Calling CUDA C kernel from PGI Visual Fortran

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

Hi,

I’m trying to call a CUDA C kernel from PGI CUDA Fortran, and I followed the instructions in this topic

http://www.pgroup.com/userforum/viewtopic.php?t=3169&start=0&postdays=0&postorder=asc&highlight=fortran+call+cuda

by using a similar Fortran interface statement as in Account Login | PGI

Unfortunately, I get link-time errors in Visual Studio. What I did was to build a static C library containing the CUDA kernel as a function (I built the library using an NVIDIA project inside Visual Studio), and then link in the resulting library in my PGI Visual Fortran project. Here are the linker errors:

2>erfCuda.lib(erf_kernel.cu.obj) : error LNK2001: unresolved external symbol __imp_frexp
2>erfCuda.lib(erf_kernel.cu.obj) : error LNK2001: unresolved external symbol __imp_ldexp
2>erfCuda.lib(erf_kernel.cu.obj) : error LNK2001: unresolved external symbol __imp_modff

I’ve searched for these symbols online, but didn’t find what libraries contain them. Has anybody here have any success in calling a CUDA C kernel from PGI Visual Fortran within MS Visual Studio? I’ve also tried building using “make” inside the PVF command prompt, but nvcc needs the Visual C compiler.

Thanks,
Chona

2

Hi Chona,

These symbols are all C99 math library routines. Since Microsoft doesn’t support C99, you may need to figure out what the Microsoft equivalent for them is or avoid using these routines. Does the code link if you remove these?

  • Mat
3

Hi Matt,

I’m not calling any of these routines, although the C kernel does use the CUDA function erf. Do you know why the kernel function needs these C99 math library routines?

To save time, will it be better if I package up my code along with the MSVS solution file, and send it to you?

Thanks,
Chona

4

Hi Matt,

I solved the problem. It turns out that using the multithreaded runtime library in my static C library project does the trick; that is, I needed to set the /MT linker option in the Properties → CUDA C/C++ → Host. I had it set to “inherit from host” before.

I hope this thread can help other MSVS developers.

Chona