nifty_mpi_data.py 63.9 KB
Newer Older
ultimanet's avatar
ultimanet committed
1
# -*- coding: utf-8 -*-
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
## NIFTY (Numerical Information Field Theory) has been developed at the
## Max-Planck-Institute for Astrophysics.
##
## Copyright (C) 2015 Max-Planck-Society
##
## Author: Theo Steininger
## Project homepage: <http://www.mpa-garching.mpg.de/ift/nifty/>
##
## 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/>.


ultimanet's avatar
ultimanet committed
24
25
26
27
28

##initialize the 'found-packages'-dictionary 
found = {}

import numpy as np
Ultimanet's avatar
Ultimanet committed
29
from nifty_about import about
ultimanet's avatar
ultimanet committed
30
31

try:
32
    from mpi4py import MPI
ultimanet's avatar
ultimanet committed
33
34
    found[MPI] = True
except(ImportError): 
35
    import mpi_dummy as MPI
ultimanet's avatar
ultimanet committed
36
37
38
39
40
41
42
43
44
    found[MPI] = False

try:
    import pyfftw
    found['pyfftw'] = True
except(ImportError):       
    found['pyfftw'] = False

try:
45
    import h5py
ultimanet's avatar
ultimanet committed
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
    found['h5py'] = True
    found['h5py_parallel'] = h5py.get_config().mpi
except(ImportError):
    found['h5py'] = False
    found['h5py_parallel'] = False

   


class distributed_data_object(object):
    """

        NIFTY class for distributed data

        Parameters
        ----------
        global_data : {tuple, list, numpy.ndarray} *at least 1-dimensional*
            Initial data which will be casted to a numpy.ndarray and then 
            stored according to the distribution strategy. The global_data's
            shape overwrites global_shape.
        global_shape : tuple of ints, *optional*
            If no global_data is supplied, global_shape can be used to
            initialize an empty distributed_data_object
        dtype : type, *optional*
            If an explicit dtype is supplied, the given global_data will be 
            casted to it.            
        distribution_strategy : {'fftw' (default), 'not'}, *optional*
            Specifies the way, how global_data will be distributed to the 
            individual nodes. 
            'fftw' follows the distribution strategy of pyfftw.
            'not' does not distribute the data at all. 
            

        Attributes
        ----------
        data : numpy.ndarray
            The numpy.ndarray in which the individual node's data is stored.
        dtype : type
            Data type of the data object.
        distribution_strategy : string
            Name of the used distribution_strategy
        distributor : distributor
            The distributor object which takes care of all distribution and 
            consolidation of the data. 
        shape : tuple of int
            The global shape of the data
            
        Raises
        ------
        TypeError : 
            If the supplied distribution strategy is not known. 
        
    """
Ultimanet's avatar
Ultimanet committed
99
100
101
    def __init__(self,  global_data=None, global_shape=None, dtype=None, 
                 distribution_strategy='fftw', hermitian=False, 
                 *args, **kwargs):
ultimanet's avatar
ultimanet committed
102
        if global_data != None:
Ultimanet's avatar
Ultimanet committed
103
            if np.isscalar(global_data):
104
105
106
107
                global_data_input = None
                dtype = np.array(global_data).dtype.type
            else:
                global_data_input = np.array(global_data, copy=True, order='C')
ultimanet's avatar
ultimanet committed
108
109
        else:
            global_data_input = None
110

Ultimanet's avatar
Ultimanet committed
111
112
113
114
115
116
        self.hermitian = False

        self.distributor = self._get_distributor(distribution_strategy)(
                            global_data=global_data_input, 
                            global_shape=global_shape, 
                            dtype=dtype, **kwargs)
Ultimanet's avatar
Ultimanet committed
117

Ultimanet's avatar
Ultimanet committed
118
119
        self.set_full_data(data=global_data_input, hermitian=hermitian, 
                           **kwargs)
ultimanet's avatar
ultimanet committed
120
        
121
            
ultimanet's avatar
ultimanet committed
122
123
124
125
        self.distribution_strategy = distribution_strategy
        self.dtype = self.distributor.dtype
        self.shape = self.distributor.global_shape
        
126
127
        self.init_args = args 
        self.init_kwargs = kwargs
128
129
        
        ## If the input data was a scalar, set the whole array to this value
Ultimanet's avatar
Ultimanet committed
130
        if global_data != None and np.isscalar(global_data):
Ultimanet's avatar
Ultimanet committed
131
132
133
            temp = np.empty(self.distributor.local_shape)
            temp.fill(global_data)
            self.set_local_data(temp)
134
            self.hermitian = True
135
        
Ultimanet's avatar
Ultimanet committed
136
137
138
139
140
141
142
143
144
    def copy(self, dtype=None, distribution_strategy=None, **kwargs):
        temp_d2o = self.copy_empty(dtype=dtype, 
                                   distribution_strategy=distribution_strategy, 
                                   **kwargs)     
        if distribution_strategy == None or \
            distribution_strategy == self.distribution_strategy:
            temp_d2o.set_local_data(self.get_local_data(), copy=True)
        else:
            temp_d2o.set_full_data(self.get_full_data())
145
        temp_d2o.hermitian = self.hermitian
146
147
        return temp_d2o
    
148
149
150
151
152
153
154
155
156
157
158
159
160
161
    def copy_empty(self, global_shape=None, dtype=None, 
                   distribution_strategy=None, **kwargs):
        if global_shape == None:
            global_shape = self.shape
        if dtype == None:
            dtype = self.dtype
        if distribution_strategy == None:
            distribution_strategy = self.distribution_strategy

        kwargs.update(self.init_kwargs)
        
        temp_d2o = distributed_data_object(global_shape=global_shape,
                                           dtype=dtype,
                                           distribution_strategy=distribution_strategy,
162
                                           *self.init_args,
163
                                           **kwargs)
164
165
        return temp_d2o
    
166
    def apply_scalar_function(self, function, inplace=False, dtype=None):
167
168
        remember_hermitianQ = self.hermitian
        
Ultimanet's avatar
Ultimanet committed
169
170
        if inplace == True:        
            temp = self
171
172
173
174
            if dtype != None and self.dtype != dtype:
                about.warnings.cprint(\
            "WARNING: Inplace dtype conversion is not possible!")
                
