From 077f415e21782c9d17b83866b7e21e112a878649 Mon Sep 17 00:00:00 2001
From: fo40225 <fo40225@hotmail.com>
Date: Thu, 3 Oct 2019 21:38:02 +0800
Subject: [PATCH 1/2] fix msvc 16.3.0 + cuda 10.1.243 build break

---
 tensorflow/core/kernels/scatter_nd_op_gpu.cu.cc | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/tensorflow/core/kernels/scatter_nd_op_gpu.cu.cc b/tensorflow/core/kernels/scatter_nd_op_gpu.cu.cc
index 209e7bb6130..2b152f44c81 100644
--- a/tensorflow/core/kernels/scatter_nd_op_gpu.cu.cc
+++ b/tensorflow/core/kernels/scatter_nd_op_gpu.cu.cc
@@ -72,7 +72,9 @@ template <typename T>
 struct LeftUpdate<std::complex<T>, scatter_nd_op::UpdateOp::SUB> {
   EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC void operator()(
       std::complex<T>* out, const std::complex<T>& val) {
-    LeftUpdate<std::complex<T>, scatter_nd_op::UpdateOp::ADD>()(out, -val);
+    T* ptr = reinterpret_cast<T*>(out);
+    GpuAtomicAdd(ptr, -val.real());
+    GpuAtomicAdd(ptr, -val.imag());
   }
 };
 

From 6a1a6925fc332c5012ba67a0d7e8e46ac6291c5d Mon Sep 17 00:00:00 2001
From: fo40225 <fo40225@hotmail.com>
Date: Sat, 19 Oct 2019 19:31:56 +0800
Subject: [PATCH 2/2] use GpuAtomicSub instead of GpuAtomicAdd

---
 tensorflow/core/kernels/scatter_nd_op_gpu.cu.cc | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/tensorflow/core/kernels/scatter_nd_op_gpu.cu.cc b/tensorflow/core/kernels/scatter_nd_op_gpu.cu.cc
index 2b152f44c81..c7b14af5bc0 100644
--- a/tensorflow/core/kernels/scatter_nd_op_gpu.cu.cc
+++ b/tensorflow/core/kernels/scatter_nd_op_gpu.cu.cc
@@ -64,7 +64,7 @@ struct LeftUpdate<std::complex<T>, scatter_nd_op::UpdateOp::ADD> {
       std::complex<T>* out, const std::complex<T>& val) {
     T* ptr = reinterpret_cast<T*>(out);
     GpuAtomicAdd(ptr, val.real());
-    GpuAtomicAdd(ptr, val.imag());
+    GpuAtomicAdd(ptr + 1, val.imag());
   }
 };
 
@@ -73,8 +73,8 @@ struct LeftUpdate<std::complex<T>, scatter_nd_op::UpdateOp::SUB> {
   EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC void operator()(
       std::complex<T>* out, const std::complex<T>& val) {
     T* ptr = reinterpret_cast<T*>(out);
-    GpuAtomicAdd(ptr, -val.real());
-    GpuAtomicAdd(ptr, -val.imag());
+    GpuAtomicSub(ptr, val.real());
+    GpuAtomicSub(ptr + 1, val.imag());
   }
 };