[Mlir-commits] [mlir] [mlir][gpu] Add pass for emulating unsupported types. (PR #138087)

[Md Abdullah Shahneous Bari]

mshahneo wrote:

Hi @kuhar, @krzysz00 , @MaheshRavishankar ,

As per discussion, BF16 support and 8 bit float support are now part of SPIR-V dialect/conversion.
However, we still need this pass for the XeVM->LLVM->LLVM SPIR-V backend path. The issue with the LLVM SPIR-V backend is that it does not support bf16 while it is supported in LLVM. The pass method is the only way for us to emulate bf16 as i16 in MLIR.

Could we please have a decision on this pass? Thank you so much for all the discussions and suggestions.

https://github.com/llvm/llvm-project/pull/138087