distributor_factory.py 95.3 KB
Newer Older
Theo Steininger's avatar
Theo Steininger committed
1
from __future__ import absolute_import
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
# D2O
# Copyright (C) 2016  Theo Steininger
#
# Author: Theo Steininger
#
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with this program.  If not, see <http://www.gnu.org/licenses/>.
theos's avatar
theos committed
19

20

Theo Steininger's avatar
Theo Steininger committed
21 22 23 24 25
from builtins import next
from builtins import str
from builtins import map
from builtins import range
from builtins import object
theos's avatar
theos committed
26 27
import numpy as np

Theo Steininger's avatar
Theo Steininger committed
28 29
from keepers import Loggable

theos's avatar
theos committed
30 31
from d2o.config import configuration as gc,\
                       dependency_injector as gdi
theos's avatar
theos committed
32

Theo Steininger's avatar
Theo Steininger committed
33
from .distributed_data_object import distributed_data_object
theos's avatar
theos committed
34

Theo Steininger's avatar
Theo Steininger committed
35
from .d2o_iter import d2o_slicing_iter,\
theos's avatar
theos committed
36
                     d2o_not_iter
Theo Steininger's avatar
Theo Steininger committed
37 38 39 40 41
from .d2o_librarian import d2o_librarian
from .dtype_converter import dtype_converter
from .cast_axis_to_tuple import cast_axis_to_tuple
from .translate_to_mpi_operator import op_translate_dict
from .slicing_generator import slicing_generator
theos's avatar
theos committed
42

Theo Steininger's avatar
Theo Steininger committed
43 44
from .strategies import STRATEGIES
from functools import reduce
theos's avatar
theos committed
45 46 47 48 49 50

MPI = gdi[gc['mpi_module']]
h5py = gdi.get('h5py')
pyfftw = gdi.get('pyfftw')


Theo Steininger's avatar
Theo Steininger committed
51
class _distributor_factory(Loggable, object):
theos's avatar
theos committed
52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74

    def __init__(self):
        self.distributor_store = {}

    def parse_kwargs(self, distribution_strategy, comm,
                     global_data=None, global_shape=None,
                     local_data=None, local_shape=None,
                     alias=None, path=None,
                     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 = {}

theos's avatar
theos committed
75
        expensive_checks = gc['mpi_init_checks']
theos's avatar
theos committed
76 77 78

        # Parse the MPI communicator
        if comm is None:
Theo Steininger's avatar
Theo Steininger committed
79 80
            raise ValueError("The distributor needs MPI-communicator object "
                             "comm!")
theos's avatar
theos committed
81 82 83 84 85 86
        else:
            return_dict['comm'] = comm

        if expensive_checks:
            # Check that all nodes got the same distribution_strategy
            strat_list = comm.allgather(distribution_strategy)
87
            if not all(x == strat_list[0] for x in strat_list):
Theo Steininger's avatar
Theo Steininger committed
88 89 90
                raise ValueError("The distribution-strategy must be the same "
                                 "on all nodes! Got: %s" %
                                 distribution_strategy)
theos's avatar
theos committed
91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114

        # Check for an hdf5 file and open it if given
        if 'h5py' in gdi and alias is not None:
            # set file path
            file_path = path if (path is not None) else alias
            # open hdf5 file
            if h5py.get_config().mpi and gc['mpi_module'] == 'MPI':
                f = h5py.File(file_path, 'r', driver='mpio', comm=comm)
            else:
                f = h5py.File(file_path, 'r')
            # open alias in file
            dset = f[alias]
        else:
            dset = None

        # Parse the datatype
        if distribution_strategy in ['not', 'equal', 'fftw'] and \
                (dset is not None):
            dtype = dset.dtype

        elif distribution_strategy in ['not', 'equal', 'fftw']:
            if dtype is None:
                if global_data is None:
                    dtype = np.dtype('float64')
Theo Steininger's avatar
Theo Steininger committed
115
                    self.logger.info("dtype was set to default.")
theos's avatar
theos committed
116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134
                else:
                    try:
                        dtype = global_data.dtype
                    except(AttributeError):
                        dtype = np.array(global_data).dtype
            else:
                dtype = np.dtype(dtype)

        elif distribution_strategy in STRATEGIES['local']:
            if dtype is None:
                if isinstance(global_data, distributed_data_object):
                    dtype = global_data.dtype
                elif local_data is not None:
                    try:
                        dtype = local_data.dtype
                    except(AttributeError):
                        dtype = np.array(local_data).dtype
                else:
                    dtype = np.dtype('float64')
Theo Steininger's avatar
Theo Steininger committed
135
                    self.logger.info("INFO: dtype set was set to default.")
theos's avatar
theos committed
136 137 138 139 140

            else:
                dtype = np.dtype(dtype)
        if expensive_checks:
            dtype_list = comm.allgather(dtype)
141
            if not all(x == dtype_list[0] for x in dtype_list):
Theo Steininger's avatar
Theo Steininger committed
142 143
                raise ValueError("The given dtype must be the same on all "
                                 "nodes!")
theos's avatar
theos committed
144 145 146 147 148 149 150
        return_dict['dtype'] = dtype

        # Parse the shape
        # Case 1: global-type slicer
        if distribution_strategy in STRATEGIES['global']:
            if dset is not None:
                global_shape = dset.shape
151
            elif global_data is not None and not np.isscalar(global_data):
theos's avatar
theos committed
152 153 154
                global_shape = global_data.shape
            elif global_shape is not None:
                global_shape = tuple(global_shape)
155 156
            elif global_data is not None:
                global_shape = ()
theos's avatar
theos committed
157
            else:
Theo Steininger's avatar
Theo Steininger committed
158 159
                raise ValueError("Neither global_data nor global_shape nor "
                                 "hdf5 file supplied!")
theos's avatar
theos committed
160 161 162 163 164

            if expensive_checks:
                global_shape_list = comm.allgather(global_shape)
                if not all(x == global_shape_list[0]
                           for x in global_shape_list):
Theo Steininger's avatar
Theo Steininger committed
165 166
                    raise ValueError("The global_shape must be the same on "
                                     "all nodes!")
theos's avatar
theos committed
167 168 169 170 171 172
            return_dict['global_shape'] = global_shape

        # Case 2: local-type slicer
        elif distribution_strategy in ['freeform']:
            if isinstance(global_data, distributed_data_object):
                local_shape = global_data.local_shape
173
            elif local_data is not None and not np.isscalar(local_data):
theos's avatar
theos committed
174 175 176 177
                local_shape = local_data.shape
            elif local_shape is not None:
                local_shape = tuple(local_shape)
            else:
Theo Steininger's avatar
Theo Steininger committed
178 179
                raise ValueError("Neither non-0-dimensional local_data nor "
                                 "local_shape nor global d2o supplied!")
theos's avatar
theos committed
180
            if local_shape == ():
Theo Steininger's avatar
Theo Steininger committed
181
                raise ValueError("local_shape == () is not a valid shape!")
theos's avatar
theos committed
182 183 184 185 186 187

            if expensive_checks:
                local_shape_list = comm.allgather(local_shape[1:])
                cleared_set = set(local_shape_list)
                cleared_set.discard(())
                if len(cleared_set) > 1:
Theo Steininger's avatar
Theo Steininger committed
188 189
                    raise ValueError("All but the first entry of local_shape "
                                     "must be the same on all nodes!")
theos's avatar
theos committed
190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217
            return_dict['local_shape'] = local_shape

        # Add the name of the distributor if needed
        if distribution_strategy in ['equal', 'fftw', 'freeform']:
            return_dict['name'] = distribution_strategy

        # close the file-handle
        if dset is not None:
            f.close()

        return return_dict

    def hash_arguments(self, distribution_strategy, **kwargs):
        kwargs = kwargs.copy()

        comm = kwargs['comm']
        kwargs['comm'] = id(comm)

        if 'global_shape' in kwargs:
            kwargs['global_shape'] = kwargs['global_shape']
        if 'local_shape' in kwargs:
            local_shape = kwargs['local_shape']
            local_shape_list = comm.allgather(local_shape)
            kwargs['local_shape'] = tuple(local_shape_list)

        kwargs['dtype'] = self.dictionize_np(kwargs['dtype'])
        kwargs['distribution_strategy'] = distribution_strategy

