[Mesa-dev] radeonsi: ppc64 glamor backtrace

[Michel Dänzer]

On 22/06/17 05:58 PM, Jack Mitchell wrote:
> On 22/06/17 03:32, Michel Dänzer wrote:
>> On 22/06/17 02:56 AM, Grazvydas Ignotas wrote:
>>> Looks like nobody tested radeonsi on BE for 5 months at least. You can
>>> try the attached patch, but I suspect there will be other places like
>>> this...
> 
> This got me further but it now looks like something fails in the kernel
> side of the driver and the card locks up. Thanks for the patch.
> 
> radeon 0001:01:00.0: ring 0 stalled for more than 10256msec
> radeon 0001:01:00.0: GPU lockup (current fence id 0x0000000000000001
> last fence id 0x0000000000000002 on ring 0)
> 
>>
>> Yeah, this is just the tip of the iceberg. There are many reasons why
>> radeonsi currently can't (and never could) work on big endian hosts.

Hmm, I guess my wording was a bit confusing. I mean "never could" as in
radeonsi has never been able to work on big endian hosts from when it
was created until now. It's probably possible to make it work, but it
would likely require a lot of effort.


> That's a shame to hear, would you mind giving an example of why it's not
> feasible?

Just as one example, the hardware doesn't support loading float values
stored in memory as big endian from shaders. So any float values (e.g.
vertex attributes) read by a shader have to be converted to little
endian first somehow.

There are probably various other similar issues.


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