how to calculate #Gflop/sec?

Accelerated Computing CUDA CUDA Programming and Performance
1

Is there any existing facility to print out the #Gflop/sec of one program? If not, how to calculate it manully?

Thanks!

2

There’s no facility to calculate this automatically. Count the number of floating point operations in your kernel and divide by the execution time.

3

I spent last night making a source to do this, is quite simple but i was confused with the measure of time that machine used.

here is the code, made in fortran 90 (yesterday i get the c compiler and start to play with the video card External Media , but i made the source in fortran to test it first).

based in this definition (Documentation – Arm Developer) i make the code above.

program flops

implicit none

!

! variables

!

integer, parameter :: n=100000001

double precision, dimension(n) :: a,b,som

integer :: i

character(7) :: tim

integer*4 , dimension(n) :: iseed

real*4 ::var =0.

integer :: count, count_rate, count_max

!

! body

!

count_rate = 1

call system_clock(count,count_rate,count_max)

print *,count

do i=1,n

iseed(i) = secnds(var) 

iseed(i) = iseed(i) + mod(iseed(i),2) + 1

a(i) = ran(iseed(i))

b(i) = ran(iseed(i))

som(i) = a(i) *b(i)

end do

call system_clock(count,count_rate,count_max)

print *,count

end program

the program works fine, but the time give by the function System_clock is very strange, and i don’t know how to interpret it. Nevertheless, i’m making the same code to C++, and for now i have a few problems with the random number generator (how to use it), if some one knows how to do that srand(), or rand() works nicely will be great.

ps: sorry for the english, without revision of a expert this time.