Pass globals as a bare pointer, same as it sued to be prior to split kernel rework.
AMD CPU platform and Intel OpenCL were complaining about this.
Perhaps we shouldn't pass globals as pointer at all, this isn't something what is
really portable and can cause issues on 32 bit perhaps.
While this compiler is not officially supported yet, getting it to work is
a nice thing because more and more AMD cards will fall under MESA driver.
It's also nice to use explicit comparison with NULL, which makes it more
clear whether variable is a boolean or pointer. Even Rust enforces this!
Patch by Ian Bruce with own modifications.
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.
Reduces memory allocation for split kernel.
This allows for faster rendering due to bigger global size,
specially when GPU memory is limited.
Perfromance results:
R9 290 total render time
Before After Change
BMW 4:37 4:34 -1.1 %
Classroom 14:43 14:30 -1.5 %
Fishy Cat 11:20 11:04 -2.4 %
Koro 12:11 12:04 -1.0 %
Pabellon Barcelona 22:01 20:44 -5.8 %
Pabellon Barcelona(*) 15:32 15:09 -2.5 %
(*) without glossy connected to volume
Decoupled ray marching is not supported yet.
Transparent shadows are always enabled for volume rendering.
Changes in kernel/bvh and kernel/geom are from Sergey.
This simiplifies code significantly, and prepares it for
record-all transparent shadow function in split kernel.
By calculating the size of the state buffer in the kernel rather than the host
less code is needed and the size actually reflects the requested features.
Will also be a little faster in some cases because of larger global work size.
Because the split kernel can render multiple samples in parallel it is
necessary to have everything initialized before rendering of any samples
begins. The code that normally handles initialization of
`rng_state` (`kernel_path_trace_setup()`) only does so for the first sample,
which was causing artifacts in the split kernel due to uninitialized
`rng_state` for some samples.
Note that because the split kernel can render samples in parallel this
means that the split kernel is incompatible with the LCG.
This was only needed for the previous implementation of parallel samples. As
we don't have that any more it can be removed.
Real reason for removal tho is this: `per_sample_output_buffers` was being
calculated too small and artifacts resulted. The tile buffer is already
the correct size and calculating the size for `per_sample_output_buffers`
is a bit difficult with the current layout of the code. As
`per_sample_output_buffers` was only needed for `sum_all_radiance`,
removing that kernel and writing output to the tile buffer directly
fixes the artifacts.
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
Doesn't currently change anything, but would need for some future
work here.
It uses existing padding in kernel BVH structure, so there is
nothing changed memory-wise.
The issue here was mainly coming from minimal pixel width feature
which is quite commonly enabled in production shots.
This feature will use some probabilistic heuristic in the curve
intersection function to check whether we need to return intersection
or not. This probability is calculated for every intersection check.
Now, when we use multiple BVH nodes for curve primitives we increase
probability of that primitive to be considered a good intersection
for us. This is similar to increasing minimal width of curve.
What is worst here is that change in the intersection probability
fully depends on exact layout of BVH, meaning probability might
change differently depending on a view angle, the way how builder
binned the primitives and such. This makes it impossible to do
simple check like dividing probability by number of BVH steps.
Other solution might have been to split BVH into fully independent
trees, but that will increase memory usage of all the static
objects in the scenes, which is also not something desirable.
For now used most simple but robust approach: store BVH primitives
time and test it in curve intersection functions. This solves the
regression, but has two downsides:
- Uses more memory.
which isn't surprising, and ANY solution to this problem will
use more memory.
What we still have to do is to avoid this memory increase for
cases when we don't use BVH motion steps.
- Reduces number of maximum available textures on pre-kepler cards.
There is not much we can do here, hardware gets old but we need
to move forward on more modern hardware..
Seems CUDA failed to de-duplicate the array across multiple inlined
versions of the shadow_blocked(). Helped it a bit with that now.
Gives about 100MB memory improvement on a scenes after previous
commit and brings up memory "regression" to only 100MB comparing to
the master branch now.
This commit enables record-all behavior of transparent shadows
rays.
Render times difference goes as following:
GTX 1080 render time
BMW -0.5%
Fishy Cat -0.0%
Pabellon Barcelona -11.6%
Classroom +1.2%
Koro -58.6%
Kernel will now use some extra VRAM memory to store the intersection
array (200MB on my configuration). This we can optimize out with some
further commits.
The idea is to record all possible transparent intersections when
shooting transparent ray on GPU (similar to what we were doing on
CPU already).
This avoids need of doing whole ray-to-scene intersections queries
for each intersection and speeds up a lot cases like transparent
hair in the cost of extra memory.
This commit is a base ground for now and this feature is kept
disabled for until some further tweaks.
Now we break the traversal cycle and then perform volume attenuation
and check with zero throughput. Not sure it makes any measurable sense
at this moment, but in the future it might help de-duplicating some
extra logic here.
