[petsc-dev] Current status: GPUs for PETSc

[Karl Rupp]

On 11/05/2012 08:23 AM, Lawrence Mitchell wrote:
> On 05/11/2012 14:18, Matthew Knepley wrote:
>>      MatSetColumnIndices(m, ...);
>>
>> This step is unnecessary.
>
> How else is the matrix going to know what the column of the data is?  I'm
> envisaging not calling MatSetValues on the device, so afaict, there's no
> other way of PETSc knowing.
>

I think the SetColumnIndices() step can possibly be dealt with in the 
same way as the entries, e.g.
   handle1->column_data = your_col_data_array;
   handle2->data = your_matrix_entries_array;
because you probably don't want to use the CPU-bound 
MatSetColumnIndices() for data that already resides on the GPU.

Best regards,
Karli