distributed_do.py 18.7 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
# 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/>.
#
14
# Copyright(C) 2013-2019 Max-Planck-Society
15
#
16
# NIFTy is being developed at the Max-Planck-Institut fuer Astrophysik.
Philipp Arras's avatar
Philipp Arras committed
17
18

import sys
Martin Reinecke's avatar
fix    
Martin Reinecke committed
19
from functools import reduce
Philipp Arras's avatar
Philipp Arras committed
20

21
22
import numpy as np
from mpi4py import MPI
Philipp Arras's avatar
Philipp Arras committed
23
24

from .random import Random
25

Martin Reinecke's avatar
Martin Reinecke committed
26
27
28
29
30
31
__all__ = ["ntask", "rank", "master", "local_shape", "data_object", "full",
           "empty", "zeros", "ones", "empty_like", "vdot", "exp",
           "log", "tanh", "sqrt", "from_object", "from_random",
           "local_data", "ibegin", "ibegin_from_shape", "np_allreduce_sum",
           "np_allreduce_min", "np_allreduce_max",
           "distaxis", "from_local_data", "from_global_data", "to_global_data",
Martin Reinecke's avatar
Martin Reinecke committed
32
           "redistribute", "default_distaxis", "is_numpy", "absmax", "norm",
Martin Reinecke's avatar
Martin Reinecke committed
33
           "lock", "locked", "uniform_full", "transpose", "to_global_data_rw",
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
34
           "ensure_not_distributed", "ensure_default_distributed",
35
           "tanh", "conjugate", "sin", "cos", "tan", "log10",
Martin Reinecke's avatar
Martin Reinecke committed
36
           "sinh", "cosh", "sinc", "absolute", "sign", "clip"]
Martin Reinecke's avatar
Martin Reinecke committed
37

Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
38
39
40
_comm = MPI.COMM_WORLD
ntask = _comm.Get_size()
rank = _comm.Get_rank()
Martin Reinecke's avatar
Martin Reinecke committed
41
master = (rank == 0)
42
43


Martin Reinecke's avatar
Martin Reinecke committed
44
45
46
47
def is_numpy():
    return False


Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
48
def _shareSize(nwork, nshares, myshare):
Martin Reinecke's avatar
Martin Reinecke committed
49
    return (nwork//nshares) + int(myshare < nwork % nshares)
Martin Reinecke's avatar
Martin Reinecke committed
50

Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
51
52

def _shareRange(nwork, nshares, myshare):
Martin Reinecke's avatar
Martin Reinecke committed
53
54
    nbase = nwork//nshares
    additional = nwork % nshares
Martin Reinecke's avatar
Martin Reinecke committed
55
    lo = myshare*nbase + min(myshare, additional)
Martin Reinecke's avatar
Martin Reinecke committed
56
    hi = lo + nbase + int(myshare < additional)
Martin Reinecke's avatar
Martin Reinecke committed
57
58
    return lo, hi

59

60
def local_shape(shape, distaxis=0):
Martin Reinecke's avatar
Martin Reinecke committed
61
    if len(shape) == 0 or distaxis == -1:
62
        return shape
Martin Reinecke's avatar
Martin Reinecke committed
63
64
    shape2 = list(shape)
    shape2[distaxis] = _shareSize(shape[distaxis], ntask, rank)
65
66
    return tuple(shape2)

Martin Reinecke's avatar
Martin Reinecke committed
67

68
69
class data_object(object):
    def __init__(self, shape, data, distaxis):
Martin Reinecke's avatar
Martin Reinecke committed
70
        self._shape = tuple(shape)
Martin Reinecke's avatar
Martin Reinecke committed
71
        if len(self._shape) == 0:
Martin Reinecke's avatar
fixes    
Martin Reinecke committed
72
            distaxis = -1
Martin Reinecke's avatar
Martin Reinecke committed
73
74
            if not isinstance(data, np.ndarray):
                data = np.full((), data)
75
76
        self._distaxis = distaxis
        self._data = data
Martin Reinecke's avatar
Martin Reinecke committed
77
78
        if local_shape(self._shape, self._distaxis) != self._data.shape:
            raise ValueError("shape mismatch")
79

