Babylon and MLIR

[Paul Sandoz]

Hi Ryan,

We have been inspired by the design of MLIR, partly (in theory) to make it easier to translate to MLIR and therefore easier to leverage native MLIR compiler toolchains, just like you indicate.

We did an experiment translating Java code to the Triton MLIR dialect and inputing into Intel’s SPIRV compiler pipeline to generate GPU kernels. That’s as far as we have explored things. It needs to be proved out more, but we don’t have any plans to do more at the moment as we (as in us folks at Oracle) are maxed out focusing on other areas. I was hoping some folks in the community might be interested in further experiments.

HAT could be a very an interesting area to focus on e.g., a HAT accelerator that generates MLIR kernels from Java code in the HAT programming model.

Paul.



On Oct 27, 2025, at 11:47 AM, Ryan Nett <jnett96 at gmail.com> wrote:

Hi folks,

I've seen Babylon's code model's similarity to MLIR be mentioned in a few conference talks, and was wondering if there were any plans to support MLIR as a "destination" via a 1st party library or something like HAT?  There are several existing projects that take MLIR as input and compile it to optimized device code (e.g. IREE, XLA, this list<https://mlir.llvm.org/users/>), so if Java code could easily be converted to MLIR it could open up many integrations and support for additional devices without requiring any work on the Java side.

Thanks,
Ryan

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/babylon-dev/attachments/20251027/880db96f/attachment.htm>