Hi Andres,
Please see inline (as [Serguei])
Sincerely yours,
Serguei Sagalovitch
From: Andres Rodriguez <andresr@xxxxxxxxxxxxxxxxx>
Sent: December 16, 2016 8:29 PM
To: Sagalovitch, Serguei; amd-gfx at lists.freedesktop.org
Subject: RE: [RFC] Mechanism for high priority scheduling in amdgpu
Â
Hi Serguei,
Thanks for the feedback. Answers inline as [AR].
Regards,
Andres
________________________________________
From: Sagalovitch, Serguei [Serguei.Sagalovitch@xxxxxxx]
Sent: Friday, December 16, 2016 8:15 PM
To: Andres Rodriguez; amd-gfx at lists.freedesktop.org
Subject: Re: [RFC] Mechanism for high priority scheduling in amdgpu
Andres,
Quick comments:
1) To minimize "bubbles", etc. we need to "force"Â CU assignments/binding
to high-priority queue when it will be in use and "free" them later
(we do not want forever take CUs from e.g. graphic task to degrade graphics
performance).
Otherwise we could have scenario when long graphics task (or low-priority
compute) will took all (extra) CUs and high--priority will wait for needed resources.
It will not be visible on "NOP "Â but only when you submit "real" compute task
so I would recommend not to use "NOP" packets at all for testing.
It (CU assignment) could be relatively easy done when everything is going via kernel
(e.g. as part of frame submission) but I must admit that I am not sure
about the best way for user level submissions (amdkfd).
[AR] I wasn't aware of this part of the programming sequence. Thanks for the heads up!
Is this similar to the CU masking programming?
[Serguei] Yes. To simplify: the problem is that "scheduler" when deciding which
queue to run will check if there is enough resources and if not then it will begin
to check other queues with lower priority.
2) I would recommend to dedicate the whole pipe to high-priority queue and have
nothing their except it.
[AR] I'm guessing in this context you mean pipe = queue? (as opposed to the MEC definition
of pipe, which is a grouping of queues). I say this because amdgpu only has access to 1 pipe,
and the rest are statically partitioned for amdkfd usage.
[Serguei] No. I mean pipe :-) as MEC define it. As far as I understand (by simplifying)
some scheduling is per pipe. I know about the current allocation scheme but I do not think
that it is ideal. I would assume that we need to switch to dynamical partition
of resources based on the workload otherwise we will have resource conflict
between Vulkan compute and OpenCL.
BTW: Which user level API do you want to use for compute: Vulkan or OpenCL?
[AR] Vulkan
[Serguei] Vulkan works via amdgpu (kernel submissions) so amdkfd will be not
involved. I would assume that in the case of VR we will have one main
application ("console" mode(?)) so we could temporally "ignore"
OpenCL/ROCm needs when VR is running.
>Â we will not be able to provide a solution compatible with GFX worloads.
I assume that you are talking about graphics? Am I right?
[AR] Yeah, my understanding is that pre-empting the currently running graphics job and scheduling in
something else using mid-buffer pre-emption has some cases where it doesn't work well. But if with
polaris10 it starts working well, it might be a better solution for us (because the whole reprojection
work uses the vulkan graphics stack at the moment, and porting it to compute is not trivial).
[Serguei] The problem with pre-emption of graphics task: (a) it may take time so
latency may suffer (b) to preempt we need to have different "context" - we want
to guarantee that submissions from the same context will be executed in order.
BTW: (a) Do you want "preempt" and later resume or do you want "preempt" and
"cancel/abort"? (b) Vulkan is generic API and could be used
for graphics as well as for plain compute tasks (VK_QUEUE_COMPUTE_BIT).
Sincerely yours,
Serguei Sagalovitch
From: amd-gfx <amd-gfx-bounces@xxxxxxxxxxxxxxxxxxxxx> on behalf of Andres Rodriguez <andresr at valvesoftware.com>
Sent: December 16, 2016 6:15 PM
To: amd-gfx at lists.freedesktop.org
Subject: [RFC] Mechanism for high priority scheduling in amdgpu
Hi Everyone,
This RFC is also available as a gist here:
https://gist.github.com/lostgoat/7000432cd6864265dbc2c3ab93204249
[RFC] Mechanism for high priority scheduling in amdgpu
gist.github.com
[RFC] Mechanism for high priority scheduling in amdgpu
[RFC] Mechanism for high priority scheduling in amdgpu
gist.github.com
[RFC] Mechanism for high priority scheduling in amdgpu
We are interested in feedback for a mechanism to effectively schedule high
priority VR reprojection tasks (also referred to as time-warping) for Polaris10
running on the amdgpu kernel driver.
Brief context:
--------------
The main objective of reprojection is to avoid motion sickness for VR users in
scenarios where the game or application would fail to finish rendering a new
frame in time for the next VBLANK. When this happens, the user's head movements
are not reflected on the Head Mounted Display (HMD) for the duration of an
extra frame. This extended mismatch between the inner ear and the eyes may
cause the user to experience motion sickness.
The VR compositor deals with this problem by fabricating a new frame using the
user's updated head position in combination with the previous frames. This
avoids a prolonged mismatch between the HMD output and the inner ear.
Because of the adverse effects on the user, we require high confidence that the
reprojection task will complete before the VBLANK interval. Even if the GFX pipe
is currently full of work from the game/application (which is most likely the case).
