def wrapper(a, axis=None, dtype=None, keepdims=False):
if not isinstance(axis, tuple):
axis = (axis,)
if axis[0] is None:
# Special case for speed
a_flat = a.flat
s = func(a_flat, a_flat.d_array, pu.c2f(a_flat.shape, 0))
if keepdims:
shape = (1,)*a.ndim
else:
shape = ()
ret = afnumpy.ndarray(tuple(shape), dtype=pu.typemap(s.dtype()),
af_array=s)
else:
shape = list(a.shape)
s = a.d_array
# Do in decreasing axis order to avoid problems with the pop
for ax in sorted(axis)[::-1]:
# Optimization
if(a.shape[ax] > 1):
s = func(a, s, pu.c2f(a.shape, ax))
if keepdims:
shape[ax] = 1
else:
shape.pop(ax)
ret = afnumpy.ndarray(tuple(shape), dtype=pu.typemap(s.dtype()),
af_array=s)
if(dtype is not None):
ret = ret.astype(dtype)
if(len(shape) == 0):
ret = ret[()]
return ret
def wrapper(a, axis=None, dtype=None, keepdims=False):
if not isinstance(axis, tuple):
axis = (axis, )
if axis[0] is None:
# Special case for speed
a_flat = a.flat
s = func(a_flat, a_flat.d_array, pu.c2f(a_flat.shape, 0))
if keepdims:
shape = (1, ) * a.ndim
else:
shape = ()
ret = afnumpy.ndarray(tuple(shape),
dtype=pu.typemap(s.dtype()),
af_array=s)
else:
shape = list(a.shape)
s = a.d_array
# Do in decreasing axis order to avoid problems with the pop
for ax in sorted(axis)[::-1]:
# Optimization
if (a.shape[ax] > 1):
s = func(a, s, pu.c2f(a.shape, ax))
if keepdims:
shape[ax] = 1
else:
shape.pop(ax)
ret = afnumpy.ndarray(pu.af_shape(s),
dtype=pu.typemap(s.dtype()),
af_array=s).reshape(tuple(shape))
if (dtype is not None):
ret = ret.astype(dtype)
if (len(shape) == 0):
ret = ret[()]
return ret
def dot(a, b):
# Arrayfire requires that the types match for dot and matmul
res_dtype = numpy.result_type(a, b)
a = a.astype(res_dtype, copy=False)
b = b.astype(res_dtype, copy=False)
if a.ndim == 1 and b.ndim == 1:
s = arrayfire.dot(a.d_array, b.d_array)
return afnumpy.ndarray((), dtype=a.dtype, af_array=s)[()]
a_shape = a.shape
b_shape = b.shape
if a.ndim == 1:
a = a.reshape((a.shape[0], 1))
if b.ndim == 1:
b = b.reshape((b.shape[0], 1))
if a.ndim == 2 and b.ndim == 2:
# Notice the order of the arguments to matmul. It's not a bug!
s = arrayfire.matmul(b.d_array, a.d_array)
return afnumpy.ndarray(pu.af_shape(s),
dtype=pu.typemap(s.dtype()),
af_array=s)
# Multidimensional dot is done with loops
# Calculate the shape of the result array
a_shape = list(a_shape)
a_shape.pop(-1)
b_shape = list(b_shape)
b_shape.pop(-2)
res_shape = a_shape + b_shape
# Make sure the arrays are at least 3D
if a.ndim < 3:
a = a.reshape((1, ) + a.shape)
if b.ndim < 3:
b = b.reshape((1, ) + b.shape)
# We're going to flatten the regions over which we're going to loop over
# to make our life easier and reduce the amount of indexing code
a = a.reshape((-1, a.shape[-2], a.shape[-1]))
b = b.reshape((-1, b.shape[-2], b.shape[-1]))
