[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
2.1 Getting started 2.2 Public types Public types declaration for Array 2.3 Constructors Array constructors 2.4 Indexing, subarrays, and slicing How to access the elements of an Array? 2.4.4 Slicing The slicing machinery 2.5 Debug mode How to debug a program that uses Blitz++? 2.6 Member functions Array member functions 2.7 Global functions Array global functions 2.8 Inputting and Outputting Arrays Inputting and outputting Array's 2.9 Array storage orders The storage of Array
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
Currently, Blitz++ provides a single array class, called
Array<T_numtype,N_rank>
. This array class provides a dynamically
allocated N-dimensional array, with reference counting, arbitrary storage
ordering, subarrays and slicing, flexible expression handling, and many
other useful features.
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
The Array
class takes two template parameters:
T_numtype
is the numeric type to be stored in the array. T_numtype
can be an
integral type (bool
, char
, unsigned char
, short
int
, short unsigned int
, int
, unsigned int
,
long
, unsigned long
), floating point type (float
,
double
, long double
), complex type (complex<float>
,
complex<double>
, complex<long double>
) or any user-defined
type with appropriate numeric semantics.
N_rank
is the rank (or dimensionality) of the array. This should be a
positive integer.
To use the Array
class, include the header <blitz/array.h>
and
use the namespace blitz
:
#include <blitz/array.h> using namespace blitz; Array<int,1> x; // A one-dimensional array of int Array<double,2> y; // A two-dimensional array of double . . Array<complex<float>, 12> z; // A twelve-dimensional array of complex<float> |
When no constructor arguments are provided, the array is empty, and no memory is allocated. To create an array which contains some data, provide the size of the array as constructor arguments:
Array<double,2> y(4,4); // A 4x4 array of double |
The contents of a newly-created array are garbage. To initialize the array, you can write:
y = 0; |
and all the elements of the array will be set to zero. If the contents of
the array are known, you can initialize it using a comma-delimited list of
values. For example, this code excerpt sets y
equal to a 4x4
identity matrix:
y = 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1; |
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
The Array<T,N>
class supports a variety of arrays:
Array<int,1>
and
Array<float,3>
Array<complex<float>,2>
Polynomial
, then Array<Polynomial,2>
is an array of
Polynomial
objects.
TinyVector
and
TinyMatrix
, in which each element is a fixed-size vector or array.
For example, Array<TinyVector<float,3>,3>
is a three-dimensional
vector field.
Array<Array<int,1>,1>
, in
which each element is a variable-length array.
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
Here's an example program which creates two 3x3 arrays, initializes them, and adds them:
#include <blitz/array.h> using namespace blitz; int main() { Array<float,2> A(3,3), B(3,3), C(3,3); A = 1, 0, 0, 2, 2, 2, 1, 0, 0; B = 0, 0, 7, 0, 8, 0, 9, 9, 9; C = A + B; cout << "A = " << A << endl << "B = " << B << endl << "C = " << C << endl; return 0; } |
and the output:
A = 3 x 3 [ 1 0 0 2 2 2 1 0 0 ] B = 3 x 3 [ 0 0 7 0 8 0 9 9 9 ] C = 3 x 3 [ 1 0 7 2 10 2 10 9 9 ] |
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
Blitz++ is very flexible about the way arrays are stored in memory.
The default storage format is row-major, C-style arrays whose indices start at zero.
Fortran-style arrays can also be created. Fortran arrays are stored in
column-major order, and have indices which start at one. To create a
Fortran-style array, use this syntax: Array<int,2> A(3, 3,
fortranArray);
The last parameter, fortranArray
, tells the
Array
constructor to use a fortran-style array format.
fortranArray
is a global object which has an automatic conversion to
type GeneralArrayStorage<N>
. GeneralArrayStorage<N>
encapsulates information about how an array is laid out in memory. By
altering the contents of a GeneralArrayStorage<N>
object, you can lay
out your arrays any way you want: the dimensions can be ordered arbitrarily
and stored in ascending or descending order, and the starting indices can be
arbitrary.
Creating custom array storage formats is described in a later section (2.9 Array storage orders).
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
The Array
class declares these public types:
T_numtype
is the element type stored in the array. For
example, the type Array<double,2>::T_numtype
would be double
.
T_index
is a vector index into the array. The class
TinyVector
is used for this purpose.
T_array
is the array type itself
(Array<T_numtype,N_rank>
)
T_iterator
is an iterator type. NB: this iterator is not
yet fully implemented, and is NOT STL compatible at the present time.
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
Array(); Array(GeneralArrayStorage<N_rank> storage) |
The default constructor creates a C-style array of zero size. Any attempt to access data in the array may result in a run-time error, because there isn't any data to access!
An optional argument specifies a storage order for the array.
Arrays created using the default constructor can subsequently be given data
by the resize()
, resizeAndPreserve()
, or reference()
member functions.
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
Array(expression...) |
You may create an array from an array expression. For example,
Array<float,2> A(4,3), B(4,3); // ... Array<float,2> C(A*2.0+B); |
This is an explicit constructor (it will not be used to perform implicit type conversions). The newly constructed array will have the same storage format as the arrays in the expression. If arrays with different storage formats appear in the expression, an error will result. (In this case, you must first construct the array, then assign the expression to it).
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
Array(int extent1); Array(int extent1, int extent2); Array(int extent1, int extent2, int extent3); ... Array(int extent1, int extent2, int extent3, ..., int extent11) |
These constructors take arguments which specify the size of the array to be constructed. You should provide as many arguments as there are dimensions in the array.(1)
An optional last parameter specifies a storage format:
Array(int extent1, GeneralArrayStorage<N_rank> storage); Array(int extent1, int extent2, GeneralArrayStorage<N_rank> storage); ... |
For high-rank arrays, it may be convenient to use this constructor:
Array(const TinyVector<int, N_rank>& extent); Array(const TinyVector<int, N_rank>& extent, GeneralArrayStorage<N_rank> storage); |
The argument extent
is a vector containing the extent (length) of the
array in each dimension. The optional second parameter indicates a storage
format. Note that you can construct TinyVector<int,N>
objects on the
fly with the shape(i1,i2,...)
global function. For example,
Array<int,2> A(shape(3,5))
will create a 3x5 array.
A similar constructor lets you provide both a vector of base index values (lbounds) and extents:
Array(const TinyVector<int, N_rank>& lbound, const TinyVector<int, N_rank>& extent); Array(const TinyVector<int, N_rank>& lbound, const TinyVector<int, N_rank>& extent, GeneralArrayStorage<N_rank> storage); |
The argument lbound
is a vector containing the base index value (or
lbound) of the array in each dimension. The argument extent
is a
vector containing the extent (length) of the array in each dimension. The
optional third parameter indicates a storage format. As with the above
constructor, you can use the shape(i1,i2,...)
global function to
create the lbound
and extent
parameters.