80
81
82
    def copy(self):
        return data_object(self._shape, self._data.copy(), self._distaxis)

Martin Reinecke's avatar
Martin Reinecke committed
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
#     def _sanity_checks(self):
#         # check whether the distaxis is consistent
#         if self._distaxis < -1 or self._distaxis >= len(self._shape):
#             raise ValueError
#         itmp = np.array(self._distaxis)
#         otmp = np.empty(ntask, dtype=np.int)
#         _comm.Allgather(itmp, otmp)
#         if np.any(otmp != self._distaxis):
#             raise ValueError
#         # check whether the global shape is consistent
#         itmp = np.array(self._shape)
#         otmp = np.empty((ntask, len(self._shape)), dtype=np.int)
#         _comm.Allgather(itmp, otmp)
#         for i in range(ntask):
#             if np.any(otmp[i, :] != self._shape):
#                 raise ValueError
#         # check shape of local data
#         if self._distaxis < 0:
#             if self._data.shape != self._shape:
#                 raise ValueError
#         else:
#             itmp = np.array(self._shape)
#             itmp[self._distaxis] = _shareSize(self._shape[self._distaxis],
#                                               ntask, rank)
#             if np.any(self._data.shape != itmp):
#                 raise ValueError
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123

    @property
    def dtype(self):
        return self._data.dtype

    @property
    def shape(self):
        return self._shape

    @property
    def size(self):
        return np.prod(self._shape)

    @property
    def real(self):
Martin Reinecke's avatar
Martin Reinecke committed
124
        return data_object(self._shape, self._data.real, self._distaxis)
125
126
127

    @property
    def imag(self):
Martin Reinecke's avatar
Martin Reinecke committed
128
        return data_object(self._shape, self._data.imag, self._distaxis)
129

Martin Reinecke's avatar
Martin Reinecke committed
130
131
132
133
134
135
    def conj(self):
        return data_object(self._shape, self._data.conj(), self._distaxis)

    def conjugate(self):
        return data_object(self._shape, self._data.conjugate(), self._distaxis)

Martin Reinecke's avatar
Martin Reinecke committed
136
    def _contraction_helper(self, op, mpiop, axis):
137
        if axis is not None:
Martin Reinecke's avatar
Martin Reinecke committed
138
            if len(axis) == len(self._data.shape):
139
140
                axis = None
        if axis is None:
Martin Reinecke's avatar
Martin Reinecke committed
141
            res = np.array(getattr(self._data, op)())
Martin Reinecke's avatar
Martin Reinecke committed
142
            if (self._distaxis == -1):
Martin Reinecke's avatar
fixes    
Martin Reinecke committed
143
                return res[()]
Martin Reinecke's avatar
Martin Reinecke committed
144
145
            res2 = np.empty((), dtype=res.dtype)
            _comm.Allreduce(res, res2, mpiop)
Martin Reinecke's avatar
fixes    
Martin Reinecke committed
146
            return res2[()]
147
148

        if self._distaxis in axis:
Martin Reinecke's avatar
Martin Reinecke committed
149
150
            res = getattr(self._data, op)(axis=axis)
            res2 = np.empty_like(res)
Martin Reinecke's avatar
Martin Reinecke committed
151
            _comm.Allreduce(res, res2, mpiop)
Martin Reinecke's avatar
Martin Reinecke committed
152
            return from_global_data(res2, distaxis=0)
153
        else:
Martin Reinecke's avatar
Martin Reinecke committed
154
            # perform the contraction on the local data
Martin Reinecke's avatar
Martin Reinecke committed
155
156
            res = getattr(self._data, op)(axis=axis)
            if self._distaxis == -1:
Martin Reinecke's avatar
Martin Reinecke committed
157
                return from_global_data(res, distaxis=0)
Martin Reinecke's avatar
Martin Reinecke committed
158
            shp = list(res.shape)
Martin Reinecke's avatar
Martin Reinecke committed
159
            shift = 0
Martin Reinecke's avatar
Martin Reinecke committed
160
            for ax in axis:
Martin Reinecke's avatar
Martin Reinecke committed
161
162
                if ax < self._distaxis:
                    shift += 1
Martin Reinecke's avatar
Martin Reinecke committed
163
164
            shp[self._distaxis-shift] = self.shape[self._distaxis]
            return from_local_data(shp, res, self._distaxis-shift)
165
166
167

    def sum(self, axis=None):
        return self._contraction_helper("sum", MPI.SUM, axis)
Martin Reinecke's avatar
Martin Reinecke committed
168

