[Intel-gfx] [RFC 0/2] Compile out integrated

[Tvrtko Ursulin]

On 02/02/2022 16:26, Lucas De Marchi wrote:
> On Wed, Feb 02, 2022 at 10:26:46AM +0000, Tvrtko Ursulin wrote:
>>
>> On 01/02/2022 17:28, Lucas De Marchi wrote:
>>> On Tue, Feb 01, 2022 at 07:09:14PM +0200, Jani Nikula wrote:
>>>> On Tue, 01 Feb 2022, Lucas De Marchi <lucas.demarchi at intel.com> wrote:
>>>>> On Tue, Feb 01, 2022 at 11:15:31AM +0000, Tvrtko Ursulin wrote:
>>>>>> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>>>>>
>>>>>> Quicky and dirty hack based on some old ideas. Thought maybe the 
>>>>>> approach might
>>>>>> interest the Arm port guys. But with IS_GEN_RANGE removed easy 
>>>>>> gains are not so
>>>>>> big so meh.. Maybe some more easy wins with IS_DISPLAY_VER but I 
>>>>>> haven't looked
>>>>>> into that side.
>>>>>>
>>>>>> 3884664  449681    6720 4341065  423d49 i915.ko.tip
>>>>>> 3599989  429034    6688 4035711  3d947f i915.ko.noigp
>>>>>
>>>>> By these numbers probably it's hard to justify. Another thing to 
>>>>> consider
>>>>> is that it's very common to have on the same system both
>>>>> integrated and discrete - doing this would remove at compile time any
>>>>> chance of driving the integrated one.
>>>>
>>>> I guess the point was, the arm systems won't have integrated, and it's
>>>> anyway going to be a separate build.
>>>
>>> so probably the focus and argument here should not be about size
>>> reduction. From patch 1 I see:
>>>
>>> +config DRM_I915_INTEGRATED_GPU_SUPPORT
>>> +       bool "Support integrated GPUs"
>>> +       default y
>>> +       depends on DRM_I915
>>> +       help
>>> +         Include support for integrated GPUs.
>>>
>>> If it's something that depends on arch rather than providing an
>>> option in menuconfig, then I think it could be some interesting
>>> investigation. However, I can't see how it would help with removing
>>> some code paths in the driver (e.g. the clflush() calls we were talking
>>> about in another patch series) since the code elimination would all
>>> happen at link time.
>>
>> Clflush class of problems is yet another orthogonal set of problems.
>>
>> Yes, idea was that the Kconfig option would be selected by Arm, or 
>> deselected by x86, whatever. But there is also a case for letting it 
>> be user visible.
>>
>> In general, I thought at least, we should look into not 
>> building/deploying binary code for irrelevant hardware on Arm builds. 
>> If that is clear and agreeable then I think the approach how to get 
>> there is really multi-pronged.
>>
>> 1)
>> What you are partly doing with "clflush" type series. Make Arm 
>> relevant code paths actually compile on Arm.
> 
> to be clear, the goal is to be more architecture independent
> and this is being tested with arm64.
> 
>>
>> 2a)
>> What I sent in this series - it's simple/easy dead code elimination 
>> from a single compilation unit.
>>
>> 2b)
>> *If* we resurrected GRAPHICS_VER check where "ver" is part of the 
>> macro, eg. not doing "if (GRAPHICS_VER <=> N)" but "if 
>> (GRAPHICS_VERN)", or "if IS_GRAPHICS_VER(N, FOREVER)", then the same 
>> approach would be more effective.
>>
>> Because if N or range is the macro parameter, we can make it dead code 
>> based on Kconfig.
> 
> yes I remember that. But when it was done the IS_GRAPHICS_VER() was
> already not much used. I doubt the handful places it was used in would
> change much of the overall picture we had at that point.

Right, I was referring to the time when N is inside the macro. Like 
before IS_GEN(p, N) was replaced with GRAPHICS_VER, when N went outside 
the macro.

>> This is what I demonstrated few years ago by being able to compile out 
>> ~3rd of a driver when selecting only execlists platforms, AFAIR.
>>
>> And which is why I was a bit unhappy this was getting removed not so 
>> long ago.
>>
>> 3)
>> Complex step is putting LTO into use to allow dead code elimination 
>> between compilation units. Think:
>>
>> file1.c:
>> RUNTIME_INFO->needs_blah = IS_PLATFORM && blah
>>
>> file2.c
>> if (RUNTIME_INFO->needs_blah)
>>   ..dead code eliminated by LTO only..
>>
>> Few years ago Andi Kleen had a proof of concept of KBuild LTO but I 
>> don't know what is the status of that. If LTO can be enabled then work 
>> from steps 2a&b automatically works much much better.
>>
>> 4)
>> If LTO is not there yet maybe see how to slightly refactor the code so 
>> that we can completely drop compilation units from the Makefile. Drop 
>> files like gt/intel_ring_submission.c, or similar if we have or can 
>> create candidates.
> 
> I think this is something we have to do and to go after. I think it's
> more future-proof to organize the code to have a better separation of
> platforms, instead of relying on the linker to do the dead code
> elimination based on user selection.

Yes, definitely high level refactor. I just think it can't be fully done 
and some dead code elimination is cheap enough to have to handle the 
remaining parts.

> I'm fine with compiling out stuff on i915 right now because "this can
> only and will ever (as far as we know) work on x86". I'm more reluctant
> about adding a build option though. For the same reasons Jani mentioned
> and that it suddenly may open the doors for a matrix of configuration
> (e.g.  build with/without display, or rc6, or dsi, or
> name-the-feature-you-dont-want).

Yeah okay, I was not saying it absolutely needs to be user visible. Just 
that it could be argued it could be. Think people building custom 
kernels and selecting AMD CPU and having Intel discrete graphics.

Regards,

Tvrtko