From 35bae513a0094f986c810c3f839e5a954caabd4b Mon Sep 17 00:00:00 2001 From: Mehdi Goli Date: Fri, 16 Dec 2016 19:46:45 +0000 Subject: Converting all parallel for lambda to functor in order to prevent kernel duplication name error; adding tensorConcatinationOp backend for sycl. --- .../test/cxx11_tensor_concatenation_sycl.cpp | 180 +++++++++++++++++++++ 1 file changed, 180 insertions(+) create mode 100644 unsupported/test/cxx11_tensor_concatenation_sycl.cpp (limited to 'unsupported/test/cxx11_tensor_concatenation_sycl.cpp') diff --git a/unsupported/test/cxx11_tensor_concatenation_sycl.cpp b/unsupported/test/cxx11_tensor_concatenation_sycl.cpp new file mode 100644 index 000000000..5a324b44c --- /dev/null +++ b/unsupported/test/cxx11_tensor_concatenation_sycl.cpp @@ -0,0 +1,180 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2016 +// Mehdi Goli Codeplay Software Ltd. +// Ralph Potter Codeplay Software Ltd. +// Luke Iwanski Codeplay Software Ltd. +// Contact: +// +// 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_concatenation_sycl +#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int +#define EIGEN_USE_SYCL + +#include "main.h" +#include + +using Eigen::Tensor; + +template +static void test_simple_concatenation(const Eigen::SyclDevice& sycl_device) +{ + Index leftDim1 = 2; + Index leftDim2 = 3; + Index leftDim3 = 1; + Eigen::array leftRange = {{leftDim1, leftDim2, leftDim3}}; + Index rightDim1 = 2; + Index rightDim2 = 3; + Index rightDim3 = 1; + Eigen::array rightRange = {{rightDim1, rightDim2, rightDim3}}; + + //Index concatDim1 = 3; +// Index concatDim2 = 3; +// Index concatDim3 = 1; + //Eigen::array concatRange = {{concatDim1, concatDim2, concatDim3}}; + + Tensor left(leftRange); + Tensor right(rightRange); + left.setRandom(); + right.setRandom(); + + DataType * gpu_in1_data = static_cast(sycl_device.allocate(left.dimensions().TotalSize()*sizeof(DataType))); + DataType * gpu_in2_data = static_cast(sycl_device.allocate(right.dimensions().TotalSize()*sizeof(DataType))); + + Eigen::TensorMap> gpu_in1(gpu_in1_data, leftRange); + Eigen::TensorMap> gpu_in2(gpu_in2_data, rightRange); + sycl_device.memcpyHostToDevice(gpu_in1_data, left.data(),(left.dimensions().TotalSize())*sizeof(DataType)); + sycl_device.memcpyHostToDevice(gpu_in2_data, right.data(),(right.dimensions().TotalSize())*sizeof(DataType)); + /// + Tensor concatenation1(leftDim1+rightDim1, leftDim2, leftDim3); + DataType * gpu_out_data1 = static_cast(sycl_device.allocate(concatenation1.dimensions().TotalSize()*sizeof(DataType))); + Eigen::TensorMap> gpu_out1(gpu_out_data1, concatenation1.dimensions()); + + //concatenation = left.concatenate(right, 0); + gpu_out1.device(sycl_device) =gpu_in1.concatenate(gpu_in2, 0); + sycl_device.memcpyDeviceToHost(concatenation1.data(), gpu_out_data1,(concatenation1.dimensions().TotalSize())*sizeof(DataType)); + + VERIFY_IS_EQUAL(concatenation1.dimension(0), 4); + VERIFY_IS_EQUAL(concatenation1.dimension(1), 3); + VERIFY_IS_EQUAL(concatenation1.dimension(2), 1); + for (int j = 0; j < 3; ++j) { + for (int i = 0; i < 2; ++i) { + VERIFY_IS_EQUAL(concatenation1(i, j, 0), left(i, j, 0)); + } + for (int i = 2; i < 4; ++i) { + VERIFY_IS_EQUAL(concatenation1(i, j, 0), right(i - 2, j, 0)); + } + } + + sycl_device.deallocate(gpu_out_data1); + Tensor concatenation2(leftDim1, leftDim2 +rightDim2, leftDim3); + DataType * gpu_out_data2 = static_cast(sycl_device.allocate(concatenation2.dimensions().TotalSize()*sizeof(DataType))); + Eigen::TensorMap> gpu_out2(gpu_out_data2, concatenation2.dimensions()); + gpu_out2.device(sycl_device) =gpu_in1.concatenate(gpu_in2, 1); + sycl_device.memcpyDeviceToHost(concatenation2.data(), gpu_out_data2,(concatenation2.dimensions().TotalSize())*sizeof(DataType)); + + //concatenation = left.concatenate(right, 