CUDA and C++ Do i need a main function in the .cu?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi, i have a .cu file with a kernel and a function that calls that kernel (and do the transfer between memories stuff).

I compile it by /usr/local/cuda/bin/nvcc myFile.cu and the other files in the project using g++.

when i

/usr/local/cuda/bin/nvcc myFile.cu

i get the error:

/usr/lib/gcc/i486-linux-gnu/4.3.2/…/…/…/…/lib/crt1.o: In function `_start’:

(.text+0x18): undefined reference to `main’

I just want to call the function from a class. Do i need a main function in the .cu file or something?

Anyone knows what am i doing wrong?

If you need some code, just ask for it.

I’ll post my makefile, just in case.

# Proyect: Paralel Reinforcement Evolutionary Artificial Neural Network

preann : XMM32.o paralelLayer.o AbstractLayer.o CudaLayer.o XmmLayer.o CppLayer.o main.o 

	g++ -o preann XMM32.o AbstractLayer.o CudaLayer.o XmmLayer.o CppLayer.o main.o paralelLayer.o

main.o : main.cpp NeuralNetwork.o

	g++ -c main.cpp

NeuralNetwork.o : NeuralNetwork.cpp NeuralNetwork.h CppLayer.o XmmLayer.o CudaLayer.o

	g++ -c NeuralNetwork.cpp

CppLayer.o : CppLayer.cpp CppLayer.h AbstractLayer.o

	g++ -c CppLayer.cpp

XmmLayer.o : XmmLayer.cpp XmmLayer.h AbstractLayer.o XMM32.h XMM32.o

	g++ -c XmmLayer.cpp

CudaLayer.o : CudaLayer.cpp CudaLayer.h AbstractLayer.o paralelLayer.o

	g++ -c CudaLayer.cpp

AbstractLayer.o : AbstractLayer.h

	g++ -c AbstractLayer.h

paralelLayer.o : paralelLayer.cu

	/usr/local/cuda/bin/nvcc paralelLayer.cu

XMM32.o : XMM32.asm

	nasm -f elf XMM32.asm

clean: 

	rm preann XMM32.o AbstractLayer.o CudaLayer.o XmmLayer.o CppLayer.o main.o paralelLayer.o

thanks

2

The command

/usr/local/cuda/bin/nvcc paralelLayer.cu

will cause nvcc to produce an executable, just like gcc. You need to let nvcc know you want to produce an object file (just like gcc):

/usr/local/cuda/bin/nvcc -c parallelLayer.cu

3

Thanks a lot!

i was stuck in that…

4

Hello,
I’m in front of the same problem !
I get the answer (thanks for it) but I did not succeed dealing with two compilers !
For example, I have a source code A (sourceA.cpp) containing a main function and a source code B (sourceB.cu) containing only a function.
In the main of A, I called the function in B. I’m compiling the sourceB.cu with nvcc -c. But after that, how can I “explain” to gcc that the function it’s looking for has been already compiled ?
Maybe my question is really easy but I can’t find the way of doing it.

5

Sorry, i was doing other things. I still didn’t try to call the function, so i’m not sure if what i’m going to tell you is right.

I think you must write in your C++ code something like:

extern "C" void yourCudaFunction ();

I did it to use assembly functions from C++ and it was enought.

Only one function?

I think you can’t call a kernel directly from C++. You need to write a “normal” function in the .cu.

In the examples, that function calls the kernel and execute all that cudaMalloc and cudaMemCpy stuff.

i hope one of this two things can help you.