The goal is to make Experimental kernel closer in performance to the
official kernel, avoiding spills and such.
There should not be big impact on official kernel, own tests showed
few percent performance drop on laptop's GPU. CPU was always the
same speed on AVX, AVX2 and SSE4.1 CPUs i've been testing here.
This seems to be the last essential step before we can get rid of
Experimental kernel and enable SSS officially on GPU without causing
some major performance issues.
Surely some more tweaks are possibly required, but that we can do
for until cows go home anyway.
This commit changes the way how we pass bounce information to the Light
Path node. Instead of manualy copying the bounces into ShaderData, we now
directly pass PathState. This reduces the arguments that we need to pass
around and also makes it easier to extend the feature.
This commit also exposes the Transmission Bounce Depth to the Light Path
node. It works similar to the Transparent Depth Output: Replace a
Transmission lightpath after X bounces with another shader, e.g a Diffuse
one. This can be used to avoid black surfaces, due to low amount of max
bounces.
Reviewed by Sergey and Brecht, thanks for some hlp with this.
I tested compilation and usage on CPU (SVM and OSL), CUDA, OpenCL Split
and Mega kernel. Hopefully this covers all devices. :)
There were multiple issues which are solved now:
- It was possible that ray wouldn't be bounced off the BSSRDF, for example
when PDF or shader eval is zero. In this case PathState might have been
left in pre-bounced state which would have been gave incorrect shading
results.
This is solved by having separate PathState for each of the hits.
- Path radiance summing wasn't happening correct as well, indirect rays
were using wrong path radiance in the case when there were more than
one hit recorded.
This is now using a bit trickier state machine which calculates path
radiance for just SSS (both direct and indirect) and then sums it back
to the final radiance.
- Previous commit wasn't totally correct either and was an induced bug
due to wrong path state left from the "un-happened" ray bounce.
There should be no special case happening here, BSSRDFs will be replaced
with diffuse ones due to PATH_RAY_DIFFUSE_ANCESTOR flag.
- Merged back codebases for "delayed" and "immediate" indirect SSS ray
tracing, hopefully making it easier to maintain the codebase.
Sure this changes brings memory usage back by about 4-5%, but overall
it's still about 2x memory reduction for the experimental kernel here.
Thanks Brecht for the review!
This is actually how it was intended to work, just didn't notice it wasn't
really happening in the main ray loop.
Solves some memory issues reported in T46880.
There are some issues to be solved with the recent optimization we did for
the indirect rays for the SSS. Those issues will take a bit of a time to
be fully solved still and we need to unlock Caminandes team now, so let's
revert some changes back.
CUDA will still use delayed indirect rays since it's an experimental
feature.
For the details about what's to be done still please refer to T46880.
This wasn't really a complete fix and only worked if there was a single scatter
event recorded only. Proper fix requires some more thoughts to make it correct
without memory use increase.
This reverts commit bf9e88bfbe.
Radiance sum and reset was happening in different order after 26f1c51.
This is a quick fix to unlock Caminandes team, perhaps we can avoid having
separate variable to detect when radiance is to be sum.
This gives much lower stack usage on GPU and reduces kernel memory size to
around 448MB on GTX560Ti (comparing to 652MB with previous commit and 946MB
with official release). There's also a barely measurable speedup of around
5%, but this is to be confirmed still.
At this stage we're using only ~3% for the experimental kernel and SSS
rendering seems to be faster by 40% and after some further testing we might
consider making SSS and CMJ official features and remove experimental
precompiled kernels.
The idea is to delay shooting indirect rays for the SSS sampling and
trace them after the main integration loop was finished.
This reduces GPU stack usage even further and brings it down to around
652MB (comparing to 722MB before the change and 946MB with previous
stable release).
This also solves the speed regression happened in the previous commit
and now simple SSS scene (SSS suzanne on the floor) renders in 0:50
(comparing to 1:16 with previous commit and 1:03 with official release).
This commit introduces a SSS-oriented intersection structure which is replacing
old logic of having separate arrays for just intersections and shader data and
encapsulates all the data needed for SSS evaluation.
This giver a huge stack memory saving on GPU. In own experiments it gave 25%
memory usage reduction on GTX560Ti (722MB vs. 946MB).
Unfortunately, this gave some performance loss of 20% which only happens on GPU.
This is perhaps due to different memory access pattern. Will be solved in the
future, hopefully.
Famous saying: won in memory - lost in time (which is also valid in other way
around).
Code there started becoming a bit too big, by splitting it up it'll make it
easier to do improvements or extending the features in there.
The layout is not totally final yet, would need to try de-duplicating parts
of code from split kernel with non-split integrators,
TODO: We might want to refactor debug passes into PASS_DEBUG and some
debug_type (similar to Blender's side passes) to avoid issue of running
out of bits.
Quite straightforward implementation, but still needs some work for the split
kernel. Includes both regular and split kernel implementation for that.
The pass is not exposed to the interface yet because it's currently not really
easy to have same pass listed in the menu multiple times.