# Initialize the output array. The shape matches the reshape a and b.
res = afnumpy.empty((a.shape[0], a.shape[-2], b.shape[0], b.shape[-1]),
dtype=a.dtype)
# Loop through the flattened indices and calculate the matmuls
for i in range(0, a.shape[0]):
for j in range(0, b.shape[0]):
res[i, :, j, :] = afnumpy.dot(a[i], b[j])
# Finally appropriately reshape the result array
return res.reshape(res_shape)
def dot(a, b):
# Arrayfire requires that the types match for dot and matmul
res_dtype = numpy.result_type(a,b)
a = a.astype(res_dtype, copy=False)
b = b.astype(res_dtype, copy=False)
if a.ndim == 1 and b.ndim == 1:
s = arrayfire.dot(a.d_array, b.d_array)
return afnumpy.ndarray((), dtype=a.dtype, af_array=s)[()]
a_shape = a.shape
b_shape = b.shape
if a.ndim == 1:
a = a.reshape((a.shape[0],1))
if b.ndim == 1:
b = b.reshape((b.shape[0],1))
if a.ndim == 2 and b.ndim == 2:
# Notice the order of the arguments to matmul. It's not a bug!
s = arrayfire.matmul(b.d_array, a.d_array)
return afnumpy.ndarray(pu.af_shape(s), dtype=pu.typemap(s.dtype()),
af_array=s)
# Multidimensional dot is done with loops
# Calculate the shape of the result array
a_shape = list(a_shape)
a_shape.pop(-1)
b_shape = list(b_shape)
b_shape.pop(-2)
res_shape = a_shape + b_shape
# Make sure the arrays are at least 3D
if a.ndim < 3:
a = a.reshape((1,)+a.shape)
if b.ndim < 3:
b = b.reshape((1,)+b.shape)
# We're going to flatten the regions over which we're going to loop over
# to make our life easier and reduce the amount of indexing code
a = a.reshape((-1,a.shape[-2],a.shape[-1]))
b = b.reshape((-1,b.shape[-2],b.shape[-1]))
# Initialize the output array. The shape matches the reshape a and b.
res = afnumpy.empty((a.shape[0], a.shape[-2], b.shape[0],
b.shape[-1]), dtype=a.dtype)
# Loop through the flattened indices and calculate the matmuls
for i in range(0,a.shape[0]):
for j in range(0,b.shape[0]):
res[i,:,j,:] = afnumpy.dot(a[i],b[j])
# Finally appropriately reshape the result array
return res.reshape(res_shape)
def test_empty_ndarray():
a = afnumpy.zeros(())
b = numpy.zeros(())
iassert(a, b)
a = afnumpy.ndarray(0)
b = numpy.ndarray(0)
iassert(a, b)
a = afnumpy.ndarray((0, ))
b = numpy.ndarray((0, ))
iassert(a, b)
a = afnumpy.zeros(3)
b = numpy.zeros(3)
iassert(a[0:0], b[0:0])
def test_empty_ndarray():
a = afnumpy.zeros(())
b = numpy.zeros(())
iassert(a,b)
a = afnumpy.ndarray(0)
b = numpy.ndarray(0)
iassert(a,b)
a = afnumpy.ndarray((0,))
b = numpy.ndarray((0,))
iassert(a,b)
a = afnumpy.zeros(3)
b = numpy.zeros(3)
iassert(a[0:0],b[0:0])
def tile(A, reps):
try:
tup = tuple(reps)
except TypeError:
tup = (reps,)
if(len(tup) > 4):
raise NotImplementedError('Only up to 4 dimensions are supported')
d = len(tup)
shape = list(A.shape)
n = max(A.size, 1)
if (d < A.ndim):
tup = (1,)*(A.ndim-d) + tup
# Calculate final shape
while len(shape) < len(tup):
shape.insert(0,1)
shape = tuple(numpy.array(shape) * numpy.array(tup))
