On Mar 4, 2011, at 8:56 PM, Chao Yang wrote:

> Hi, Barry,
> 
> So it's just like MatSolve, too sequential.

  Yup.

   BTW: we hope to have a MatSolve() for SeqAIJ fairly soon. It won't be great 
but hopefully a bit better than using the CPU.

   Barry

> Thanks for the hint!
> 
> Chao
> 
> On Mar 4, 2011, at 9:43 PM, Barry Smith wrote:
> 
>> 
>> On Mar 4, 2011, at 3:42 AM, Chao Yang wrote:
>> 
>>> Dear PETSc team,
>>> 
>>> I am developing a code using PETSc on GPU. In the code an SOR 
>>> preconditioner is needed.
>>> Is there an available version of SOR for matrix of seqcusp type in the 
>>> current petsc-dev?
>> 
>>   No. The problem is that SOR is a sequential algorithm, there is not much 
>> parallelism in it and GPUs require/use a great deal of parallelism to get 
>> good performance. 
>> Do you have alternatives? How about diagonal scaled CG or Bi-CG-stab?
>> 
>> 
>>  Barry
>> 
>> 
>> 
>>> Thanks!
>>> 
>>> Best regards,
>>> 
>>> Chao Yang
>> 
> 

Reply via email to