Not too long ago we released the new Alea GPU 2.0 release, which was a major step forward for GPU computing on .NET. Today we can announce Alea GPU 2.1. It is a maintenance release but also brings some new interesting features.
First of all Alea GPU 2.1 has integrated cuDNN, a GPU-accelerated library of primitives for deep neural networks, which are very much in vogue these days.
The new version also supports printing from GPU kernels, either with printf/printfn in F# or Console.Write/Console.WriteLine in C# based GPU kernels. This is a very handy tool for quickly debugging GPU kernels or understand them more thoroughly.
Also important is supporting IntPtr in malloc and int64 indexing, which allows to address device memory beyond the 4GB boundary.
Finally, some experts requested support for
atomicCAS has an issue on Linux which could not be resolved in time. We hope it will be fixed with the upgrade to CUDA 7.5, which will be released this summer.
The Alea Tutorial will be updated soon as well with some examples how to use cuDNN directly with Alea GPU.