[Mesa-dev] [PATCH v3 1/8] nir: evaluate if condition uses inside the if branches

[Marek Olšák]

On Thu, Aug 2, 2018 at 8:52 PM, Dieter Nützel <Dieter at nuetzel-hh.de> wrote:
> Am 02.08.2018 05:14, schrieb Marek Olšák:
>>
>> On Wed, Aug 1, 2018 at 5:02 PM, Dieter Nützel <Dieter at nuetzel-hh.de>
>> wrote:
>>>
>>> Am 01.08.2018 19:53, schrieb Marek Olšák:
>>>>
>>>>
>>>> On Wed, Aug 1, 2018 at 12:02 PM, Dieter Nützel <Dieter at nuetzel-hh.de>
>>>> wrote:
>>>>>
>>>>>
>>>>> Am 01.08.2018 17:53, schrieb Marek Olšák:
>>>>>>
>>>>>>
>>>>>>
>>>>>> On Wed, Aug 1, 2018 at 2:39 AM, Dieter Nützel <Dieter at nuetzel-hh.de>
>>>>>> wrote:
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Am 01.08.2018 05:56, schrieb Timothy Arceri:
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> On 01/08/18 13:09, Dieter Nützel wrote:
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Am 31.07.2018 13:34, schrieb Timothy Arceri:
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> On 31/07/18 13:50, Dieter Nützel wrote:
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> Am 30.07.2018 05:24, schrieb Dieter Nützel:
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>> For the series
>>>>>>>>>>>>
>>>>>>>>>>>> Tested-by: Dieter Nützel <Dieter at nuetzel-hh.de>
>>>>>>>>>>>>
>>>>>>>>>>>> with glmark2, glxgears, UH, UV, Blender 2.79 and FreeCAD 0.17
>>>>>>>>>>>>
>>>>>>>>>>>> on RX 580.
>>>>>>>>>>>>
>>>>>>>>>>>> With UH I saw some small light blue triangles spreading around.
>>>>>>>>>>>> Have to bisect which patch set was the culprit. (If I have some
>>>>>>>>>>>> time.)
>>>>>>>>>>>> Tried your's above
>>>>>>>>>>>> configure-bump-libdrm-for-AMDGPU-to-2.4.92.mbox (Samuel)
>>>>>>>>>>>> ASTC-texture-compression-for-all-Gallium-drivers.mbox (Marek)
>>>>>>>>>>>>
>>>>>>>>>>>> Dieter
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> It has something to do with tessellation.
>>>>>>>>>>> If I disable it (F3) all small light blue triangles are _gone_.
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> Are you able to bisect?
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> In short, it has _nothing_ to do with _this_ series.
>>>>>>>>> Tested on e6ff5ac446.
>>>>>>>>> So my tb stands.
>>>>>>>>>
>>>>>>>>> After some sleep (Nothern Germany is _very_ hot and _dry_ this
>>>>>>>>> year)
>>>>>>>>> I'll
>>>>>>>>> do my very best... ;-)
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> hehe. Yes I've experienced a "heat wave" in Northern Europe before,
>>>>>>>> but some of us have very different definitions of _very_ hot and
>>>>>>>> _dry_
>>>>>>>> :P
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> With tonight's shower we had only _4_ (really 3, first day was 10
>>>>>>> minutes
>>>>>>> with only some drops)
>>>>>>> days with rain since last week of April!!!
>>>>>>> Our fields and forests need badly water.
>>>>>>> We never had such dry spring/early summer/summer for ages.
>>>>>>>
>>>>>>> So back to work.
>>>>>>>
>>>>>>> /opt/mesa> git bisect bad
>>>>>>> 86b52d42368ac496fe24bc6674e754c323381635 is the first bad commit
>>>>>>> commit 86b52d42368ac496fe24bc6674e754c323381635
>>>>>>> Author: Marek Olšák <marek.olsak at amd.com>
>>>>>>> Date:   Fri Jul 13 00:23:36 2018 -0400
>>>>>>>
>>>>>>>     radeonsi: reduce LDS stalls by 40% for tessellation
>>>>>>>
>>>>>>>     40% is the decrease in the LGKM counter (which includes SMEM too)
>>>>>>>     for the GFX9 LSHS stage.
>>>>>>>
>>>>>>>     This will make the LDS size slightly larger, but I wasn't able to
>>>>>>> increase
>>>>>>>     the patch stride without corruption, so I'm increasing the vertex
>>>>>>> stride.
>>>>>>>
>>>>>>> :040000 040000 8d457bbd593fe44ca0c7080820f3428f64d5ba19
>>>>>>> 56feae85455f50d6e5ab7a0285db7319d0601afe M src
>>>>>>>
>>>>>>> Reverting it (on top of master + _this_ series) makes all smooth and
>>>>>>> dandy, again - at least on RX580.
>>>>>>>
>>>>>>> Going to bed.
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>> Can you provide more info? Is it only happening with NIR?
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> Yes, only if I set R600_DEBUG sisched,nir (I'm running sisched all the
>>>>> time).
>>>>> Do you need more info?
>>>>
>>>>
>>>>
>>>> It could be a bug in sisched.
>>>
>>>
>>>
>>> It seems to be NOT the case.
>>> R600_DEBUG=nir show the corruption. - But much slower...;-)
>>> You can see it 'best' in scene 4 (cannon) and 1.
>>>
>>> git revert 86b52d4236
>>> fix it (for NIR with and without sisched)
>>
>>
>> What's the LLVM version?
>
>
> I had the same idea.
>
> LLVM 7.0.0svn git 234c23e0a02
> (Sorry, I have some trouble with these rXXXXXX numbers.)
>
> Should I try with the last commit _before_ LLVM 7.0 branch (31.07.2018)
> or with LLVM 8.0.0svn?

The issue only occurs with NIR on Polaris10, not with TGSI. I can't
reproduce it on Vega10.

Marek