This fixes issue [#105100](https://projects.blender.org/blender/blender/issues/105100) where multi-pass renders can be incorrect due to kernels using stale specialisation constants (e.g. when rendering Pokedstudio).
This patch adds a new group of md5 hashes (`global_defines_md5`) to track whether the injected block of #defines is stale and regenerate the source string as appropriate. It also renames the existing group of md5 hashes from `source_md5` to `kernels_md5` to clarify that these refer to a specific kernel set rather than just the source (which might build an arbitrarily large number of kernel sets).
Pull Request #105103
The internal compiler error appears to be gone. Unclear why it appeared in the
first place and why it's gone now. Just random kernel code changes causing it.
Pull Request #104719
(Follow on from D17043)
On AMD Navi2 devices the MetalRT checkbox was not hooked up properly and had no effect. This patch fixes it.
Co-authored-by: Michael Jones <michael_p_jones@apple.com>
Pull Request #104520
Host memory fallback in CUDA and HIP devices is almost identical.
We remove duplicated code and create a shared generic version that
other devices (oneAPI) will be able to use.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D17173
This patch optimises subsurface intersection queries on MetalRT. Currently intersect_local traverses from the scene root, retrospectively discarding all non-local hits. Using a lookup of bottom level acceleration structures, we can explicitly query only the relevant instance. On M1 Max, with MetalRT selected, this can give a render speedup of 15-20% for scenes like Monster which make heavy use of subsurface scattering.
Patch authored by Marco Giordano.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D17153
This patch adds two new kernels: SORT_BUCKET_PASS and SORT_WRITE_PASS. These replace PREFIX_SUM and SORTED_PATHS_ARRAY on supported devices (currently implemented on Metal, but will be trivial to enable on the other backends). The new kernels exploit sort partitioning (see D15331) by sorting each partition separately using local atomics. This can give an overall render speedup of 2-3% depending on architecture. As before, we fall back to the original non-partitioned sorting when the shader count is "too high".
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D16909
This patch removes the option to select both AMD and Intel GPUs on system that have both. Currently both devices will be selected by default which results in crashes and other poorly understood behaviour. This patch adds precedence for using any discrete AMD GPU over an integrated Intel one. This can be overridden with CYCLES_METAL_FORCE_INTEL.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D17166
This patch fixes T103393 by undefining `__LIGHT_TREE__` on Metal/AMD as it has an unexpected & major impact on performance even when light trees are not in use.
Patch authored by Prakash Kamliya.
Reviewed By: brecht
Maniphest Tasks: T103393
Differential Revision: https://developer.blender.org/D17167
A noticeable (>5%) performance regression in oneAPI backend came with
a501a2dbff. Updating to latest graphics
compiler from driver 101.4032 fixes it.
I've tested it with current min-supported drivers and it runs well but
since compatibility of graphics compiler with older drivers isn't
guaranteed, I'm also bumping the min-supported driver versions.
If end-users consider latest drivers too fresh to switch to (version
isn't released as stable on Linux as of today but should be before
Blender 3.5 release), CYCLES_ONEAPI_ALL_DEVICES=1 env variable can be
used.
Intel Graphics Compiler on Linux will be updated in a later commit
so we can then close D16984.
Reviewed By: sergey, LazyDodo
This patch adds markup to specify that certain kernel data constants should not be specialised. Currently it is used for `tabulated_sobol_sequence_size` and `sobol_index_mask` which change frequently based on the aa sample count, trash the shader cache, and have little bearing on performance.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D16968
This patch adds occupancy tuning for the newly announced high-end M2 machines, giving 10-15% render speedup over a pre-tuned build.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D17037
While keeping SSE2, SSE4.1 and AVX2. This does not affect hardware support, it
only slightly reduces performance for some older CPUs.
To reduce maintenance cost and improve compile times.
Differential Revision: https://developer.blender.org/D16978
This patch adds a new "Kernel Optimization Level" dropdown menu to control Metal kernel specialisation. Currently this defaults to "full" optimisation, on the assumption that the changes proposed in D16371 will address usability concerns around app responsiveness and shader cache housekeeping.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D16514
All kernel specialisation is now performed in the background regardless of kernel type, meaning that the first render will be visible a few seconds sooner. The only exception is during benchmark warm up, in which case we wait for all kernels to be cached. When stopping a render, we call a new `cancel()` method on the device which causes any outstanding compilation work to be cancelled, and we destroy the device in a detached thread so that any stale queued compilations can be safely purged without blocking the UI for longer than necessary.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D16371
This functionality is related only to debugging of SYCL implementation
via single-threaded CPU execution and is disabled by default.
