From 95d90ab2e0994127ffc42b80e16a3f532895cf6d Mon Sep 17 00:00:00 2001
From: Luke Iwanski <luke@codeplay.com>
Date: Thu, 1 Jun 2017 03:30:39 +0100
Subject: [PATCH] [OpenCL] Fixes Split op (#10322)

* [OpenCL] Fixes Split op

  Split should alway go through SYCL device

* [OpenCL] Removes half from registred types
---
 tensorflow/core/kernels/split_lib_cpu.cc | 6 +-----
 tensorflow/core/kernels/split_op.cc      | 6 ++----
 2 files changed, 3 insertions(+), 9 deletions(-)

diff --git a/tensorflow/core/kernels/split_lib_cpu.cc b/tensorflow/core/kernels/split_lib_cpu.cc
index e377e4d97a4..6583f96a917 100644
--- a/tensorflow/core/kernels/split_lib_cpu.cc
+++ b/tensorflow/core/kernels/split_lib_cpu.cc
@@ -50,16 +50,12 @@ void Split<Eigen::SyclDevice, T>::operator()(
     typename TTypes<T, 3>::ConstTensor input,
     const Eigen::DSizes<Eigen::DenseIndex, 3>& slice_indices,
     const Eigen::DSizes<Eigen::DenseIndex, 3>& slice_sizes) {
-  if (output.size() < 131072) {
-    output = input.slice(slice_indices, slice_sizes);
-  } else {
     output.device(d) = input.slice(slice_indices, slice_sizes);
-  }
 }
 
 #define DEFINE_SYCL_KERNELS(T) template struct Split<Eigen::SyclDevice, T>;
 
-TF_CALL_GPU_NUMBER_TYPES(DEFINE_SYCL_KERNELS)
+TF_CALL_GPU_NUMBER_TYPES_NO_HALF(DEFINE_SYCL_KERNELS);
 #endif // TENSORFLOW_USE_SYCL
 
 }  // namespace functor
diff --git a/tensorflow/core/kernels/split_op.cc b/tensorflow/core/kernels/split_op.cc
index cf22a22fa3a..1bccd5b53fe 100644
--- a/tensorflow/core/kernels/split_op.cc
+++ b/tensorflow/core/kernels/split_op.cc
@@ -247,7 +247,6 @@ class SplitOpGPU : public SplitOpBase<GPUDevice, T> {
 #endif  // GOOGLE_CUDA
 
 #ifdef TENSORFLOW_USE_SYCL
-
 template <typename T>
 class SplitOpSYCL : public SplitOpBase<SYCLDevice, T> {
  public:
@@ -312,8 +311,7 @@ class SplitOpSYCL : public SplitOpBase<SYCLDevice, T> {
     }
   }
 };
-
-#endif  // TENSORFLOW_USE_SYCL
+#endif // TENSORFLOW_USE_SYCL
 
 #define REGISTER_SPLIT(type)                             \
   REGISTER_KERNEL_BUILDER(Name("Split")                  \
@@ -351,7 +349,7 @@ TF_CALL_complex128(REGISTER_GPU);
                               .HostMemory("split_dim"),   \
                           SplitOpSYCL<type>)
 
-TF_CALL_GPU_NUMBER_TYPES(REGISTER_SYCL);
+TF_CALL_GPU_NUMBER_TYPES_NO_HALF(REGISTER_SYCL);
 #undef REGISTER_SYCL
 
 #endif  // TENSORFLOW_USE_SYCL