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
These constructors allow arbitrary bases (starting indices) to be set:
Array(Range r1); Array(Range r1, Range r2); Array(Range r1, Range r2, Range r3); ... Array(Range r1, Range r2, Range r3, ..., Range r11); |
For example, this code:
Array<int,2> A(Range(10,20), Range(20,30)); |
will create an 11x11 array whose indices are 10..20 and 20..30. An optional last parameter provides a storage order:
Array(Range r1, GeneralArrayStorage<N_rank> storage); Array(Range r1, Range r2, GeneralArrayStorage<N_rank> storage); ... |
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
This constructor makes a shared view of another array's data:
Array(Array<T_numtype, N_rank>& array); |
After this constructor is used, both Array
objects refer to the
same data. Any changes made to one array will appear in the other
array. If you want to make a duplicate copy of an array, use the
copy()
member function.
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
Arrays may be constructed from expressions, which are described in 3.1 Expression evaluation order. The syntax is:
Array(...array expression...); |
For example, this code creates an array B which contains the square roots of the elements in A:
Array<float,2> A(N,N); // ... Array<float,2> B(sqrt(A)); |
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
When creating an array using a pointer to already existing data, you have
three choices for how Blitz++ will handle the data. These choices are
enumerated by the enum type preexistingMemoryPolicy
:
enum preexistingMemoryPolicy { duplicateData, deleteDataWhenDone, neverDeleteData }; |
If you choose duplicateData
, Blitz++ will create an array object
using a copy of the data you provide. If you choose
deleteDataWhenDone
, Blitz++ will not create a copy of the data; and
when no array objects refer to the data anymore, it will deallocate the data
using delete []
. Note that to use deleteDataWhenDone
, your
array data must have been allocated using the C++ new
operator -- for
example, you cannot allocate array data using Fortran or malloc
, then
create a Blitz++ array from it using the deleteDataWhenDone
flag.
The third option is neverDeleteData
, which means that Blitz++ will
not never deallocate the array data. This means it is your responsibility
to determine when the array data is no longer needed, and deallocate it.
You should use this option for memory which has not been allocated using the
C++ new
operator.
These constructors create array objects from pre-existing data:
Array(T_numtype* dataFirst, TinyVector<int, N_rank> shape, preexistingMemoryPolicy deletePolicy); Array(T_numtype* dataFirst, TinyVector<int, N_rank> shape, preexistingMemoryPolicy deletePolicy, GeneralArrayStorage<N_rank> storage); |
The first argument is a pointer to the array data. It should point to the
element of the array which is stored first in memory. The second argument
indicates the shape of the array. You can create this argument using the
shape()
function. For example:
double data[] = { 1, 2, 3, 4 }; Array<double,2> A(data, shape(2,2), neverDeleteData); // Make a 2x2 array |
The shape()
function takes N integer arguments and returns a
TinyVector<int,N>
.
By default, Blitz++ arrays are row-major. If you want to work with data which is stored in column-major order (e.g. a Fortran array), use the second version of the constructor:
Array<double,2> B(data, shape(2,2), neverDeleteData, FortranArray<2>()); |
This is a tad awkward, so Blitz++ provides the global object
fortranArray
which will convert to an instance of
GeneralArrayStorage<N_rank>
:
Array<double,2> B(data, shape(2,2), neverDeleteData, fortranArray); |
Another version of this constructor allows you to pass an arbitrary vector of strides:
Array(T_numtype* _bz_restrict dataFirst, TinyVector<int, N_rank> shape, TinyVector<int, N_rank> stride, preexistingMemoryPolicy deletePolicy, GeneralArrayStorage<N_rank> storage = GeneralArrayStorage<N_rank>()) |
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
For some platforms, it can be advantageous to store a set of arrays
interlaced together in memory. Blitz++ provides support for this through
the routines interlaceArrays()
and allocateArrays()
. An
example:
Array<int,2> A, B; interlaceArrays(shape(10,10), A, B); |
The first parameter of interlaceArrays()
is the shape for the arrays
(10x10). The subsequent arguments are the set of arrays to be interlaced
together. Up to 11 arrays may be interlaced. All arrays must store the
same data type and be of the same rank. In the above example, storage is
allocated so that A(0,0)
is followed immediately by B(0,0)
in
memory, which is folloed by A(0,1)
and B(0,1)
, and so on.
A related routine is allocateArrays()
, which has identical syntax:
Array<int,2> A, B; allocateArrays(shape(10,10), A, B); |
Unlike interlaceArrays()
, which always interlaces the arrays, the
routine allocateArrays()
may or may not interlace them, depending on
whether interlacing is considered advantageous for your platform. If the
tuning flag BZ_INTERLACE_ARRAYS
is defined in
<blitz/tuning.h>
, then the arrays are interlaced.
Note that the performance effects of interlacing are unpredictable: in some
situations it can be a benefit, and in most others it can slow your code
down substantially. You should only use interlaceArrays()
after
running some benchmarks to determine whether interlacing is beneficial for
your particular algorithm and architecture.
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
Blitz++ arrays use reference counting. When you create a new array, a
memory block is allocated. The Array
object acts like a handle for
this memory block. A memory block can be shared among multiple Array
objects -- for example, when you take subarrays and slices. The memory
block keeps track of how many Array
objects are referring to it.
When a memory block is orphaned -- when no Array
objects are
referring to it -- it automatically deletes itself and frees the allocated
memory.
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
This section describes how to access the elements of an array. There are three main ways:
Indexing, subarrays and slicing all use the overloaded parenthesis
operator()
.
As a running example, we'll consider the three dimensional array pictured below, which has index ranges (0..7, 0..7, 0..7). Shaded portions of the array show regions which have been obtained by indexing, creating a subarray, and slicing.
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
There are two ways to get a single element from an array. The simplest is
to provide a set of integer operands to operator()
:
A(7,0,0) = 5; cout << "A(7,0,0) = " << A(7,0,0) << endl; |
This version of indexing is available for arrays of rank one through eleven.
If the array object isn't const
, the return type of
operator()
is a reference; if the array object is const
, the
return type is a value.
You can also get an element by providing an operand of type
TinyVector<int,N_rank>
where N_rank
is the rank of the array
object:
TinyVector<int,3> index; index = 7, 0, 0; A(index) = 5; cout << "A(7,0,0) = " << A(index) << endl; |
This version of operator()
is also available in a const-overloaded
version.
