Fixed-point PoC

[KFAFSP]

We want to create a fixed-point only PoC implementation that incorporates the new design and specification elements.

• Make a branch for the PoC
• Remove floating-point and constant folding components to be able to change the interface quickly
• Remove limit on E for fixed-point (32bits?)
• Implement Elaborator for issues Add the base2.fixed.* exact operations #2 , Implement a base2.* to arith.* lowering #5 , Implement a base2.* to base2.* conversion for rounding modes #6
  • Can we already show the similarity between fixed and floating by following the frexp/ldexp scheme?
  • Is there a way to turn constant results into type parameters while elaborating (make fixed from const exp)?
  • Is there 1 best method of rounding? If not, how do we allow the user to select it?
  • How many bits can we use and how can the user chose a different path (promotion pass)?
  • How to seperate the responsibilities (add alias operations vs elaborator method)?
• Create experiments for the demo

[KFAFSP]

This would also have to incorporate the change to the limits on E for fixed-point!

[jhbi826c]

What I've done:

• Creat a new branch heart23 and remove floating point semantics and folders for ops.

• Change FixedPointType from <integerBits, fractionalBits> to <bitWidth, exponent> and modify the FixedPointSemantics interface and FixedPointInterpreter to adjust this.

• Create Pass ConvertBase2ToArith, which now converts

• base2.bit_cast to arith.bitcast

• base2.value_cast to arith.trunci/extsi/extui

• base2.fixed.add to base2.value_cast and arith.addi

• example of --convert-base2-to-arith:

!type1 = !base2.fixed_point<signed 6, -3>
!type2 = !base2.fixed_point<unsigned 4, -2>
!type3 = !base2.fixed_point<signed 7, -3>

func.func @test(%arg0: !type1, %arg1: !type2) -> !type3
{
    %0 = base2.fixed.add %arg0, %arg1 : !type1, !type2
    return %0 : !type3
}

func.func @test(%arg0: i6, %arg1: i4) -> i7 {
    %0 = arith.extsi %arg0 : i6 to i7
    %1 = arith.extui %arg1 : i4 to i7
    %2 = arith.addi %0, %1 : i7
    return %2 : i7
  }

What still needs to think of:

• fixed.op's return type
• Implement a base2.* to base2.* conversion for rounding modes #6 -> How to deal with rounding modes.
• elaborator methods
• other ops to add