Adding Vulkan support?

[Ulfgard]

Is possible. The computational model fits quite well for compute queues and the interop with graphics is naturally quite a lot better. It could maybe give better performance as one can write really gpu dependent code.

[TGlas]

Sure, but it would be a huge overhead. Imho we should better wait for an abstraction layer for the compute functionality to pop up somewhere.

[Ulfgard]

the overhead is not as big as it might sound. Remora is pretty extensible. And it would be a strong selling point.

here are the steps needed

1. rename gpu_tag to opencl_tag, add vulkan_tag
2. move all ///gpu folders -> opencl folder, fix all includes
3. add the kernels/vulkan implementation (around 2k loc?), include them in kernels/*.hpp
4. implement the remora/vulkan folder (around 3k loc maybe some wrapper code is more expensive idk,).

this also holds for cuda support, etc.