It's possible to use fewer than N_rank
indices. However, missing
indices are assumed to be zero, which will cause bounds errors if
the valid index range does not include zero (e.g. Fortran arrays). For this
reason, and for code clarity, it's a bad idea to omit indices.
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
You can obtain a subarray by providing Range
operands to
operator()
. A Range
object represents a set of regularly
spaced index values. For example,
Array<int,3> B = A(Range(5,7), Range(5,7), Range(0,2)); |
The object B now refers to elements (5..7,5..7,0..2) of the array A.
The returned subarray is of type Array<T_numtype,N_rank>
. This means
that subarrays can be used wherever arrays can be: in expressions, as
lvalues, etc. Some examples:
// A three-dimensional stencil (used in solving PDEs) Range I(1,6), J(1,6), K(1,6); B = (A(I,J,K) + A(I+1,J,K) + A(I-1,J,K) + A(I,J+1,K) + A(I,J-1,K) + A(I,J+1,K) + A(I,J,K+1) + A(I,J,K-1)) / 7.0; // Set a subarray of A to zero A(Range(5,7), Range(5,7), Range(5,7)) = 0.; |
The bases of the subarray are equal to the bases of the original array:
Array<int,2> D(Range(1,5), Range(1,5)); // 1..5, 1..5 Array<int,2> E = D(Range(2,3), Range(2,3)); // 1..2, 1..2 |
An array can be used on both sides of an expression only if the subarrays don't overlap. If the arrays overlap, the result may depend on the order in which the array is traversed.
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
The classes RectDomain
and StridedDomain
, defined in
blitz/domain.h
, offer a dimension-independent notation for subarrays.
RectDomain
and StridedDomain
can be thought of as a
TinyVector<Range,N>
. Both have a vector of lower- and upper-bounds;
StridedDomain
has a stride vector. For example, the subarray:
Array<int,2> B = A(Range(4,7), Range(8,11)); // 4..7, 8..11 |
could be obtained using RectDomain
this way:
TinyVector<int,2> lowerBounds(4, 8); TinyVector<int,2> upperBounds(7, 11); RectDomain<2> subdomain(lowerBounds, upperBounds); Array<int,2> B = A(subdomain); |
Here are the prototypes of RectDomain
and StridedDomain
.
template<int N_rank> class RectDomain { public: RectDomain(const TinyVector<int,N_rank>& lbound, const TinyVector<int,N_rank>& ubound); const TinyVector<int,N_rank>& lbound() const; int lbound(int i) const; const TinyVector<int,N_rank>& ubound() const; int ubound(int i) const; Range operator[](int rank) const; void shrink(int amount); void shrink(int dim, int amount); void expand(int amount); void expand(int dim, int amount); }; template<int N_rank> class StridedDomain { public: StridedDomain(const TinyVector<int,N_rank>& lbound, const TinyVector<int,N_rank>& ubound, const TinyVector<int,N_rank>& stride); const TinyVector<int,N_rank>& lbound() const; int lbound(int i) const; const TinyVector<int,N_rank>& ubound() const; int ubound(int i) const; const TinyVector<int,N_rank>& stride() const; int stride(int i) const; Range operator[](int rank) const; void shrink(int amount); void shrink(int dim, int amount); void expand(int amount); void expand(int dim, int amount); }; |
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
A combination of integer and Range operands produces a slice. Each integer operand reduces the rank of the array by one. For example:
Array<int,2> F = A(Range::all(), 2, Range::all()); Array<int,1> G = A(2, 7, Range::all()); |
Range and integer operands can be used in any combination, for arrays up to rank 11.
Note: Using a combination of integer and Range operands requires a
newer language feature (partial ordering of member templates) which not all
compilers support. If your compiler does provide this feature,
BZ_PARTIAL_ORDERING
will be defined in <blitz/config.h>
. If
not, you can use this workaround:
Array<int,3> F = A(Range::all(), Range(2,2), Range::all()); Array<int,3> G = A(Range(2,2), Range(7,7), Range::all()); |
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
A Range
object represents an ordered set of uniformly spaced
integers. Here are some examples of using Range objects to obtain
subarrays:
#include <blitz/array.h> using namespace blitz; int main() { Array<int,1> A(7); A = 0, 1, 2, 3, 4, 5, 6; cout << A(Range::all()) << endl // [ 0 1 2 3 4 5 6 ] << A(Range(3,5)) << endl // [ 3 4 5 ] << A(Range(3,toEnd)) << endl // [ 3 4 5 6 ] << A(Range(fromStart,3)) << endl // [ 0 1 2 3 ] << A(Range(1,5,2)) << endl // [ 1 3 5 ] << A(Range(5,1,-2)) << endl // [ 5 3 1 ] << A(Range(fromStart,toEnd,2)) << endl; // [ 0 2 4 6 ] return 0; } |
The optional third constructor argument specifies a stride. For example,
Range(1,5,2)
refers to elements [1 3 5]. Strides can also be
negative: Range(5,1,-2)
refers to elements [5 3 1].
Note that if you use the same Range frequently, you can just construct one object and use it multiple times. For example:
Range all = Range::all(); A(0,all,all) = A(N-1,all,all); A(all,0,all) = A(all,N-1,all); A(all,all,0) = A(all,all,N-1); |
Here's an example of using strides with a two-dimensional array:
#include <blitz/array.h> using namespace blitz; int main() { Array<int,2> A(8,8); A = 0; Array<int,2> B = A(Range(1,7,3), Range(1,5,2)); B = 1; cout << "A = " << A << endl; return 0; } |
Here's an illustration of the B
subarray:
And the program output:
A = 8 x 8 [ 0 0 0 0 0 0 0 0 0 1 0 1 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 1 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 1 0 1 0 0 ] |
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
The assignment operator (=
) always results in the expression on the
right-hand side (rhs) being copied to the lhs (i.e. the data on the
lhs is overwritten with the result from the rhs). This is different from
some array packages in which the assignment operator makes the lhs a
reference (or alias) to the rhs. To further confuse the issue, the copy
constructor for arrays does have reference semantics. Here's an
example which should clarify things:
Array<int,1> A(5), B(10); A = B(Range(0,4)); // Statement 1 Array<int,1> C = B(Range(0,4)); // Statement 2 |
Statement 1 results in a portion of B
's data being copied into
A
. After Statement 1, both A
and B
have their own
(nonoverlapping) blocks of data. Contrast this behaviour with that of
Statement 2, which is not an assignment (it uses the copy
constructor). After Statement 2 is executed, the array C
is a
reference (or alias) to B
's data.
So to summarize: If you want to copy the rhs, use an assignment operator.
If you want to reference (or alias) the rhs, use the copy constructor (or
alternately, the reference()
member function in 2.6 Member functions).