Theo Steininger's avatar
Theo Steininger committed
218
        return frozenset(list(kwargs.items()))
theos's avatar
theos committed
219 220

    def dictionize_np(self, x):
Theo Steininger's avatar
Theo Steininger committed
221
        dic = list(x.type.__dict__.items())
theos's avatar
theos committed
222 223 224 225 226 227 228 229 230
        if x is np.float:
            dic[24] = 0
            dic[29] = 0
            dic[37] = 0
        return frozenset(dic)

    def get_distributor(self, distribution_strategy, comm, **kwargs):
        # check if the distribution strategy is known
        if distribution_strategy not in STRATEGIES['all']:
Theo Steininger's avatar
Theo Steininger committed
231
            raise ValueError("Unknown distribution strategy supplied.")
theos's avatar
theos committed
232 233 234 235 236 237 238

        # parse the kwargs
        parsed_kwargs = self.parse_kwargs(
            distribution_strategy=distribution_strategy,
            comm=comm,
            **kwargs)

239 240
        if parsed_kwargs.get('global_shape') == ():
            distribution_strategy = 'not'
Theo Steininger's avatar
Theo Steininger committed
241 242
            self.logger.warn("Distribution strategy was set to "
                             "'not' because of global_shape == ()")
243

theos's avatar
theos committed
244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284
        hashed_kwargs = self.hash_arguments(distribution_strategy,
                                            **parsed_kwargs)
        # check if the distributors has already been produced in the past
        if hashed_kwargs in self.distributor_store:
            return self.distributor_store[hashed_kwargs]
        else:
            # produce new distributor
            if distribution_strategy == 'not':
                produced_distributor = _not_distributor(**parsed_kwargs)

            elif distribution_strategy == 'equal':
                produced_distributor = _slicing_distributor(
                    slicer=_equal_slicer,
                    **parsed_kwargs)

            elif distribution_strategy == 'fftw':
                produced_distributor = _slicing_distributor(
                    slicer=_fftw_slicer,
                    **parsed_kwargs)
            elif distribution_strategy == 'freeform':
                produced_distributor = _slicing_distributor(
                    slicer=_freeform_slicer,
                    **parsed_kwargs)

            self.distributor_store[hashed_kwargs] = produced_distributor
            return self.distributor_store[hashed_kwargs]


distributor_factory = _distributor_factory()


def _infer_key_type(key):
    if key is None:
        return (None, None)
    found_boolean = False
    # Check which case we got:
    if isinstance(key, slice) or np.isscalar(key):
        found = 'slicetuple'
    elif isinstance(key, tuple) or isinstance(key, list):
        # Check if there is something different in the array than
        # scalars and slices
Theo Steininger's avatar
Theo Steininger committed
285 286
        scalarQ = np.array(list(map(np.isscalar, key)))
        sliceQ = np.array([isinstance(z, slice) for z in key])
theos's avatar
theos committed
287 288 289 290 291 292 293 294 295 296 297
        if np.all(scalarQ + sliceQ):
            found = 'slicetuple'
        else:
            found = 'indexinglist'
    elif isinstance(key, np.ndarray):
        found = 'ndarray'
        found_boolean = (key.dtype.type == np.bool_)
    elif isinstance(key, distributed_data_object):
        found = 'd2o'
        found_boolean = (key.dtype == np.bool_)
    else:
Theo Steininger's avatar
Theo Steininger committed
298
        raise ValueError("Unknown keytype!")
theos's avatar
theos committed
299 300 301 302 303
    return (found, found_boolean)


class distributor(object):

theos's avatar
theos committed
304 305 306 307 308 309 310 311 312 313 314
    def distribute_data(self, data=None, alias=None,
                        path=None, copy=True, **kwargs):
        '''
        distribute data checks
        - whether the data is located on all nodes or only on node 0
        - that the shape of 'data' matches the global_shape
        '''

        if 'h5py' in gdi and alias is not None:
            data = self.load_data(alias=alias, path=path)

315 316 317 318
        h5py_dataset_Q = False
        if 'h5py' in gdi:
            h5py_dataset_Q = isinstance(data, h5py.Dataset)

theos's avatar
theos committed
319
        if data is None:
320
            result_data = np.empty(self.local_shape, dtype=self.dtype)
theos's avatar
theos committed
321
        elif np.isscalar(data):
322 323
            result_data = np.empty(self.local_shape, dtype=self.dtype)
            result_data[:] = data
theos's avatar
theos committed
324
        elif isinstance(data, np.ndarray) or \
325 326
                isinstance(data, distributed_data_object) or \
                h5py_dataset_Q:
theos's avatar
theos committed
327 328
            data = self.extract_local_data(data)

329
            if data.shape != self.local_shape:
theos's avatar
theos committed
330 331 332 333 334 335 336 337
                copy = True

            if copy:
                result_data = np.empty(self.local_shape, dtype=self.dtype)
                result_data[:] = data
            else:
                result_data = data

338 339
            # fix dtype
            result_data = result_data.astype(self.dtype, copy=False)
theos's avatar
theos committed
340 341
        else:
            new_data = np.array(data)
342 343 344 345
            result_data = new_data.astype(self.dtype,
                                          copy=copy).reshape(self.local_shape)

        return result_data
theos's avatar
theos committed
346

theos's avatar
theos committed
347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389
    def disperse_data(self, data, to_key, data_update, from_key=None,
                      local_keys=False, copy=True, **kwargs):
        # Check which keys we got:
        (to_found, to_found_boolean) = _infer_key_type(to_key)
        (from_found, from_found_boolean) = _infer_key_type(from_key)

        comm = self.comm
        if local_keys is False:
            return self._disperse_data_primitive(
                                         data=data,
                                         to_key=to_key,
                                         data_update=data_update,
                                         from_key=from_key,
                                         copy=copy,
                                         to_found=to_found,
                                         to_found_boolean=to_found_boolean,
                                         from_found=from_found,
                                         from_found_boolean=from_found_boolean,
                                         **kwargs)

        else:
            # assert that all to_keys are from same type
            to_found_list = comm.allgather(to_found)
            assert(all(x == to_found_list[0] for x in to_found_list))
            to_found_boolean_list = comm.allgather(to_found_boolean)
            assert(all(x == to_found_boolean_list[0] for x in
                       to_found_boolean_list))
            from_found_list = comm.allgather(from_found)
            assert(all(x == from_found_list[0] for x in from_found_list))
            from_found_boolean_list = comm.allgather(from_found_boolean)
            assert(all(x == from_found_boolean_list[0] for
                       x in from_found_boolean_list))

            # gather the local to_keys into a global to_key_list
            # Case 1: the to_keys are not distributed_data_objects
            # -> allgather does the job
            if to_found != 'd2o':
                to_key_list = comm.allgather(to_key)
            # Case 2: if the to_keys are distributed_data_objects, gather
            # the index of the array and build the to_key_list with help
            # from the librarian
            else:
                to_index_list = comm.allgather(to_key.index)
Theo Steininger's avatar
Theo Steininger committed
390
                to_key_list = [d2o_librarian[z] for z in to_index_list]
theos's avatar
theos committed
391 392 393 394 395 396

