Methods for writing multiple LevelData to an HDF5 file. More...

#include <CH_HDF5.H>

Public Member Functions
	WriteMultiData (HDF5Handle &a_handle, const BoxLayout &a_layout, const int a_numIntv, const string &a_name, const int a_policyFlags=CH_HDF5::IOPolicyDefault, const IntVect &a_outputGhost=IntVect::Zero, const bool a_newForm=false)
	Constructor writes boxes and allocates dataset for LevelData. More...

	~WriteMultiData ()
	Destructor. More...

int	writeData (const BoxLayoutData< T > &a_data, const Interval &a_intvMem, const Interval &a_intvFile)
	Write an interval of LevelData to the dataset. More...

Protected Attributes
const int	m_policyFlags
	Policies. More...

const IntVect	m_outputGhost
	Number of ghost cells written. More...

const bool	m_newForm
	? More...

char	m_dataname [128]
	Name for level data dataset. More...

hid_t	m_dataSet
	Dataset for level data. More...

Interval	m_allIntvFile
	Interval for components in file. More...

long	m_maxBoxPerProc
	Maximum boxes written by any proc. More...

Vector< Vector< long long > >	m_offsets

Vector< hid_t >	m_types
	Type of data written. More...

Private Member Functions
	WriteMultiData (const WriteMultiData &)

WriteMultiData &	operator= (const WriteMultiData)

Detailed Description

template<class T>
class WriteMultiData< T >

Methods for writing multiple LevelData to an HDF5 file.

While the write functions can be used to write a single LevelData, they do not support writing intervals from multiple LevelData. This is often the case when you want to add diagnostic information to the output file. The purpose of this class is to handle writing multiple LevelData to a single file. The dataset for the LevelData is created during construction, when the boxes are written. Any number of intervals from LevelData can then be written.

Special behavior can be selected with the policy flags:

CH_HDF5::IOPolicyMultiDimHyperslab - The memory dataspace will be set up so that HDF5 linearizes the data. Using this requires T.dataPtr() and contiguous memory (so things like T=FluxBox will not work).
CH_HDF5::IOPolicyCollectiveWrite - The write for each dataset will be collective. The write is always parallel but otherwise will be done independently. May or may not provide a speedup but it should work. Rumor has it the BG requires this.

With MPI-everywhere on x86 architectures, it is highly likely that the default policy CH_HDF5::IOPolicyDefault gives the best performance (i.e., use internal linearization and independent I/O). In some limited testing (2 procs, 20 cores each): IOPolicyDefault - 6.77 IOPolicyCollectiveWrite - 33.33 IOPolicyMultiDimHyperslab - 168.14 IOPolicyCollectiveWrite | IOPolicyMultiDimHyperslab - 30.94 which is difficult to make sense of.

Note

Users are resposible for matching the total number of components used to allocate the dataset match the number of levelData added to the dataset
Only a single datatype is supported per file
Since the dataset must be created in advance, T can only be statically allocatable (T::preAllocatable() == 0)
API 1.6 is not supported

Constructor & Destructor Documentation

◆ WriteMultiData() [1/2]

template<class T >

WriteMultiData< T >::WriteMultiData	(	HDF5Handle &	a_handle,
		const BoxLayout &	a_layout,
		const int	a_numIntv,
		const string &	a_name,
		const int	a_policyFlags = `CH_HDF5::IOPolicyDefault`,
		const IntVect &	a_outputGhost = `IntVect::Zero`,
		const bool	a_newForm = `false`
	)

Constructor writes boxes and allocates dataset for LevelData.

Write boxes, processors, attributes and creates a dataset for all the level data to be written. The dataset is closed on destruction.

Parameters

[in]	a_handle	Chombo HDF5 handle holding open file for writing
[in]	a_layout	Box layout for all data to be written
[in]	a_numIntv	Total number of components that will be written from all BoxLayoutData or LevelData
[in]	a_name	Name of the dataset for the level data.
[in]	a_policyFlags	Either CH_HDF5::IOPolicyDefault - Use internal T.linearOut(...) to linearize data and processes independently write to file, but still in parallel. or a union (\|) of the following flags: CH_HDF5::IOPolicyMultiDimHyperslab - The memory dataspace will be set up so that HDF5 linearizes the data. Using this requires T.dataPtr() and contiguous memory (so things like T=FluxBox will not work). CH_HDF5::IOPolicyCollectiveWrite - The write for each dataset will be collective.
[in]	a_outputGhost	Number of ghost cells that will be written to the data file. Any data written must have this many ghosts
[in]	a_newForm	?

◆ ~WriteMultiData()

template<class T >

WriteMultiData< T >::~WriteMultiData ( )

inline

Destructor.

◆ WriteMultiData() [2/2]

template<class T >

WriteMultiData< T >::WriteMultiData ( const WriteMultiData< T > & )

private

Member Function Documentation

◆ operator=()

template<class T >

WriteMultiData& WriteMultiData< T >::operator= ( const WriteMultiData< T > )

private

◆ writeData()

template<class T >

int WriteMultiData< T >::writeData	(	const BoxLayoutData< T > &	a_data,
		const Interval &	a_intvMem,
		const Interval &	a_intvFile
	)

Write an interval of LevelData to the dataset.

Call this as many times as you want to write level data. User is resposible for making sure 'a_intvFile' has relevance to 'a_numIntv' specified during construction.

Parameters