[mlir][tosa] Fix tosa.cast UiToFp32 for tosa-to-linalg

[rsuderman]

Summary

Part of the arith update broke UiToFp32. Fixed the lowering and included a new
test to detect a regression.

[GMNGeoffrey]

mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp

575

I wonder if for the future it would be better to do this with the type conversion part of the dialect conversion framework. Also man this is a huge function...

[rsuderman]

mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp

575

In this case I am not certain the type conversion part would help that much. We actually need the signedness preserved after this executes and arguably other compilers may want to preserve it for their version. It would not be as bad if SIToFP and UIToFP operating signed/unsigned integers (not just signless).

But yeah, the function is a bit huge but it represents around 1/2 of all TOSA lowerings (all elementwise operations).

mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp

if (isa<tosa::CastOp>(op)) {
     if (srcTy.isInteger(1) && arith::UIToFPOp::areCastCompatible(srcTy, dstTy))
       return rewriter.create<arith::UIToFPOp>(loc, resultTypes, args,
                                               mlir::None);
 
     if (srcTy.isInteger(1) && dstTy.isa<IntegerType>() && bitExtend)
       return rewriter.create<arith::ExtUIOp>(loc, resultTypes, args,
                                              mlir::None);
 
-    // All other si-to-fp conversions should be handled by SIToFP.
-    if (arith::SIToFPOp::areCastCompatible(srcTy, dstTy))
-      return rewriter.create<arith::SIToFPOp>(loc, resultTypes, args,
-                                              mlir::None);
-
     // Unsigned integers need an unrealized cast so that they can be passed
     // to UIToFP.
     if (srcTy.isUnsignedInteger() && dstTy.isa<FloatType>()) {
       auto unrealizedCast =
           rewriter
               .create<UnrealizedConversionCastOp>(
                   loc, rewriter.getIntegerType(srcTy.getIntOrFloatBitWidth()),
                   args[0])
               .getResult(0);
       return rewriter.create<arith::UIToFPOp>(loc, resultTypes[0],
                                               unrealizedCast);
     }
 
+    // All other si-to-fp conversions should be handled by SIToFP.
+    if (arith::SIToFPOp::areCastCompatible(srcTy, dstTy))
+      return rewriter.create<arith::SIToFPOp>(loc, resultTypes, args,
+                                              mlir::None);
+
     // Casting to boolean, floats need to only be checked as not-equal to zero.
     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);
     }
 

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

func @test_simple_i16(%arg0: tensor<1xi16>) -> () {
   // CHECK: arith.muli
   %0 = "tosa.mul"(%arg0, %arg0) {shift = 0 : i32} : (tensor<1xi16>, tensor<1xi16>) -> tensor<1xi32>
 
   return
 }
 
 // -----
 
+// CHECK-LABEL: @test_simple_ui8
+func @test_simple_ui8(%arg0: tensor<1xui8>) -> () {
+  // CHECK: arith.uitofp
+  %0 = "tosa.cast"(%arg0) : (tensor<1xui8>) -> tensor<1xf32>
+  return
+}
+
+// -----
+
 // CHECK-LABEL: @test_simple_i32
 func @test_simple_i32(%arg0: tensor<1xi32>) -> () {
   // CHECK: linalg.generic
   // CHECK: arith.addi
   %0 = "tosa.add"(%arg0, %arg0) : (tensor<1xi32>, tensor<1xi32>) -> tensor<1xi32>
 
   // CHECK: linalg.generic
   // CHECK: arith.subi