hardware-acceleration?

[gol]

Hi. This isn't about something specific, I will just share thoughts about how IPP could be in the future.

I like to see the IPP lib as a set of hardware-accelerated functions, a bit like a driver for some DSP coprocessor.
Even though it's not really the case (although one could see the SSE units as specialized processors), it could be in the future.

That is, nVidia recently released CUDA, and if Intel was partnering with nVidia, you could see, on systems with nVidia graphic boards, functions using CUDA. You wouldn't have to change your code and, for ex, your FFT's would be accelerated by your graphic card instead of SSE.

The same could apply to (I'm working with audio) soundcards. Right now there's zero chance anyone will bother programming soundcards, while some have good onboard DSP caps, there's no standard, no doc, etc.

Only problem (beside marketing, and also accuracy) I can imagine is the latency inherent to hardware-acceleration. But it could still benefit for paralellization (for ex, you would ask to perform 20 FFT's in parallel, some of them would be threaded, some would be hardware-accelerated).

[Vladimir_Dudnik]

Thanks for the interesting question. To be honest I do not expect much benefits from hw assistance for every single IPP function (you mentioned about latency already). Instead of that, I think some middle-level solution, providing some standard and consistent API would be more preferable. One of example could be DXVA API, but this is not cross platform API.

Regards,
Vladimir