Code::Blocks Semi-ignoring .cu files

Accelerated Computing CUDA CUDA Programming and Performance
1

Let’s say I have 3 files: test.c, test.h and test_kernels.cu. I’m also using Code::Blocks with all the nvcc setup requirements.

I have a function: void cuda_setup(int x) initialized in test.h, declared in test_kernels.cu and referenced in test.c . Everything compiles fine BEFORE I put in the reference into test.c but after I call the function I get an ‘undefined reference to cuda_setup’ error. test_kernels.cu IS being compiled into an object and if I plant an error in test_kernels.cu the compiler complains. However, if i copy/paste cuda_setup(int x) into test.c (or even a fourth file) everything compiles and links fine.

ARgh, this is frustrating. I can’t figure out what is wrong. I’m actually building on another person’s project and their code is split across 20+ .c files. I can put cuda_setup() into any other file and it works fine. It’s only ignored when it’s defined in test_kernels.cu. Any ideas why? Help would be saintly.

2

Try adding a few extern “C” lines:

/* e.g. when declaring */
extern “C” {
void cuda_setup(int x);
}

/* e.g. when defining */
extern “C”
void cuda_setup(int x)
{
…

}

3

That didn’t work :(

It’s compiling the .cu file into a .o, it complains when anything in the .cu file is wrong, and JUST now it even complained when I mismatched an ‘extern function()’ in the header with a ‘extern “C” function()’ in the actual declaration.

It just refuses to link the .cu object file with the rest of the same project. If i rename the .cu file to .c and remove the kernel everything works fine even without the ‘extern’. This must be a very very simple problem… that I apparently haven’t been able to figure out in days…

EDIT: I think i should have mentioned that the project I’m working off of is written in C and converting to C++ seems like it’ll be too much of a hassle. But, if there are no other alternatives…

4

Just a bump update: This isn’t a code::blocks error. I’ve got a seperate makefile and this problem still arises. The linker just refuses to see the functions within the .cu->object file.

5

Check the symbol names in the .o using nm.

If you have definitions like this:

extern “C” void cudafunction_(cuComplex *a, cuComplex *b, int *Np)

The symbol will have the same name:

nm Cuda_function.o |grep cudafunction
0000000000000260 T cudafunction_

If you omit the extern “C”, there will be C++ name mangling.

6

Thanks for the help. This was part of the problem. The solution (in my case at least) was that the main() function HAS to be in my .cu file. Once I switched that, defined external function calls within an extern ‘C’ the problem was fixed.