Calling Thrust from CUDA Fortran

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

Hi,

I am trying to call Thrust from CUDA Fortran.

I am having trouble compiling the C code to get the .obj file.

For the Fortran code, I am using PVF 10.13 with CUDA 5.5 on a Windows system.

For the c code, I am trying to compile from Visual Studio 2012. Here is the code:

#include <thrust/device_vector.h>
#include <thrust/device_vector.h>
#include <thrust/sort.h>
extern "C" {
//Sort for integer arrays
void sort_int_wrapper( int *data, int N)
{
// Wrap raw pointer with a device_ptr
thrust::device_ptr <int> dev_ptr(data);
// Use device_ptr in Thrust sort algorithm
thrust::sort(dev_ptr, dev_ptr+N);
}
//Sort for float arrays
void sort_float_wrapper( float *data, int N)
{
thrust::device_ptr <float> dev_ptr(data);
thrust::sort(dev_ptr, dev_ptr+N);
}
//Sort for double arrays
void sort_double_wrapper( double *data, int N)
{
thrust::device_ptr <double> dev_ptr(data);
thrust::sort(dev_ptr, dev_ptr+N);
}
}

I get an error when I try to build:

error LNK1561: entry point must be defined

I also get warnings:

warning: Cannot tell what pointer points to, assuming global memory space

Can someone help me with this? The .obj file is not being created.

Thank you,

Kirk

2

This is what my linker command line looks like for the c code:

/OUT:“C:\Users\kirk\Desktop\CUDA Fortran\Thrust\Sort\Debug\Sort.exe” /MANIFEST /NXCOMPAT /PDB:“C:\Users\kirk\Desktop\CUDA Fortran\Thrust\Sort\Debug\Sort.pdb” /DYNAMICBASE “kernel32.lib” “user32.lib” “gdi32.lib” “winspool.lib” “comdlg32.lib” “advapi32.lib” “shell32.lib” “ole32.lib” “oleaut32.lib” “uuid.lib” “odbc32.lib” “odbccp32.lib” /DEBUG /MACHINE:X86 /INCREMENTAL /PGD:“C:\Users\kirk\Desktop\CUDA Fortran\Thrust\Sort\Debug\Sort.pgd” /MANIFESTUAC:“level=‘asInvoker’ uiAccess=‘false’” /ManifestFile:“Debug\Sort.exe.intermediate.manifest” /ERRORREPORT:PROMPT /NOLOGO /LIBPATH:“C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v5.5\lib\Win32” /TLBID:1

Kirk

3

Since I only want the .obj file, maybe I need to suppress the linking phase when building the c code?

Kirk

4

Hi Kirk,

I haven’t tried using Thrust on Windows but on Linux I compiled the Thrust wrapper using NVCC -c. It sounds like you’re trying using CL.

Compiling multi-language in Visual Studio is a bit of a pain since VS doesn’t allow you to mix languages in the same project. Hence, you need to create a library of all the object of one language and then link the library in with another.

What I’d try is compiling the CUDA C thrust interface source file using NVCC on the command line and then link this object into the PVF project.

Personally I was never happy with the performance I got out of thrust when calling it from CUDA Fortran. At least for me, any performance gains that thrust provided was off-set by the fact that I had to use this interface.

If you don’t mind blazing trail, there is a new library out of NVLabs called CUB (CUB: Main Page) that producing much better performance than Thrust. You’d still have to provide an CUDA C interface but it may be worth a try. Granted, I have not tried using it with CUDA Fortan so have no idea what problems you might encounter.

  • Mat
5

OK, I got the .cu file to compile. I finally have an .obj file ready for use with my Fortran code.

The trick was that I had to have Visual Studio Pro installed, otherwise nvcc complained about a bunch of stuff.

I am now able to use Thrust!

I will try to use CUB as well, if I succeed, I will post a comparison between sorting in Thrust and CUB.

Kirk