169
170
171
    def prod(self, axis=None):
        return self._contraction_helper("prod", MPI.PROD, axis)

172
173
#    def min(self, axis=None):
#        return self._contraction_helper("min", MPI.MIN, axis)
Martin Reinecke's avatar
Martin Reinecke committed
174

175
176
#    def max(self, axis=None):
#        return self._contraction_helper("max", MPI.MAX, axis)
177

178
179
180
181
182
183
    def mean(self, axis=None):
        if axis is None:
            sz = self.size
        else:
            sz = reduce(lambda x, y: x*y, [self.shape[i] for i in axis])
        return self.sum(axis)/sz
Martin Reinecke's avatar
Martin Reinecke committed
184

185
186
    def std(self, axis=None):
        return np.sqrt(self.var(axis))
Martin Reinecke's avatar
Martin Reinecke committed
187

Martin Reinecke's avatar
Martin Reinecke committed
188
    # FIXME: to be improved!
189
190
191
    def var(self, axis=None):
        if axis is not None and len(axis) != len(self.shape):
            raise ValueError("functionality not yet supported")
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
192
193
        return (abs(self-self.mean())**2).mean()

194
    def _binary_helper(self, other, op):
Martin Reinecke's avatar
Martin Reinecke committed
195
        a = self
196
        if isinstance(other, data_object):
Martin Reinecke's avatar
Martin Reinecke committed
197
            b = other
198
199
200
201
            if a._shape != b._shape:
                raise ValueError("shapes are incompatible.")
            if a._distaxis != b._distaxis:
                raise ValueError("distributions are incompatible.")
Martin Reinecke's avatar
Martin Reinecke committed
202
203
            a = a._data
            b = b._data
Martin Reinecke's avatar
Martin Reinecke committed
204
205
206
207
208
        elif np.isscalar(other):
            a = a._data
            b = other
        else:
            return NotImplemented
209
210

        tval = getattr(a, op)(b)
Martin Reinecke's avatar
Martin Reinecke committed
211
212
213
214
        if tval is a:
            return self
        else:
            return data_object(self._shape, tval, self._distaxis)
215

Martin Reinecke's avatar
Martin Reinecke committed
216
217
218
    def clip(self, min=None, max=None):
        return data_object(self._shape, np.clip(self._data, min, max))

219
    def __neg__(self):
Martin Reinecke's avatar
Martin Reinecke committed
220
        return data_object(self._shape, -self._data, self._distaxis)
221
222

    def __abs__(self):
223
        return data_object(self._shape, abs(self._data), self._distaxis)
224
225

    def all(self):
Martin Reinecke's avatar
fixes    
Martin Reinecke committed
226
        return self.sum() == self.size
227
228

    def any(self):
Martin Reinecke's avatar
fixes    
Martin Reinecke committed
229
        return self.sum() != 0
230

Martin Reinecke's avatar
fixes    
Martin Reinecke committed
231
232
    def fill(self, value):
        self._data.fill(value)
233

234

235
236
237
238
239
240
241
242
243
244
245
246
247
for op in ["__add__", "__radd__", "__iadd__",
           "__sub__", "__rsub__", "__isub__",
           "__mul__", "__rmul__", "__imul__",
           "__truediv__", "__rtruediv__", "__itruediv__",
           "__floordiv__", "__rfloordiv__", "__ifloordiv__",
           "__pow__", "__rpow__", "__ipow__",
           "__lt__", "__le__", "__gt__", "__ge__", "__eq__", "__ne__"]:
    def func(op):
        def func2(self, other):
            return self._binary_helper(other, op=op)
        return func2
    setattr(data_object, op, func(op))

Martin Reinecke's avatar
Martin Reinecke committed
248

Martin Reinecke's avatar
Martin Reinecke committed
249
def full(shape, fill_value, dtype=None, distaxis=0):
Martin Reinecke's avatar
Martin Reinecke committed
250
251
    return data_object(shape, np.full(local_shape(shape, distaxis),
                                      fill_value, dtype), distaxis)
252
253


