happy to see you're doing OpenCL. Since you explicitly said you're looking for feedback, I'll describe my situation/needs, in the hope that it might be useful.
I'm working on a high quality video renderer for HTPCs (home theater PCs). For this purpose CPU powered OpenCL doesn't work. GPU support is required. One of the main reasons for this is that after DXVA decoding the video frames are stored in GPU RAM. Transfering them back to system RAM, processing them there and then transfering them once again to the GPU costs so much performance that it's all not useful, anymore. I've measured that with Intel GPUs transfering NV12 DXVA 1080p frames from GPU RAM to system RAM runs with less than 24fps. So already the simple GPU -> system RAM transfer is so slow that I can't use it at all for real time video playback.
My exact requirements are these:
- OpenCL kernels need to run on GPU
- Direct3D9 interop is necessary (DirectShow/MediaFoundation is based on DX9)
- need ability for OpenCL kernels to access NV12 DXVA decoding surfaces
P.S: A bit OT, but I'm desperate: Is there any way to sign an NDA to get access to confidential specs and SDKs? E.g. I've been trying hard for *months* to get access to the Intel CUI SDK. No chance in hell, it seems, which really is a pain. Generally Intel GPU related information (e.g. DXVA decoding, h264 MVC decoding, activating frame packed 3D output etc) is nearly impossible to get. If you can do anything to improve that, that would be wonderful.
Best regards, Mathias.
Systemsoftware Mathias Rauen