simple question on CUBLAS cublasInit

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi guys,

When compiling the following code:

#include <cublas.h>

...

cublasStatus status;

status = cublasInit();

I get

Undefined reference to `cublasInit’…

Any ideas why?

Thanks

2

you need to link with cublas.lib which loates in C:\CUDA\lib64 or libcublas.so in /usr/local/cuda/lib

3

brilliant!

Thanks. If someone else has the same problem, what I did was to add to the Makefile:

# Additional libraries needed by the project

USECUBLAS	   := 1

and now it compiles ok

4

Hi there! Got the same problem. Which makefile did you edit? My CUDA-Program does not have a makefile, but I think i set the LD_LIBRARY_PATH correctly in my .bashrc:

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

The .so files are all there :)

My program is not that big (just starting to make myself comfortable with cuBLAS):

#include <vector>

#include <cublas.h>

#include <stdio.h>

#include <stdlib.h>

#include <string.h>

using namespace std;

vector<int> cVector;

int main(int argc, const char* argv[]) {

	cublasStatus status;

	status = cublasInit();

		if (status != CUBLAS_STATUS_SUCCESS) {

			fprintf (stderr, "!!!! CUBLAS initialization error\n");

			return EXIT_FAILURE;

		}

}

;

and i get the following error:

/tmp/tmpxft_00000d92_00000000-1_01_vector.o: In function `main':

01_vector.cc:(.text+0x10): undefined reference to `cublasInit'

collect2: ld returned 1 exit status

Does anybody have a solution? I’m using CUDA 2.3.

5

LD_LIBRARY_PATH doesn’t effect compilation behaviour. You need to explicitly specify you want to link CUBLAS, something like this:

cc -o mycode mycode.c -lcublas -lcudart

and you should be in business,…

6

Great! Thanks!

nvcc mycode.cc -lcublas

did the job!

7

Hello all.

I have the same problem when I compile the simpleCuda example under Linux CentOS 5.2 Cuda 3.0 version.

I use this commands to compile the example:

nvcc ./simpleCUBLAS.c -o simpleCUBLAS -I/usr/local/cuda/include -L/usr/local/cuda/lib64 -lcudartemu -lcublasemu -deviceemu

and

nvcc ./simpleCUBLAS.c -o simpleCUBLAS -I/usr/local/cuda/include -L/usr/local/cuda/lib64 -lcudartemu -lcublasemu

All dependenses are correct

libcudartemu.so.3 => /usr/local/cuda/lib64/libcudartemu.so.3 (0x00002ab850e86000)

libcublasemu.so.3 => /usr/local/cuda/lib64/libcublasemu.so.3 (0x00002ab8510c1000)

But still it gives

simpleCUBLAS test running..

!!!! CUBLAS initialization error
8

No you don’t.

You are not compiling for device emulation, but are linking with the device emulation libraries. That will not work.

9

Thanks for reply.

I`ve tried to compile the example with cublasemu library.

I use such command:

nvcc ./simpleCUBLAS.c -o simpleCUBLAS -I/usr/local/cuda/include -L/usr/local/cuda/lib64 -lcudart -lcublasemu -lcuda

But the program says:

“Segmentation fault”

10

Let me try again. If you are trying to use device emulation, the you need to tell nvcc using the -deviceemu flag, otherwise it won’t work. If you are trying to compile for a device, don’t use the device emulation libraries. Do one or the other, but don’t mix the two because, as you have found out, it doesn’t work.

11

I pretty understand what emulation mode does. When I tried to compile it for a device with the help of this command:

nvcc ./simpleCUBLAS.c -o simpleCUBLAS -I/usr/local/cuda/include -L/usr/local/cuda/lib64 -lcudart -lcublas

it still says:

simpleCUBLAS test running..

!!!! CUBLAS initialization error
12

And you have a working, cuda compatible video card?

13

Thanks for reply. There is something strang with my video card.

14

What card, driver and CUDA version are you using?

15

I use NVIDIA Geforce GTX 280, Cuda version 2.3, driver version is 190.18.

I`ve reinstalled all software and it started working in emulation mode, but when I try to run it as for device it says, that there is no device found. Probably I must install the latest software and configure nvidia card correctly. Because I have to cards connected to one PC(nVidia and AMD).

P.S. I`ve noticed another problem. When I start the my program compiled for cuda in emulation mode on my laptop(Acer 5632 WLMI) the function cudaGetDeviceCount returns 0. But when I start the same example on PC in emulation mode all is working ok. :(