Ultimanet's avatar
Ultimanet committed
175
        else:
176
            temp = self.copy_empty(dtype=dtype)
Ultimanet's avatar
Ultimanet committed
177
178
179
180
181

        try: 
            temp.data[:] = function(self.data)
        except:
            temp.data[:] = np.vectorize(function)(self.data)
182
        
183
184
185
186
        if function in (np.exp, np.log):
            temp.hermitian = remember_hermitianQ
        else:
            temp.hermitian = False
Ultimanet's avatar
Ultimanet committed
187
188
189
190
191
192
        return temp
    
    def apply_generator(self, generator):
        self.set_local_data(generator(self.distributor.local_shape))
        self.hermitian = False
            
ultimanet's avatar
ultimanet committed
193
194
195
196
197
198
    def __str__(self):
        return self.data.__str__()
    
    def __repr__(self):
        return '<distributed_data_object>\n'+self.data.__repr__()
    
Ultimanet's avatar
Ultimanet committed
199
    def __eq__(self, other):
Ultimanet's avatar
Ultimanet committed
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
        result = self.copy_empty(dtype = np.bool)
        ## Case 1: 'other' is a scalar
        ## -> make point-wise comparison
        if np.isscalar(other):
            result.set_local_data(self.get_local_data(copy = False) == other)
            return result        

        ## Case 2: 'other' is a numpy array or a distributed_data_object
        ## -> extract the local data and make point-wise comparison
        elif isinstance(other, np.ndarray) or\
        isinstance(other, distributed_data_object):
            temp_data = self.distributor.extract_local_data(other)
            result.set_local_data(self.get_local_data(copy=False) == temp_data)
            return result
        
        ## Case 3: 'other' is None
        elif other == None:
            return False
        
        ## Case 4: 'other' is something different
        ## -> make a numpy casting and make a recursion
        else:
            temp_other = np.array(other)
            return self.__eq__(temp_other)
            
            
        
    
    def equal(self, other):
Ultimanet's avatar
Ultimanet committed
229
230
231
232
233
234
235
236
237
        if other is None:
            return False
        try:
            assert(self.dtype == other.dtype)
            assert(self.shape == other.shape)
            assert(self.init_args == other.init_args)
            assert(self.init_kwargs == other.init_kwargs)
            assert(self.distribution_strategy == other.distribution_strategy)
            assert(np.all(self.data == other.data))
Ultimanet's avatar
Ultimanet committed
238
        except(AssertionError, AttributeError):
Ultimanet's avatar
Ultimanet committed
239
240
241
242
243
244
245
            return False
        else:
            return True
        

            
    
246
    def __pos__(self):
247
        temp_d2o = self.copy_empty()
248
249
250
        temp_d2o.set_local_data(data = self.get_local_data())
        return temp_d2o
        
ultimanet's avatar
ultimanet committed
251
    def __neg__(self):
252
        temp_d2o = self.copy_empty()
ultimanet's avatar
ultimanet committed
253
254
255
        temp_d2o.set_local_data(data = self.get_local_data().__neg__()) 
        return temp_d2o
    
256
    def __abs__(self):
Ultimanet's avatar
Ultimanet committed
257
258
259
260
261
262
263
264
265
266
267
268
        ## translate complex dtypes
        if self.dtype == np.complex64:
            new_dtype = np.float32
        elif self.dtype == np.complex128:
            new_dtype = np.float64
        elif self.dtype == np.complex:
            new_dtype = np.float
        elif issubclass(self.dtype, np.complexfloating):
            new_dtype = np.float
        else:
            new_dtype = self.dtype
        temp_d2o = self.copy_empty(dtype = new_dtype)
269
270
        temp_d2o.set_local_data(data = self.get_local_data().__abs__()) 
        return temp_d2o
ultimanet's avatar
ultimanet committed
271
            
272
    def __builtin_helper__(self, operator, other, inplace=False):
Ultimanet's avatar
Ultimanet committed
273
274
275
276
277
        ## Case 1: other is not a scalar
        if not (np.isscalar(other) or np.shape(other) == (1,)):
##            if self.shape != other.shape:            
##                raise AttributeError(about._errors.cstring(
##                    "ERROR: Shapes do not match!")) 
278
279
280
281
            try:            
                hermitian_Q = other.hermitian
            except(AttributeError):
                hermitian_Q = False
Ultimanet's avatar
Ultimanet committed
282
283
284
            ## extract the local data from the 'other' object
            temp_data = self.distributor.extract_local_data(other)
            temp_data = operator(temp_data)
Ultimanet's avatar
Ultimanet committed
285
            
286
287
288
289
        ## Case 2: other is a real scalar -> preserve hermitianity
        elif np.isreal(other) or (self.dtype not in (np.complex, np.complex128,
                                                np.complex256)):
            hermitian_Q = self.hermitian
ultimanet's avatar
ultimanet committed
290
            temp_data = operator(other)
291
292
293
294
        ## Case 3: other is complex
        else:
            hermitian_Q = False
            temp_data = operator(other)        
Ultimanet's avatar
Ultimanet committed
295
        ## write the new data into a new distributed_data_object        
296
297
298
299
        if inplace == True:
            temp_d2o = self
        else:
            temp_d2o = self.copy_empty()        
ultimanet's avatar
ultimanet committed
300
        temp_d2o.set_local_data(data=temp_data)
301
        temp_d2o.hermitian = hermitian_Q
ultimanet's avatar
ultimanet committed
302
        return temp_d2o
303
    """
Ultimanet's avatar
Ultimanet committed
304
    def __inplace_builtin_helper__(self, operator, other):
305
        ## Case 1: other is not a scalar
Ultimanet's avatar
Ultimanet committed
306
307
308
        if not (np.isscalar(other) or np.shape(other) == (1,)):        
            temp_data = self.distributor.extract_local_data(other)
            temp_data = operator(temp_data)
309
310
311
        ## Case 2: other is a real scalar -> preserve hermitianity
        elif np.isreal(other):
            hermitian_Q = self.hermitian
Ultimanet's avatar
Ultimanet committed
312
            temp_data = operator(other)
313
314
315
        ## Case 3: other is complex
        else:
            temp_data = operator(other)        
Ultimanet's avatar
Ultimanet committed
316
        self.set_local_data(data=temp_data)
317
        self.hermitian = hermitian_Q
Ultimanet's avatar
Ultimanet committed
318
        return self
319
    """ 
Ultimanet's avatar
Ultimanet committed
320
    
ultimanet's avatar
ultimanet committed
321
322
323
324
325
    def __add__(self, other):
        return self.__builtin_helper__(self.get_local_data().__add__, other)

    def __radd__(self, other):
        return self.__builtin_helper__(self.get_local_data().__radd__, other)
Ultimanet's avatar
Ultimanet committed
326
327

    def __iadd__(self, other):
328
329
330
        return self.__builtin_helper__(self.get_local_data().__iadd__, 
                                               other,
                                               inplace = True)
Ultimanet's avatar
Ultimanet committed
331

ultimanet's avatar
ultimanet committed
332
333
334
335
336
337
338
    def __sub__(self, other):
        return self.__builtin_helper__(self.get_local_data().__sub__, other)
    
    def __rsub__(self, other):
        return self.__builtin_helper__(self.get_local_data().__rsub__, other)
    
    def __isub__(self, other):
339
340
341
        return self.__builtin_helper__(self.get_local_data().__isub__, 
                                               other,
                                               inplace = True)
ultimanet's avatar
ultimanet committed
342
343
344
345
346
347
348
        
    def __div__(self, other):
        return self.__builtin_helper__(self.get_local_data().__div__, other)
    
    def __rdiv__(self, other):
        return self.__builtin_helper__(self.get_local_data().__rdiv__, other)

