[mlir][tosa] TosaToLinalg: Lower TOSA.Cast via RoundEven according to TOSA spec 0.60.0

[mgehre-amd]

Summary

TOSA now specifies rounding of ties to even in section 1.8.2., "Main Inference Profile"

[eric-k256]

Added Rob as an additional reviewer. As mentioned, this matches with the TOSA specification, and looks like a nice simplification at the same time.

mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp

if (isa<tosa::CastOp>(op)) {
     if (srcTy.isa<FloatType>() && dstTy.isInteger(1)) {
       Value zero = rewriter.create<arith::ConstantOp>(
           loc, rewriter.getFloatAttr(srcTy, 0.0));
       return rewriter.create<arith::CmpFOp>(loc, arith::CmpFPredicate::UNE,
                                             args.front(), zero);
     }
 
     if (arith::FPToSIOp::areCastCompatible(srcTy, dstTy)) {
-      auto zero = rewriter.create<arith::ConstantOp>(
-          loc, rewriter.getF32FloatAttr(0.0f));
-      auto half = rewriter.create<arith::ConstantOp>(
-          loc, rewriter.getF32FloatAttr(0.5f));
-
       auto intMin = rewriter.create<arith::ConstantOp>(
           loc, rewriter.getF32FloatAttr(
                    APInt::getSignedMinValue(dstTy.getIntOrFloatBitWidth())
                        .getSExtValue()));
 
       auto intMax = rewriter.create<arith::ConstantOp>(
           loc, rewriter.getF32FloatAttr(
                    APInt::getSignedMaxValue(dstTy.getIntOrFloatBitWidth())
                        .getSExtValue()));
 
-      auto added = rewriter.create<arith::AddFOp>(loc, args[0], half);
-      auto subbed = rewriter.create<arith::SubFOp>(loc, args[0], half);
-      auto negative = rewriter.create<arith::CmpFOp>(
-          loc, arith::CmpFPredicate::OLT, args[0], zero);
-      auto rounded =
-          rewriter.create<arith::SelectOp>(loc, negative, subbed, added);
+      auto rounded = rewriter.create<math::RoundEvenOp>(loc, args[0]);
 
       auto clamped = clampFloatHelper(loc, rounded, intMin, intMax, rewriter);
 
       return rewriter.create<arith::FPToSIOp>(loc, dstTy, clamped);
     }
 
     // Casting to boolean, integers need to only be checked as not-equal to
     // zero.

mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir

func.func @test_simple_f32(%arg0: tensor<1xf32>) -> () {
   // CHECK: linalg.generic
   // CHECK: arith.negf
   // CHECK: exp
   // CHECK: arith.addf
   // CHECK: arith.divf
   %19 = "tosa.sigmoid"(%0) : (tensor<1xf32>) -> tensor<1xf32>
 
   // CHECK: linalg.generic
-  // CHECK: arith.constant 0.000000e+00
-  // CHECK: arith.constant 5.000000e-01
   // CHECK: arith.constant -2.14748365E+9
   // CHECK: arith.constant 2.14748365E+9
-  // CHECK: arith.addf
-  // CHECK: arith.subf
-  // CHECK: arith.cmpf olt
-  // CHECK: select
+  // CHECK: math.roundeven
   // CHECK: arith.minf
   // CHECK: arith.maxf
   // CHECK: arith.fptosi
   %20 = "tosa.cast"(%0) : (tensor<1xf32>) -> tensor<1xi32>
 
   // CHECK: linalg.generic
   // CHECK: arith.constant 0
   // CHECK: arith.cmpf