This file is indexed.

/usr/lib/python2.7/dist-packages/pyopencl/array.py is in python-pyopencl 2015.1-2build3.

This file is owned by root:root, with mode 0o644.

The actual contents of the file can be viewed below.

   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
 444
 445
 446
 447
 448
 449
 450
 451
 452
 453
 454
 455
 456
 457
 458
 459
 460
 461
 462
 463
 464
 465
 466
 467
 468
 469
 470
 471
 472
 473
 474
 475
 476
 477
 478
 479
 480
 481
 482
 483
 484
 485
 486
 487
 488
 489
 490
 491
 492
 493
 494
 495
 496
 497
 498
 499
 500
 501
 502
 503
 504
 505
 506
 507
 508
 509
 510
 511
 512
 513
 514
 515
 516
 517
 518
 519
 520
 521
 522
 523
 524
 525
 526
 527
 528
 529
 530
 531
 532
 533
 534
 535
 536
 537
 538
 539
 540
 541
 542
 543
 544
 545
 546
 547
 548
 549
 550
 551
 552
 553
 554
 555
 556
 557
 558
 559
 560
 561
 562
 563
 564
 565
 566
 567
 568
 569
 570
 571
 572
 573
 574
 575
 576
 577
 578
 579
 580
 581
 582
 583
 584
 585
 586
 587
 588
 589
 590
 591
 592
 593
 594
 595
 596
 597
 598
 599
 600
 601
 602
 603
 604
 605
 606
 607
 608
 609
 610
 611
 612
 613
 614
 615
 616
 617
 618
 619
 620
 621
 622
 623
 624
 625
 626
 627
 628
 629
 630
 631
 632
 633
 634
 635
 636
 637
 638
 639
 640
 641
 642
 643
 644
 645
 646
 647
 648
 649
 650
 651
 652
 653
 654
 655
 656
 657
 658
 659
 660
 661
 662
 663
 664
 665
 666
 667
 668
 669
 670
 671
 672
 673
 674
 675
 676
 677
 678
 679
 680
 681
 682
 683
 684
 685
 686
 687
 688
 689
 690
 691
 692
 693
 694
 695
 696
 697
 698
 699
 700
 701
 702
 703
 704
 705
 706
 707
 708
 709
 710
 711
 712
 713
 714
 715
 716
 717
 718
 719
 720
 721
 722
 723
 724
 725
 726
 727
 728
 729
 730
 731
 732
 733
 734
 735
 736
 737
 738
 739
 740
 741
 742
 743
 744
 745
 746
 747
 748
 749
 750
 751
 752
 753
 754
 755
 756
 757
 758
 759
 760
 761
 762
 763
 764
 765
 766
 767
 768
 769
 770
 771
 772
 773
 774
 775
 776
 777
 778
 779
 780
 781
 782
 783
 784
 785
 786
 787
 788
 789
 790
 791
 792
 793
 794
 795
 796
 797
 798
 799
 800
 801
 802
 803
 804
 805
 806
 807
 808
 809
 810
 811
 812
 813
 814
 815
 816
 817
 818
 819
 820
 821
 822
 823
 824
 825
 826
 827
 828
 829
 830
 831
 832
 833
 834
 835
 836
 837
 838
 839
 840
 841
 842
 843
 844
 845
 846
 847
 848
 849
 850
 851
 852
 853
 854
 855
 856
 857
 858
 859
 860
 861
 862
 863
 864
 865
 866
 867
 868
 869
 870
 871
 872
 873
 874
 875
 876
 877
 878
 879
 880
 881
 882
 883
 884
 885
 886
 887
 888
 889
 890
 891
 892
 893
 894
 895
 896
 897
 898
 899
 900
 901
 902
 903
 904
 905
 906
 907
 908
 909
 910
 911
 912
 913
 914
 915
 916
 917
 918
 919
 920
 921
 922
 923
 924
 925
 926
 927
 928
 929
 930
 931
 932
 933
 934
 935
 936
 937
 938
 939
 940
 941
 942
 943
 944
 945
 946
 947
 948
 949
 950
 951
 952
 953
 954
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
"""CL device arrays."""