Ultimanet's avatar
Ultimanet committed
349
    def __idiv__(self, other):
350
351
352
        return self.__builtin_helper__(self.get_local_data().__idiv__, 
                                               other,
                                               inplace = True)
Ultimanet's avatar
Ultimanet committed
353

ultimanet's avatar
ultimanet committed
354
    def __floordiv__(self, other):
Ultimanet's avatar
Ultimanet committed
355
356
        return self.__builtin_helper__(self.get_local_data().__floordiv__, 
                                       other)    
ultimanet's avatar
ultimanet committed
357
    def __rfloordiv__(self, other):
Ultimanet's avatar
Ultimanet committed
358
359
360
        return self.__builtin_helper__(self.get_local_data().__rfloordiv__, 
                                       other)
    def __ifloordiv__(self, other):
361
362
363
        return self.__builtin_helper__(
                    self.get_local_data().__ifloordiv__, other,
                                               inplace = True)
ultimanet's avatar
ultimanet committed
364
365
366
367
368
369
370
371
    
    def __mul__(self, other):
        return self.__builtin_helper__(self.get_local_data().__mul__, other)
    
    def __rmul__(self, other):
        return self.__builtin_helper__(self.get_local_data().__rmul__, other)

    def __imul__(self, other):
372
373
374
        return self.__builtin_helper__(self.get_local_data().__imul__, 
                                               other,
                                               inplace = True)
Ultimanet's avatar
Ultimanet committed
375

ultimanet's avatar
ultimanet committed
376
377
378
379
380
381
382
    def __pow__(self, other):
        return self.__builtin_helper__(self.get_local_data().__pow__, other)
 
    def __rpow__(self, other):
        return self.__builtin_helper__(self.get_local_data().__rpow__, other)

    def __ipow__(self, other):
383
384
385
        return self.___builtin_helper__(self.get_local_data().__ipow__, 
                                               other,
                                               inplace = True)
Ultimanet's avatar
Ultimanet committed
386
   
387
388
    def __len__(self):
        return self.shape[0]
389
    
390
391
392
    def dim(self):
        return np.prod(self.shape)
        
393
394
395
396
397
398
399
400
    def vdot(self, other):
        if isinstance(other, distributed_data_object):        
            other = other.get_local_data()
        local_vdot = np.vdot(self.get_local_data(), other)
        local_vdot_list = self.distributor._allgather(local_vdot)
        global_vdot = np.sum(local_vdot_list)
        return global_vdot
            
Ultimanet's avatar
Ultimanet committed
401

402
    
ultimanet's avatar
ultimanet committed
403
    def __getitem__(self, key):
Ultimanet's avatar
Ultimanet committed
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
        ## Case 1: key is a boolean array.
        ## -> take the local data portion from key, use this for data 
        ## extraction, and then merge the result in a flat numpy array
        if 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:
            found = 'other'
                
        if (found == 'ndarray' or found == 'd2o') and found_boolean == True:
            ## extract the data of local relevance
            local_bool_array = self.distributor.extract_local_data(key)
            local_results = self.get_local_data(copy=False)[local_bool_array]
            global_results = self.distributor._allgather(local_results)
            global_results = np.concatenate(global_results)
            return global_results            
            
        else:
            return self.get_data(key)
ultimanet's avatar
ultimanet committed
426
427
428
429
    
    def __setitem__(self, key, data):
        self.set_data(data, key)
        
430
    def _contraction_helper(self, function, **kwargs):
431
432
433
434
435
436
        local = function(self.data, **kwargs)
        local_list = self.distributor._allgather(local)
        global_ = function(local_list, axis=0)
        return global_
        
    def amin(self, **kwargs):
437
        return self._contraction_helper(np.amin, **kwargs)
438
439

    def nanmin(self, **kwargs):
440
        return self._contraction_helper(np.nanmin, **kwargs)
441
442
        
    def amax(self, **kwargs):
443
        return self._contraction_helper(np.amax, **kwargs)
444
445
    
    def nanmax(self, **kwargs):
446
        return self._contraction_helper(np.nanmax, **kwargs)
Ultimanet's avatar
Ultimanet committed
447
    
448
449
450
451
452
453
    def sum(self, **kwargs):
        return self._contraction_helper(np.sum, **kwargs)

    def prod(self, **kwargs):
        return self._contraction_helper(np.prod, **kwargs)        
        
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
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
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
    def mean(self, power=1):
        ## compute the local means and the weights for the mean-mean. 
        local_mean = np.mean(self.data**power)
        local_weight = np.prod(self.data.shape)
        ## collect the local means and cast the result to a ndarray
        local_mean_weight_list = self.distributor._allgather((local_mean, 
                                                              local_weight))
        local_mean_weight_list =np.array(local_mean_weight_list)   
        ## compute the denominator for the weighted mean-mean                                                           
        global_weight = np.sum(local_mean_weight_list[:,1])
        ## compute the numerator
        numerator = np.sum(local_mean_weight_list[:,0]*\
            local_mean_weight_list[:,1])
        global_mean = numerator/global_weight
        return global_mean

    def var(self):
        mean_of_the_square = self.mean(power=2)
        square_of_the_mean = self.mean()**2
        return mean_of_the_square - square_of_the_mean
    
    def std(self):
        return np.sqrt(self.var())
        
    def _argmin_argmax_flat_helper(self, function):
        local_argmin = function(self.data)
        local_argmin_value = self.data[np.unravel_index(local_argmin, 
                                                        self.data.shape)]
        globalized_local_argmin = self.distributor.globalize_flat_index(local_argmin)                                                       
        local_argmin_list = self.distributor._allgather((local_argmin_value, 
                                                         globalized_local_argmin))
        local_argmin_list = np.array(local_argmin_list, dtype=[('value', int),
                                                               ('index', int)])    
        return local_argmin_list
        
    def argmin_flat(self):
        local_argmin = np.argmin(self.data)
        local_argmin_value = self.data[np.unravel_index(local_argmin, 
                                                        self.data.shape)]
        globalized_local_argmin = self.distributor.globalize_flat_index(local_argmin)                                                       
        local_argmin_list = self.distributor._allgather((local_argmin_value, 
                                                         globalized_local_argmin))
        local_argmin_list = np.array(local_argmin_list, dtype=[('value', int),
                                                               ('index', int)])    
        local_argmin_list = np.sort(local_argmin_list, order=['value', 'index'])        
        return local_argmin_list[0][1]
    
    def argmax_flat(self):
        local_argmax = np.argmax(self.data)
        local_argmax_value = -self.data[np.unravel_index(local_argmax, 
                                                        self.data.shape)]
        globalized_local_argmax = self.distributor.globalize_flat_index(local_argmax)                                                       
        local_argmax_list = self.distributor._allgather((local_argmax_value, 
                                                         globalized_local_argmax))
        local_argmax_list = np.array(local_argmax_list, dtype=[('value', int),
                                                               ('index', int)])         
        return local_argmax_list[0][1]
        

    def argmin(self):    
        return np.unravel_index(self.argmin_flat(), self.shape)
    
    def argmax(self):
        return np.unravel_index(self.argmax_flat(), self.shape)
    
    def conjugate(self):
        temp_d2o = self.copy_empty()
        temp_data = np.conj(self.get_local_data())
        temp_d2o.set_local_data(temp_data)
        return temp_d2o

    
    def conj(self):
        return self.conjugate()      
        
    def median(self):
Ultimanet's avatar
Ultimanet committed
530
        about.warnings.cprint(\
531
532
533
534
            "WARNING: The current implementation of median is very expensive!")
        median = np.median(self.get_full_data())
        return median
        
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
    def iscomplex(self):
        temp_d2o = self.copy_empty(dtype=bool)
        temp_d2o.set_local_data(np.iscomplex(self.data))
        return temp_d2o
    
    def isreal(self):
        temp_d2o = self.copy_empty(dtype=bool)
        temp_d2o.set_local_data(np.isreal(self.data))
        return temp_d2o
    
    def is_completely_real(self):
        local_realiness = np.all(self.isreal())
        global_realiness = self.distributor._allgather(local_realiness)
        return np.all(global_realiness)
    
Ultimanet's avatar
Ultimanet committed
550
    def set_local_data(self, data, hermitian=False, copy=False):
ultimanet's avatar
ultimanet committed
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
        """
            Stores data directly in the local data attribute. No distribution 
            is done. The shape of the data must fit the local data attributes
            shape.

