benchdnn and rdpmc

[kruus]

I set out this week to run (and then, of course, port) benchdnn/ timings.
executing the rdpmc instruction was always segfaulting for me on Ubuntu 16.04,
kernel 4.4.0-93-generic.

I worked around by following kernel tools "perf" code. After opening a fd and mmap,
the rdpmc instruction was OK to use, and gave reasonable timing values.

Q1: Is there a reason to avoid usual rdtsc (optionally w/ cpuid to flush pipeline)?
Q2: Is there simpler "magic" to get rdpmc to work? (after all, I do not really even have
to use the mmaped page, because you're using the 'magic number' approach in your
rdpmc routine.)

Note3: If you did want to use the mmap'ed perf page, it seems it gives you the wall-time
conversion algorithm for free.

[vpirogov]

I'm summoning @emfomenk to this thread.

[kruus]

Well, benchdnn compiling but segfaulting on some linux systems might be a bug. Granted it is not critical, because it is in a tests/benchdnn/ subdirectory and can be fixed in several ways.

[vpirogov]

Fixed in 8d9fa58