Nouveau Working On Video Decoding & More

[MùPùF]

Sorry, but I have to correct you here. There is still a good deal of work to do in the rewritten nvfx driver, as it is not only porting nvfx to the new libdrm, but also adapting the nv50 driver model. This will provide a far more cleaner codebase (maintainability greaty improved) and hopefully will enable some new features. The rewritten driver will follow the naming scheme of the other nouveau gallium drivers and so will be most likely be called nv30.

So anybody reading here: don't expect anything too soon. The new libdrm is a big break for all pieces of the userspace stack, so we need to do a good deal of validation before pushing out anything to the users.

/me should have been clearer. I was only talking about the libdrm rewrite, not nvfx.

[whizse]

Heh, was that a sonic rainboom in the presentation?

[Kano]

@Lynxeye

You find that easyly in the net: "GF10x series cards have their double precision (FP64) performance capped at 1/8-th of the single precision (FP32) performance, while the architecture is capable of 1/2."

http://stackoverflow.com/questions/3...vs-geforce-480

Btw. that ECC support is a SOFTWARE not hardware feature.

[Nobu]

Thanks MùPùF, Lynxeye. Can't wait to see the results.

Of course, I'll try my best to report bugs properly when the time comes.

[mirza]

I would like to test OpenCL. Can any nouveau developer tell me, what is size of private memory available to individual computing unit? I cannot find that information anywhere in documentation for individual cards, which is not that surprisingc OTOH.

[calim]

I would like to test OpenCL. Can any nouveau developer tell me, what is size of private memory available to individual computing unit? I cannot find that information anywhere in documentation for individual cards, which is not that surprisingc OTOH.

That should be documented in the CUDA guide (where it's called shared memory).

For Fermis it's 64 KiB per MP (multi-processor), partitioned between L1 cache and shared memory, one gets 16 KiB and the other gets 48 KiB, user's choice which should be larger.