Preallocate imfilter kernel memory

[ClaroHenrique]

This PR aims to optimize the kernel padding in GPU imfilter. The idea is to preallocate a CuArray with the same size of TI and copy the kernel data to this array, avoiding the full padded kernel allocation in every imfilter operation.

[codecov-commenter]

Codecov Report

Merging #45 (1d935d2) into master (3eebf4b) will decrease coverage by 0.62%.
The diff coverage is 0.00%.

❗ Current head 1d935d2 differs from pull request most recent head ce926cc. Consider uploading reports for the commit ce926cc to get more accurate results

@@            Coverage Diff             @@
##           master      #45      +/-   ##
==========================================
- Coverage   85.71%   85.08%   -0.63%     
==========================================
  Files           9        9              
  Lines         420      456      +36     
==========================================
+ Hits          360      388      +28     
- Misses         60       68       +8     

Impacted Files	Coverage Δ
src/imfilter.jl	11.76% <0.00%> (-4.91%)	⬇️
src/relaxation.jl	93.10% <0.00%> (-3.20%)	⬇️
src/iqsim.jl	96.85% <0.00%> (-0.63%)	⬇️
src/geostats.jl	100.00% <0.00%> (ø)
src/graphcut.jl	100.00% <0.00%> (ø)
src/taumodel.jl	100.00% <0.00%> (ø)
src/plotrecipes.jl	0.00% <0.00%> (ø)
src/utils.jl	94.87% <0.00%> (+0.42%)	⬆️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

[juliohm]

@ClaroHenrique I was looking into the code and perhaps we can improve the situation by refactoring the imfilter_kernel and fastdistance functions to accept the img and the krn with the same size already. So instead of adding a third argument to all functions we could simply pass the two arrays with the same size (krn padded) and a third argument with the cartesian indices if necessary. That will be more intuitive to maintain. Do you understand what I mean? Can you work on that?

[ClaroHenrique]

@ClaroHenrique I was looking into the code and perhaps we can improve the situation by refactoring the imfilter_kernel and fastdistance functions to accept the img and the krn with the same size already. So instead of adding a third argument to all functions we could simply pass the two arrays with the same size (krn padded) and a third argument with the cartesian indices if necessary. That will be more intuitive to maintain. Do you understand what I mean? Can you work on that?

Working on that.

[juliohm]

Hi @ClaroHenrique I think we are taking too long to address these optimizations in the code. Can we set a meeting to discuss a plan that is effective and quick to execute?

[juliohm]

@ClaroHenrique I merged it assuming that you saw good speedups with the change. I will test it locally now to see how it goes on my NVIDIA GPU.