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
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
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
|
/*
* This file implements assignment from an ndarray to another ndarray.
*
* Written by Mark Wiebe (mwwiebe@gmail.com)
* Copyright (c) 2011 by Enthought, Inc.
*
* See LICENSE.txt for the license.
*/
#define PY_SSIZE_T_CLEAN
#include <Python.h>
#define NPY_NO_DEPRECATED_API NPY_API_VERSION
#define _MULTIARRAYMODULE
#include <numpy/ndarraytypes.h>
#include "npy_config.h"
#include "npy_pycompat.h"
#include "convert_datatype.h"
#include "methods.h"
#include "shape.h"
#include "lowlevel_strided_loops.h"
#include "array_assign.h"
/*
* Check that array data is both uint-aligned and true-aligned for all array
* elements, as required by the copy/casting code in lowlevel_strided_loops.c
*/
NPY_NO_EXPORT int
copycast_isaligned(int ndim, npy_intp *shape,
PyArray_Descr *dtype, char *data, npy_intp *strides)
{
int aligned;
int big_aln, small_aln;
int uint_aln = npy_uint_alignment(dtype->elsize);
int true_aln = dtype->alignment;
/* uint alignment can be 0, meaning not uint alignable */
if (uint_aln == 0) {
return 0;
}
/*
* As an optimization, it is unnecessary to check the alignment to the
* smaller of (uint_aln, true_aln) if the data is aligned to the bigger of
* the two and the big is a multiple of the small aln. We check the bigger
* one first and only check the smaller if necessary.
*/
if (true_aln >= uint_aln) {
big_aln = true_aln;
small_aln = uint_aln;
}
else {
big_aln = uint_aln;
small_aln = true_aln;
}
aligned = raw_array_is_aligned(ndim, shape, data, strides, big_aln);
if (aligned && big_aln % small_aln != 0) {
aligned = raw_array_is_aligned(ndim, shape, data, strides, small_aln);
}
return aligned;
}
/*
* Assigns the array from 'src' to 'dst'. The strides must already have
* been broadcast.
*
* Returns 0 on success, -1 on failure.
*/
NPY_NO_EXPORT int
raw_array_assign_array(int ndim, npy_intp *shape,
PyArray_Descr *dst_dtype, char *dst_data, npy_intp *dst_strides,
PyArray_Descr *src_dtype, char *src_data, npy_intp *src_strides)
{
int idim;
npy_intp shape_it[NPY_MAXDIMS];
npy_intp dst_strides_it[NPY_MAXDIMS];
npy_intp src_strides_it[NPY_MAXDIMS];
npy_intp coord[NPY_MAXDIMS];
PyArray_StridedUnaryOp *stransfer = NULL;
NpyAuxData *transferdata = NULL;
int aligned, needs_api = 0;
npy_intp src_itemsize = src_dtype->elsize;
NPY_BEGIN_THREADS_DEF;
aligned =
copycast_isaligned(ndim, shape, dst_dtype, dst_data, dst_strides) &&
copycast_isaligned(ndim, shape, src_dtype, src_data, src_strides);
/* Use raw iteration with no heap allocation */
if (PyArray_PrepareTwoRawArrayIter(
ndim, shape,
dst_data, dst_strides,
src_data, src_strides,
&ndim, shape_it,
&dst_data, dst_strides_it,
&src_data, src_strides_it) < 0) {
return -1;
}
/*
* Overlap check for the 1D case. Higher dimensional arrays and
* opposite strides cause a temporary copy before getting here.
