vulkan?

[Neal Becker]

I have a laptop with (recent) intel gpu. I see it is supported by vulkan,
and I have a driver (linux).

I wonder if there's plans to support vulkan on numpy?

[Neal Becker]

I meant, support vulkan on numba?

[Kevin Sheppard]

Vulcan is a Kronos-led common low-level 3D and compute API.  I would wager that it is sufficiently limited when compared to CUDA that targeting Vulkan isn't likely.

--
You received this message because you are subscribed to the Google Groups "Numba Public Discussion - Public" group.
To unsubscribe from this group and stop receiving emails from it, send an email to numba-users...@continuum.io.
To post to this group, send email to numba...@continuum.io.
To view this discussion on the web visit https://groups.google.com/a/continuum.io/d/msgid/numba-users/na4esr%24as2%242%40ger.gmane.org.
For more options, visit https://groups.google.com/a/continuum.io/d/optout.

[Stanley Seibert]

This is something we are keeping an eye on, but it is not a high priority at the moment.

Some background: There was an experimental branch of Numba in summer 2014 where we (by which, I mean Siu) implemented an OpenCL target based on SPIR (which is based on the LLVM IR).  It was very challenging to get working because we found that SPIR support was incomplete or buggy, and the memory model was very limiting.  Pointers had to be tagged with which memory region they operated on (local, global, constant, etc), and this turned out to be very complicated to manage in a compiler designed to work with a flat memory space.  Ultimately, we had to abandon the experiment.

SPIR-V supports generic memory spaces but is not based on the LLVM IR.  However, it sounds like there are tools to translate, so it is possible we could have a Vulkan target in the future, assuming that SPIR-V is well supported.  We are hoping to find ways to share more code between our GPU targets (CUDA and HSA), so that it will be easier to add new GPU targets (like Vulkan) in the future.

To view this discussion on the web visit https://groups.google.com/a/continuum.io/d/msgid/numba-users/CAE8Ss2G6qxL%3DzA07L0BZ4vMBU9qa3U64OZEOkOJjgv35%3DmfTFA%40mail.gmail.com.

[Hugh Perkins]

I might take a look at this. I have a compiler taht can run NVIDIA® CUDA™ code on OpenCL 1.2 devices, ie https://github.com/hughperkins/coriander .  Since it uses llvm, and handles the static address space declarations etc, seems like not impossible it could handle numba perhaps?