Commit e5f8538c authored by Ultima's avatar Ultima

First working and performant version of new los_response!!

Fixed allreduce and Allreduce in MPI_dummy.py.
Added Option to d2o to skip the parsing of the input parameters. Added cumsum function to d2o.
parent 3d9281f9
......@@ -79,11 +79,11 @@ class Intracomm(Comm):
return self._scattergather_helper(*args, **kwargs)
def Allreduce(self, sendbuf, recvbuf, op, **kwargs):
recvbuf[:] = op(sendbuf)
recvbuf[:] = sendbuf
return recvbuf
def allreduce(self, sendbuf, recvbuf, op, **kwargs):
recvbuf[:] = op(sendbuf)
recvbuf[:] = sendbuf
return recvbuf
def sendrecv(self, sendobj, **kwargs):
......
......@@ -87,7 +87,6 @@ class distributed_data_object(object):
If the supplied distribution strategy is not known.
"""
def __init__(self, global_data=None, global_shape=None, dtype=None,
local_data=None, local_shape=None,
distribution_strategy='fftw', hermitian=False,
......@@ -904,6 +903,14 @@ class distributed_data_object(object):
else:
return work_d2o
def cumsum(self, axis=None):
cumsum_data = self.distributor.cumsum(self.data, axis=axis)
result_d2o = self.copy_empty()
if axis is None:
result_d2o = result_d2o.flatten(inplace=True)
result_d2o.set_local_data(cumsum_data)
return result_d2o
def save(self, alias, path=None, overwriteQ=True):
"""
Saves a distributed_data_object to disk utilizing h5py.
......@@ -948,10 +955,20 @@ class _distributor_factory(object):
global_data=None, global_shape=None,
local_data=None, local_shape=None,
alias=None, path=None,
dtype=None, **kwargs):
dtype=None, skip_parsing=False, **kwargs):
if skip_parsing:
return_dict = {'comm': comm,
'dtype': dtype,
'name': distribution_strategy
}
if distribution_strategy in STRATEGIES['global']:
return_dict['global_shape'] = global_shape
elif distribution_strategy in STRATEGIES['local']:
return_dict['local_shape'] = local_shape
return return_dict
return_dict = {}
# Check that all nodes got the same distribution_strategy
strat_list = comm.allgather(distribution_strategy)
if all(x == strat_list[0] for x in strat_list) == False:
......@@ -998,7 +1015,7 @@ class _distributor_factory(object):
else:
dtype = np.dtype(dtype)
elif distribution_strategy in ['freeform']:
elif distribution_strategy in STRATEGIES['local']:
if dtype is None:
if isinstance(global_data, distributed_data_object):
dtype = global_data.dtype
......@@ -1021,7 +1038,7 @@ class _distributor_factory(object):
# Parse the shape
# Case 1: global-type slicer
if distribution_strategy in ['not', 'equal', 'fftw']:
if distribution_strategy in STRATEGIES['global']:
if dset is not None:
global_shape = dset.shape
elif global_data is not None and np.isscalar(global_data) == False:
......@@ -1376,6 +1393,9 @@ class _slicing_distributor(distributor):
hermitian = True
elif local_data is None:
local_data = np.empty(self.local_shape, dtype=self.dtype)
elif isinstance(local_data, np.ndarray):
local_data = local_data.astype(
self.dtype, copy=copy).reshape(self.local_shape)
else:
local_data = np.array(local_data).astype(
self.dtype, copy=copy).reshape(self.local_shape)
......@@ -2179,6 +2199,22 @@ class _slicing_distributor(distributor):
local_where)
return global_where
def cumsum(self, data, axis):
# compute the local np.cumsum
local_cumsum = np.cumsum(data, axis=axis)
if axis is None or axis == 0:
# communicate the highest value from node to node
rank = self.comm.rank
if local_cumsum.shape[0] == 0:
local_shift = np.zeros((), dtype=local_cumsum.dtype)
else:
local_shift = local_cumsum[-1]
local_shift_list = self.comm.allgather(local_shift)
local_sum_of_shift = np.sum(local_shift_list[:rank],
axis=0)
local_cumsum += local_sum_of_shift
return local_cumsum
def _sliceify(self, inp):
sliceified = []
result = []
......@@ -2196,7 +2232,10 @@ class _slicing_distributor(distributor):
else:
if x[i] >= self.global_shape[i]:
raise IndexError('Index out of bounds!')
