32-bit CUDA version is insufficient for CUDART version

Accelerated Computing CUDA CUDA Programming and Performance
1

I’m using CUDA 3.0 in 64-bit linux. I have installed driver version 195.36.15, and the latest toolkit, which has libcudart.so.3.0.14. This is a new install, so this is the only version on the computer. When I compile a 64-bit version of the program, it works correctly. However, when I add -m32 to the compiler (after cleaning, of course) to produce a 32-bit version, when run it gives the error “CUDA version is insufficient for CUDART version” when run. ldd confirms that it is pulling libcudart from the correct location:
libcudart.so.3 => /usr/local/cuda/lib/libcudart.so.3 (0xf7ef7000)

Is this a known problem when using the 32-bit runtime library with the 64-bit driver? Any work-arounds?

Edit: I discovered the Linux 64-bit version 195.36.24 driver on the drivers page and upgraded to it. No change.

2

Are you aware that cudart dll is versioned now too?

3

I took the deviceQuery source, removed all references to the shr* commands, replacing them with standard commands, and recompiled as such:

/usr/local/cuda/bin/nvcc -m64 -I/home/cluster/CUDA/3.0/sdk/sdk/C/common/inc -I/usr/local/cuda/include -o deviceQuery deviceQuery.cpp -L/usr/local/cuda/lib64 -lcudart

/usr/local/cuda/bin/nvcc -m32 -I/home/cluster/CUDA/3.0/sdk/sdk/C/common/inc -I/usr/local/cuda/include -o deviceQuery32 deviceQuery.cpp -L/usr/local/cuda/lib -lcudart

Here are the outputs:

./deviceQuery Starting…

CUDA Device Query (Runtime API) version (CUDART static linking)

There is 1 device supporting CUDA

Device 0: “GeForce GTX 480”

CUDA Driver Version: 3.0

CUDA Runtime Version: 3.0

CUDA Capability Major revision number: 2

CUDA Capability Minor revision number: 0

Total amount of global memory: 1609760768 bytes

Number of multiprocessors: 15

Total amount of constant memory: 65536 bytes

Total amount of shared memory per block: 49152 bytes

Total number of registers available per block: 32768

Warp size: 32

Maximum number of threads per block: 1024

Maximum sizes of each dimension of a block: 1024 x 1024 x 64

Maximum sizes of each dimension of a grid: 65535 x 65535 x 1

Maximum memory pitch: 2147483647 bytes

Texture alignment: 512 bytes

Clock rate: 1.40 GHz

Concurrent copy and execution: Yes

Run time limit on kernels: No

Integrated: No

Support host page-locked memory mapping: Yes

Compute mode: Default (multiple host threads can use this device simultaneously)

deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 4200131, CUDA Runtime Version = 3.0, NumDevs = 1, Device = GeForce GTX 480

PASSED

./deviceQuery32 Starting…

CUDA Device Query (Runtime API) version (CUDART static linking)

cudaGetDeviceCount FAILED CUDA Driver and Runtime version may be mismatched.

FAILED

I also tried using the 32-bit toolkit explicitly, with the same result. The modified deviceQuery source is attached.

I also compiled and ran this on another computer with a Tesla S1070 using CUDA 2.3 and driver 195.36.15. Here’s the output

64-bit version:

deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 4200171, CUDA Runtime Version = 2.30, NumDevs = 4, Device = Tesla T10 Processor, Device = Tesla T10 Processor

32-bit version:

deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 134518067, CUDA Runtime Version = 2.30, NumDevs = 4, Device = Tesla T10 Processor, Device = Tesla T10 Processor

Should the reported driver versions be different?
deviceQuery.cpp (6.78 KB)

4

There’s a typo in the original source in the sprintf statement for the driver version. Once that’s fixed, and looking only at the versions,

64-bit:

CUDA Driver Version = 3.0, CUDA Runtime Version = 3.0

32-bit:

CUDA Driver Version = 0.0, CUDA Runtime Version = 3.0

And that looks like the problem. When using 32-bit code with the 64-bit driver, cudaDriverGetVersion() returns version 0.0.

5

In windows, you have to put cudart32_(version number).dll with your programm, cause in 3.0 cudart.dll is versioned. Could it be an issue?

6

No, libcudart.so.3 links to the correct version.

7

This is fixed in CUDA 3.1 beta. Thanks!