Timing

[wschwanghart]

I just tested the run times of FLOWobj with so-far's implementation of libtopotoolbox of the single flow algorithm. The new implementation still lags behind MATLAB's image processing toolbox, but we are on a good way. The option 'mex', true, uses a self-written steepest gradient and topological sorting which actually does a quite good job.

DEM = GRIDobj('srtm_bigtujunga30m_utm11.tif');

fun = @() FLOWobj(DEM,'single','uselibtt',true,'mex',false);
timeit(fun)

% Returns: 0.5012 seconds

fun = @() FLOWobj(DEM,'single','uselibtt',true,'mex',true);
timeit(fun)

% Returns: 0.3272 seconds

fun = @() FLOWobj(DEM,'single','uselibtt',false,'mex',false);
timeit(fun)

% Returns: 0.3423 seconds

fun = @() FLOWobj(DEM,'single','uselibtt',false,'mex',true);
timeit(fun)

% Returns: 0.2069 seconds

Why is MATLAB so fast? Looking at imreconstruct, the function behind sink filling, uses IPP, Intel's Integrated Performance Primitives (link) which is probably hard to beat.

[wkearn]

We should keep track of these performance numbers. I have an issue (TopoToolbox/libtopotoolbox#132) for this in libtopotoolbox, but it will be much easier to do that from here. I'll create a separate issue for that, and we can use this issue to track performance improvements.