            Parameters
            ----------
            data : tuple, list, numpy.ndarray 
                The data which should be stored in the local data attribute.
            
            Returns
            -------
            None
        
        """
Ultimanet's avatar
Ultimanet committed
566
567
        self.hermitian = hermitian
        self.data = np.array(data, dtype=self.dtype, copy=copy, order='C')
ultimanet's avatar
ultimanet committed
568
    
Ultimanet's avatar
Ultimanet committed
569
    def set_data(self, data, key, hermitian=False, *args, **kwargs):
ultimanet's avatar
ultimanet committed
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
        """
            Stores the supplied data in the region which is specified by key. 
            The data is distributed according to the distribution strategy. If
            the individual nodes get different key-arguments. Their data is 
            processed one-by-one.
            
            Parameters
            ----------
            data : tuple, list, numpy.ndarray 
                The data which should be distributed.
            key : int, slice, tuple of int or slice
                The key is the object which specifies the region, where data 
                will be stored in.                
            
            Returns
            -------
            None
        
        """
Ultimanet's avatar
Ultimanet committed
589
        self.hermitian = hermitian
ultimanet's avatar
ultimanet committed
590
        (slices, sliceified) = self.__sliceify__(key)        
Ultimanet's avatar
Ultimanet committed
591
592
593
594
        self.distributor.disperse_data(data=self.data, 
                        to_slices = slices,
                        data_update = self.__enfold__(data, sliceified), 
                        *args, **kwargs)        
ultimanet's avatar
ultimanet committed
595
    
Ultimanet's avatar
Ultimanet committed
596
    def set_full_data(self, data, hermitian=False, **kwargs):
ultimanet's avatar
ultimanet committed
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
        """
            Distributes the supplied data to the nodes. The shape of data must 
            match the shape of the distributed_data_object.
            
            Parameters
            ----------
            data : tuple, list, numpy.ndarray 
                The data which should be distributed.
            
            Notes
            -----
            set_full_data(foo) is equivalent to set_data(foo,slice(None)) but 
            faster.
        
            Returns
            -------
            None
        
        """
Ultimanet's avatar
Ultimanet committed
616
        self.hermitian = hermitian
617
        self.data = self.distributor.distribute_data(data=data, **kwargs)
ultimanet's avatar
ultimanet committed
618
619
    

Ultimanet's avatar
Ultimanet committed
620
    def get_local_data(self, key=(slice(None),), copy=True):
ultimanet's avatar
ultimanet committed
621
622
623
624
625
626
627
628
629
630
631
632
633
        """
            Loads data directly from the local data attribute. No consolidation 
            is done. 

            Parameters
            ----------
            key : int, slice, tuple of int or slice
                The key which will be used to access the data. 
            
            Returns
            -------
            self.data[key] : numpy.ndarray
        
Ultimanet's avatar
Ultimanet committed
634
        """
Ultimanet's avatar
Ultimanet committed
635
636
637
638
        if copy == True:
            return self.data[key]        
        if copy == False:
            return self.data
ultimanet's avatar
ultimanet committed
639
        
640
    def get_data(self, key, **kwargs):
ultimanet's avatar
ultimanet committed
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
        """
            Loads data from the region which is specified by key. The data is 
            consolidated according to the distribution strategy. If the 
            individual nodes get different key-arguments, they get individual
            data. 
            
            Parameters
            ----------
        
            key : int, slice, tuple of int or slice
                The key is the object which specifies the region, where data 
                will be loaded from.                 
            
            Returns
            -------
            global_data[key] : numpy.ndarray
        
        """
659
660
        (slices, sliceified) = self.__sliceify__(key)
        result = self.distributor.collect_data(self.data, slices, **kwargs)        
ultimanet's avatar
ultimanet committed
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
        return self.__defold__(result, sliceified)
        
    
    
    def get_full_data(self, target_rank='all'):
        """
            Fully consolidates the distributed data. 
            
            Parameters
            ----------
            target_rank : 'all' (default), int *optional*
                If only one node should recieve the full data, it can be 
                specified here.
            
            Notes
            -----
            get_full_data() is equivalent to get_data(slice(None)) but 
            faster.
        
            Returns
            -------
            None
        """

        return self.distributor.consolidate_data(self.data, target_rank)

Ultimanet's avatar
Ultimanet committed
687
688
689
690
691
692
693
694
    def inject(self, to_slices=(slice(None),), data=None, 
               from_slices=(slice(None),)):
        if data == None:
            return self
        
        self.distributor.inject(self.data, to_slices, data, from_slices)
        
        
ultimanet's avatar
ultimanet committed
695
696
697
698
699
700
    def _get_distributor(self, distribution_strategy):
        '''
            Comments:
              - The distributor's get_data and set_data functions MUST be 
                supplied with a tuple of slice objects. In case that there was 
                a direct integer involved, the unfolding will be done by the
701
                helper functions __sliceify__, __enfold__ and __defold__.
ultimanet's avatar
ultimanet committed
702
703
704
705
706
707
708
        '''
        
        distributor_dict={
            'fftw':     _fftw_distributor,
            'not':      _not_distributor
        }
        if not distributor_dict.has_key(distribution_strategy):
Ultimanet's avatar
Ultimanet committed
709
            raise TypeError(about._errors.cstring("ERROR: Unknown distribution strategy supplied."))
ultimanet's avatar
ultimanet committed
710
711
712
713
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
        return distributor_dict[distribution_strategy]
      
    def save(self, alias, path=None, overwriteQ=True):
        
        """
            Saves a distributed_data_object to disk utilizing h5py.
            
            Parameters
            ----------
            alias : string
                The name for the dataset which is saved within the hdf5 file.
         
            path : string *optional*
                The path to the hdf5 file. If no path is given, the alias is 
                taken as filename in the current path.
            
            overwriteQ : Boolean *optional*
                Specifies whether a dataset may be overwritten if it is already
                present in the given hdf5 file or not.
        """
        self.distributor.save_data(self.data, alias, path, overwriteQ)

    def load(self, alias, path=None):
        """
            Loads a distributed_data_object from disk utilizing h5py.
            
