RetypedArrayEngine.h

Classes

RetypedArrayEngine -- Virtual column engine to retype and reshape arrays. (full description)

template<class VirtualType, class StoredType> class RetypedArrayEngine : public BaseMappedArrayEngine<VirtualType,StoredType>

Interface

Public Members
RetypedArrayEngine (const String& virtualColumnName, const String& storedColumnName)
RetypedArrayEngine (const String& virtualColumnName, const String& storedColumnName, const IPosition& virtualElementShape, const TableRecord& extraInformation)
RetypedArrayEngine (const Record& spec)
~RetypedArrayEngine()
virtual String dataManagerType() const
virtual String dataManagerName() const
virtual Record dataManagerSpec() const
static String className()
static void registerClass()
Private Members
RetypedArrayEngine (const RetypedArrayEngine<VirtualType,StoredType>&)
RetypedArrayEngine<VirtualType,StoredType>& operator= (const RetypedArrayEngine<VirtualType,StoredType>&)
DataManager* clone() const
void create (uInt initialNrrow)
void prepare()
void setShapeColumn (const IPosition& shape)
void setShape (uInt rownr, const IPosition& shape)
uInt ndim (uInt rownr)
IPosition shape (uInt rownr)
void getArray (uInt rownr, Array<VirtualType>& array)
void putArray (uInt rownr, const Array<VirtualType>& array)
void getSlice (uInt rownr, const Slicer& slicer, Array<VirtualType>& array)
void putSlice (uInt rownr, const Slicer& slicer, const Array<VirtualType>& array)
void getArrayColumn (Array<VirtualType>& array)
void putArrayColumn (const Array<VirtualType>& array)
void getColumnSlice (const Slicer& slicer, Array<VirtualType>& array)
void putColumnSlice (const Slicer& slicer, const Array<VirtualType>& array)
IPosition checkShape (const Array<VirtualType>& source, const Array<StoredType>& target)
void copyOnGet (Array<VirtualType>& array, const Array<StoredType>& stored)
void copyOnPut (const Array<VirtualType>& array, Array<StoredType>& stored)
IPosition storedShape (uInt rownr, const IPosition& virtualShape)
Slicer storedSlicer (const Slicer& virtualSlicer) const
Public Members
static DataManager* makeObject (const String& dataManagerType, const Record& spec)

Description

Review Status

Reviewed By:
Brian Glendenning
Date Reviewed:
1995/12/20
Programs:
Demos:
Tests:

Prerequisite

Synopsis

RetypedArrayEngine maps a virtual column containing arrays of objects to a stored column containing arrays of data of another type. Usually the dimensionality of the arrays get smaller during this mapping process. The engine makes it possible to store an array of any type in a table.
For example, a column with 2D arrays of StokesVector's can be mapped to a column with 3D arrays of floats (of which the first axes has, say, length 4). Another example is mapping a 2D array of StokesMatrix's to a 4D array of floats.

The mapping process has to be done by a (static) set and get function in the VirtualType class. When a RetypedArrayEngine object is constructed, it is possible to pass information in a TableRecord. This TableRecord is indirectly passed to the set/get functions. This is done by means of the function newCopyInfo, which can preprocess the information in the TableRecord and store it in another object. That object is passed to the set and get functions. At the end a function deleteCopyInfo is called to delete the object. Of course, it is not needed to allocate such an object; newCopyInfo can return a null pointer.

Tip Because the variables have to be generic and because of limitations in the CFront compiler, several variables have to be passed as void* and need to be casted in the set/get functions.