*/
if (ndim == 1 && src_data < dst_data &&
src_data + shape_it[0] * src_strides_it[0] > dst_data) {
src_data += (shape_it[0] - 1) * src_strides_it[0];
dst_data += (shape_it[0] - 1) * dst_strides_it[0];
src_strides_it[0] = -src_strides_it[0];
dst_strides_it[0] = -dst_strides_it[0];
}
/* Get the function to do the casting */
if (PyArray_GetDTypeTransferFunction(aligned,
src_strides_it[0], dst_strides_it[0],
src_dtype, dst_dtype,
0,
&stransfer, &transferdata,
&needs_api) != NPY_SUCCEED) {
return -1;
}
if (!needs_api) {
NPY_BEGIN_THREADS;
}
NPY_RAW_ITER_START(idim, ndim, coord, shape_it) {
/* Process the innermost dimension */
stransfer(dst_data, dst_strides_it[0], src_data, src_strides_it[0],
shape_it[0], src_itemsize, transferdata);
} NPY_RAW_ITER_TWO_NEXT(idim, ndim, coord, shape_it,
dst_data, dst_strides_it,
src_data, src_strides_it);
NPY_END_THREADS;
NPY_AUXDATA_FREE(transferdata);
return (needs_api && PyErr_Occurred()) ? -1 : 0;
}
/*
* Assigns the array from 'src' to 'dst, wherever the 'wheremask'
* value is True. The strides must already have been broadcast.
*
* Returns 0 on success, -1 on failure.
*/
NPY_NO_EXPORT int
raw_array_wheremasked_assign_array(int ndim, npy_intp *shape,
PyArray_Descr *dst_dtype, char *dst_data, npy_intp *dst_strides,
PyArray_Descr *src_dtype, char *src_data, npy_intp *src_strides,
PyArray_Descr *wheremask_dtype, char *wheremask_data,
npy_intp *wheremask_strides)
{
int idim;
npy_intp shape_it[NPY_MAXDIMS];
npy_intp dst_strides_it[NPY_MAXDIMS];
npy_intp src_strides_it[NPY_MAXDIMS];
npy_intp wheremask_strides_it[NPY_MAXDIMS];
npy_intp coord[NPY_MAXDIMS];
PyArray_MaskedStridedUnaryOp *stransfer = NULL;
NpyAuxData *transferdata = NULL;
int aligned, needs_api = 0;
npy_intp src_itemsize = src_dtype->elsize;
NPY_BEGIN_THREADS_DEF;
aligned =
copycast_isaligned(ndim, shape, dst_dtype, dst_data, dst_strides) &&
copycast_isaligned(ndim, shape, src_dtype, src_data, src_strides);
/* Use raw iteration with no heap allocation */
if (PyArray_PrepareThreeRawArrayIter(
ndim, shape,
dst_data, dst_strides,
src_data, src_strides,
wheremask_data, wheremask_strides,
&ndim, shape_it,
&dst_data, dst_strides_it,
&src_data, src_strides_it,
&wheremask_data, wheremask_strides_it) < 0) {
return -1;
}
/*
* Overlap check for the 1D case. Higher dimensional arrays cause
* a temporary copy before getting here.
*/
if (ndim == 1 && src_data < dst_data &&
src_data + shape_it[0] * src_strides_it[0] > dst_data) {
src_data += (shape_it[0] - 1) * src_strides_it[0];
dst_data += (shape_it[0] - 1) * dst_strides_it[0];
wheremask_data += (shape_it[0] - 1) * wheremask_strides_it[0];
src_strides_it[0] = -src_strides_it[0];
dst_strides_it[0] = -dst_strides_it[0];
wheremask_strides_it[0] = -wheremask_strides_it[0];
}
/* Get the function to do the casting */
if (PyArray_GetMaskedDTypeTransferFunction(aligned,
src_strides_it[0],
dst_strides_it[0],
wheremask_strides_it[0],
src_dtype, dst_dtype, wheremask_dtype,
0,
&stransfer, &transferdata,
&needs_api) != NPY_SUCCEED) {
return -1;
}
if (!needs_api) {
NPY_BEGIN_THREADS;
}
NPY_RAW_ITER_START(idim, ndim, coord, shape_it) {
/* Process the innermost dimension */
stransfer(dst_data, dst_strides_it[0], src_data, src_strides_it[0],
(npy_bool *)wheremask_data, wheremask_strides_it[0],
shape_it[0], src_itemsize, transferdata);
} NPY_RAW_ITER_THREE_NEXT(idim, ndim, coord, shape_it,
dst_data, dst_strides_it,
src_data, src_strides_it,
wheremask_data, wheremask_strides_it);
NPY_END_THREADS;
NPY_AUXDATA_FREE(transferdata);
return (needs_api && PyErr_Occurred()) ? -1 : 0;
}
/*
* An array assignment function for copying arrays, broadcasting 'src' into
* 'dst'. This function makes a temporary copy of 'src' if 'src' and
* 'dst' overlap, to be able to handle views of the same data with
* different strides.
