Intel documentation for Ubuntu 22.04 does list all runtime components
needed by the driver and oneAPI Cycles device but end-users getting
drivers from (other) sources can easily end-up missing required
Level-Zero Loader and struggle root causing what's wrong in their
system. Calling this requirement out in the UI will hopefull help them.
oneAPI Level-Zero incl. Loader: https://github.com/oneapi-src/level-zero
Common package names: level-zero, level-zero-loader
This fixes a 15% performance regression silently introduced by
79ab76e156 that aligned the compact
float3 on 16 bytes for oneAPI.
Current change is minimalist, there are further cleanup opportunities
such as removing packed_float3 definition for oneAPI but for some
reason, it cuts the recovered speedup in half, so we're starting with
this small fix for now.
Reviewed by: brecht
Differential Revision: https://developer.blender.org/D16340
This patch fixes T101790 by adding a macOS version check for deciding whether to show the caustics settings in the UI (MNEE kernels don't compile on macOS < 13.0)
Reviewed By: brecht
Maniphest Tasks: T101790
Differential Revision: https://developer.blender.org/D16339
This patch tunes the integrator state sizing for Metal (`num_concurrent_states` and `num_concurrent_busy_states`).
On all GPUs architecture, we adjust the busy:total states ratio to be 1:4 which gives better rendering performance than the previous 1:16 ratio (independent of total state count). This gives a small performance uplift (e.g. 2-3% on M1 Ultra).
Additionally for M2 architectures, we double the overall state size if there is available headroom. Inclusive of the first change, we can expect uplift of close to 10% in future, as this results in larger dispatch sizes and minimises work submission overheads. In order to make an accurate determination of available headroom, we defer the calculation of `num_concurrent_states` and `num_concurrent_busy_states` until the time of integrator state allocation (i.e. after all of the scene data has been allocated). We also refactor `alloc_integrator_soa` to calculate an *exact* single-state-size in a first pass, right before allocating the integrator SoA buffers in a second pass.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D16313
This is to help ensure buildbot builds are correct, while still gracefully
disabling features in user/developer builds.
* Add WITH_STRICT_BUILD_OPTIONS to give an error when features can't be
enabled due to missing libraries or other reasons. Add new macro
set_and_warn_library_found used everywhere features were being
automatically disabled.
* Remove code from Windows and macOS for various libraries that would
automatically disable features. set_and_warn_library_found could be
used here also, but we are generally assuming the precompiled libraries
are complete and only test for availability when libraries are just
added.
Differential Revision: https://developer.blender.org/D16104
Buildbot infrastructure relies on the fact that it can enable and
disable `WITH_CYCLES_<COMPUTE>_BINARIES` without affecting speed of
incremental builds. This allows buildbot to skip GPU kernels when
doing CI regression tests which do not need GPU kernels, as well as
it allows to move GPU kernels compilation to a separate step where
all the resources are available to the GPU kernel builders.
For the oneAPI compute enabling and disabling AoT kernels has much
higher implications due to the kernels being a part of the device
implementation from the build target perspective.
This change makes it so different target names are used for JIT and
AoT configurations, which allows CMake to more fully benefit from
"caching" the compiled result.
The end goal of this change is to make it so sequential build of the
same code base on the buildbot happens super fast,
Blender binary still needs to be re-linked when the AOT of oneAPI
option is toggled, but that's already the case in the buildbot due
to the WITH_BUILDINFO.
Differential Revision: https://developer.blender.org/D16312
sycl::info::device::ext_intel_* descriptors are deprecated,
replaced with sycl::ext::intel::info::device:: that are available from
6.0+, for which we now check version in CMake.
In T93382, the problem was that the Blender-side rendering code was
still generating the subsurface passes because the old render pass
flags were set, even though Cycles doesn't generate them anymore.
After a closer look, it turns out that the entire hardcoded pass
creation code can be removed. We already have an Engine API function
to query the list of render passes from the engine, so we might as
well just call that and create the returned passes.
Turns out that Eevee already did this anyways. On the Cycles side, it
allows to deduplicate a lot of `BlenderSync::sync_render_passes`.
Before, passes were defined in engine.py and in sync.cpp. Now, all
passes that engine.py returns are created automatically, so sync.cpp
only needs to handle a few special cases.
I'm not really concerned about affecting external renderer addons,
since they already needed to handle the old "builtin passes" in
their Engine API implementation anyways to make them show up in the
compositor. So, unless they missed that for like 10 releases, they
should not notice any difference.
Differential Revision: https://developer.blender.org/D16295
Known Issues:
- Command buffer failures when using binary archives (binary archives is disabled for Intel GPUs as a workaround)
- Wrong texture sampler being applied (to be addressed in the future)
Ref T92212
Reviewed By: brecht
Maniphest Tasks: T92212
Differential Revision: https://developer.blender.org/D16253
Previously, a first build using ninja would throw "ninja: error:
'intern/cycles/kernel/cycles_kernel_oneapi.lib', needed by
'bin/blender.exe', missing and no known rule to make it".
MSVC Tools version doesn't match MSVC Redist version on some systems and
it's not populated when using Ninja outside of Visual Studio shell,
trying another way.
DDS files coming through OIIO needed a similar treatment as TGA in
T99565; just for DDS OIIO just never set the "unassociated alpha"
attribute. Fixes T101850.