The virtual column data type class has to contain several functions. The example shows how they can be implemented.

static String dataTypeId();
has to give the (unique) name of the class.
static IPosition shape();
This has to return the full shape of the elements in the virtual. E.g. StokesVector will return [4]. StokesMatrix will return [4,4].
static void* newCopyInfo (const TableRecord& record, const IPosition& virtualElementShape);
This function has to setup the set/get functions by preprocessing the information contained in the TableRecord and storing it in a so-called "copyInfo" object. A pointer to that object has to be returned, which is kept by the engine and passed to the set/get functions. The "copyInfo" class can be a nested class in the VirtualType (as shown in the StokesVector example), but it can also be an independent class.
The supplied TableRecord is the TableRecord given when constructing the engine. When no TableRecord was given, it will be empty. The supplied shape is the shape of a virtual element as given to the constructor of the engine. This can be a full or partial shape. E.g. for a StokesVector it will usually be [4], but it can also, say, [1] if only U is used. The function could check if the information in the TableRecord and the shape match.
Of course, a VirtualType may not need any extra information. Therefore it is possible to return a null "copyInfo" pointer.
static void deleteCopyInfo (void* copyInfo);
This function has to delete the "copyInfo" object allocated by newCopyInfo. To do so, it needs to cast the pointer to the correct type.
static void set (void* copyInfo, void* out, const Array<StoredType>& in, const IPosition& virtualElementShape);
This function is called when an Array<VirtualType> is read. It has to convert the StoredType array to the VirtualType array. In principle, there are two different cases (which can be deduced from the given shape):
  1. The stored information is complete. For example: suppose the VirtualType is a StokesVector object (containing I, Q, U and V), When the stored array contains 4 values per StokesVector, it is complete.
    In this case the entire virtual array can be directly copied from the stored array when the VirtualType object contains no virtual functions and the data are directly contained in it. The function
    retypedArrayEngineSet (Array<VirtualType>& out, const Array<StoredType>& in);
    can be used for this purpose.
  2. When in the example above the stored array contains less than 4 values per StokesVector, the stored information is incomplete. In this case the set function has to fill the data in one way or another. The information in the "copyInfo" object can assist in it.
    Each VirtualType element has to be set individually, so a loop through the array is required. To assist in this, the loop has been implemented in the function
    retypedArrayEngineSet (Array<VirtualType>& out, const Array<StoredType>& in, const void* extraArgument);
    It calls the VirtualType function
                 void setElem (const StoredType* data, const IPosition& shape,
                               const void* extraArgument);
    
    for each VirtualType element. This set function has to fill the VirtualType object from the data. It can use the shape and the extraArgument to know how it should do it.
    Note that the 3-argument function retypedArrayEngineSet is only a convenience function. For optimal performance it may be needed to handcode the loop instead of using this function.
Warning Note that the given virtual element shape does not need to match the shape given to the constructor of the engine. It is possible that the user sets the shape of the stored array before putting the virtual array. In that case the system uses the relevant part of the stored array shape as the virtual element shape.
Tip When the out argument is declared (as it should be) as Array<VirtualType>& out, the CFront compiler complains about unknown size of VirtualType when instantiating Array. Therefore it has to be declared as void* and the set function needs to cast it to Array<VirtualType>*.
static void get (void* copyInfo, Array<float>& out, const void* in, const IPosition& virtualElementShape);
This function is similar to the set function described above, but is called when an Array<VirtualType> is written. It has to convert the VirtualType array to the StoredType array.


E.g.: A StokesVector has 4 float elements.

    // Construct the column object for the Stokes column.
    ArrayColumn<StokesVector> stokesColumn (table, "StokesVirtualColumn");
    // Put an array of StokesVector's with shape 512,512.
    // This will implicitly set the shape of the underlying
    // data column to 4,512,512.
    // This put is very quick (it can copy all data in one go).
    Array<StokesVector> stokesData (IPosition(2,512,512));
    stokesColumn.put (rownr, stokesData);

    // Get the column object for the Data column.
    // Set its shape explicitly to 1,512,512,
    ArrayColumn<float> dataColumn (table, "DataColumn");
    dataColumn.setShape (rownr, IPosition(3,1,512,512));
    // Now a put of the data results in calling the StokesVector::getElem
    // function for each element with an IPosition(1,1); i.e. the
    // data array needs only one value for each StokesVector.
    stokesColumn.put (rownr, stokesData);

When reading a table back, the engine has to be registered. Otherwise it will be unknown to the table system. Similarly, the appropriate ArrayColumnDesc object has to be registered. This can be done as follows:

    RetypedArrayEngine::registerClass();
    ArrayColumnDesc tmp(ColumnDesc::registerMap);
 
When they are not registered, the open of the table will fail telling which class could not be found.

