Re: [Starpu-devel] Question about matrix splicing

[christian eder <Christian.Eder@inria.fr>]

Dear Samuel,

thanks a lot, this helped me.

It seems that the strange offset behaviour was also due to my
different attempts trying to get the code working.

If I am using starpu_matrix_filter_block I would get
ptr -> A[0]
dev_handle -> A[0]+1
offset -> 4
and this I will get for any matrix sizes. Due to your explanations
this should be the correct behaviour.

Best regards,
Christian

On Tue, Jul 2, 2013 at 1:58 PM, Samuel Thibault
<samuel.thibault@ens-lyon.org> wrote:
> christian eder, le Mon 01 Jul 2013 18:08:44 +0200, a écrit :
>>   fa.filter_func    = starpu_matrix_filter_block;
>>   fa.nchildren      = l;
>>   fb.filter_func    = starpu_matrix_filter_block;
>>   fb.nchildren      = n;
>>   fc.filter_func    = starpu_matrix_filter_vertical_block;
>>   fc.nchildren      = n;
>>
>> I only get offsets of 1, so sub_a and sub_b are just shifted by 1
>> entry, but not by one row.
>
> See the reference documentation for starpu_matrix_filter_block: “This
> partitions a dense Matrix along the x dimension”. So getting 1 offsets
> is expected. You need to use starpu_matrix_filter_vertical_block to
> partition along the y dimension.
>
>> If I try the same doing by hand, i.e. using
>>
>> STARPU_MATRIX_GET_DEV_HANDLE and STARPU_MATRIX_GET_OFFSET
>>
>> the offset is not correct. So, for example, for two 8x8 matrices I get
>> with the above setting offsets of 4, with 16x16 matrices an offset of
>> 12.
>
> This, however, is more surprising. Could you provide the exact values of
> ptr, dev_handle and offset?
>
> Samuel



-- 
gpg: http://pgp.mit.edu:11371/pks/lookup?op=get&search=0x6E0D034B959FDE6A