            Parameters
            ----------
            alias : string
                The name of the dataset which is loaded from the hdf5 file.
 
            path : string *optional*
                The path to the hdf5 file. If no path is given, the alias is 
                taken as filename in the current path.
        """
        self.data = self.distributor.load_data(alias, path)
           
    def __sliceify__(self, inp):
        sliceified = []
        result = []
        if isinstance(inp, tuple):
            x = inp
Ultimanet's avatar
Ultimanet committed
752
753
        elif isinstance(inp, list):
            x = tuple(inp)
ultimanet's avatar
ultimanet committed
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
        else:
            x = (inp, )
        
        for i in range(len(x)):
            if isinstance(x[i], slice):
                result += [x[i], ]
                sliceified += [False, ]
            else:
                result += [slice(x[i], x[i]+1), ]
                sliceified += [True, ]
    
        return (tuple(result), sliceified)
                
                
    def __enfold__(self, in_data, sliceified):
        data = np.array(in_data, copy=False)    
        temp_shape = ()
        j=0
        for i in sliceified:
            if i == True:
                temp_shape += (1,)
775
776
777
778
779
                try:
                    if data.shape[j] == 1:
                        j +=1
                except(IndexError):
                    pass
ultimanet's avatar
ultimanet committed
780
            else:
781
782
783
784
                try:
                    temp_shape += (data.shape[j],)
                except(IndexError):
                    temp_shape += (1,)
ultimanet's avatar
ultimanet committed
785
786
787
788
789
790
                j += 1
        ## take into account that the sliceified tuple may be too short, because 
        ## of a non-exaustive list of slices
        for i in range(len(data.shape)-j):
            temp_shape += (data.shape[j],)
            j += 1
Ultimanet's avatar
Ultimanet committed
791
        
ultimanet's avatar
ultimanet committed
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
        return data.reshape(temp_shape)
    
    def __defold__(self, data, sliceified):
        temp_slice = ()
        for i in sliceified:
            if i == True:
                temp_slice += (0,)
            else:
                temp_slice += (slice(None),)
        return data[temp_slice]

    

   
class _fftw_distributor(object):
807
808
    def __init__(self, global_data=None, global_shape=None, dtype=None, 
                 comm=MPI.COMM_WORLD, alias=None, path=None):
ultimanet's avatar
ultimanet committed
809
810
811
812
813
814
815
816
817
818
819
        
        if alias != None:
            file_path = path if path != None else alias 
            if found['h5py_parallel']:
                f = h5py.File(file_path, 'r', driver='mpio', comm=comm)
            else:
                f= h5py.File(file_path, 'r')        
            dset = f[alias]        

        
        if comm.rank == 0:        
Ultimanet's avatar
Ultimanet committed
820
            ## Case 1: hdf5 path supplied
ultimanet's avatar
ultimanet committed
821
822
            if alias != None:
                self.global_shape = dset.shape
Ultimanet's avatar
Ultimanet committed
823
824
825
826
            ## Case 2: no hdf5 path supplied
            else:           
                ## subcase 1: input data is scalar or None
                if global_data == None or np.isscalar(global_data):
ultimanet's avatar
ultimanet committed
827
                    if global_shape == None:
Ultimanet's avatar
Ultimanet committed
828
                        raise TypeError(about._errors.\
Ultimanet's avatar
Ultimanet committed
829
                cstring("ERROR: Neither non-scalar data nor shape supplied!"))
ultimanet's avatar
ultimanet committed
830
831
                    else:
                        self.global_shape = global_shape
Ultimanet's avatar
Ultimanet committed
832
833
                ## subcase 2: input data is non-scalar 
                ## -> Take the shape of the input data
ultimanet's avatar
ultimanet committed
834
835
836
837
                else:
                    self.global_shape = global_data.shape
        else:
            self.global_shape = None
Ultimanet's avatar
Ultimanet committed
838
            
ultimanet's avatar
ultimanet committed
839
840
841
842
843
844
845
846
847
848
849
850
        self.global_shape = comm.bcast(self.global_shape, root = 0)
        self.global_shape = tuple(self.global_shape)
        
        if comm.rank == 0:        
            if alias != None:
                self.dtype = dset.dtype.type
            else:    
                if dtype != None:        
                    self.dtype = dtype
                elif global_data != None:
                    self.dtype = np.array(global_data).dtype.type
                else:
Ultimanet's avatar
Ultimanet committed
851
                    raise TypeError(about._errors.\
852
853
                    cstring("ERROR: Failed setting datatype. Neither data, "+\
                     "nor datatype supplied."))
ultimanet's avatar
ultimanet committed
854
855
856
857
858
859
860
861
862
        else:
            self.dtype=None
        self.dtype = comm.bcast(self.dtype, root=0)
        if alias != None:        
            f.close()        
        
        self._my_dtype_converter = dtype_converter()
        
        if not self._my_dtype_converter.known_np_Q(self.dtype):
Ultimanet's avatar
Ultimanet committed
863
            raise TypeError(about._errors.cstring(\
864
            "ERROR: The datatype "+str(self.dtype)+" is not known to mpi4py."))
ultimanet's avatar
ultimanet committed
865
866
867
868
869
870
871
872
873
874

        self.mpi_dtype  = self._my_dtype_converter.to_mpi(self.dtype)
        
        self._local_size = pyfftw.local_size(self.global_shape)
        self.local_start = self._local_size[2]
        self.local_end = self.local_start + self._local_size[1]
        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)
875
876
        comm.Allgather([np.array(self.local_dim,dtype=np.int), MPI.INT],\
            [self.local_dim_list, MPI.INT])
ultimanet's avatar
ultimanet committed
877
878
        self.local_dim_offset = np.sum(self.local_dim_list[0:comm.rank])
        
879
880
881
        self.local_slice = np.array([self.local_start, self.local_end,\
            self.local_length, self.local_dim, self.local_dim_offset],\
            dtype=np.int)
ultimanet's avatar
ultimanet committed
882
883
884
        ## collect all local_slices 
        ## [start, stop, length=stop-start, dimension, dimension_offset]
        self.all_local_slices = np.empty((comm.size,5),dtype=np.int)
885
886
        comm.Allgather([np.array((self.local_slice,),dtype=np.int), MPI.INT],\
            [self.all_local_slices, MPI.INT])
ultimanet's avatar
ultimanet committed
887
        
888
        self.comm = comm
ultimanet's avatar
ultimanet committed
889
        
890
891
892
893
894
895
    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),):
Ultimanet's avatar
Ultimanet committed
896
            raise TypeError(about._errors.cstring("ERROR: Length\
897
898
899
900
901
902
903
904
905
                of index tuple does not match the array's shape!"))                 
        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):
Ultimanet's avatar
Ultimanet committed
906
            about.warnings.cprint("WARNING: Indices were clipped!")
907
908
909
910
911
912
913
914
915
916
917
        globalized_index = tuple(globalized_index)
        return globalized_index
    
    def _allgather(self, thing, comm=None):
        if comm == None:
            comm = self.comm            
        gathered_things = comm.allgather(thing)
        return gathered_things
    
    def distribute_data(self, data=None, comm = None, alias=None,
                        path=None, **kwargs):
ultimanet's avatar
ultimanet committed
918
919
920
921
922
        '''
        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
        '''
923
924
        if comm == None:
            comm = self.comm            
925
926
927
928
        rank = comm.Get_rank()
        size = comm.Get_size()        
        local_data_available_Q = np.array((int(data != None), ))
        data_available_Q = np.empty(size,dtype=int)
929
930
        comm.Allgather([local_data_available_Q, MPI.INT], 
                       [data_available_Q, MPI.INT])        
931
932
        
        if data_available_Q[0]==False and found['h5py']:
ultimanet's avatar
ultimanet committed
933
934
935
936
937
938
939
            try: 
                file_path = path if path != None else alias 
                if found['h5py_parallel']:
                    f = h5py.File(file_path, 'r', driver='mpio', comm=comm)
                else:
                    f= h5py.File(file_path, 'r')        
                dset = f[alias]
940
941
                if dset.shape == self.global_shape and \
                 dset.dtype.type == self.dtype:
ultimanet's avatar
ultimanet committed
942
943
944
945
                    temp_data = dset[self.local_start:self.local_end]
                    f.close()
                    return temp_data
                else:
Ultimanet's avatar
Ultimanet committed
946
                    raise TypeError(about._errors.cstring("ERROR: \
947
                    Input data has the wrong shape or wrong dtype!"))                 
ultimanet's avatar
ultimanet committed
948
949
950
            except(IOError, AttributeError):
                pass
            
