cuda-memcheck --tool racecheck Internal Memcheck Error: Device not supported

Has anyone else seen this error reported by
/usr/local/cuda-5.0/bin/cuda-memcheck CUDA-MEMCHECK version 5.0 (17) ?

The documentation says racecheck is supported by compute level 1.1 and above
but I am seeing it on nvidia GTX 295 (SM_13) and Tesla K20c (SM_35).

On both GPUs the error message includes:
Host Frame:/usr/local/cuda/lib64/libcudart.so.5.0 (cudaDeviceSetCacheConfig + 0x120) [0x38000]
and is reported when my program successfully finishes.
cudaDeviceSetCacheConfig(cudaFuncCachePreferL1) appears to be ok and as expected returns zero.

As always any help or comments would be most welcome.
Thank you
Bill
http://www.cs.ucl.ac.uk/staff/W.Langdon/

Hi Bill, The racecheck tool in cuda-memcheck is not supported on SM 1.x devices. SM 3.5 support was added to racecheck in CUDA 5.5. Can you download the latest CUDA toolkit and driver and see if the problem still persists ?

Dear Vyas,
Thank you for your prompt and helpful reply.

Ok some part of this is my mistake (I blame small fonts…)
but http://docs.nvidia.com/cuda/cuda-memcheck/ says
“The CUDA-MEMCHECK tool suite is supported on all CUDA capable
GPUs with SM versions 1.1 and above.” This is immediately
qualified by Table 4. However Table 4 (on second
reading;-) suggests that Racecheck should work with SM_20.

If racecheck does not work on SM 2.0… 3.0 perhaps
http://docs.nvidia.com/cuda/cuda-memcheck/ should be
updated?

Thanks again
Bill

Hi Bill, as of CUDA 5.5., racecheck works on SM 2.x, SM 3.0 and SM 3.5. Only SM 1.x devices are unsupported by the racecheck tool.