result += [slice(x[i], x[i] + 1), ]
if x[i] == -1:
result += [slice(-1, None)]
else:
result += [slice(x[i], x[i] + 1), ]
sliceified += [True, ]
return (tuple(result), sliceified)
......@@ -2412,7 +2451,6 @@ class _slicing_distributor(distributor):
"the distributed_data_object."))
# check dtype
if dset.dtype != self.dtype:
print ('dsets', dset.dtype, self.dtype)
raise TypeError(about._errors.cstring(
"ERROR: The datatype of the given dataset does not " +
"match the one of the distributed_data_object."))
......@@ -2542,6 +2580,8 @@ class _not_distributor(distributor):
return np.empty(self.global_shape, dtype=self.dtype)
elif isinstance(data, distributed_data_object):
new_data = data.get_full_data()
elif isinstance(data, np.ndarray):
new_data = data
else:
new_data = np.array(data)
return new_data.astype(self.dtype,
......@@ -2608,6 +2648,11 @@ class _not_distributor(distributor):
local_where)
return global_where
def cumsum(self, data, axis):
# compute the local results from np.cumsum
cumsum = np.cumsum(data, axis=axis)
return cumsum
if 'h5py' in gdi:
def save_data(self, data, alias, path=None, overwriteQ=True):
comm = self.comm
......
# -*- coding: utf-8 -*-
#cython: nonecheck=False
#cython: boundscheck=False
#cython: wraparound=False
#cython: cdivision=True
import numpy as np
cimport numpy as np
cimport cython
from scipy.special import erf
FLOAT = np.float
ctypedef np.float_t FLOAT_t
......@@ -13,16 +14,18 @@ ctypedef np.float_t FLOAT_t
INT = np.int
ctypedef np.int_t INT_t
ctypedef np.ndarray[FLOAT_t, ndim=1] (*error_function_type)(np.ndarray[FLOAT_t,
ndim=1])
cdef extern from "numpy/npy_math.h":
bint isnan(double x)
bint signbit(double x)
double ceil(double x)
double floor(double x)
double sqrt(double x)
#from libc.math cimport isnan, signbit, ceil, floor, sqrt
cdef FLOAT_t NAN = float("NaN")
cdef class line_integrator(object):
cdef tuple shape
cdef list start
......@@ -37,22 +40,57 @@ cdef class line_integrator(object):
assert(np.all(np.array(self.shape) != 0))
assert(len(self.shape) == len(self.start) == len(self.end))
cpdef tuple integrate(self):
cpdef tuple integrate(self, bint with_cumsum=False):
cdef list indices
cdef np.ndarray[FLOAT_t, ndim=1] weights, normalized_cumsum_of_weights
if list_equal_Q(self.start, self.end):
return self._empty_results()
return self._empty_results(with_cumsum)
try:
projected_start = self._project_to_cuboid('start')
projected_end = self._project_to_cuboid('end')
except ValueError:
return self._empty_results()
return self._empty_results(with_cumsum)
(indices, weights) = self._integrate_through_cuboid(projected_start,
projected_end)
return (indices, weights)
if with_cumsum:
normalized_cumsum_of_weights = self._cumsum(weights,
projected_start)
return (indices, weights, normalized_cumsum_of_weights)
else:
return (indices, weights)
def _empty_results(self):
return ([np.array([], dtype=INT)] * len(self.shape),
np.array([], dtype=FLOAT))
def _empty_results(self, with_cumsum):
if with_cumsum:
return ([np.array([], dtype=INT)] * len(self.shape),
np.array([], dtype=FLOAT),
np.array([], dtype=FLOAT))
else:
return ([np.array([], dtype=INT)] * len(self.shape),
np.array([], dtype=FLOAT))
cpdef np.ndarray[FLOAT_t, ndim=1] _cumsum(self,
np.ndarray[FLOAT_t, ndim=1] weights,
list projected_start):
cdef list distance = list_sub(self.end, self.start)
cdef FLOAT_t total_length = list_norm(distance)
cdef list initial_skip = list_sub(projected_start, self.start)
cdef FLOAT_t skipped_length = list_norm(initial_skip)
cdef int i
