On Mon, Oct 10, 2011 at 10:06:09PM +0530, Sumit Semwal wrote:
On 10 October 2011 02:13, Clark, Rob rob@ti.com wrote:
hmm, random though..
should we define CPU_PREP/CPU_FINI type ioctl's?
At least in the GStreamer case, we could cache the userspace mmap'ing, and use the extra ioctls so the kernel side would know when cache clean/invalidate is needed. That would be a bit more efficient than forcing userspace to map/unmap all the time, and could hook into sync-objects of the allocator in cases where you have CPU access after GPU renders to a buffer..
So, you think we could post the RFC to upstream lists without this, mentioning it as the immediate next TODO item? Or is it essential to add it for v1 of RFC?
I think th "That would be a bit more efficient than ..." is a dead giveaway: Imo performance optimisation for future extensions ;-)
Otherwise I'm pretty satisfied with what the patch currently looks like now. I think the next step is to proof-of-concept actual use-cases. I'm hoping to be able to help there by porting Dave Airlie's PRIME drm buffer sharing work onto this (the necessary hw is already here to make something work on i915). But I can't promise much, unfortunately - too much other stuff going on.
Yours, Daniel