From f36c0c2c65a78959f6ccbbc29c6e80f86b062bc8 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Fri, 19 Feb 2016 06:23:28 +0000
Subject: [PATCH] Added regression test for float16

---
 .../test/cxx11_tensor_of_float16_cuda.cu      | 60 +++++++++++++++++++
 1 file changed, 60 insertions(+)
 create mode 100644 unsupported/test/cxx11_tensor_of_float16_cuda.cu

diff --git a/unsupported/test/cxx11_tensor_of_float16_cuda.cu b/unsupported/test/cxx11_tensor_of_float16_cuda.cu
new file mode 100644
index 000000000..e9f5dd968
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_of_float16_cuda.cu
@@ -0,0 +1,60 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#define EIGEN_TEST_NO_LONGDOUBLE
+#define EIGEN_TEST_NO_COMPLEX
+#define EIGEN_TEST_FUNC cxx11_tensor_of_float16_cuda
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
+#define EIGEN_USE_GPU
+
+
+#include "main.h"
+#include <unsupported/Eigen/CXX11/Tensor>
+
+using Eigen::Tensor;
+
+void test_cuda_conversion() {
+  Eigen::CudaStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+  int num_elem = 101;
+  
+  float* d_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
+  half* d_half = (half*)gpu_device.allocate(num_elem * sizeof(half));
+  float* d_conv = (float*)gpu_device.allocate(num_elem * sizeof(float));
+
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(
+      d_float, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<half, 1>, Eigen::Aligned> gpu_half(
+      d_half, num_elem);
+  Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_conv(
+      d_conv, num_elem);
+
+  gpu_float.device(gpu_device) = gpu_float.random();
+  gpu_half.device(gpu_device) = gpu_float.cast<half>();
+  gpu_conv.device(gpu_device) = gpu_half.cast<float>();
+
+  Tensor<float, 1> initial(num_elem);
+  Tensor<float, 1> final(num_elem);
+  gpu_device.memcpyDeviceToHost(initial.data(), d_float, num_elem*sizeof(float));
+  gpu_device.memcpyDeviceToHost(final.data(), d_conv, num_elem*sizeof(float));
+
+  for (int i = 0; i < num_elem; ++i) {
+    VERIFY_IS_APPROX(initial(i), final(i));
+  }
+
+  gpu_device.deallocate(d_float);
+  gpu_device.deallocate(d_half);
+  gpu_device.deallocate(d_conv);
+}
+
+
+void test_cxx11_tensor_of_float16_cuda()
+{
+  CALL_SUBTEST_1(test_cuda_conversion());
+}