Tutorial - How To Generate A Call Graph Using xSIM and gprof

[johned]

The combination of xSIM and gprof provides a powerful combination for analyzing the runtime performance of an application.

Notes :
gprof2dot (https://github.com/jrfonseca/gprof2dot) is a python application. I downloaded the zip and extracted to the folders :

Code: Select all

Windows : %HOMEPATH%\python\
Linux : ~/python/gprof2dot

You may wish to change this path to suit.
XCORE-200-EXPLORER.xn can be found in the xTIMEcomposer targets folder.
When you run the xsim command, hit <ctrl>-C after a few seconds to halt the simulation.

The attached example can be analyzed using the following commands :

Windows

Code: Select all

xcc SimpleTimerBiDi.xc XCORE-200-EXPLORER.xn -o SimpleTimerBiDi.xe
xsim --gprof SimpleTimerBiDi.xe
xobjdump --split SimpleTimerBiDi.xe
xgprof image_n0c0_2.elf tile[0]_core0.gprof | python %HOMEPATH%\python\gprof2dot\gprof2dot.py > SimpleTimerBiDi.dot
dot -Tpng -oSimpleTimerBiDi.png SimpleTimerBiDi.dot

Linux

Code: Select all

xcc SimpleTimerBiDi.xc XCORE-200-EXPLORER.xn -o SimpleTimerBiDi.xe
xsim --gprof SimpleTimerBiDi.xe
xobjdump --split SimpleTimerBiDi.xe
xgprof image_n0c0_2.elf tile\[0\]_core0.gprof | ~/python/gprof2dot/gprof2dot.py > SimpleTimerBiDi.dot
dot -Tpng -oSimpleTimerBiDi.png SimpleTimerBiDi.dot

You will now see the call graph in the file : SimpleTimerBiDi.png.

[peter]

Hi John,

Really helpful to share this. Would you mind adding the .png output as well so that people can see what they might expect if they follow these steps. Thanks.

Also, if anyone is interested in having absolute times in their output I've found a way of tweaking gprof2dot to do it. I've added a comment on how to get absolute times here

Peter

[johned]

Here's the png

Cheers,
John

[peter]

I've produced a very basic test application:

Code: Select all

int main()
{
  double x = 1.2;
  double y = 1.5;
  double z = 19.5;
  for (unsigned i = 0; i < 100; i++) {
    y *= x;
    z += y;
  }

  return (y+z < 100.0) ? 0 : 1;
}

and produced the call graph.