[Mlir-commits] [mlir] [MLIR][XeGPU] Allow some nd ops to have argument shapes mismatch for … (PR #120566)

[Chao Chen]

chencha3 wrote:

It looks to me the current implementation is following option 2, am I right? 
I agree with adam that the sgMap should be gone in a clean design. And I also agree with jianhui that it is not convenient to distribute TensorDesc at this stage due to its nature. I like the idea of option 4, it keeps the original TensorDesc info and also gets rid of the sgMap after lowering. But at the same time, I feel it also has a draw back. Can users create arbitrary, or a set, subviews? Based on what I know so far, I feel the answer is no. It seems there is only one view available, which is implicitly defined the intrinsic. Which means the sgMap is also fixed, and there is only one valid mapping. Can we drop the sgMap in future and derive it from kernel attributes? We may use the presence of `VectorComputeINTEL` attribute to indicate the kernel is for SIMD, which will trigger the SIMD logic, otherwise it is a SIMT kernel and trigger the SIMT logic. Is it safe? 

https://github.com/llvm/llvm-project/pull/120566