Very important: whenever you have an assignment operator (=
,
+=
, -=
, etc.) the lhs must have the same shape as the
rhs. If you want the array on the left hand side to be resized to
the proper shape, you must do so by calling the resize
method, for
example:
A.resize(B.shape()); // Make A the same size as B A = B; |
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
#include <blitz/array.h> using namespace blitz; int main() { Array<int,2> A(6,6), B(3,3); // Set the upper left quadrant of A to 5 A(Range(0,2), Range(0,2)) = 5; // Set the upper right quadrant of A to an identity matrix B = 1, 0, 0, 0, 1, 0, 0, 0, 1; A(Range(0,2), Range(3,5)) = B; // Set the fourth row to 1 A(3, Range::all()) = 1; // Set the last two rows to 0 A(Range(4, Range::toEnd), Range::all()) = 0; // Set the bottom right element to 8 A(5,5) = 8; cout << "A = " << A << endl; return 0; } |
The output:
A = 6 x 6 [ 5 5 5 1 0 0 5 5 5 0 1 0 5 5 5 0 0 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 8 ] |
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
The Blitz++ library has a debugging mode which is enabled by defining the
preprocessor symbol BZ_DEBUG
. For most compilers, the command line
argument -DBZ_DEBUG
should work.
In debugging mode, your programs will run very slowly. This is because Blitz++ is doing lots of precondition checking and bounds checking. When it detects something fishy, it will likely halt your program and display an error message.
For example, this program attempts to access an element of a 4x4 array which doesn't exist:
#include <blitz/array.h> using namespace blitz; int main() { Array<complex<float>, 2> Z(4,4); Z = complex<float>(0.0, 1.0); Z(4,4) = complex<float>(1.0, 0.0); return 0; } |
When compiled with -DBZ_DEBUG
, the out of bounds indices are detected
and an error message results:
[Blitz++] Precondition failure: Module ../../blitz/array-impl.h line 1295 Array index out of range: (4, 4) Lower bounds: 2 [ 0 0 ] Length: 2 [ 4 4 ] debug: ../../blitz/array-impl.h:1295: bool blitz::Array<T, N>::assertInRange(int, int) const [with P_numtype = std::complex<float>, int N_rank = 2]: Assertion `0' failed. |
Precondition failures send their error messages to the standard error stream
(cerr
). After displaying the error message, assert(0)
is
invoked.
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
Several of the member functions take a dimension parameter which is
an integer in the range 0 .. N_rank
-1. For example, the method
extent(int n)
returns the extent (or length) of the array in
dimension n
.
These parameters are problematic:
reverse()
member function won't stand a chance of understanding what
A.reverse(2)
does.
N_rank
, rather
than 0 .. N_rank
-1. This makes dimension numbers inherently
error-prone. Even though I'm a experienced C/C++ programmer, I still
want to think of the first dimension as 1 -- it doesn't make sense to talk
about the "zeroth" dimension.
As a solution to this problem, Blitz++ provides a series of symbolic constants which you can use to refer to dimensions:
const int firstDim = 0; const int secondDim = 1; const int thirdDim = 2; . . const int eleventhDim = 10; |
These symbols should be used in place of the numerals 0, 1, ... N_rank
-1.
For example:
A.reverse(thirdDim); |
This code is clearer: you can see that the parameter refers to a dimension, and it isn't much of a leap to realize that it's reversing the element ordering in the third dimension.
If you find firstDim
, secondDim
, ... aesthetically unpleasing,
there are equivalent symbols firstRank
, secondRank
,
thirdRank
, ..., eleventhRank
.
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
The symbols had to stop somewhere, and eleven seemed an appropriate place to stop. Besides, if you're working in more than eleven dimensions your code is going to be confusing no matter what help Blitz++ provides.
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
const TinyVector<int, N_rank>& base() const; int base(int dimension) const; |
The base of a dimension is the first valid index value. A typical C-style array will have base of zero; a Fortran-style array will have base of one. The base can be different for each dimension, but only if you deliberately use a Range-argument constructor or design a custom storage ordering.
The first version returns a reference to the vector of base values.
The second version returns the base for just one dimension; it's
equivalent to the lbound()
member function. See the
note on dimension parameters such as firstDim
above.
Array<T,N>::iterator begin(); Array<T,N>::const_iterator begin() const; |
These functions return STL-style forward and input iterators, respectively,
positioned at the first element of the array. Note that the array data is
traversed in memory order (i.e. by rows for C-style arrays, and by columns
for Fortran-style arrays). The Array<T,N>::const_iterator
has these
methods:
const_iterator(const Array<T,N>&); T operator*() const; const T* [restrict] operator->() const; const_iterator& operator++(); void operator++(int); bool operator==(const const_iterator<T,N>&) const; bool operator!=(const const_iterator<T,N>&) const; const TinyVector<int,N>& position() const; |
Note that postfix ++ returns void (this is not STL-compliant, but is done
for efficiency). The method position()
returns a vector containing
current index positions of the iterator. The Array<T,N>::iterator
has the same methods as const_iterator
, with these exceptions:
iterator& operator++(); T& operator*(); T* [restrict] operator->();
The iterator
type may be used to modify array elements. To obtain
iterator positioned at the end of the array, use the end()
methods.
int cols() const; int columns() const; |
Both of these functions return the extent of the array in the
second dimension. Equivalent to extent(secondDim)
.
See also rows()
and depth()
.
Array<T_numtype, N_rank> copy() const; |
This method creates a copy of the array's data, using the same storage ordering as the current array. The returned array is guaranteed to be stored contiguously in memory, and to be the only object referring to its memory block (i.e. the data isn't shared with any other array object).
const T_numtype* [restrict] data() const; T_numtype* [restrict] data(); const T_numtype* [restrict] dataZero() const; T_numtype* [restrict] dataZero(); const T_numtype* [restrict] dataFirst() const; T_numtype* [restrict] dataFirst(); |
These member functions all return pointers to the array data. The NCEG
restrict
qualifier is used only if your compiler supports it. If
you're working with the default storage order (C-style arrays with base
zero), you'll only need to use data()
. Otherwise, things get
complicated:
data()
returns a pointer to the element whose indices are equal to
the array base. With a C-style array, this means the element (0,0,...,0);
with a Fortran-style array, this means the element (1,1,...,1). If A
is an array object, A.data()
is equivalent to (&A(A.base(firstDim),
A.base(secondDim), ...)). If any of the dimensions are stored in reverse
order, data()
will not refer to the element which comes first in
memory.
dataZero()
returns a pointer to the element (0,0,...,0), even if such
an element does not exist in the array. What's the point of having such a
pointer? Say you want to access the element (i,j,k). If you add to the
pointer the dot product of (i,j,k) with the stride vector
(A.stride()
), you get a pointer to the element (i,j,k).
dataFirst()
returns a pointer to the element of the array which comes
first in memory. Note however, that under some circumstances (e.g.
subarrays), the data will not be stored contiguously in memory. You have to
be very careful when meddling directly with an array's data.
