[Mlir-commits] [mlir] [mlir][vector] add tensor.concat, bitcast, expand_shape, collapse_shape vectorization support (PR #97297)

[llvmlistbot at llvm.org]

BRUCE11111 wrote:

> I believe it is this one:
> 
> * https://discourse.llvm.org/t/rfc-vector-distribution-for-cpu-convert-vector-to-physical-register-size-vector/
> 
> We haven't really agreed on any specific path forward.

Yes, the current work is about this RFC.

**Motivation**:

The current community discussion has reached a certain degree of consensus on the RFC work, such as vector-to-loops, etc. This work is necessary. But some detailed designs are still under discussion.

For the current PR, we only need to know, before converting large virtual vector to `scf.for + tiled vector`, we first need to convert the operation to a vector operation. So I want to expand the existing functions of the community.

_____

> vectorize expand_shape, collapse_shape into vector.shape_cast.

In the current vector dialect, we only have shape_cast that can align semantically with tensor.expand_shape, so I currently use shape_cast.

>  tensor.bitcast is not equivalent to vector.bitcast.

Thanks for pointing out that, I noticed that differentce between tensor.bitcast and vector.bitcast, I will convert tensor.bitcast → vector.tranfer_read + arith.bitcast + vector.transfer_write.

____

In addition, I have a mask-related question for @hanhanW , which is written in another PR[#97248](https://github.com/llvm/llvm-project/pull/97248#discussion_r1665236084). 

Sincere Thanks~


https://github.com/llvm/llvm-project/pull/97297