I wanted to post this observation to the group to see if I get a hit.
I updated Bug #569 today with an additional observation:
Basically, on my ARM target, if I enable cpu core #1 (two cores total),
the qa_mpsk_receiver test fails about 1/3rd of the time, when it
compares 0.5 to sometimes -.492 or sometimes -.503 (i.e., the sign in
the real part of a complex result flipped). If I keep cpu #1 disabled,
the test passes every time. Keeping the CPUs limited to their minimum
speed does not help.
I’m compiling with gcc 4.7.3, and -mcpu=cortex-a15 -mfpu=neon
-mvectorize-with-neon-quad -ffast-math -funsafe-loop-optimizations
I’ll be varying the build flags to see if that makes a difference in the