distributed_do.py 18.6 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
20
21
import numpy as np
from mpi4py import MPI
Philipp Arras's avatar
Philipp Arras committed
22
23

from .random import Random
24

Martin Reinecke's avatar
Martin Reinecke committed
25
26
27
28
29
30
__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
31
           "redistribute", "default_distaxis", "is_numpy", "absmax", "norm",
Martin Reinecke's avatar
Martin Reinecke committed
32
           "lock", "locked", "uniform_full", "transpose", "to_global_data_rw",
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
33
           "ensure_not_distributed", "ensure_default_distributed",
Martin Reinecke's avatar
Martin Reinecke committed
34
           "tanh", "conjugate", "sin", "cos", "tan",
Martin Reinecke's avatar
Martin Reinecke committed
35
           "sinh", "cosh", "sinc", "absolute", "sign", "clip"]
Martin Reinecke's avatar
Martin Reinecke committed
36

Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
37
38
39
_comm = MPI.COMM_WORLD
ntask = _comm.Get_size()
rank = _comm.Get_rank()
Martin Reinecke's avatar
Martin Reinecke committed
40
master = (rank == 0)
41
42


Martin Reinecke's avatar
Martin Reinecke committed
43
44
45
46
def is_numpy():
    return False


Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
47
def _shareSize(nwork, nshares, myshare):
Martin Reinecke's avatar
Martin Reinecke committed
48
    return (nwork//nshares) + int(myshare < nwork % nshares)
Martin Reinecke's avatar
Martin Reinecke committed
49

Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
50
51

def _shareRange(nwork, nshares, myshare):
Martin Reinecke's avatar
Martin Reinecke committed
52
53
    nbase = nwork//nshares
    additional = nwork % nshares
Martin Reinecke's avatar
Martin Reinecke committed
54
    lo = myshare*nbase + min(myshare, additional)
Martin Reinecke's avatar
Martin Reinecke committed
55
    hi = lo + nbase + int(myshare < additional)
Martin Reinecke's avatar
Martin Reinecke committed
56
57
    return lo, hi

58

59
def local_shape(shape, distaxis=0):
Martin Reinecke's avatar
Martin Reinecke committed
60
    if len(shape) == 0 or distaxis == -1:
61
        return shape
Martin Reinecke's avatar
Martin Reinecke committed
62
63
    shape2 = list(shape)
    shape2[distaxis] = _shareSize(shape[distaxis], ntask, rank)
64
65
    return tuple(shape2)

Martin Reinecke's avatar
Martin Reinecke committed
66

67
68
class data_object(object):
    def __init__(self, shape, data, distaxis):
Martin Reinecke's avatar
Martin Reinecke committed
69
        self._shape = tuple(shape)
Martin Reinecke's avatar
Martin Reinecke committed
70
        if len(self._shape) == 0:
Martin Reinecke's avatar
fixes    
Martin Reinecke committed
71
            distaxis = -1
Martin Reinecke's avatar
Martin Reinecke committed
72
73
            if not isinstance(data, np.ndarray):
                data = np.full((), data)
74
75
        self._distaxis = distaxis
        self._data = data
Martin Reinecke's avatar
Martin Reinecke committed
76
77
        if local_shape(self._shape, self._distaxis) != self._data.shape:
            raise ValueError("shape mismatch")
78

79
80
81
    def copy(self):
        return data_object(self._shape, self._data.copy(), self._distaxis)

Martin Reinecke's avatar
Martin Reinecke committed
82
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
#     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
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122

    @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
123
        return data_object(self._shape, self._data.real, self._distaxis)
124
125
126

    @property
    def imag(self):
Martin Reinecke's avatar
Martin Reinecke committed
127
        return data_object(self._shape, self._data.imag, self._distaxis)
128

Martin Reinecke's avatar
Martin Reinecke committed
129
130
131
132
133
134
    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
135
    def _contraction_helper(self, op, mpiop, axis):
136
        if axis is not None:
Martin Reinecke's avatar
Martin Reinecke committed
137
            if len(axis) == len(self._data.shape):
138
139
                axis = None
        if axis is None:
Martin Reinecke's avatar
Martin Reinecke committed
140
            res = np.array(getattr(self._data, op)())
Martin Reinecke's avatar
Martin Reinecke committed
141
            if (self._distaxis == -1):
Martin Reinecke's avatar
fixes    
Martin Reinecke committed
142
                return res[()]
Martin Reinecke's avatar
Martin Reinecke committed
143
144
            res2 = np.empty((), dtype=res.dtype)
            _comm.Allreduce(res, res2, mpiop)
Martin Reinecke's avatar
fixes    
Martin Reinecke committed
145
            return res2[()]
146
147

