• Locked Locked
  • Replies 1 reply
  • Subscribers 54 subscribers
  • Views 518 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

Significant variations in timing when using TSCL

Ry Gr
Prodigy 40 points

Greetings

I am using CCS 3.3 and c6455 running at 1GHz. I am employing the TSCL registers to measure the time it takes (in terms of number of cycles) to execute a function.

My code is as follows:

#include <c6x.h>

...

maxiters = 50;   

for (j=0;j<maxiters;j++)

{

start = TSCL;
stop = TSCL;
overhead = stop-start;

start_count = TSCL;
my_function;
stop_count = TSCL;

count_time = stop_count-start_count;

printf("Execution time = %d ns\n",count_time-overhead);

}

...

When running the code, the Execution time for "my_function" reported varies significantly between approximately 100,000 cycles and 511,000,000 cycles. On the contrary, the overhead is always the same and equal to 4 cycles.

Any ideas what could cause this variation? It is quite a complicated function but I don't see why the variation should vary this greatly.

Even better, is there an alternative, more accurate way of measuring the time it takes to execute a function?

Many thanks in advance

Ry

  • TI__Guru**** 171729 points

    Hi,

    Although I don't necessarily konw the intrinsics of how TSCL registers measure time (nor the intrinsics of the my_function()), I imagine two scenarios that may be happening between the start_count / stop_count lines:

    - Either the my_function() is experiencing cache hits/misses depending if data is pre-loaded from a previous run, or

    - The counter is wrapping around the 32-bit boundary and therefore giving very different readings every time the code runs (I don't see a reset of TSCL register).

    For this device you can also try to use the CCS Profile Clock - it will still impacted by the cache, though.

    I've hit both scenarios in the past (unfortunately none using the specific registers/device as yours), therefore I hope the tips may help you. I will reply to this thread if I think of any additional scenarios that may be relevant.

    Regards,

    Rafael