[Mlir-commits] [mlir] [MLIR][NVGPU] Add convert.fpext and convert.fptrunc Ops (PR #199700)

[llvmlistbot at llvm.org]

================
@@ -699,6 +699,138 @@ LogicalResult RcpOp::verify() {
   return success();
 }
 
+//===----------------------------------------------------------------------===//
+// NVGPU_ConvertFPTruncOp
+//===----------------------------------------------------------------------===//
+
+static bool isShapedContainerType(Type t) {
+  return llvm::isa<VectorType, RankedTensorType, UnrankedTensorType>(t);
----------------
yifeihe-007 wrote:

 Originally we wanted the NVGPU conversion ops to have tensor-level expressiveness like arith.truncf/arith.extf. The current pipeline doesn't really use the tensor form yet. Down the line it could be nice to actually use nvgpu at the tensor layer, so fp conversion has one unified representation from tensor down to the lower levels. But open on this, happy to hear your thoughts on the layering here.

https://github.com/llvm/llvm-project/pull/199700