Martin Reinecke's avatar
Martin Reinecke committed
254
255
256
257
258
259
def uniform_full(shape, fill_value, dtype=None, distaxis=0):
    return data_object(
        shape, np.broadcast_to(fill_value, local_shape(shape, distaxis)),
        distaxis)


Martin Reinecke's avatar
fixes    
Martin Reinecke committed
260
def empty(shape, dtype=None, distaxis=0):
Martin Reinecke's avatar
Martin Reinecke committed
261
262
    return data_object(shape, np.empty(local_shape(shape, distaxis),
                                       dtype), distaxis)
263
264


Martin Reinecke's avatar
fixes    
Martin Reinecke committed
265
def zeros(shape, dtype=None, distaxis=0):
Martin Reinecke's avatar
Martin Reinecke committed
266
267
    return data_object(shape, np.zeros(local_shape(shape, distaxis), dtype),
                       distaxis)
268
269


Martin Reinecke's avatar
fixes    
Martin Reinecke committed
270
def ones(shape, dtype=None, distaxis=0):
Martin Reinecke's avatar
Martin Reinecke committed
271
272
    return data_object(shape, np.ones(local_shape(shape, distaxis), dtype),
                       distaxis)
273
274
275
276
277
278
279


def empty_like(a, dtype=None):
    return data_object(np.empty_like(a._data, dtype))


def vdot(a, b):
Martin Reinecke's avatar
fixes    
Martin Reinecke committed
280
    tmp = np.array(np.vdot(a._data, b._data))
Martin Reinecke's avatar
Martin Reinecke committed
281
    if a._distaxis == -1:
Martin Reinecke's avatar
Martin Reinecke committed
282
        return tmp[()]
Martin Reinecke's avatar
Martin Reinecke committed
283
284
    res = np.empty((), dtype=tmp.dtype)
    _comm.Allreduce(tmp, res, MPI.SUM)
Martin Reinecke's avatar
fixes    
Martin Reinecke committed
285
    return res[()]
286
287
288


def _math_helper(x, function, out):
289
    function = getattr(np, function)
290
291
292
293
    if out is not None:
        function(x._data, out=out._data)
        return out
    else:
Martin Reinecke's avatar
Martin Reinecke committed
294
        return data_object(x.shape, function(x._data), x._distaxis)
295
296


297
_current_module = sys.modules[__name__]
Martin Reinecke's avatar
Martin Reinecke committed
298

299
for f in ["sqrt", "exp", "log", "tanh", "conjugate", "sin", "cos", "tan",
300
          "sinh", "cosh", "sinc", "absolute", "sign", "log10"]:
301
302
303
304
305
    def func(f):
        def func2(x, out=None):
            return _math_helper(x, f, out)
        return func2
    setattr(_current_module, f, func(f))
306
307


Martin Reinecke's avatar
Martin Reinecke committed
308
def clip(x, a_min=None, a_max=None):
309
    return data_object(x.shape, np.clip(x._data, a_min, a_max), x._distaxis)
310
311


Martin Reinecke's avatar
Martin Reinecke committed
312
313
314
315
316
317
318
319
320
321
322
323
def from_object(object, dtype, copy, set_locked):
    if dtype is None:
        dtype = object.dtype
    dtypes_equal = dtype == object.dtype
    if set_locked and dtypes_equal and locked(object):
        return object
    if not dtypes_equal and not copy:
        raise ValueError("cannot change data type without copying")
    if set_locked and not copy:
        raise ValueError("cannot lock object without copying")
    data = np.array(object._data, dtype=dtype, copy=copy)
    if set_locked:
Martin Reinecke's avatar
fix    
Martin Reinecke committed
324
        data.flags.writeable = False
Martin Reinecke's avatar
Martin Reinecke committed
325
    return data_object(object._shape, data, distaxis=object._distaxis)
326
327


