2016-09-19 19:44:13 +08:00
|
|
|
// This file is part of Eigen, a lightweight C++ template library
|
|
|
|
// for linear algebra.
|
|
|
|
//
|
2016-09-19 21:09:25 +08:00
|
|
|
// Copyright (C) 2016
|
|
|
|
// Mehdi Goli Codeplay Software Ltd.
|
|
|
|
// Ralph Potter Codeplay Software Ltd.
|
|
|
|
// Luke Iwanski Codeplay Software Ltd.
|
|
|
|
// Contact: <eigen@codeplay.com>
|
2016-09-19 19:44:13 +08:00
|
|
|
//
|
|
|
|
// 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
|
2016-10-06 06:00:32 +08:00
|
|
|
#define EIGEN_TEST_FUNC cxx11_tensor_broadcast_sycl
|
2016-11-19 05:44:20 +08:00
|
|
|
#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int64_t
|
2016-09-19 19:44:13 +08:00
|
|
|
#define EIGEN_USE_SYCL
|
|
|
|
|
|
|
|
#include "main.h"
|
|
|
|
#include <unsupported/Eigen/CXX11/Tensor>
|
|
|
|
|
|
|
|
using Eigen::array;
|
|
|
|
using Eigen::SyclDevice;
|
|
|
|
using Eigen::Tensor;
|
|
|
|
using Eigen::TensorMap;
|
|
|
|
|
2016-11-19 05:44:20 +08:00
|
|
|
template <typename DataType, int DataLayout, typename IndexType>
|
2016-11-15 02:13:53 +08:00
|
|
|
static void test_broadcast_sycl_fixed(const Eigen::SyclDevice &sycl_device){
|
2016-09-19 19:44:13 +08:00
|
|
|
|
2016-11-09 01:08:02 +08:00
|
|
|
// BROADCAST test:
|
2016-11-19 05:44:20 +08:00
|
|
|
IndexType inDim1=2;
|
|
|
|
IndexType inDim2=3;
|
|
|
|
IndexType inDim3=5;
|
|
|
|
IndexType inDim4=7;
|
|
|
|
IndexType bDim1=2;
|
|
|
|
IndexType bDim2=3;
|
|
|
|
IndexType bDim3=1;
|
|
|
|
IndexType bDim4=4;
|
|
|
|
array<IndexType, 4> in_range = {{inDim1, inDim2, inDim3, inDim4}};
|
|
|
|
array<IndexType, 4> broadcasts = {{bDim1, bDim2, bDim3, bDim4}};
|
|
|
|
array<IndexType, 4> out_range; // = in_range * broadcasts
|
2016-11-09 01:08:02 +08:00
|
|
|
for (size_t i = 0; i < out_range.size(); ++i)
|
|
|
|
out_range[i] = in_range[i] * broadcasts[i];
|
|
|
|
|
2016-11-19 05:44:20 +08:00
|
|
|
Tensor<DataType, 4, DataLayout, IndexType> input(in_range);
|
|
|
|
Tensor<DataType, 4, DataLayout, IndexType> out(out_range);
|
2016-09-19 19:44:13 +08:00
|
|
|
|
2016-11-09 01:08:02 +08:00
|
|
|
for (size_t i = 0; i < in_range.size(); ++i)
|
|
|
|
VERIFY_IS_EQUAL(out.dimension(i), out_range[i]);
|
2016-09-19 19:44:13 +08:00
|
|
|
|
|
|
|
|
2016-11-19 05:44:20 +08:00
|
|
|
for (IndexType i = 0; i < input.size(); ++i)
|
2016-11-19 00:20:42 +08:00
|
|
|
input(i) = static_cast<DataType>(i);
|
2016-09-19 19:44:13 +08:00
|
|
|
|
2016-11-19 00:20:42 +08:00
|
|
|
DataType * gpu_in_data = static_cast<DataType*>(sycl_device.allocate(input.dimensions().TotalSize()*sizeof(DataType)));
|
|
|
|
DataType * gpu_out_data = static_cast<DataType*>(sycl_device.allocate(out.dimensions().TotalSize()*sizeof(DataType)));
