If you are doing only hardware-accelerated h264 decoding on GPU, it's better to use the video memory. If you additionally want to do some work on CPU you, we recommend to use system memory. So you don't need to copy something by yourself.
Okay, and when I do additional work on the CPU, why is the ipp memcpy not available via the media sdk API?
As far as I understand it is used internally by the SDK.
For copying frames you can use VPP without any filters (just init it with the same input and output parameters). Also you can look at mfxExtVPPFieldProcessing structure. You can find more detailed information at the mediasdk-man.pdf.