[Mesa-dev] No luck with tstellar/llvm on HD4890

[Andy Furniss]

Andy Furniss wrote:
> Andy Furniss wrote:
>> Tom Stellard wrote:
>>> On Wed, Nov 07, 2012 at 09:24:13PM +0000, Andy Furniss wrote:
>>>> Vincent Lejeune wrote:
>>>>> git seems to have trouble sending my patch to ML atm, can you
>>>>> manually apply it ? It fixes lock up here
>>>>>
>>>>> diff --git a/src/gallium/drivers/r600/r600_asm.c
>>>>> b/src/gallium/drivers/r600/r600_asm.c
>>>>> index 5f2548e..f06af44 100644
>>>>> --- a/src/gallium/drivers/r600/r600_asm.c
>>>>> +++ b/src/gallium/drivers/r600/r600_asm.c
>>>>> @@ -2961,6 +2961,6 @@ void r600_bytecode_export_read(struct
>>>>> r600_bytecode_output *output, uint32_t wor
>>>>>    output->swizzle_w = G_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_W(word1);
>>>>>    output->burst_count =
>>>>> G_SQ_CF_ALLOC_EXPORT_WORD1_BURST_COUNT(word1);
>>>>>    output->end_of_program =
>>>>> G_SQ_CF_ALLOC_EXPORT_WORD1_END_OF_PROGRAM(word1);
>>>>> -output->inst =
>>>>> EG_S_SQ_CF_ALLOC_EXPORT_WORD1_CF_INST(G_SQ_CF_ALLOC_EXPORT_WORD1_CF_INST(word1));
>>>>>
>>>>>
>>>>> +output->inst =
>>>>> R600_S_SQ_CF_ALLOC_EXPORT_WORD1_CF_INST(G_SQ_CF_ALLOC_EXPORT_WORD1_CF_INST(word1));
>>>>>
>>>>>
>>>>>    output->barrier = G_SQ_CF_ALLOC_EXPORT_WORD1_BARRIER(word1);
>>>>>    }
>>>>> --
>>>>
>>>> This helps - on mesa head, demos/openarena/nexuiz run OK (nexuiz
>>>> stutters a bit initially compared to R600_LLVM=0, but then I am running
>>>> a debug build of llvm).
>>>>
>>>> etqw now locks GPU as soon as it loads when using llvm.
>>>>
>>>
>>> Is this a regression caused by the same commit?
>>
>> Yes, but etqw has problems anyway.
>>
>> Reset mesa to 84b437213294ff4e1a3bcae2f9cbb36a9b4955c4 applied patch and
>> it gives GPU reset.
>>
>> Reset mesa to the commit before that + no patch and it runs, but gives
>> the page allocation fails etc. that I posted earlier in the thread.
>>
>> Even with R600_LLVM=0 or llvm 3.1 etqw has some isssues with current
>> mesa, but they are not the same as these.
>
> I've found a more trivial case to reset GPU.
>
> I see the patch went in and I am testing on mesa head -
>
> The mesa demo dissolve will cause a reset, all other demos in that dir
> work.

Dissolve and etqw are working now with current llvm and mesa.