[driver api][ptx]: cuModuleGetFunction fails with CUDA_ERROR_NOT_FOUND ... but cuModuleGetGlobal wor

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi!

I have a PTX file that’s successfully loaded by cuModuleLoad(). If I run cuModuleGetFunction() on that module with an existing kernel name, I get CUDA_ERROR_NOT_FOUND. However, if I run cuModuleGetGlobal() on the very same loaded module, it works like a charm. How’s that possible? Both symbols (the kernel and the global memory variable) can be found in the PTX file, so why does cuModuleGetFunction() fail? FYI, I use the “pure” kernel name as a string literal during lookup, not the mangled version found in the PTX. Using CUDA 3.1 on 64-bit Linux.

Any idea?

Thanks in advance!

2

I think you should trj using mangled name (at least that works for me )

Sergey.

3

I think you should trj using mangled name (at least that works for me )

Sergey.

4

That does work indeed! I consider this a bug as it’s neither documented that way nor does it seem sensible to me…

Cheers

5

That does work indeed! I consider this a bug as it’s neither documented that way nor does it seem sensible to me…

Cheers

6

CUDA files are compiled as C++, so unless functions are declared as extern “C” they have mangled names.

7

CUDA files are compiled as C++, so unless functions are declared as extern “C” they have mangled names.

8

In that case, they indeed get mangled names. Is the fact documented that they’re compiled as C++? If not, it should probably be added to the driver API section on kernel execution. Anyhow, I declared my kernels as extern “C” and now the function lookup works as expected.

Thanks Tim!

9

In that case, they indeed get mangled names. Is the fact documented that they’re compiled as C++? If not, it should probably be added to the driver API section on kernel execution. Anyhow, I declared my kernels as extern “C” and now the function lookup works as expected.

Thanks Tim!