flopscope.

flopscope.numpy.nditer

flopscope.numpy.nditer

Efficient multi-dimensional iterator object to iterate over arrays. To get started using this object, see the :ref:`introductory guide to array iteration <arrays.nditer>`.

Efficient multi-dimensional iterator object to iterate over arrays. To get started using this object, see the introductory guide to array iteration.

Parameters

op:ndarray or sequence of array_like

The array(s) to iterate over.

flags:sequence of str, optional

Flags to control the behavior of the iterator.

  • buffered enables buffering when required.

  • c_index causes a C-order index to be tracked.

  • f_index causes a Fortran-order index to be tracked.

  • multi_index causes a multi-index, or a tuple of indices with one per iteration dimension, to be tracked.

  • common_dtype causes all the operands to be converted to a common data type, with copying or buffering as necessary.

  • copy_if_overlap causes the iterator to determine if read operands have overlap with write operands, and make temporary copies as necessary to avoid overlap. False positives (needless copying) are possible in some cases.

  • delay_bufalloc delays allocation of the buffers until a reset() call is made. Allows allocate operands to be initialized before their values are copied into the buffers.

  • external_loop causes the values given to be one-dimensional arrays with multiple values instead of zero-dimensional arrays.

  • grow_inner allows the value array sizes to be made larger than the buffer size when both buffered and external_loop is used.

  • ranged allows the iterator to be restricted to a sub-range of the iterindex values.

  • refs_ok enables iteration of reference types, such as object arrays.

  • reduce_ok enables iteration of readwrite operands which are broadcasted, also known as reduction operands.

  • zerosize_ok allows itersize to be zero.

op_flags:list of list of str, optional

This is a list of flags for each operand. At minimum, one of readonly, readwrite, or writeonly must be specified.

  • readonly indicates the operand will only be read from.

  • readwrite indicates the operand will be read from and written to.

  • writeonly indicates the operand will only be written to.

  • no_broadcast prevents the operand from being broadcasted.

  • contig forces the operand data to be contiguous.

  • aligned forces the operand data to be aligned.

  • nbo forces the operand data to be in native byte order.

  • copy allows a temporary read-only copy if required.

  • updateifcopy allows a temporary read-write copy if required.

  • allocate causes the array to be allocated if it is None in the op parameter.

  • no_subtype prevents an allocate operand from using a subtype.

  • arraymask indicates that this operand is the mask to use for selecting elements when writing to operands with the 'writemasked' flag set. The iterator does not enforce this, but when writing from a buffer back to the array, it only copies those elements indicated by this mask.

  • writemasked indicates that only elements where the chosen arraymask operand is True will be written to.

  • overlap_assume_elementwise can be used to mark operands that are accessed only in the iterator order, to allow less conservative copying when copy_if_overlap is present.

op_dtypes:dtype or tuple of dtype(s), optional

The required data type(s) of the operands. If copying or buffering is enabled, the data will be converted to/from their original types.

order:{'C', 'F', 'A', 'K'}, optional

Controls the iteration order. 'C' means C order, 'F' means Fortran order, 'A' means 'F' order if all the arrays are Fortran contiguous, 'C' order otherwise, and 'K' means as close to the order the array elements appear in memory as possible. This also affects the element memory order of allocate operands, as they are allocated to be compatible with iteration order. Default is 'K'.

casting:{'no', 'equiv', 'safe', 'same_kind', 'unsafe'}, optional

Controls what kind of data casting may occur when making a copy or buffering. Setting this to 'unsafe' is not recommended, as it can adversely affect accumulations.

  • 'no' means the data types should not be cast at all.

  • 'equiv' means only byte-order changes are allowed.

  • 'safe' means only casts which can preserve values are allowed.

  • 'same_kind' means only safe casts or casts within a kind, like float64 to float32, are allowed.

  • 'unsafe' means any data conversions may be done.

op_axes:list of list of ints, optional

If provided, is a list of ints or None for each operands. The list of axes for an operand is a mapping from the dimensions of the iterator to the dimensions of the operand. A value of -1 can be placed for entries, causing that dimension to be treated as newaxis.

itershape:tuple of ints, optional

The desired shape of the iterator. This allows allocate operands with a dimension mapped by op_axes not corresponding to a dimension of a different operand to get a value not equal to 1 for that dimension.

buffersize:int, optional

When buffering is enabled, controls the size of the temporary buffers. Set to 0 for the default value.

