Disable neon dot on Unisoc UMS312 #265
Conversation
GregoryComer
force-pushed
the
ums312-dot-disable
branch
from
a72f52c
to
5ca8ae1
Compare
| @@ -0,0 +1,31 @@ | |||
| # begin build properties | |||
| # autogenerated by buildinfo.sh | |||
There was a problem hiding this comment.
It should be located under /system/build.props. However, I can't read it without root, so I pulled the template file and populated this by running adb shell getprop. I can remove this file, if desired.
| @@ -147,6 +147,8 @@ void cpuinfo_arm_linux_decode_isa_from_proc_cpuinfo( | |||
| "VDOT instructions disabled: cause occasional SIGILL on Spreadtrum SC9863A"); | |||
| } else if (chipset->series == cpuinfo_arm_chipset_series_unisoc_t && chipset->model == 310) { | |||
| cpuinfo_log_warning("VDOT instructions disabled: cause occasional SIGILL on Unisoc T310"); | |||
| } else if (chipset->series == cpuinfo_arm_chipset_series_unisoc_ums && chipset->model == 312) { | |||
There was a problem hiding this comment.
we you able to repro this on armv8 kernel? And does this fix it? Ok with this change even if you couldn't TBH.
There was a problem hiding this comment.
Unfortunately. I don't have a device with UMS312 and armv8 kernel. From the crash logs, they were mostly (maybe all?) armv7 kernels, so it may be okay. If we see issues on armv8 kernels, I can take as a follow-up.
Yeah, I confirmed that it is working as intended on the Astro 55R by checking cpuinfo debug logging. With this change, it does disable dot instructions. |
This PR adds logic to recognize Unisoc UMS312 devices and disable neon dot instructions. This is due to SIGILL crashes on UMS312-based phones. We already disable dot instructions on the Unisoc T310, so this is not unprecedented. It is unclear to me if the UMS312 is actually a SOC or if is a board/platform including the T310. Regardless, it is identified separately, so I have added logic to handle it.
I've also added cpuinfo and ro.props for the Astro 55R, which is a UMS312-based device.
To validate this change, I added tests for matching T301 and UMS312 hardware strings and ran all tests on the Astro 55R. I noticed that there are a large number of existing failing tests, so I re-ran all tests on main and diffed the results to confirm that there are no regressions. I also built XNNPACK with this CPUINFO commit and validated that dot instructions were disabled via observing the relevant log message.