Open-Source OpenCL Adoption Is Sadly An Issue In 2017

[boboviz]

Well, its AMD. They have good engineers, but their management is awful, and has been like this for a while. Creating OpenCL? Oh, it has been a great idea, it even got traction across opensource devs, who were eager to adopt it instead of using proprietary CUDA and so on.

AMD did NOT create OpenCl.

They've created standard, it got some traction across opensource devs instead of cuda

I repeat: they did NOT create this standard. The standard is up to Khronos Group.

[TheLexMachine]

OpenCL was created by Apple, working with several other parties, including AMD. Basically, all the major OpenCL players. They submitted their collective work to the Khronos Group, which further connected the proposal to other CPU and GPU manufacturers (Embedded/Mobile) who helped refine and add to it and then they rolled that into OpenCL 1.0. The rest is history.

[vasc]

OpenCL 2.x is needlessly complicated to implement by the vendors. They need to add C++ support to the compiler and libraries and validate it. Some of the features like SVM are also not useful on many kinds of hardware and troublesome to implement. We do need SPIR-V though. Because we do not have it, closed source application developers (games and utilities) won't touch OpenCL. The lack of cross platform binaries also makes library development a pain. Because we don't have good libraries, like the CUDA app devs, this makes application development a pain. We application developers only need an OpenCL 1.4 with SPIR-V support and some basic 2.x features which can be easily implemented (e.g. C11 atomics).

Target the coin miner application guys. That should be a great target market since they run their own code and have the resources to buy top end hardware.

He complained that image support is too hard. Well, the thing is, OpenCL images are total crap. I don't use it as an application developer. They don't even support texture arrays, or atlases in most versions. I also think none of the OpenCL versions support mipmapping. This makes them totally useless! I just use plain 2D byte arrays. Images should just be an extension. Fact is we don't need it. So you can just ignore it. FP16 support is much more important in my view.

The guy doing the presentation sounded awfully defeatist to me. Things aren't that bad! Also it's a bad idea to expect OpenCL Next to solve things. It won't and it will take too much time for the spec to get out, let alone for us to see ever it implemented.

FWIW I'm one of the BRL-CAD OpenCL render backend developers. We use the AMD, Intel, and NVIDIA OpenCL implementation and target OpenCL 1.2 since that's the lowest common denominator.

[foadsf]

Here you may find a list of available open source implementations of OpenCL:

https://github.com/fakenmc/cf4ocl/wiki/OpenCL-implementations