            # gather the local from_keys. It is the same procedure as above
            if from_found != 'd2o':
                from_key_list = comm.allgather(from_key)
            else:
                from_index_list = comm.allgather(from_key.index)
Theo Steininger's avatar
Theo Steininger committed
397
                from_key_list = [d2o_librarian[z] for z in from_index_list]
theos's avatar
theos committed
398 399 400

            local_data_update_is_scalar = np.isscalar(data_update)
            local_scalar_list = comm.allgather(local_data_update_is_scalar)
Theo Steininger's avatar
Theo Steininger committed
401
            for i in range(len(to_key_list)):
theos's avatar
theos committed
402 403 404 405 406
                if np.all(np.array(local_scalar_list) == True):
                    scalar_list = comm.allgather(data_update)
                    temp_data_update = scalar_list[i]
                elif isinstance(data_update, distributed_data_object):
                    data_update_index_list = comm.allgather(data_update.index)
Theo Steininger's avatar
Theo Steininger committed
407
                    data_update_list = [d2o_librarian[z] for z in data_update_index_list]
theos's avatar
theos committed
408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449
                    temp_data_update = data_update_list[i]
                else:
                    # build a temporary freeform d2o which only contains data
                    # from node i
                    if comm.rank == i:
                        temp_shape = np.shape(data_update)
                        try:
                            temp_dtype = np.dtype(data_update.dtype)
                        except(TypeError):
                            temp_dtype = np.array(data_update).dtype
                    else:
                        temp_shape = None
                        temp_dtype = None
                    temp_shape = comm.bcast(temp_shape, root=i)
                    temp_dtype = comm.bcast(temp_dtype, root=i)

                    if comm.rank != i:
                        temp_shape = list(temp_shape)
                        temp_shape[0] = 0
                        temp_shape = tuple(temp_shape)
                        temp_data = np.empty(temp_shape, dtype=temp_dtype)
                    else:
                        temp_data = data_update
                    temp_data_update = distributed_data_object(
                                        local_data=temp_data,
                                        distribution_strategy='freeform',
                                        copy=False,
                                        comm=self.comm)
                # disperse the data one after another
                self._disperse_data_primitive(
                                      data=data,
                                      to_key=to_key_list[i],
                                      data_update=temp_data_update,
                                      from_key=from_key_list[i],
                                      copy=copy,
                                      to_found=to_found,
                                      to_found_boolean=to_found_boolean,
                                      from_found=from_found,
                                      from_found_boolean=from_found_boolean,
                                      **kwargs)
                i += 1

450
    def bincount(self, obj, length, weights=None, axis=None):
451 452
        if obj.shape == ():
            raise ValueError("object of too small depth for desired array")
453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468
        data = obj.get_local_data(copy=False)
        # this implementation fits all distribution strategies where the
        # axes of the global array correspond to the axes of the local data

        if weights is not None:
            local_weights = self.extract_local_data(weights)
        else:
            local_weights = None

        # if present, parse the axis keyword and transpose/reorder self.data
        # such that all affected axes follow each other. Only if they are in a
        # sequence flattening will be possible
        if axis is not None:
            # do the reordering
            ndim = len(self.global_shape)
            axis = sorted(cast_axis_to_tuple(axis, length=ndim))
Theo Steininger's avatar
Theo Steininger committed
469
            reordering = [x for x in range(ndim) if x not in axis]
470 471 472 473 474 475
            reordering += axis

            data = np.transpose(data, reordering)
            if local_weights is not None:
                local_weights = np.transpose(local_weights, reordering)

Theo Steininger's avatar
Theo Steininger committed
476
            reord_axis = list(range(ndim-len(axis), ndim))
477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502

            # semi-flatten the dimensions in `axis`, i.e. after reordering
            # the last ones.
            semi_flat_dim = reduce(lambda x, y: x*y,
                                   data.shape[ndim-len(reord_axis):])
            flat_shape = data.shape[:ndim-len(reord_axis)] + (semi_flat_dim, )
        else:
            flat_shape = (reduce(lambda x, y: x*y, data.shape), )

        data = np.ascontiguousarray(data.reshape(flat_shape))
        if local_weights is not None:
            local_weights = np.ascontiguousarray(
                                local_weights.reshape(flat_shape))

        # compute the local bincount results
        # -> prepare the local result array
        if local_weights is None:
            result_dtype = np.int
        else:
            result_dtype = np.float
        local_counts = np.empty(flat_shape[:-1] + (length, ),
                                dtype=result_dtype)
        # iterate over all entries in the surviving axes and compute the local
        # bincounts
        for slice_list in slicing_generator(flat_shape,
                                            axes=(len(flat_shape)-1, )):
theos's avatar
theos committed
503 504 505 506 507 508 509 510
            if local_weights is not None:
                current_weights = local_weights[slice_list]
            else:
                current_weights = None
            local_counts[slice_list] = np.bincount(
                                            data[slice_list],
                                            weights=current_weights,
                                            minlength=length)
511 512 513 514 515 516

        # restore the original ordering
        # place the bincount stuff at the location of the first `axis` entry
        if axis is not None:
            # axis has been sorted above
            insert_position = axis[0]
theos's avatar
theos committed
517
            new_ndim = len(local_counts.shape)
Theo Steininger's avatar
Theo Steininger committed
518
            return_order = (list(range(0, insert_position)) +
theos's avatar
theos committed
519
                            [new_ndim-1, ] +
Theo Steininger's avatar
Theo Steininger committed
520
                            list(range(insert_position, new_ndim-1)))
521 522 523 524
            local_counts = np.ascontiguousarray(
                                local_counts.transpose(return_order))
        return self._combine_local_bincount_counts(obj, local_counts, axis)

theos's avatar
theos committed
525 526 527 528 529 530 531 532 533 534 535

class _slicing_distributor(distributor):
    def __init__(self, slicer, name, dtype, comm, **remaining_parsed_kwargs):

        self.comm = comm
        self.distribution_strategy = name
        self.dtype = np.dtype(dtype)

        self._my_dtype_converter = dtype_converter

        if not self._my_dtype_converter.known_np_Q(self.dtype):
Theo Steininger's avatar
Theo Steininger committed
536 537
            raise TypeError("The datatype " + str(self.dtype.__repr__()) +
                            " is not known to mpi4py.")
theos's avatar
theos committed
538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596

        self.mpi_dtype = self._my_dtype_converter.to_mpi(self.dtype)

        self.slicer = slicer
        self._local_size = self.slicer(comm=comm, **remaining_parsed_kwargs)
        self.local_start = self._local_size[0]
        self.local_end = self._local_size[1]
        self.global_shape = self._local_size[2]
        self.global_dim = reduce(lambda x, y: x*y, self.global_shape)

        self.local_length = self.local_end - self.local_start
        self.local_shape = (self.local_length,) + tuple(self.global_shape[1:])
        self.local_dim = np.product(self.local_shape)
        self.local_dim_list = np.empty(comm.size, dtype=np.int)
        comm.Allgather([np.array(self.local_dim, dtype=np.int), MPI.INT],
                       [self.local_dim_list, MPI.INT])
        self.local_dim_offset = np.sum(self.local_dim_list[0:comm.rank])

        self.local_slice = np.array([self.local_start, self.local_end,
                                     self.local_length, self.local_dim,
                                     self.local_dim_offset],
                                    dtype=np.int)
        # collect all local_slices
        self.all_local_slices = np.empty((comm.size, 5), dtype=np.int)
        comm.Allgather([np.array((self.local_slice,), dtype=np.int), MPI.INT],
                       [self.all_local_slices, MPI.INT])

    def initialize_data(self, global_data, local_data, alias, path, hermitian,
                        copy, **kwargs):
        if 'h5py' in gdi and alias is not None:
            local_data = self.load_data(alias=alias, path=path)
            return (local_data, hermitian)

        if self.distribution_strategy in ['equal', 'fftw']:
            if np.isscalar(global_data):
                local_data = np.empty(self.local_shape, dtype=self.dtype)
                local_data.fill(global_data)
                hermitian = True
            else:
                local_data = self.distribute_data(data=global_data,
                                                  copy=copy)
        elif self.distribution_strategy in ['freeform']:
            if isinstance(global_data, distributed_data_object):
                local_data = global_data.get_local_data(copy=copy)
            elif np.isscalar(local_data):
                temp_local_data = np.empty(self.local_shape,
                                           dtype=self.dtype)
                temp_local_data.fill(local_data)
                local_data = temp_local_data
                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)
        else:
Theo Steininger's avatar
Theo Steininger committed
597
            raise TypeError("Unknown istribution strategy")
theos's avatar
theos committed
598 599 600 601 602 603 604 605
        return (local_data, hermitian)

    def globalize_flat_index(self, index):
        return int(index) + self.local_dim_offset

