valgrind error on simpleCUBLAS

Hej,

when I call valgrind on the NVIDIA SDK example simpleCUBLAS, valgrind finds a few errors. Do I have to care about them?
[agruber@plastilin release]$ valgrind ./simpleCUBLAS
==8998== Memcheck, a memory error detector.
==8998== Copyright © 2002-2007, and GNU GPL’d, by Julian Seward et al.
==8998== Using LibVEX rev 1804, a library for dynamic binary translation.
==8998== Copyright © 2004-2007, and GNU GPL’d, by OpenWorks LLP.
==8998== Using valgrind-3.3.0, a dynamic binary instrumentation framework.
==8998== Copyright © 2000-2007, and GNU GPL’d, by Julian Seward et al.
==8998== For more details, rerun with: -v
==8998==
simpleCUBLAS test running…
==8998== Conditional jump or move depends on uninitialised value(s)
==8998== at 0x667AD98: (within /usr/lib64/libcuda.so.185.18.08)
==8998== by 0x667AEFD: (within /usr/lib64/libcuda.so.185.18.08)
==8998== by 0x667AFEE: (within /usr/lib64/libcuda.so.185.18.08)
==8998== by 0x6666E09: (within /usr/lib64/libcuda.so.185.18.08)
==8998== by 0x6668332: (within /usr/lib64/libcuda.so.185.18.08)
==8998== by 0x6660B22: cuLaunchGridAsync (in /usr/lib64/libcuda.so.185.18.08)
==8998== by 0x4C26540: (within /opt/cuda/lib/libcudart.so.2.2)
==8998== by 0x4C1F69E: cudaLaunch (in /opt/cuda/lib/libcudart.so.2.2)
==8998== by 0x4E667D1: (within /opt/cuda/lib/libcublas.so.2.2)
==8998== by 0x4E667ED: (within /opt/cuda/lib/libcublas.so.2.2)
==8998== by 0x4E6739A: (within /opt/cuda/lib/libcublas.so.2.2)
==8998== by 0x4E6566F: cublasSgemm (in /opt/cuda/lib/libcublas.so.2.2)
==8998==
==8998== Conditional jump or move depends on uninitialised value(s)
==8998== at 0x667AD9A: (within /usr/lib64/libcuda.so.185.18.08)
==8998== by 0x667AEFD: (within /usr/lib64/libcuda.so.185.18.08)
==8998== by 0x667AFEE: (within /usr/lib64/libcuda.so.185.18.08)
==8998== by 0x6666E09: (within /usr/lib64/libcuda.so.185.18.08)
==8998== by 0x6668332: (within /usr/lib64/libcuda.so.185.18.08)
==8998== by 0x6660B22: cuLaunchGridAsync (in /usr/lib64/libcuda.so.185.18.08)
==8998== by 0x4C26540: (within /opt/cuda/lib/libcudart.so.2.2)
==8998== by 0x4C1F69E: cudaLaunch (in /opt/cuda/lib/libcudart.so.2.2)
==8998== by 0x4E667D1: (within /opt/cuda/lib/libcublas.so.2.2)
==8998== by 0x4E667ED: (within /opt/cuda/lib/libcublas.so.2.2)
==8998== by 0x4E6739A: (within /opt/cuda/lib/libcublas.so.2.2)
==8998== by 0x4E6566F: cublasSgemm (in /opt/cuda/lib/libcublas.so.2.2)
Test PASSED

Press ENTER to exit…

==8998==
==8998== ERROR SUMMARY: 4 errors from 2 contexts (suppressed: 10 from 1)
==8998== malloc/free: in use at exit: 14,143,510 bytes in 12,552 blocks.
==8998== malloc/free: 473,474 allocs, 460,922 frees, 223,736,395 bytes allocated.
==8998== For counts of detected errors, rerun with: -v
==8998== searching for pointers to 12,552 not-freed blocks.
==8998== checked 14,492,416 bytes.
==8998==
==8998== LEAK SUMMARY:
==8998== definitely lost: 4 bytes in 1 blocks.
==8998== possibly lost: 0 bytes in 0 blocks.
==8998== still reachable: 14,143,506 bytes in 12,551 blocks.
==8998== suppressed: 0 bytes in 0 blocks.
==8998== Rerun with --leak-check=full to see details of leaked memory.

Works without problems on my machine (Fedora 8/CUDA 2.0), in both release and debug builds of the example.

No, you don’t care about them.