cxx11_tensor_scan_sycl.cpp File Reference

#include "main.h"
#include <unsupported/Eigen/CXX11/Tensor>

Macros
#define	EIGEN_TEST_NO_LONGDOUBLE
#define	EIGEN_TEST_NO_COMPLEX
#define	EIGEN_DEFAULT_DENSE_INDEX_TYPE   int64_t
#define	EIGEN_USE_SYCL

Typedefs
typedef Tensor< float, 1 >::DimensionPair	DimPair

Functions
template<typename DataType , int DataLayout, typename IndexType >
void	test_sycl_cumsum (const Eigen::SyclDevice &sycl_device, IndexType m_size, IndexType k_size, IndexType n_size, int consume_dim, bool exclusive)
template<typename DataType , typename Dev >
void	sycl_scan_test_exclusive_dim0_per_device (const Dev &sycl_device)
template<typename DataType , typename Dev >
void	sycl_scan_test_exclusive_dim1_per_device (const Dev &sycl_device)
template<typename DataType , typename Dev >
void	sycl_scan_test_exclusive_dim2_per_device (const Dev &sycl_device)
template<typename DataType , typename Dev >
void	sycl_scan_test_inclusive_dim0_per_device (const Dev &sycl_device)
template<typename DataType , typename Dev >
void	sycl_scan_test_inclusive_dim1_per_device (const Dev &sycl_device)
template<typename DataType , typename Dev >
void	sycl_scan_test_inclusive_dim2_per_device (const Dev &sycl_device)
	EIGEN_DECLARE_TEST (cxx11_tensor_scan_sycl)

Macro Definition Documentation

EIGEN_DEFAULT_DENSE_INDEX_TYPE

#define EIGEN_DEFAULT_DENSE_INDEX_TYPE   int64_t

EIGEN_TEST_NO_COMPLEX

#define EIGEN_TEST_NO_COMPLEX

EIGEN_TEST_NO_LONGDOUBLE

#define EIGEN_TEST_NO_LONGDOUBLE

EIGEN_USE_SYCL

#define EIGEN_USE_SYCL

Typedef Documentation

DimPair

typedef Tensor<float, 1>::DimensionPair DimPair

Function Documentation

EIGEN_DECLARE_TEST()

EIGEN_DECLARE_TEST

(

cxx11_tensor_scan_sycl

)

                                           {
  for (const auto& device : Eigen::get_sycl_supported_devices()) {
    std::cout << "Running on " << device.template get_info<cl::sycl::info::device::name>() << std::endl;
    QueueInterface queueInterface(device);
    auto sycl_device = Eigen::SyclDevice(&queueInterface);
    CALL_SUBTEST_1(sycl_scan_test_exclusive_dim0_per_device<float>(sycl_device));
    CALL_SUBTEST_2(sycl_scan_test_exclusive_dim1_per_device<float>(sycl_device));
    CALL_SUBTEST_3(sycl_scan_test_exclusive_dim2_per_device<float>(sycl_device));
    CALL_SUBTEST_4(sycl_scan_test_inclusive_dim0_per_device<float>(sycl_device));
    CALL_SUBTEST_5(sycl_scan_test_inclusive_dim1_per_device<float>(sycl_device));
    CALL_SUBTEST_6(sycl_scan_test_inclusive_dim2_per_device<float>(sycl_device));
  }
}

References CALL_SUBTEST_1, CALL_SUBTEST_2, CALL_SUBTEST_3, CALL_SUBTEST_4, CALL_SUBTEST_5, and CALL_SUBTEST_6.

sycl_scan_test_exclusive_dim0_per_device()

template<typename DataType , typename Dev >

void sycl_scan_test_exclusive_dim0_per_device

(

const Dev &

sycl_device

)

                                                                      {
  test_sycl_cumsum<DataType, ColMajor, int64_t>(sycl_device, 2049, 1023, 127, 0, true);
  test_sycl_cumsum<DataType, RowMajor, int64_t>(sycl_device, 2049, 1023, 127, 0, true);
}

sycl_scan_test_exclusive_dim1_per_device()

template<typename DataType , typename Dev >

void sycl_scan_test_exclusive_dim1_per_device

(

const Dev &

sycl_device

)

                                                                      {
  test_sycl_cumsum<DataType, ColMajor, int64_t>(sycl_device, 1023, 2049, 127, 1, true);
  test_sycl_cumsum<DataType, RowMajor, int64_t>(sycl_device, 1023, 2049, 127, 1, true);
}

sycl_scan_test_exclusive_dim2_per_device()

template<typename DataType , typename Dev >

void sycl_scan_test_exclusive_dim2_per_device

(

const Dev &

sycl_device

)

