|
From: | Nowlan, Sean |
Subject: | Re: [Discuss-gnuradio] [VOLK] volk_profile failures |
Date: | Sat, 6 Jun 2015 20:51:52 +0000 |
That makes sense. I installed ORC 0.4.23 and tried to recompile GR, but the build is failing on Volk. GCC segfaults a couple times before failing with an error on volk_16ic_magnitude_16i_a_orc_impl. Full output here: https://gist.github.com/nowls/2f6ddf5734e152ca21a7 ORC 0.4.23 was built with: CFLAGS="-mcpu=cortex-a15.cortex-a7 -mtune=cortex-a15.cortex-a7 -mfpu=neon-vfpv4 -mfloat-abi=hard -funsafe-math-optimizations" CPPFLAGS="-mcpu=cortex-a15.cortex-a7 -mtune=cortex-a15.cortex-a7 -mfpu=neon-vfpv4 -mfloat-abi=hard -funsafe-math-optimizations"
UHD built fine w/ ORC 0.4.23.
Sean
From: West, Nathan <address@hidden>
Sent: Saturday, June 6, 2015 12:57 PM To: Nowlan, Sean Cc: address@hidden Subject: Re: [Discuss-gnuradio] [VOLK] volk_profile failures Most likely.
What we've seen before is that ORC didn't preserve registers according to ARM specifications. This basically results in unpredictable data corruption. The commit I pointed to is the fix (vpush certain registers if used, then vpop after the routine).The visible corruption is seen by corrupting the test tolerance. If you look at ctest -V you should see that the "error"s are actually the same result, but since tolerance gets corrupted the test erroneously is false. The only QA failures should be on kernels that have an ORC protokernel, but this bug can affect non-ORC protokernels for that kernel. On Sat, Jun 6, 2015 at 12:47 PM, Nowlan, Sean
<address@hidden> wrote:
|
[Prev in Thread] | Current Thread | [Next in Thread] |