[Intel-gfx] Supporting Intel GPU tracing in gpuvis

[Tvrtko Ursulin]

On 20/11/2017 21:19, Pierre-Loup A. Griffais wrote:
> On 11/20/2017 06:01 AM, Tvrtko Ursulin wrote:
>>
>> Hi,
>>
>> On 16/11/2017 20:42, Michael Sartain wrote:
>>> On Wed, Sep 6, 2017, at 02:09 AM, Chris Wilson wrote:
>>>> Quoting Daniel Vetter (2017-09-06 08:46:50)
>>>>> Hi Pierre,
>>>>>
>>>>> On Tue, Sep 5, 2017 at 11:15 PM, Pierre-Loup A. Griffais
>>>>> <pgriffais at valvesoftware.com> wrote:
>>>>>> Hi Daniel,
>>>>>>
>>>>>> In the past couple of months we've been working on gpuvis, a GPU 
>>>>>> tracing
>>>>>> tool similar to GPUView on Windows. It's lower level than other 
>>>>>> API-based
>>>>>> tracing tools and lets you debug system-wide GPU scheduling 
>>>>>> issues, eg.
>>>>>> interaction between several processes using the GPU, which is pretty
>>>>>> critical for VR usecases.
>>>>>>
>>>>>> It's all based on ftrace; we've initially developped it with 
>>>>>> support for
>>>>>> amdgpu, and had to patch the kernel code there to change what 
>>>>>> tracing events
>>>>>> are reported and how. Now that we have a good idea of what's 
>>>>>> needed and it's
>>>>>> more or less proven in production, we were wondering if you had 
>>>>>> any interest
>>>>>> in adding a similar set of events for Intel GPUs so we could read 
>>>>>> them and
>>>>>> present them the same way? We have pretty specific requirements, 
>>>>>> but this
>>>>>> work-in-progress documentation should give a good idea of what 
>>>>>> they are:
>>>>>>
>>>>>> https://github.com/mikesart/gpuvis/wiki/Overview
>>>>
>>>> We already have those tracepoint equivs and a script to generate a
>>>> similar visualisation: intel-gpu-tools/scripts/trace.pl, but only
>>>> looking at the scheduling issue from the gpu pov. But it's really 
>>>> only a
>>>> dev toy atm, plugging the gap between userspace and the gpu has been on
>>>> the perennial wishlist.
>>>> -Chris
>>>
>>> I added Intel event visualization to gpuvis based on your trace.pl
>>> script. Screenshot at the top of the wiki page here:
>>>
>>> https://github.com/mikesart/gpuvis/wiki/TechDocs-Intel
>>>
>>> In that screenshot the mouse is hovering over the ctx=30,seqno=1900 bar
>>> which selects those events in the event list and shows a tooltip with
>>> the submit, execute, etc info.
>>
>> It certainly looks immensely better than my browser based hack. But 
>> unfortunately I still did not get round actually trying your tool.
>>
>> How scalable it is - meaning - can it handle very busy and huge traces?
> 
> The typical SteamVR trace in our "DVR" plumbing (always tracing in the 
> background) has about 500k events over ~20 seconds. I can zoom and scrub 
> through it at 60fps without issues here.

Sounds quite good!

>> Is there any outlook of it getting packaged in some distro?
> 
> The tool itself is standalone and very easy to package; Mike recently 
> wrote some plumbing example scripts around it that can get useful 
> captures, so I would think we're in better shape for packaging now than 
> before they existed. We bundle it with SteamVR so haven't really spent 
> any time looking into packaging.

Okay, I was just curious and there probably isn't any direct need from 
our side for this. Only that it can help with adding new interesting 
bits to the kernel. Tracepoints are kind of neither here or there in 
this respect.

>>> For the amdgpu driver, we're able to get the submit information from
>>> user space and associate those events to specific processes. Example of
>>> that is here:
>>>
>>> https://github.com/mikesart/gpuvis/wiki/TechDocs-AMDGpu
>>>
>>> If you ever get a chance to try gpuvis and have any feedback, we'd love
>>> to hear it.  Also if you ever get userspace tracepoint data in, let me
>>> know and I'd be happy to hook that up as well.
>>
>> What kind of information is missing to wire up this missing bit? I 
>> mean the thing you are referring to as user space submit data, what is 
>> that?
> 
> The main thing is to be able to associate a chunk of GPU work with a 
> userspace process. Currently, there's no tracepoint we're aware of in 
> the work submission ioctl, which means that while the GPU tasks are 
> properly displayed, they're anonymous. The main intent behind gpuvis is 
> to be able to debug multi-process GPU interaction and timing problems, 
> so having a tracepoint in the submit ioctl lets us associate the user 
> context of the application that submitted the work with the work itself 
> if it shares the same identifying seqnos/ids. gpuvis can then show you 
> thread information about that process, color-code for easy 
> disambiguation, etc.

We have i915_gem_request_queue which is at the submit time (execbuf 
ioctl), can you have a look and see if that would be enough for gpuvis?

>> Btw one new thing we are close to merging to i915 is the perf PMU 
>> support. That will enable real-time monitoring of per-engine busyness, 
>> waits, frequency, power, maybe more in the future like queue-depth. I 
>> don't know if things like that would be interesting for gpuvis? Some 
>> of it can be inferred from the tracepoints already in post-processing 
>> so there is some overlap. I am not sure, but thought to mention it. 
>> This is the series: https://patchwork.freedesktop.org/series/27488/. 
>> It is used via existing perf userspace API.
> 
> Seems like the kind of thing gpuvis could happily overlay on top of the 
> trace with its plotting functionality.

Cool. And it even sounds like we are close to merging this so keep an 
eye on it!

Regards,

Tvrtko