[Mesa-dev] [PATCH 1/2] radeonsi: always use compute rings for clover on CI and newer (v2)

[Michel Dänzer]

On 2019-04-02 4:00 p.m., Michel Dänzer wrote:
> On 2019-04-02 2:57 p.m., Marek Olšák wrote:
>> On Tue, Apr 2, 2019 at 4:57 AM Michel Dänzer <michel at daenzer.net> wrote:
>>> On 2019-04-02 12:39 a.m., Marek Olšák wrote:
>>>> On Mon, Apr 1, 2019 at 1:28 PM Jan Vesely <jan.vesely at rutgers.edu>
>>> wrote:
>>>>> On Mon, 2019-04-01 at 12:30 -0400, Marek Olšák wrote:
>>>>>> Does the attached patch fix the copy-buffer test?
>>>>>
>>>>> it does thanks.
>>>>> Won't the compute only context still need some synchronization?
>>>>> Is there anything else to guarantee that the data is in place after
>>>>> return from resource_copy_region ?
>>>>
>>>> The synchronization is the same on gfx and compute rings.
>>>
>>> BTW, did you see https://bugs.freedesktop.org/show_bug.cgi?id=110214#c24
>>> ? It does indicate some kind of synchronization issue between
>>> si_resource_copy_region using a compute ring and other operations using
>>> a GFX ring.
>>
>> Only OpenCL uses the compute ring. xterm doesn't invoke OpenCL AFAIK.
> 
> That bugzilla comment is about the GTK menu issue, not about xterm.
> Anyway, I doubt GTK uses OpenCL either, and neither does glamor, so it
> was probably an invalid bisect result then.

Apparently not, after all:

https://bugs.freedesktop.org/110355

Looks like there is some issue with si_compute_copy_image, even if it
doesn't use a compute ring.


-- 
Earthling Michel Dänzer               |              https://www.amd.com
Libre software enthusiast             |             Mesa and X developer