Documentation about AMD's HSA implementation?

Sat Mar 17 20:17:34 UTC 2018

>-----Original Message-----
>From: Ming Yang [mailto:minos.future at gmail.com]
>Sent: Saturday, March 17, 2018 12:35 PM
>To: Kuehling, Felix; Bridgman, John
>Cc: amd-gfx at lists.freedesktop.org
>Subject: Re: Documentation about AMD's HSA implementation?
>
>Hi,
>
>After digging into documents and code, our previous discussion about GPU
>workload scheduling (mainly HWS and ACE scheduling) makes a lot more
>sense to me now.  Thanks a lot!  I'm writing this email to ask more questions.
>Before asking, I first share a few links to the documents that are most helpful
>to me.
>
>GCN (1st gen.?) architecture whitepaper
>https://www.amd.com/Documents/GCN_Architecture_whitepaper.pdf
>Notes: ACE scheduling.
>
>Polaris architecture whitepaper (4th gen. GCN)
>http://radeon.com/_downloads/polaris-whitepaper-4.8.16.pdf
>Notes: ACE scheduling; HWS; quick response queue (priority assignment);
>compute units reservation.
>
>AMDKFD patch cover letters:
>v5: https://lwn.net/Articles/619581/
>v1: https://lwn.net/Articles/605153/
>
>A comprehensive performance analysis of HSA and OpenCL 2.0:
>http://ieeexplore.ieee.org/document/7482093/
>
>Partitioning resources of a processor (AMD patent)
>https://patents.google.com/patent/US8933942B2/
>Notes: Compute resources are allocated according to the resource
>requirement percentage of the command.
>
>Here come my questions about ACE scheduling:
>Many of my questions are about ACE scheduling because the firmware is
>closed-source and how ACE schedules commands (queues) is not detailed
>enough in these documents.  I'm not able to run experiments on Raven Ridge
>yet.
>
>1. Wavefronts of one command scheduled by an ACE can be spread out to
>multiple compute engines (shader arrays)?  This is quite confirmed by the
>cu_mask setting, as cu_mask for one queue can cover CUs over multiple
>compute engines.

Correct, assuming the work associated with the command is not trivially small
and so generates enough wavefronts to require multiple CU's. 

>
>2.  If so, how is the competition resolved between commands scheduled by
>ACEs?  What's the scheduling scheme?  For example, when each ACE has a
>command ready to occupy 50% compute resources, are these 4 commands
>each occupies 25%, or they execute in the round-robin with 50% resources at
>a time?  Or just the first two scheduled commands execute and the later two
>wait?

Depends on how you measure compute resources, since each SIMD in a CU can
have up to 10 separate wavefronts running on it as long as total register usage
for all the threads does not exceed the number available in HW. 

If each ACE (let's say pipe for clarity) has enough work to put a single wavefront
on 50% of the SIMDs then all of the work would get scheduled to the SIMDs (4
SIMDs per CU) and run in a round-robin-ish manner as each wavefront was 
blocked waiting for memory access.

If each pipe has enough work to fill 50% of the CPUs and all pipes/queues were
assigned the same priority (see below) then the behaviour would be more like
"each one would get 25% and each time a wavefront finished another one would
be started". 

>
>3. If the barrier bit of the AQL packet is not set, does ACE schedule the
>following command using the same scheduling scheme in #2?

Not sure, barrier behaviour has paged so far out of my head that I'll have to skip
this one.

>
>4. ACE takes 3 pipe priorities: low, medium, and high, even though AQL queue
>has 7 priority levels, right?

Yes-ish. Remember that there are multiple levels of scheduling going on here. At
any given time a pipe is only processing work from one of the queues; queue 
priorities affect the pipe's round-robin-ing between queues in a way that I have
managed to forget (but will try to find). There is a separate pipe priority, which
IIRC is actually programmed per queue and takes effect when the pipe is active
on that queue. There is also a global (IIRC) setting which adjusts how compute
work and graphics work are prioritized against each other, giving options like
making all compute lower priority than graphics or making only high priority
compute get ahead of graphics.

I believe the pipe priority is also referred to as SPI priority, since it affects
the way SPI decides which pipe (graphics/compute) to accept work from 
next.

