Max191

Integrate llvm/llvm-project@c54064de80e93494d1d44550b56ce8f2f3cf9c4b

1

Bump LLVM to include https://github.com/llvm/llvm-project/commit/205dce6029bed302f354c0bde5d8c5804f214051 and https://github.com/llvm/llvm-project/commit/3f18f6a2cfecb080f006477c46d3626102841a17

[GlobalOpt][Flow] Add GlobalOp folding to FoldUnitExtentDims

4

This adds a new pass to fold all unit dims on mutable globals. Reshape ops are inserted at the loads and stores, which should ideally fold with unit dim reshapes...

[GlobalOpt] Add pass with basic global data layout propagation analysis

14

This adds a new pass to do propagation of data layouts all the way to mutable GlobalOps. This is just the basic analysis flow for now, only capable of handling...

Canonicalizer dropping dynamic dims in tensor.pack op

3

Running the canonicalizer on this pack drops a dynamic dim on the result shape: ``` module { func.func @main(%arg0: tensor

GRAPH: Possible missing fusions of element-wise producer with reduction consumer

3

The two dispatches in the following gist come from SDXL ([mlir file](https://storage.cloud.google.com/shark-public/ean/sdxl-turbine/SDXL1_0/BS1_len64/stable_diffusion_xl_base_1_0_64_unet_attn.mlir)) https://gist.github.com/Max191/49ef6fda457959cf4888897f4b0df8e7 We see a `dequant-like op -> element-wise op -> reduction -> element-wise op`. The indexing maps line...

Aten_LocalScalarDense conversion to linalg

6

This is added as a ConvertAtenTensorToScalarLikeOp conversion in torch to linalg, lowering into a tensor.extract.

The AtenTriuOp is decomposed into ops with shapeless and typeless vtensors, which causes some conversion failures in TorchToLinalg. This PR explicitly sets the shape and dtype on the resulting ops...

This op is just a no-op, so it is simply erased in the conversion.

[LLVMGPU] Add Winograd pipeline for LLVMGPU

3

This PR adds a vectorization pass pipeline and configurations for linalg_ext.winograd ops on LLVMGPU. The pipeline follows the base vectorization pipeline, but uses scf.for thread tiling and distribution, and includes...

[Winograd] Use output_tile_size for more static output transform tiling

2

In the winograd output transform, the output tiles will always write in bounds, and will always have the same static shape. This PR uses this in the tiledImplementation of the...