    def globalize_index(self, index):
        index = np.array(index, dtype=np.int).flatten()
        if index.shape != (len(self.global_shape),):
Theo Steininger's avatar
Theo Steininger committed
606 607
            raise TypeError("Length of index tuple does not match the array's "
                            "shape!")
theos's avatar
theos committed
608 609 610 611 612 613 614 615
        globalized_index = index
        globalized_index[0] = index[0] + self.local_start
        # ensure that the globalized index list is within the bounds
        global_index_memory = globalized_index
        globalized_index = np.clip(globalized_index,
                                   -np.array(self.global_shape),
                                   np.array(self.global_shape) - 1)
        if np.any(global_index_memory != globalized_index):
Theo Steininger's avatar
Theo Steininger committed
616
            self.logger.warn("Indices were clipped!")
theos's avatar
theos committed
617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666
        globalized_index = tuple(globalized_index)
        return globalized_index

    def _allgather(self, thing, comm=None):
        if comm is None:
            comm = self.comm
        gathered_things = comm.allgather(thing)
        return gathered_things

    def _Allreduce_helper(self, sendbuf, recvbuf, op):
        send_dtype = self._my_dtype_converter.to_mpi(sendbuf.dtype)
        recv_dtype = self._my_dtype_converter.to_mpi(recvbuf.dtype)
        self.comm.Allreduce([sendbuf, send_dtype],
                            [recvbuf, recv_dtype],
                            op=op)
        return recvbuf

    def _selective_allreduce(self, data, op, bufferQ=False):
        size = self.comm.size
        rank = self.comm.rank

        if size == 1:
            if data is None:
                raise ValueError("ERROR: No process with non-None data.")
            result_data = data
        else:

            # infer which data should be included in the allreduce and if its
            # array data
            if data is None:
                got_array = np.array([0])
            elif not isinstance(data, np.ndarray):
                got_array = np.array([2])
            elif reduce(lambda x, y: x*y, data.shape) == 0:
                got_array = np.array([1])
            elif np.issubdtype(data.dtype, np.complexfloating):
                # MPI.MAX and MPI.MIN do not support complex data types
                got_array = np.array([3])
            else:
                got_array = np.array([4])

            got_array_list = np.empty(size, dtype=np.int)
            self.comm.Allgather([got_array, MPI.INT],
                                [got_array_list, MPI.INT])

            if reduce(lambda x, y: x & y, got_array_list == 1):
                return data

            # get first node with non-None data
            try:
Theo Steininger's avatar
Theo Steininger committed
667
                start = next(i for i in range(size) if got_array_list[i] > 1)
theos's avatar
theos committed
668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685
            except(StopIteration):
                raise ValueError("ERROR: No process with non-None data.")

            # check if the Uppercase function can be used or not
            # -> check if op supports buffers and if we got real array-data
            if bufferQ and got_array_list[start] == 4:
                # Send the dtype and shape from the start process to the others
                (new_dtype,
                 new_shape) = self.comm.bcast((data.dtype,
                                               data.shape), root=start)
                mpi_dtype = self._my_dtype_converter.to_mpi(new_dtype)
                if rank == start:
                    result_data = data
                else:
                    result_data = np.empty(new_shape, dtype=new_dtype)

                self.comm.Bcast([result_data, mpi_dtype], root=start)

Theo Steininger's avatar
Theo Steininger committed
686
                for i in range(start+1, size):
theos's avatar
theos committed
687 688 689 690 691 692 693 694 695 696
                    if got_array_list[i] > 1:
                        if rank == i:
                            temp_data = data
                        else:
                            temp_data = np.empty(new_shape, dtype=new_dtype)
                        self.comm.Bcast([temp_data, mpi_dtype], root=i)
                        result_data = op(result_data, temp_data)

            else:
                result_data = self.comm.bcast(data, root=start)
Theo Steininger's avatar
Theo Steininger committed
697
                for i in range(start+1, size):
theos's avatar
theos committed
698 699 700 701 702 703 704 705 706 707 708
                    if got_array_list[i] > 1:
                        temp_data = self.comm.bcast(data, root=i)
                        result_data = op(result_data, temp_data)
        return result_data

    def contraction_helper(self, parent, function, allow_empty_contractions,
                           axis=None, **kwargs):
        if axis == ():
            return parent.copy()

        old_shape = parent.shape
709
        axis = cast_axis_to_tuple(axis, length=len(self.global_shape))
theos's avatar
theos committed
710 711 712
        if axis is None:
            new_shape = ()
        else:
Theo Steininger's avatar
Theo Steininger committed
713
            new_shape = tuple([old_shape[i] for i in range(len(old_shape))
theos's avatar
theos committed
714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753
                               if i not in axis])

        local_data = parent.data

        try:
            contracted_local_data = function(local_data, axis=axis, **kwargs)
        except(ValueError):
            contracted_local_data = None

        # check if additional contraction along the first axis must be done
        if axis is None or 0 in axis:
            (mpi_op, bufferQ) = op_translate_dict[function]
            contracted_global_data = self._selective_allreduce(
                                        contracted_local_data,
                                        mpi_op,
                                        bufferQ)
            new_dist_strategy = 'not'
        else:
            if contracted_local_data is None:
                # raise the exception implicitly
                function(local_data, axis=axis, **kwargs)
            contracted_global_data = contracted_local_data
            new_dist_strategy = parent.distribution_strategy

        if new_shape == ():
            result = contracted_global_data
        else:
            new_dtype = contracted_global_data.dtype
            # try to store the result in a distributed_data_object with the
            # distribution_strategy as parent
            result = parent.copy_empty(global_shape=new_shape,
                                       dtype=new_dtype,
                                       distribution_strategy=new_dist_strategy)

            # However, there are cases where the contracted data does not any
            # longer follow the prior distribution scheme.
            # Example: FFTW distribution on 4 MPI processes
            # Contracting (4, 4) to (4,).
            # (4, 4) was distributed (1, 4)...(1, 4)
            # (4, ) is not distributed like (1,)...(1,) but like (2,)(2,)()()!
754 755 756
            equalQ = (result.local_shape == contracted_global_data.shape)
            all_equalQ = np.all(self._allgather(equalQ))
            if not all_equalQ:
theos's avatar
theos committed
757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782
                result = parent.copy_empty(
                                    local_shape=contracted_global_data.shape,
                                    dtype=new_dtype,
                                    distribution_strategy='freeform')
            result.set_local_data(contracted_global_data, copy=False)

        return result

    def _disperse_data_primitive(self, data, to_key, data_update, from_key,
                                 copy, to_found, to_found_boolean, from_found,
                                 from_found_boolean, **kwargs):
        if np.isscalar(data_update):
            from_key = None

        # Case 1: to_key is a slice-tuple. Hence, the basic indexing/slicing
        # machinery will be used
        if to_found == 'slicetuple':
            if from_found == 'slicetuple':
                return self.disperse_data_to_slices(data=data,
                                                    to_slices=to_key,
                                                    data_update=data_update,
                                                    from_slices=from_key,
                                                    copy=copy,
                                                    **kwargs)
            else:
                if from_key is not None:
Theo Steininger's avatar
Theo Steininger committed
783 784 785
                    self.logger.info("Advanced injection is not available for "
                                     "this combination of to_key and "
                                     "from_key.")
theos's avatar
theos committed
786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801
                    prepared_data_update = data_update[from_key]
                else:
                    prepared_data_update = data_update

                return self.disperse_data_to_slices(
                                            data=data,
                                            to_slices=to_key,
                                            data_update=prepared_data_update,
                                            copy=copy,
                                            **kwargs)

        # Case 2: key is an array
        elif (to_found == 'ndarray' or to_found == 'd2o'):
            # Case 2.1: The array is boolean.
            if to_found_boolean:
                if from_key is not None:
Theo Steininger's avatar
Theo Steininger committed
802 803 804
                    self.logger.info("Advanced injection is not available for "
                                     "this combination of to_key and "
                                     "from_key.")
theos's avatar
theos committed
805 806 807 808 809 810 811 812 813 814 815 816 817
                    prepared_data_update = data_update[from_key]
                else:
                    prepared_data_update = data_update
                return self.disperse_data_to_bool(
                                              data=data,
                                              to_boolean_key=to_key,
                                              data_update=prepared_data_update,
                                              copy=copy,
                                              **kwargs)
            # Case 2.2: The array is not boolean. Only 1-dimensional
            # advanced slicing is supported.
            else:
                if len(to_key.shape) != 1:
Theo Steininger's avatar
Theo Steininger committed
818 819
                    raise ValueError("Only one-dimensional advanced indexing "
                                     "is supported")
theos's avatar
theos committed
820 821 822 823 824 825 826 827 828 829
                # Make a recursive call in order to trigger the 'list'-section
                return self.disperse_data(data=data, to_key=[to_key],
                                          data_update=data_update,
                                          from_key=from_key, copy=copy,
                                          **kwargs)