1); + VERIFY_IS_EQUAL(concatenation2.dimension(0), 2); + VERIFY_IS_EQUAL(concatenation2.dimension(1), 6); + VERIFY_IS_EQUAL(concatenation2.dimension(2), 1); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + VERIFY_IS_EQUAL(concatenation2(i, j, 0), left(i, j, 0)); + } + for (int j = 3; j < 6; ++j) { + VERIFY_IS_EQUAL(concatenation2(i, j, 0), right(i, j - 3, 0)); + } + } + sycl_device.deallocate(gpu_out_data2); + Tensor concatenation3(leftDim1, leftDim2, leftDim3+rightDim3); + DataType * gpu_out_data3 = static_cast(sycl_device.allocate(concatenation3.dimensions().TotalSize()*sizeof(DataType))); + Eigen::TensorMap> gpu_out3(gpu_out_data3, concatenation3.dimensions()); + gpu_out3.device(sycl_device) =gpu_in1.concatenate(gpu_in2, 2); + sycl_device.memcpyDeviceToHost(concatenation3.data(), gpu_out_data3,(concatenation3.dimensions().TotalSize())*sizeof(DataType)); + + //concatenation = left.concatenate(right, 2); + VERIFY_IS_EQUAL(concatenation3.dimension(0), 2); + VERIFY_IS_EQUAL(concatenation3.dimension(1), 3); + VERIFY_IS_EQUAL(concatenation3.dimension(2), 2); + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + VERIFY_IS_EQUAL(concatenation3(i, j, 0), left(i, j, 0)); + VERIFY_IS_EQUAL(concatenation3(i, j, 1), right(i, j, 0)); + } + } + sycl_device.deallocate(gpu_out_data3); + sycl_device.deallocate(gpu_in1_data); + sycl_device.deallocate(gpu_in2_data); +} +template +static void test_concatenation_as_lvalue(const Eigen::SyclDevice& sycl_device) +{ + + Index leftDim1 = 2; + Index leftDim2 = 3; + Eigen::array leftRange = {{leftDim1, leftDim2}}; + + Index rightDim1 = 2; + Index rightDim2 = 3; + Eigen::array rightRange = {{rightDim1, rightDim2}}; + + Index concatDim1 = 4; + Index concatDim2 = 3; + Eigen::array resRange = {{concatDim1, concatDim2}}; + + Tensor left(leftRange); + Tensor right(rightRange); + Tensor result(resRange); + + left.setRandom(); + right.setRandom(); + result.setRandom(); + + DataType * gpu_in1_data = static_cast(sycl_device.allocate(left.dimensions().TotalSize()*sizeof(DataType))); + DataType * gpu_in2_data = static_cast(sycl_device.allocate(right.dimensions().TotalSize()*sizeof(DataType))); + DataType * gpu_out_data = static_cast(sycl_device.allocate(result.dimensions().TotalSize()*sizeof(DataType))); + + + Eigen::TensorMap> gpu_in1(gpu_in1_data, leftRange); + Eigen::TensorMap> gpu_in2(gpu_in2_data, rightRange); + Eigen::TensorMap> gpu_out(gpu_out_data, resRange); + + sycl_device.memcpyHostToDevice(gpu_in1_data, left.data(),(left.dimensions().TotalSize())*sizeof(DataType)); + sycl_device.memcpyHostToDevice(gpu_in2_data, right.data(),(right.dimensions().TotalSize())*sizeof(DataType)); + sycl_device.memcpyHostToDevice(gpu_out_data, result.data(),(result.dimensions().TotalSize())*sizeof(DataType)); + +// t1.concatenate(t2, 0) = result; + gpu_in1.concatenate(gpu_in2, 0).device(sycl_device) =gpu_out; + sycl_device.memcpyDeviceToHost(left.data(), gpu_in1_data,(left.dimensions().TotalSize())*sizeof(DataType)); + sycl_device.memcpyDeviceToHost(right.data(), gpu_in2_data,(right.dimensions().TotalSize())*sizeof(DataType)); + + for (int i = 0; i < 2; ++i) { + for (int j = 0; j < 3; ++j) { + VERIFY_IS_EQUAL(left(i, j), result(i, j)); + VERIFY_IS_EQUAL(right(i, j), result(i+2, j)); + } + } + sycl_device.deallocate(gpu_in1_data); + sycl_device.deallocate(gpu_in2_data); + sycl_device.deallocate(gpu_out_data); +} + + +template void tensorConcat_perDevice(Dev_selector s){ + QueueInterface queueInterface(s); + auto sycl_device = Eigen::SyclDevice(&queueInterface); + test_simple_concatenation(sycl_device); + test_simple_concatenation(sycl_device); + test_concatenation_as_lvalue(sycl_device); +} +void test_cxx11_tensor_concatenation_sycl() { + for (const auto& device :Eigen::get_sycl_supported_devices()) { + CALL_SUBTEST(tensorConcat_perDevice(device)); + } +} -- cgit v1.2.3