distributed_do.py 18.7 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
# 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/>.
#
# Copyright(C) 2013-2018 Max-Planck-Society
#
# NIFTy is being developed at the Max-Planck-Institut fuer Astrophysik
# and financially supported by the Studienstiftung des deutschen Volkes.

19
from __future__ import absolute_import, division, print_function
Philipp Arras's avatar
Philipp Arras committed
20 21 22

import sys

23 24
import numpy as np
from mpi4py import MPI
Philipp Arras's avatar
Philipp Arras committed
25 26 27

from ..compat import *
from .random import Random
28

Martin Reinecke's avatar
Martin Reinecke committed
29 30 31 32 33 34
__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
35
           "redistribute", "default_distaxis", "is_numpy", "absmax", "norm",
Martin Reinecke's avatar
Martin Reinecke committed
36
           "lock", "locked", "uniform_full", "transpose", "to_global_data_rw",
Martin Reinecke's avatar
tweaks  
Martin Reinecke committed
37
           "ensure_not_distributed", "ensure_default_distributed",
38 39
           "clipped_exp", "tanh", "conjugate", "sin", "cos", "tan",
           "sinh", "cosh", "sinc", "absolute", "sign"]
Martin Reinecke's avatar
Martin Reinecke committed
40

Martin Reinecke's avatar
tweaks  
Martin Reinecke committed
41 42 43
_comm = MPI.COMM_WORLD
ntask = _comm.Get_size()
rank = _comm.Get_rank()
Martin Reinecke's avatar
Martin Reinecke committed
44
master = (rank == 0)
45 46


Martin Reinecke's avatar
Martin Reinecke committed
47 48 49 50
def is_numpy():
    return False


Martin Reinecke's avatar
tweaks  
Martin Reinecke committed
51
def _shareSize(nwork, nshares, myshare):
Martin Reinecke's avatar
Martin Reinecke committed
52
    return (nwork//nshares) + int(myshare < nwork % nshares)
Martin Reinecke's avatar
Martin Reinecke committed
53

Martin Reinecke's avatar
tweaks  
Martin Reinecke committed
54 55

def _shareRange(nwork, nshares, myshare):
Martin Reinecke's avatar
Martin Reinecke committed
56 57
    nbase = nwork//nshares
    additional = nwork % nshares
Martin Reinecke's avatar
Martin Reinecke committed
58
    lo = myshare*nbase + min(myshare, additional)
Martin Reinecke's avatar
Martin Reinecke committed
59
    hi = lo + nbase + int(myshare < additional)
Martin Reinecke's avatar
Martin Reinecke committed
60 61
    return lo, hi

62

63
def local_shape(shape, distaxis=0):
Martin Reinecke's avatar
Martin Reinecke committed
64
    if len(shape) == 0 or distaxis == -1:
65
        return shape
Martin Reinecke's avatar
Martin Reinecke committed
66 67
    shape2 = list(shape)
    shape2[distaxis] = _shareSize(shape[distaxis], ntask, rank)
68 69
    return tuple(shape2)

Martin Reinecke's avatar
Martin Reinecke committed
70

71 72
class data_object(object):
    def __init__(self, shape, data, distaxis):
Martin Reinecke's avatar
Martin Reinecke committed
73
        self._shape = tuple(shape)
Martin Reinecke's avatar
Martin Reinecke committed
74
        if len(self._shape) == 0:
Martin Reinecke's avatar
fixes  
Martin Reinecke committed
75
            distaxis = -1
Martin Reinecke's avatar
Martin Reinecke committed
76 77
            if not isinstance(data, np.ndarray):
                data = np.full((), data)
78 79
        self._distaxis = distaxis
        self._data = data
Martin Reinecke's avatar
Martin Reinecke committed
80 81
        if local_shape(self._shape, self._distaxis) != self._data.shape:
            raise ValueError("shape mismatch")
82

83 84 85
    def copy(self):
        return data_object(self._shape, self._data.copy(), self._distaxis)

Martin Reinecke's avatar
Martin Reinecke committed
86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111
#     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
112 113 114 115 116 117 118 119 120 121 122 123 124 125 126

    @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
127
        return data_object(self._shape, self._data.real, self._distaxis)
128 129 130

    @property
    def imag(self):
Martin Reinecke's avatar
Martin Reinecke committed
131
        return data_object(self._shape, self._data.imag, self._distaxis)
132