Other relevant functions are: isStorageContiguous()
and
zeroOffset()
.
int depth() const; |
Returns the extent of the array in the third dimension. This function is
equivalent to extent(thirdDim)
. See also rows()
and
columns()
.
int dimensions() const; |
Returns the number of dimensions (rank) of the array. The return value is
the second template parameter (N_rank
) of the Array
object.
Same as rank()
.
RectDomain<N_rank> domain() const; |
Returns the domain of the array. The domain consists of a vector of lower
bounds and a vector of upper bounds for the indices. NEEDS_WORK-- need a
section to explain methods of RectDomain<N>
.
Array<T,N>::iterator end(); Array<T,N>::const_iterator end() const; |
Returns STL-style forward and input iterators (respectively) for the array, positioned at the end of the array.
int extent(int dimension) const; |
The first version the extent (length) of the array in the specified
dimension. See the note about dimension parameters such as firstDim
in the previous section.
Array<T_numtype2,N_rank> extractComponent(T_numtype2, int componentNumber, int numComponents); |
This method returns an array view of a single component of a multicomponent
array. In a multicomponent array, each element is a tuple of fixed size.
The components are numbered 0, 1, ..., numComponents-1
. Example:
Array<TinyVector<int,3>,2> A(128,128); // A 128x128 array of int[3] Array<int,2> B = A.extractComponent(int(), 1, 3); |
Now the B array refers to the 2nd component of every element in A. Note:
for complex arrays, special global functions real(A)
and
imag(A)
are provided to obtain real and imaginary components of an
array. See the Global Functions section.
void free(); |
This method resizes an array to zero size. If the array data is not being shared with another array object, then it is freed.
bool isMajorRank(int dimension) const; |
Returns true if the dimension has the largest stride. For C-style arrays
(the default), the first dimension always has the largest stride. For
Fortran-style arrays, the last dimension has the largest stride. See also
isMinorRank()
below and the note about dimension parameters such as
firstDim
in the previous section.
bool isMinorRank(int dimension) const; |
Returns true if the dimension does not have the largest stride. See
also isMajorRank()
.
bool isRankStoredAscending(int dimension) const; |
Returns true if the dimension is stored in ascending order in memory. This
is the default. It will only return false if you have reversed a dimension
using reverse()
or have created a custom storage order with a
descending dimension.
bool isStorageContiguous() const; |
Returns true if the array data is stored contiguously in memory. If you
slice the array or work on subarrays, there can be skips -- the array data
is interspersed with other data not part of the array. See also the various
data..()
functions. If you need to ensure that the storage is
contiguous, try reference(copy())
.
int lbound(int dimension) const; TinyVector<int,N_rank> lbound() const; |
The first version returns the lower bound of the valid index range for a
dimension. The second version returns a vector of lower bounds for all
dimensions. The lower bound is the first valid index value. If you're
using a C-style array (the default), the lbound will be zero; Fortran-style
arrays have lbound equal to one. The lbound can be different for each
dimension, but only if you deliberately set them that way using a Range
constructor or a custom storage ordering. This function is equivalent to
base(dimension)
. See the note about dimension parameters such as
firstDim
in the previous section.
void makeUnique(); |
If the array's data is being shared with another Blitz++ array object, this member function creates a copy so the array object has a unique view of the data.
int numElements() const; |
Returns the total number of elements in the array, calculated by taking the
product of the extent in each dimension. Same as size()
.
const TinyVector<int, N_rank>& ordering() const; int ordering(int storageRankIndex) const; |
These member functions return information about how the data is ordered in
memory. The first version returns the complete ordering vector; the second
version returns a single element from the ordering vector. The argument for
the second version must be in the range 0 .. N_rank
-1. The ordering
vector is a list of dimensions in increasing order of stride;
ordering(0)
will return the dimension number with the smallest
stride, and ordering(N_rank-1)
will return the dimension number with
largest stride. For a C-style array, the ordering vector contains the
elements (N_rank
-1, N_rank
-2, ..., 0). For a Fortran-style
array, the ordering vector is (0, 1, ..., N_rank
-1). See also the
description of custom storage orders in section 2.9 Array storage orders.
int rank() const; |
Returns the rank (number of dimensions) of the array. The return value is
equal to N_rank
. Equivalent to dimensions()
.
void reference(Array<T_numtype,N_rank>& A); |
This causes the array to adopt another array's data as its own. After this
member function is used, the array object and the array A
are
indistinguishable -- they have identical sizes, index ranges, and data. The
data is shared between the two arrays.
void reindexSelf(const TinyVector<int,N_rank>&); Array<T,N> reindex(const TinyVector<int,N_rank>&); |
These methods reindex an array to use a new base vector. The first version reindexes the array, and the second just returns a reindexed view of the array, leaving the original array unmodified.
void resize(int extent1, ...); void resize(const TinyVector<int,N_rank>&); |
These functions resize an array to the specified size. If the array is
already the size specified, then no memory is allocated. After resizing,
the contents of the array are garbage. See also resizeAndPreserve()
.
void resizeAndPreserve(int extent1, ...); void resizeAndPreserve(const TinyVector<int,N_rank>&); |
These functions resize an array to the specified size. If the array is already the size specified, then no change occurs (the array is not reallocated and copied). The contents of the array are preserved whenever possible; if the new array size is smaller, then some data will be lost. Any new elements created by resizing the array are left uninitialized.
Array<T,N> reverse(int dimension); void reverseSelf(int dimension); |
This method reverses the array in the specified dimension. For example, if
reverse(firstDim)
is invoked on a 2-dimensional array, then the
ordering of rows in the array will be reversed; reverse(secondDim)
would reverse the order of the columns. Note that this is implemented by
twiddling the strides of the array, and doesn't cause any data copying. The
first version returns a reversed "view" of the array data; the second
version applies the reversal to the array itself.
int rows() const; |
Returns the extent (length) of the array in the first dimension. This
function is equivalent to extent(firstDim)
. See also
columns()
, and depth()
.
int size() const; |
Returns the total number of elements in the array, calculated by taking the
product of the extent in each dimension. Same as numElements()
.
const TinyVector<int, N_rank>& shape() const; |
Returns the vector of extents (lengths) of the array.
const TinyVector<int, N_rank>& stride() const; int stride(int dimension) const; |
The first version returns the stride vector; the second version returns the
stride associated with a dimension. A stride is the distance between
pointers to two array elements which are adjacent in a dimension. For
example, A.stride(firstDim)
is equal to &A(1,0,0) - &A(0,0,0)
.
