Various fixes in the HIP-RT BVH building related on making sure
curves motion blur is supported and is working correctly, as well
as properly handle motion pass configuration when path tracing is
to ignore motion blur (and instead write vector pass).
This PR contains #134797 with fixes needed to fully finish it:
moving commits from that PR here made it easier to ensure all
moving parts are tested without mental overhead.
Fixes#134510
Co-authored-by: Sahar A. Kashi <sahar.alipourkashi@amd.com>
Co-authored-by: Sergey Sharybin <sergey@blender.org>
Co-authored-by: Brecht Van Lommel <brecht@blender.org>
Pull Request: https://projects.blender.org/blender/blender/pulls/135125
Rewrite the ONEAPI Blender texture allocation code to make use of
1D images backed by linear USM memory. This increases parity
with the CUDA implementation and sets the ground work for enabling
host USM allocations in Blender. By enabling this functionality,
previously failing benchmarks are now passing.
Together with the previous commit, no functional changes are expected.
These have bugs in with the latest HIP-RT and HIP SDK, so just disable them
as we do not expect a fix in time, and rolling back would re-introduce other
bugs. As RDNA1 does not have hardware raytracing, it is also less important
to use HIP-RT.
Note that only RDNA2+ is officially supported by HIP, so these GPUs working
at all is somewhat lucky.
Fix#134979Fix#134978Fix#134975
Pull Request: https://projects.blender.org/blender/blender/pulls/135179
After the recent HIP SDK 6.3 update on Windows, the minimum GPU driver
required to use HIP in Cycles has increased.
This commit increases the required driver version listed in the UI and
adds a check to avoid showing HIP devices if they're below a certain
driver version number as they don't work properly.
Pull Request: https://projects.blender.org/blender/blender/pulls/134965
This change brings the following improvements on the user level
- Support of GPUs with gfx12 architecture
- New HIP-RT library which in addition to the gfx12 support brings
various bug-fixes.
The known limitation of gfx12 is that OpenImageDenoiser does not yet
support this GPU architecture. This means that while Cycles will use the
full advantage of the gfx12 (including hardware accelerated ray-tracing),
denoising will only be possible on CPU, or secondary gfx11 or below GPU.
This is something that requires a change in OIDN and it is to late to do
it for Blender 4.4, but it is something to look forward for Blender 4.5.
The gfx12 changes for the pre-compiled kernels is rather trivial,
so it comes together (in the same PR) as the bigger HIP-RT change.
On the development side this change brings the following improvements:
- One step compile and link (much simpler CMake rules)
- Embedding BVH binaries in hiprt dll (which makes it easier to package
and load, without relying on special path configuration)
Co-authored-by: Sahar Kashi <sahar.kashi@amd.com>
Co-authored-by: Sergey Sharybin <sergey@blender.org>
Co-authored-by: Brecht Van Lommel <brecht@blender.org>
Pull Request: https://projects.blender.org/blender/blender/pulls/133129
There is a bug in Embree that makes BVH updates crash. Disabling multithreaded
BVH updates after the initial BVH build appears to work around it, at the cost
of some performance.
This will not affect performance of the initial BVH build, transforming objects
or editing a single mesh. It will only affect performance when multiple smaller
meshes are edited together, as those can no longer have their BVH updated in
parallel or benefit from parallellization over many primitives.
Pull Request: https://projects.blender.org/blender/blender/pulls/134747
free_memory queries were disabled due to runtime driver issues on Linux
when using jemalloc.
compute-runtime introduced a fix for these issues with
8527779778
which is part of versions 31740 and higher, and matches the currently
required min-driver version, so we can restore this feature.
Pull Request: https://projects.blender.org/blender/blender/pulls/134542
* Do oneAPI copy optimization as part of host memory alloc and free, so
it is properly released before host memory is freed.
* Synchronize after loading texture info, like CUDA and HIP.
https://projects.blender.org/blender/blender/pulls/134412
This may be used for device to do host memory allocation in a way that
is more efficient for copy the host memory to the device.
Also rename and group device memory allocation functions for clarity.
Pull Request: https://projects.blender.org/blender/blender/pulls/134412
The issue was caused by c0ba800f64.
Simple solution for now: check the data size and free the memory to
allow the device memory to be re-allocated. Seems the safest for the
upcoming release.
Ideally we'd need to avoid having these manual tricks with the device
memory pointers.
Pull Request: https://projects.blender.org/blender/blender/pulls/134516
The current usage of software-based texture operations in
the oneAPI implementation puts additional register pressure on
the GPU compiler during register allocation. And it also creates
code that requires maintenance. This commit is intended to address
this situation by utilizing a recently productized SYCL bindless
texture API to enable HW-based texture operations using
Intel GPUs' hardware sampler.
This currently translates to 1-11% rendering speedups (scene-specific)
on my Arc A770 and Arc B580. At the moment, there are small
performance regressions with NanoVDB texture operations on Arc B580
and small performance regressions in shade surface MNEE and Raytrace
kernels on Arc A770, but they look recoverable and will be handled
in the future.
Pull Request: https://projects.blender.org/blender/blender/pulls/133457
The issue also happens on macOS 15.3.
This is a Metal driver bug, a fix is coming in macOS 15.4. Until then
disable refitting the viewport. There is no perceptible benefit from
refitting, so while it might be less that ideal it allows to side step
the problem and still benefit from the HWRT.
Pull Request: https://projects.blender.org/blender/blender/pulls/134399
Regardless of what mem info reports. We can't move this to the host, so
might as well try because the free memory might not be a reliable predictor
of success.
Pull Request: https://projects.blender.org/blender/blender/pulls/132912
With host mapped memory these can be shared, and we can't get back the
original host pointer unless we make a copy which is inefficient.
Also add asserts to verify this doesn't happen.
Pull Request: https://projects.blender.org/blender/blender/pulls/132912
This was not thread safe. And it's better to do them one by one to avoid
moving more than is needed, when another thread already freed up enough.
Thanks to Jorn Visser for investigating and finding this problem.
Pull Request: https://projects.blender.org/blender/blender/pulls/132912
If one of the devices already used host happed memory but another not,
it would previously realloc both.
Thanks to Jorn Visser for investigating and finding this problem.
Pull Request: https://projects.blender.org/blender/blender/pulls/132912
Should be a bit more efficient, and it fixes host memory fallback bugs,
where host memory was incorrectly freed during re-copy. For the case
where memory should get reallocated on the host, a new mem_move_to_host
was added.
Thanks to Jorn Visser for investigating and finding this problem.
Pull Request: https://projects.blender.org/blender/blender/pulls/132912
