It should consistently use the Cycles pirmitive ID for self intersection detection,
not the one from the OptiX or Embree acceleration structure.
Differential Revision: https://developer.blender.org/D15632
Recently, performance with oneAPI have regressed due some recent
changes in Blender itself. This commit's changes is resolving this
and also improve compilation time for oneAPI backend first
execution (or Blender compilation time in case of AoT).
Regression have appeared after 5152c7c152 and not related to the
changes itself, but increase of kernels complexity introduced with
it. Changes in this commit is marking some Blender functions as
noinlined for oneAPI backend, which helps GPU compiler to deal with
this complexity without any negative side-effects on performance.
Simplifies intersection code a little and slightly improves precision regarding
self intersection.
The parametric texture coordinate in shader nodes is still the same as before
for compatibility.
The number of Execution Units and resident "threads" (simd width * threads
per EUs) are now exposed and used to select the number of states using
a simplified heuristic.
float8 is a reserved type in Metal, but is not implemented. So rename to
float8_t for now.
Also move back intersection handlers to kernel.metal, they can't be in the
class that encapsulates the other Metal kernel functions.
This was tested in some places to check if code was being compiled for the
CPU, however this is only defined in the kernel. Checking __KERNEL_GPU__
always works.
Having the OptiX/MetalRT/Embree/MetalRT implementations all in one file with
many #ifdefs became too confusing. Instead split it up per device, and also
move it together with device specific hit/filter/intersect functions and
associated data types.
All our intersections functions now work with unnormalized ray direction,
which means we no longer need to transform ray distance between world and
object space, they can all remain in world space.
There doesn't seem to be any real performance difference one way or the
other, but it does simplify the code.
For transparency, volume and light intersection rays, adjust these distances
rather than the ray start position. This way we increment the start distance
by the smallest possible float increment to avoid self intersections, and be
sure it works as the distance compared to be will be exactly the same as
before, due to the ray start position and direction remaining the same.
Fix T98764, T96537, hair ray tracing precision issues.
Differential Revision: https://developer.blender.org/D15455
The Metal backend now compiles and caches a second set of kernels which are
optimized for scene contents, enabled for Apple Silicon.
The implementation supports doing this both for intersection and shading
kernels. However this is currently only enabled for intersection kernels that
are quick to compile, and already give a good speedup. Enabling this for
shading kernels would be faster still, however this also causes a long wait
times and would need a good user interface to control this.
M1 Max samples per minute (macOS 13.0):
PSO_GENERIC PSO_SPECIALIZED_INTERSECT PSO_SPECIALIZED_SHADE
barbershop_interior 83.4 89.5 93.7
bmw27 1486.1 1671.0 1825.8
classroom 175.2 196.8 206.3
fishy_cat 674.2 704.3 719.3
junkshop 205.4 212.0 257.7
koro 310.1 336.1 342.8
monster 376.7 418.6 424.1
pabellon 273.5 325.4 339.8
sponza 830.6 929.6 1142.4
victor 86.7 96.4 96.3
wdas_cloud 111.8 112.7 183.1
Code contributed by Jason Fielder, Morteza Mostajabodaveh and Michael Jones
Differential Revision: https://developer.blender.org/D14645
The current specific CentOS7 workaround we have for AoT, which is to
disable __FAST_MATH__ by using -fhonor-nans, now also fixes the
compilation issue for JIT as well since at least driver 23570.
with a very high min-driver version requirement, placeholder until JIT
CentOS runtime compilation issue gets fixed in a defined version.
min-driver version check can be worked around by setting
CYCLES_ONEAPI_ALL_DEVICES environment variable.
We used it only to access device id for explicitly allowing Arc GPUs.
It made the backend require ze_loader.dll which could be problematic if
we end up using direct linking.
I've replaced filtering based on PCI device id by using other HW properties
instead (EUs, threads per EU), that are now available through Level-Zero.
Initially oneAPI implementation have waited after each memory
operation, even if there was no need for this. Now, the implementation
will wait only if it is really necessary - it have improved
performance noticeble for some scenes and a bit for the rest of them.
This patch adds a new Cycles device with similar functionality to the
existing GPU devices. Kernel compilation and runtime interaction happen
via oneAPI DPC++ compiler and SYCL API.
This implementation is primarly focusing on Intel® Arc™ GPUs and other
future Intel GPUs. The first supported drivers are 101.1660 on Windows
and 22.10.22597 on Linux.