# Prepend ones to simplify calling
if (d < 4):
tup = (1,)*(4-d) + tup
tup = pu.c2f(tup)
s = arrayfire.tile(A.d_array, int(tup[0]), int(tup[1]), int(tup[2]), int(tup[3]))
return afnumpy.ndarray(shape, dtype=pu.typemap(s.dtype()), af_array=s)
def tile(A, reps):
try:
tup = tuple(reps)
except TypeError:
tup = (reps,)
if(len(tup) > 4):
raise NotImplementedError('Only up to 4 dimensions are supported')
d = len(tup)
shape = list(A.shape)
n = max(A.size, 1)
if (d < A.ndim):
tup = (1,)*(A.ndim-d) + tup
# Calculate final shape
while len(shape) < len(tup):
shape.insert(0,1)
shape = tuple(numpy.array(shape) * numpy.array(tup))
# Prepend ones to simplify calling
if (d < 4):
tup = (1,)*(4-d) + tup
tup = pu.c2f(tup)
s = arrayfire.tile(A.d_array, int(tup[0]), int(tup[1]), int(tup[2]), int(tup[3]))
return afnumpy.ndarray(shape, dtype=pu.typemap(s.dtype()), af_array=s)
def where(condition, x=pu.dummy, y=pu.dummy):
a = condition
s = arrayfire.where(a.d_array)
# numpy uses int64 while arrayfire uses uint32
s = afnumpy.ndarray(pu.af_shape(s), dtype=numpy.uint32, af_array=s).astype(numpy.int64)
# Looks like where goes through the JIT??
s.eval()
if(x is pu.dummy and y is pu.dummy):
idx = []
mult = 1
for i in a.shape[::-1]:
mult = i
idx = [s % mult] + idx
s //= mult
idx = tuple(idx)
return idx
elif(x is not pu.dummy and y is not pu.dummy):
if(x.dtype != y.dtype):
raise TypeError('x and y must have same dtype')
if(x.shape != y.shape):
raise ValueError('x and y must have same shape')
ret = afnumpy.array(y)
if(len(ret.shape) > 1):
ret = ret.flatten()
ret[s] = x.flatten()[s]
ret = ret.reshape(x.shape)
else:
ret[s] = x[s]
return ret;
else:
raise ValueError('either both or neither of x and y should be given')
def arctan2(x1, x2):
if isinstance(x1, afnumpy.ndarray) and isinstance(x2, afnumpy.ndarray):
s = arrayfire.atan2(x1.d_array, x2.d_array)
return afnumpy.ndarray(x1.shape,
dtype=pu.typemap(s.dtype()),
af_array=s)
else:
return numpy.arctan(x1, x2)
def sin(x):
if isinstance(x, afnumpy.ndarray):
s = arrayfire.sin(x.d_array)
return afnumpy.ndarray(x.shape,
dtype=pu.typemap(s.dtype()),
af_array=s)
else:
return numpy.sin(x)
def array(object, dtype=None, copy=True, order=None, subok=False, ndmin=0):
# We're going to ignore this for now
# if(subok is not False):
# raise NotImplementedError
if(order is not None and order is not 'K' and order is not 'C'):
raise NotImplementedError
# If it's not a numpy or afnumpy array first create a numpy array from it
if(not isinstance(object, afnumpy.ndarray) and
not isinstance(object, numpy.ndarray) and
not isinstance(object, arrayfire.array.Array)):
object = numpy.array(object, dtype=dtype, copy=copy, order=order, subok=subok, ndmin=ndmin)
if isinstance(object, arrayfire.array.Array):
shape = pu.c2f(object.dims())
else:
shape = object.shape
while(ndmin > len(shape)):
shape = (1,)+shape
if(dtype is None):
