1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
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
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
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
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
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
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
|
from __future__ import division
import numpy as np
from .elemwise import elemwise1, elemwise2, ielemwise2, compare, arg, GpuElemwise, as_argument
from .reduction import reduce1
from .dtypes import dtype_to_ctype, get_np_obj, get_common_dtype
from . import gpuarray
class ndgpuarray(gpuarray.GpuArray):
"""
Extension class for gpuarray.GpuArray to add numpy mathematical
operations between arrays. These operations are all performed on
the GPU but this is not the most efficient way since it will
involve the creation of temporaries (just like numpy) for all
intermediate results.
This class may help transition code from numpy to pygpu by acting
more like a drop-in replacement for numpy.ndarray than the raw
GpuArray class.
"""
# add
def __add__(self, other):
return elemwise2(self, '+', other, self, broadcast=True)
def __radd__(self, other):
return elemwise2(other, '+', self, self, broadcast=True)
def __iadd__(self, other):
return ielemwise2(self, '+', other, broadcast=True)
# sub
def __sub__(self, other):
return elemwise2(self, '-', other, self, broadcast=True)
def __rsub__(self, other):
return elemwise2(other, '-', self, self, broadcast=True)
def __isub__(self, other):
return ielemwise2(self, '-', other, broadcast=True)
# mul
def __mul__(self, other):
return elemwise2(self, '*', other, self, broadcast=True)
def __rmul__(self, other):
return elemwise2(other, '*', self, self, broadcast=True)
def __imul__(self, other):
return ielemwise2(self, '*', other, broadcast=True)
# div
def __div__(self, other):
return elemwise2(self, '/', other, self, broadcast=True)
def __rdiv__(self, other):
return elemwise2(other, '/', self, self, broadcast=True)
def __idiv__(self, other):
return ielemwise2(self, '/', other, broadcast=True)
# truediv
def __truediv__(self, other):
np1 = get_np_obj(self)
np2 = get_np_obj(other)
res = (np1.__truediv__(np2)).dtype
return elemwise2(self, '/', other, self, odtype=res, broadcast=True)
def __rtruediv__(self, other):
np1 = get_np_obj(self)
np2 = get_np_obj(other)
res = (np2.__truediv__(np1)).dtype
return elemwise2(other, '/', self, self, odtype=res, broadcast=True)
def __itruediv__(self, other):
np2 = get_np_obj(other)
kw = {'broadcast': True}
if self.dtype == np.float32 or np2.dtype == np.float32:
kw['op_tmpl'] = "a = (float)a / (float)b"
if self.dtype == np.float64 or np2.dtype == np.float64:
kw['op_tmpl'] = "a = (double)a / (double)b"
return ielemwise2(self, '/', other, **kw)
# floordiv
def __floordiv__(self, other):
out_dtype = get_common_dtype(self, other, True)
kw = {'broadcast': True}
if out_dtype.kind == 'f':
kw['op_tmpl'] = "res = floor((%(out_t)s)a / (%(out_t)s)b)"
return elemwise2(self, '/', other, self, odtype=out_dtype, **kw)
def __rfloordiv__(self, other):
out_dtype = get_common_dtype(other, self, True)
kw = {'broadcast': True}
if out_dtype.kind == 'f':
kw['op_tmpl'] = "res = floor((%(out_t)s)a / (%(out_t)s)b)"
return elemwise2(other, '/', self, self, odtype=out_dtype, **kw)
def __ifloordiv__(self, other):
out_dtype = self.dtype
kw = {'broadcast': True}
if out_dtype == np.float32:
kw['op_tmpl'] = "a = floor((float)a / (float)b)"
if out_dtype == np.float64:
kw['op_tmpl'] = "a = floor((double)a / (double)b)"
return ielemwise2(self, '/', other, **kw)
# mod
def __mod__(self, other):
out_dtype = get_common_dtype(self, other, True)
kw = {'broadcast': True}
if out_dtype.kind == 'f':
kw['op_tmpl'] = "res = fmod((%(out_t)s)a, (%(out_t)s)b)"
return elemwise2(self, '%', other, self, odtype=out_dtype, **kw)
def __rmod__(self, other):
out_dtype = get_common_dtype(other, self, True)
kw = {'broadcast': True}
if out_dtype.kind == 'f':
kw['op_tmpl'] = "res = fmod((%(out_t)s)a, (%(out_t)s)b)"
return elemwise2(other, '%', self, self, odtype=out_dtype, **kw)
def __imod__(self, other):
out_dtype = get_common_dtype(self, other, self.dtype == np.float64)
kw = {'broadcast': True}
if out_dtype == np.float32:
kw['op_tmpl'] = "a = fmod((float)a, (float)b)"
if out_dtype == np.float64:
kw['op_tmpl'] = "a = fmod((double)a, (double)b)"
return ielemwise2(self, '%', other, **kw)
# divmod
def __divmod__(self, other):
if not isinstance(other, gpuarray.GpuArray):
other = np.asarray(other)
odtype = get_common_dtype(self, other, True)
a_arg = as_argument(self, 'a', read=True)
b_arg = as_argument(other, 'b', read=True)
args = [arg('div', odtype, write=True), arg('mod', odtype, write=True), a_arg, b_arg]
div = self._empty_like_me(dtype=odtype)
mod = self._empty_like_me(dtype=odtype)
if odtype.kind == 'f':
tmpl = ("div = floor((%(out_t)s)a / (%(out_t)s)b),"
"mod = fmod((%(out_t)s)a, (%(out_t)s)b)")
else:
tmpl = ("div = (%(out_t)s)a / (%(out_t)s)b,"
"mod = a %% b")
ksrc = tmpl % {'out_t': dtype_to_ctype(odtype)}
k = GpuElemwise(self.context, ksrc, args)
k(div, mod, self, other, broadcast=True)
return (div, mod)
def __rdivmod__(self, other):
if not isinstance(other, gpuarray.GpuArray):
other = np.asarray(other)
odtype = get_common_dtype(other, self, True)
a_arg = as_argument(other, 'a', read=True)
b_arg = as_argument(self, 'b', read=True)
args = [arg('div', odtype, write=True), arg('mod', odtype, write=True), a_arg, b_arg]
div = self._empty_like_me(dtype=odtype)
mod = self._empty_like_me(dtype=odtype)
if odtype.kind == 'f':
tmpl = ("div = floor((%(out_t)s)a / (%(out_t)s)b),"
"mod = fmod((%(out_t)s)a, (%(out_t)s)b)")
else:
tmpl = ("div = (%(out_t)s)a / (%(out_t)s)b,"
"mod = a %% b")
ksrc = tmpl % {'out_t': dtype_to_ctype(odtype)}
k = GpuElemwise(self.context, ksrc, args)
k(div, mod, other, self, broadcast=True)
return (div, mod)
def __neg__(self):
return elemwise1(self, '-')
def __pos__(self):
return elemwise1(self, '+')
def __abs__(self):
if self.dtype.kind == 'u':
return self.copy()
if self.dtype.kind == 'f':
oper = "res = fabs(a)"
elif self.dtype.itemsize < 4:
# cuda 5.5 finds the c++ stdlib definition if we don't cast here.
oper = "res = abs((int)a)"
else:
oper = "res = abs(a)"
return elemwise1(self, None, oper=oper)
# richcmp
def __lt__(self, other):
return compare(self, '<', other, broadcast=True)
def __le__(self, other):
return compare(self, '<=', other, broadcast=True)
def __eq__(self, other):
return compare(self, '==', other, broadcast=True)
def __ne__(self, other):
return compare(self, '!=', other, broadcast=True)
def __ge__(self, other):
return compare(self, '>=', other, broadcast=True)
def __gt__(self, other):
return compare(self, '>', other, broadcast=True)
# misc other things
@property
def T(self):
if self.ndim < 2:
return self
return self.transpose()
"""
Since these functions are untested (thus probably wrong), we disable them.
def clip(self, a_min, a_max, out=None):
oper=('res = a > %(max)s ? %(max)s : '
'(a < %(min)s ? %(min)s : a)' % dict(min=a_min, max=a_max))
return elemwise1(self, '', oper=oper, out=out)
def fill(self, value):
self[...] = value
"""
# reductions
def all(self, axis=None, out=None):
if self.ndim == 0:
return self.copy()
return reduce1(self, '&&', '1', np.dtype('bool'),
axis=axis, out=out)
def any(self, axis=None, out=None):
if self.ndim == 0:
return self.copy()
return reduce1(self, '||', '0', np.dtype('bool'),
axis=axis, out=out)
def prod(self, axis=None, dtype=None, out=None):
if dtype is None:
dtype = self.dtype
# we only upcast integers that are smaller than the plaform default
if dtype.kind == 'i':
di = np.dtype('int')
if di.itemsize > dtype.itemsize:
dtype = di
if dtype.kind == 'u':
di = np.dtype('uint')
if di.itemsize > dtype.itemsize:
dtype = di
return reduce1(self, '*', '1', dtype, axis=axis, out=out)
# def max(self, axis=None, out=None);
# nd = self.ndim
# if nd == 0:
# return self.copy()
# idx = (0,) * nd
# n = str(self.__getitem__(idx).__array__())
# return reduce1(self, '', n, self.dtype, axis=axis, out=out,
# oper='max(a, b)')
# def min(self, axis=None, out=None):
# nd = self.ndim
# if nd == 0:
# return self.copy()
# idx = (0,) * nd
# n = str(self.__getitem__(idx).__array__())
# return reduce1(self, '', n, self.dtype, axis=axis, out=out,
# oper='min(a, b)')
def sum(self, axis=None, dtype=None, out=None):
if dtype is None:
dtype = self.dtype
# we only upcast integers that are smaller than the plaform default
if dtype.kind == 'i':
di = np.dtype('int')
if di.itemsize > dtype.itemsize:
dtype = di
if dtype.kind == 'u':
di = np.dtype('uint')
if di.itemsize > dtype.itemsize:
dtype = di
return reduce1(self, '+', '0', dtype, axis=axis, out=out)
|