The necessary tools for compilation are:
- A SYCL compiler such as oneAPI DPC++ compiler or
https://github.com/intel/llvm
- Intel® oneAPI Level Zero which is used for low level device queries:
https://github.com/oneapi-src/level-zero
- To optionally generate prebuilt graphics binaries: Intel® Graphics
Compiler All are included in Linux precompiled libraries on svn:
https://svn.blender.org/svnroot/bf-blender/trunk/lib The same goes for
Windows precompiled binaries but for the graphics compiler, available
as "Intel® Graphics Offline Compiler for OpenCL™ Code" from
https://www.intel.com/content/www/us/en/developer/articles/tool/oneapi-standalone-components.html,
for which path can be set as OCLOC_INSTALL_DIR.
Being based on the open SYCL standard, this implementation could also be
extended to run on other compatible non-Intel hardware in the future.
Reviewed By: sergey, brecht
Differential Revision: https://developer.blender.org/D15254
Co-authored-by: Nikita Sirgienko <nikita.sirgienko@intel.com>
Co-authored-by: Stefan Werner <stefan.werner@intel.com>
Enables Vega and Vega II GPUs as well as Vega APU, using changes in HIP code
to support 64-bit waves and a new HIP SDK version.
Tested with Radeon WX9100, Radeon VII GPUs and Ryzen 7 PRO 5850U with Radeon
Graphics APU.
Ref T96740, T91571
Differential Revision: https://developer.blender.org/D15242
* Rename "texture" to "data array". This has not used textures for a long time,
there are just global memory arrays now. (On old CUDA GPUs there was a cache
for textures but not global memory, so we used to put all data in textures.)
* For CUDA and HIP, put globals in KernelParams struct like other devices.
* Drop __ prefix for data array names, no possibility for naming conflict now that
these are in a struct.
Move MNEE to own kernel, separate from shader ray-tracing. This does introduce
the limitation that a shader can't use both MNEE and AO/bevel, but that seems
like the better trade-off for now.
We can experiment with bigger kernel organization changes later.
Differential Revision: https://developer.blender.org/D15070
This patch makes it possible to change the precision with which to
store volume data in the NanoVDB data structure (as float, half, or
using variable bit quantization) via the previously unused precision
field in the volume data block.
It makes it possible to further reduce memory usage during
rendering, at a slight cost to the visual detail of a volume.
Differential Revision: https://developer.blender.org/D10023
Found those missing casts while looking into a crash report made in
the Blender Chat. Was unable to reproduce the crash, but the casts
should totally be there to avoid integer overflow.
This is a stripped down version of D14645 without the scene specialisation optimisations.
The two major changes in this patch are:
- Enables more aggressive inlining on Apple Silicon resulting in a 1.1x speedup and 10% reduction in spill, at the cost of longer pipeline build times
- Revival of shader binary archives through a new ShaderCache which is shared between MetalDevice instances using the same physical MTLDevice. This mitigates the extra compile times via explicit caching (rather than, as before, relying on the implicit system shader cache which can be purged without notice)
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D14763
This reverts commit 390b9f1305. It seems to
break things on Linux for unknown reasons, so leave it out for now. A solution
to this will be required for Vega cards though.
Explicit template specialization has to happen outside of class
definition (some compilers are more lenient). Since it is not possible to
specialize the method without also specializing the enclosing class for
all of its possible types, the method is moved outside of the class, and
specialized there.
Use templates to optimize the CPU texture sampler to interpolate using
float for single component datatypes instead of using float4 for all types.
Differential Revision: https://developer.blender.org/D14424
CPU code for cubic interpolation with clip texture extension only performed
texture interpolation inside the range of [0,1]. As a result, even though the
volume's color is sampled using cubic interpolation, the boundary is not
being interpolated. The GPU appears was interpolating samples that span the
clip boundary softening the edge, which the CPU now does also.
This commit also includes refactoring of 2D and 3D texture sampling in
preparation of adding new extension modes.
Differential Revision: https://developer.blender.org/D14295
To make porting to other architectures easier, clarifying that this does not
need to be supported. The unused parallel_reduce implementation assumed warp
size 32, but is easy to update if we ever need it in the future.
Workaround for a compilation issue preventing kernels compiling for AMD GPUs: Avoid problematic use of templates on Metal by making `gpu_parallel_active_index_array` a wrapper macro, and moving `blocksize` to be a macro parameter.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D14081
