hello, I am recently interested in polyhedral technology in mlir. I had read this paper “High Performance Code Generation in MLIR: An Early Case Study with GEMM” – uday. In this paper, @bondhugula mentioned: There are two ways of achieving this tiling in MLIR: one by calling [mlir::tile] (which …

isl is able to generate code given any arbitrary affine schedule. MLIR doesn’t have such a functionality. Instead, each affine transformation utility / pass in MLIR has the code generation step it needs built into it (say tiling, fusion, permute, unroll, unroll-and-jam, separate full and partial til…

Thanks very much for the reply. This makes me more clear about this problem. [image] bondhugula: isl is able to generate code given any arbitrary affine schedule. MLIR doesn’t have such a functionality. Instead, each affine transformation utility / pass in MLIR has the code generation step it n…

First of all, we have a rationale document for the MLIR’s take on polyhedral compilation. It uses an extremely outdated syntax (basically ignore all code in that), but the arguments are still valid. MLIR does not consider ISL since the question of including ISL into LLVM has been discussed repeated…

Thanks for the detailed reply by @ftynse . It is very very helpful. [image] ftynse: First of all, we have a rationale document for the MLIR’s take on polyhedral compilation. It uses an extremely outdated syntax (basically ignore all code in that), but the arguments are still valid. This argu…

[image] xux: As far as I know, the Linalg / affine dialect has different objective from TC / affine scheduler. Is there existing / potential project based on MLIR polyhedral expression and target end-to-end code generation ? It depends on the perspective. The ultimate goal is to generate code…

[image] ftynse: It depends on the perspective. The ultimate goal is to generate code that performs well without spending too much time on compilation. We are not there yet. For context, it took 16 years of research between Featurier’s formulation of Farkas lemma-based affine scheduler and Pluto …

[image] xux: May I ask for some links or references about these topics? What specifically are you interested in? Affine scheduling? Uses of the polyhedral model that are not scheduling? I am not at liberty to discuss the work-in-progress I am aware of, unless all concerned people agree.

[image] ftynse: What specifically are you interested in? Affine scheduling? Uses of the polyhedral model that are not scheduling? I am not at liberty to discuss the work-in-progress I am aware of, unless all concerned people agree. [image] ftynse: There is a growing trend in the literatu…

[image] xux: polyhedral model without affine scheduling, e.g. for memory/dependence analysis, task splitting, etc. Here’s a random assortment of recent things: Stripe (compilation using polyhedral but without affine scheduling) - [1903.06498] Stripe: Tensor Compilation via the Nested Polyh…

That is very kind of @ftynse . I need to dive into these references for a while. Thanks again.

About MLIR polyhedral optimization

MLIR

xux June 19, 2020, 8:00am 9

I am interested in the new trend…

polyhedral model without affine scheduling, e.g. for memory/dependence analysis, task splitting, etc.
compilation with various MLIR flows.(or MLIR DL compiler users)

Some published references / links will be enough if existing…

Topic		Replies	Views
How to join the open MLIR Project MLIR	30	3615	February 11, 2020
RFC: Polygeist LLVM Incubator Proposal MLIR clang	30	4575	April 5, 2022
Non hyper-rectangular loop tiling in the Affine dialect MLIR	4	929	July 27, 2020
Request from a complete noob who is coming to LLVM/MLIR world. Asking for advice/suggestions MLIR gpu , clang	9	934	August 23, 2023
Any front-end framework support such as TF, Caffe? MLIR	15	3710	February 21, 2020

About MLIR polyhedral optimization

Related topics