Project Code Review

[riggskevinp]

I liked reading through your code. It's really cool how much of image processing you were able to touch on. A lot of things like variational diffusion go right over my head but they've piqued my interest and left me with a list of things to read up on.

Smart pointers and Cuda

Something I haven't had time to look into but has consistently caught my eye when it comes to modern c++ is memory management.
I think a part of your code provides a good opportunity to explore this.

In main, where you use raw pointers for the input and output pixel arrays, I'd be interested to see how using a unique point might make the usage cleaner.

imageProcessing/imageProcessing.cu

Line 289 in 2a1ee69

inputPixelArrays = imageHandling::allocateImage(depth, width, height);

imageProcessing/imageProcessing.cu

Line 307 in 2a1ee69

imageHandling::freeImage(inputPixelArrays, depth);

What I'm wondering is if there is a clean way to do something like

auto inputPixelArrays = std::make_unique<imageHandleObject>(depth, width, height);

Which would call the imageHandle object's destructor which does the cudafree part. That would require turning your functional library into a class though.

To avoid that, I think it would be possible to create a unique pointer where you pass your function,

imageProcessing/imageHandling.cu

Line 61 in 2a1ee69

void imageHandling::freeImage(double **pixelArrays, int depth) {

,
as the deleter.

I wonder if it would look something like

auto inputPixelArrays = std::make_unique<double>(imageHandling::allocateImage(depth, width, height), imageHandling::freeImage( depth))

Anyway, this still doesn't solve what I'd really like. I'm more interested in cudafree, and it's variants, getting called when the memory goes out of scope. More generally, a way to avoid using cuda malloc and free. Just something I think could be handy.

kSVD

imageProcessing/imageAlgebra.cu

Line 77 in 2a1ee69

cusolverDnDgesvd(

I saw that the api you are using was marked legacy by nvidia in their documentation. What I thought was weird is there isn't a link to a newer or preferred method. Did you happen to find anything like that?

[matthague]

Thanks for pointing out these issues and suggesting make_unique as a way to make my usages cleaner. I'll have to explore that some more and see what I can do!

It does appear that all of the dense eigenvalue solver methods are marked as legacy, cusolverDnXgesvd() should be an updated method that uses the current dense linear solver library. I'll check into using this.