Martin Reinecke's avatar
Martin Reinecke committed
133 134 135 136 137 138
    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
139
    def _contraction_helper(self, op, mpiop, axis):
140
        if axis is not None:
Martin Reinecke's avatar
Martin Reinecke committed
141
            if len(axis) == len(self._data.shape):
142 143
                axis = None
        if axis is None:
Martin Reinecke's avatar
Martin Reinecke committed
144
            res = np.array(getattr(self._data, op)())
Martin Reinecke's avatar
Martin Reinecke committed
145
            if (self._distaxis == -1):
Martin Reinecke's avatar
fixes  
Martin Reinecke committed
146
                return res[()]
Martin Reinecke's avatar
Martin Reinecke committed
147 148
            res2 = np.empty((), dtype=res.dtype)
            _comm.Allreduce(res, res2, mpiop)
Martin Reinecke's avatar
fixes  
Martin Reinecke committed
149
            return res2[()]
150 151

        if self._distaxis in axis:
Martin Reinecke's avatar
Martin Reinecke committed
152 153
            res = getattr(self._data, op)(axis=axis)
            res2 = np.empty_like(res)
Martin Reinecke's avatar
Martin Reinecke committed
154
            _comm.Allreduce(res, res2, mpiop)
Martin Reinecke's avatar
Martin Reinecke committed
155
            return from_global_data(res2, distaxis=0)
156
        else:
Martin Reinecke's avatar
Martin Reinecke committed
157
            # perform the contraction on the local data
Martin Reinecke's avatar
Martin Reinecke committed
158 159
            res = getattr(self._data, op)(axis=axis)
            if self._distaxis == -1:
Martin Reinecke's avatar
Martin Reinecke committed
160
                return from_global_data(res, distaxis=0)
Martin Reinecke's avatar
Martin Reinecke committed
161
            shp = list(res.shape)
Martin Reinecke's avatar
Martin Reinecke committed
162
            shift = 0
Martin Reinecke's avatar
Martin Reinecke committed
163
            for ax in axis:
Martin Reinecke's avatar
Martin Reinecke committed
164 165
                if ax < self._distaxis:
                    shift += 1
Martin Reinecke's avatar
Martin Reinecke committed
166 167
            shp[self._distaxis-shift] = self.shape[self._distaxis]
            return from_local_data(shp, res, self._distaxis-shift)
168 169 170

    def sum(self, axis=None):
        return self._contraction_helper("sum", MPI.SUM, axis)
Martin Reinecke's avatar
Martin Reinecke committed
171

172 173 174
    def prod(self, axis=None):
        return self._contraction_helper("prod", MPI.PROD, axis)

175 176
#    def min(self, axis=None):
#        return self._contraction_helper("min", MPI.MIN, axis)
Martin Reinecke's avatar
Martin Reinecke committed
177

178 179
#    def max(self, axis=None):
#        return self._contraction_helper("max", MPI.MAX, axis)
180

181 182 183 184 185 186
    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
187

188 189
    def std(self, axis=None):
        return np.sqrt(self.var(axis))
Martin Reinecke's avatar
Martin Reinecke committed
190

Martin Reinecke's avatar
Martin Reinecke committed
191
    # FIXME: to be improved!
192 193 194
    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
195 196
        return (abs(self-self.mean())**2).mean()

197
    def _binary_helper(self, other, op):
Martin Reinecke's avatar
Martin Reinecke committed
198
        a = self
199
        if isinstance(other, data_object):
Martin Reinecke's avatar
Martin Reinecke committed
200
            b = other
201 202 203 204
            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
205 206
            a = a._data
            b = b._data
Martin Reinecke's avatar
Martin Reinecke committed
207 208 209 210 211
        elif np.isscalar(other):
            a = a._data
            b = other
        else:
            return NotImplemented
212 213

        tval = getattr(a, op)(b)
Martin Reinecke's avatar
Martin Reinecke committed
214 215 216 217
        if tval is a:
            return self
        else:
            return data_object(self._shape, tval, self._distaxis)
218 219

    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

235 236 237 238 239 240 241 242 243 244 245 246 247 248
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
249

Martin Reinecke's avatar
Martin Reinecke committed
250
def full(shape, fill_value, dtype=None, distaxis=0):
Martin Reinecke's avatar
Martin Reinecke committed
251 252
    return data_object(shape, np.full(local_shape(shape, distaxis),
                                      fill_value, dtype), distaxis)
253 254


Martin Reinecke's avatar
Martin Reinecke committed
255 256 257 258 259 260
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
261
def empty(shape, dtype=None, distaxis=0):
Martin Reinecke's avatar
Martin Reinecke committed
262 263
    return data_object(shape, np.empty(local_shape(shape, distaxis),
                                       dtype), distaxis)
264 265


