On Fri, Oct 7, 2011 at 3:42 PM, Jerome Glisse j.glisse@gmail.com wrote:
Forget to mention that we would need a new wait_fence ioctl to wait for a specific fence seq on a specific ring
We need a wait_fence ioctl anyway and even with a timeout parameter. It's a mandatory OpenGL feature. r600g uses a loop with sched_yield right now, which is like the worst implementation of wait_fence we can have. r300g uses bo_wait for timeout == infinite, and a loop with usleep for timeout != infinite.
Marek