[PATCH 5/5] amdgpu: use drm sync objects for shared semaphores (v4)
airlied at gmail.com
Fri May 12 04:17:46 UTC 2017
On 12 May 2017 at 13:34, zhoucm1 <david1.zhou at amd.com> wrote:
> 1. generally, functions in amdgpu_cs.c should be with amdgpu_cs_ as prefix.
Okay I've fixed this and previous patch up locally.
> 2. If I'm not wrong to your proposal, SYNCOBJ_IN is to semaphore wait while
> SYNCOBJ_OUT is to semaphore signal. SYNCOBJ_IN/OUT both are based on command
> submission ioctl, that means user space must generate CS when using
> semaphore? but with my understand, they should not be dependent with that,
> they can be used independently, right?
Yes in is WAIT and out is signal, however OUT could also be used to
write a syncobj as a fence if needed, hence why I moved away from
The only place I can see them being used independently is a possible
signal operation after present, due not being able to pass the
semaphores over dri3 yet. I think I've said this before and Christian
has confirmed that doing anything with semaphores not via the command
submission ioctl is going to be messy as they have to queue jobs in
the scheduler, so if we need to tune the command submission ioctl to
take empty CS or add a flag to just do semaphore operations we should
do so in the future when we have a clear use case for it (and we see
the need to optimise for it).
More information about the amd-gfx