freenode :: #aomedia

12 Jul 2017
13:21xiphmont_nofishWhat is a good way to run the ASAN tests locally? I'm getting failures due to a test that's configured not to run, need to figure out why.
13:29ltrudeauxiphmont_nofish: I use the following script the --sanitize option runs both ASAN and UBSAN
13:38xiphmont_nofishltrudeau: thanks
13:40ltrudeauxiphmont_nofish:This is where I got the values for CFlags
13:58xiphmont_nofishis the output logged? Or do I need to rerun with a redirection?
14:04ltrudeauxiphmont_nofish: output should be in stdout/stderr
14:04xiphmont_nofish[ RUN ] SSE2/AV1Trans4x4HT.MemCheck/9
14:04xiphmont_nofish./test/transform_test_base.h:259: Failure
14:05xiphmont_nofishthat claims to be test_shard.0
14:05xiphmont_nofishstill deciphering the build system, not sure what code is getting run and with what flags
14:05ltrudeauoh I have never tried sanitizers on the tests
14:05xiphmont_nofishi'm getting fails in the SSE. Which I've not touched.
14:06ltrudeaudo you have a link for your build?
14:06xiphmont_nofishof course, I might have accidentally clobbered some input
14:07xiphmont_nofishthe output is useful, I just need to track down what is actually triggering it.
14:08ltrudeauconsole output gives more info:
14:09xiphmont_nofishyes, read it--- but that test should not even be run for my changes. That's a fail in the stock code.
14:09xiphmont_nofish[and the stock code is disabled by my change-- all of MMX/SSE/AVX is turned off]
14:09xiphmont_nofishso it's testing the default config.
14:09xiphmont_nofishand that's what's fialing.
14:10xiphmont_nofishI want to know what piece of code that is actually fingering.
14:10xiphmont_nofishie, what's actually being tested
14:10xiphmont_nofishthe name could signify one of several codepaths
14:11xiphmont_nofish"Knowing is half the battle"
14:11xiphmont_nofisheven narrowing it down to something as simple as 'is that the forward or inverse transform path fialing?'
14:12xiphmont_nofishI believe it is telling the truth, once I understand what it is telling :-)
14:12ltrudeaucan I have the link to your patch?
14:13xiphmont_nofishNote that it assumes 7960 as well (which passes tests and is waiting to be re- +2ed
14:14ltrudeauI think this is your problem
14:15xiphmont_nofishoh, it got changed again?
14:16ltrudeaumaybe more the return statement
14:16xiphmont_nofishit was a style nit fix
14:21xiphmont_nofishit is functionally equivalent to the code path that was there and passes in 7960
14:23xiphmont_nofishOoooh, I wonder if someone got lazy.....
14:24ltrudeauyou are right I misread that
14:24xiphmont_nofishnope, that wasn't it
14:26xiphmont_nofishI bet the SSE and C versions of the fwd transform no longer agree.
14:27xiphmont_nofishno, wait, that's preexisting too and passed
14:27ltrudeauit appears to be an off by 1 in 4X4
14:27xiphmont_nofishwhen my code isn't even running
14:27xiphmont_nofishso an accidental funcitonal change in the existing C
14:28xiphmont_nofishfoind it
14:29xiphmont_nofishyup, that's it.
14:29xiphmont_nofishaccidentally reoved one of the google rounding hacks that triggered in 4x4.
14:35xiphmont_nofishhooray for unit tests.
13 Jul 2017
No messages
Last message: 74 days and 5 hours ago