[mlir][Tosa] Allow non-fp32 tosa.cast to integers

[krzysz00]

Summary

Fix the lowering of tosa.cast to create attributes of the input source
type when casting from floats to integers.

This is motivated by the need to cast fp16 to i9, which we have
encountered in certain quantized models.

[eric-k256]

Looks good to me.

mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp

if (isa<FloatType>(srcTy) && 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 intMin = rewriter.create<arith::ConstantOp>(
-          loc, rewriter.getF32FloatAttr(
+          loc, rewriter.getFloatAttr(
+                   getElementTypeOrSelf(srcTy),
                    APInt::getSignedMinValue(dstTy.getIntOrFloatBitWidth())
                        .getSExtValue()));
 
       auto intMax = rewriter.create<arith::ConstantOp>(
-          loc, rewriter.getF32FloatAttr(
+          loc, rewriter.getFloatAttr(
+                   getElementTypeOrSelf(srcTy),
                    APInt::getSignedMaxValue(dstTy.getIntOrFloatBitWidth())
                        .getSExtValue()));
 
       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);

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

 
 // CHECK-LABEL: @test_simple_f16
 func.func @test_simple_f16(%arg0: tensor<1xf16>) -> () {
 
   // CHECK: linalg.generic
   // CHECK: arith.extf
   %0 = tosa.cast %arg0 : (tensor<1xf16>) -> tensor<1xf32>
 
+  // CHECK: linalg.generic
+  // CHECK: arith.constant -1.280000e+02
+  // CHECK: arith.constant 1.270000e+02
+  // CHECK: math.roundeven
+  // CHECK: arith.minf
+  // CHECK: arith.maxf
+  // CHECK: arith.fptosi
+  %1 = "tosa.cast"(%arg0) : (tensor<1xf16>) -> tensor<1xi8>
   return
 }
 
 // -----
 
 // CHECK-LABEL: @test_simple_i16
 func.func @test_simple_i16(%arg0: tensor<1xi16>) -> () {
   // CHECK: linalg.generic