951
        if np.all(data_available_Q==False):
Ultimanet's avatar
Ultimanet committed
952
            return np.empty(self.local_shape, dtype=self.dtype, order='C')
ultimanet's avatar
ultimanet committed
953
        ## if all nodes got data, we assume that it is the right data and 
954
955
        ## store it individually. If not, take the data on node 0 and scatter 
        ## it...
ultimanet's avatar
ultimanet committed
956
        if np.all(data_available_Q):
957
958
            return data[self.local_start:self.local_end].astype(self.dtype,\
                copy=False)    
959
960
        ## ... but only if node 0 has actually data!
        elif data_available_Q[0] == False:# or np.all(data_available_Q==False):
Ultimanet's avatar
Ultimanet committed
961
            return np.empty(self.local_shape, dtype=self.dtype, order='C')
962
        
ultimanet's avatar
ultimanet committed
963
964
965
966
967
        else:
            if data == None:
                data = np.empty(self.global_shape)            
            if rank == 0:
                if np.all(data.shape != self.global_shape):
Ultimanet's avatar
Ultimanet committed
968
                    raise TypeError(about._errors.cstring(\
969
                        "ERROR: Input data has the wrong shape!"))
ultimanet's avatar
ultimanet committed
970
            ## Scatter the data!            
Ultimanet's avatar
Ultimanet committed
971
            _scattered_data = np.empty(self.local_shape, dtype = self.dtype)
ultimanet's avatar
ultimanet committed
972
973
            _dim_list = self.all_local_slices[:,3]
            _dim_offset_list = self.all_local_slices[:,4]
974
975
            comm.Scatterv([data, _dim_list, _dim_offset_list, self.mpi_dtype],\
                [_scattered_data, self.mpi_dtype], root=0)
ultimanet's avatar
ultimanet committed
976
977
978
            return _scattered_data
        return None
    
Ultimanet's avatar
Ultimanet committed
979
980
    def _disperse_data_primitive(self, data, to_slices, data_update, 
                                 from_slices, source_rank='all', comm=None):
981
982
        if comm == None:
            comm = self.comm            
983
984
        ## compute the part of the slice which is relevant for the 
        ## individual node      
ultimanet's avatar
ultimanet committed
985
        localized_start, localized_stop = self._backshift_and_decycle(
Ultimanet's avatar
Ultimanet committed
986
            to_slices[0], self.local_start, self.local_end,\
987
988
                self.global_shape[0])
        local_slice = (slice(localized_start, localized_stop,\
Ultimanet's avatar
Ultimanet committed
989
                        to_slices[0].step),) + to_slices[1:]
ultimanet's avatar
ultimanet committed
990
991
992
993
994
        
        ## compute the parameter sets and list for the data splitting
        local_slice_shape = data[local_slice].shape        
        local_affected_data_length = local_slice_shape[0]
        local_affected_data_length_list=np.empty(comm.size, dtype=np.int)        
995
996
997
998
999
        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)[:-1])
ultimanet's avatar
ultimanet committed
1000
1001
1002
1003
1004
1005
1006
1007
        
        
        if source_rank == 'all':
            ## only take the relevant part out of data_update and plug it into 
            ## data[local_slice]
            r = comm.rank
            o = local_affected_data_length_offset_list
            l = local_affected_data_length
Ultimanet's avatar
Ultimanet committed
1008
1009
1010
1011
1012
1013
            
            ## if the from_slices object is not None, i.e. only a part from
            ## the data source is used, form the update_slice accordingly
            if from_slices == None:
                update_slice = (slice(o[r], o[r]+l),)
            else:
Ultimanet's avatar
Ultimanet committed
1014
                    
Ultimanet's avatar
Ultimanet committed
1015
1016
1017
                f_step = from_slices[0].step
                if f_step == None:
                    f_step = 1
Ultimanet's avatar
Ultimanet committed
1018
                    
Ultimanet's avatar
Ultimanet committed
1019
                f_direction = np.sign(f_step)
Ultimanet's avatar
Ultimanet committed
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032

                f_relative_start = from_slices[0].start
                if f_relative_start != None:
                    f_start = f_relative_start + f_direction*o[r]
                else:
                    f_start = None
                    f_relative_start = 0
                    
                f_stop = f_relative_start + f_direction*(o[r]+l*np.abs(f_step))
                if f_stop < 0:
                    f_stop = None


Ultimanet's avatar
Ultimanet committed
1033
                ## combine the slicing for the first dimension 
Ultimanet's avatar
Ultimanet committed
1034
1035
                update_slice = (slice(f_start,
                                      f_stop,
Ultimanet's avatar
Ultimanet committed
1036
1037
1038
1039
                                      f_step),
                                )
                ## add the rest of the from_slicing
                update_slice += from_slices[1:]
Ultimanet's avatar
Ultimanet committed
1040

1041
1042
            data[local_slice] = np.array(data_update[update_slice],\
                                    copy=False).astype(self.dtype)
ultimanet's avatar
ultimanet committed
1043
1044
1045
1046
            
        else:
            ## Scatterv the relevant part from the source_rank to the others 
            ## and plug it into data[local_slice]
1047
1048
1049
            
            ## if the first slice object has a negative step size, the ordering 
            ## of the Scatterv function must be reversed         
Ultimanet's avatar
Ultimanet committed
1050
            order = to_slices[0].step
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
            if order == None:
                order = 1
            else:
                order = np.sign(order)

            local_affected_data_dim_list = \
                np.array(local_affected_data_length_list) *\
                    np.product(local_slice_shape[1:])                    

            local_affected_data_dim_offset_list = np.append([0],\
                np.cumsum(local_affected_data_dim_list[::order])[:-1])[::order]
                
            local_dispersed_data = np.zeros(local_slice_shape,\
                dtype=self.dtype)
            comm.Scatterv(\
Ultimanet's avatar
Ultimanet committed
1066
1067
                [np.array(data_update[from_slices],copy=False).\
                                                        astype(self.dtype),\
1068
1069
                    local_affected_data_dim_list,\
                    local_affected_data_dim_offset_list, self.mpi_dtype],
ultimanet's avatar
ultimanet committed
1070
1071
1072
1073
1074
1075
1076
                          [local_dispersed_data, self.mpi_dtype], 
                          root=source_rank)                            
            data[local_slice] = local_dispersed_data
        return None
        
    
    
