[Mlir-commits] [mlir] [mlir][linalg] Fix isaConvolutionOpInterface logic (PR #102087)

[llvmlistbot at llvm.org]

yifeizh2 wrote:

> Thanks for working on this! For those of us a bit less familiar with the logic being updated here ...
> 
> > Fix isaConvolutionOpInterface logic
> 
> ... what's broken and how does this PR fix it? :)

Currently, the `isaConvolutionOpInterface` utility function returns false positive for all linalg binary element ops (e.g. `linalg.add`, `linalg.mul`, ...). The purpose of this PR is to eliminate such false positives.

The `isaConvolutionOpInterface` analyzes the dimensions of input/filter/output and categorizes each dimension into one of the following: batch dimension/output image dimension/output channel dimension/filter loop dimension/input channel dimension/depth multiplier. The original logic only required that each dimension must belong to one of the five categories. The current PR further requires there must be convolving dims, so that "filter loop dimension" and "output image dimension" shall be non-empty.

https://github.com/llvm/llvm-project/pull/102087