if isinstance(object, arrayfire.array.Array):
dtype = pu.typemap(object.dtype())
else:
dtype = object.dtype
if(isinstance(object, afnumpy.ndarray)):
if(copy):
s = arrayfire.cast(object.d_array.copy(), pu.typemap(dtype))
else:
s = arrayfire.cast(object.d_array, pu.typemap(dtype))
a = afnumpy.ndarray(shape, dtype=dtype, af_array=s)
a._eval()
return a
elif(isinstance(object, arrayfire.array.Array)):
if(copy):
s = arrayfire.cast(object.copy(), pu.typemap(dtype))
else:
s = arrayfire.cast(object, pu.typemap(dtype))
a = afnumpy.ndarray(shape, dtype=dtype, af_array=s)
a._eval()
return a
elif(isinstance(object, numpy.ndarray)):
return afnumpy.ndarray(shape, dtype=dtype, buffer=numpy.ascontiguousarray(object.astype(dtype, copy=copy)))
else:
raise AssertionError
def concatenate(arrays, axis=0):
if(len(arrays) < 1):
raise ValueError('need at least one array to concatenate')
if(axis > 3):
raise NotImplementedError('only up to 4 axis as currently supported')
arr = arrays[0].d_array.copy()
axis = pu.c2f(arrays[0].shape, axis)
for a in arrays[1:]:
arr = arrayfire.join(axis, arr, a.d_array)
return afnumpy.ndarray(pu.af_shape(arr), dtype=arrays[0].dtype, af_array=arr)
def concatenate(arrays, axis=0):
if (len(arrays) < 1):
raise ValueError('need at least one array to concatenate')
if (axis > 3):
raise NotImplementedError('only up to 4 axis as currently supported')
arr = arrays[0].d_array.copy()
axis = pu.c2f(arrays[0].shape, axis)
for a in arrays[1:]:
arr = arrayfire.join(axis, arr, a.d_array)
return afnumpy.ndarray(pu.af_shape(arr),
dtype=arrays[0].dtype,
af_array=arr)
def test_ndarray_constructor():
a = afnumpy.arrayfire.randu(3,2)
with pytest.raises(ValueError):
b = afnumpy.ndarray(a.dims(), dtype='f', af_array = a)
# This one should be fine
b = afnumpy.ndarray(a.dims()[::-1], dtype='f', af_array = a)
c = afnumpy.ndarray(a.dims()[::-1], dtype='f', buffer=a.raw_ptr(),
buffer_type=afnumpy.arrayfire.get_active_backend())
d = afnumpy.ndarray(a.dims()[::-1], dtype='f', buffer=a.raw_ptr(),
buffer_type=afnumpy.arrayfire.get_active_backend())
# Make sure they share the same underlying data
d[0,0] = 3
assert d[0,0] == c[0,0]
if afnumpy.arrayfire.get_active_backend() == 'cpu':
c = numpy.ones((3,2))
d = afnumpy.ndarray(c.shape, dtype=c.dtype, buffer=c.ctypes.data,
buffer_type=afnumpy.arrayfire.get_active_backend())
# Make sure they share the same underlying data
d[0,0] = -1
assert d[0,0] == c[0,0]
with pytest.raises(ValueError):
# Check for wrong backend
b = afnumpy.ndarray(c.shape, dtype=c.dtype, buffer=c.ctypes.data,
buffer_type='cuda')
def test_ndarray_constructor():
a = afnumpy.arrayfire.randu(3,2)
with pytest.raises(ValueError):
b = afnumpy.ndarray(a.dims(), dtype='f', af_array = a)
# This one should be fine
b = afnumpy.ndarray(a.dims()[::-1], dtype='f', af_array = a)
c = afnumpy.ndarray(a.dims()[::-1], dtype='f', buffer=a.raw_ptr(),
buffer_type=afnumpy.arrayfire.get_active_backend())
