From fffe63045cb48da3fd3ff0e5e8b7ab4180db8f64 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Mon, 29 Jun 2015 14:10:32 -0700
Subject: [PATCH] Added a test for full reductions on GPU

---
 unsupported/test/CMakeLists.txt               |  1 +
 .../test/cxx11_tensor_reduction_cuda.cpp      | 55 +++++++++++++++++++
 2 files changed, 56 insertions(+)
 create mode 100644 unsupported/test/cxx11_tensor_reduction_cuda.cpp

diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index f438d4107..64e26683e 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -141,5 +141,6 @@ if(EIGEN_TEST_CXX11)
 #  ei_add_test(cxx11_tensor_device "-std=c++0x")
 #  ei_add_test(cxx11_tensor_cuda "-std=c++0x")
 #  ei_add_test(cxx11_tensor_contract_cuda "-std=c++0x")
+#  ei_add_test(cxx11_tensor_reduction_cuda "-std=c++0x")
 
 endif()
diff --git a/unsupported/test/cxx11_tensor_reduction_cuda.cpp b/unsupported/test/cxx11_tensor_reduction_cuda.cpp
new file mode 100644
index 000000000..a7eb7ac75
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_reduction_cuda.cpp
@@ -0,0 +1,55 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 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_reduction_cuda
+#define EIGEN_USE_GPU
+
+#include "main.h"
+#include <unsupported/Eigen/CXX11/Tensor>
+
+
+template<int DataLayout>
+static void test_full_reductions() {
+
+  Eigen::GpuDevice gpu_device;
+
+  const int num_rows = internal::random<int>(1024, 5*1024);
+  const int num_cols = internal::random<int>(1024, 5*1024);
+
+  Tensor<float, 2, DataLayout> in(num_rows, num_cols);
+  in.setRandom();
+
+  Tensor<float, 1, DataLayout> full_redux(1);
+  full_redux = in.sum();
+
+  std::size_t in_bytes = in.size() * sizeof(float);
+  std::size_t out_bytes = full_redux.size() * sizeof(float);
+  float* gpu_in_ptr = static_cast<float*>(gpu_device.allocate(in_bytes));
+  float* gpu_out_ptr = static_cast<float*>(gpu_device.allocate(out_bytes));
+  gpu_device.memcpyHostToDevice(gpu_in_ptr, in.data(), in_bytes);
+
+  TensorMap<Tensor<float, 2, DataLayout> > in_gpu(gpu_in_ptr, num_rows, num_cols);
+  TensorMap<Tensor<float, 1, DataLayout> > out_gpu(gpu_out_ptr, 1);
+
+  out_gpu.device(gpu_device) = in_gpu.sum();
+
+  Tensor<float, 1, DataLayout> full_redux_gpu(1);
+  gpu_device.memcpyDeviceToHost(full_redux_gpu.data(), gpu_out_ptr, out_bytes);
+  gpu_device.synchronize();
+
+  // Check that the CPU and GPU reductions return the same result.
+  VERIFY_IS_APPROX(full_redux(0), full_redux_gpu(0));
+}
+
+void test_cxx11_tensor_reduction_cuda() {
+  CALL_SUBTEST(test_full_reductions<ColMajor>());
+  CALL_SUBTEST(test_full_reductions<RowMajor>());
+}