gfortran w/ CUDA C How to link a gfortran w/ CUDA C

Accelerated Computing CUDA CUDA Programming and Performance
1

I’ve created a glue kernel routine in C that is callable from my gfortran module.

How do I link my gfortran objects w/ nvcc objects? Do I just compile them w/ gfortran (-c) and
then use

nvcc mykernel.c f1.o f2.o -o my_executable

I imagine there are library issues.

Anybody do this?

Vince Schuster

2

Use fortran to compile your fortran source code like this

gfortran -c fmain.f

then use gfortran linker to link your cuda binary files like this

gfortran -o test.exe fmain.o cmain.cu.o -lgcc -L/usr/local/cuda/lib64/ -lcuda -lcudart

The library path depends on your machine. Also you can use nvcc linker to do the same thing, but then you have to specify the Fortran library file.

3

Essentially you are doing it backwards.

Fist just compile your cuda code, “nvcc -c mykernel”

Also compile your fortran code with “gfortran -c myfortrancode”

Link all the object files together with gfortran, be sure to include a library path and include path to libcuda, and libcudart

Here is a Makefile that I have used to do this:

[codebox]

NVCC := /usr/local/cuda/bin/nvcc

LD_LIBRARY_PATH := /usr/local/cuda/lib64

LD_INCLUDE_PATH := /usr/local/cuda/include,/usr/local/pgplot

CUDAFORTRAN_FLAGS := -L $(LD_LIBRARY_PATH) -I $(LD_INCLUDE_PATH) -lcudart -lcuda

PGPLOT_FLAGS := -L/usr/local/pgplot -lcpgplot -lpgplot -lX11 -lgcc -lm

PGPLOT_DIR = /usr/local/pgplot/

NVCCFLAGS := -gencode arch=compute_20,code=sm_20 -I $(LD_INCLUDE_PATH)

all: cuda plottest run

cuda: pspline.cu

$(NVCC) -c pspline.cu $(NVCCFLAGS) $(PGPLOT_FLAGS)

plottest:

 gfortran -o pspline pspline.o $(PGPLOT_FLAGS) $(CUDAFORTRAN_FLAGS)

run:

./pspline

[/codebox]

You can do this with any fortran compiler, I’ve personally done it with both gfortran and ifort.

There is a very good guide here:

http://www.computationalmathematics.org/to…om_fortran.html

4

The technically sophisticated way to do this is use the iso_c_binding module in the Fortran compiler. It avoids the issues with name mangling and padding. It may have a somewhat longer learning curve, but for the long run is the better approach.

MMB