This is all a bit complicated by a separate (global IIRC) option which randomizes
priority settings in order to avoid deadlock in certain conditions. We used to 
have that enabled by default (believe it was needed for specific OpenCL 
programs) but not sure if it is still enabled - if so then most of the above gets
murky because of the randomization.

At first glance we do not enable randomization for Polaris or Vega but do for
all of the older parts. Haven't looked at Raven yet.

>
>5. Is this patent (https://patents.google.com/patent/US8933942B2/)
>implemented?  How to set resource allocation percentage for
>commands/queues?

I don't remember seeing that being implemented in the drivers.

>
>If these features work well, I have confidence in AMD GPUs of providing very
>nice real-time predictability.
>
>
>Thanks,
>Ming
>
>On Wed, Feb 14, 2018 at 1:05 AM, Ming Yang <minos.future at gmail.com>
>wrote:
>> Thanks for all the inputs.  Very helpful!  I think I have a general
>> understanding of the queue scheduling now and it's time for me to read
>> more code and materials and do some experiments.
>>
>> I'll come back with more questions hopefully. :-)
>>
>> Hi David, please don't hesitate to share more documents.  I might find
>> helpful information from them eventually.  People like me may benefit
>> from them someway in the future.
>>
>>
>> Best,
>> Ming (Mark)
>>
>> On Tue, Feb 13, 2018 at 7:14 PM, Panariti, David <David.Panariti at amd.com>
>wrote:
>>> I found a bunch of doc whilst spelunking info for another project.
>>> I'm not sure what's up-to-date, correct, useful, etc.
>>> I've attached one.
>>> Let me know if you want any more.
>>>
>>> davep
>>>
>>>> -----Original Message-----
>>>> From: amd-gfx [mailto:amd-gfx-bounces at lists.freedesktop.org] On
>>>> Behalf Of Bridgman, John
>>>> Sent: Tuesday, February 13, 2018 6:45 PM
>>>> To: Bridgman, John <John.Bridgman at amd.com>; Ming Yang
>>>> <minos.future at gmail.com>; Kuehling, Felix <Felix.Kuehling at amd.com>
>>>> Cc: Deucher, Alexander <Alexander.Deucher at amd.com>; amd-
>>>> gfx at lists.freedesktop.org
>>>> Subject: RE: Documentation about AMD's HSA implementation?
>>>>
>>>>
>>>> >-----Original Message-----
>>>> >From: amd-gfx [mailto:amd-gfx-bounces at lists.freedesktop.org] On
>>>> >Behalf Of Bridgman, John
>>>> >Sent: Tuesday, February 13, 2018 6:42 PM
>>>> >To: Ming Yang; Kuehling, Felix
>>>> >Cc: Deucher, Alexander; amd-gfx at lists.freedesktop.org
>>>> >Subject: RE: Documentation about AMD's HSA implementation?
>>>> >
>>>> >
>>>> >
>>>> >>-----Original Message-----
>>>> >>From: amd-gfx [mailto:amd-gfx-bounces at lists.freedesktop.org] On
>>>> Behalf
>>>> >>Of Ming Yang
>>>> >>Sent: Tuesday, February 13, 2018 4:59 PM
>>>> >>To: Kuehling, Felix
>>>> >>Cc: Deucher, Alexander; amd-gfx at lists.freedesktop.org
>>>> >>Subject: Re: Documentation about AMD's HSA implementation?
>>>> >>
>>>> >>That's very helpful, thanks!
>>>> >>
>>>> >>On Tue, Feb 13, 2018 at 4:17 PM, Felix Kuehling
>>>> >><felix.kuehling at amd.com>
>>>> >>wrote:
>>>> >>> On 2018-02-13 04:06 PM, Ming Yang wrote:
>>>> >>>> Thanks for the suggestions!  But I might ask several specific
>>>> >>>> questions, as I can't find the answer in those documents, to
>>>> >>>> give myself a quick start if that's okay. Pointing me to the
>>>> >>>> files/functions would be good enough.  Any explanations are
>>>> >>>> appreciated.   My purpose is to hack it with different scheduling
>>>> >>>> policy with real-time and predictability consideration.
>>>> >>>>
>>>> >>>> - Where/How is the packet scheduler implemented?  How are
>>>> >>>> packets from multiple queues scheduled?  What about scheduling
