Adding support for using Eigen in HIP kernels.

This commit enables the use of Eigen on HIP kernels / AMD GPUs. Support has been added along the same lines as what already exists for using Eigen in CUDA kernels / NVidia GPUs.

Application code needs to explicitly define EIGEN_USE_HIP when using Eigen in HIP kernels. This is because some of the CUDA headers get picked up by default during Eigen compile (irrespective of whether or not the underlying compiler is CUDACC/NVCC, for e.g. Eigen/src/Core/arch/CUDA/Half.h). In order to maintain this behavior, the EIGEN_USE_HIP macro is used to switch to using the HIP version of those header files (see Eigen/Core and unsupported/Eigen/CXX11/Tensor)


Use the "-DEIGEN_TEST_HIP" cmake option to enable the HIP specific unit tests.

1 files changed, 76 insertions, 0 deletions

diff --git a/unsupported/test/cxx11_tensor_scan_hip.cu b/unsupported/test/cxx11_tensor_scan_hip.cu
new file mode 100644
index 000000000..f4c4f59b9
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_scan_hip.cu
@@ -0,0 +1,76 @@
+// 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_scan_hip
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
+#define EIGEN_USE_GPU
+
+#include "main.h"
+#include <unsupported/Eigen/CXX11/Tensor>
+
+using Eigen::Tensor;
+typedef Tensor<float, 1>::DimensionPair DimPair;
+
+template<int DataLayout>
+void test_hip_cumsum(int m_size, int k_size, int n_size)
+{
+  std::cout << "Testing for (" << m_size << "," << k_size << "," << n_size << ")" << std::endl;
+  Tensor<float, 3, DataLayout> t_input(m_size, k_size, n_size);
+  Tensor<float, 3, DataLayout> t_result(m_size, k_size, n_size);
+  Tensor<float, 3, DataLayout> t_result_gpu(m_size, k_size, n_size);
+
+  t_input.setRandom();
+
+  std::size_t t_input_bytes = t_input.size()  * sizeof(float);
+  std::size_t t_result_bytes = t_result.size() * sizeof(float);
+
+  float* d_t_input;
+  float* d_t_result;
+
+  hipMalloc((void**)(&d_t_input), t_input_bytes);
+  hipMalloc((void**)(&d_t_result), t_result_bytes);
+
+  hipMemcpy(d_t_input, t_input.data(), t_input_bytes, hipMemcpyHostToDevice);
+
+  Eigen::HipStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<float, 3, DataLayout> >
+      gpu_t_input(d_t_input, Eigen::array<int, 3>(m_size, k_size, n_size));
+  Eigen::TensorMap<Eigen::Tensor<float, 3, DataLayout> >
+      gpu_t_result(d_t_result, Eigen::array<int, 3>(m_size, k_size, n_size));
+
+  gpu_t_result.device(gpu_device) = gpu_t_input.cumsum(1);
+  t_result = t_input.cumsum(1);
+
+  hipMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, hipMemcpyDeviceToHost);
+  for (DenseIndex i = 0; i < t_result.size(); i++) {
+    if (fabs(t_result(i) - t_result_gpu(i)) < 1e-4f) {
+      continue;
+    }
+    if (Eigen::internal::isApprox(t_result(i), t_result_gpu(i), 1e-4f)) {
+      continue;
+    }
+    std::cout << "mismatch detected at index " << i << ": " << t_result(i)
+              << " vs " <<  t_result_gpu(i) << std::endl;
+    assert(false);
+  }
+
+  hipFree((void*)d_t_input);
+  hipFree((void*)d_t_result);
+}
+
+
+void test_cxx11_tensor_scan_hip()
+{
+  CALL_SUBTEST(test_hip_cumsum<ColMajor>(128, 128, 128));
+  CALL_SUBTEST(test_hip_cumsum<RowMajor>(128, 128, 128));
+}