implement overloaded cast codegen for type conversion

Author: nicunxiao

src/op/copy.cc

@@ -325,10 +325,6 @@ For CopyNode::MakeSIMTLoop(arith::Analyzer *analyzer) const {
 
   PrimExpr value = BufferLoad(src, src_indices);
   if (src->dtype != dst->dtype) {
-    // If dst is fp8 and src is bf16, first cast dst to fp32.
-    if (src->dtype.is_bfloat16() && dst->dtype.is_float8_e4m3()) {
-      value = Cast(DataType::Float(32), value);
-    }
     value = Cast(dst->dtype, value);
   }
   if (src_predicate.defined())

src/target/codegen_cuda.cc

@@ -953,12 +953,15 @@ void CodeGenTileLangCUDA::VisitExpr_(const CastNode *op, std::ostream &os) {
     }
   }
 
+  const char *convert_part =
+      (from_ty.is_bfloat16() || target_ty.is_float8_e4m3()) ? ")(half)(" : ")(";
+
   // Fallback: elementwise cast
   for (int i = 0, lanes = from_ty.lanes(); i < lanes; ++i) {
     std::ostringstream val;
     val << "(";
     PrintType(target_ty.element_of(), val);
-    val << ")(";
+    val << convert_part;
     PrintVecElemLoad(src, from_ty, i, val);
     val << ")";
     PrintVecElemStore(sret, target_ty, i, val.str());