Re: Fence array patchset
From: Alex Deucher
Date: Wed Jun 01 2016 - 09:42:35 EST
On Wed, Jun 1, 2016 at 9:10 AM, Christian KÃnig <deathsimple@xxxxxxxxxxx> wrote:
> Hi guys,
> this is the next iteration of the fence array patch set.
> Daniel suggested that I provide an example on how this functionality might be
> used by a driver. So I added a few additional patches in this series to show
> what I want to do with this in the amdgpu driver.
> The main idea is that for each VMID we have a set of hardware fences which are
> currently using this VMID. Now when a new command submission needs a VMID we
> construct a fence array which should signal when any of the VMIDs becomes
> available and gives that back to our the scheduler.
For those that are not familiar, a VMID = Virtual Memory ID. AMD GPUs
have multiple virtual address space contexts which can be in flight on
the GPU at any given time. The VMID is used to select which VM
context you want to use for a specific GPU operation.
> This effort and my testing also found a rather stupid typo in the code and I
> also tried to incorporate the comments from Chris and Daniel as well.
> I think it's ready to land now, but as usual feel free to take it apart.
> dri-devel mailing list