Max191

There were a few ONNX test suite failures in https://github.com/iree-org/iree/pull/17770, but it was decided to accept the failures for now, and fix them retroactively. This issue is for tracking these...

[Codegen] Do not consider parallel regions in bufferization analysis

2

When there is a buffer used inside of an `scf.forall` op that is defined outside of the `scf.forall`, bufferization will unconditionally bufferize out of place by default in order to...

[Codegen] Add interface tensor reshape foldings to TileAndDistribute

1

This PR adds reshape into interface tensor folding patterns to TileAndDistributeToWorkgroups. If there are reshapes between interface tensors and their users, then TileAndDistributeToWorkgroups can fail, so these patterns help to...

[Codegen] Add vector transfer + slice foldings in GenericVectorization

3

Vectorizing a `linalg.copy` op can result in a sequence of ``` %extract = tensor.extract_slice %source %read = vector.transfer_read %extract %write = vector.transfer_read %dest %insert = tensor.insert_slice %write into %dest ```...

[LLVMGPU] Don't create copies when subviews are equivalent

4

The LLVMGPU bufferization copy function inserts barriers for shared memory copies. Some of the copies copy to and from equivalent subviews, which gets folded away by canonicalizations. When this happens...

[Im2col] Add option to unroll decomposed im2col loops

1

This adds an option to the DecomposeIm2colPass to unroll the resulting loop nest of the decomposition, and sets it to true by default. This is an easier form to handle...

This PR is the follow up to https://github.com/iree-org/iree/pull/18063, implementing the fusion pass to move set_encoding ops into producer dispatch regions when when the producer op is a LinalgOp. If there...

[Codegen][GPU] Add pass to reuse shared memory buffers in simple cases

1

This PR adds a new pass that tries to reuse shared memory allocations in functions. This pass only does a very basic analysis, assuming no control flow operations (and is...

[Flow][SDXL] Numerics different with vs. without aggressive fusion on SDXL

7

Running SDXL int8 with aggressive fusion enabled produces different results from running without aggressive fusion enabled. ### Repro Instructions ### 1. Checkout https://github.com/iree-org/iree/tree/shared/sdxl_quantized in IREE 2. Clone https://github.com/nod-ai/sdxl-scripts and `cd...

[GPU] Support multiple contraction dims in MmaSchedules

3

This adds support for multiple M, N, and K dims in problems when deducing a GPUMMASchedule. The new heuristic is similar to the old one, but works on pairs of...