d = afnumpy.ndarray(a.dims()[::-1], dtype='f', buffer=a.raw_ptr(),
buffer_type=afnumpy.arrayfire.get_active_backend())
# Make sure they share the same underlying data
d[0,0] = 3
assert d[0,0] == c[0,0]
if afnumpy.arrayfire.get_active_backend() == 'cpu':
c = numpy.ones((3,2))
d = afnumpy.ndarray(c.shape, dtype=c.dtype, buffer=c.ctypes.data,
buffer_type=afnumpy.arrayfire.get_active_backend())
# Make sure they share the same underlying data
d[0,0] = -1
assert d[0,0] == c[0,0]
with pytest.raises(ValueError):
# Check for wrong backend
b = afnumpy.ndarray(c.shape, dtype=c.dtype, buffer=c.ctypes.data,
buffer_type='cuda')
def roll(a, shift, axis=None):
shape = a.shape
if(axis is None):
axis = 0
a = a.flatten()
axis = pu.c2f(a.shape, axis)
if axis == 0:
s = arrayfire.shift(a.d_array, shift, 0, 0, 0)
elif axis == 1:
s = arrayfire.shift(a.d_array, 0, shift, 0, 0)
elif axis == 2:
s = arrayfire.shift(a.d_array, 0, 0, shift, 0)
elif axis == 3:
s = arrayfire.shift(a.d_array, 0, 0, 0, shift)
else:
raise NotImplementedError
return afnumpy.ndarray(shape, dtype=a.dtype, af_array=s)
def sort(a, axis=-1, kind='quicksort', order=None):
try:
if kind != 'quicksort':
print( "sort 'kind' argument ignored" )
if order is not None:
raise ValueError('order argument is not supported')
if(axis is None):
input = a.flatten()
axis = 0
else:
input = a
if(axis < 0):
axis = a.ndim+axis
s = arrayfire.sort(input.d_array, pu.c2f(input.shape, axis))
return afnumpy.ndarray(input.shape, dtype=pu.typemap(s.dtype()), af_array=s)
except AttributeError:
return numpy.argsort(a, axis, kind, order)
def roll(a, shift, axis=None):
shape = a.shape
if (axis is None):
axis = 0
a = a.flatten()
axis = pu.c2f(a.shape, axis)
if axis == 0:
s = arrayfire.shift(a.d_array, shift, 0, 0, 0)
elif axis == 1:
s = arrayfire.shift(a.d_array, 0, shift, 0, 0)
elif axis == 2:
s = arrayfire.shift(a.d_array, 0, 0, shift, 0)
elif axis == 3:
s = arrayfire.shift(a.d_array, 0, 0, 0, shift)
else:
raise NotImplementedError
return afnumpy.ndarray(a.shape, dtype=a.dtype, af_array=s).reshape(shape)
def full(shape, fill_value, dtype=None, order='C'):
"""
Return a new array of given shape and type, filled with `fill_value`.
Parameters
----------
shape : int or sequence of ints
Shape of the new array, e.g., ``(2, 3)`` or ``2``.
fill_value : scalar
Fill value.
dtype : data-type, optional
The desired data-type for the array The default, `None`, means
`np.array(fill_value).dtype`.
order : {'C', 'F'}, optional
Whether to store multidimensional data in C- or Fortran-contiguous
(row- or column-wise) order in memory.
Returns
-------
out : ndarray
Array of `fill_value` with the given shape, dtype, and order.
See Also
--------
full_like : Return a new array with shape of input filled with value.
empty : Return a new uninitialized array.
ones : Return a new array setting values to one.
zeros : Return a new array setting values to zero.