Motivation

This class allows one to store arrays of arbitrary objects in a table. It also allows it to be done it in a very efficient way.

The class had to be doubly templated. There were 2 reasons:

  1. The typedef trick described on page 321 in Barton/Nackman did not work with the CFront-based ObjectCenter compiler.
  2. It was needed to allow derivation from BaseMappedArrayEngine.

Originally it was the idea to have a mandatory nested CopyInfo class in the VirtualType class and use syntax like VirtualType::CopyInfo to access functions in it and to keep a pointer to such an object. Alas, the CFront compiler could not handle this.

Because the engine can serve only one column, it was possible to combine the engine and the column functionality in one class. This has been achieved using multiple inheritance. The advantage of this is that only one templated class is used, so fewer template instantiations are needed.

Example

The following example shows how a StokesVector could be implemented. It doesn't check whether the mask is correct. Two more examples are contained in the demo/test program dRetypedArrayEngine.h and its .cc file. Their second example (class RetypedArrayEx2) is similar to the StokesVector example below, but contains more extensive checking.
    //# Forward Declarations
    template<class T> class Array;
    template<class T> class Vector;
    
    class StokesVector
    {
    public:
        StokesVector(): I_p(0), Q_p(0), U_p(0), V_p(0) {}
        StokesVector(double i, double q, double u, double v)
             : I_p(i), Q_p(q), U_p(u), V_p(v) {}
        StokesVector(const StokesVector& that): I_p(that.I_p), Q_p(that.Q_p),
                                                U_p(that.U_p), V_p(that.V_p) {}
        StokesVector& operator= (const StokesVector& that)
           { I_p=that.I_p; Q_p=that.Q_p; U_p=that.U_p; V_p=that.V_p;
             return *this; }
    
        static String dataTypeId()
            { return "StokesVector"; }
    
        // A StokesVector is 1-dim and contains 4 elements.
        static IPosition shape()
            { return IPosition (1,4); }
    
        // Preprocess possible information in the TableRecord.
        static void* newCopyInfo (const TableRecord& record,
                                  const IPosition& shape)
            { return new CopyInfo(record, shape); }
    
        // Delete the object containing preprocessed information.
        static void* deleteSetDet (void* copyInfo)
            { delete (CopyInfo*)copyInfo; }
    
        // Convert a StoredType array to a VirtualType array.
        // Do this in a CopyInfo function to use its preprocessed information.
        static void set (void* copyInfo, void* out,
                         const Array<double>& in, const IPosition& shape)
            { ((CopyInfo*)copyInfo)->set (out, in, shape); }
    
        // Convert a VirtualType array to a StoredType array.
        // Do this in a CopyInfo function to use its preprocessed information.
        static void get (void* copyInfo, Array<double>& out,
                         const void* in, const IPosition& shape)
            { ((CopyInfo*)copyInfo)->get (out, in, shape); }
    
        // This nested class is used to hold preprocessed information. It
        // holds a mask extracted from the TableRecord supplied to the engine.
        // One can imagine that it could also extract a flag telling
        // whether the stored data is stored as I,Q,U,V or as XX,YY,XY,YX
        // (although such a conversion would probably be better handled
        // by a separate virtual column engine).
        class CopyInfo {
        public:
            // The constructor extracts the mask from the record.
            void CopyInfo (const TableRecord& record)
                {
                    RORecordFieldRef<Array<Bool> > field (record, 0);
                    mask_p = new Vector<Bool>;
                    *mask_p = *field;
                }
            // The set function fills the StokesVector.
            // It uses the general functions for that purpose.
            void set (void* vout, const Array<double>& in,
                      const IPosition& shape)
                {
                    Array<StokesVector>& out = *(Array<StokesVector>*)vout;
                    if (shape.nelements() == 1  &&  shape(0) == 4) {
                        // All values available, copy in one go.
                        // This can be done because a StokesVector object
                        // only contains 4 double values (and no virtual
                        // function table).
                        retypedArrayEngineSet (out, in);
                    }else{
                        // Only some values available. Fill each
                        // StokesVector object using the shape and mask.
                        // The set function below is called for each object.
                        retypedArrayEngineSet (out, in, shape, (void*)mask_p);
                    }
                }
            // get is the opposite of set.
            void get (Array<double>& out, const void* vin,
                      const IPosition& shape)
                {
                    const Array<StokesVector>& in =
                                             *(const Array<StokesVector>*)vin;
                    if (shape.nelements() == 1  &&  shape(0) == 4) {
                        retypedArrayEngineGet (out, in);
                    }else{
                        retypedArrayEngineGet (out, in, shape, (void*)mask_p);
                    }
        private:
            Vector<Bool>* mask_p;
        };
    