Reviewed By: Brecht Van Lommel
Differential Revision: https://developer.blender.org/D16270
This was a floating point precision issue - or, to be more precise,
an issue with how Cycles split floats into the integer and fractional
parts for Perlin noise.
For coordinates below -2^24, the integer could be wrong, leading to
the fractional part being outside of 0-1 range, which breaks all sorts
of other things. 2^24 sounds like a lot, but due to how the detail
octaves work, it's not that hard to reach when combined with a large
scale.
Since this code is originally based on OSL, I checked if they changed
it in the meantime, and sure enough, there's a fix for it:
https://github.com/OpenImageIO/oiio/commit/5c9dc68391e9
So, this basically just ports over that change to Cycles.
The original code mentions being faster, but as pointed out in the
linked commit, the performance impact is actually irrelevant.
I also checked in a simple scene with eight Noise textures at
detail 15 (with >90% of render time being spent on the noise), and
the render time went from 13.06sec to 13.05sec. So, yeah, no issue.
Test kernel will now test functionalities related to kernel execution
with USM memory allocations instead of with SYCL buffers and accessors
as these aren't currently used in the backend.
This change removes CMake code for automatic calculation of the number
of offline device compiler instances, to hand over control to developers
instead as it incurs a rather large memory usage with around 8GB per
instance at peak.
Use SYCL_OFFLINE_COMPILER_PARALLEL_JOBS CMake variable to configure it.
Currently Cycles uses zero as a default radius when no "radius"
attribute exists. This is more confusing than helpful. Instead,
use the same default radius as EEVEE and Workbench.
This patch enables MNEE on macOS >= 13. There was an inefficiency in the calculation of spill requirements, fixed as of macOS 13. This patch also adds a temporary inlining workaround for a Metal compiler bug which causes `mnee_compute_constraint_derivatives` to behave incorrectly.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D16235
The issue here was that PathTraceWork was set up before checking if
any error occurred, and it didn't account for the dummy device so
it called a non-implemented function.
This fix therefore avoids creating PathTraceWork for dummy devices
and checks for device creation errors earlier in the process.
JIT compilation of oneAPI kernels now happens during load stage
and proper message gets shown in the GUI during compilation.
Also, this implementation skips kernels that aren't needed for
the used scene, reducing overall (re)compilation time.
OSL (like Cycles) has no internal boolean type, instead an integer
input can be flagged to be shown as a boolean in the UI.
Cycles reacts to this by creating a boolean socket on the Blender
side, but as a result incorrectly called the boolean overload of the
set function even though the internal type is an integer.
There's another unrelated crash in the GPU viewport shader code that
appears to apply to every OSL node that outputs a shader, and the file
in T101702 triggers both, so this is only a partial fix for the report.
The attribute node already allows accessing attributes associated
with objects and meshes, which allows changing the behavior of the
same material between different objects or instances. The same idea
can be extended to an even more global level of layers and scenes.
Currently view layers provide an option to replace all materials
with a different one. However, since the same material will be applied
to all objects in the layer, varying the behavior between layers while
preserving distinct materials requires duplicating objects.
Providing access to properties of layers and scenes via the attribute
node enables making materials with built-in switches or settings that
can be controlled globally at the view layer level. This is probably
most useful for complex NPR shading and compositing. Like with objects,
the node can also access built-in scene properties, like render resolution
or FOV of the active camera. Lookup is also attempted in World, similar
to how the Object mode checks the Mesh datablock.
In Cycles this mode is implemented by replacing the attribute node with
the attribute value during sync, allowing constant folding to take the
values into account. This means however that materials that use this
feature have to be re-synced upon any changes to scene, world or camera.
The Eevee version uses a new uniform buffer containing a sorted array
mapping name hashes to values, with binary search lookup. The array
is limited to 512 entries, which is effectively limitless even
considering it is shared by all materials in the scene; it is also
just 16KB of memory so no point trying to optimize further.
The buffer has to be rebuilt when new attributes are detected in a
material, so the draw engine keeps a table of recently seen attribute
names to minimize the chance of extra rebuilds mid-draw.
Differential Revision: https://developer.blender.org/D15941
It fixes SYCL runtime issues in Debug builds that were due to mixing
Release and Debug MSVC runtimes.
This commit also removes specific handling of dpcpp compiler executable
to simplify the CMake implementation. Using it like clang++ works and
clang++ executable is also available from Intel oneAPI DPC++ compiler in
case it doesn't.
This is a minimal set of changes, allowing a lot of cleanup that can
happen afterward as it allows sycl method and objects to be used outside
of kernel.cpp.
Reviewed By: brecht, sergey
Differential Revision: https://developer.blender.org/D15397
Changing volume parameters during rendering could cause a crash
when guiding was enabled. It was due to an unintialized state paramter
at the beginning of the path tracing process.
In addition guiding is disabled when dealing with almost delta volumes
(i.e., g close to 1.0 or -1.0).
To avoid issues with install_deps. If we more generally switch to using
CMake configs then perhaps this code can be deduplicated again or at
least simplified.
This change speeds up the compilation at the cost of higher memory usage.
CMake implementation checks the amount of available memory to spawn a
reasonable number of parallel compiler jobs.