Examples
--------
>>> np.full((2, 2), np.inf)
array([[ inf, inf],
[ inf, inf]])
>>> np.full((2, 2), 10)
array([[10, 10],
[10, 10]])
"""
if dtype is None:
dtype = array(fill_value).dtype
s = arrayfire.constant(fill_value, *shape[::-1], dtype=pu.typemap(dtype))
return afnumpy.ndarray(shape, dtype=pu.typemap(s.dtype()), af_array=s)
def ones(shape, dtype=float, order='C'):
b = numpy.ones(shape, dtype, order)
return afnumpy.ndarray(b.shape, b.dtype, buffer=b,order=order)
def isinf(x):
if not isinstance(x, afnumpy.ndarray):
return numpy.isinf(x)
s = arrayfire.isinf(x.d_array)
return afnumpy.ndarray(x.shape, dtype=pu.typemap(s.dtype()), af_array=s)
def sinh(x):
if isinstance(x, afnumpy.ndarray):
s = arrayfire.sinh(x.d_array)
return afnumpy.ndarray(x.shape, dtype=pu.typemap(s.dtype()), af_array=s)
else:
return numpy.sinh(x)
def arctan2(x1, x2):
if isinstance(x1, afnumpy.ndarray) and isinstance(x2, afnumpy.ndarray):
s = arrayfire.atan2(x1.d_array, x2.d_array)
return afnumpy.ndarray(x1.shape, dtype=pu.typemap(s.dtype()), af_array=s)
else:
return numpy.arctan(x1, x2)
#! /usr/bin/env python
import gnufft
import numpy as np
import afnumpy as afnp
import arrayfire as af
import time
from matplotlib import pyplot as plt
pts = afnp.ndarray((512, 128),
dtype=np.complex64,
af_array=af.randu(128, 512, dtype=af.Dtype.c32))
cplx = afnp.ndarray((512, 512),
dtype=np.complex64,
af_array=af.randu(512, 512, dtype=af.Dtype.c32))
kblut = afnp.ndarray((128, 1),
dtype=np.float32,
af_array=af.randu(1, 128, dtype=af.Dtype.f32))
grid = [512, 512]
scale = 12
k_r = 5
t0 = time.time()
res = gnufft.polarsample_transpose(pts, cplx, grid, kblut, scale, k_r)
t1 = time.time() - t0
print "walltime: " + str(t1)
print res.shape
print res.size
print res.dtype
def zeros(shape, dtype=float, order='C'):
b = numpy.zeros(shape, dtype, order)
return afnumpy.ndarray(b.shape, b.dtype, buffer=b,order=order)
#! /usr/bin/env python import numpy as np import afnumpy as afnp import afnumpy.fft as fft import arrayfire as af import gnufft a = af.randu(512) real = afnp.ndarray((512, 1), dtype=np.float32, af_array=a) a = af.randu(512) imag = afnp.ndarray((512, 1), dtype=np.float32, af_array=a) cplx = real + 1j * imag a = af.constant(1, 512, dtype=af.Dtype.f32) ones = afnp.ndarray((512, 1), dtype=np.float32, af_array=a) res = gnufft.debug(real, imag) t1 = res * ones t2 = fft.ifft(t1)
def vdot(a, b):
s = arrayfire.dot(arrayfire.conjg(a.flat.d_array), b.flat.d_array)
return afnumpy.ndarray((), dtype=a.dtype, af_array=s)[()]
def ceil(x, out=None):
s = arrayfire.ceil(x.d_array)
a = afnumpy.ndarray(x.shape, dtype=pu.typemap(s.dtype()), af_array=s)
if out is not None:
out[:] = a[:]
return a
def empty(shape, dtype=float, order='C'):
return afnumpy.ndarray(shape, dtype=dtype, order=order)
def ceil(x, out=None):
s = arrayfire.ceil(x.d_array)
a = afnumpy.ndarray(x.shape, dtype=pu.typemap(s.dtype()), af_array=s)
if out is not None:
out[:] = a[:]
return a
def __mul__(self, other):
other = afnp.asarray(other).astype(self.dtype)
s = arrayfire.matmul(self.d_array, other.d_array)
a = afnp.ndarray(pu.af_shape(s), dtype=pu.typemap(s.dtype()), af_array=s)
a._eval()
return a
def isinf(x):
if not isinstance(x, afnumpy.ndarray):
return numpy.isinf(x)
s = arrayfire.isinf(x.d_array)
return afnumpy.ndarray(x.shape, dtype=pu.typemap(s.dtype()), af_array=s)
def vdot(a, b):
s = arrayfire.dot(arrayfire.conjg(a.flat.d_array), b.flat.d_array)
return afnumpy.ndarray((), dtype=a.dtype, af_array=s)[()]