[Mlir-commits] [mlir] [mlir][bufferization] Test tensor encoding -> memref layout conversion (PR #161166)

[Andrei Golubev]

andrey-golubev wrote:

@matthias-springer I've redone the tests completely (luckily, I don't need to test in unit C++ but can do everything in LIT) and also dropped the API update. This is effectively "new" PR but I repurposed the old one.

https://github.com/llvm/llvm-project/pull/161166#discussion_r2406069568 shows where I would rely on memref layout inference function in bufferization options (that is configured by a pass). But at this point it seems to be a "convenience" - since custom operations provide own bufferization anyway.

Anyhow, since we already have RankedTensorType::getBufferType() and function to get default memory space, maybe it makes sense to provide default layout inference.

https://github.com/llvm/llvm-project/pull/161166