        # Case 3 : to_key is a list. This list is interpreted as
        # one-dimensional advanced indexing list.
        elif to_found == 'indexinglist':
            if from_key is not None:
Theo Steininger's avatar
Theo Steininger committed
830 831
                self.logger.info("Advanced injection is not available for "
                                 "this combination of to_key and from_key.")
theos's avatar
theos committed
832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913
                prepared_data_update = data_update[from_key]
            else:
                prepared_data_update = data_update
            return self.disperse_data_to_list(data=data,
                                              to_list_key=to_key,
                                              data_update=prepared_data_update,
                                              copy=copy,
                                              **kwargs)

    def disperse_data_to_list(self, data, to_list_key, data_update,
                              copy=True, **kwargs):

        if to_list_key == []:
            return data

        local_to_list_key = self._advanced_index_decycler(to_list_key)
        return self._disperse_data_to_list_and_bool_helper(
            data=data,
            local_to_key=local_to_list_key,
            data_update=data_update,
            copy=copy,
            **kwargs)

    def disperse_data_to_bool(self, data, to_boolean_key, data_update,
                              copy=True, **kwargs):
        # Extract the part of the to_boolean_key which corresponds to the
        # local data
        local_to_boolean_key = self.extract_local_data(to_boolean_key)
        return self._disperse_data_to_list_and_bool_helper(
            data=data,
            local_to_key=local_to_boolean_key,
            data_update=data_update,
            copy=copy,
            **kwargs)

    def _disperse_data_to_list_and_bool_helper(self, data, local_to_key,
                                               data_update, copy, **kwargs):
        comm = self.comm
        rank = comm.rank
        # Infer the length and offset of the locally affected data
        locally_affected_data = data[local_to_key]
        data_length = np.shape(locally_affected_data)[0]
        data_length_list = comm.allgather(data_length)
        data_length_offset_list = np.append([0],
                                            np.cumsum(data_length_list)[:-1])

        # Update the local data object with its very own portion
        o = data_length_offset_list
        l = data_length

        if isinstance(data_update, distributed_data_object):
            local_data_update = data_update.get_data(
                                          slice(o[rank], o[rank] + l),
                                          local_keys=True
                                          ).get_local_data(copy=False)
            data[local_to_key] = local_data_update.astype(self.dtype,
                                                          copy=False)
        elif np.isscalar(data_update):
            data[local_to_key] = data_update
        else:
            data[local_to_key] = np.array(data_update[o[rank]:o[rank] + l],
                                          copy=copy).astype(self.dtype,
                                                            copy=False)
        return data

    def disperse_data_to_slices(self, data, to_slices,
                                data_update, from_slices=None, copy=True):
        comm = self.comm
        (to_slices, sliceified) = self._sliceify(to_slices)

        # parse the to_slices object
        localized_to_start, localized_to_stop = self._backshift_and_decycle(
            to_slices[0], self.local_start, self.local_end,
            self.global_shape[0])
        local_to_slice = (slice(localized_to_start, localized_to_stop,
                                to_slices[0].step),) + to_slices[1:]
        local_to_slice_shape = data[local_to_slice].shape

        to_step = to_slices[0].step
        if to_step is None:
            to_step = 1
        elif to_step == 0:
Theo Steininger's avatar
Theo Steininger committed
914
            raise ValueError("to_step size == 0!")
theos's avatar
theos committed
915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950

        # Compute the offset of the data the individual node will take.
        # The offset is free of stepsizes. It is the offset in terms of
        # the purely transported data. If to_step < 0, the offset will
        # be calculated in reverse order
        order = np.sign(to_step)

        local_affected_data_length = local_to_slice_shape[0]
        local_affected_data_length_list = np.empty(comm.size, dtype=np.int)
        comm.Allgather(
            [np.array(local_affected_data_length, dtype=np.int), MPI.INT],
            [local_affected_data_length_list, MPI.INT])
        local_affected_data_length_offset_list = np.append([0],
                                                           np.cumsum(
            local_affected_data_length_list[::order])[:-1])[::order]

        if np.isscalar(data_update):
            data[local_to_slice] = data_update
        else:
            # construct the locally adapted from_slice object
            r = comm.rank
            o = local_affected_data_length_offset_list
            l = local_affected_data_length

            data_update = self._enfold(data_update, sliceified)

            # parse the from_slices object
            if from_slices is None:
                from_slices = (slice(None, None, None),)
            (from_slices_start, from_slices_stop) = \
                self._backshift_and_decycle(
                                            slice_object=from_slices[0],
                                            shifted_start=0,
                                            shifted_stop=data_update.shape[0],
                                            global_length=data_update.shape[0])
            if from_slices_start is None:
Theo Steininger's avatar
Theo Steininger committed
951 952
                raise ValueError("_backshift_and_decycle should never return "
                                 "None for local_start!")
theos's avatar
theos committed
953 954 955 956 957 958

            # parse the step sizes
            from_step = from_slices[0].step
            if from_step is None:
                from_step = 1
            elif from_step == 0:
Theo Steininger's avatar
Theo Steininger committed
959
                raise ValueError("from_step size must not be equal to 0!")
theos's avatar
theos committed
960 961 962 963 964 965 966 967 968 969 970 971 972 973 974

            localized_from_start = from_slices_start + from_step * o[r]
            localized_from_stop = localized_from_start + from_step * l
            if localized_from_stop < 0:
                localized_from_stop = None

            localized_from_slice = (slice(localized_from_start,
                                          localized_from_stop,
                                          from_step),)

            update_slice = localized_from_slice + from_slices[1:]

            if isinstance(data_update, distributed_data_object):
                selected_update = data_update.get_data(
                                 key=update_slice,
975 976
                                 local_keys=True,
                                 copy=copy)
theos's avatar
theos committed
977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025
                local_data_update = selected_update.get_local_data(copy=False)
                local_data_update = local_data_update.astype(self.dtype,
                                                             copy=False)
                if np.prod(np.shape(local_data_update)) != 0:
                    data[local_to_slice] = local_data_update
            # elif np.isscalar(data_update):
            #    data[local_to_slice] = data_update
            else:
                local_data_update = np.array(data_update)[update_slice]
                if np.prod(np.shape(local_data_update)) != 0:
                    data[local_to_slice] = np.array(
                                                local_data_update,
                                                copy=copy).astype(self.dtype,
                                                                  copy=False)

    def collect_data(self, data, key, local_keys=False, copy=True, **kwargs):
        # collect_data supports three types of keys
        # Case 1: key is a slicing/index tuple
        # Case 2: key is a boolean-array of the same shape as self
        # Case 3: key is a list of shape (n,), where n is
        #         0<n<len(self.shape). The entries of the list must be a
        #         scalar/list/tuple/ndarray. If not scalar the length must be
        #         the same for all of the lists. This is essentially
        #         numpy advanced indexing in one dimension, only.