from __future__ import division

__copyright__ = "Copyright (C) 2009 Andreas Kloeckner"

__license__ = """
Permission is hereby granted, free of charge, to any person
obtaining a copy of this software and associated documentation
files (the "Software"), to deal in the Software without
restriction, including without limitation the rights to use,
copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the
Software is furnished to do so, subject to the following
conditions:

The above copyright notice and this permission notice shall be
included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
OTHER DEALINGS IN THE SOFTWARE.
"""


import numpy as np
import pyopencl.elementwise as elementwise
import pyopencl as cl
from pytools import memoize_method
from pyopencl.compyte.array import (
        as_strided as _as_strided,
        f_contiguous_strides as _f_contiguous_strides,
        c_contiguous_strides as _c_contiguous_strides,
        ArrayFlags as _ArrayFlags,
        get_common_dtype as _get_common_dtype_base)
from pyopencl.characterize import has_double_support


def _get_common_dtype(obj1, obj2, queue):
    return _get_common_dtype_base(obj1, obj2,
            has_double_support(queue.device))

# Work around PyPy not currently supporting the object dtype.
# (Yes, it doesn't even support checking!)
# (as of May 27, 2014 on PyPy 2.3)
try:
    np.dtype(object)

    def _dtype_is_object(t):
        return t == object
except:
    def _dtype_is_object(t):
        return False


# {{{ vector types

class vec:
    pass


def _create_vector_types():
    field_names = ["x", "y", "z", "w"]

    from pyopencl.tools import get_or_register_dtype

    vec.types = {}
    vec.type_to_scalar_and_count = {}

    counts = [2, 3, 4, 8, 16]

    for base_name, base_type in [
            ('char', np.int8),
            ('uchar', np.uint8),
            ('short', np.int16),
            ('ushort', np.uint16),
            ('int', np.int32),
            ('uint', np.uint32),
            ('long', np.int64),
            ('ulong', np.uint64),
            ('float', np.float32),
            ('double', np.float64),
            ]:
        for count in counts:
            name = "%s%d" % (base_name, count)

            titles = field_names[:count]

            padded_count = count
            if count == 3:
                padded_count = 4

            names = ["s%d" % i for i in range(count)]
            while len(names) < padded_count:
                names.append("padding%d" % (len(names)-count))

            if len(titles) < len(names):
                titles.extend((len(names)-len(titles))*[None])

            try:
                dtype = np.dtype(dict(
                    names=names,
                    formats=[base_type]*padded_count,
                    titles=titles))
            except NotImplementedError:
                try:
                    dtype = np.dtype([((n, title), base_type)
                                      for (n, title) in zip(names, titles)])
                except TypeError:
                    dtype = np.dtype([(n, base_type) for (n, title)
                                      in zip(names, titles)])

            get_or_register_dtype(name, dtype)

            setattr(vec, name, dtype)

            def create_array(dtype, count, padded_count, *args, **kwargs):
                if len(args) < count:
                    from warnings import warn
                    warn("default values for make_xxx are deprecated;"
                            " instead specify all parameters or use"
                            " array.vec.zeros_xxx", DeprecationWarning)
                padded_args = tuple(list(args)+[0]*(padded_count-len(args)))
                array = eval("array(padded_args, dtype=dtype)",
                        dict(array=np.array, padded_args=padded_args,
                        dtype=dtype))
                for key, val in kwargs.items():
                    array[key] = val
                return array

            setattr(vec, "make_"+name, staticmethod(eval(
                    "lambda *args, **kwargs: create_array(dtype, %i, %i, "
                    "*args, **kwargs)" % (count, padded_count),
                    dict(create_array=create_array, dtype=dtype))))
            setattr(vec, "filled_"+name, staticmethod(eval(
                    "lambda val: vec.make_%s(*[val]*%i)" % (name, count))))
            setattr(vec, "zeros_"+name,
                    staticmethod(eval("lambda: vec.filled_%s(0)" % (name))))
            setattr(vec, "ones_"+name,
                    staticmethod(eval("lambda: vec.filled_%s(1)" % (name))))

            vec.types[np.dtype(base_type), count] = dtype
            vec.type_to_scalar_and_count[dtype] = np.dtype(base_type), count

_create_vector_types()