|
2016-09-19 19:44:13 +08:00
|
|
|
|
2016-11-19 05:44:20 +08:00
|
|
|
TensorMap<TensorFixedSize<DataType, Sizes<2, 3, 5, 7>, DataLayout, IndexType>> gpu_in(gpu_in_data, in_range);
|
|
|
|
TensorMap<Tensor<DataType, 4, DataLayout, IndexType>> gpu_out(gpu_out_data, out_range);
|
2016-11-19 00:20:42 +08:00
|
|
|
sycl_device.memcpyHostToDevice(gpu_in_data, input.data(),(input.dimensions().TotalSize())*sizeof(DataType));
|
2016-11-15 02:13:53 +08:00
|
|
|
gpu_out.device(sycl_device) = gpu_in.broadcast(broadcasts);
|
2016-11-19 00:20:42 +08:00
|
|
|
sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(DataType));
|
2016-11-15 02:13:53 +08:00
|
|
|
|
2016-11-19 05:44:20 +08:00
|
|
|
for (IndexType i = 0; i < inDim1*bDim1; ++i) {
|
|
|
|
for (IndexType j = 0; j < inDim2*bDim2; ++j) {
|
|
|
|
for (IndexType k = 0; k < inDim3*bDim3; ++k) {
|
|
|
|
for (IndexType l = 0; l < inDim4*bDim4; ++l) {
|
2016-11-15 02:13:53 +08:00
|
|
|
VERIFY_IS_APPROX(input(i%2,j%3,k%5,l%7), out(i,j,k,l));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
printf("Broadcast Test with fixed size Passed\n");
|
|
|
|
sycl_device.deallocate(gpu_in_data);
|
|
|
|
sycl_device.deallocate(gpu_out_data);
|
|
|
|
}
|
|
|
|
|
2016-11-19 05:44:20 +08:00
|
|
|
template <typename DataType, int DataLayout, typename IndexType>
|
2016-11-15 02:13:53 +08:00
|
|
|
static void test_broadcast_sycl(const Eigen::SyclDevice &sycl_device){
|
|
|
|
|
|
|
|
// BROADCAST test:
|
2016-11-19 05:44:20 +08:00
|
|
|
IndexType inDim1=2;
|
|
|
|
IndexType inDim2=3;
|
|
|
|
IndexType inDim3=5;
|
|
|
|
IndexType inDim4=7;
|
|
|
|
IndexType bDim1=2;
|
|
|
|
IndexType bDim2=3;
|
|
|
|
IndexType bDim3=1;
|
|
|
|
IndexType bDim4=4;
|
|
|
|
array<IndexType, 4> in_range = {{inDim1, inDim2, inDim3, inDim4}};
|
|
|
|
array<IndexType, 4> broadcasts = {{bDim1, bDim2, bDim3, bDim4}};
|
|
|
|
array<IndexType, 4> out_range; // = in_range * broadcasts
|
2016-11-15 02:13:53 +08:00
|
|
|
for (size_t i = 0; i < out_range.size(); ++i)
|
|
|
|
out_range[i] = in_range[i] * broadcasts[i];
|
|
|
|
|
2016-11-19 05:44:20 +08:00
|
|
|
Tensor<DataType, 4, DataLayout, IndexType> input(in_range);
|
|
|
|
Tensor<DataType, 4, DataLayout, IndexType> out(out_range);
|
2016-11-15 02:13:53 +08:00
|
|
|
|
|
|
|
for (size_t i = 0; i < in_range.size(); ++i)
|
|
|
|
VERIFY_IS_EQUAL(out.dimension(i), out_range[i]);
|
|
|
|
|
|
|
|
|
2016-11-19 05:44:20 +08:00
|
|
|
for (IndexType i = 0; i < input.size(); ++i)
|
2016-11-19 00:20:42 +08:00
|
|
|
input(i) = static_cast<DataType>(i);
|
2016-11-15 02:13:53 +08:00
|
|
|
|
2016-11-19 00:20:42 +08:00
