Halide icon indicating copy to clipboard operation
Halide copied to clipboard

Published 20 hours ago • verified publisher icon halide

Some SVE2 intrinsics are not enabled due to required interleaving.

Open zvookin opened this issue 1 year ago • 3 comments

zvookin avatar Feb 13 '24 03:02 zvookin

See 417d7626dc6fbca970647709fa9314b82014e9a1 .

zvookin avatar Feb 13 '24 03:02 zvookin

It sounds like SVE does things the way Hexagon does. There's quite a non-trivial amount of code in HexagonOptimize.cpp that aims to insert the necessary shuffles, and then to simplify them away as much as possible. It could probably be adapted and generalized for use with SVE and Hexagon.

dsharlet avatar Feb 13 '24 04:02 dsharlet

It sounds like SVE does things the way Hexagon does.

It is not as widely used in SVE2. I need to quantify which instructions are involved. (TODO comment is from Steve Suzuki.) I don't think this is as important re: performance as it is for HVX.

zvookin avatar Feb 13 '24 18:02 zvookin