Is it possible to use GPU from single application both via OpenCL and Brook+ provided no memory buffers shared between OpenCL and Brook+ kernel invocations?
No one still tried it? AFAIK current OpenCL lack of images/texture fetches support. Brook+ has it (but has no read/write arrays). It would be nice to use GPU via OpenCL device for some parts of code and to use GPU via Brook+ interface for other parts of code. If those parts have no common GPU memory buffers, should it be possible to maintain 2 different GPU device descriptions (for OpenCL and for Brook+ runtime) inside single process/thread or app should use different threads or even should consist of different cooperating processes to implement such approach?
I already have Brook+ portion. I can rather easy port another parts via OpenCL. But CAL/IL.... just another technology to learn + no hope to easy get cross-platform build. I would say it too costly for now to make such attempt. Maybe I will try to get OpenCL/Brook+ build. If it's possible to keep two independent descriptions of same GPU device for single thread see no reasons why it should fail...
You already have all the CAL/IL kernels if you have coded the Brook+ and OpenCL kernels.... so you just need to learn the "CAL" part (which can also be taken from the above, with some adjustments)..
Answering on my own question: yes, it's possible. It's possible to build app that uses both OpenCL SDK 2.01 and Brook+ 1.4 beta.
EDIT: but then one could get "Stream Write : Failed to Initiate DMA " in brook part. If no data returned to host memory in brook part all works fine.... EDIT2: not all data writes back to host fail. Only "async" ones.