|
|
|
DataType * gpu_in_data = static_cast<DataType*>(sycl_device.allocate(input.dimensions().TotalSize()*sizeof(DataType)));
|
|
|
|
DataType * gpu_out_data = static_cast<DataType*>(sycl_device.allocate(out.dimensions().TotalSize()*sizeof(DataType)));
|
2016-11-15 02:13:53 +08:00
|
|
|
|
2016-11-19 05:44:20 +08:00
|
|
|
TensorMap<Tensor<DataType, 4, DataLayout, IndexType>> gpu_in(gpu_in_data, in_range);
|
|
|
|
TensorMap<Tensor<DataType, 4, DataLayout, IndexType>> gpu_out(gpu_out_data, out_range);
|
2016-11-19 00:20:42 +08:00
|
|
|
sycl_device.memcpyHostToDevice(gpu_in_data, input.data(),(input.dimensions().TotalSize())*sizeof(DataType));
|
2016-11-09 01:08:02 +08:00
|
|
|
gpu_out.device(sycl_device) = gpu_in.broadcast(broadcasts);
|
2016-11-19 00:20:42 +08:00
|
|
|
sycl_device.memcpyDeviceToHost(out.data(), gpu_out_data,(out.dimensions().TotalSize())*sizeof(DataType));
|
2016-11-09 01:08:02 +08:00
|
|
|
|
2016-11-19 05:44:20 +08:00
|
|
|
for (IndexType i = 0; i < inDim1*bDim1; ++i) {
|
|
|
|
for (IndexType j = 0; j < inDim2*bDim2; ++j) {
|
|
|
|
for (IndexType k = 0; k < inDim3*bDim3; ++k) {
|
|
|
|
for (IndexType l = 0; l < inDim4*bDim4; ++l) {
|
2016-11-19 00:20:42 +08:00
|
|
|
VERIFY_IS_APPROX(input(i%inDim1,j%inDim2,k%inDim3,l%inDim4), out(i,j,k,l));
|
2016-11-09 01:08:02 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
printf("Broadcast Test Passed\n");
|
|
|
|
sycl_device.deallocate(gpu_in_data);
|
|
|
|
sycl_device.deallocate(gpu_out_data);
|
2016-09-19 19:44:13 +08:00
|
|
|
}
|
|
|
|
|
2016-11-19 08:26:50 +08:00
|
|
|
template<typename DataType> void sycl_broadcast_test_per_device(const cl::sycl::device& d){
|
|
|
|
std::cout << "Running on " << d.template get_info<cl::sycl::info::device::name>() << std::endl;
|
|
|
|
QueueInterface queueInterface(d);
|
2016-11-19 00:20:42 +08:00
|
|
|
auto sycl_device = Eigen::SyclDevice(&queueInterface);
|
2016-11-19 05:44:20 +08:00
|
|
|
test_broadcast_sycl<DataType, RowMajor, int64_t>(sycl_device);
|
|
|
|
test_broadcast_sycl<DataType, ColMajor, int64_t>(sycl_device);
|
2016-11-26 00:19:07 +08:00
|
|
|
test_broadcast_sycl_fixed<DataType, RowMajor, int64_t>(sycl_device);
|
|
|
|
test_broadcast_sycl_fixed<DataType, ColMajor, int64_t>(sycl_device);
|
2016-11-19 00:20:42 +08:00
|
|
|
}
|
2016-11-15 02:13:53 +08:00
|
|
|
|
2016-10-06 06:00:32 +08:00
|
|
|
void test_cxx11_tensor_broadcast_sycl() {
|
2016-11-26 00:19:07 +08:00
|
|
|
for (const auto& device :Eigen::get_sycl_supported_devices()) {
|
2016-11-19 08:26:50 +08:00
|
|
|
CALL_SUBTEST(sycl_broadcast_test_per_device<float>(device));
|
|
|
|
}
|
2016-09-19 19:44:13 +08:00
|
|
|
}
|