C++ integration: 'undefined reference to main'

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello,

Apologies if this has been posted elsewhere, I couldn’t find anything using search.

I am trying to integrate working CUDA code into a working C++ program, but am faced with a compile error I can’t seem to work out.

Without posting all the code, here’s how I have it set up. It’s based on the ‘C++ Integration’ sample code from here.

== main.h ==

// includes

int main(int argc, char **argv);

extern "C" void gpuDivide(float* image, unsigned int width, unsigned int height, float divisor);

== main.cpp ==

#include "main.h"

int main(int argc, char **argv)

{

	// code

}

== GpuIntegration.cu ==

#include "GpuKernel.cu"

extern "C" void gpuDivide(float* image, unsigned int width, unsigned int height, float divisor)

{

	divide<<< 1,1 >>>(image, width, height, divisor);

}

== GpuKernel.cu ==

__global__ void divide(float *image, unsigned int width, unsigned int height, float divisor)

{

	// code

}

Compile command:

/usr/local/cuda/bin/nvcc -I/usr/local/cuda/include -O2 -g -c

The compiler comes up with the following:

make all 

Building file: ../GpuIntegration.cu

Invoking: C Compiler

/usr/local/cuda/bin/nvcc -o "GpuIntegration.o" "../GpuIntegration.cu"

/usr/lib/gcc/x86_64-linux-gnu/4.4.1/../../../../lib/crt1.o: In function `_start':

/build/buildd/eglibc-2.10.1/csu/../sysdeps/x86_64/elf/start.S:109: undefined reference to `main'

collect2: ld returned 1 exit status

make: *** [GpuIntegration.o] Error 1

I’ve been struggling with this since last week, which has stifled my progress. Any suggestions on how to fix this are most welcome!

Thanks for your help,

Chris

2

Probably function name mangling inside main.cpp . It is a bit hard to tell, but it looks like you are linking with gcc. If the C++ compiler has mangled main() into something else in the output object file, the plain C linker won’t know what to do. It should be trivial to confirm this using objdump (presuming this is linux/elf, if it is OS X then there is a mach tool which does the same thing but I can’t remember what it is called off the top of my head) to look at the symbols in the output.

The remedy is either to link with g++, as opposed to gcc, or declare main as extern “C” as well, which should defeat the symbol name mangling and make it work as expected.

3

The solution turned out to be much simpler (more stupid) than I suspected.

Am using Eclipse as an IDE. In the compiler command line settings, I removed the ${flags} parameter as it was adding -MMD as an option which nvcc didn’t like. However, having removed the flags, I forgot to put the -c option back. The compiler was trying to link straight after compiling the first source file, and it couldn’t find a main function.

D’oh! Thanks for your help avidday.