Question about optimizations related to profiler

[mickey_mouse]

Hi all
I have only one simple question: when I profile mi application with visual profiler and I look at the “Profiler Counter Plot” I see that the most of the time spent in executing my kernel is used for “instructions” (about 80%). The second place is occupied by “branches” (about 10%).
My question is: to optimizing my application I need to focus on branches? Or, in other words, the “instructions” counter says only how many time is spent in processing data, am I right?

thanks for your answers
mickey

[Sanjiv.Satoor]

Hi all

I have only one simple question: when I profile mi application with visual profiler and I look at the “Profiler Counter Plot” I see that the most of the time spent in executing my kernel is used for “instructions” (about 80%). The second place is occupied by “branches” (about 10%).

My question is: to optimizing my application I need to focus on branches? Or, in other words, the “instructions” counter says only how many time is spent in processing data, am I right?

thanks for your answers

mickey

The percentage values in “Profiler Counter Plot” do not represent time spent for that activity.

Percentage for a profiler counter_x = ((count for profiler counter_x) / (sum of counts for all profiler counters)) * 100.0

[benry]

It is not very clear to me why to report this value as a percentage… How should this percentages be used to optimize the code? Is it bad to have a lot of branches? If they are not divergent I thought it should not be a problem, right?