The stride for the second dimension, A.stride(secondDim)
, is equal to
&A(0,1,0) - &A(0,0,0)
, and so on. For more information about
strides, see the description of custom storage formats in Section
2.9 Array storage orders. See also the description of parameters like
firstDim
and secondDim
in the previous section.
Array<T,N> transpose(int dimension1, int dimension2, ...); void transposeSelf(int dimension1, int dimension2, ...); |
These methods permute the dimensions of the array. The dimensions of the
array are reordered so that the first dimension is dimension1
, the
second is dimension2
, and so on. The arguments should be a
permutation of the symbols firstDim, secondDim, ...
. Note that this
is implemented by twiddling the strides of the array, and doesn't cause any
data copying. The first version returns a transposed "view" of the array
data; the second version transposes the array itself.
int ubound(int dimension) const; TinyVector<int,N_rank> ubound() const; |
The first version returns the upper bound of the valid index range for a
dimension. The second version returns a vector of upper bounds for all
dimensions. The upper bound is the last valid index value. If you're using
a C-style array (the default), the ubound will be equal to the
extent(dimension)-1
. Fortran-style arrays will have ubound equal to
extent(dimension)
. The ubound can be different for each dimension.
The return value of ubound(dimension)
will always be equal to
lbound(dimension)+extent(dimension)-1
. See the note about
dimension parameters such as firstDim
in the previous section.
int zeroOffset() const; |
This function has to do with the storage of arrays in memory. You may want
to refer to the description of the data..()
member functions and of
custom storage orders in Section 2.9 Array storage orders for
clarification. The return value of zeroOffset()
is the distance from
the first element in the array to the (possibly nonexistant) element
(0,0,...,0)
. In this context, "first element" returns to the element
(base(firstDim),base(secondDim),...)
.
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
void allocateArrays(TinyVector<int,N>& shape, Array<T,N>& A, Array<T,N>& B, ...); |
This function will allocate interlaced arrays, but only if interlacing is
desirable for your architecture. This is controlled by the
BZ_INTERLACE_ARRAYS
flag in `blitz/tuning.h'. You can provide up to
11 arrays as parameters. Any views currently associated with the array
objects are lost. Here is a typical use:
Array<int,2> A, B, C; allocateArrays(shape(64,64),A,B,C); |
If array interlacing is enabled, then the arrays are stored in memory like
this: A(0,0)
, B(0,0)
, C(0,0)
, A(0,1)
,
B(0,1)
, ... If interlacing is disabled, then the arrays are
allocated in the normal fashion: each array has its own block of memory.
Once interlaced arrays are allocated, they can be used just like regular
arrays.
#include <blitz/array/convolve.h> Array<T,1> convolve(const Array<T,1>& B, const Array<T,1>& C); |
This function computes the 1-D convolution of the arrays B and C: A[i] = sum(B[j] * C[i-j], j) If the array B has domain b_l \ldots b_h, and array C has domain c_l \ldots c_h, then the resulting array has domain a_l \ldots a_h, with l = b_l + c_l and a_h = b_h + c_h.
A new array is allocated to contain the result. To avoid copying the result
array, you should use it as a constructor argument. For example:
Array<float,1> A = convolve(B,C);
The convolution is computed in the
spatial domain. Frequency-domain transforms are not used. If you are
convolving two large arrays, then this will be slower than using a Fourier
transform.
Note that if you need a cross-correlation, you can use the convolve function with one of the arrays reversed. For example:
Array<float,1> A = convolve(B,C.reverse()); |
Autocorrelation can be performed using the same approach.
void cycleArrays(Array<T,N>& A, Array<T,N>& B); void cycleArrays(Array<T,N>& A, Array<T,N>& B, Array<T,N>& C); void cycleArrays(Array<T,N>& A, Array<T,N>& B, Array<T,N>& C, Array<T,N>& D); void cycleArrays(Array<T,N>& A, Array<T,N>& B, Array<T,N>& C, Array<T,N>& D, Array<T,N>& E); |
These routines are useful for time-stepping PDEs. They take a set of arrays
such as [A,B,C,D
] and cyclically rotate them to [B,C,D,A
];
i.e. the A
array then refers to what was B
's data, the
B
array refers to what was C
's data, and the D
array
refers to what was A
's data. These functions operate in constant
time, since only the handles change (i.e. no data is copied; only pointers
change).
Array<T,N> imag(Array<complex<T>,N>&); |
This method returns a view of the imaginary portion of the array.
void interlaceArrays(TinyVector<int,N>& shape, Array<T,N>& A, Array<T,N>& B, ...); |
This function is similar to allocateArrays()
above, except that the
arrays are always interlaced, regardless of the setting of the
BZ_INTERLACE_ARRAYS
flag.
Array<T,N> real(Array<complex<T>,N>&); |
This method returns a view of the real portion of the array.
TinyVector<int,1> shape(int L); TinyVector<int,2> shape(int L, int M); TinyVector<int,3> shape(int L, int M, int N); TinyVector<int,4> shape(int L, int M, int N, int O); ... [up to 11 dimensions] |
These functions may be used to create shape parameters. They package the
set of integer arguments as a TinyVector
of appropriate length. For
an example use, see allocateArrays()
above.
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
The current version of Blitz++ includes rudimentary output formatting for arrays. Here's an example:
#include <blitz/array.h> using namespace blitz; int main() { Array<int,2> A(4,5,FortranArray<2>()); firstIndex i; secondIndex j; A = 10*i + j; cout << "A = " << A << endl; Array<float,1> B(20); B = exp(-i/100.); cout << "B = " << endl << B << endl; return 0; } |
And the output:
A = 4 x 5 [ 11 12 13 14 15 21 22 23 24 25 31 32 33 34 35 41 42 43 44 45 ] B = 20 [ 1 0.99005 0.980199 0.970446 0.960789 0.951229 0.941765 0.932394 0.923116 0.913931 0.904837 0.895834 0.88692 0.878095 0.869358 0.860708 0.852144 0.843665 0.83527 0.826959 ] |
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
Arrays may be restored from an istream using the >>
operator.
Note: you must know the dimensionality of the array being restored
from the stream. The >>
operator expects an array in the same input
format as generated by the <<
operator, namely:
'['
indicating the start of the array data
']'
indicating the end of the array data
The operator prototype is:
template<class T, int N> istream& operator>>(istream&, Array<T,N>&); |
Here is an example of saving and restoring arrays from files. You can find this example in the Blitz++ distribution as `examples/io.cpp'.