        if self._distaxis in axis:
Martin Reinecke's avatar
Martin Reinecke committed
148
149
            res = getattr(self._data, op)(axis=axis)
            res2 = np.empty_like(res)
Martin Reinecke's avatar
Martin Reinecke committed
150
            _comm.Allreduce(res, res2, mpiop)
Martin Reinecke's avatar
Martin Reinecke committed
151
            return from_global_data(res2, distaxis=0)
152
        else:
Martin Reinecke's avatar
Martin Reinecke committed
153
            # perform the contraction on the local data
Martin Reinecke's avatar
Martin Reinecke committed
154
155
            res = getattr(self._data, op)(axis=axis)
            if self._distaxis == -1:
Martin Reinecke's avatar
Martin Reinecke committed
156
                return from_global_data(res, distaxis=0)
Martin Reinecke's avatar
Martin Reinecke committed
157
            shp = list(res.shape)
Martin Reinecke's avatar
Martin Reinecke committed
158
            shift = 0
Martin Reinecke's avatar
Martin Reinecke committed
159
            for ax in axis:
Martin Reinecke's avatar
Martin Reinecke committed
160
161
                if ax < self._distaxis:
                    shift += 1
Martin Reinecke's avatar
Martin Reinecke committed
162
163
            shp[self._distaxis-shift] = self.shape[self._distaxis]
            return from_local_data(shp, res, self._distaxis-shift)
164
165
166

    def sum(self, axis=None):
        return self._contraction_helper("sum", MPI.SUM, axis)
Martin Reinecke's avatar
Martin Reinecke committed
167

168
169
170
    def prod(self, axis=None):
        return self._contraction_helper("prod", MPI.PROD, axis)

171
172
#    def min(self, axis=None):
#        return self._contraction_helper("min", MPI.MIN, axis)
Martin Reinecke's avatar
Martin Reinecke committed
173

174
175
#    def max(self, axis=None):
#        return self._contraction_helper("max", MPI.MAX, axis)
176

177
178
179
180
181
182
    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
183

184
185
    def std(self, axis=None):
        return np.sqrt(self.var(axis))
Martin Reinecke's avatar
Martin Reinecke committed
186

Martin Reinecke's avatar
Martin Reinecke committed
187
    # FIXME: to be improved!
188
189
190
    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
191
192
        return (abs(self-self.mean())**2).mean()

193
    def _binary_helper(self, other, op):
Martin Reinecke's avatar
Martin Reinecke committed
194
        a = self
195
        if isinstance(other, data_object):
Martin Reinecke's avatar
Martin Reinecke committed
196
            b = other
197
198
199
200
            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
201
202
            a = a._data
            b = b._data
Martin Reinecke's avatar
Martin Reinecke committed
203
204
205
206
207
        elif np.isscalar(other):
            a = a._data
            b = other
        else:
            return NotImplemented
208
209

        tval = getattr(a, op)(b)
Martin Reinecke's avatar
Martin Reinecke committed
210
211
212
213
        if tval is a:
            return self
        else:
            return data_object(self._shape, tval, self._distaxis)
214

Martin Reinecke's avatar
Martin Reinecke committed
215
216
217
    def clip(self, min=None, max=None):
        return data_object(self._shape, np.clip(self._data, min, max))

218
    def __neg__(self):
Martin Reinecke's avatar
Martin Reinecke committed
219
        return data_object(self._shape, -self._data, self._distaxis)
220
221

    def __abs__(self):
222
        return data_object(self._shape, abs(self._data), self._distaxis)
223
224

    def all(self):
Martin Reinecke's avatar
fixes    
Martin Reinecke committed
225
        return self.sum() == self.size
226
227

    def any(self):
Martin Reinecke's avatar
fixes    
Martin Reinecke committed
228
        return self.sum() != 0
229

Martin Reinecke's avatar
fixes    
Martin Reinecke committed
230
231
    def fill(self, value):
        self._data.fill(value)
232

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__",
           "__div__", "__rdiv__", "__idiv__",
           "__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
300
for f in ["sqrt", "exp", "log", "tanh", "conjugate", "sin", "cos", "tan",
          "sinh", "cosh", "sinc", "absolute", "sign"]:
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
309
def clip(x, a_min=None, a_max=None):
    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):
Martin Reinecke's avatar
Martin Reinecke committed
389
390
391
    return data_object(shape, arr, distaxis)


