced anyways.
To avoid that, this patch adds probabilistic termination to light samples - if the contribution before checking for shadowing is below a user-defined threshold, the sample will be discarded with probability (1 - (contribution / threshold)) and otherwise kept, but weighted more to remain unbiased.
This is the same approach that's also used in path termination based on length.
Note that the rendering remains unbiased with this option, it just adds a bit of noise - but if the setting is used moderately, the speedup gained easily outweighs the additional noise.
Reviewers: #cycles
Subscribers: sergey, brecht
Differential Revision: https://developer.blender.org/D2217
When using the Normal output of the Texture Coordinate node on Point and Spot lamps, the coordinates now depend on the rotation of the lamp.
On Area lamps, the Parametric output of the Geometry node now returns UV coordinates on the area lamp.
Credit for the Area lamp part goes to Stefan Werner (from D1995).
Oh man, is it a compiler bug? Is it something we do stupid?
For now more crap to prevent crashes. During the conference will talk to
Maxyn about how can we troubleshoot such weird issues.
Basically don't use rcp() in areas which seems to be critical after
second look. Also disabled some multiplication operators, not sure
yet why they might be a problem.
Tomorrow will be setting up a full test with all cases which were
buggy in our farm to see if this fix is complete.
There is some precision issues for big magnitude coordinates which started
to give weird behavior of release builds. Some weird memory usage in BVH
which is tricky to nail down because only happens in release builds and GDB
reports all variables as optimized out when trying to use RelWithDebInfo.
There are two things in this commit:
- Attempt to make vectorized code closer to original one, hoping that it'll
eliminate precision issue.
This seems to work for transform_point().
- Similar trick did not work for transform_direction() even tho absolute
error here is much smaller. For now disabled that function, need a more
careful look here.
Several ideas here:
- Optimize calculation of near_{x,y,z} in a way that does not require
3 if() statements per update, which avoids negative effect of wrong
branch prediction.
- Optimization of direction clamping for BVH.
- Optimization of point/direction transform.
Brings ~1.5% speedup again depending on a scene (unfortunately, this
speedup can't be sum across all previous commits because speedup of
each of the changes varies from scene to scene, but it still seems to
be nice solid speedup of few percent on Linux and bigger speedup was
reported on Windows).
Once again ,thanks Maxym for inspiration!
Still TODO: We have multiple places where we need to calculate near
x,y,z indices in BVH, for now it's only done for main BVH traversal.
Will try to move this calculation to an utility function and see if
that can be easily re-used across all the BVH flavors.
The idea here is to avoid if statements which could cause wrong
branch prediction.
Gives a bit of measurable speedup up to ~1%. Still nice :)
Inspired by Maxym Dmytrychenko, thanks!
Initialization order of global stats and node types was not strictly
defined and it was possible to have node types initialized first and
stats after that. This will zero out memory which was allocated from
the statistics causing assert failure when de-initializing node types.
This was giving some speedup but made intersection tests to fail
from watertight point of view.
Needs deeper investigation, but need to quickly get it fixed for
the studio.
This gives about 5% speedup on AVX2 kernels (other kernels still
have SSE disabled for math operations) and this solves the slowdown
of koro scene mention in the previous commit.
The title says it all actually. This commit also contains
changes to pass float3 as const reference in affected functions.
This should make MSVC happier without breaking OpenCL because it's
only done in areas which are ifdef-ed for non-OpenCL.
Another patch based on inspiration from Maxym Dmytrychenko, thanks!
Based on existing ssef data type and to my knowledge it's also what happens in
Embree nowadays.
Inspired by Maxym Dmytrychenko and required for the upcoming triangle
intersection commit.
Hopefully the copyright message is correct.
Mostly this is making inlining match CUDA 7.5 in a few performance critical
places. The end result is that performance is now better than before, possibly
due to less register spilling or other CUDA 8.0 compiler improvements.
On benchmarks scenes, there are 3% to 35% render time reductions. Stack memory
usage is reduced a little too.
Reviewed By: sergey
Differential Revision: https://developer.blender.org/D2269