Martin Reinecke's avatar
Martin Reinecke committed
328
329
# This function draws all random numbers on all tasks, to produce the same
# array independent on the number of tasks
Martin Reinecke's avatar
Martin Reinecke committed
330
331
332
# MR FIXME: depending on what is really wanted/needed (i.e. same result
# independent of number of tasks, performance etc.) we need to adjust the
# algorithm.
Martin Reinecke's avatar
Martin Reinecke committed
333
def from_random(random_type, shape, dtype=np.float64, **kwargs):
334
    generator_function = getattr(Random, random_type)
Martin Reinecke's avatar
Martin Reinecke committed
335
    if len(shape) == 0:
Martin Reinecke's avatar
Martin Reinecke committed
336
337
338
        ldat = generator_function(dtype=dtype, shape=shape, **kwargs)
        ldat = _comm.bcast(ldat)
        return from_local_data(shape, ldat, distaxis=-1)
Martin Reinecke's avatar
Martin Reinecke committed
339
340
341
342
343
344
345
    for i in range(ntask):
        lshape = list(shape)
        lshape[0] = _shareSize(shape[0], ntask, i)
        ldat = generator_function(dtype=dtype, shape=lshape, **kwargs)
        if i == rank:
            outdat = ldat
    return from_local_data(shape, outdat, distaxis=0)
346

Martin Reinecke's avatar
Martin Reinecke committed
347

Martin Reinecke's avatar
Martin Reinecke committed
348
349
350
351
def local_data(arr):
    return arr._data


352
353
def ibegin_from_shape(glob_shape, distaxis=0):
    res = [0] * len(glob_shape)
Martin Reinecke's avatar
Martin Reinecke committed
354
    if distaxis < 0:
355
356
357
358
359
        return res
    res[distaxis] = _shareRange(glob_shape[distaxis], ntask, rank)[0]
    return tuple(res)


Martin Reinecke's avatar
fixes    
Martin Reinecke committed
360
361
def ibegin(arr):
    res = [0] * arr._data.ndim
Martin Reinecke's avatar
Martin Reinecke committed
362
    res[arr._distaxis] = _shareRange(arr._shape[arr._distaxis], ntask, rank)[0]
Martin Reinecke's avatar
fixes    
Martin Reinecke committed
363
    return tuple(res)
Martin Reinecke's avatar
Martin Reinecke committed
364
365


Martin Reinecke's avatar
fixes    
Martin Reinecke committed
366
367
def np_allreduce_sum(arr):
    res = np.empty_like(arr)
Martin Reinecke's avatar
Martin Reinecke committed
368
    _comm.Allreduce(arr, res, MPI.SUM)
Martin Reinecke's avatar
fixes    
Martin Reinecke committed
369
    return res
Martin Reinecke's avatar
Martin Reinecke committed
370
371


372
373
374
375
376
377
def np_allreduce_min(arr):
    res = np.empty_like(arr)
    _comm.Allreduce(arr, res, MPI.MIN)
    return res


Martin Reinecke's avatar
fixes    
Martin Reinecke committed
378
379
380
381
382
383
def np_allreduce_max(arr):
    res = np.empty_like(arr)
    _comm.Allreduce(arr, res, MPI.MAX)
    return res


Martin Reinecke's avatar
Martin Reinecke committed
384
385
386
387
def distaxis(arr):
    return arr._distaxis


Martin Reinecke's avatar
Martin Reinecke committed
388
def from_local_data(shape, arr, distaxis=0):
389
390
    if arr.dtype.kind not in "fciub":
        raise TypeError
Martin Reinecke's avatar
Martin Reinecke committed
391
392
393
    return data_object(shape, arr, distaxis)


394
def from_global_data(arr, sum_up=False, distaxis=0):
395
396
    if arr.dtype.kind not in "fciub":
        raise TypeError
397
398
    if sum_up:
        arr = np_allreduce_sum(arr)
Martin Reinecke's avatar
Martin Reinecke committed
399
    if distaxis == -1:
Martin Reinecke's avatar
Martin Reinecke committed
400
        return data_object(arr.shape, arr, distaxis)
Martin Reinecke's avatar
Martin Reinecke committed
401
    lo, hi = _shareRange(arr.shape[distaxis], ntask, rank)
Martin Reinecke's avatar
Martin Reinecke committed
402
    sl = [slice(None)]*len(arr.shape)
Martin Reinecke's avatar
Martin Reinecke committed
403
    sl[distaxis] = slice(lo, hi)
404
    return data_object(arr.shape, arr[tuple(sl)], distaxis)
Martin Reinecke's avatar
Martin Reinecke committed
405
406