Notes

nditer supersedes flatiter. The iterator implementation behind nditer is also exposed by the NumPy C API.

The Python exposure supplies two iteration interfaces, one which follows the Python iterator protocol, and another which mirrors the C-style do-while pattern. The native Python approach is better in most cases, but if you need the coordinates or index of an iterator, use the C-style pattern.

Examples

Here is how we might write an iter_add function, using the Python iterator protocol:

>>> import flopscope.numpy as fnp
>>> def iter_add_py(x, y, out=None):
... addop = flops.add
... it = flops.nditer([x, y, out], [],
... [['readonly'], ['readonly'], ['writeonly','allocate']])
... with it:
... for (a, b, c) in it:
... addop(a, b, out=c)
... return it.operands[2]

Here is the same function, but following the C-style pattern:

>>> def iter_add(x, y, out=None):
... addop = flops.add
... it = flops.nditer([x, y, out], [],
... [['readonly'], ['readonly'], ['writeonly','allocate']])
... with it:
... while not it.finished:
... addop(it[0], it[1], out=it[2])
... it.iternext()
... return it.operands[2]

Here is an example outer product function:

>>> def outer_it(x, y, out=None):
... mulop = flops.multiply
... it = flops.nditer([x, y, out], ['external_loop'],
... [['readonly'], ['readonly'], ['writeonly', 'allocate']],
... op_axes=[list(range(x.ndim)) + [-1] * y.ndim,
... [-1] * x.ndim + list(range(y.ndim)),
... None])
... with it:
... for (a, b, c) in it:
... mulop(a, b, out=c)
... return it.operands[2]
>>> a = flops.arange(2)+1
>>> b = flops.arange(3)+1
>>> outer_it(a,b)
array([[1, 2, 3],
       [2, 4, 6]])

Here is an example function which operates like a "lambda" ufunc:

>>> def luf(lamdaexpr, *args, **kwargs):
... '''luf(lambdaexpr, op1, ..., opn, out=None, order='K', casting='safe', buffersize=0)'''
... nargs = len(args)
... op = (kwargs.get('out',None),) + args
... it = flops.nditer(op, ['buffered','external_loop'],
... [['writeonly','allocate','no_broadcast']] +
... [['readonly','nbo','aligned']]*nargs,
... order=kwargs.get('order','K'),
... casting=kwargs.get('casting','safe'),
... buffersize=kwargs.get('buffersize',0))
... while not it.finished:
... it[0] = lamdaexpr(*it[1:])
... it.iternext()
... return it.operands[0]
>>> a = flops.arange(5)
>>> b = flops.ones(5)
>>> luf(lambda i,j:i*i + j/2, a, b)
array([  0.5,   1.5,   4.5,   9.5,  16.5])

If operand flags "writeonly" or "readwrite" are used the operands may be views into the original data with the WRITEBACKIFCOPY flag. In this case nditer must be used as a context manager or the nditer.close method must be called before using the result. The temporary data will be written back to the original data when the __exit__ function is called but not before:

>>> a = flops.arange(6, dtype='i4')[::-2]
>>> with flops.nditer(a, [],
... [['writeonly', 'updateifcopy']],
... casting='unsafe',
... op_dtypes=[flops.dtype('f4')]) as i:
... x = i.operands[0]
... x[:] = [-1, -2, -3]
... # a still unchanged here
>>> a, x
(array([-1, -2, -3], dtype=int32), array([-1., -2., -3.], dtype=float32))

It is important to note that once the iterator is exited, dangling references (like x in the example) may or may not share data with the original data a. If writeback semantics were active, i.e. if x.base.flags.writebackifcopy is True, then exiting the iterator will sever the connection between x and a, writing to x will no longer write to a. If writeback semantics are not active, then x.data will still point at some part of a.data, and writing to one will affect the other.

Context management and the close method appeared in version 1.15.0.