2014-08-28 08:41, Richardson, Bruce:
As for rte_rdtsc_precise, I'm not sure about its origins, but I'm
surprised to see that it does not correspond to the rdtscp instruction.
Can anyone else comment on this one? I would assume it's designed
to be used to get more accurate measurements of smaller blocks of
code that we want to benchmark, since rdtsc works best when timing
larger blocks (in terms of cycle counts, that is, not source lines :-) ).
The good thing with git history (and well written commit logs) is that
we can easily get such answer:
http://dpdk.org/browse/dpdk/commit/?id=3314648f83c3dc06d7d9a
Bruce, do you know how rdtscp is supported across Intel processors?
--
Thomas
-----Original Message-----
From: Thomas Monjalon [mailto:thomas.monjalon-pdR9zngts4EAvxtiuMwx3w@public.gmane.org]
Sent: Thursday, August 28, 2014 10:06 AM
To: Richardson, Bruce
Cc: Cyril Chemparathy; dev-VfR2kkLFssw@public.gmane.org
Subject: Re: [dpdk-dev] next releases
2014-08-28 08:41, Richardson, Bruce:
quoted
As for rte_rdtsc_precise, I'm not sure about its origins, but I'm
surprised to see that it does not correspond to the rdtscp instruction.
Can anyone else comment on this one? I would assume it's designed
to be used to get more accurate measurements of smaller blocks of
code that we want to benchmark, since rdtsc works best when timing
larger blocks (in terms of cycle counts, that is, not source lines :-) ).
The good thing with git history (and well written commit logs) is that
we can easily get such answer:
http://dpdk.org/browse/dpdk/commit/?id=3314648f83c3dc06d7d9a
Indeed! I had mistakenly assumed that the function was older than the dpdk.org history. My bad on that one, and thanks for the link. :-)
Bruce, do you know how rdtscp is supported across Intel processors?
Sorry, no I don't, and the instruction set manuals only say to check the relevant cpuid bit to see if it's supported. :-(. If the existing code works ok, there is probably no compelling reason to look to change it though.
--
Thomas