Hi all, We have been working on the openmp “worksharing loop” construct (aka “do”) for MLIR and would like to start a discussion on a couple of design topics we have encountered. A separate RFC on the lowering process for this construct will be posted soon. The OpenMP workshare loop represents a…

Mentioning everyone who commented in D86071 or was pinged there. @mehdi_amini @River707 @jdoerfert @ftynse @bondhugula @nicolasvasilache @clementval @SouraVX

The first has the advantage of allowing us to take advantage of optimisations that already exist in the scf.for operation, or other loop-style operations that we might want to be contained inside the omp operation (for example fir.do). I have to stress this again, a loop “inside” an omp.do is n…

Thanks @jdoerfert for highlighting the issue about illegal transformations. Thanks @DavidTruby for the RFC. Just wanted to add two more points. The openmp do construct can apply to more than one loop, like when it has the collapse(n) clause. In this case what would be the representation? Keeping …

Keeping omp.do separate from the loop to be workshared allows us to remove the omp.do without changing the semantics of the loop. This was suggested as a nice to have by @schweitz . This I don’t understand. In general, omp.do cannot be removed from a loop. The most obvious example uses a firstpriv…

Thanks for the feedback @jdoerfert , given what you’ve outlined here it might be that the simple option of just having one omp loop operation that does not contain another loop is the best approach, since sharing optimisations might be an issue anyway. I wonder if there is a way to share some optimi…

I guess the next question is how to represent collapse if we go this way. One way is to store lower bound, upper bound, and step of each associated loop with the omp.do and only nest the body (plus intermediate statements that we implicitly sink into the innermost body). We then allow OMPIRBuilde…

Thanks for bringing the proposal up for discussion @DavidTruby , and thanks for the feedback @jdoerfert ! A couple of general observations first: [image] DavidTruby: The first has the advantage of allowing us to take advantage of optimisations that already exist in the scf.for operation, or oth…

I added the lowering flow RFC (link below). Had prepared it before reading @ftynse ’s comments. [image] [OpenMP Dialect] Workshare loop lowering flow MLIR The OpenMP worksharing loop specifies that the iterations of a loop will be executed in parallel by the threads in the …

[… reuse optimizations …] Can someone please give me an example of an optimization you want to reuse for omp.do loops? I feel this discussion needs some more concrete details rather than “ideas”. [image] ftynse: Generally, semantic differences lead to a new op being introduced. FWIW, that i…

[image] ftynse: With all this considered, I would suggest to have a separate omp.do operation that has a semantics of a worksharing loop nest , similarly to scf.parallel . That sounds fine. I have a question in general about loops in MLIR. Most of the loops (scf, affine) seem to have only a s…

OpenMP Worksharing Loop RFC

MLIR

kiranchandramohan September 24, 2020, 1:35pm 9

I added the lowering flow RFC (link below). Had prepared it before reading @ftynse’s comments.

Topic		Replies	Views
[OpenMP Dialect] Workshare loop lowering flow MLIR	8	898	September 30, 2020
[RFC] Disambiguation between loop and block associated omp::ParallelOp MLIR openmp	21	630	August 2, 2024
About OpenMP dialect in MLIR LLVM Dev List Archives	23	512	February 19, 2020
[RFC][MLIR] OpenMP Loop Transformation : tile and unroll directive operation support for omp.dialect MLIR llvm	14	1266	October 21, 2022
Flang Technical Call : Summary of presentation on OpenMP for Flang Flang	12	291	September 27, 2019

OpenMP Worksharing Loop RFC

Related topics