        // Set values of StokesVector using the mask.
        // The shape is not used here.
        void setElem (const double* data, const IPosition&, const void* maskPtr)
            {
                 const Vector<Bool>& mask = *(const Vector<Bool>*)maskPtr;
                 I_p = Q_p = U_p = V_p = 0;
                 if (mask(0)) {
                     I_p = *data++;
                 }
                 if (mask(1)) {
                     Q_p = *data++;
                 }
                 if (mask(2)) {
                     U_p = *data++;
                 }
                 if (mask(3)) {
                     V_p = *data;
                 }
            }
        // Get values of StokesVector using the mask (opposite of setElem).
        void getElem (double* data, const IPosition&, const void* maskPtr);
    private:
       double I_p, Q_p, U_p, V_p;
    };
    
    main() {
       // First register the virtual column engine.
       RetypedArrayEngine<StokesVector,double>::registerClass();
       // Add ArrayColumnDesc<StokesVector> to column type map.
       ArrayColumnDesc<StokesVector> tmp(ColumnDesc::registerMap);
    
       // Build the table description.
       TableDesc td("", "1", TableDesc::Scratch);
       td.addColumn (ArrayColumnDesc<double> ("Data"));
       td.addColumn (ArrayColumnDesc<StokesVector> ("Stokes"));
    
       // Now create a new table from the description.
       SetupNewTable newtab("tRetypedArrayEngine_tmp.data", td, Table::New);
       // Create the virtual column engine with the stored columns Data.
       RetypedArrayEngine<StokesVector,double> engine ("Stokes", "Data");
       newtab.bindColumn ("Stokes", engine);
       Table tab(newtab, 50);
    
       // Fill the table via the virtual columns.
       ArrayColumn<StokesVector> stokesColumn (tab, "Stokes");
       Vector<StokesVector> vec(10);
       uInt i;
       for (i=0; i<tab.nrow(); i++) {
           stokesColumn.put (i, vec);
       }
    }
    
Caution Due to instantiation problems with the CFront-based ObjectCenter compiler (and probably other CFront-based compilers as well) the Array and Vector have to be forward declared. Array.h and Vector.h should NOT be included in this StokesVector.h, thus the implementations should not be inlined (they are too large anyway), but put in a separate .cc file where Array.h and Vector.h can be included.

Another compiler problem is that the variable mask_p is not automatically converted to a void*, so an explicit cast has to be done.

Template Type Argument Requirements (VirtualType)

Template Type Argument Requirements (StoredType)

Member Description

RetypedArrayEngine (const String& virtualColumnName, const String& storedColumnName)

Construct an engine to map a virtual column containing arrays with an arbitrary data type to arrays in a stored column. StoredColumnName is the name of the column where the converted data will be put and must have data type StoredType. The virtual column using this engine must have data type VirtualType.

RetypedArrayEngine (const String& virtualColumnName, const String& storedColumnName, const IPosition& virtualElementShape, const TableRecord& extraInformation)

Construct an engine to map a virtual column containing arrays with an arbitrary data type to arrays in a stored column. StoredColumnName is the name of the column where the converted data will be put and must have data type StoredType. The virtual column using this engine must have data type VirtualType. The shape and record provided is handed to the newCopyInfo function in the VirtualType class. It can be used to determine how an element has to be handled when the stored data is incomplete.

RetypedArrayEngine (const Record& spec)

Construct from a record specification as created by getmanagerSpec().