This commit contains all the work related on the AMD megakernel split work
which was mainly done by Varun Sundar, George Kyriazis and Lenny Wang, plus
some help from Sergey Sharybin, Martijn Berger, Thomas Dinges and likely
someone else which we're forgetting to mention.
Currently only AMD cards are enabled for the new split kernel, but it is
possible to force split opencl kernel to be used by setting the following
environment variable: CYCLES_OPENCL_SPLIT_KERNEL_TEST=1.
Not all the features are supported yet, and that being said no motion blur,
camera blur, SSS and volumetrics for now. Also transparent shadows are
disabled on AMD device because of some compiler bug.
This kernel is also only implements regular path tracing and supporting
branched one will take a bit. Branched path tracing is exposed to the
interface still, which is a bit misleading and will be hidden there soon.
More feature will be enabled once they're ported to the split kernel and
tested.
Neither regular CPU nor CUDA has any difference, they're generating the
same exact code, which means no regressions/improvements there.
Based on the research paper:
https://research.nvidia.com/sites/default/files/publications/laine2013hpg_paper.pdf
Here's the documentation:
https://docs.google.com/document/d/1LuXW-CV-sVJkQaEGZlMJ86jZ8FmoPfecaMdR-oiWbUY/edit
Design discussion of the patch:
https://developer.blender.org/T44197
Differential Revision: https://developer.blender.org/D1200
The issue was caused by the way how we shoot the ray to see which rays we're
inside which might start bouncing back-n-forth between two close to parallel
intersecting faces.
Real solution would be to record all the intersections when shooting the ray,
but it's kinda tricky on GPU because of needed sorting and uncertainty of
how huge intersection array should be.
For now we'll just limit number of steps in the check so in worst case we'll
have some samples not being correct which will be compensated with further
sampling. Shouldn't be an issue since probability of such a lock is quite
small actually.
basically we skip all non-volume objects now in the volume stack function.
Depending on the show it might give some percent of speedup.
Most of the speedup would be gained in the scenes when having SSS object
intersecting the volume and taking a reasonable amount of frame space.
Quite straightforward change, the only annoying thing is that we can't use
indentation for include directive just because of the way headers inlineing
works for OpenCL.
Might do smarter job in path_source_replace_includes() but don't want to
spend time on this yet.
Currently only summed number of traversal steps and intersections used by the
camera ray intersection pass is implemented, but in the future we will support
more debug passes which would help checking what things makes the scene slow.
Example of such extra passes could be number of bounces, time spent on the
shader tree evaluation and so.
Implementation from the Cycles side is pretty much straightforward, could only
mention here that it's a build-time option disabled by default.
From the blender side it's implemented as a PASS_DEBUG with several subtypes
possible. This way we don't need to create an extra DNA pass type for each of
the debug passes, saving us a bits.
Reviewers: campbellbarton
Reviewed By: campbellbarton
Differential Revision: https://developer.blender.org/D813
This adds an AABB collision check for objects with volumes and if there's a
collision detected then the object will have SD_OBJECT_INTERSECTS_VOLUME flag.
This solves a speed regression introduced by the fix for T39823 by skipping
volume stack update in cases no volumes intersects the current SSS object.
Basically the title says it all, volume stack initialization now is aware that
camera might be inside of the volume. This gives quite noticeable render time
regressions in cases camera is in the volume (didn't measure them yet) because
this requires quite a few of ray-casting per camera ray in order to check which
objects we're inside. Not quite sure if this might be optimized.
But the good thing is that we can do quite a good job on detecting whether
camera is outside of any of the volumes and in this case there should be no
time penalty at all (apart from some extra checks during the sync state).
For now we're only doing rather simple AABB checks between the viewplane and
volume objects. This could give some false-positives, but this should be good
starting point.
Need to mention panoramic cameras here, for them it's only check for whether
there are volumes in the scene, which would lead to speed regressions even if
the camera is outside of the volumes. Would need to figure out proper check
for such cameras.
There are still quite a few of TODOs in the code, but the patch is good enough
to start playing around with it checking whether there are some obvious mistakes
somewhere.
Currently the feature is only available in the Experimental feature sey, need
to solve some of the TODOs and look into making things faster before considering
the feature is ready for the official feature set. This would still likely
happen in current release cycle.
Reviewers: brecht, juicyfruit, dingto
Differential Revision: https://developer.blender.org/D794
Basically the title says it all, we need to update volume stack when doing ray
scatter for SSS. This leads to speed regressions in cases scene does have both
volume and SSS (performance in case there's no SSS or no volume should be the
same).
We might try optimizing kernel_path_subsurface_update_volume_stack() a bit by
either recording all intersections or using some more appropriate visibility
flags.
Reviewers: brecht, juicyfruit, dingto
Differential Revision: https://developer.blender.org/D795
* __VOLUME__ is basic volume support with Emission and Absorption.
* __VOLUME_SCATTER__ enables volume Scattering support.
* __VOLUME_DECOUPLED__ enables Decoupled Ray Marching.