Martin Reinecke's avatar
fixes  
Martin Reinecke committed
266
def zeros(shape, dtype=None, distaxis=0):
Martin Reinecke's avatar
Martin Reinecke committed
267 268
    return data_object(shape, np.zeros(local_shape(shape, distaxis), dtype),
                       distaxis)
269 270


Martin Reinecke's avatar
fixes  
Martin Reinecke committed
271
def ones(shape, dtype=None, distaxis=0):
Martin Reinecke's avatar
Martin Reinecke committed
272 273
    return data_object(shape, np.ones(local_shape(shape, distaxis), dtype),
                       distaxis)
274 275 276 277 278 279 280


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
281
    tmp = np.array(np.vdot(a._data, b._data))
Martin Reinecke's avatar
Martin Reinecke committed
282
    if a._distaxis == -1:
Martin Reinecke's avatar
Martin Reinecke committed
283
        return tmp[()]
Martin Reinecke's avatar
Martin Reinecke committed
284 285
    res = np.empty((), dtype=tmp.dtype)
    _comm.Allreduce(tmp, res, MPI.SUM)
Martin Reinecke's avatar
fixes  
Martin Reinecke committed
286
    return res[()]
287 288 289


def _math_helper(x, function, out):
290
    function = getattr(np, function)
291 292 293 294
    if out is not None:
        function(x._data, out=out._data)
        return out
    else:
Martin Reinecke's avatar
Martin Reinecke committed
295
        return data_object(x.shape, function(x._data), x._distaxis)
296 297


298
_current_module = sys.modules[__name__]
Martin Reinecke's avatar
Martin Reinecke committed
299

300 301
for f in ["sqrt", "exp", "log", "tanh", "conjugate", "sin", "cos", "tan",
          "sinh", "cosh", "sinc", "absolute", "sign"]:
302 303 304 305 306
    def func(f):
        def func2(x, out=None):
            return _math_helper(x, f, out)
        return func2
    setattr(_current_module, f, func(f))
307 308


309
def clipped_exp(x):
Martin Reinecke's avatar
fix  
Martin Reinecke committed
310
    return data_object(x.shape, np.exp(np.clip(x.data, -300, 300), x.distaxis))
Martin Reinecke's avatar
tweaks  
Martin Reinecke committed
311 312


313 314 315 316
def hardplus(x):
    return data_object(x.shape, np.clip(x.data, 1e-20, None), x.distaxis)


Martin Reinecke's avatar
Martin Reinecke committed
317 318 319 320 321 322 323 324 325 326 327 328
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
329
        data.flags.writeable = False
Martin Reinecke's avatar
Martin Reinecke committed
330
    return data_object(object._shape, data, distaxis=object._distaxis)
331 332


Martin Reinecke's avatar
Martin Reinecke committed
333 334
# 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
335 336 337
# 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
338
def from_random(random_type, shape, dtype=np.float64, **kwargs):
339
    generator_function = getattr(Random, random_type)
Martin Reinecke's avatar
Martin Reinecke committed
340
    if len(shape) == 0:
Martin Reinecke's avatar
Martin Reinecke committed
341 342 343
        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
344 345 346 347 348 349 350
    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)
351

Martin Reinecke's avatar
Martin Reinecke committed
352

Martin Reinecke's avatar
Martin Reinecke committed
353 354 355 356
def local_data(arr):
    return arr._data


357 358
def ibegin_from_shape(glob_shape, distaxis=0):
    res = [0] * len(glob_shape)
Martin Reinecke's avatar
Martin Reinecke committed
359
    if distaxis < 0:
360 361 362 363 364
        return res
    res[distaxis] = _shareRange(glob_shape[distaxis], ntask, rank)[0]
    return tuple(res)


Martin Reinecke's avatar
fixes  
Martin Reinecke committed
365 366
def ibegin(arr):
    res = [0] * arr._data.ndim
Martin Reinecke's avatar
Martin Reinecke committed
367
    res[arr._distaxis] = _shareRange(arr._shape[arr._distaxis], ntask, rank)[0]
Martin Reinecke's avatar
fixes  
Martin Reinecke committed
368
    return tuple(res)
Martin Reinecke's avatar
Martin Reinecke committed
369 370


Martin Reinecke's avatar
fixes  
Martin Reinecke committed
371 372
def np_allreduce_sum(arr):
    res = np.empty_like(arr)
Martin Reinecke's avatar
Martin Reinecke committed
373
    _comm.Allreduce(arr, res, MPI.SUM)
Martin Reinecke's avatar
fixes  
Martin Reinecke committed
374
    return res
Martin Reinecke's avatar
Martin Reinecke committed
375 376