#include <blitz/array.h> #ifdef BZ_HAVE_STD #include <fstream> #else #include <fstream.h> #endif BZ_USING_NAMESPACE(blitz) const char* filename = "io.data"; void write_arrays() { ofstream ofs(filename); if (ofs.bad()) { cerr << "Unable to write to file: " << filename << endl; exit(1); } Array<float,3> A(3,4,5); A = 111 + tensor::i + 10 * tensor::j + 100 * tensor::k; ofs << A << endl; Array<float,2> B(3,4); B = 11 + tensor::i + 10 * tensor::j; ofs << B << endl; Array<float,1> C(4); C = 1 + tensor::i; ofs << C << endl; } int main() { write_arrays(); ifstream ifs(filename); if (ifs.bad()) { cerr << "Unable to open file: " << filename << endl; exit(1); } Array<float,3> A; Array<float,2> B; Array<float,1> C; ifs >> A >> B >> C; cout << "Arrays restored from file: " << A << B << C << endl; return 0; } |
Note: The storage order and starting indices are not restored from the input stream. If you are restoring (for example) a Fortran-style array, you must create a Fortran-style array, and then restore it. For example, this code restores a Fortran-style array from the standard input stream:
Array<float,2> B(fortranArray); cin >> B; |
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
Blitz++ is very flexible about the way arrays are stored in memory. Starting indices can be 0, 1, or arbitrary numbers; arrays can be stored in row major, column major or an order based on any permutation of the dimensions; each dimension can be stored in either ascending or descending order. An N dimensional array can be stored in N! 2^N possible ways.
Before getting into the messy details, a review of array storage formats is useful. If you're already familiar with strides and bases, you might want to skip on to the next section.
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
Suppose we want to store this two-dimensional array in memory:
[ 1 2 3 ] [ 4 5 6 ] [ 7 8 9 ] |
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
To lay the array out in memory, it's necessary to map the indices (i,j) into a one-dimensional block. Here are two ways the array might appear in memory:
[ 1 2 3 4 5 6 7 8 9 ] [ 1 4 7 2 5 8 3 6 9 ] |
The first order corresponds to a C or C++ style array, and is called row-major ordering: the data is stored first by row, and then by column. The second order corresponds to a Fortran style array, and is called column-major ordering: the data is stored first by column, and then by row.
The simplest way of mapping the indices (i,j) into one-dimensional memory is to take a linear combination.(2) Here's the appropriate linear combination for row major ordering:
memory offset = 3*i + 1*j |
And for column major ordering:
memory offset = 1*i + 3*j |
The coefficients of the (i,j) indices are called strides. For a row major storage of this array, the row stride is 3 -- you have to skip three memory locations to move down a row. The column stride is 1 -- you move one memory location to move to the next column. This is also known as unit stride. For column major ordering, the row and column strides are 1 and 3, respectively.
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
To throw another complication into this scheme, C-style arrays have indices which start at zero, and Fortran-style arrays have indices which start at one. The first valid index value is called the base. To account for a non-zero base, it's necessary to include an offset term in addition to the linear combination. Here's the mapping for a C-style array with i=0..3 and j=0..3:
memory offset = 0 + 3*i + 1*j |
No offset is necessary since the indices start at zero for C-style arrays. For a Fortran-style array with i=1..4 and j=1..4, the mapping would be:
memory offset = -4 + 3*i + 1*j |
By default, Blitz++ creates arrays in the C-style storage format (base zero, row major ordering). To create a Fortran-style array, you can use this syntax:
Array<int,2> A(3, 3, FortranArray<2>()); |
The third parameter, FortranArray<2>()
, tells the Array
constructor to use a storage format appropriate for two-dimensional Fortran
arrays (base one, column major ordering).
A similar object, ColumnMajor<N>
, tells the Array
constructor
to use column major ordering, with base zero:
Array<int,2> B(3, 3, ColumnMajor<2>()); |
This creates a 3x3 array with indices i=0..2 and j=0..2.
In addition to supporting the 0 and 1 conventions for C and Fortran-style arrays, Blitz++ allows you to choose arbitrary bases, possibly different for each dimension. For example, this declaration creates an array whose indices have ranges i=5..8 and j=2..5:
Array<int,2> A(Range(5,8), Range(2,5)); |
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
All Array
constructors take an optional parameter of type
GeneralArrayStorage<N_rank>
. This parameter encapsulates a complete
description of the storage format. If you want a storage format other than
C or Fortran-style, you have two choices:
GeneralArrayStorage<N_rank>
, customize the storage format, and use
the object as a argument for the Array
constructor.
GeneralArrayStorage<N_rank>
. This is useful if you will be using the
storage format many times. This approach (inheriting from
GeneralArrayStorage<N_rank>
) was used to create the
FortranArray<N_rank>
objects. If you want to take this approach, you
can use the declaration of FortranArray<N_rank>
in
<blitz/array.h>
as a guide.
The next sections describe how to modify a
GeneralArrayStorage<N_rank>
object to suit your needs.
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
In more than two dimensions, the choice of storage order becomes more complicated. Suppose we had a 3x3x3 array. To map the indices (i,j,k) into memory, we might choose one of these mappings:
memory offset = 9*i + 3*j + 1*k memory offset = 1*i + 3*j + 9*k |
The first corresponds to a C-style array, and the second to a Fortran-style array. But there are other choices; we can permute the strides (1,3,9) any which way:
memory offset = 1*i + 9*j + 3*k memory offset = 3*i + 1*j + 9*k memory offset = 3*i + 9*j + 1*k memory offset = 9*i + 1*j + 3*k |
For an N dimensional array, there are N! such permutations. Blitz++ allows
you to select any permutation of the dimensions as a storage order. First
you need to create an object of type GeneralArrayStorage<N_rank>
:
GeneralArrayStorage<3> storage; |
GeneralArrayStorage<N_rank>
contains a vector called ordering
which controls the order in which dimensions are stored in memory. The
ordering
vector will contain a permutation of the numbers 0, 1, ...,
N_rank-1. Since some people are used to the first dimension being 1 rather
than 0, a set of symbols (firstDim, secondDim, ..., eleventhDim) are
provided which make the code more legible.
