Transformations of interest

[aqjune]

Here is the list of transformations that we'd like to verify primarily:

• linalg.conv->matmul
• linalg-fusion-for-tensor-ops(https://github.com/llvm/llvm-project/blob/d6467e00dfccf10aa91e5a1d55d981a9b13642a3/mlir/include/mlir/Dialect/Linalg/Passes.td#L59)
• tileLinalgOp code
• linalg.matmul to a loop of mul+add
• Folding linalg.subview+load
• Folding tensor_load + extract to load
• linalg to llvm plan conv loop order
• tosa-to-linalg

What else? @makesource @strike-f

[aqjune]

LinAlg is probably the right place to begin with. The very first step could be formalizing LingAlg dialect esp. structured ones
linalg-fusion-for-tensor-ops could be an initial example that we can investigate
tileLinalgOp could be one of the real examples where translation validation could be quite useful

[aqjune]

@makesource will look into either linalg.subview+load or tensor_load + extract to load.
This will require encoding memref type, which is a memory buffer that is either read-only or not.

[aqjune]

We can support a few bufferize transformations as well. They are not changing function signatures.

• -linalg-bufferize (https://github.com/llvm/llvm-project/blob/bc8acf2ce8ad6e8c9b1d97b2e02d3f4ad26e1d9d/mlir/test/Dialect/Linalg/bufferize.mlir#L1)
• -tensor-bufferize (https://github.com/llvm/llvm-project/blob/bc8acf2ce8ad6e8c9b1d97b2e02d3f4ad26e1d9d/mlir/test/Dialect/Tensor/bufferize.mlir#L1)
• -tensor-constant-bufferize (https://github.com/llvm/llvm-project/blob/bc8acf2ce8ad6e8c9b1d97b2e02d3f4ad26e1d9d/mlir/test/Dialect/Standard/tensor-constant-bufferize.mlir#L1)

[aqjune]

To match loops doing reduction, we can use matchReduction (https://reviews.llvm.org/D110303)

[aqjune]

Tiling: -linalg-tile

[aqjune]

TOSA to LinAlg: -tosa-to-linalg (updated)

[aqjune]

Buffer opts:

• buffer-deallocation
• buffer-hoisting
• buffer-loop-hoisting
• promote-buffers-to-stack
• buffer-results-to-out-params

[aqjune]

-normalize-memrefs