Are NPP routines CPU-thread-safe ?

I just wonder whether I can assume that I can call a NPP routine safely simultanously from multiple CPU threads (of course each CPU thread operates on a different set of images) ?

So in short, are the NPP routines thread-safe (for CPU threads) ?

I know that one can get problems with CPU-thread-safety when using stuff like texture references, and constant memory, and I suppose that some of the NPP routines might internally use such things.
For this topic see also

Ok, if found already a thread handling this …

In the thread it is mentioned that the NPP developers are working on decreasing the usage of texture references - any update on this from the NPP developers ? I see the usage of the ‘nppSetStream’ method less critical, because it can be avoided. But I cannot control how the NPP routines are implemented internally …

Any news regarding CPU-threadsafety in NPP library from Cuda Toolkit 6.5 ?