>>>> >>>> packets from queues in different address spaces?
>>>> >>>
>>>> >>> This is done mostly in firmware. The CP engine supports up to 32
>>>> queues.
>>>> >>> We share those between KFD and AMDGPU. KFD gets 24 queues to
>use.
>>>> >>> Usually that is 6 queues times 4 pipes. Pipes are threads in the
>>>> >>> CP micro engine. Within each pipe the queues are time-multiplexed.
>>>> >>
>>>> >>Please correct me if I'm wrong.  CP is computing processor, like
>>>> >>the Execution Engine in NVIDIA GPU. Pipe is like wavefront (warp)
>>>> >>scheduler multiplexing queues, in order to hide memory latency.
>>>> >
>>>> >CP is one step back from that - it's a "command processor" which
>>>> >reads command packets from driver (PM4 format) or application (AQL
>>>> >format) then manages the execution of each command on the GPU. A
>>>> >typical
>>>> packet
>>>> >might be "dispatch", which initiates a compute operation on an
>>>> >N-dimensional array, or "draw" which initiates the rendering of an
>>>> >array of triangles. Those compute and render commands then generate
>>>> >a
>>>> >(typically) large number of wavefronts which are multiplexed on the
>>>> >shader core (by SQ IIRC). Most of our recent GPUs have one micro
>>>> >engine for graphics ("ME") and two for compute ("MEC"). Marketing
>>>> >refers to each
>>>> pipe on an MEC block as an "ACE".
>>>>
>>>> I missed one important point - "CP" refers to the combination of ME,
>>>> MEC(s) and a few other related blocks.
>>>>
>>>> >>
>>>> >>>
>>>> >>> If we need more than 24 queues, or if we have more than 8
>>>> >>> processes, the hardware scheduler (HWS) adds another layer
>>>> >>> scheduling, basically round-robin between batches of 24 queues or 8
>processes.
>>>> >>> Once you get into such an over-subscribed scenario your
>>>> >>> performance and GPU utilization can suffers quite badly.
>>>> >>
>>>> >>HWS is also implemented in the firmware that's closed-source?
>>>> >
>>>> >Correct - HWS is implemented in the MEC microcode. We also include
>>>> >a simple SW scheduler in the open source driver code, however.
>>>> >>
>>>> >>>
>>>> >>>>
>>>> >>>> - I noticed the new support of concurrency of multi-processes
>>>> >>>> in the archive of this mailing list.  Could you point me to the
>>>> >>>> code that implements this?
>>>> >>>
>>>> >>> That's basically just a switch that tells the firmware that it
>>>> >>> is allowed to schedule queues from different processes at the same
>time.
>>>> >>> The upper limit is the number of VMIDs that HWS can work with.
>>>> >>> It needs to assign a unique VMID to each process (each VMID
>>>> >>> representing a separate address space, page table, etc.). If
>>>> >>> there are more processes than VMIDs, the HWS has to time-
>multiplex.
>>>> >>
>>>> >>HWS dispatch packets in their order of becoming the head of the
>>>> >>queue, i.e., being pointed by the read_index? So in this way it's
>>>> >>FIFO.  Or round-robin between queues? You mentioned round-robin
>>>> >>over batches
>>>> in
>>>> >>the over- subscribed scenario.
>>>> >
>>>> >Round robin between sets of queues. The HWS logic generates sets as
>>>> >follows:
>>>> >
>>>> >1. "set resources" packet from driver tells scheduler how many
>>>> >VMIDs and HW queues it can use
>>>> >
>>>> >2. "runlist" packet from driver provides list of processes and list
>>>> >of queues for each process
>>>> >
>>>> >3. if multi-process switch not set, HWS schedules as many queues
>>>> >from the first process in the runlist as it has HW queues (see #1)
>>>> >
>>>> >4. at the end of process quantum (set by driver) either switch to
>>>> >next process (if all queues from first process have been scheduled)
