Like the discussion (here)[ Adding integer ceil and floor in Std Dialect - #4 by AlexEichenberger ], when we deal with dynamic shapes, we cannot eliminate the first select. ‘The first selects could be eliminated by having their test being a compile time constant “true” or “false”.’ #map = affine_m…

Same observation here as well. I have quite a few cases that can’t avoid using signed div/mod which makes further optimization unavailable. I suppose you meant to add unsigned operation in the affine expression, not in std dialect. More specifically, ‘uceildiv, ufloordiv and umod’ are needed. and…

Thank you, I very much agree with your point of view. I don’t think we need to add new types. I think we need to add some operations like SignedFloorDivIOp. We can know what types of data are processed and avoid unnecessary instructions.

I am not sure what is the concrete proposal here. Standard dialect no longer contains arithmetic operations. I won’t be opposed to adding unsigned ceildiv and floordiv to the arithmetic dialect, at least for symmetry with the signed versions it already contains. The remainder operation is already av…

imo, we need ‘uceildiv, ufloordiv and umod’ in affine expression. 'affine' Dialect - MLIR Multi-Level IR Compiler Framework @lipracer do you think this can address your issue?

As much as adding operations to the arithmetic dialect is fine with me, extending affine expressions sounds quite problematic. How do you plan to support these new expressions in FlatAffineConstraints? What is the semantics of the PresburgerSet defined with such expressions? cc: @bondhugula

I can try to provide a PR to add the CeilDivSIOp and FloorDivUIOp operation in the arithmetic dialect.

Currently we lack of ability to express unsigned operation in the affine expression itself. For example, ceildiv in affine expression is only treated as signed op and simply lowered using sdiv currently. How can we make it unsigned using unsigned ceildiv in arithmetic dialect? I didn’t expect a bi…

[image] lipracer: We can also add unsigned_ceildiv operation. I prefer the latter, which may have fewer changes. I am very interested in doing this. +1 on this, especially since you have a good motivating example. [image] ftynse: The remainder operation is already available in the arithm…

[image] jungpark: Currently we lack of ability to express unsigned operation in the affine expression itself. For example, ceildiv in affine expression is only treated as signed op and simply lowered using sdiv currently. How can we make it unsigned using unsigned ceildiv in arithmetic dialect? …

Adding unsigned integer ceil and floor in Std Dialect

MLIR

jungpark November 2, 2021, 10:51am 5

imo, we need ‘uceildiv, ufloordiv and umod’ in affine expression.

@lipracer do you think this can address your issue?

Topic		Replies	Views
Adding integer ceil and floor in Std Dialect MLIR	6	1052	October 20, 2020
[RFC] Range analysis when evaluting/lowering affine maps MLIR	11	1125	April 19, 2022
[RFC] Arith dialect versions of affine.apply/min/max MLIR	20	1515	February 3, 2022
[RFC] Define precise arith semantics MLIR	36	2410	October 10, 2022
[Affine] Is it reasonable to extend Affine to support non-index arithmetic? MLIR affine	10	222	December 12, 2025

Adding unsigned integer ceil and floor in Std Dialect

Related topics