392
393
394
def from_global_data(arr, sum_up=False, distaxis=0):
    if sum_up:
        arr = np_allreduce_sum(arr)
Martin Reinecke's avatar
Martin Reinecke committed
395
    if distaxis == -1:
Martin Reinecke's avatar
Martin Reinecke committed
396
        return data_object(arr.shape, arr, distaxis)
Martin Reinecke's avatar
Martin Reinecke committed
397
    lo, hi = _shareRange(arr.shape[distaxis], ntask, rank)
Martin Reinecke's avatar
Martin Reinecke committed
398
    sl = [slice(None)]*len(arr.shape)
Martin Reinecke's avatar
Martin Reinecke committed
399
    sl[distaxis] = slice(lo, hi)
400
    return data_object(arr.shape, arr[tuple(sl)], distaxis)
Martin Reinecke's avatar
Martin Reinecke committed
401
402


Martin Reinecke's avatar
Martin Reinecke committed
403
404
def to_global_data(arr):
    if arr._distaxis == -1:
Martin Reinecke's avatar
fixes    
Martin Reinecke committed
405
406
407
408
409
        return arr._data
    tmp = redistribute(arr, dist=-1)
    return tmp._data


410
411
412
413
414
415
416
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
417
def redistribute(arr, dist=None, nodist=None):
Martin Reinecke's avatar
Martin Reinecke committed
418
419
420
    if dist is not None:
        if nodist is not None:
            raise ValueError
Martin Reinecke's avatar
Martin Reinecke committed
421
        if dist == arr._distaxis:
Martin Reinecke's avatar
Martin Reinecke committed
422
423
424
425
426
427
            return arr
    else:
        if nodist is None:
            raise ValueError
        if arr._distaxis not in nodist:
            return arr
Martin Reinecke's avatar
Martin Reinecke committed
428
        dist = -1
Martin Reinecke's avatar
Martin Reinecke committed
429
430
        for i in range(len(arr.shape)):
            if i not in nodist:
Martin Reinecke's avatar
Martin Reinecke committed
431
                dist = i
Martin Reinecke's avatar
Martin Reinecke committed
432
                break
Martin Reinecke's avatar
Martin Reinecke committed
433

Martin Reinecke's avatar
Martin Reinecke committed
434
    if arr._distaxis == -1:  # all data available, just pick the proper subset
435
        return from_global_data(arr._data, distaxis=dist)
Martin Reinecke's avatar
Martin Reinecke committed
436
    if dist == -1:  # gather all data on all tasks
Martin Reinecke's avatar
Martin Reinecke committed
437
        tmp = np.moveaxis(arr._data, arr._distaxis, 0)
Martin Reinecke's avatar
Martin Reinecke committed
438
439
        slabsize = np.prod(tmp.shape[1:])*tmp.itemsize
        sz = np.empty(ntask, dtype=np.int)
Martin Reinecke's avatar
Martin Reinecke committed
440
        for i in range(ntask):
Martin Reinecke's avatar
Martin Reinecke committed
441
442
443
444
            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
445
        tmp = np.require(tmp, requirements="C")
Martin Reinecke's avatar
Martin Reinecke committed
446
447
        out = np.empty(arr.size, dtype=arr.dtype)
        _comm.Allgatherv(tmp, [out, sz, disp, MPI.BYTE])
Martin Reinecke's avatar
Martin Reinecke committed
448
449
450
451
        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
452
        out = np.moveaxis(out, 0, arr._distaxis)
Martin Reinecke's avatar
Martin Reinecke committed
453
        return from_global_data(out, distaxis=-1)
Martin Reinecke's avatar
Martin Reinecke committed
454