>>>> >or schedule next set of queues from the same process
>>>> >
>>>> >5. when all queues from all processes have been scheduled and run
>>>> >for a process quantum, go back to the start of the runlist and
>>>> >repeat
>>>> >
>>>> >If the multi-process switch is set, and the number of queues for a
>>>> >process is less than the number of HW queues available, then in
>>>> >step #3 above HWS will start scheduling queues for additional
>>>> >processes, using a different VMID for each process, and continue
>>>> >until it either runs out of VMIDs or HW queues (or reaches the end
>>>> >of the runlist). All of the queues and processes would then run
>>>> >together for a process quantum
>>>> before switching to the next queue set.
>>>> >
>>>> >>
>>>> >>This might not be a big deal for performance, but it matters for
>>>> >>predictability and real-time analysis.
>>>> >
>>>> >Agreed. In general you would not want to overcommit either VMIDs or
>>>> >HW queues in a real-time scenario, and for hard real time you would
>>>> >probably want to limit to a single queue per pipe since the MEC
>>>> >also multiplexes between HW queues on a pipe even without HWS.
>>>> >
>>>> >>
>>>> >>>
>>>> >>>>
>>>> >>>> - Also another related question -- where/how is the
>>>> >>>> preemption/context switch between packets/queues
>implemented?
>>>> >>>
>>>> >>> As long as you don't oversubscribe the available VMIDs, there is
>>>> >>> no real context switching. Everything can run concurrently. When
>>>> >>> you start oversubscribing HW queues or VMIDs, the HWS firmware
>>>> >>> will start multiplexing. This is all handled inside the firmware
>>>> >>> and is quite transparent even to KFD.
>>>> >>
>>>> >>I see.  So the preemption in at least AMD's implementation is not
>>>> >>switching out the executing kernel, but just letting new kernels
>>>> >>to run concurrently with the existing ones.  This means the
>>>> >>performance is degraded when too many workloads are submitted.
>>>> >>The running kernels leave the GPU only when they are done.
>>>> >
>>>> >Both - you can have multiple kernels executing concurrently (each
>>>> >generating multiple threads in the shader core) AND switch out the
>>>> >currently executing set of kernels via preemption.
>>>> >
>>>> >>
>>>> >>Is there any reason for not preempting/switching out the existing
>>>> >>kernel, besides context switch overheads?  NVIDIA is not providing
>>>> >>this
>>>> >option either.
>>>> >>Non-preemption hurts the real-time property in terms of priority
>>>> >>inversion.  I understand preemption should not be massively used
>>>> >>but having such an option may help a lot for real-time systems.
>>>> >
>>>> >If I understand you correctly, you can have it either way depending
>>>> >on the number of queues you enable simultaneously. At any given
>>>> >time you are typically only going to be running the kernels from
>>>> >one queue on each pipe, ie with 3 pipes and 24 queues you would
>>>> >typically only be running 3 kernels at a time. This seemed like a
>>>> >good compromise between
>>>> scalability and efficiency.
>>>> >
>>>> >>
>>>> >>>
>>>> >>> KFD interacts with the HWS firmware through the HIQ (HSA
>>>> >>> interface queue). It supports packets for unmapping queues, we
>>>> >>> can send it a new runlist (basically a bunch of map-process and map-
>queue packets).
>>>> >>> The interesting files to look at are kfd_packet_manager.c,
>>>> >>> kfd_kernel_queue_<hw>.c and kfd_device_queue_manager.c.
>>>> >>>
>>>> >>
>>>> >>So in this way, if we want to implement different scheduling
>>>> >>policy, we should control the submission of packets to the queues
