Writing elegant host-side CUDA code in Modern C++
Детали
Hello CUDA community,
We're happy to announce our next online meetup!
On February 24th (19:00 UTC+2), we'll talk about writing elegant host-side CUDA code in Modern C++ and more. We will add a link to the Zoom meeting closer to the event so stay tuned.
Speaker:
Eyal Rozenberg
Abstract:
C++ developers who learn to program GPUs using CUDA notice a strange
discrepancy: You can write your kernels in proper C++, and in recent
years - in modern versions of the language (C++11/14/17). But the
host-side code - CUDA's runtime and driver API - looks like bad
spaghetti from the 1980s: Pointers everywhere, including double-pointers
as output parameters; numeric IDs used instead of classes; manual,
explicit management of all resources and errors; etc. And the only
alternative seems to be higher-level parallel frameworks (e.g. thrust)
which force you into a set of abstractions and restrictions before you
can do anything with them.
In this talk we will gradually transform a sample ugly CUDA
caterpillar of a program into a beautiful modern-C++ butterfly (which
does the same thing). The "secret sauce" will be a lightweight,
header-only, expresiness-preserving wrapper library, named
cuda-api-wrappers.
The talk will focus on coding practices, coding style, and some
C++11/C++14 idioms in context of CUDA programming. No in-depth knowledge
is required.
Agenda:
- Welcome message
- The talk
- Post meetup discussions