cdef np.ndarray[FLOAT_t, ndim=1] cumsum = np.empty_like(weights)
cdef FLOAT_t[:] cumsum_view = cumsum
cdef FLOAT_t[:] weights_view = weights
cumsum_view[0] = (skipped_length + weights_view[0]/2.)/total_length
for i in xrange(1, len(weights)):
cumsum_view[i] = (cumsum_view[i-1] +
(weights_view[i] +
weights_view[i-1])/2./total_length)
return cumsum
cpdef list _project_to_cuboid(self, str mode):
cdef list a, b, c, p, surface_list, translator_list,\
......@@ -153,23 +191,25 @@ cdef class line_integrator(object):
cdef np.ndarray[FLOAT_t, ndim=1] weight_list = np.empty(num_estimate,
FLOAT)
current_position = start
i = 0
while True:
next_position, weight = self._get_next_position(current_position,
end)
floor_current_position = list_floor(current_position)
floor_next_position = list_floor(next_position)
for j in xrange(len(start)):
index_list[i, j] = floor_current_position[j]
if floor_current_position[j] < floor_next_position[j]:
index_list[i, j] = floor_current_position[j]
else:
index_list[i, j] = floor_next_position[j]
weight_list[i] = weight
if floor_current_position == list_floor(end):
if next_position == end:
break
current_position = next_position
i += 1
return (list(index_list[:i].T), weight_list[:i])
return (list(index_list[:i+1].T), weight_list[:i+1])
cdef tuple _get_next_position(self,
......@@ -209,15 +249,114 @@ cdef class line_integrator(object):
if isnan(best_translator_norm):
floor_position = list_floor(position)
# check if position is in the same cell as the endpoint
assert(floor_position == list_floor(end_position))
#assert(floor_position == list_floor(end_position))
weight = list_norm(list_sub(end_position, position))
next_position = position
next_position = end_position
else:
next_position = list_add(position, best_translator)
weight = list_norm(best_translator)
return (next_position, weight)
cpdef list multi_integrator(list starts,
list ends,
list sigmas_low,
list sigmas_up,
tuple shape,
list distances,
object error_function):
cdef INT_t i, j, total_number = len(starts[0]), dim = len(shape)
cdef list indices_and_weights_list = []
cdef list start = [None]*dim, end = [None]*dim
cdef list sigma_low = [None]*dim, sigma_up = [None]*dim
for i in xrange(total_number):
for j in xrange(dim):
start[j] = starts[j][i]
end[j] = ends[j][i]
sigma_low[j] = sigmas_low[j][i]
sigma_up[j] = sigmas_up[j][i]
indices_and_weights_list += single_integrator(i,
start,
end,
sigma_low,
sigma_up,
shape,
distances,
error_function)
return indices_and_weights_list
cdef list single_integrator(INT_t index,
list start,
list end,
list sigma_low,
list sigma_up,
tuple shape,
list distances,
object error_function):
cdef np.ndarray[FLOAT_t, ndim=1] pure_weights, low_weights, up_weights, \
low_lengths, up_lengths
cdef list pure_indices, low_indices, up_indices
# compute the three parts of a full line of sight
(pure_indices, pure_weights) = line_integrator(
shape, start, sigma_low).integrate()
(low_indices, low_weights, low_lengths) = line_integrator(
shape, sigma_low, end).integrate(True)
(up_indices, up_weights, up_lengths) = line_integrator(
shape, end, sigma_up).integrate(True)
# apply the error function on the sigma_low and sigma_up intervalls
low_weights = _apply_error_function(low_weights, low_lengths, False,
error_function)
up_weights = _apply_error_function(up_weights, up_lengths, True,
error_function)
# correct the volume distortion
cdef list direction = list_sub(end, start)
cdef FLOAT_t rescaler = (list_norm(list_mult(direction, distances))/
list_norm(direction))
pure_weights *= rescaler
