Using Matconvnet on recent GPU platforms

[jacqu...@gmail.com]

Trying to compile Matconvnet with recent versions of Cuda (11.x) to make it usable on the latest GPU cards, I ran into the following issues. Compilation generally works fine, bur when I try to run some training iterations a bunch of formatting errors occur, e.g. in functions vl_nnrelu.m, Layer.m, and others.

It always occur because some gpu array storing one of the network variable gets its format changed, making it incompatible for further operations (addition, multiplication,...). The issue can be partly fixed by inserting simple tests to do the operations only when the formats are compatible, but this seems rather shaky.

I wonder where these format changes can come from: is it the different behaviors of different Cuda versions, or something else? Of course the best solution would be to update the Matconvnet compiler. Is there any chance that this might be done in the near future?

Thanks in advance for any answer,

Jacques Boutet de Monvel