[SoC] Parallelization of RST modules of GRASS GIS on GPU

[Helena Mitasova]

Do the system of linear equations first - it should help on several levels
because it will also make it possible to run the interpolation with fewer larger segments
and larger neighborhood reducing the instances when segments may be visible
in the resulting grid.

Computing grid may be time consuming for certain values of tension - I would have to look at it.
Checking the interpolating function at given points should be made optional - it depends on how
many points there are - it could take as much time as computing the grid.

Finding points should not take too much time unless the spatial distribution of points
is very heterogeneous, but I would have to look at it.

you can write me off-this list (maybe cc to GRASS dev list) as SoC list is mostly
to organize GSoC (at least that is how I understand its purpose),

Helena


Helena Mitasova
Associate Professor
Department of Marine, Earth and Atmospheric Sciences
North Carolina State University
Raleigh, NC 27695
hmitaso at unity.ncsu.edu



On Mar 15, 2010, at 12:09 PM, Abhishek Shukla wrote:

> By what I understood about RST, I concluded that these two steps (after performing segmentation) consume most of the time.
> 
> (1). Finding points from neighbouring segments so that the total number of points for calculation of interpolation function is between some given minimum and maximum.
> (2). Creating and solving the system of linear equations using these points.
> 
>  First step is also affected by segmentation algorithm.
> 
>  Are steps which include "checking the interpolating function values at given points" and "computing grid for each segment" also time
> consuming ? (should not be , though)
> 
> Anyways, I hope all steps would get some improvement when we'll do it on GPU. 
> Please inform me what I got correct and what I got wrong.
> I have good understanding of GPU parallelization and I have got good guidance on it in my university.
> 
> 
> Most importantly , please inform me which steps are most time consuming ( in v.surf.rst  & v.vol.rst ) so that I can concentrate on parallelizing those steps.
> 
> ---
> Abhishek Shukla
> Center for Security, Theory And Algorithm Research (CSTAR)
> International Institute of Information Technology
> Hyderabad
> India
> http://www.iiit.ac.in/