*
* dst: The destination array.
* src: The source array.
* wheremask: If non-NULL, a boolean mask specifying where to copy.
* casting: An exception is raised if the copy violates this
* casting rule.
*
* Returns 0 on success, -1 on failure.
*/
NPY_NO_EXPORT int
PyArray_AssignArray(PyArrayObject *dst, PyArrayObject *src,
PyArrayObject *wheremask,
NPY_CASTING casting)
{
int copied_src = 0;
npy_intp src_strides[NPY_MAXDIMS];
/* Use array_assign_scalar if 'src' NDIM is 0 */
if (PyArray_NDIM(src) == 0) {
return PyArray_AssignRawScalar(
dst, PyArray_DESCR(src), PyArray_DATA(src),
wheremask, casting);
}
/*
* Performance fix for expressions like "a[1000:6000] += x". In this
* case, first an in-place add is done, followed by an assignment,
* equivalently expressed like this:
*
* tmp = a[1000:6000] # Calls array_subscript in mapping.c
* np.add(tmp, x, tmp)
* a[1000:6000] = tmp # Calls array_assign_subscript in mapping.c
*
* In the assignment the underlying data type, shape, strides, and
* data pointers are identical, but src != dst because they are separately
* generated slices. By detecting this and skipping the redundant
* copy of values to themselves, we potentially give a big speed boost.
*
* Note that we don't call EquivTypes, because usually the exact same
* dtype object will appear, and we don't want to slow things down
* with a complicated comparison. The comparisons are ordered to
* try and reject this with as little work as possible.
*/
if (PyArray_DATA(src) == PyArray_DATA(dst) &&
PyArray_DESCR(src) == PyArray_DESCR(dst) &&
PyArray_NDIM(src) == PyArray_NDIM(dst) &&
PyArray_CompareLists(PyArray_DIMS(src),
PyArray_DIMS(dst),
PyArray_NDIM(src)) &&
PyArray_CompareLists(PyArray_STRIDES(src),
PyArray_STRIDES(dst),
PyArray_NDIM(src))) {
/*printf("Redundant copy operation detected\n");*/
return 0;
}
if (PyArray_FailUnlessWriteable(dst, "assignment destination") < 0) {
goto fail;
}
/* Check the casting rule */
if (!PyArray_CanCastTypeTo(PyArray_DESCR(src),
PyArray_DESCR(dst), casting)) {
PyObject *errmsg;
errmsg = PyUString_FromString("Cannot cast scalar from ");
PyUString_ConcatAndDel(&errmsg,
PyObject_Repr((PyObject *)PyArray_DESCR(src)));
PyUString_ConcatAndDel(&errmsg,
PyUString_FromString(" to "));
PyUString_ConcatAndDel(&errmsg,
PyObject_Repr((PyObject *)PyArray_DESCR(dst)));
PyUString_ConcatAndDel(&errmsg,
PyUString_FromFormat(" according to the rule %s",
npy_casting_to_string(casting)));
PyErr_SetObject(PyExc_TypeError, errmsg);
Py_DECREF(errmsg);
goto fail;
}
/*
* When ndim is 1 and the strides point in the same direction,
* the lower-level inner loop handles copying
* of overlapping data. For bigger ndim and opposite-strided 1D
* data, we make a temporary copy of 'src' if 'src' and 'dst' overlap.'