Martin Reinecke's avatar
Martin Reinecke committed
407
408
def to_global_data(arr):
    if arr._distaxis == -1:
Martin Reinecke's avatar
fixes    
Martin Reinecke committed
409
410
411
412
413
        return arr._data
    tmp = redistribute(arr, dist=-1)
    return tmp._data


414
415
416
417
418
419
420
def to_global_data_rw(arr):
    if arr._distaxis == -1:
        return arr._data.copy()
    tmp = redistribute(arr, dist=-1)
    return tmp._data


Martin Reinecke's avatar
Martin Reinecke committed
421
def redistribute(arr, dist=None, nodist=None):
Martin Reinecke's avatar
Martin Reinecke committed
422
423
424
    if dist is not None:
        if nodist is not None:
            raise ValueError
Martin Reinecke's avatar
Martin Reinecke committed
425
        if dist == arr._distaxis:
Martin Reinecke's avatar
Martin Reinecke committed
426
427
428
429
430
431
            return arr
    else:
        if nodist is None:
            raise ValueError
        if arr._distaxis not in nodist:
            return arr
Martin Reinecke's avatar
Martin Reinecke committed
432
        dist = -1
Martin Reinecke's avatar
Martin Reinecke committed
433
434
        for i in range(len(arr.shape)):
            if i not in nodist:
Martin Reinecke's avatar
Martin Reinecke committed
435
                dist = i
Martin Reinecke's avatar
Martin Reinecke committed
436
                break
Martin Reinecke's avatar
Martin Reinecke committed
437

Martin Reinecke's avatar
Martin Reinecke committed
438
    if arr._distaxis == -1:  # all data available, just pick the proper subset
439
        return from_global_data(arr._data, distaxis=dist)
Martin Reinecke's avatar
Martin Reinecke committed
440
    if dist == -1:  # gather all data on all tasks
Martin Reinecke's avatar
Martin Reinecke committed
441
        tmp = np.moveaxis(arr._data, arr._distaxis, 0)
Martin Reinecke's avatar
Martin Reinecke committed
442
443
        slabsize = np.prod(tmp.shape[1:])*tmp.itemsize
        sz = np.empty(ntask, dtype=np.int)
Martin Reinecke's avatar
Martin Reinecke committed
444
        for i in range(ntask):
Martin Reinecke's avatar
Martin Reinecke committed
445
446
447
448
            sz[i] = slabsize*_shareSize(arr.shape[arr._distaxis], ntask, i)
        disp = np.empty(ntask, dtype=np.int)
        disp[0] = 0
        disp[1:] = np.cumsum(sz[:-1])
Martin Reinecke's avatar
Martin Reinecke committed
449
        tmp = np.require(tmp, requirements="C")
Martin Reinecke's avatar
Martin Reinecke committed
450
451
        out = np.empty(arr.size, dtype=arr.dtype)
        _comm.Allgatherv(tmp, [out, sz, disp, MPI.BYTE])
Martin Reinecke's avatar
Martin Reinecke committed
452
453
454
455
        shp = np.array(arr._shape)
        shp[1:arr._distaxis+1] = shp[0:arr._distaxis]
        shp[0] = arr.shape[arr._distaxis]
        out = out.reshape(shp)
Martin Reinecke's avatar
Martin Reinecke committed
456
        out = np.moveaxis(out, 0, arr._distaxis)
Martin Reinecke's avatar
Martin Reinecke committed
457
        return from_global_data(out, distaxis=-1)
Martin Reinecke's avatar
Martin Reinecke committed
458

