Cuda

From Mandriva Community Wiki

Using the RPM package

As of 2009.1 and up, nvidia-cuda-* packages can be found in the non-free repository.

A backport seems to exist for Mandriva 2009.0 on some ftp servers, but you should only use it for testing. It is very bad for production and security (read ahead).

The nvidia way

I have used Fedora 9 packages for Mandriva 2008.1 : works ok.

You can download files from Cuda's page on the nVidia website.

chmod 777 *

urpmi before you start

• gcc
  
• gcc-c++
  

graphic driver : NVIDIA-Linux-x86-180.22-pkg1.run

this is the usual Nvidia graphic driver, so :

• you need to be in text mode (init level 3)
• as root
• No need for this driver on text only emulation CUDA servers

Toolkit : ./cudatoolkit_X.X_linux32_fedora9.run

• as root
• destination : /usr/local/cuda
• no problems

SDK : ./cuda-sdk-linux-X.YY.ZZZZ.ZZZZ-ZZZZ.run

• no need to be root
• destination : ~/NVIDIA_CUDA_SDK/
• this package seems optional, I did not need it to compile. this could be only tutorials.

Before you continue :

• Please make sure your PATH includes /usr/local/cuda/bin
• Please make sure your LD_LIBRARY_PATH includes /usr/local/cuda/lib

PATH=$PATH:/usr/local/cuda/bin
LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib

or the definitive method :

#/etc/profile add :
PATH=$PATH:/usr/local/cuda/bin

#/etc/ld.so.conf add :
/usr/local/cuda/lib

To uninstall the NVIDIA CUDA SDK, please delete ~/NVIDIA_CUDA_SDK

Then you have to run as root :

ldconfig

Please note!

f you don't run this at root, you'll have the message at run time :

./run: error while loading shared libraries: libcudart.so.2: cannot open shared object file: No such file or directory

I had to logout - login too (?)

Your installation should be complete.

Compilation

I succeeded in compiling this program (remember I set the path) :

nvcc -o run prog.cu

Caution !
The file extension must be .cu, not .cpp !

#include <stdio.h>
#include <stdlib.h>
#include <cuda.h>
#include <cuda_runtime.h>


__global__ void mykernel(float * A, float * B, float * C)
{
	int p = threadIdx.x;
	C[p] = A[p] + B[p];
}

int main()
{
	float A[]={1,2,3,4,5,6,7,8,9};
	float B[]={10,20,30,40,50,60,70,80,90};
	float C[9];

	//9 sums, no loop !
	mykernel<<<1,9>>>(A,B,C);

	//output
	for(int i=0; i<9 ; i++)
		printf("%f\n",C[i]);

}

CUDA and Qt

Caution !
Qt is used on CPU only, not GPU !

Create an object file using -c option :

nvcc -c -o kernel.o kernel.cu

Then include your kernel.o file into your Qt Makefile as usual (line OBJECTS=).

Caution !

CUDA requires to add to the line LIBS :

LIBS= .... -L/usr/local/cuda/lib -lcudart

Execution

Tip !
To run the program in emulation mode (CPU instead of real GPGPU), compile as :

nvcc -deviceemu -o run prog.cu

libcudart.so.2: No such file or directory

-> see installation note *

./run
11.000000
22.000000
33.000000
44.000000
55.000000
66.000000
77.000000
88.000000
99.000000

It works fine in emulation mode using Mandriva 2008.1.

Original author : dsant_0 (Lyon, France). E-mail : forum (at) votreservice (point) com