Cuda 3.2RC not compatible with 260.19.12 driver?

Accelerated Computing CUDA CUDA Programming and Performance
1

I tried to compile cuda sdk sample programs after installing cuda 3.2RC.
The compiler is unable to compile some of the programs, e.g. simpleTextureDrv,
vectorAddDrv, threadMigration, matrixMulDrv, deviceQueryDrv

Typical linker error:

simpleTextureDrv.cpp:(.text+0x152): undefined reference to `cuCtxCreate_v2’

my workaround: modification of some Makefiles to comment out the offending cpp file and
adding a trivial cpp (with emty main) file to please the make and compiler.

My platform:
Ubuntu 10.4
Cuda compilation tools, release 3.2, V0.2.1221
NVidia 260.19.12 driver installed from UBUNTU packages

2

I tried to compile cuda sdk sample programs after installing cuda 3.2RC.
The compiler is unable to compile some of the programs, e.g. simpleTextureDrv,
vectorAddDrv, threadMigration, matrixMulDrv, deviceQueryDrv

Typical linker error:

simpleTextureDrv.cpp:(.text+0x152): undefined reference to `cuCtxCreate_v2’

my workaround: modification of some Makefiles to comment out the offending cpp file and
adding a trivial cpp (with emty main) file to please the make and compiler.

My platform:
Ubuntu 10.4
Cuda compilation tools, release 3.2, V0.2.1221
NVidia 260.19.12 driver installed from UBUNTU packages

3

There is your problem. The linking error you are seeing is because the linker cannot find libcuda.so in the standard paths it is looking to find it.

I am willing to get that the repackaged drivers in the Ubuntu repository either don’t include the CUDA libraries or have moved them to a non-standard place. The simple solution is only use the NVIDIA official driver distribution - then things like CUDA and OpenCL will work as advertised.

4

There is your problem. The linking error you are seeing is because the linker cannot find libcuda.so in the standard paths it is looking to find it.

I am willing to get that the repackaged drivers in the Ubuntu repository either don’t include the CUDA libraries or have moved them to a non-standard place. The simple solution is only use the NVIDIA official driver distribution - then things like CUDA and OpenCL will work as advertised.

5

This explanation makes sense.

In my system libcuda.so can be found in 4 places under /usr.

I exclude 2 of them as their names suggests 32 bits, while my system is 64 bit.

I’m left with two “suspects”

/usr/lib/libcuda.so

/usr/lib/nvidia-current/libcuda.so

These are symbolic links, which after 1 or two steps lead to (respectively)

libcuda.so.256.53

libcuda.so.260.19.12

So I guess the linker used library from the old driver, while X uses the new one. I’ll check it.

One problem is that my ubuntu repository I get the driver from is unofficial.

Attempt of installing 260.* manually left me in the console, even after many hours of trying to get things right :-(

Hence google and unofficial repo.

Looks I’ve got a mess in the system :-(

Thanks

6

This explanation makes sense.

In my system libcuda.so can be found in 4 places under /usr.

I exclude 2 of them as their names suggests 32 bits, while my system is 64 bit.

I’m left with two “suspects”

/usr/lib/libcuda.so

/usr/lib/nvidia-current/libcuda.so

These are symbolic links, which after 1 or two steps lead to (respectively)

libcuda.so.256.53

libcuda.so.260.19.12

So I guess the linker used library from the old driver, while X uses the new one. I’ll check it.

One problem is that my ubuntu repository I get the driver from is unofficial.

Attempt of installing 260.* manually left me in the console, even after many hours of trying to get things right :-(

Hence google and unofficial repo.

Looks I’ve got a mess in the system :-(

Thanks