function define in .h declared in .cu

Question

I made a project split in three folders : src, include, obj I am using CImg.h and cuda.h library. I declare void convolve(cimg_library::CImg&, cimg_libra

Answer 1

1. please don't post code (or Makefile contents) with line numbers. It just gets in the way, usually.
2. The compilation process here follows a compile...link process. The first stage is the compile stage, and you should be using -c for both g++ and nvcc during this stage (i.e. for both types of Makefile targets)
3. nvcc has no -cu option. I think maybe you meant -x cu

I also removed the ampersands on the function call.

Here is a simplified example based on what you have shown, primarily removing the CImg stuff and some other unnecessary items:

$ cat main.cpp
#include "convolve.h"

int main(){

int var1 = 0;
int var2 = 1;
convolve(var1, var2);
}
$ cat convolve.h
void convolve(int &, int &);
$ cat convolve.cu
#include "convolve.h"

__global__ void kernel(){};

void convolve(int & i1, int & i2){

kernel<<<1,1>>>();
}
$ cat Makefile
CC=nvcc
CX=g++
IDIR = .
special_IDIR = /usr/local/cuda-9.1/include
LDIR = /usr/local/cuda-9.1/lib64
CFLAGS=-I$(special_IDIR) -L$(LDIR)
LIBS =
ODIR = .
_DEPS = convolve.h
DEPS = $(patsubst %,$(IDIR)/%,$(_DEPS))
_OBJ = main.o convolve.o
OBJ = $(patsubst %,$(ODIR)/%,$(_OBJ))
$(ODIR)/%.o: %.cu $(DEPS)
        $(CC) -c -x cu -o $@ $< $(CFLAGS)
$(ODIR)/%.o: %.cpp $(DEPS)
        $(CX) -c -x c++ -o $@ $< $(CFLAGS) $(LIBS)
all: $(OBJ)
        $(CC) -o $@ $^ $(CFLAGS)
.PHONY: clean
clean:
        rm -f $(ODIR)/*.o *~ core $(INCDIR)/*~
$ make clean
rm -f ./*.o *~ core /*~
$ make
g++ -c -x c++ -o main.o main.cpp -I/usr/local/cuda-9.1/include -L/usr/local/cuda-9.1/lib64
nvcc -c -x cu -o convolve.o convolve.cu -I/usr/local/cuda-9.1/include -L/usr/local/cuda-9.1/lib64
nvcc -o all main.o convolve.o -I/usr/local/cuda-9.1/include -L/usr/local/cuda-9.1/lib64
$

change the instances of cuda-9.1 to cuda-9.0 in the above Makefile if you have CUDA 9.0 in your setup. The linker libs (-L) are not really needed for the compile commands, however. When linking with nvcc as we are doing here, it's also unnecessary to pass the -I and -L switches that refer to the CUDA includes and CUDA libraries that are part of the toolkit. nvcc already knows how to find those.

There may be any number of additional recommendations or tweaks to the Makefile, for instance the -x cu and -x c++ switches are probably unnecessary (at least they are unnecessary for my simplified example) but my objective here is not to create the perfect Makefile, but to give you a roadmap to get past the issue you are currently witnessing.