[llvm-dev] [mlir] Re: "MLIR meets LLVM" Round-Table at the LLVM Dev Meeting

[Vedant Kumar via llvm-dev]

> On Oct 21, 2019, at 11:34 AM, Mehdi AMINI via llvm-dev <llvm-dev at lists.llvm.org> wrote:
> 
> Hi all,
> 
> As a reminder, the round-table will be Tuesday, Oct 22, 10:20-11:30, see here: http://llvm.org/devmtg/2019-10/#round <http://llvm.org/devmtg/2019-10/#round>
> 
> Here is a potential list of all the topics I gathered from everyone, let me know if I missed anything.
> 
> - Intro, practical aspect of the integration (repository, etc.)
> - Can we have a machine-readable definition of LLVM IR in LLVM itself: instructions, tracking changes, altering the semantics in the dialect wrt LLVM IR. We could reuse this for the MLIR LLVM Dialect.

The need for a machine-readable description of LLVM IR in LLVM itself has come up in past discussions re: the function merging pass. When LLVM IR is extended, it would be great if the function comparator utility automatically learned which IR properties to diff instead of silently miscompiling.

vedant


> - Clang: would the static analyzer be a good candidate for a project to be playing with a MLIR Dialect?
> - Using MLIR for SCEV, VPLAN, etc. Isn’t SCEV an IR?
> - MIR: could/should it be reimplemented as a MLIR dialect? 
> - What about Alive? (If Alive folks are around?)
> - Representing parallelism in MLIR
> - Heterogeneous compilation: this includes GPU of course, but also more exotic platforms / third-party accelerator: how to plug all this into a codegen story?
> - Using MLIR dialects as a testbed for proposed LLVM IR changes (there were proposals of tensor types for LLVM for example). 
> - Using MLIR for loop optimization and other higher-level transformations, e.g. doing LLVM->Polly->MLIR Affine->LLVM.- 
> 
> Note also that at shortly after the round-table, at 12:05pm the talk about An MLIR Dialect for High-Level Optimization of Fortran <https://llvm.org/devmtg/2019-10/talk-abstracts.html#tech19> will take place.
> 
> Best,
> 
> -- 
> Mehdi
> 
> 
> On Fri, Oct 18, 2019 at 7:21 AM 'James Molloy' via MLIR <mlir at tensorflow.org <mailto:mlir at tensorflow.org>> wrote:
> > Some existing bits in LLVM that could have used MLIR are ScalarEvolution expressions, VPlan, SelectionDAG.  Moving these would be a lot of work obviously but what are some new things under development that could use MLIR from day one?
> 
> I'd add to this list the MIR format, which is clearly crying out for MLIR-ification.
> 
> On Fri, 18 Oct 2019 at 14:00, Lei Zhang <antiagainst at google.com <mailto:antiagainst at google.com>> wrote:
> FYI: related to GPU compilation, there will also be a "Challenges using LLVM for GPU compilation" roundtable covering
> 
>  - Canonicalization vs. GPUs: Type mutation;
>  - Control flow mutation (graphics shaders are more sensitive to this);
>  - Divergence/reconvergence sensitivity;
>  - Address-space awareness;
> 
> Time slot hasn't been finalized but preliminary planned for Tue 3-4pm or Wed 2-3pm.
> 
> http://lists.llvm.org/pipermail/llvm-dev/2019-October/136033.html <http://lists.llvm.org/pipermail/llvm-dev/2019-October/136033.html>
> 
> Thanks,
> Lei
> 
> 
> On Tue, Oct 15, 2019 at 12:01 PM Sanjoy Das <sanjoy at google.com <mailto:sanjoy at google.com>> wrote:
> I won't be present during the round table, but it might be interesting to discuss if there are ways to use MLIR within LLVM.
> 
> Some existing bits in LLVM that could have used MLIR are ScalarEvolution expressions, VPlan, SelectionDAG.  Moving these would be a lot of work obviously but what are some new things under development that could use MLIR from day one?
> 
> -- Sanjoy
> 
> On Tue, Oct 15, 2019 at 3:59 AM 'Alex Zinenko' via MLIR <mlir at tensorflow.org <mailto:mlir at tensorflow.org>> wrote:
> I would be interested in practical aspects:
> - practical connection between MLIR's LLVM dialect and actual LLVM IR: having a machine-readable definition of instructions, tracking changes, altering the semantics in the dialect wrt LLVM IR;
> - using MLIR dialects as a testbed for proposed LLVM IR changes (there were proposals of tensor types for LLVM for example).
> 
> This may be partially orthogonal, but may be interesting:
> - using MLIR for loop optimization and other higher-level transformations, e.g. doing LLVM->Polly->MLIR Affine->LLVM.
> 
> 
> On Tue, Oct 15, 2019 at 1:27 AM Mehdi Amini <aminim at google.com <mailto:aminim at google.com>> wrote:
> Hey all,
> 
> FYI the round-table will be Tuesday, Oct 22, 10:20-11:30, see here: http://llvm.org/devmtg/2019-10/#round <http://llvm.org/devmtg/2019-10/#round>
> 
> It would be nice to have an agenda now, what are your thoughts on topics we should cover?
> 
> Cheers,
> 
> -- 
> Mehdi
> 
> 
> 
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20191021/1fcbb10b/attachment.html>