Am 16.08.23 um 14:30 schrieb Danilo Krummrich:
On 8/16/23 16:05, Christian König wrote:
Am 16.08.23 um 13:30 schrieb Danilo Krummrich:
Hi Matt,
On 8/11/23 04:31, Matthew Brost wrote:
In XE, the new Intel GPU driver, a choice has made to have a 1 to 1
mapping between a drm_gpu_scheduler and drm_sched_entity. At first
this
seems a bit odd but let us explain the reasoning below.
1. In XE the submission order from multiple drm_sched_entity is not
guaranteed to be the same completion even if targeting the same
hardware
engine. This is because in XE we have a firmware scheduler, the GuC,
which allowed to reorder, timeslice, and preempt submissions. If a
using
shared drm_gpu_scheduler across multiple drm_sched_entity, the TDR
falls
apart as the TDR expects submission order == completion order. Using a
dedicated drm_gpu_scheduler per drm_sched_entity solve this problem.
2. In XE submissions are done via programming a ring buffer (circular
buffer), a drm_gpu_scheduler provides a limit on number of jobs, if
the
limit of number jobs is set to RING_SIZE / MAX_SIZE_PER_JOB we get
flow
control on the ring for free.
In XE, where does the limitation of MAX_SIZE_PER_JOB come from?
In Nouveau we currently do have such a limitation as well, but it is
derived from the RING_SIZE, hence RING_SIZE / MAX_SIZE_PER_JOB would
always be 1. However, I think most jobs won't actually utilize the
whole ring.
Well that should probably rather be RING_SIZE / MAX_SIZE_PER_JOB =
hw_submission_limit (or even hw_submission_limit - 1 when the hw
can't distinct full vs empty ring buffer).
Not sure if I get you right, let me try to clarify what I was trying
to say: I wanted to say that in Nouveau MAX_SIZE_PER_JOB isn't really
limited by anything other than the RING_SIZE and hence we'd never
allow more than 1 active job.
But that lets the hw run dry between submissions. That is usually a
pretty horrible idea for performance.
However, it seems to be more efficient to base ring flow control on
the actual size of each incoming job rather than the worst case,
namely the maximum size of a job.
That doesn't sounds like a good idea to me. See we don't limit the
number of submitted jobs based on the ring size, but rather we calculate
the ring size based on the number of submitted jobs.
In other words the hw_submission_limit defines the ring size, not the
other way around. And you usually want the hw_submission_limit as low as
possible for good scheduler granularity and to avoid extra overhead.
Christian.
Otherwise your scheduler might just overwrite the ring buffer by
pushing things to fast.
Christian.
Given that, it seems like it would be better to let the scheduler
keep track of empty ring "slots" instead, such that the scheduler
can deceide whether a subsequent job will still fit on the ring and
if not re-evaluate once a previous job finished. Of course each
submitted job would be required to carry the number of slots it
requires on the ring.
What to you think of implementing this as alternative flow control
mechanism? Implementation wise this could be a union with the
existing hw_submission_limit.
- Danilo
A problem with this design is currently a drm_gpu_scheduler uses a
kthread for submission / job cleanup. This doesn't scale if a large
number of drm_gpu_scheduler are used. To work around the scaling
issue,
use a worker rather than kthread for submission / job cleanup.
v2:
- (Rob Clark) Fix msm build
- Pass in run work queue
v3:
- (Boris) don't have loop in worker
v4:
- (Tvrtko) break out submit ready, stop, start helpers into own
patch
Signed-off-by: Matthew Brost <matthew.br...@intel.com>