>>>> >>in runtime/KFD, before getting to the firmware.  Because it's out
>>>> >>of access once it's submitted to the HWS in the firmware.
>>>> >
>>>> >Correct - there is a tradeoff between "easily scheduling lots of work"
>>>> >and fine- grained control. Limiting the number of queues you run
>>>> >simultaneously is another way of taking back control.
>>>> >
>>>> >You're probably past this, but you might find the original
>>>> >introduction to KFD useful in some way:
>>>> >
>>>> >https://lwn.net/Articles/605153/
>>>> >
>>>> >>
>>>> >>Best,
>>>> >>Mark
>>>> >>
>>>> >>> Regards,
>>>> >>>   Felix
>>>> >>>
>>>> >>>>
>>>> >>>> Thanks in advance!
>>>> >>>>
>>>> >>>> Best,
>>>> >>>> Mark
>>>> >>>>
>>>> >>>>> On 13 Feb 2018, at 2:56 PM, Felix Kuehling
>>>> >>>>> <felix.kuehling at amd.com>
>>>> >>wrote:
>>>> >>>>> There is also this: https://gpuopen.com/professional-compute/,
>>>> >>>>> which give pointer to several libraries and tools that built
>>>> >>>>> on top of
>>>> >ROCm.
>>>> >>>>>
>>>> >>>>> Another thing to keep in mind is, that ROCm is diverging from
>>>> >>>>> the strict HSA standard in some important ways. For example
>>>> >>>>> the HSA standard includes HSAIL as an intermediate
>>>> >>>>> representation that gets finalized on the target system,
>>>> >>>>> whereas ROCm compiles directly to native
>>>> >>GPU ISA.
>>>> >>>>>
>>>> >>>>> Regards,
>>>> >>>>>   Felix
>>>> >>>>>
>>>> >>>>> On Tue, Feb 13, 2018 at 9:40 AM, Deucher, Alexander
>>>> >><Alexander.Deucher at amd.com> wrote:
>>>> >>>>>> The ROCm documentation is probably a good place to start:
>>>> >>>>>>
>>>> >>>>>> https://rocm.github.io/documentation.html
>>>> >>>>>>
>>>> >>>>>>
>>>> >>>>>> Alex
>>>> >>>>>>
>>>> >>>>>> ________________________________
>>>> >>>>>> From: amd-gfx <amd-gfx-bounces at lists.freedesktop.org> on
>>>> >>>>>> behalf
>>>> >of
>>>> >>>>>> Ming Yang <minos.future at gmail.com>
>>>> >>>>>> Sent: Tuesday, February 13, 2018 12:00 AM
>>>> >>>>>> To: amd-gfx at lists.freedesktop.org
>>>> >>>>>> Subject: Documentation about AMD's HSA implementation?
>>>> >>>>>>
>>>> >>>>>> Hi,
>>>> >>>>>>
>>>> >>>>>> I'm interested in HSA and excited when I found AMD's fully
>>>> >>>>>> open-stack ROCm supporting it. Before digging into the code,
>>>> >>>>>> I wonder if there's any documentation available about AMD's
>>>> >>>>>> HSA implementation, either book, whitepaper, paper, or
>documentation.
>>>> >>>>>>
>>>> >>>>>> I did find helpful materials about HSA, including HSA
>>>> >>>>>> standards on this page
>>>> >>>>>> (http://www.hsafoundation.com/standards/) and a nice book
>>>> >>>>>> about
>>>> >>HSA
>>>> >>>>>> (Heterogeneous System Architecture A New Compute Platform
>>>> >>Infrastructure).
>>>> >>>>>> But regarding the documentation about AMD's implementation, I
>>>> >>>>>> haven't found anything yet.
>>>> >>>>>>
>>>> >>>>>> Please let me know if there are ones publicly accessible. If
>>>> >>>>>> no, any suggestions on learning the implementation of
>>>> >>>>>> specific system components, e.g., queue scheduling.
>>>> >>>>>>
>>>> >>>>>> Best,
>>>> >>>>>> Mark
>>>> >>>
>>>> >>_______________________________________________
>>>> >>amd-gfx mailing list
>>>> >>amd-gfx at lists.freedesktop.org
>>>> >>https://lists.freedesktop.org/mailman/listinfo/amd-gfx
>>>> >_______________________________________________
>>>> >amd-gfx mailing list
>>>> >amd-gfx at lists.freedesktop.org
>>>> >https://lists.freedesktop.org/mailman/listinfo/amd-gfx
>>>> _______________________________________________
>>>> amd-gfx mailing list
>>>> amd-gfx at lists.freedesktop.org
>>>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx