distributed_do.py 10.2 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
import numpy as np
from .random import Random
from mpi4py import MPI

comm = MPI.COMM_WORLD
ntask = comm.Get_size()
rank = comm.Get_rank()


def shareSize(nwork, nshares, myshare):
    nbase = nwork//nshares
    return nbase if myshare>=nwork%nshares else nbase+1
Martin Reinecke's avatar
Martin Reinecke committed
13
14
15
16
17
18
def shareRange(nwork, nshares, myshare):
    nbase = nwork//nshares;
    additional = nwork%nshares;
    lo = myshare*nbase + min(myshare, additional)
    hi = lo+nbase+ (1 if myshare<additional else 0)
    return lo,hi
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45

def get_locshape(shape, distaxis):
    if distaxis==-1:
        return shape
    shape2=list(shape)
    shape2[distaxis]=shareSize(shape[distaxis],ntask,rank)
    return tuple(shape2)

class data_object(object):
    def __init__(self, shape, data, distaxis):
        """Must not be called directly by users"""
        self._shape = shape
        self._distaxis = distaxis
        lshape = get_locshape(self._shape, self._distaxis)
        self._data = data

    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)
Martin Reinecke's avatar
Martin Reinecke committed
46
        otmp=np.empty((ntask,len(self._shape)),dtype=np.int)
47
48
        comm.Allgather(itmp,otmp)
        for i in range(ntask):
Martin Reinecke's avatar
Martin Reinecke committed
49
            if np.any(otmp[i,:]!=self._shape):
50
51
52
53
54
55
56
                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)
Martin Reinecke's avatar
Martin Reinecke committed
57
58
            itmp[self._distaxis] = shareSize(self._shape[self._distaxis],ntask,rank)
            if np.any(self._data.shape!=itmp):
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
                raise ValueError

    @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
75
        return data_object(self._shape, self._data.real, self._distaxis)
76
77
78

    @property
    def imag(self):
Martin Reinecke's avatar
Martin Reinecke committed
79
        return data_object(self._shape, self._data.imag, self._distaxis)
80

Martin Reinecke's avatar
Martin Reinecke committed
81
    def _contraction_helper(self, op, mpiop, axis):
82
83
84
85
        if axis is not None:
            if len(axis)==len(self._data.shape):
                axis = None
        if axis is None:
Martin Reinecke's avatar
Martin Reinecke committed
86
87
88
89
            res = np.array(getattr(self._data, op)())
            if (self._distaxis==-1):
                return res
            res2 = np.empty(1,dtype=res.dtype)
90
            MPI.COMM_WORLD.Allreduce(res,res2,mpiop)
Martin Reinecke's avatar
Martin Reinecke committed
91
            return res2[0]
92
93
94
95

        if self._distaxis in axis:
            pass# reduce globally, redistribute the result along axis 0(?)
        else:
Martin Reinecke's avatar
Martin Reinecke committed
96
97
98
            # perform the contraction on the local data
            data = getattr(self._data, op)(axis=axis)
            #shp =
99
100
101
102
103
104
105
106
107
108
109

        # check if the result is scalar or if a result_field must be constr.
        if np.isscalar(data):
            return data
        else:
            return data_object(data)

    def sum(self, axis=None):
        return self._contraction_helper("sum", MPI.SUM, axis)

    def _binary_helper(self, other, op):
Martin Reinecke's avatar
Martin Reinecke committed
110
        a = self
111
        if isinstance(other, data_object):
Martin Reinecke's avatar
Martin Reinecke committed
112
            b = other
113
114
115
116
            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
117
118
            a = a._data
            b = b._data
119
        else:
Martin Reinecke's avatar
Martin Reinecke committed
120
            a = a._data
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
            b = other

        tval = getattr(a, op)(b)
        return self if tval is a else data_object(self._shape, tval, self._distaxis)

    def __add__(self, other):
        return self._binary_helper(other, op='__add__')

    def __radd__(self, other):
        return self._binary_helper(other, op='__radd__')

    def __iadd__(self, other):
        return self._binary_helper(other, op='__iadd__')

    def __sub__(self, other):
        return self._binary_helper(other, op='__sub__')

    def __rsub__(self, other):
        return self._binary_helper(other, op='__rsub__')

    def __isub__(self, other):
        return self._binary_helper(other, op='__isub__')

    def __mul__(self, other):
        return self._binary_helper(other, op='__mul__')

    def __rmul__(self, other):
        return self._binary_helper(other, op='__rmul__')

    def __imul__(self, other):
        return self._binary_helper(other, op='__imul__')

    def __div__(self, other):
        return self._binary_helper(other, op='__div__')

    def __rdiv__(self, other):
        return self._binary_helper(other, op='__rdiv__')

    def __truediv__(self, other):
        return self._binary_helper(other, op='__truediv__')

    def __rtruediv__(self, other):
        return self._binary_helper(other, op='__rtruediv__')

    def __pow__(self, other):
        return self._binary_helper(other, op='__pow__')

    def __rpow__(self, other):
        return self._binary_helper(other, op='__rpow__')

    def __ipow__(self, other):
        return self._binary_helper(other, op='__ipow__')

    def __eq__(self, other):
        return self._binary_helper(other, op='__eq__')

    def __ne__(self, other):
        return self._binary_helper(other, op='__ne__')

    def __neg__(self):
        return data_object(-self._data)

    def __abs__(self):
        return data_object(np.abs(self._data))

    def ravel(self):
        return data_object(self._data.ravel())

    def reshape(self, shape):
        return data_object(self._data.reshape(shape))

    def all(self):
        return self._data.all()

    def any(self):
        return self._data.any()


Martin Reinecke's avatar
Martin Reinecke committed
199
200
def full(shape, fill_value, dtype=None, distaxis=0):
    return data_object(shape, np.full(shape, local_shape(shape, distaxis), fill_value, dtype))
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
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268


def empty(shape, dtype=np.float):
    return data_object(np.empty(shape, dtype))


def zeros(shape, dtype=np.float):
    return data_object(np.zeros(shape, dtype))


def ones(shape, dtype=np.float):
    return data_object(np.ones(shape, dtype))


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


def vdot(a, b):
    return np.vdot(a._data, b._data)


def _math_helper(x, function, out):
    if out is not None:
        function(x._data, out=out._data)
        return out
    else:
        return data_object(function(x._data))


def abs(a, out=None):
    return _math_helper(a, np.abs, out)


def exp(a, out=None):
    return _math_helper(a, np.exp, out)


def log(a, out=None):
    return _math_helper(a, np.log, out)


def sqrt(a, out=None):
    return _math_helper(a, np.sqrt, out)


def bincount(x, weights=None, minlength=None):
    if weights is not None:
        weights = weights._data
    res = np.bincount(x._data, weights, minlength)
    return data_object(res)


def from_object(object, dtype=None, copy=True):
    return data_object(np.array(object._data, dtype=dtype, copy=copy))


def from_random(random_type, shape, dtype=np.float64, **kwargs):
    generator_function = getattr(Random, random_type)
    return data_object(generator_function(dtype=dtype, shape=shape, **kwargs))


def to_ndarray(arr):
    return arr._data


def from_ndarray(arr):
    return data_object(arr.shape,arr,-1)
Martin Reinecke's avatar
Martin Reinecke committed
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362


def local_data(arr):
    return arr._data


#def ibegin(arr):
#    return (0,)*arr._data.ndim


#def np_allreduce_sum(arr):
#    return arr


def distaxis(arr):
    return arr._distaxis


def from_local_data (shape, arr, distaxis):
    return data_object(shape, arr, distaxis)


def from_global_data (arr, distaxis=0):
    if distaxis==-1:
        return data_object(arr.shape, arr, distaxis)
    lo, hi = shareRange(arr.shape[distaxis],ntask,rank)
    sl = [slice(None)]*len(arr.shape)
    sl[distaxis]=slice(lo,hi)
    return data_object(arr.shape, arr[sl], distaxis)


def redistribute (arr, dist=None, nodist=None):
    if dist is not None:
        if nodist is not None:
            raise ValueError
        if dist==arr._distaxis:
            return arr
    else:
        if nodist is None:
            raise ValueError
        if arr._distaxis not in nodist:
            return arr
        dist=-1
        for i in range(len(arr.shape)):
            if i not in nodist:
                dist=i
                break
    if arr._distaxis==-1:  # just pick the proper subset
        return from_global_data(arr._data, dist)
    if dist==-1: # gather data
        tmp = np.moveaxis(arr._data, arr._distaxis, 0)
        slabsize=np.prod(tmp.shape[1:])*tmp.itemsize
        sz=np.empty(ntask,dtype=np.int)
        for i in range(ntask):
            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])
        tmp=tmp.flatten()
        out = np.empty(arr.size,dtype=arr.dtype)
        print tmp.shape, out.shape, sz, disp
        comm.Allgatherv(tmp,[out,sz,disp,MPI.BYTE])
        out = out.reshape(arr._shape)
        out = np.moveaxis(out, 0, arr._distaxis)
        return from_global_data (out, distaxis=-1)
    # real redistribution via Alltoallv
    tmp = np.moveaxis(arr._data, (dist, arr._distaxis), (0, 1))
    tshape = tmp.shape
    slabsize=np.prod(tmp.shape[2:])*tmp.itemsize
    ssz=np.empty(ntask,dtype=np.int)
    rsz=np.empty(ntask,dtype=np.int)
    for i in range(ntask):
        ssz[i]=slabsize*tmp.shape[1]*shareSize(arr.shape[dist],ntask,i)
        rsz[i]=slabsize*shareSize(arr.shape[dist],ntask,rank)*shareSize(arr.shape[arr._distaxis],ntask,i)
    sdisp=np.empty(ntask,dtype=np.int)
    rdisp=np.empty(ntask,dtype=np.int)
    sdisp[0]=0
    rdisp[0]=0
    sdisp[1:]=np.cumsum(ssz[:-1])
    rdisp[1:]=np.cumsum(rsz[:-1])
    print ssz, rsz
    tmp=tmp.flatten()
    out = np.empty(np.prod(get_locshape(arr.shape,dist)),dtype=arr.dtype)
    s_msg = [tmp, (ssz, sdisp), MPI.BYTE]
    r_msg = [out, (rsz, rdisp), MPI.BYTE]
    comm.Alltoallv(s_msg, r_msg)
    new_shape = [shareSize(arr.shape[dist],ntask,rank), arr.shape[arr._distaxis]] +list(tshape[2:])
    out=out.reshape(new_shape)
    out = np.moveaxis(out, (0, 1), (dist, arr._distaxis))
    return from_local_data (arr.shape, out, dist)


def default_distaxis():
    return 0