Jerry Wu

Reland Pixel8 benchmark migration

1

We have cooling plate for Pixel 8 and expect to get more stable performance now. Reland #16087

In the new OneShotBufferization, the `linalg.buffer_layout` and `linalg.inplaceable` are replaced by `bufferization.buffer_layout` and `bufferization.writable`. This fixes the bufferization after we bumped the LLVM in #459.

When testing `python.examples.conv.conv_2d_bench`, I got an error with DoubleTiling methods: `error: replacement operation is already associated with another key`, while the runs with SingleTiling methods succeeded. From the stack trace,...

[spirv] vector.shape_cast is not handled in ConvertToSPIRV

7

`OptimizeVectorTransferPass` used in the SPIR-V pipeline might generate `vector.shape_cast` during optimizations. For example: https://github.com/openxla/iree/blob/6c016cac1c94ddf72314ae85053142f4b9babf87/compiler/src/iree/compiler/Codegen/Common/OptimizeVectorTransferPass.cpp#L102-L104 But [ConvertToSPIRVPass](https://github.com/openxla/iree/blob/6c016cac1c94ddf72314ae85053142f4b9babf87/compiler/src/iree/compiler/Codegen/SPIRV/ConvertToSPIRVPass.cpp#L362) calls [populateVectorToSPIRVPatterns](https://github.com/llvm/llvm-project/blob/deb53102a7ec90790556c732b2f954816548308d/mlir/lib/Conversion/VectorToSPIRV/VectorToSPIRV.cpp#L552) during lowering, which doesn't handle `vector.shape_cast` We run into this issue when...