On Wednesday, 17 December 2014 at 12:58:23 UTC, ponce wrote:
Hi, I'm kind of embarrassed by my bitter post, must have been a
bad day :).
On Tuesday, 16 December 2014 at 19:49:37 UTC, Shehzan wrote:
We also support CPU and OpenCL backends along with CUDA. This
way, you can use the same ArrayFire code to run across any of
those technologies without changes. All you need to do is link
the correct library.
Cool, this was reason enough to avoid using NPP until now.
I've certainly found desirable to be able to target OpenCL, CPU
or CUDA indifferently from the same codebase. What I'd like
more than a library of functions though is an abstracted
compute API. This would be a compiler from your own compute
language to OpenCL C or CUDA C++ also an API wrapper. That
would probably mean to leave some features behind to have the
intersection. Similar to bgfx but for compute APIs, which has a
shader compiler to many targets.
We used a BSD 3-Clause license to make it easy for everyone to
use in their own projects.
Here is a blog I made about implementing Conway's Game of Life
demonstrates how easy it is to use ArrayFire.
Our goal is to make it easy for people to get started with GPU
programming and break down the barrier for non-programmers to
use the hardware efficiently. I agree that complex algorithms
require more custom solutions, but once you get started,
things become much easier.
Your example is indeed very simple, so I guess it has its uses.
I know this is a really old post, but just to add to what Shehzan
already mentioned, we have double precision support (both real
and complex) since day one (and quite a long time before that as
well). Our documentation does not make it obvious immediately
because we just have a single array class. The array class holds
the metadata of the data types and we eventually launch the
ArrayFire can also integrate with existing CUDA or OpenCL code.
The goal of libraries (be it Thrust or Bolt or ArrayFire) is to
not take back control, but to make sure users are not
re-inventing the wheel over and over again. Having access to
highly optimized, pre-existing GPU kernels for commonly used
algorithms can only increase productivity.