[Mesa-dev] [Mesa-stable] [PATCH] radv: adjust the VGT workaround for prim restart on GFX9

[Samuel Pitoiset]

On 11/8/18 5:49 PM, Emil Velikov wrote:
> On Thu, 8 Nov 2018 at 16:01, Samuel Pitoiset <samuel.pitoiset at gmail.com> wrote:
>>
>> On 11/8/18 4:43 PM, Emil Velikov wrote:
>>> HI guys,
>>>
>>> On Tue, 16 Oct 2018 at 22:06, Marek Olšák <maraeo at gmail.com> wrote:
>>>>
>>>> On Thu, Oct 11, 2018 at 4:43 AM Samuel Pitoiset
>>>> <samuel.pitoiset at gmail.com> wrote:
>>>>>
>>>>> WD_SWITCH_ON_EOP seems to be the only workaround that fixes
>>>>> the GPU hangs with Yakuza and The Evil Within on Vega. I don't
>>>>> like as it might decrease geometry performance as pointed out
>>>>> by Marek, but I don't know how to implement a better one.
>>>>>
>>>>> Cc: mesa-stable at lists.freedesktop.org
>>>>> Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
>>>>> ---
>>>>>    src/amd/vulkan/radv_pipeline.c | 13 +++++++++++--
>>>>>    1 file changed, 11 insertions(+), 2 deletions(-)
>>>>>
>>>>> diff --git a/src/amd/vulkan/radv_pipeline.c b/src/amd/vulkan/radv_pipeline.c
>>>>> index 426b417e172..2256b2c58e9 100644
>>>>> --- a/src/amd/vulkan/radv_pipeline.c
>>>>> +++ b/src/amd/vulkan/radv_pipeline.c
>>>>> @@ -3412,14 +3412,23 @@ radv_compute_ia_multi_vgt_param_helpers(struct radv_pipeline *pipeline,
>>>>>           }
>>>>>
>>>>>           /* Workaround for a VGT hang when strip primitive types are used with
>>>>> -        * primitive restart.
>>>>> +        * primitive restart. This fixes GPU hangs with Yakuza and The Evil
>>>>> +        * Within, at least. Not sure if we can implement a better workaround.
>>>>>            */
>>>>>           if (pipeline->graphics.prim_restart_enable &&
>>>>>               (prim == V_008958_DI_PT_LINESTRIP ||
>>>>>                prim == V_008958_DI_PT_TRISTRIP ||
>>>>>                prim == V_008958_DI_PT_LINESTRIP_ADJ ||
>>>>>                prim == V_008958_DI_PT_TRISTRIP_ADJ)) {
>>>>
>>>> Adjacency primitive types should already have wd_switch_on_eop set to true.
>>>>
>>>> The workaround I'm going to use is:
>>>>
>>>>           if (!wd_switch_on_eop && key->u.primitive_restart)
>>>>               partial_vs_wave = true;
>>>>
>>>> Our docs say we should do this. I don't know why it still hangs with
>>>> this workaround.
>>>>
>>> I can see that Marek committed the above suggestion to radeonsi as
>>> commit eae8f49fc65e6e625f5e05d38c3bf1b61b84bd3d
>>> Yet I cannot see an equivalent patch for radv.
>>>
>>> Samuel I'm assuming that you're working on something else and didn't
>>> get back to this one?
>>> Alternatively, please let me know which commit addresses the GPU
>>> hangs, and if we'd want it in stable.
>>
>> Hi,
>>
>> "[PATCH v2] radv: add a workaround for a VGT hang with prim restart and
>> strips" should already be in stable.
>>
>> This one hasn't been pushed because it might decrease performance and we
>> might want to find a better solution.
>>
> The patch mentioned should address the GPU hang for radeonsi.
> Although eop/eoi <> vs_wave handling is flipped in radv, relative to radeonsi.
> 
> If one toggles that, the radeonsi fix could be trivially ported.

There is still a difference between RADV and RadeonSI.

That game (ie. Yakuza 0) is now fixed with D3D/RadeonSI and with Marek's 
patch on Polaris/Vega.

But, it stills hangs with DXVK/RADV on Vega for some reasons. Setting 
partial_vs_wave doesn't seem to be enough, while it correctly fixes 
Polaris...

Needs more investigation.

> 
> HTH
> Emil
>