[Mlir-commits] [mlir] [BugFix] : Fix padding calculation for tosa.transpose_conv2d decomposition. (PR #68167)

[Georgios Pinitas]

GeorgeARM wrote:

I think the current calculation is correct.

This might be confusing but seems that the padding in the transposed convolution in the tosa spec has been negated compared to the traditional calculation; not sure on the reasoning behind this. Overall I think that `k - 1 + p` is correct. Note that according to the spec the `output_padding` can be negative as well. Sure @eric-k256 will know more on this.
 
FYI the latest TOSA spec can be found [here](https://www.mlplatform.org/tosa/tosa_spec.html#_transpose_conv2d) 

https://github.com/llvm/llvm-project/pull/68167