Martin Reinecke's avatar
Martin Reinecke committed
455
    # real redistribution via Alltoallv
Martin Reinecke's avatar
Martin Reinecke committed
456
    ssz0 = arr._data.size//arr.shape[dist]
Martin Reinecke's avatar
Martin Reinecke committed
457
    ssz = np.empty(ntask, dtype=np.int)
Martin Reinecke's avatar
Martin Reinecke committed
458
459
460
    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
461
    rsz = np.empty(ntask, dtype=np.int)
Martin Reinecke's avatar
Martin Reinecke committed
462
463
464
465
466
467
468
469
470
471
472
473
    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
474
            sslice[dist] = slice(lo, hi)
Martin Reinecke's avatar
Martin Reinecke committed
475
            ssz[i] = ssz0*(hi-lo)
476
            sbuf[ofs:ofs+ssz[i]] = arr._data[tuple(sslice)].flat
Martin Reinecke's avatar
Martin Reinecke committed
477
478
479
480
            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
481
482
    sdisp = np.append(0, np.cumsum(ssz[:-1]))
    rdisp = np.append(0, np.cumsum(rsz[:-1]))
Martin Reinecke's avatar
Martin Reinecke committed
483
484
    s_msg = [sbuf, (ssz, sdisp), MPI.BYTE]
    r_msg = [rbuf, (rsz, rdisp), MPI.BYTE]
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
485
    _comm.Alltoallv(s_msg, r_msg)
Martin Reinecke's avatar
Martin Reinecke committed
486
    del sbuf  # free memory
Martin Reinecke's avatar
Martin Reinecke committed
487
488
489
490
    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
491
        arrnew = np.empty(local_shape(arr.shape, dist), dtype=arr.dtype)
Martin Reinecke's avatar
Martin Reinecke committed
492
493
494
495
        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
496
            rslice[arr._distaxis] = slice(lo, hi)
Martin Reinecke's avatar
Martin Reinecke committed
497
            sz = rsz[i]//arr._data.itemsize
498
            arrnew[tuple(rslice)].flat = rbuf[ofs:ofs+sz]
Martin Reinecke's avatar
Martin Reinecke committed
499
            ofs += sz
Martin Reinecke's avatar
Martin Reinecke committed
500
        arrnew = from_local_data(arr.shape, arrnew, distaxis=dist)
Martin Reinecke's avatar
Martin Reinecke committed
501
    return arrnew
Martin Reinecke's avatar
Martin Reinecke committed
502
503


Martin Reinecke's avatar
Martin Reinecke committed
504
505
def transpose(arr):
    if len(arr.shape) != 2 or arr._distaxis != 0:
Martin Reinecke's avatar
Martin Reinecke committed
506
        raise ValueError("bad input")
Martin Reinecke's avatar
Martin Reinecke committed
507
508
509
510
511
512
513
514
515
516
517
    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
518
        sbuf[ofs:ofs+ssz[i]] = arr._data[:, lo:hi].flat
Martin Reinecke's avatar
Martin Reinecke committed
519
520
521
522
523
524
525
526
527
528
        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
529
    sz2 = _shareSize(arr.shape[1], ntask, rank)
Martin Reinecke's avatar
Martin Reinecke committed
530
531
    arrnew = np.empty((sz2, arr.shape[0]), dtype=arr.dtype)
    ofs = 0
Martin Reinecke's avatar
Martin Reinecke committed
532
533
534
    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
535
        arrnew[:, lo:hi] = rbuf[ofs:ofs+sz].reshape(hi-lo, sz2).T
Martin Reinecke's avatar
Martin Reinecke committed
536
        ofs += sz
537
    return from_local_data((arr.shape[1], arr.shape[0]), arrnew, 0)
Martin Reinecke's avatar
Martin Reinecke committed
538
539


Martin Reinecke's avatar
Martin Reinecke committed
540
541
def default_distaxis():
    return 0
542
543
544
545
546
547
548
549


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


def locked(arr):
    return not arr._data.flags.writeable
Martin Reinecke's avatar
Martin Reinecke committed
550
551
552
553
554
555
556
557
558
559
560
561


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
562
563
564
565
566
567


def absmax(arr):
    if arr._data.size == 0:
        tmp = np.array(0, dtype=arr._data.dtype)
    else:
568
        tmp = np.asarray(np.linalg.norm(arr._data.reshape(-1), ord=np.inf))
Martin Reinecke's avatar
Martin Reinecke committed
569
570
571
572
573
574
575
576
    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)
577
    tmp = np.asarray(np.linalg.norm(arr._data.reshape(-1), ord=ord) ** ord)
Martin Reinecke's avatar
Martin Reinecke committed
578
    res = np.empty_like(tmp)
Martin Reinecke's avatar
Martin Reinecke committed
579
580
581
582
    if len(arr._data.shape) == 0:
        res = tmp
    else:
        _comm.Allreduce(tmp, res, MPI.SUM)
Martin Reinecke's avatar
Martin Reinecke committed
583
    return res[()] ** (1./ord)