Ultimanet's avatar
Ultimanet committed
1077
1078
    def disperse_data(self, data, to_slices, data_update, from_slices=None,
                      comm=None, **kwargs):
1079
1080
        if comm == None:
            comm = self.comm            
Ultimanet's avatar
Ultimanet committed
1081
        to_slices_list = comm.allgather(to_slices)
ultimanet's avatar
ultimanet committed
1082
        ## check if all slices are the same. 
Ultimanet's avatar
Ultimanet committed
1083
        if all(x == to_slices_list[0] for x in to_slices_list):
ultimanet's avatar
ultimanet committed
1084
1085
            ## in this case, the _disperse_data_primitive can simply be called 
            ##with target_rank = 'all'
Ultimanet's avatar
Ultimanet committed
1086
1087
1088
1089
1090
1091
            self._disperse_data_primitive(data = data, 
                                          to_slices = to_slices,
                                          data_update=data_update,
                                          from_slices=from_slices, 
                                          source_rank='all', 
                                          comm=comm)
1092
1093
        ## if the different nodes got different slices, disperse the data 
        ## individually
ultimanet's avatar
ultimanet committed
1094
1095
        else:
            i = 0        
Ultimanet's avatar
Ultimanet committed
1096
            for temp_to_slices in to_slices_list:
ultimanet's avatar
ultimanet committed
1097
                ## make the collect_data call on all nodes            
Ultimanet's avatar
Ultimanet committed
1098
1099
1100
1101
1102
1103
                self._disperse_data_primitive(data=data,
                                              to_slices=temp_to_slices,
                                              data_update=data_update,
                                              from_slices=from_slices,
                                              source_rank=i, 
                                              comm=comm)
ultimanet's avatar
ultimanet committed
1104
1105
1106
                i += 1
                 
        
1107
1108
1109
1110
    def _collect_data_primitive(self, data, slice_objects, target_rank='all', comm=None):
        if comm == None:
            comm = self.comm            
            
ultimanet's avatar
ultimanet committed
1111
        localized_start, localized_stop = self._backshift_and_decycle(
1112
            slice_objects[0], self.local_start, self.local_end, self.global_shape[0])
ultimanet's avatar
ultimanet committed
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
        local_slice = (slice(localized_start,localized_stop,slice_objects[0].step),)+slice_objects[1:]
        local_collected_data = np.ascontiguousarray(data[local_slice])

        local_collected_data_length = local_collected_data.shape[0]
        local_collected_data_length_list=np.empty(comm.size, dtype=np.int)        
        comm.Allgather([np.array(local_collected_data_length, dtype=np.int), MPI.INT], [local_collected_data_length_list, MPI.INT])        
             
        collected_data_length = np.sum(local_collected_data_length_list) 
        collected_data_shape = (collected_data_length,)+local_collected_data.shape[1:]
        local_collected_data_dim_list= np.array(local_collected_data_length_list) * np.product(local_collected_data.shape[1:])        
        
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
        ## if the first slice object has a negative step size, the ordering 
        ## of the Gatherv functions must be reversed         
        order = slice_objects[0].step
        if order == None:
            order = 1
        else:
            order = np.sign(order)
            
        local_collected_data_dim_offset_list = np.append([0],np.cumsum(local_collected_data_dim_list[::order])[:-1])[::order]

        local_collected_data_dim_offset_list = local_collected_data_dim_offset_list
ultimanet's avatar
ultimanet committed
1135
        collected_data = np.empty(collected_data_shape, dtype=self.dtype)
1136
        
ultimanet's avatar
ultimanet committed
1137
1138
1139
1140
1141
1142
1143
1144
1145

        if target_rank == 'all':
            comm.Allgatherv([local_collected_data, self.mpi_dtype], 
                         [collected_data, local_collected_data_dim_list, local_collected_data_dim_offset_list, self.mpi_dtype])                
        else:
            comm.Gatherv([local_collected_data, self.mpi_dtype], 
                         [collected_data, local_collected_data_dim_list, local_collected_data_dim_offset_list, self.mpi_dtype], root=target_rank)                            
        return collected_data

1146
1147
1148
    def collect_data(self, data, slice_objects, comm=None, **kwargs):
        if comm == None:
            comm = self.comm                    
ultimanet's avatar
ultimanet committed
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
        slice_objects_list = comm.allgather(slice_objects)
        ## check if all slices are the same. 
        if all(x == slice_objects_list[0] for x in slice_objects_list):
            ## in this case, the _collect_data_primitive can simply be called 
            ##with target_rank = 'all'
            return self._collect_data_primitive(data=data, slice_objects=slice_objects, target_rank='all', comm=comm)
        
        ## if the different nodes got different slices, collect the data individually
        i = 0        
        for temp_slices in slice_objects_list:
            ## make the collect_data call on all nodes            
            temp_data = self._collect_data_primitive(data=data, slice_objects=temp_slices, target_rank=i, comm=comm)
            ## save the result only on the pulling node            
            if comm.rank == i:
                individual_data = temp_data
            i += 1
        return individual_data
        
    
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
    def _backshift_and_decycle(self, slice_object, shifted_start, shifted_stop, global_length):
        ## Crop the start value
        if slice_object.start > global_length-1:
            slice_object = slice(global_length-1, slice_object.stop,
                                 slice_object.step)
                                 
        ## Reformulate negative indices                                  
        if slice_object.start < 0 and slice_object.start != None:
            temp_start = slice_object.start + global_length
            if temp_start < 0:
Ultimanet's avatar
Ultimanet committed
1178
                raise ValueError(about._errors.cstring(\
1179
1180
1181
1182
1183
1184
1185
                "ERROR: Index is out of bounds!"))
            slice_object = slice(temp_start, slice_object.stop,\
            slice_object.step) 

        if slice_object.stop < 0 and slice_object.stop != None:
            temp_stop = slice_object.stop + global_length
            if temp_stop < 0:
Ultimanet's avatar
Ultimanet committed
1186
                raise ValueError(about._errors.cstring(\
1187
1188
1189
1190
1191
                "ERROR: Index is out of bounds!"))
            slice_object = slice(slice_object.start, temp_stop,\
            slice_object.step) 
                
        ## initialize the step
ultimanet's avatar
ultimanet committed
1192
1193
1194
1195
        if slice_object.step == None:
            step = 1
        else:
            step = slice_object.step
1196
1197
1198
1199
1200
1201
1202
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
        
        if step > 0:
            shift = shifted_start
            ## calculate the start index
            if slice_object.start == 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
            ## calculate the stop index
            if slice_object.stop == None:
                local_stop = None
            else:
                local_stop = slice_object.stop - shift
                ## if local_stop is negative, pull it up to zero
                local_stop = 0 if local_stop < 0 else local_stop
                
        else: # if step < 0
            step = -step
            local_length = shifted_stop - shifted_start
            ## calculate the start index. (Here, local_start > local_stop!)
            if slice_object.start == None:
                local_start = (local_length-1) -\
                    (global_length-shifted_stop)%step #stepsize compensation
            else:
                local_start = slice_object.start - shifted_start
                ## if the local_start is negative, pull it up to zero
                local_start = 0 if local_start < 0 else local_start                
                ## 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
            ## calculate the stop index
            if slice_object.stop == None:
                local_stop = None
            else:
                local_stop = slice_object.stop - shifted_start
                ## if local_stop is negative, pull it up to zero
                local_stop = 0 if local_stop < 0 else local_stop    
1238
        ## Note: if start or stop are greater than the array length,
ultimanet's avatar
ultimanet committed
1239
1240
1241
        ## numpy will automatically cut the index value down into the 
        ## array's range 
        return local_start, local_stop        
1242
    