low_weights *= rescaler
up_weights *= rescaler
# construct the result tuple
cdef list result_list = []
if pure_weights.shape[0] != 0:
result_list += [[index, pure_indices, pure_weights],]
if low_weights.shape[0] != 0:
result_list += [[index, low_indices, low_weights],]
if up_weights.shape[0] != 0:
result_list += [[index, up_indices, up_weights],]
return result_list
cdef np.ndarray[FLOAT_t, ndim=1] _apply_error_function(
np.ndarray[FLOAT_t, ndim=1] weights,
np.ndarray[FLOAT_t, ndim=1] lengths,
bint up_Q,
object error_function):
cdef np.ndarray[FLOAT_t, ndim=1] output_weights
if up_Q:
output_weights = weights * error_function(lengths)
else:
output_weights = weights * error_function(-1 + lengths)
return output_weights
cpdef np.ndarray[FLOAT_t, ndim=1] gaussian_error_function(
np.ndarray[FLOAT_t, ndim=1] x):
cdef FLOAT_t sigma = 0.5
return 0.5*(1 - erf(x / (sqrt(2.)*sigma)))
cdef INT_t strong_floor(FLOAT_t x):
cdef FLOAT_t floor_x
floor_x = floor(x)
......@@ -226,7 +365,7 @@ cdef INT_t strong_floor(FLOAT_t x):
else:
return INT(floor_x)
cpdef INT_t strong_ceil(FLOAT_t x):
cdef INT_t strong_ceil(FLOAT_t x):
cdef FLOAT_t ceil_x
ceil_x = ceil(x)
if ceil_x == x:
......@@ -278,33 +417,40 @@ cdef bint list_all_le(list list1, list list2):
return True
cdef list list_add(list list1, list list2):
cdef int ndim = len(list1)
cdef int i, ndim = len(list1)
cdef list result = [None]*ndim
for i in xrange(ndim):
result[i] = list1[i] + list2[i]
return result
cdef list list_sub(list list1, list list2):
cdef int ndim = len(list1)
cdef int i, ndim = len(list1)
cdef list result = [None]*ndim
for i in xrange(ndim):
result[i] = list1[i] - list2[i]
return result
cdef list list_mult(list list1, list list2):
cdef int i, ndim = len(list1)
cdef list result = [None]*ndim
for i in xrange(ndim):
result[i] = list1[i] * list2[i]
return result
#def test2():
# print ceil(1.5)
# print floor(1.5)
#
#def test():
# l = los_integrator_pure((1000,1000,1000), (-1,-1,-10), (1001, 1002, 1003))
# for i in xrange(30000):
# l._get_next_position([1.,1.1,1.2], [10.,10.,11.])
#
#def test3():
# print sqrt(3)
#
cdef list list_scalar_mult(list list1, FLOAT_t scaler):
cdef int ndim = len(list1)
cdef list result = [None]*ndim
for i in xrange(ndim):
result[i] = list1[i]*scaler
return result
cdef list list_scalar_div(list list1, FLOAT_t scaler):
cdef int ndim = len(list1)
cdef list result = [None]*ndim
for i in xrange(ndim):
result[i] = list1[i]/scaler
return result
......
# -*- coding: utf-8 -*-
import numpy as np
from line_integrator import line_integrator
from nifty.keepers import about
from line_integrator import multi_integrator, \
gaussian_error_function
from nifty.keepers import about,\
global_dependency_injector as gdi
from nifty.nifty_mpi_data import distributed_data_object,\
STRATEGIES
from nifty.nifty_core import point_space,\
field
from nifty.rg import rg_space
from nifty.operators import operator
MPI = gdi['MPI']
class los_response(operator):
def __init__(self, domain, starts, ends, sigmas_low=None, sigmas_up=None,
zero_point=None, error_function=lambda x: 0.5):
zero_point=None, error_function=gaussian_error_function,
target=None):
if not isinstance(domain, rg_space):
raise TypeError(about._errors.cstring(
"ERROR: The domain must be a rg_space instance."))
self.domain = domain
self.codomain = self.domain.get_codomain()
if not callable(error_function):
if callable(error_function):
self.error_function = error_function
else:
raise ValueError(about._errors.cstring(
"ERROR: error_function must be callable."))
