How to measure time IN a kernel?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi!

I want to measure the time elapsed IN a kernel, not how long the kernel execution overall is.

Something like this:

some_kernel(){

start timer

do_something1

<-----measure the time elapsed till here------->

do_something2

<-----measure the time elapsed till here------->

do_something3

<-----measure the time elapsed till here------->

}

Is there any way to do that?

2

You can use a bit of inline assembly to read the PTX cycle counter:

__device__ unsigned int get_clock(void)

{

  unsigned int clock;

  asm("mov.u32 %0, %clock;" : "=r"(clock) );

  return clock;

}

It seems to run at half the shader clock.

3

You can use a bit of inline assembly to read the PTX cycle counter:

__device__ unsigned int get_clock(void)

{

  unsigned int clock;

  asm("mov.u32 %0, %clock;" : "=r"(clock) );

  return clock;

}

It seems to run at half the shader clock.

4

Thank you very much!
That works very well for me

5

Thank you very much!
That works very well for me

6

You could also use the built-in method!

clock_t clock();

See section B.10 of the CUDA 3.1 programming guide for more info. The inline PTX is just is effective though…

7

You could also use the built-in method!

clock_t clock();

See section B.10 of the CUDA 3.1 programming guide for more info. The inline PTX is just is effective though…

8

Thanks, Profquail!

I somehow had in the back of my head that there was a CUDA function for this already, but apparently missed it when I quickly scanned the Programming Guide appendices to check.

9

Thanks, Profquail!

I somehow had in the back of my head that there was a CUDA function for this already, but apparently missed it when I quickly scanned the Programming Guide appendices to check.

10

Thank you both very much, love this board :)

11

Thank you both very much, love this board :)