[Mesa-dev] [PATCH 7/7] intel: tools: add aubinator_error_decode tool

[Lionel Landwerlin]

On 30/03/17 20:09, Chris Wilson wrote:
> On Thu, Mar 30, 2017 at 11:27:26AM -0700, Matt Turner wrote:
>> I think we should figure out how to make this not just a fork of
>> intel_error_decode. Should intel_error_decode do away?
>>
>> There are various tools in i-g-t that I'm definitely in favor of
>> moving into Mesa (like the assembler and disassembler, and aubdump).
>> Should this one move too? Do we have buy-in from Chris?
> Sure, if it means that we do get an actively maintained decoder - just
> hope that gen2-3 is forthcoming, and perhaps some inferred type
> decoding.

I guess that depends on whether we can get somebody to write genxml for 
them.
Not sure it's a big priority for us :/
Maybe a rainy weekend project?

>
> What's more important is that it is fairly flexible in error-state file
> format - mostly the lists and order of registers change, different
> amount of detail in the kernel state trackers etc. I was thinking that
> maybe we wanted a more parseable format like json (error.json),
> something a bit more flexible?

That would be great.

> What I also want is a use-case for EXEC_OBJECT_CAPTURE. Aside from
> plain listing in the error state, it would be nice if the decoder could
> parse 3DSTATE_BASE to work out which buffer was, for example, the
> instruction buffer and then decode the relevant kernels from 3DSTATE_PS.

If I understood things right, I think that's what Matt's working on.

> Decoding vertex buffers is perhaps less interesting in general that it
> was for me, but similarly if they could be found within the error state,
> dumping the vertices for each PRIMITIVE -- though with GS/HS/TS that
> again is probably useless -- so really just an example of being
> flexible, if the data is there in the error state, decode it inline.

That's something I'm also interested in.
I would really like to be able to inspect the state of the GPU for each 
draw call in aubinator for example.
Lines of text aren't the best way to inspect though.

A few months back, I started to write an aubinator.py script in the hope 
of having some UI on top.
I think we need a slightly better genxml description to have that 
completely automatic though, as it stands we miss a some relationships 
between offsets/address & STATE_BASE_ADDRESS for example.

>
> I was won over at having pager support. I was getting close to moving
> the decoder to mesa myself, though it would have been a much less
> refined effort!
> -Chris
>