Martin Reinecke's avatar
Martin Reinecke committed
459
    # real redistribution via Alltoallv
Martin Reinecke's avatar
Martin Reinecke committed
460
    ssz0 = arr._data.size//arr.shape[dist]
Martin Reinecke's avatar
Martin Reinecke committed
461
    ssz = np.empty(ntask, dtype=np.int)
Martin Reinecke's avatar
Martin Reinecke committed
462
463
464
    rszall = arr.size//arr.shape[dist]*_shareSize(arr.shape[dist], ntask, rank)
    rbuf = np.empty(rszall, dtype=arr.dtype)
    rsz0 = rszall//arr.shape[arr._distaxis]
Martin Reinecke's avatar
Martin Reinecke committed
465
    rsz = np.empty(ntask, dtype=np.int)
Martin Reinecke's avatar
Martin Reinecke committed
466
467
468
469
470
471
472
473
474
475
476
477
    if dist == 0:  # shortcut possible
        sbuf = np.ascontiguousarray(arr._data)
        for i in range(ntask):
            lo, hi = _shareRange(arr.shape[dist], ntask, i)
            ssz[i] = ssz0*(hi-lo)
            rsz[i] = rsz0*_shareSize(arr.shape[arr._distaxis], ntask, i)
    else:
        sbuf = np.empty(arr._data.size, dtype=arr.dtype)
        sslice = [slice(None)]*arr._data.ndim
        ofs = 0
        for i in range(ntask):
            lo, hi = _shareRange(arr.shape[dist], ntask, i)
Martin Reinecke's avatar
Martin Reinecke committed
478
            sslice[dist] = slice(lo, hi)
Martin Reinecke's avatar
Martin Reinecke committed
479
            ssz[i] = ssz0*(hi-lo)
480
            sbuf[ofs:ofs+ssz[i]] = arr._data[tuple(sslice)].flat
Martin Reinecke's avatar
Martin Reinecke committed
481
482
483
484
            ofs += ssz[i]
            rsz[i] = rsz0*_shareSize(arr.shape[arr._distaxis], ntask, i)
    ssz *= arr._data.itemsize
    rsz *= arr._data.itemsize
Martin Reinecke's avatar
Martin Reinecke committed
485
486
    sdisp = np.append(0, np.cumsum(ssz[:-1]))
    rdisp = np.append(0, np.cumsum(rsz[:-1]))
Martin Reinecke's avatar
Martin Reinecke committed
487
488
    s_msg = [sbuf, (ssz, sdisp), MPI.BYTE]
    r_msg = [rbuf, (rsz, rdisp), MPI.BYTE]
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
489
    _comm.Alltoallv(s_msg, r_msg)
Martin Reinecke's avatar
Martin Reinecke committed
490
    del sbuf  # free memory
Martin Reinecke's avatar
Martin Reinecke committed
491
492
493
494
    if arr._distaxis == 0:
        rbuf = rbuf.reshape(local_shape(arr.shape, dist))
        arrnew = from_local_data(arr.shape, rbuf, distaxis=dist)
    else:
Martin Reinecke's avatar
Martin Reinecke committed
495
        arrnew = np.empty(local_shape(arr.shape, dist), dtype=arr.dtype)
Martin Reinecke's avatar
Martin Reinecke committed
496
497
498
499
        rslice = [slice(None)]*arr._data.ndim
        ofs = 0
        for i in range(ntask):
            lo, hi = _shareRange(arr.shape[arr._distaxis], ntask, i)
Martin Reinecke's avatar
Martin Reinecke committed
500
            rslice[arr._distaxis] = slice(lo, hi)
Martin Reinecke's avatar
Martin Reinecke committed
501
            sz = rsz[i]//arr._data.itemsize
502
            arrnew[tuple(rslice)].flat = rbuf[ofs:ofs+sz]
Martin Reinecke's avatar
Martin Reinecke committed
503
            ofs += sz
Martin Reinecke's avatar
Martin Reinecke committed
504
        arrnew = from_local_data(arr.shape, arrnew, distaxis=dist)
Martin Reinecke's avatar
Martin Reinecke committed
505
    return arrnew
Martin Reinecke's avatar
Martin Reinecke committed
506
507