        # Check which case we got:
        (found, found_boolean) = _infer_key_type(key)
        comm = self.comm
        if local_keys is False:
            return self._collect_data_primitive(data, key, found,
                                                found_boolean, copy=copy,
                                                **kwargs)
        else:
            # assert that all keys are from same type
            found_list = comm.allgather(found)
            assert(all(x == found_list[0] for x in found_list))
            found_boolean_list = comm.allgather(found_boolean)
            assert(all(x == found_boolean_list[0] for x in found_boolean_list))

            # gather the local_keys into a global key_list
            # Case 1: the keys are no distributed_data_objects
            # -> allgather does the job
            if found != 'd2o':
                key_list = comm.allgather(key)
            # Case 2: if the keys are distributed_data_objects, gather
            # the index of the array and build the key_list with help
            # from the librarian
            else:
                index_list = comm.allgather(key.index)
Theo Steininger's avatar
Theo Steininger committed
1026
                key_list = [d2o_librarian[z] for z in index_list]
theos's avatar
theos committed
1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067
            i = 0
            for temp_key in key_list:
                # build the locally fed d2o
                temp_d2o = self._collect_data_primitive(data, temp_key, found,
                                                        found_boolean,
                                                        copy=copy, **kwargs)
                # collect the data stored in the d2o to the individual target
                # rank
                temp_data = temp_d2o.get_full_data(target_rank=i)
                if comm.rank == i:
                    individual_data = temp_data
                i += 1
            return_d2o = distributed_data_object(
                            local_data=individual_data,
                            distribution_strategy='freeform',
                            copy=False,
                            comm=self.comm)
            return return_d2o

    def _collect_data_primitive(self, data, key, found, found_boolean,
                                copy=True, **kwargs):

        # Case 1: key is a slice-tuple. Hence, the basic indexing/slicing
        # machinery will be used
        if found == 'slicetuple':
            return self.collect_data_from_slices(data=data,
                                                 slice_objects=key,
                                                 copy=copy,
                                                 **kwargs)
        # Case 2: key is an array
        elif (found == 'ndarray' or found == 'd2o'):
            # Case 2.1: The array is boolean.
            if found_boolean:
                return self.collect_data_from_bool(data=data,
                                                   boolean_key=key,
                                                   copy=copy,
                                                   **kwargs)
            # Case 2.2: The array is not boolean. Only 1-dimensional
            # advanced slicing is supported.
            else:
                if len(key.shape) != 1:
Theo Steininger's avatar
Theo Steininger committed
1068 1069
                    raise ValueError("Only one-dimensional advanced indexing "
                                     "is supported")
theos's avatar
theos committed
1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083
                # Make a recursive call in order to trigger the 'list'-section
                return self.collect_data(data=data, key=[key], copy=copy,
                                         **kwargs)

        # Case 3 : key is a list. This list is interpreted as one-dimensional
        # advanced indexing list.
        elif found == 'indexinglist':
            return self.collect_data_from_list(data=data,
                                               list_key=key,
                                               copy=copy,
                                               **kwargs)

    def collect_data_from_list(self, data, list_key, copy=True, **kwargs):
        if list_key == []:
Theo Steininger's avatar
Theo Steininger committed
1084
            raise ValueError("key == [] is an unsupported key!")
theos's avatar
theos committed
1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108
        local_list_key = self._advanced_index_decycler(list_key)
        local_result = data[local_list_key]
        global_result = distributed_data_object(
                                            local_data=local_result,
                                            distribution_strategy='freeform',
                                            copy=copy,
                                            comm=self.comm)
        return global_result

    def _advanced_index_decycler(self, from_list_key):
        global_length = self.global_shape[0]
        local_length = self.local_length
        shift = self.local_start
        rank = self.comm.rank

        zeroth_key = from_list_key[0]
        # Check if from_list_key is a scalar
        if np.isscalar(zeroth_key):
            # decycle negative index
            if zeroth_key < 0:
                zeroth_key += global_length
            # if the index is still negative, or it is greater than
            # global_length the index is ill-choosen
            if zeroth_key < 0 or zeroth_key >= global_length:
Theo Steininger's avatar
Theo Steininger committed
1109
                raise ValueError("Index out of bounds!")
theos's avatar
theos committed
1110 1111 1112 1113 1114
            # shift the index
            local_zeroth_key = zeroth_key - shift
            # if the index lies within the local nodes' data-range
            # take the shifted index, combined with rest of from_list_key
            result = [local_zeroth_key]
Theo Steininger's avatar
Theo Steininger committed
1115
            for ii in range(1, len(from_list_key)):
theos's avatar
theos committed
1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132
                current = from_list_key[ii]
                if isinstance(current, distributed_data_object):
                    result.append(current.get_full_data())
                else:
                    result.append(current)
            if (local_zeroth_key < 0) or (local_zeroth_key >= local_length):
                result = (np.array([], dtype=np.dtype('int')),) * \
                    len(from_list_key)

        elif isinstance(zeroth_key, distributed_data_object):
            zeroth_key = zeroth_key.copy()
            # decycle negative indices
            zeroth_key[zeroth_key < 0] = zeroth_key[zeroth_key < 0] + \
                global_length
            # if there are still negative indices, or indices greater than
            # global_length the indices are ill-choosen
            if (zeroth_key < 0).any() or (zeroth_key >= global_length).any():
Theo Steininger's avatar
Theo Steininger committed
1133
                raise ValueError("Index out of bounds!")
theos's avatar
theos committed
1134 1135
            # shift the indices according to shift
            shift_list = self.comm.allgather(shift)
Theo Steininger's avatar
Theo Steininger committed
1136
            local_zeroth_key_list = [zeroth_key - z for z in shift_list]
theos's avatar
theos committed
1137 1138
            # discard all entries where the indices are negative or larger
            # than local_length
Theo Steininger's avatar
Theo Steininger committed
1139
            greater_than_lower_list = [z >= 0 for z in local_zeroth_key_list]
theos's avatar
theos committed
1140 1141
            # -> build up a list with the local selection d2o's
            local_length_list = self.comm.allgather(local_length)
Theo Steininger's avatar
Theo Steininger committed
1142 1143 1144 1145 1146 1147 1148 1149
            less_than_upper_list = list(map(lambda z, zz: z < zz,
                                            local_zeroth_key_list,
                                            local_length_list))
            local_selection_list = list(map(lambda z, zz: z * zz,
                                            less_than_upper_list,
                                            greater_than_lower_list))

            for j in range(len(local_zeroth_key_list)):
theos's avatar
theos committed
1150 1151 1152 1153 1154 1155 1156 1157
                temp_result = local_zeroth_key_list[j].\
                    get_data(local_selection_list[j]).\
                    get_full_data(target_rank=j)
                if j == rank:
                    result = temp_result
# TODO: Implement fast check!
#            if not all(result[i] <= result[i + 1]
#                       for i in xrange(len(result) - 1)):
theos's avatar
theos committed
1158
#                raise ValueError(about_cstring(
theos's avatar
theos committed
1159 1160 1161
#                   "ERROR: The first dimemnsion of list_key must be sorted!"))

            result = [result]
Theo Steininger's avatar
Theo Steininger committed
1162
            for ii in range(1, len(from_list_key)):
theos's avatar
theos committed
1163 1164 1165 1166 1167 1168 1169 1170
                current = from_list_key[ii]
                if np.isscalar(current):
                    result.append(current)
                elif isinstance(current, distributed_data_object):
                    result.append(current.get_data(
                                   local_selection_list[rank],
                                   local_keys=True).get_local_data(copy=False))
                else:
Theo Steininger's avatar
Theo Steininger committed
1171
                    for j in range(len(local_selection_list)):
theos's avatar
theos committed
1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185
                        temp_select = local_selection_list[j].\
                            get_full_data(target_rank=j)
                        if j == rank:
                            temp_result = current[temp_select]
                    result.append(temp_result)

        else:
            zeroth_key = zeroth_key.copy()
            # decycle negative indices
            zeroth_key[zeroth_key < 0] = zeroth_key[zeroth_key < 0] + \
                global_length
            # if there are still negative indices, or indices greater than
            # global_length the indices are ill-choosen
            if (zeroth_key < 0).any() or (zeroth_key >= global_length).any():
Theo Steininger's avatar
Theo Steininger committed
1186
                raise ValueError("Index out of bounds!")
theos's avatar
theos committed
1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198
            # shift the indices according to shift
            local_zeroth_key = zeroth_key - shift
            # discard all entries where the indices are negative or larger
            # than local_length
            greater_than_lower = (local_zeroth_key >= 0)
            less_than_upper = (local_zeroth_key < local_length)
            local_selection = greater_than_lower * less_than_upper

            result = [local_zeroth_key[local_selection]]
# TODO: Implement fast check!
#            if not all(result[0][i] <= result[0][i + 1]
#                       for i in xrange(len(result[0]) - 1)):
theos's avatar
theos committed
1199
#                raise ValueError(about_cstring(
theos's avatar
theos committed
1200 1201
#                   "ERROR: The first dimemnsion of list_key must be sorted!"))