Ultimanet's avatar
Ultimanet committed
1243
1244
    def inject(self, data, to_slices, data_update, from_slices, comm=None, 
               **kwargs):
Ultimanet's avatar
Ultimanet committed
1245
        ## check if to_key and from_key are completely built of slices 
Ultimanet's avatar
Ultimanet committed
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
        if not np.all(
            np.vectorize(lambda x: isinstance(x, slice))(to_slices)):
            raise ValueError(about._errors.cstring(
            "ERROR: The to_slices argument must be a list or tuple of slices!")
            )

        if not np.all(
            np.vectorize(lambda x: isinstance(x, slice))(from_slices)):
            raise ValueError(about._errors.cstring(
            "ERROR: The from_slices argument must be a list or tuple of slices!")
            )
            
        to_slices = tuple(to_slices)
        from_slices = tuple(from_slices)
        self.disperse_data(data = data, 
                           to_slices = to_slices,
                           data_update = data_update,
                           from_slices = from_slices,
                           comm=comm,
                           **kwargs)
Ultimanet's avatar
Ultimanet committed
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
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

    def extract_local_data(self, data_object):
        ## if data_object is not a ndarray or a d2o, cast it to a ndarray
        if not (isinstance(data_object, np.ndarray) or 
                isinstance(data_object, distributed_data_object)):
            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 locall 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[:] 
            ## if not, allgather the local data pieces and extract from this
            else:
                allgathered_data = self._allgather(data_object)
                allgathered_data = np.concatenate(allgathered_data)
                if allgathered_data.shape != self.global_shape:
                    raise ValueError(
                            about._errors.cstring(
            "ERROR: supplied shapes do neither match globally nor locally"))
                return self.extract_local_data(allgathered_data)
            
        ## if shape-casting was successfull, extract the data
        else:
            ## If the first dimension matches only via broadcasting...
            ## Case 1: ...do broadcasting. This procedure does not depend on the
            ## array type (ndarray or d2o)
            if matching_dimensions[0] == False:
                extracted_data = data_object[0:1]
    
    
            ## Case 2: First dimension fits directly and data_object is a d2o
            elif isinstance(data_object, distributed_data_object):
                ## Check if the distribution_strategy and the comm match 
                ## the own ones.            
                if type(self) == type(data_object.distributor) and\
                    self.comm == data_object.distributor.comm:
                    ## Case 1: yes. Simply take the local data
                    extracted_data = data_object.data
                else:            
                    ## Case 2: no. All nodes extract their local slice from the 
                    ## data_object
                    extracted_data =\
                        data_object[self.local_start:self.local_end]
            
            ## Case 3: First dimension fits directly and data_object is an generic
            ## array        
            else:
                extracted_data =\
                    data_object[self.local_start:self.local_end]
            
        return extracted_data

    def _reshape_foreign_data(self, foreign):
        ## 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):
           raise ValueError(
               about._errors.cstring("ERROR: unequal number of dimensions!")) 
        ## 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):
            raise ValueError(
                about._errors.cstring("ERROR: incompatible shapes!")) 
        matching_dimensions = tuple(direct_match)
        return (foreign, matching_dimensions)
Ultimanet's avatar
Ultimanet committed
1361
        
Ultimanet's avatar
Ultimanet committed
1362
                
1363
1364
1365
    def consolidate_data(self, data, target_rank='all', comm = None):
        if comm == None:
            comm = self.comm            
ultimanet's avatar
ultimanet committed
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
        _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]
        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
    
    if found['h5py']:
1379
1380
1381
        def save_data(self, data, alias, path=None, overwriteQ=True, comm=None):
            if comm == None:
                comm = self.comm            
ultimanet's avatar
ultimanet committed
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
            ## if no path and therefore no filename was given, use the alias as filename        
            use_path = alias if path==None else path
            
            ## create the file-handle
            if found['h5py_parallel']:
                f = h5py.File(use_path, 'a', driver='mpio', comm=comm)
            else:
                f= h5py.File(use_path, 'a')
            ## check if dataset with name == alias already exists
            try: 
                f[alias]
                if overwriteQ == False: #if yes, and overwriteQ is set to False, raise an Error
Ultimanet's avatar
Ultimanet committed
1394
                    raise KeyError(about._errors.cstring("ERROR: overwriteQ == False, but alias already in use!"))
ultimanet's avatar
ultimanet committed
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
                else: # if yes, remove the existing dataset
                    del f[alias]
            except(KeyError):
                pass
            
            ## create dataset
            dset = f.create_dataset(alias, shape=self.global_shape, dtype=self.dtype)
            ## write the data
            dset[self.local_start:self.local_end] = data
            ## close the file
            f.close()
        
1407
1408
1409
        def load_data(self, alias, path, comm=None):
            if comm == None:
                comm = self.comm            
ultimanet's avatar
ultimanet committed
1410
1411
1412
1413
1414
1415
1416
1417
            ## create the file-handle
            if found['h5py_parallel']:
                f = h5py.File(path, 'r', driver='mpio', comm=comm)
            else:
                f= h5py.File(path, 'r')        
            dset = f[alias]        
            ## check shape
            if dset.shape != self.global_shape:
Ultimanet's avatar
Ultimanet committed
1418
                raise TypeError(about._errors.cstring("ERROR: The shape of the given dataset does not match the distributed_data_object."))
ultimanet's avatar
ultimanet committed
1419
1420
            ## check dtype
            if dset.dtype.type != self.dtype:
Ultimanet's avatar
Ultimanet committed
1421
                raise TypeError(about._errors.cstring("ERROR: The datatype of the given dataset does not match the distributed_data_object."))
ultimanet's avatar
ultimanet committed
1422
1423
1424
1425
1426
1427
1428
            ## if everything seems to fit, load the data
            data = dset[self.local_start:self.local_end]
            ## close the file
            f.close()
            return data
    else:
        def save_data(self, *args, **kwargs):
Ultimanet's avatar
Ultimanet committed
1429
            raise ImportError(about._errors.cstring("ERROR: h5py was not imported")) 
ultimanet's avatar
ultimanet committed
1430
        def load_data(self, *args, **kwargs):
Ultimanet's avatar
Ultimanet committed
1431
            raise ImportError(about._errors.cstring("ERROR: h5py was not imported")) 
ultimanet's avatar
ultimanet committed
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
        
        
        
        

class _not_distributor(object):
    def __init__(self, global_data=None, global_shape=None, dtype=None, *args,  **kwargs):
        if dtype != None:        
            self.dtype = dtype
        elif global_data != None:
            self.dtype = np.array(global_data).dtype.type
            
1444
        if global_data != None and np.array(global_data).shape != ():
ultimanet's avatar
ultimanet committed
1445
1446
1447
1448
            self.global_shape = np.array(global_data).shape
        elif global_shape != None:
            self.global_shape = global_shape
        else:
Ultimanet's avatar
Ultimanet committed
1449
            raise TypeError(about._errors.cstring("ERROR: Neither data nor shape supplied!")) 
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
    
    def globalize_flat_index(self, index):
        return index
    
    def globalize_index(self, index):
        return index
    
    def _allgather(self, thing):
        return [thing,]
        
ultimanet's avatar
ultimanet committed
1460
    def distribute_data(self, data, **kwargs):
Ultimanet's avatar
Ultimanet committed