A CUDA Kernel Author's Toolkit

On June 19th (19:00 MSK (UTC+3)), we talked about utility functions and classes for writing device-side CUDA code and more. Speaker: Eyal Rozenberg Abstract: While C++ has such a library - CUDA does not: The C++ standard library is intended for host-side code; and while some of it relevant for device-side use - most of it really isn't, even with an adapted implementation. On the other hand - for those of us writing kernels, the needs fulfilled by a language's standard library are not met: We often find ourselves lacking adequate abstraction and generification of hardware features; vocabulary and idioms which would otherwise be repeating snippets in our code; and data structures and algorithms whose semantics involve too much GPU "special sauce" to be written outside of kernels. The CUDA KAT library is a limited collection of contributions towards filling that gap: Those pieces of code you may have found yourself writing again and again into your kernels, which first find their way into your boilerplate includes, then finally become worthy of sprucing up and rounding out into a public library. Note: In the discussion/Q&A following the talk, participants are encouraged to compare and contrast some of their own kernel-authoring idioms with those suggested. https://github.com/eyalroz/cuda-kat