*/
if (((PyArray_NDIM(dst) == 1 && PyArray_NDIM(src) >= 1 &&
PyArray_STRIDES(dst)[0] *
PyArray_STRIDES(src)[PyArray_NDIM(src) - 1] < 0) ||
PyArray_NDIM(dst) > 1 || PyArray_HASFIELDS(dst)) &&
arrays_overlap(src, dst)) {
PyArrayObject *tmp;
/*
* Allocate a temporary copy array.
*/
tmp = (PyArrayObject *)PyArray_NewLikeArray(dst,
NPY_KEEPORDER, NULL, 0);
if (tmp == NULL) {
goto fail;
}
if (PyArray_AssignArray(tmp, src, NULL, NPY_UNSAFE_CASTING) < 0) {
Py_DECREF(tmp);
goto fail;
}
src = tmp;
copied_src = 1;
}
/* Broadcast 'src' to 'dst' for raw iteration */
if (PyArray_NDIM(src) > PyArray_NDIM(dst)) {
int ndim_tmp = PyArray_NDIM(src);
npy_intp *src_shape_tmp = PyArray_DIMS(src);
npy_intp *src_strides_tmp = PyArray_STRIDES(src);
/*
* As a special case for backwards compatibility, strip
* away unit dimensions from the left of 'src'
*/
while (ndim_tmp > PyArray_NDIM(dst) && src_shape_tmp[0] == 1) {
--ndim_tmp;
++src_shape_tmp;
++src_strides_tmp;
}
if (broadcast_strides(PyArray_NDIM(dst), PyArray_DIMS(dst),
ndim_tmp, src_shape_tmp,
src_strides_tmp, "input array",
src_strides) < 0) {
goto fail;
}
}
else {
if (broadcast_strides(PyArray_NDIM(dst), PyArray_DIMS(dst),
PyArray_NDIM(src), PyArray_DIMS(src),
PyArray_STRIDES(src), "input array",
src_strides) < 0) {
goto fail;
}
}
/* optimization: scalar boolean mask */
if (wheremask != NULL &&
PyArray_NDIM(wheremask) == 0 &&
PyArray_DESCR(wheremask)->type_num == NPY_BOOL) {
npy_bool value = *(npy_bool *)PyArray_DATA(wheremask);
if (value) {
/* where=True is the same as no where at all */
wheremask = NULL;
}
else {
/* where=False copies nothing */
return 0;
}
}
if (wheremask == NULL) {
/* A straightforward value assignment */
/* Do the assignment with raw array iteration */
if (raw_array_assign_array(PyArray_NDIM(dst), PyArray_DIMS(dst),
PyArray_DESCR(dst), PyArray_DATA(dst), PyArray_STRIDES(dst),
PyArray_DESCR(src), PyArray_DATA(src), src_strides) < 0) {
goto fail;
}
}
else {
npy_intp wheremask_strides[NPY_MAXDIMS];
/* Broadcast the wheremask to 'dst' for raw iteration */
if (broadcast_strides(PyArray_NDIM(dst), PyArray_DIMS(dst),
PyArray_NDIM(wheremask), PyArray_DIMS(wheremask),
PyArray_STRIDES(wheremask), "where mask",
wheremask_strides) < 0) {
goto fail;
}
/* A straightforward where-masked assignment */
/* Do the masked assignment with raw array iteration */
if (raw_array_wheremasked_assign_array(
PyArray_NDIM(dst), PyArray_DIMS(dst),
PyArray_DESCR(dst), PyArray_DATA(dst), PyArray_STRIDES(dst),
PyArray_DESCR(src), PyArray_DATA(src), src_strides,
PyArray_DESCR(wheremask), PyArray_DATA(wheremask),
wheremask_strides) < 0) {
goto fail;
}
}
if (copied_src) {
Py_DECREF(src);
}
return 0;
fail:
if (copied_src) {
Py_DECREF(src);
}
return -1;
}
|