Theo Steininger's avatar
Theo Steininger committed
1202
            for ii in range(1, len(from_list_key)):
theos's avatar
theos committed
1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279
                current = from_list_key[ii]
                if np.isscalar(current):
                    result.append(current)
                elif isinstance(current, distributed_data_object):
                    result.append(current.get_data(
                                   local_selection,
                                   local_keys=True).get_local_data(copy=False))
                else:
                    result.append(current[local_selection])

        return result

    def collect_data_from_bool(self, data, boolean_key, copy=True, **kwargs):
        local_boolean_key = self.extract_local_data(boolean_key)
        local_result = data[local_boolean_key]
        global_result = distributed_data_object(
                                            local_data=local_result,
                                            distribution_strategy='freeform',
                                            copy=copy,
                                            comm=self.comm)
        return global_result

    def _invert_mpi_data_ordering(self, data):
        data = np.ascontiguousarray(data)

        comm = self.comm
        s = comm.size
        r = comm.rank
        if s == 1:
            return data

        partner = s - 1 - r

        new_shape = comm.sendrecv(sendobj=data.shape,
                                  dest=partner,
                                  source=partner)
        new_data = np.empty(new_shape,
                            dtype=self.dtype)

        comm.Sendrecv(sendbuf=[data, self.mpi_dtype],
                      recvbuf=[new_data, self.mpi_dtype],
                      dest=partner,
                      source=partner)

        return new_data

    def collect_data_from_slices(self, data, slice_objects, copy=True,
                                 target_rank='all'):

        (slice_objects, sliceified) = self._sliceify(slice_objects)

        localized_start, localized_stop = self._backshift_and_decycle(
            slice_objects[0],
            self.local_start,
            self.local_end,
            self.global_shape[0])
        first_step = slice_objects[0].step
        local_slice = (slice(localized_start,
                             localized_stop,
                             first_step),) + slice_objects[1:]

        local_result = data[local_slice]
        if (first_step is not None) and (first_step < 0):
            local_result = self._invert_mpi_data_ordering(local_result)

        global_result = distributed_data_object(
                                            local_data=local_result,
                                            distribution_strategy='freeform',
                                            copy=copy,
                                            comm=self.comm)

        return self._defold(global_result, sliceified)

    def _backshift_and_decycle(self, slice_object, shifted_start, shifted_stop,
                               global_length):

        # Reformulate negative indices
Theo Steininger's avatar
Theo Steininger committed
1280 1281 1282 1283 1284
        if slice_object.start is not None:
            if slice_object.start < 0:
                temp_start = slice_object.start + global_length
                if temp_start < 0:
                    temp_start = 0
theos's avatar
theos committed
1285

Theo Steininger's avatar
Theo Steininger committed
1286 1287
                slice_object = slice(temp_start, slice_object.stop,
                                     slice_object.step)
theos's avatar
theos committed
1288

Theo Steininger's avatar
Theo Steininger committed
1289 1290 1291 1292 1293
        if slice_object.stop is not None:
            if slice_object.stop < 0:
                temp_stop = slice_object.stop + global_length
                if temp_stop < 0:
                    temp_stop = None
theos's avatar
theos committed
1294

Theo Steininger's avatar
Theo Steininger committed
1295 1296
                slice_object = slice(slice_object.start, temp_stop,
                                     slice_object.step)
theos's avatar
theos committed
1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382

        # initialize the step
        if slice_object.step is None:
            step = 1
        else:
            step = slice_object.step

        # compute local_length
        local_length = shifted_stop - shifted_start
        if step > 0:
            shift = shifted_start
            # calculate the start index
            if slice_object.start is None:
                local_start = (-shift) % step  # step size compensation
            else:
                local_start = slice_object.start - shift
                # if the local_start is negative, pull it up to zero
                local_start = \
                    local_start % step if local_start < 0 else local_start

            if local_start >= local_length:
                return (0, 0)

            # calculate the stop index
            if slice_object.stop is None:
                local_stop = None
            else:
                local_stop = slice_object.stop - shift
                # if local_stop is negative, the slice is empty
                if local_stop < 0:
                    return (0, 0)
                if local_stop > local_length:
                    local_stop = None

        else:  # if step < 0
            step = -step
            # calculate the start index. (Here, local_start > local_stop!)
            if slice_object.start is None:
                local_start = (local_length - 1) -\
                    (-(global_length - shifted_stop)
                     ) % step  # stepsize compensation
                # if local_start becomes negative here, it means, that the
                # step size is bigger than the length of the local slice and
                # that no relevant data is in this slice
                if local_start < 0:
                    return (0, 0)
            else:
                if slice_object.start > global_length - 1:
                    slice_object = slice(global_length - 1,
                                         slice_object.stop,
                                         slice_object.step)
                local_start = slice_object.start - shifted_start
                # if the local_start is negative, immediately return the
                # values for an empty slice
                if local_start < 0:
                    return (0, 0)

                # if the local_start is greater than the local length, pull
                # it down
                if local_start > local_length - 1:
                    overhead = local_start - (local_length - 1)
                    overhead = overhead - overhead % (-step)
                    local_start = local_start - overhead
                    # if local_start becomes negative here, it means, that the
                    # step size is bigger than the length of the localslice and
                    # that no relevant data is in this slice
                    if local_start < 0:
                        return (0, 0)

            # calculate the stop index
            if slice_object.stop is None:
                local_stop = None
            else:
                local_stop = slice_object.stop - shifted_start
                # if local_stop is negative, pull it up to None
                local_stop = None if local_stop < 0 else local_stop
        # Note: if start or stop are greater than the array length,
        # numpy will automatically cut the index value down into the
        # array's range
#        if local_start > local_length:
#            local_start = local_length
#        if local_stop > local_length:
#            local_stop = local_length
        return (local_start, local_stop)

    def extract_local_data(self, data_object):
1383 1384 1385
        h5py_dataset_Q = False
        if 'h5py' in gdi:
            h5py_dataset_Q = isinstance(data_object, h5py.Dataset)
theos's avatar
theos committed
1386 1387
        # if data_object is not a ndarray or a d2o, cast it to a ndarray
        if not (isinstance(data_object, np.ndarray) or
1388 1389
                isinstance(data_object, distributed_data_object) or
                h5py_dataset_Q):
theos's avatar
theos committed
1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405
            data_object = np.array(data_object)
        # check if the shapes are remotely compatible, reshape if possible
        # and determine which dimensions match only via broadcasting
        try:
            (data_object, matching_dimensions) = \
                self._reshape_foreign_data(data_object)
        # if the shape-casting fails, try to fix things via local data
        # matching
        except(ValueError):
            # Check if all the local shapes match the supplied data
            local_matchQ = (self.local_shape == data_object.shape)
            global_matchQ = self._allgather(local_matchQ)
            # if the local shapes match, simply return the data_object
            if np.all(global_matchQ):
                extracted_data = data_object[:]
            else:
Theo Steininger's avatar
Theo Steininger committed
1406 1407
                raise ValueError("supplied shapes do neither match globally "
                                 "nor locally")
theos's avatar
theos committed
1408 1409 1410

        # if shape-casting was successfull, extract the data
        else:
theos's avatar
theos committed
1411 1412 1413
            if isinstance(data_object, distributed_data_object):
                # If the first dimension matches only via broadcasting...
                # Case 1: ...do broadcasting.
1414
                if not matching_dimensions[0]:
theos's avatar
theos committed
1415 1416 1417 1418 1419 1420 1421 1422 1423
                    extracted_data = data_object.get_full_data()
                    extracted_data = extracted_data[0]
                else:
                    # Case 2: First dimension fits directly and data_object is
                    # a d2o
                    # Check if both d2os have the same slicing
                    # If the distributor is exactly the same, extract the data
                    if self is data_object.distributor:
                        # Simply take the local data
theos's avatar
theos committed
1424
                        extracted_data = data_object.data
theos's avatar
theos committed
1425 1426 1427 1428 1429 1430 1431 1432 1433
                    # If the distributor is not exactly the same, check if the
                    # geometry matches if it is a slicing distributor
                    # -> comm and local shapes
                    elif (isinstance(data_object.distributor,
                                     _slicing_distributor) and
                          (self.comm is data_object.distributor.comm) and
                          (np.all(self.all_local_slices ==
                                  data_object.distributor.all_local_slices))):
                            extracted_data = data_object.data
theos's avatar
theos committed
1434

theos's avatar
theos committed
1435
                    else:
1436 1437
                        # Case 2: no.
                        # All nodes extract their local slice from the
theos's avatar
theos committed
1438 1439 1440 1441 1442 1443
                        # data_object
                        extracted_data =\
                            data_object.get_data(slice(self.local_start,
                                                       self.local_end),
                                                 local_keys=True)
                        extracted_data = extracted_data.get_local_data()
theos's avatar
theos committed
1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463


