Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR adds ARM64 support by way of the reference implementations.
NOTE: This goes in tandem with RenderKit/mkl-dnn#1
Performance is ....not great.
From oidnBenchmark.exe:
Windows ARM64 (using clang-cl 16.0.6):
RT.hdr_alb_nrm.1920x1080 ... 1.04168e+06 msec/image
Linux ARM64 (using GCC 12.2.0):
RT.hdr_alb_nrm.1920x1080 ... 635797 msec/image
Windows Emulated x64 (using release fom github):
RT.hdr_alb_nrm.1920x1080 ... 4174.18 msec/image
Command lines used:
Windows (within a vcvarsall ARM64 window, VS has clang installed as part of it):
Linux:
And then both built with:
Clearly, there is serious need for improvement here (250x slowdown!). What are the options? Would a PR of say, DirectML enablement, for Windows be acceptable?
I tried enabling Arm CL, which (after some faff) compiled and linked in, but none of the implementations seemed to be a match - it always fell back to the reference implementation, and if that was disabled, a segfault occurred.
The other ARM64 jit implementations present are not applicable in this case, as they use SVE512 (which is not widely implemented yet outside of a few select/recent CPUs).