Fortran codes in CUDA

Accelerated Computing CUDA CUDA Programming and Performance
#1

Hey Guys

I'm new here. And i have some doubts how i must proceed to make my fortran codes run with CUDA. I make codes in fortran90 and reading the manual i don't see any keyword or something like to use when i'm compiling the program.

Let us assume that is a code:

program test
implicit none

! Declaring variables

integer :: var1
real :: var 2
doubleprecision :: var3
character :: var4

! The code

assume this is a program

end program

save as test.f90, what must i do to convert my program to CUDA. Exist any keyword, or some library that must i add to the source. I’m completely newbie with this, i try to make some programs run with CUDA and none works. Thanks for all help.

PS: sorry for my english, isn’t good enough, iḿ a brazilian user.

#2

I develop F90+CUDA programs, and here’s what I do…

In your F90 code, I’ll assume somewhere you have a function line like

[font=“Lucida Console”]call launchgpustuff(arguments)[/font]

Then in a *.cu file, you’ll put the declaration of that function with the appropriate name-mangling convention. For example, if using gfortran on linux, I’d define (note the underscore after the function name)

[font=“Lucida Console”]global void launchgpustuff_(arguments)
{
.
.
.
kernelfunction<<dimGrid,dimBlock>>(more arguments);
.
.
.
}[/font]

To compile this code, you first compile the fortran code using [font=“Lucida Console”]gfortran -c[/font], then compile the CUDA code using [font=“Lucida Console”]nvcc -c[/font]. Finally, link it all together using gfortran: [font=“Lucida Console”]gfortran -lstdc++ -L/usr/local/cuda/lib/ -lcudart *.o -o executable_name[/font]

This is what has worked for me. Unfortunately, I am unable to use the visual profiler. I think it may be a CUDA+F90 problem. So I’d be interested to know if you can get that to work for you.

#3

Fortran is not supported by CUDA now. You can only write CUDA code in C/C++. However, you can write the CPU code in Fortran, and write GPU code in C/C++, then link them together using the technique described by faircdl.

#4

I am using CUDA+90 the same way you are doing and the visual profiler works pretty well for me.

#5

I’m curious about your setup. I have a RHEL 5.3 box with a GTX 295 and 2 Teslas. I’ve tried CUDA 2.1 and 2.2b to no avail. I’m not doing any fancy multi-GPU stuff for now. All I can see when I run the profiler are the memcopy operations. Any thoughts?

#6

I have Cuda 2.0 and RHEL 5.1 installed with C1060 GPUs. Actually, I can see all the kernel calls but memcopies doesn’t appear in the profiler. May be there is a problem when profiling operations that may overlap. I dont think it is a fortran+Cuda issue.

#7

Well, for some reason unknown to me, the profiler is working now. I’ll just keep my fingers crossed that everything remains happy. Thanks for your reply.

#8

Thanks for all the help. But, I use programs who are code, completely in fortran, including the functions, which makes the use of C/C++ functions impossible. What I’m trying to do is running a fortran source natively in my GPU, but, this appears to be impossible. I thought about converting the hole program to C/C++, but I Know too little about C/C++ to do this, and the program isn’t mine, i just had access to the source, but the new modifications and all new things implemented in the program i don’t know how it works. But, thanks a lot, I’ll study C!

#9

You should look into Flagon (just google it). It’s a fortran wrapper for most of CUDA. You could also look into the latest beta of Portland Group’s fortran compiler. It attempts to use some OpenMP-like flags to auto-parallelize parts of your code for the GPU.

#10

The Flagon looks nice. I’ll think about it. Thanks again

#11