#                # Check if the distributor and the comm match
#                # the own ones. Checking equality via 'is' is ok, as the
#                # distributor factory caches simmilar distributors
#                if self is data_object.distributor and\
#                        self.comm is data_object.distributor.comm:
#                    # Case 1: yes. Simply take the local data
#                    extracted_data = data_object.data
#                # If the distributors do not match directly, check
#                else:
#                    # Case 2: no. All nodes extract their local slice from the
#                    # data_object
#                    extracted_data =\
#                        data_object.get_data(slice(self.local_start,
#                                                   self.local_end),
#                                             local_keys=True)
#                    extracted_data = extracted_data.get_local_data()
#
#
theos's avatar
theos committed
1464 1465 1466
            # Case 2: np-array
            # If the first dimension matches only via broadcasting
            # ...do broadcasting.
1467
            elif not matching_dimensions[0]:
theos's avatar
theos committed
1468
                extracted_data = data_object[0:1]
theos's avatar
theos committed
1469
            # Case 3: First dimension fits directly and data_object is an
1470
            # generic array or h5py dataset
theos's avatar
theos committed
1471 1472 1473 1474 1475 1476 1477
            else:
                extracted_data =\
                    data_object[self.local_start:self.local_end]

        return extracted_data

    def _reshape_foreign_data(self, foreign):
theos's avatar
theos committed
1478

theos's avatar
theos committed
1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497
        # Case 1:
        # check if the shapes match directly
        if self.global_shape == foreign.shape:
            matching_dimensions = [True, ] * len(self.global_shape)
            return (foreign, matching_dimensions)
        # Case 2:
        # if not, try to reshape the input data.
        # In particular, this will fail when foreign is a d2o as long as
        # reshaping is not implemented
        try:
            output = foreign.reshape(self.global_shape)
            matching_dimensions = [True, ] * len(self.global_shape)
            return (output, matching_dimensions)
        except(ValueError, AttributeError):
            pass
        # Case 3:
        # if this does not work, try to broadcast the shape
        # check if the dimensions match
        if len(self.global_shape) != len(foreign.shape):
Theo Steininger's avatar
Theo Steininger committed
1498
            raise ValueError("unequal number of dimensions!")
theos's avatar
theos committed
1499 1500 1501 1502 1503 1504 1505 1506
        # check direct matches
        direct_match = (np.array(self.global_shape) == np.array(foreign.shape))
        # check broadcast compatibility
        broadcast_match = (np.ones(len(self.global_shape), dtype=int) ==
                           np.array(foreign.shape))
        # combine the matches and assert that all are true
        combined_match = (direct_match | broadcast_match)
        if not np.all(combined_match):
Theo Steininger's avatar
Theo Steininger committed
1507
            raise ValueError("incompatible shapes!")
theos's avatar
theos committed
1508 1509 1510 1511 1512 1513 1514 1515 1516
        matching_dimensions = tuple(direct_match)
        return (foreign, matching_dimensions)

    def consolidate_data(self, data, target_rank='all', comm=None):
        if comm is None:
            comm = self.comm
        _gathered_data = np.empty(self.global_shape, dtype=self.dtype)
        _dim_list = self.all_local_slices[:, 3]
        _dim_offset_list = self.all_local_slices[:, 4]
1517
        data = np.ascontiguousarray(data)
theos's avatar
theos committed
1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541
        if target_rank == 'all':
            comm.Allgatherv([data, self.mpi_dtype],
                            [_gathered_data, _dim_list, _dim_offset_list,
                             self.mpi_dtype])
        else:
            comm.Gatherv([data, self.mpi_dtype],
                         [_gathered_data, _dim_list, _dim_offset_list,
                          self.mpi_dtype],
                         root=target_rank)
        return _gathered_data

    def flatten(self, data, inplace=False):
        if inplace:
            return data.ravel()
        else:
            return data.flatten()

    def where(self, data):
        # compute np.where result from the node's local data
        local_where = list(np.where(data))
        # shift the first dimension
        local_where[0] = local_where[0] + self.local_start
        local_where = tuple(local_where)

Theo Steininger's avatar
Theo Steininger committed
1542 1543 1544 1545
        global_where = \
            [distributed_data_object(local_data=z,
                                     distribution_strategy='freeform')
             for z in local_where]
theos's avatar
theos committed
1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563
        return global_where

    def unique(self, data):
        # if the size of the MPI communicator is equal to 1, the
        # reduce operator will not be applied. -> Cover this case directly.
        comm = self.comm
        size = self.comm.size
        rank = self.comm.rank
        if size == 1:
            global_unique_data = np.unique(data)
        else:
            local_unique_data = np.unique(data)
            local_data_length = np.array([local_unique_data.shape[0]])
            local_data_length_list = np.empty(size, dtype=np.int)
            comm.Allgather([local_data_length, MPI.INT],
                           [local_data_length_list, MPI.INT])

            global_unique_data = np.array([], dtype=self.dtype)
Theo Steininger's avatar
Theo Steininger committed
1564
            for i in range(size):
theos's avatar
theos committed
1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578
                # broadcast data to the other nodes
                # prepare the recv array
                if rank != i:
                    work_shape = local_data_length_list[i]
                    work_array = np.empty(work_shape, dtype=self.dtype)
                else:
                    work_array = local_unique_data
                # do the actual broadcasting
                comm.Bcast([work_array, self.mpi_dtype], root=i)
                global_unique_data = np.unique(
                                        np.concatenate([work_array.flatten(),
                                                        global_unique_data]))
        return global_unique_data

1579 1580 1581 1582 1583 1584 1585
    def _combine_local_bincount_counts(self, obj, local_counts, axis):
        if axis is None or 0 in axis:
            global_counts = np.empty_like(local_counts)
            self._Allreduce_helper(local_counts, global_counts, MPI.SUM)
            result_object = obj.copy_empty(global_shape=global_counts.shape,
                                           dtype=global_counts.dtype,
                                           distribution_strategy='not')
1586 1587
            result_object.set_local_data(global_counts, copy=False)

theos's avatar
theos committed
1588
        else:
1589
            global_counts = local_counts
1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612
            global_shape = list(global_counts.shape)
            global_shape[0] = self.global_shape[0]

            # Try to preserve the distribution_strategy if possible.
            # Choose the local distribution_strategy if...
            # -> if the distribution_strategy was local in the beginning
            # -> if the slicing of the first axis changes even though
            #    the bincounting wasn't performed on it. This can be the case
            #    if axis=0 is extremely small and collapsing axis=1,2,3...
            #    forces fftw to use a different slicing.
            # Both cases are triggered by an ValueError exception

            try:
                if self.distribution_strategy not in STRATEGIES['global']:
                    raise ValueError

                result_object = obj.copy_empty(global_shape=global_shape,
                                               dtype=global_counts.dtype)
                result_object.set_local_data(global_counts, copy=False)

            except ValueError:
                result_object = obj.copy_empty(
                                           local_shape=global_counts.shape,
1613 1614 1615
                                           dtype=global_counts.dtype,
                                           distribution_strategy='freeform')
        return result_object
theos's avatar
theos committed
1616

1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634

#    def bincount(self, local_data, local_weights, minlength, axis=None):
#        if local_weights is None:
#            result_dtype = np.int
#        else:
#            result_dtype = np.float
#
#        local_counts = np.bincount(local_data,
#                                   weights=local_weights,
#                                   minlength=minlength)
#
#        # cast the local_counts to the right dtype while avoiding copying
#        local_counts = np.array(local_counts, copy=False, dtype=result_dtype)
#        global_counts = np.empty_like(local_counts)
#        self._Allreduce_helper(local_counts,
#                               global_counts,
#                               MPI.SUM)
#        return global_counts
theos's avatar
theos committed
1635 1636 1637 1638 1639 1640 1641 1642 1643 1644