Martin Reinecke's avatar
Martin Reinecke committed
508
509
def transpose(arr):
    if len(arr.shape) != 2 or arr._distaxis != 0:
Martin Reinecke's avatar
Martin Reinecke committed
510
        raise ValueError("bad input")
Martin Reinecke's avatar
Martin Reinecke committed
511
512
513
514
515
516
517
518
519
520
521
    ssz0 = arr._data.size//arr.shape[1]
    ssz = np.empty(ntask, dtype=np.int)
    rszall = arr.size//arr.shape[1]*_shareSize(arr.shape[1], ntask, rank)
    rbuf = np.empty(rszall, dtype=arr.dtype)
    rsz0 = rszall//arr.shape[0]
    rsz = np.empty(ntask, dtype=np.int)
    sbuf = np.empty(arr._data.size, dtype=arr.dtype)
    ofs = 0
    for i in range(ntask):
        lo, hi = _shareRange(arr.shape[1], ntask, i)
        ssz[i] = ssz0*(hi-lo)
Martin Reinecke's avatar
Martin Reinecke committed
522
        sbuf[ofs:ofs+ssz[i]] = arr._data[:, lo:hi].flat
Martin Reinecke's avatar
Martin Reinecke committed
523
524
525
526
527
528
529
530
531
532
        ofs += ssz[i]
        rsz[i] = rsz0*_shareSize(arr.shape[0], ntask, i)
    ssz *= arr._data.itemsize
    rsz *= arr._data.itemsize
    sdisp = np.append(0, np.cumsum(ssz[:-1]))
    rdisp = np.append(0, np.cumsum(rsz[:-1]))
    s_msg = [sbuf, (ssz, sdisp), MPI.BYTE]
    r_msg = [rbuf, (rsz, rdisp), MPI.BYTE]
    _comm.Alltoallv(s_msg, r_msg)
    del sbuf  # free memory
Martin Reinecke's avatar
Martin Reinecke committed
533
    sz2 = _shareSize(arr.shape[1], ntask, rank)
Martin Reinecke's avatar
Martin Reinecke committed
534
535
    arrnew = np.empty((sz2, arr.shape[0]), dtype=arr.dtype)
    ofs = 0
Martin Reinecke's avatar
Martin Reinecke committed
536
537
538
    for i in range(ntask):
        lo, hi = _shareRange(arr.shape[0], ntask, i)
        sz = rsz[i]//arr._data.itemsize
Martin Reinecke's avatar
Martin Reinecke committed
539
        arrnew[:, lo:hi] = rbuf[ofs:ofs+sz].reshape(hi-lo, sz2).T
Martin Reinecke's avatar
Martin Reinecke committed
540
        ofs += sz
541
    return from_local_data((arr.shape[1], arr.shape[0]), arrnew, 0)
Martin Reinecke's avatar
Martin Reinecke committed
542
543


Martin Reinecke's avatar
Martin Reinecke committed
544
545
def default_distaxis():
    return 0
546
547
548
549
550
551
552
553


def lock(arr):
    arr._data.flags.writeable = False


def locked(arr):
    return not arr._data.flags.writeable
Martin Reinecke's avatar
Martin Reinecke committed
554
555
556
557
558
559
560
561
562
563
564
565


def ensure_not_distributed(arr, axes):
    if arr._distaxis in axes:
        arr = redistribute(arr, nodist=axes)
    return arr, arr._data


def ensure_default_distributed(arr):
    if arr._distaxis != 0:
        arr = redistribute(arr, dist=0)
    return arr
Martin Reinecke's avatar
Martin Reinecke committed
566
567
568
569
570
571


def absmax(arr):
    if arr._data.size == 0:
        tmp = np.array(0, dtype=arr._data.dtype)
    else:
572
        tmp = np.asarray(np.linalg.norm(arr._data.reshape(-1), ord=np.inf))
Martin Reinecke's avatar
Martin Reinecke committed
573
574
575
576
577
578
579
580
    res = np.empty_like(tmp)
    _comm.Allreduce(tmp, res, MPI.MAX)
    return res[()]


def norm(arr, ord=2):
    if ord == np.inf:
        return absmax(arr)
581
    tmp = np.asarray(np.linalg.norm(arr._data.reshape(-1), ord=ord) ** ord)
Martin Reinecke's avatar
Martin Reinecke committed
582
    res = np.empty_like(tmp)
Martin Reinecke's avatar
Martin Reinecke committed
583
584
585
586
    if len(arr._data.shape) == 0:
        res = tmp
    else:
        _comm.Allreduce(tmp, res, MPI.SUM)
Martin Reinecke's avatar
Martin Reinecke committed
587
    return res[()] ** (1./ord)