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

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.

Reviewed By: eric-k256, jpienaar

Differential Revision: https://reviews.llvm.org/D158738

Author: krzysz00

mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp

@@ -481,12 +481,14 @@ createLinalgBodyCalculationForElementwiseOp(Operation *op, ValueRange args,
 
     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()));
 

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

@@ -551,6 +551,14 @@ func.func @test_simple_f16(%arg0: tensor<1xf16>) -> () {
   // 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
 }