377 378 379 380 381 382
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
383 384 385 386 387 388
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
389 390 391 392
def distaxis(arr):
    return arr._distaxis


Martin Reinecke's avatar
Martin Reinecke committed
393
def from_local_data(shape, arr, distaxis=0):
Martin Reinecke's avatar
Martin Reinecke committed
394 395 396
    return data_object(shape, arr, distaxis)


397 398 399
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
400
    if distaxis == -1:
Martin Reinecke's avatar
Martin Reinecke committed
401
        return data_object(arr.shape, arr, distaxis)
Martin Reinecke's avatar
Martin Reinecke committed
402
    lo, hi = _shareRange(arr.shape[distaxis], ntask, rank)
Martin Reinecke's avatar
Martin Reinecke committed
403
    sl = [slice(None)]*len(arr.shape)
Martin Reinecke's avatar
Martin Reinecke committed
404
    sl[distaxis] = slice(lo, hi)
405
    return data_object(arr.shape, arr[tuple(sl)], distaxis)
Martin Reinecke's avatar
Martin Reinecke committed
406 407


Martin Reinecke's avatar
Martin Reinecke committed
408 409
def to_global_data(arr):
    if arr._distaxis == -1:
Martin Reinecke's avatar
fixes  
Martin Reinecke committed
410 411 412 413 414
        return arr._data
    tmp = redistribute(arr, dist=-1)
    return tmp._data


415 416 417 418 419 420 421
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
422
def redistribute(arr, dist=None, nodist=None):
Martin Reinecke's avatar
Martin Reinecke committed
423 424 425
    if dist is not None:
        if nodist is not None:
            raise ValueError
Martin Reinecke's avatar
Martin Reinecke committed
426
        if dist == arr._distaxis:
Martin Reinecke's avatar
Martin Reinecke committed
427 428 429 430 431 432
            return arr
    else:
        if nodist is None:
            raise ValueError
        if arr._distaxis not in nodist:
            return arr
Martin Reinecke's avatar
Martin Reinecke committed
433
        dist = -1
Martin Reinecke's avatar
Martin Reinecke committed
434 435
        for i in range(len(arr.shape)):
            if i not in nodist:
Martin Reinecke's avatar
Martin Reinecke committed
436
                dist = i
Martin Reinecke's avatar
Martin Reinecke committed
437
                break
Martin Reinecke's avatar
Martin Reinecke committed
438

Martin Reinecke's avatar
Martin Reinecke committed
439
    if arr._distaxis == -1:  # all data available, just pick the proper subset
440
        return from_global_data(arr._data, distaxis=dist)
Martin Reinecke's avatar
Martin Reinecke committed
441
    if dist == -1:  # gather all data on all tasks
Martin Reinecke's avatar
Martin Reinecke committed
442
        tmp = np.moveaxis(arr._data, arr._distaxis, 0)
Martin Reinecke's avatar
Martin Reinecke committed
443 444
        slabsize = np.prod(tmp.shape[1:])*tmp.itemsize
        sz = np.empty(ntask, dtype=np.int)
Martin Reinecke's avatar
Martin Reinecke committed
445
        for i in range(ntask):
Martin Reinecke's avatar
Martin Reinecke committed
446 447 448 449
            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
450
        tmp = np.require(tmp, requirements="C")
Martin Reinecke's avatar
Martin Reinecke committed
451 452
        out = np.empty(arr.size, dtype=arr.dtype)
        _comm.Allgatherv(tmp, [out, sz, disp, MPI.BYTE])
Martin Reinecke's avatar
Martin Reinecke committed
453 454 455 456
        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
457
        out = np.moveaxis(out, 0, arr._distaxis)
Martin Reinecke's avatar
Martin Reinecke committed
458
        return from_global_data(out, distaxis=-1)
Martin Reinecke's avatar
Martin Reinecke committed
459