......@@ -29,6 +42,20 @@ class los_response(operator):
starts, ends, sigmas_low,
sigmas_up, zero_point)
self.local_weights_and_indices = self._compute_weights_and_indices()
self.number_of_los = len(self.sigmas_low)
if target is None:
self.target = point_space(num=self.number_of_los,
dtype=self.domain.dtype,
datamodel=self.domain.datamodel,
comm=self.domain.comm)
else:
self.target = target
self.cotarget = self.target.get_codomain()
self.imp = True
self.uni = False
self.sym = False
......@@ -96,12 +123,11 @@ class los_response(operator):
parsed_ends = self._parse_startsends(ends, number_of_los)
# check that sigmas_up/lows have the right shape and parse scalars
parsed_sigmas_up = self._parse_sigmas_uplows(sigmas_up, number_of_los)
parsed_sigmas_low = self._parse_sigmas_uplows(sigmas_low,
number_of_los)
return (parsed_starts, parsed_ends, parsed_sigmas_up,
parsed_sigmas_low, parsed_zero_point)
parsed_sigmas_up = self._parse_sigmas_uplows(sigmas_up, number_of_los)
return (parsed_starts, parsed_ends, parsed_sigmas_low,
parsed_sigmas_up, parsed_zero_point)
def _parse_startsends(self, coords, number_of_los):
result_coords = [None]*len(coords)
......@@ -129,42 +155,162 @@ class los_response(operator):
"numpy ndarray."))
return parsed_sig
def convert_indices_to_physical(self, pixel_coordinates):
# first of all, compute the phyiscal distance of the given pixel
# from the zeroth-pixel
phyiscal_distance = np.array(pixel_coordinates) * \
np.array(self.domain.distances)
# add the offset of the zeroth pixel with respect to the coordinate
# system
physical_position = phyiscal_distance + np.array(self.zero_point)
return physical_position.tolist()
def convert_physical_to_indices(self, physical_position):
# compute the distance to the zeroth pixel
relative_position = np.array(physical_position) - \
np.array(self.zero_point)
# rescale the coordinates to the uniform grid
pixel_coordinates = relative_position / np.array(self.domain.distances)
return pixel_coordinates.tolist()
def convert_physical_to_indices(self, physical_positions):
pixel_coordinates = [None]*len(physical_positions)
local_zero_point = self._get_local_zero_point()
for i in xrange(len(pixel_coordinates)):
# Compute the distance to the zeroth pixel.
# Then rescale the coordinates to the uniform grid.
pixel_coordinates[i] = ((physical_positions[i] -
local_zero_point[i]) /
self.domain.distances[i]) + 0.5
return pixel_coordinates
def _convert_physical_to_pixel_lengths(self, lengths, starts, ends):
directions = np.array(ends) - np.array(starts)
distances = np.array(self.domain.distances)[:, None]
rescalers = (np.linalg.norm(directions / distances, axis=0) /
np.linalg.norm(directions, axis=0))
return lengths * rescalers
def _convert_sigmas_to_physical_coordinates(self, starts, ends,
sigmas_low, sigmas_up):
starts = np.array(starts)
ends = np.array(ends)
c = ends - starts
abs_c = np.linalg.norm(c, axis=0)
sigmas_low_coords = list(starts + (abs_c - sigmas_low)*c/abs_c)
sigmas_up_coords = list(starts + (abs_c + sigmas_up)*c/abs_c)
return (sigmas_low_coords, sigmas_up_coords)
def _get_local_zero_point(self):
if self.domain.datamodel == 'np':
return self.zero_point
elif self.domain.datamodel in STRATEGIES['not']:
return self.zero_point
elif self.domain.datamodel in STRATEGIES['slicing']:
dummy_d2o = distributed_data_object(
global_shape=self.domain.get_shape(),
dtype=np.dtype('int16'),