Math Functions sqrt, pow ^ and log

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

pgfortran compiling my CUDA Fortran code is ignoring the operator ^ while it says pow cannot be used in device kernel (allowed only in emulation mode).

I found conflicting information about support for math funcitons
PGI Webpage says

Q Does the compiler support IEEE standard-floating point arithmetic?

A The GPU accelerators available today support most of the IEEE floating-point standard. However, they do not support all the rounding modes, and some operations, notably square root, exponential, logarithm, and other transcendental functions, may not deliver full precision results. > This is a hardware limitation > that compilers cannot overcome.

But CUDA Toolkit Documentation 12.3 Update 1 page from NVIDIA seems to have support for the said functions.

Can somebody shed more light about how to use math functionalities in device kernels?

Thanks.

2

Hi iamaditya,

“pow” is a C Math library function so can’t be called from a CUDA Fortran device routine. The Fortran equivalent for “pow” is the “**” operator.

Hope this helps,
Mat

3

That’s right, of course. I was rather vague in posting the issue.

Actually even ** was not accepted in a device kernel by the compiler(pgfortran).
PGF90-S-0000-Internal compiler error. unsupported operation 185

It sort of puts me in a fix because our fortran code uses the following functionalities (double precision)
**
sqrt
exp

So without them, I will not be able to port it to CUDA Fortran. Would you mind providing some pointers to literature, if they are already supported (double precision)?

4

Hi iamaditya,

These routines are all supported in CUDA Fortran so something else is going on. Can you please either post or send to PGI Customer Support (trs@pgroup.com) a reproducing example?

Thanks,
Mat

5

Hi Mat,

I tweaked the kernel a little in the meanwhile and fixed the compile error. For some reason, 10zn2 is not valid, while 10.0zn2 is proper

Unsupported kernel:

  attributes(global) subroutine addnum_kernel( zdarr1, zn2, zdarr3 )
    integer, device ::  zn2
    integer,device :: zdarr1(10), zdarr3(10)
    integer :: ix

    do ix = 1,10
       zdarr3(ix) = zdarr1(ix)*(10**zn2) 
    end do

  end subroutine addnum_kernel

Proper, supported kernel:


  attributes(global) subroutine addnum_kernel( zdarr1, zn2, zdarr3 )
    integer, device ::  zn2
    integer,device :: zdarr1(10), zdarr3(10)
    integer :: ix

    do ix = 1,10
       zdarr3(ix) = zdarr1(ix)*(10.0**zn2) 
    end do

  end subroutine addnum_kernel
  • Aditya

PS.: There are issues with declaring zdarr3 etc as integer, just to warn other fellow CUDA programmers

6

Hi Aditya,

It looks like a compiler issue to me so I created a problem report (TPR#18956) and sent it on to engineering. You’ve already determined that the work around is to use a real instead of an integer for the base value.

Thanks for the report!
Mat

7

18956 - CUDA Fortran: Using integer as base value of an exponential gets Unexpected runtime function call

has been fixed in 13.6, available now.

dave