Martin Reinecke's avatar
Martin Reinecke committed
460
    # real redistribution via Alltoallv
Martin Reinecke's avatar
Martin Reinecke committed
461
    ssz0 = arr._data.size//arr.shape[dist]
Martin Reinecke's avatar
Martin Reinecke committed
462
    ssz = np.empty(ntask, dtype=np.int)
Martin Reinecke's avatar
Martin Reinecke committed
463 464 465
    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
466
    rsz = np.empty(ntask, dtype=np.int)
Martin Reinecke's avatar
Martin Reinecke committed
467 468 469 470 471 472 473 474 475 476 477 478
    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
479
            sslice[dist] = slice(lo, hi)
Martin Reinecke's avatar
Martin Reinecke committed
480
            ssz[i] = ssz0*(hi-lo)
481
            sbuf[ofs:ofs+ssz[i]] = arr._data[tuple(sslice)].flat
Martin Reinecke's avatar
Martin Reinecke committed
482 483 484 485
            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
486 487
    sdisp = np.append(0, np.cumsum(ssz[:-1]))
    rdisp = np.append(0, np.cumsum(rsz[:-1]))
Martin Reinecke's avatar
Martin Reinecke committed
488 489
    s_msg = [sbuf, (ssz, sdisp), MPI.BYTE]
    r_msg = [rbuf, (rsz, rdisp), MPI.BYTE]
Martin Reinecke's avatar
tweaks  
Martin Reinecke committed
490
    _comm.Alltoallv(s_msg, r_msg)
Martin Reinecke's avatar
Martin Reinecke committed
491
    del sbuf  # free memory
Martin Reinecke's avatar
Martin Reinecke committed
492 493 494 495
    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
496
        arrnew = np.empty(local_shape(arr.shape, dist), dtype=arr.dtype)
Martin Reinecke's avatar
Martin Reinecke committed
497 498 499 500
        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
501
            rslice[arr._distaxis] = slice(lo, hi)
Martin Reinecke's avatar
Martin Reinecke committed
502
            sz = rsz[i]//arr._data.itemsize
503
            arrnew[tuple(rslice)].flat = rbuf[ofs:ofs+sz]
Martin Reinecke's avatar
Martin Reinecke committed
504
            ofs += sz
Martin Reinecke's avatar
Martin Reinecke committed
505
        arrnew = from_local_data(arr.shape, arrnew, distaxis=dist)
Martin Reinecke's avatar
Martin Reinecke committed
506
    return arrnew
Martin Reinecke's avatar
Martin Reinecke committed
507 508


Martin Reinecke's avatar
Martin Reinecke committed
509 510
def transpose(arr):
    if len(arr.shape) != 2 or arr._distaxis != 0:
Martin Reinecke's avatar
Martin Reinecke committed
511
        raise ValueError("bad input")
Martin Reinecke's avatar
Martin Reinecke committed
512 513 514 515 516 517 518 519 520 521 522
    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
523
        sbuf[ofs:ofs+ssz[i]] = arr._data[:, lo:hi].flat
Martin Reinecke's avatar
Martin Reinecke committed
524 525 526 527 528 529 530 531 532 533
        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
534
    sz2 = _shareSize(arr.shape[1], ntask, rank)
Martin Reinecke's avatar
Martin Reinecke committed
535 536
    arrnew = np.empty((sz2, arr.shape[0]), dtype=arr.dtype)
    ofs = 0
Martin Reinecke's avatar
Martin Reinecke committed
537 538 539
    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
540
        arrnew[:, lo:hi] = rbuf[ofs:ofs+sz].reshape(hi-lo, sz2).T
Martin Reinecke's avatar
Martin Reinecke committed
541
        ofs += sz
542
    return from_local_data((arr.shape[1], arr.shape[0]), arrnew, 0)
Martin Reinecke's avatar
Martin Reinecke committed
543 544


Martin Reinecke's avatar
Martin Reinecke committed
545 546
def default_distaxis():
    return 0
547 548 549 550 551 552 553 554


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


def locked(arr):
    return not arr._data.flags.writeable
Martin Reinecke's avatar
Martin Reinecke committed
555 556 557 558 559 560 561 562 563 564 565 566


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
567 568 569 570 571 572


def absmax(arr):
    if arr._data.size == 0:
        tmp = np.array(0, dtype=arr._data.dtype)
    else:
573
        tmp = np.asarray(np.linalg.norm(arr._data.reshape(-1), ord=np.inf))
Martin Reinecke's avatar
Martin Reinecke committed
574 575 576 577 578 579 580 581
    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)
582
    tmp = np.asarray(np.linalg.norm(arr._data.reshape(-1), ord=ord) ** ord)
Martin Reinecke's avatar
Martin Reinecke committed
583
    res = np.empty_like(tmp)
Martin Reinecke's avatar
Martin Reinecke committed
584 585 586 587
    if len(arr._data.shape) == 0:
        res = tmp
    else:
        _comm.Allreduce(tmp, res, MPI.SUM)
Martin Reinecke's avatar
Martin Reinecke committed
588
    return res[()] ** (1./ord)