Hi -
Thanks Michael for the reply. I agree that I do not perceive an MLIR
dialect significantly helping a current Chapel user. As you have rightly
pointed, I was coming more from the angle of an ML programmer who may be
interested in using Chapel (or Chapel-like syntax/semantics) to target a modern
hardware accelerator built for ML (many of them resemble distributed meshes).
In which case an MLIR dialect may help map the Chapel syntax/semantics to the
target hardware more efficiently.
Please correct me if I am misunderstanding - but I think you are
saying that you see MLIR support as a more likely path towards
supporting hardware built for machine learning. It's not clear to
me personally why that would necessarily be the case, but we
can certainly imagine it potentially working if e.g. the Chapel
MLIR dialect was a subset of the tensorflow one.
For a compiler to support any hardware platform, somebody
has to do the work to enable the support. For things like LLVM,
the hardware support can be shared by lots of compilers.
But, because MLIR is intended to be very customizeable, and in
particular the MLIR dialects are not necessarily related to each other,
it's not clear to me why using MLIR would help to re-use hardware
support that other compilers (e.g. tensorflow) are using.
Thanks,
-michael
-----Original Message-----
From: Michael Ferguson <[email protected]>
Sent: Thursday, January 2, 2020 7:40 PM
To: Das, Dibyendu <[email protected]>
Cc: [email protected]
Subject: Re: [Chapel-developers] Chapel and MLIR
Hi -
Creating an MLIR dialect is worth considering. I think that conceptually,
MLIR is a reasonable fit. It's less clear to me if in practice the MLIR code
base and community will be ready for other use cases besides tensorflow. To
some degree I'd like to see how it goes with f18.
However, most of the benefits I can imagine of using MLIR or something like
it are at the level of making the compiler easier to write. I wasn't aware of
situations in which MLIR would significantly change the experience of a Chapel
user (as compared to, say, other approaches to encode Chapel semantics into
LLVM IR). Do you know of such a case, where having some MLIR integration would
make a difference for somebody using Chapel?
Thanks,
-michael
Hi-
Is there any plan of creating an MLIR dialect for Chapel ? Because I
think such an effort may be useful in making Chapel a programming language that
is useful for machine learning - especially for training on distributed
hardware.
Thx
-Dibyendu
_______________________________________________
Chapel-developers mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/chapel-developers
