[Mlir-commits] [mlir] [MLIR][NVVM] Add support for Convert Ops with rs-rounding mode (PR #165736)

[Stefan Mada]

================
@@ -365,6 +365,83 @@ LogicalResult ConvertF4x2ToF16x2Op::verify() {
   return success();
 }
 
+//===----------------------------------------------------------------------===//
+// Stochastic Rounding Conversion Ops
+//===----------------------------------------------------------------------===//
+
+LogicalResult ConvertF32x2ToF16x2Op::verify() {
+  if (getRnd() != FPRoundingMode::RS)
+    return emitOpError("Only RS rounding mode is supported for "
+                       "conversions from f32x2 to f16x2.");
+  return success();
+}
+
+LogicalResult ConvertF32x2ToBF16x2Op::verify() {
+  if (getRnd() != FPRoundingMode::RS)
+    return emitOpError("Only RS rounding mode is supported for "
+                       "conversions from f32x2 to bf16x2.");
+  return success();
+}
+
+LogicalResult ConvertF32x4ToF8x4Op::verify() {
+  mlir::MLIRContext *ctx = getContext();
+
+  if (getRnd() != FPRoundingMode::RS)
+    return emitOpError("Only RS rounding mode is supported for "
+                       "conversions from f32x4 to f8x4.");
+
+  if (getSat() == SaturationMode::NONE)
----------------
smada3 wrote:

Removed this check entirely - fp32x4->fp8x4 (and all of the other variants) only exist as .rs and .satfinite variants. With removing the attributes, no need to check it here.

https://github.com/llvm/llvm-project/pull/165736