From b8392684999b738d0d4cb367381075dc3b925fef Mon Sep 17 00:00:00 2001
From: Trevor Morris <tmorris@nvidia.com>
Date: Thu, 14 Feb 2019 11:20:58 -0800
Subject: [PATCH] Change deprecated setHalf2Mode -> setFp16Mode. Also allow
 int8 mode to fp16 kernels when they are more performant.

---
 tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc
index 002526c04bb..4cc72e50b24 100644
--- a/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc
+++ b/tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc
@@ -3770,8 +3770,11 @@ tensorflow::Status ConvertGraphDefToEngine(
   builder->setMaxWorkspaceSize(max_workspace_size_bytes);
   builder->setGpuAllocator(allocator);
   if (precision_mode == TrtPrecisionMode::FP16) {
-    builder->setHalf2Mode(true);
+    builder->setFp16Mode(true);
   } else if (precision_mode == TrtPrecisionMode::INT8) {
+    // Setting FP16 mode as well allows TRT to also consider FP16 kernels and
+    // use them in situations where they are faster than INT8.
+    builder->setFp16Mode(true);
     builder->setInt8Mode(true);
     if (use_calibration) {
       builder->setInt8Calibrator(calibrator);