[Mlir-commits] [mlir] [mlir][linalg] Implement TilingInterface for winograd operators (PR #96184)

[llvmlistbot at llvm.org]

Max191 wrote:

> @Max191 what are the benefits with the NCHW layout and winograd if I may?

The inner loop of the winograd input transformation reads a tile from the input image dimensions. This means that for an NHWC convolution it is reading a tile of size `[1, hTile, wTile, 1]`, and for an NCHW convolution it reads `[1, 1, hTile, wTile]`. The second layout is better because the inner shape of the load is contiguous in memory and can become a vectorized load. This is also the case for the write accesses of the winograd output transformation.

https://github.com/llvm/llvm-project/pull/96184