                                                                      {
  test_sycl_cumsum<DataType, ColMajor, int64_t>(sycl_device, 1023, 127, 2049, 2, true);
  test_sycl_cumsum<DataType, RowMajor, int64_t>(sycl_device, 1023, 127, 2049, 2, true);
}

sycl_scan_test_inclusive_dim0_per_device()

template<typename DataType , typename Dev >

void sycl_scan_test_inclusive_dim0_per_device

(

const Dev &

sycl_device

)

                                                                      {
  test_sycl_cumsum<DataType, ColMajor, int64_t>(sycl_device, 2049, 1023, 127, 0, false);
  test_sycl_cumsum<DataType, RowMajor, int64_t>(sycl_device, 2049, 1023, 127, 0, false);
}

sycl_scan_test_inclusive_dim1_per_device()

template<typename DataType , typename Dev >

void sycl_scan_test_inclusive_dim1_per_device

(

const Dev &

sycl_device

)

                                                                      {
  test_sycl_cumsum<DataType, ColMajor, int64_t>(sycl_device, 1023, 2049, 127, 1, false);
  test_sycl_cumsum<DataType, RowMajor, int64_t>(sycl_device, 1023, 2049, 127, 1, false);
}

sycl_scan_test_inclusive_dim2_per_device()

template<typename DataType , typename Dev >

void sycl_scan_test_inclusive_dim2_per_device

(

const Dev &

sycl_device

)

                                                                      {
  test_sycl_cumsum<DataType, ColMajor, int64_t>(sycl_device, 1023, 127, 2049, 2, false);
  test_sycl_cumsum<DataType, RowMajor, int64_t>(sycl_device, 1023, 127, 2049, 2, false);
}

test_sycl_cumsum()

template<typename DataType , int DataLayout, typename IndexType >

void test_sycl_cumsum	(	const Eigen::SyclDevice &	sycl_device,
		IndexType	m_size,
		IndexType	k_size,
		IndexType	n_size,
		int	consume_dim,
		bool	exclusive
	)

                                                       {
  static const DataType error_threshold = 1e-4f;
  std::cout << "Testing for (" << m_size << "," << k_size << "," << n_size << " consume_dim : " << consume_dim << ")"
            << std::endl;
  Tensor<DataType, 3, DataLayout, IndexType> t_input(m_size, k_size, n_size);
  Tensor<DataType, 3, DataLayout, IndexType> t_result(m_size, k_size, n_size);
  Tensor<DataType, 3, DataLayout, IndexType> t_result_gpu(m_size, k_size, n_size);
 
  t_input.setRandom();
  std::size_t t_input_bytes = t_input.size() * sizeof(DataType);
  std::size_t t_result_bytes = t_result.size() * sizeof(DataType);
 
  DataType* gpu_data_in = static_cast<DataType*>(sycl_device.allocate(t_input_bytes));
  DataType* gpu_data_out = static_cast<DataType*>(sycl_device.allocate(t_result_bytes));
 
  array<IndexType, 3> tensorRange = {{m_size, k_size, n_size}};
  TensorMap<Tensor<DataType, 3, DataLayout, IndexType>> gpu_t_input(gpu_data_in, tensorRange);
  TensorMap<Tensor<DataType, 3, DataLayout, IndexType>> gpu_t_result(gpu_data_out, tensorRange);
  sycl_device.memcpyHostToDevice(gpu_data_in, t_input.data(), t_input_bytes);
  sycl_device.memcpyHostToDevice(gpu_data_out, t_input.data(), t_input_bytes);
 
  gpu_t_result.device(sycl_device) = gpu_t_input.cumsum(consume_dim, exclusive);
 
  t_result = t_input.cumsum(consume_dim, exclusive);
 
  sycl_device.memcpyDeviceToHost(t_result_gpu.data(), gpu_data_out, t_result_bytes);
  sycl_device.synchronize();
 
  for (IndexType i = 0; i < t_result.size(); i++) {
    if (static_cast<DataType>(std::fabs(static_cast<DataType>(t_result(i) - t_result_gpu(i)))) < error_threshold) {
      continue;
    }
    if (Eigen::internal::isApprox(t_result(i), t_result_gpu(i), error_threshold)) {
      continue;
    }
    std::cout << "mismatch detected at index " << i << " CPU : " << t_result(i) << " vs SYCL : " << t_result_gpu(i)
              << std::endl;
    assert(false);
  }
  sycl_device.deallocate(gpu_data_in);
  sycl_device.deallocate(gpu_data_out);
}

References assert, Eigen::Tensor< Scalar_, NumIndices_, Options_, IndexType_ >::data(), Eigen::TensorBase< Derived, AccessLevel >::device(), error_threshold, boost::multiprecision::fabs(), i, Eigen::internal::isApprox(), Eigen::TensorBase< Derived, AccessLevel >::setRandom(), and Eigen::Tensor< Scalar_, NumIndices_, Options_, IndexType_ >::size().