Host device has been deprecated in SYCL 2020 spec and we removed it
in 305b92e05f.
Since this is still very useful for debugging, we're restoring a
similar functionality here through SYCL 2020 Host Task.
Materials without connections to the output node would crash with OSL
in OptiX, since the Cycles `OSLCompiler` generates an empty shader
group reference for them, which resulted in the OptiX device
implementation setting an empty SBT entry for the corresponding direct
callables, which then crashed when calling those direct callables was
attempted in `osl_eval_nodes`. This fixes that by setting the SBT entries
for empty shader groups to a dummy direct callable that does nothing.
Switching viewport denoising causes kernels to be reloaded with a new
feature mask, which would destroy the existing OptiX pipelines. But OSL
kernels were not reloaded as well, leaving the shading pipeline
uninitialized and therefore causing an error when it is later attempted to
execute it. This fixes that by ensuring OSL kernels are always reloaded
when the normal kernels are too.
This adds a new mirror image extension type for shaders and
geometry nodes (next to the existing repeat, extend and clip
options).
See D16432 for a more detailed explanation of `wrap_mirror`.
This also adds a new sampler flag `GPU_SAMPLER_MIRROR_REPEAT`.
It acts as a modifier to `GPU_SAMPLER_REPEAT`, so any `REPEAT`
flag must be set for the `MIRROR` flag to have an effect.
Differential Revision: https://developer.blender.org/D16432
* preempt_attr was copied from CUDA, but not used in HIP.
* Remove shadowed variable before conditional in EnvironmentTextureNode code.
Differential Revision: https://developer.blender.org/D16741
This patch adds a new `max_working_set_exceeded()` check on Metal so that we can display a "System is out of GPU memory" message to the user. Without this, we get obtuse "CommandBuffer failed" errors at render time due to exceeding the size limit of resident resources.
Likely fix for T101787 & T102786.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D16713
Uses a light tree to more effectively sample scenes with many lights. This can
significantly reduce noise, at the cost of a somewhat longer render time per
sample.
Light tree sampling is enabled by default. It can be disabled in the Sampling >
Lights panel. Scenes using light clamping or ray visibility tricks may render
different as these are biased techniques that depend on the sampling strategy.
The implementation is currently disabled on AMD HIP. This is planned to be fixed
before the release.
Implementation by Jeffrey Liu, Weizhen Huang, Alaska and Brecht Van Lommel.
Ref T77889
Cycles already treats denoising fairly separate in its code, with a
dedicated `Denoiser` base class used to describe denoising
behavior. That class has been fully implemented for OIDN
(`denoiser_oidn.cpp`), but for OptiX was mostly empty
(`denoiser_optix.cpp`) and denoising was instead implemented in
the OptiX device. That meant denoising code was split over various
files and directories, making it a bit awkward to work with. This
patch moves the OptiX denoising implementation into the existing
`OptiXDenoiser` class, so that everything is in one place. There are
no functional changes, code has been mostly moved as-is. To
retain support for potential other denoiser implementations based
on a GPU device in the future, the `DeviceDenoiser` base class was
kept and slightly extended (and its file renamed to
`denoiser_gpu.cpp` to follow similar naming rules as
`path_trace_work_*.cpp`).
Differential Revision: https://developer.blender.org/D16502
This patch adapts D14754 for the Metal backend. Kernels of the same type are already organised into subdirectories which simplifies type matching.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D16469
This patch generalizes the OSL support in Cycles to include GPU
device types and adds an implementation for that in the OptiX
device. There are some caveats still, including simplified texturing
due to lack of OIIO on the GPU and a few missing OSL intrinsics.
Note that this is incomplete and missing an update to the OSL
library before being enabled! The implementation is already
committed now to simplify further development.
Maniphest Tasks: T101222
Differential Revision: https://developer.blender.org/D15902
If no OPTIX_ROOT is set, nvcc fails to compile because there is a stray "-I"
in the arguments. Detect if the include path is empty and act accordingly.
Differential Revision: https://developer.blender.org/D16308
This patch tunes maximum threads-per-threadgroup and threads-per-block for faster renders on Apple GPUs. Appropriate tuning is selected based on the GPU architecture (M1 or M2). We see a benchmark uplift of around 5-10% on M1 family chips. Similar uplift is expected on M2 with upcoming OS changes. (Ref T101931)
Reviewed By: brecht
Maniphest Tasks: T101931
Differential Revision: https://developer.blender.org/D16299
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
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
