Using custom bitwidth MLIR integers.
Hi
If we were to create custom types that wrap around custom bitwidth integers using MLIR types and operations, does it improve performance?
Theoretically, it should not right? Because if the register is 64 bits wide, it would be a waste to use only 4 bits to store a
ui4 wasting the entire 60 bits. Unless, the compiler is smart enough to look at the function signature and then group many e.g
UInt[4] together.0 Replies