[Starpu-devel] Implementation question regarding output type

[Johan Mazel <johan.mazel@gmail.com>]

Hello
I am using StarPU to test ranges of values and identify "good" ones inside these ranges.
I however cannot predict the number of "good" values inside a range.
The output of the method is thus a list of "good" values which is of variable size.

My current naive implementation returns a vector of size equal to the number of elements in the range. Inside this output vector, 0 means "bad" and 1 means "good".
This is quite slow because I guess that StarPU has to transfer back all the result vectors to the RAM.

I identified two possible solutions :

* Define a new data interface (http://starpu.gforge.inria.fr/doc/html/DataManagement.html#DefiningANewDataInterface) with bitset or list.

If I use a list, the best solution would be an extendable list but I may also be able to use a fixed length list if I make an hypothesis on the number of "good" values in a range.

* Use the CSR data interface.

I found two examples that uses it for input data but none that use it for output. One of these examples (spmv.c) also implements row filtering which fits my use case.

If I understand correctly, it may be possible to use CSR as output if I can make an hypothesis on the total number of "good" values in each partitioned subpart of data.

Which one of these two methods looks do you think is the most appropriate for my use case? Do you know any examples that use a similar method?

Thank you for your time.

Best regards,

Johan