[Mlir-commits] [mlir] [mlir][linalg] Introduce new `linalg.conv` op (PR #117688)

[Felix Schneider]

ubfx wrote:

Thanks, I'll copy the description of the current state from the discours thread here for reference:

Currently, we have a proposed linalg.conv op using the above mentioned principle of 4 named dimensions + up to 3 spatial dimensions. We can also convert the old (non-quantized) convolution ops to the new one.

The different “subtypes” of convolution emerge only from the dimensions of the tensors and can be queried via e.g. ConvOp::isDepthwise() and ConvOps::getSpatialDims().

“Quantized” (i.e. equivalent to the old ..._q ops is not implemented yet but could be implemented via a separate op or via an additional operand.

There’s a lot of room for improvement, e.g. expanding the number of possible spatial dimensions from 3 to N, and the ConvDimsAttr assembly format. But I’d like to see how happy people are with the general idea of this op.

https://github.com/llvm/llvm-project/pull/117688