For more details and illustrations, please refer to the following document:
https://community.amd.com/community/gaming/blog/2016/03/28/asynchronous-shaders-evolved
Gaming: Asynchronous Shaders Evolved | Community
community.amd.com
One of the most exciting new developments in GPU technology over the past year has been the adoption of asynchronous shaders, which can make more efficient use of ...
Gaming: Asynchronous Shaders Evolved | Community
community.amd.com
One of the most exciting new developments in GPU technology over the past year has been the adoption of asynchronous shaders, which can make more efficient use of ...
Requirements:
-------------
The mechanism must expose the following functionaility:
   * Job round trip time must be predictable, from submission to fence signal
   * The mechanism must support compute workloads.
Goals:
------
   * The mechanism should provide low submission latencies
Test: submitting a NOP packet through the mechanism on busy hardware should
be equivalent to submitting a NOP on idle hardware.
Nice to have:
-------------
   * The mechanism should also support GFX workloads.
My understanding is that with the current hardware capabilities in Polaris10 we
will not be able to provide a solution compatible with GFX worloads.
But I would love to hear otherwise. So if anyone has an idea, approach or
suggestion that will also be compatible with the GFX ring, please let us know
about it.
   * The above guarantees should also be respected by amdkfd workloads
Would be good to have for consistency, but not strictly necessary as users running
games are not traditionally running HPC workloads in the background.
Proposed approach:
------------------
Similar to the windows driver, we could expose a high priority compute queue to
userspace.
Submissions to this compute queue will be scheduled with high priority, and may
acquire hardware resources previously in use by other queues.
This can be achieved by taking advantage of the 'priority' field in the HQDs
and could be programmed by amdgpu or the amdgpu scheduler. The relevant
register fields are:
       * mmCP_HQD_PIPE_PRIORITY
       * mmCP_HQD_QUEUE_PRIORITY
Implementation approach 1 - static partitioning:
------------------------------------------------
The amdgpu driver currently controls 8 compute queues from pipe0. We can
statically partition these as follows:
       * 7x regular
       * 1x high priority
The relevant priorities can be set so that submissions to the high priority
ring will starve the other compute rings and the GFX ring.
The amdgpu scheduler will only place jobs into the high priority rings if the
context is marked as high priority. And a corresponding priority should be
added to keep track of this information:
    * AMD_SCHED_PRIORITY_KERNEL
    * -> AMD_SCHED_PRIORITY_HIGH
    * AMD_SCHED_PRIORITY_NORMAL
The user will request a high priority context by setting an appropriate flag
in drm_amdgpu_ctx_in (AMDGPU_CTX_HIGH_PRIORITY or similar):
https://github.com/torvalds/linux/blob/master/include/uapi/drm/amdgpu_drm.h#L163
The setting is in a per context level so that we can:
   * Maintain a consistent FIFO ordering of all submissions to a context
   * Create high priority and non-high priority contexts in the same process
Implementation approach 2 - dynamic priority programming:
---------------------------------------------------------
Similar to the above, but instead of programming the priorities and
amdgpu_init() time, the SW scheduler will reprogram the queue priorities
dynamically when scheduling a task.
This would involve having a hardware specific callback from the scheduler to
set the appropriate queue priority: set_priority(int ring, int index, int priority)
During this callback we would have to grab the SRBM mutex to perform the appropriate
HW programming, and I'm not really sure if that is something we should be doing from
the scheduler.
On the positive side, this approach would allow us to program a range of
priorities for jobs instead of a single "high priority" value", achieving
something similar to the niceness API available for CPU scheduling.
I'm not sure if this flexibility is something that we would need for our use
case, but it might be useful in other scenarios (multiple users sharing compute
time on a server).
This approach would require a new int field in drm_amdgpu_ctx_in, or repurposing
of the flags field.
Known current obstacles:
------------------------
The SQ is currently programmed to disregard the HQD priorities, and instead it picks
jobs at random. Settings from the shader itself are also disregarded as this is
considered a privileged field.
Effectively we can get our compute wavefront launched ASAP, but we might not get the
time we need on the SQ.
The current programming would have to be changed to allow priority propagation
from the HQD into the SQ.
Generic approach for all HW IPs:
--------------------------------
For consistency purposes, the high priority context can be enabled for all HW IPs
with support of the SW scheduler. This will function similarly to the current
AMD_SCHED_PRIORITY_KERNEL priority, where the job can jump ahead of anything not
commited to the HW queue.
The benefits of requesting a high priority context for a non-compute queue will
be lesser (e.g. up to 10s of wait time if a GFX command is stuck in front of
you), but having the API in place will allow us to easily improve the implementation
in the future as new features become available in new hardware.
Future steps:
-------------
Once we have an approach settled, I can take care of the implementation.
Also, once the interface is mostly decided, we can start thinking about
exposing the high priority queue through radv.
Request for feedback:
---------------------
We aren't married to any of the approaches outlined above. Our goal is to
obtain a mechanism that will allow us to complete the reprojection job within a
predictable amount of time. So if anyone anyone has any suggestions for
improvements or alternative strategies we are more than happy to hear them.
If any of the technical information above is also incorrect, feel free to point
out my misunderstandings.
Looking forward to hearing from you.
Regards,
Andres
_______________________________________________
amd-gfx mailing list
amd-gfx at lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
amd-gfx Info Page - lists.freedesktop.org
lists.freedesktop.org
To see the collection of prior postings to the list, visit the amd-gfx Archives. Using amd-gfx: To post a message to all the list members, send email ...
