Forum: GNU Radio Profiling with oprofile

Announcement (2017-05-07): www.ruby-forum.com is now read-only since I unfortunately do not have the time to support and maintain the forum any more. Please see rubyonrails.org/community and ruby-lang.org/en/community for other Rails- und Ruby-related community platforms.
3e5d1707dd5f4807ed9938d5c37b2649?d=identicon&s=25 Patrick Strasser (Guest)
on 2006-06-07 21:22
(Received via mailing list)
Hello group!

I tried oprofile with usrp_wfm_rx_nogui.py [1] and got interresting
results.

= Preparations =

First start oprofile as root (required) with

# oprof_start

Disable the kernel image (only needed for kerenl profiling which we do
not), enable "per application profiles" and let the remaining options
disabled and start the profiler.
On a second shell prepare (as root again) a call to opcontrol to get a
exact shot of my symple run:

# opcontrol --session=some_session_name

This will save all previous data to session "some_session_name". To be
exact, the generic session "current" will be saved to that name, all
counters reset and a new session "current" will be created.
In a third shell (screen is your friend) prepare to start the FM
receiver:

$ ./usrp_wfm_rcv_nogui.py -f89.2

Note: Until now we have a running profiler which dumps data about the
whole system to the "current" session, a prepared command to create a
new session an a command line for a program ready to run.

= Profiling =

Profiling is easy:
* Start the programm
* Start a new session: # opcontrol --session=some_session_name
* Listen and/or watch GNURadio do its work for a reasonable time, say 1
minute [2]
* Save all your recorded data in new session: # opcontrol
--session=gnuradio_wfm_nogui_1
* Stop the programm (^C)

= Analysis =

== opreport ==

opreport gives us info about the data we collected. Without a session
the "current" session is used, other sessions can be examined with
appended "session:my_session_name". To get statistics for a certain
binary, just add the path to that binary. For GNURadio the binary is
python which calls GNURadio as library. The option -r reverses sorting,
printing most active parts last. The option -l lists all symbols with
the corresponding data (performance counters and percentage)

$ opreport  -rl session:gnuradio_wfm_nogui_1 /usr/lib/python

== opannotate ==

Another really useful tool: opannotate. It locates the corresponding
source code via symbol tables and adds information about runtime.
It can produce annotated assembler code, or, when the binaries are not
stripped, annotated C/C++ code. Quite nice. Just add -o
my_output_directory to tell it where to store the annotated code (no, it
wont mess with the originals :-)), and of course the binary, we want our
date to be relative to the program that was running, not the whole
system:

$ mkdir annotated
$ opannotate -D smart session:gnuradio_wfm_nogui_1 --source -o annotated
/usr/bin/python

VoilĂ ! Have a look at the files in the annotated directory. You'll find
something like that from
gr-build/gnuradio-core/src/lib/filter/float_dotprod_sse.S :

                :.loop2:
    414  1.3286 :        mulps   (%edx), %xmm0 /* .loop2 total:   7982
25.6154 */  1054  3.3824 :        addps   %xmm2, %xmm6
     71  0.2278 :        movaps  0x20(%eax), %xmm2
                :
    587  1.8838 :        mulps   0x10(%edx), %xmm1
   1318  4.2296 :        addps   %xmm3, %xmm7
    159  0.5103 :        movaps  0x30(%eax), %xmm3
                :
    334  1.0719 :        mulps   0x20(%edx), %xmm2
   1261  4.0467 :        addps   %xmm0, %xmm4
    262  0.8408 :        movaps  0x40(%eax), %xmm0
                :
    203  0.6515 :        mulps   0x30(%edx), %xmm3

and so on.

You see, .loop2 in float_dotprod_sse.S has quite a lot to do.

= Interpreting data =

oprofile is a statistical profiler, so the gathered data is not exact.
You probably don't want to rely too much on fractional portions of
numbers given. But it gives a good overview.

= Results from ./usrp_wfm_rcv_nogui.py =

I let ./usrp_wfm_rcv_nogui.py run for one minute, and the most
interresting symbols where:

1929      6.1904  libgnuradio-core.so.0.0.0 .loop1
2503      8.0325  libgnuradio-core.so.0.0.0
gr_fir_ccf_simd::filter(std::complex<float> const*)
5656     18.1509  libgnuradio-core.so.0.0.0 .loop2
5720     18.3563  libgnuradio-core.so.0.0.0 gr_fast_atan2f(float, float)
7982     25.6154  libgnuradio-core.so.0.0.0 .loop2

Which add up to some 76 percent of run time.

* .loop1 is in gnuradio-core/src/lib/filter/fcomplex_dotprod_sse.S
* gr_fir_ccf_simd::filter(std::complex<float> const*) is in
gnuradio-core/src/lib/filter/gr_fir_ccf_simd.cc
* first .loop2 (18.1509) is in
gnuradio-core/src/lib/filter/fcomplex_dotprod_sse.S
* gr_fast_atan2f(float, float) is in
gnuradio-core/src/lib/general/gr_fast_atan2f.cc
* second .loop2 () is in
gnuradio-core/src/lib/filter/fcomplex_dotprod_sse.S

Patrick

[1] the FM receiver that runs on my PIII 450 without hiccups...
[2] oprofile is a statistical profiling system: It has a look at the
system from time to time, resulting in _statistical_ data, preferably
from a long time period to get reasonable good results.
--
Engineers motto: cheap, good, fast: choose any two
Patrick Strasser <patrick dot strasser at  tugraz dot at>
Student of Telematik, Techn. University Graz, Austria
745d8202ef5a58c1058d0e5395a78f9c?d=identicon&s=25 Eric Blossom (Guest)
on 2006-06-07 22:50
(Received via mailing list)
On Wed, Jun 07, 2006 at 09:20:13PM +0200, Patrick Strasser wrote:
> Hello group!
>
> I tried oprofile with usrp_wfm_rx_nogui.py [1] and got interresting results.
>

Thanks for the nice write up.  Oprofile is a great tool.

Eric
This topic is locked and can not be replied to.