The ordering
vector lists the dimensions in increasing order of
stride. You can access this vector using the member function
ordering()
. A C-style array, the default, would have:
storage.ordering() = thirdDim, secondDim, firstDim; |
meaning that the third index (k) is associated with the smallest stride, and the first index (i) is associated with the largest stride. A Fortran-style array would have:
storage.ordering() = firstDim, secondDim, thirdDim; |
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
To add yet another wrinkle, there are some applications where the rows or columns need to be stored in reverse order.(3)
Blitz++ allows you to store each dimension in either ascending or descending
order. By default, arrays are always stored in ascending order. The
GeneralArrayStorage<N_rank>
object contains a vector called
ascendingFlag
which indicates whether each dimension is stored
ascending (true
) or descending (false
). To alter the contents
of this vector, use the ascendingFlag()
method:
// Store the third dimension in descending order storage.ascendingFlag() = true, true, false; // Store all the dimensions in descending order storage.ascendingFlag() = false, false, false; |
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
GeneralArrayStorage<N_rank>
also has a base
vector which
contains the base index value for each dimension. By default, the base
vector is set to zero. FortranArray<N_rank>
sets the base vector to
one.
To set your own set of bases, you have two choices:
base
vector inside the
GeneralArrayStorage<N_rank>
object. The method base()
returns
a mutable reference to the base
vector which you can use to set the
bases.
Range
arguments to the
Array
constructor.
Here are some examples of the first approach:
// Set all bases equal to 5 storage.base() = 5; // Set the bases to [ 1 0 1 ] storage.base() = 1, 0, 1; |
And of the second approach:
// Have bases of 5, but otherwise C-style storage Array<int,3> A(Range(5,7), Range(5,7), Range(5,7)); // Have bases of [ 1 0 1 ] and use a custom storage Array<int,3> B(Range(1,4), Range(0,3), Range(1,4), storage); |
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
Once you have created an array object, you will probably never have to worry about its storage order. Blitz++ should handle arrays of different storage orders transparently. It's possible to mix arrays of different storage orders in one expression, and still get the correct result.
Note however, that mixing different storage orders in an expression may incur a performance penalty, since Blitz++ will have to pay more attention to differences in indexing than it normally would.
You may not mix arrays with different domains in the same expression. For example, adding a base zero to a base one array is a no-no. The reason for this restriction is that certain expressions become ambiguous, for example:
Array<int,1> A(Range(0,5)), B(Range(1,6)); A=0; B=0; using namespace blitz::tensor; int result = sum(A+B+i); |
Should the index i
take its domain from array A
or array
B
? To avoid such ambiguities, users are forbidden from mixing arrays
with different domains in an expression.
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
-DBZ_DEBUG
), class Array
provides a member
function dumpStructureInformation()
which displays information about
the array storage:
Array<float,4> A(3,7,8,2,FortranArray<4>()); A.dumpStructureInformation(cerr); |
The optional argument is an ostream
to dump information to. It
defaults to cout
. Here's the output:
Dump of Array<f, 4>: ordering_ = 4 [ 0 1 2 3 ] ascendingFlag_ = 4 [ 1 1 1 1 ] base_ = 4 [ 1 1 1 1 ] length_ = 4 [ 3 7 8 2 ] stride_ = 4 [ 1 3 21 168 ] zeroOffset_ = -193 numElements() = 336 isStorageContiguous() = 1 |
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
When initializing arrays with comma delimited lists, note that the array is filled in storage order: from the first memory location to the last memory location. This won't cause any problems if you stick with C-style arrays, but it can be confusing for Fortran-style arrays:
Array<int,2> A(3, 3, FortranArray<2>()); A = 1, 2, 3, 4, 5, 6, 7, 8, 9; cout << A << endl; |
The output from this code excerpt will be:
A = 3 x 3 1 4 7 2 5 8 3 6 9 |
This is because Fortran-style arrays are stored in column major order.
[ < ] | [ > ] | [ << ] | [ Up ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |
#include <blitz/array.h> BZ_USING_NAMESPACE(blitz) int main() { // 3x3 C-style row major storage, base zero Array<int,2> A(3, 3); // 3x3 column major storage, base zero Array<int,2> B(3, 3, ColumnMajorArray<2>()); // A custom storage format: // Indices have range 0..3, 0..3 // Column major ordering // Rows are stored ascending, columns stored descending GeneralArrayStorage<2> storage; storage.ordering() = firstRank, secondRank; storage.base() = 0, 0; storage.ascendingFlag() = true, false; Array<int,2> C(3, 3, storage); // Set each array equal to // [ 1 2 3 ] // [ 4 5 6 ] // [ 7 8 9 ] A = 1, 2, 3, 4, 5, 6, 7, 8, 9; cout << "A = " << A << endl; // Comma-delimited lists initialize in memory-storage order only. // Hence we list the values in column-major order to initialize B: B = 1, 4, 7, 2, 5, 8, 3, 6, 9; cout << "B = " << B << endl; // Array C is stored in column major, plus the columns are stored // in descending order! C = 3, 6, 9, 2, 5, 8, 1, 4, 7; cout << "C = " << C << endl; Array<int,2> D(3,3); D = A + B + C; #ifdef BZ_DEBUG A.dumpStructureInformation(); B.dumpStructureInformation(); C.dumpStructureInformation(); D.dumpStructureInformation(); #endif cout << "D = " << D << endl; return 0; } |
And the output:
A = 3 x 3 [ 1 2 3 4 5 6 7 8 9 ] B = 3 x 3 [ 1 2 3 4 5 6 7 8 9 ] C = 3 x 3 [ 1 2 3 4 5 6 7 8 9 ] Dump of Array<i, 2>: ordering_ = 2 [ 1 0 ] ascendingFlag_ = 2 [ 1 1 ] base_ = 2 [ 0 0 ] length_ = 2 [ 3 3 ] stride_ = 2 [ 3 1 ] zeroOffset_ = 0 numElements() = 9 isStorageContiguous() = 1 Dump of Array<i, 2>: ordering_ = 2 [ 0 1 ] ascendingFlag_ = 2 [ 1 1 ] base_ = 2 [ 0 0 ] length_ = 2 [ 3 3 ] stride_ = 2 [ 1 3 ] zeroOffset_ = 0 numElements() = 9 isStorageContiguous() = 1 Dump of Array<i, 2>: ordering_ = 2 [ 0 1 ] ascendingFlag_ = 2 [ 1 0 ] base_ = 2 [ 0 0 ] length_ = 2 [ 3 3 ] stride_ = 2 [ 1 -3 ] zeroOffset_ = 6 numElements() = 9 isStorageContiguous() = 1 Dump of Array<i, 2>: ordering_ = 2 [ 1 0 ] ascendingFlag_ = 2 [ 1 1 ] base_ = 2 [ 0 0 ] length_ = 2 [ 3 3 ] stride_ = 2 [ 3 1 ] zeroOffset_ = 0 numElements() = 9 isStorageContiguous() = 1 D = 3 x 3 [ 3 6 9 12 15 18 21 24 27 ] |
[ << ] | [ >> ] | [Top] | [Contents] | [Index] | [ ? ] |