~RetypedArrayEngine()

Destructor is mandatory.

virtual String dataManagerType() const

Return the type name of the engine (i.e. its class name).

virtual String dataManagerName() const

Get the name given to the engine (is the virtual column name).

virtual Record dataManagerSpec() const

Record a record containing data manager specifications.

static String className()

Return the name of the class. This includes the names of the template arguments.

static void registerClass()

Register the class name and the static makeObject "constructor". This will make the engine known to the table system. The automatically invoked registration function in DataManReg.cc contains RetypedArrayEngine. Any other instantiation of this class must be registered "manually" (or added to DataManReg.cc).

RetypedArrayEngine (const RetypedArrayEngine<VirtualType,StoredType>&)

Copy constructor is only used by clone(). (so it is made private).

RetypedArrayEngine<VirtualType,StoredType>& operator= (const RetypedArrayEngine<VirtualType,StoredType>&)

Assignment is not needed and therefore forbidden (so it is made private and not implemented).

DataManager* clone() const

Clone the engine object.

void create (uInt initialNrrow)

Initialize the object for a new table. It defines the keywords containing the engine parameters.

void prepare()

Preparing consists of setting the writable switch and adding the initial number of rows in case of create. Furthermore it reads the keywords containing the engine parameters and allocates a CopyInfo object for the VirtualType.

void setShapeColumn (const IPosition& shape)

Set the shape of the FixedShape arrays in the column. This function only gets called if the column has FixedShape arrays. The shape gets saved and used to set the shape of the arrays in the stored in case the stored has non-FixedShape arrays.

void setShape (uInt rownr, const IPosition& shape)

Define the shape of the array in the given row. When the shape of the (underlying) stored array has already been defined, it checks whether its latter dimensions match the given virtual shape. When matching, nothing will be done. When mismatching or when the stored shape has not been defined yet, the stored shape will be defined from the virtual shape and the virtual element shape. E.g. in case of a StokesVector a virtual shape of (512,512) results in a stored shape of (4,512,512).

uInt ndim (uInt rownr)

Get the dimensionality of the array in the given row.

IPosition shape (uInt rownr)

Get the shape of the array in the given row. This is done by stripping the first dimension(s) from the shape of the underlying stored array.

void getArray (uInt rownr, Array<VirtualType>& array)

Get an array in the given row.

void putArray (uInt rownr, const Array<VirtualType>& array)

Put an array in the given row.

void getSlice (uInt rownr, const Slicer& slicer, Array<VirtualType>& array)

Get a section of the array in the given row.

void putSlice (uInt rownr, const Slicer& slicer, const Array<VirtualType>& array)

Put into a section of the array in the given row.

void getArrayColumn (Array<VirtualType>& array)

Get an entire column.

void putArrayColumn (const Array<VirtualType>& array)

Put an entire column.

void getColumnSlice (const Slicer& slicer, Array<VirtualType>& array)

Get a section of all arrays in the column.

void putColumnSlice (const Slicer& slicer, const Array<VirtualType>& array)

Put a section of all arrays in the column.

IPosition checkShape (const Array<VirtualType>& source, const Array<StoredType>& target)

Check if the shapes of virtual and stored match. Determine the shape of the virtual elements in the stored.

void copyOnGet (Array<VirtualType>& array, const Array<StoredType>& stored)

Copy the stored array to the virtual array. It tries to optimize as much as possible.

void copyOnPut (const Array<VirtualType>& array, Array<StoredType>& stored)

Copy the virtual array to the stored array. It tries to optimize as much as possible.

IPosition storedShape (uInt rownr, const IPosition& virtualShape)

Determine the shape of a cell in the stored column from the shape of the cell in the virtual column.

Slicer storedSlicer (const Slicer& virtualSlicer) const

Convert the Slicer for a virtual to a Slicer for the stored.

static DataManager* makeObject (const String& dataManagerType, const Record& spec)

Define the "constructor" to construct this engine when a table is read back. This "constructor" has to be registered by the user of the engine. If the engine is commonly used, its registration can be added to the registerAllCtor function in DataManReg.cc. That function gets automatically invoked by the table system.