# }}}


# {{{ helper functionality

def splay(queue, n, kernel_specific_max_wg_size=None):
    dev = queue.device
    max_work_items = _builtin_min(128, dev.max_work_group_size)

    if kernel_specific_max_wg_size is not None:
        from __builtin__ import min
        max_work_items = min(max_work_items, kernel_specific_max_wg_size)

    min_work_items = _builtin_min(32, max_work_items)
    max_groups = dev.max_compute_units * 4 * 8
    # 4 to overfill the device
    # 8 is an Nvidia constant--that's how many
    # groups fit onto one compute device

    if n < min_work_items:
        group_count = 1
        work_items_per_group = min_work_items
    elif n < (max_groups * min_work_items):
        group_count = (n + min_work_items - 1) // min_work_items
        work_items_per_group = min_work_items
    elif n < (max_groups * max_work_items):
        group_count = max_groups
        grp = (n + min_work_items - 1) // min_work_items
        work_items_per_group = (
                (grp + max_groups - 1) // max_groups) * min_work_items
    else:
        group_count = max_groups
        work_items_per_group = max_work_items

    #print "n:%d gc:%d wipg:%d" % (n, group_count, work_items_per_group)
    return (group_count*work_items_per_group,), (work_items_per_group,)


def elwise_kernel_runner(kernel_getter):
    """Take a kernel getter of the same signature as the kernel
    and return a function that invokes that kernel.

    Assumes that the zeroth entry in *args* is an :class:`Array`.
    """

    def kernel_runner(*args, **kwargs):
        repr_ary = args[0]
        queue = kwargs.pop("queue", None) or repr_ary.queue
        wait_for = kwargs.pop("wait_for", None)

        # wait_for must be a copy, because we modify it in-place below
        if wait_for is None:
            wait_for = []
        else:
            wait_for = list(wait_for)

        knl = kernel_getter(*args, **kwargs)

        gs, ls = repr_ary.get_sizes(queue,
                knl.get_work_group_info(
                    cl.kernel_work_group_info.WORK_GROUP_SIZE,
                    queue.device))

        assert isinstance(repr_ary, Array)

        actual_args = []
        for arg in args:
            if isinstance(arg, Array):
                if not arg.flags.forc:
                    raise RuntimeError("only contiguous arrays may "
                            "be used as arguments to this operation")
                actual_args.append(arg.base_data)
                actual_args.append(arg.offset)
                wait_for.extend(arg.events)
            else:
                actual_args.append(arg)
        actual_args.append(repr_ary.size)

        return knl(queue, gs, ls, *actual_args, **dict(wait_for=wait_for))

    try:
        from functools import update_wrapper
    except ImportError:
        return kernel_runner
    else:
        return update_wrapper(kernel_runner, kernel_getter)


class DefaultAllocator(cl.tools.DeferredAllocator):
    def __init__(self, *args, **kwargs):
        from warnings import warn
        warn("pyopencl.array.DefaultAllocator is deprecated. "
                "It will be continue to exist throughout the 2013.x "
                "versions of PyOpenCL.",
                DeprecationWarning, 2)
        cl.tools.DeferredAllocator.__init__(self, *args, **kwargs)


def _make_strides(itemsize, shape, order):
    if order in "fF":
        return _f_contiguous_strides(itemsize, shape)
    elif order in "cC":
        return _c_contiguous_strides(itemsize, shape)
    else:
        raise ValueError("invalid order: %s" % order)

# }}}


# {{{ array class

class ArrayHasOffsetError(ValueError):
    """
    .. versionadded:: 2013.1
    """

    def __init__(self, val="The operation you are attempting does not yet "
                "support arrays that start at an offset from the beginning "
                "of their buffer."):
        ValueError.__init__(self, val)


class _copy_queue:
    pass


class Array(object):
    """A :class:`numpy.ndarray` work-alike that stores its data and performs
    its computations on the compute device.  *shape* and *dtype* work exactly
    as in :mod:`numpy`.  Arithmetic methods in :class:`Array` support the
    broadcasting of scalars. (e.g. `array+5`)

    *cqa* must be a :class:`pyopencl.CommandQueue` or a :class:`pyopencl.Context`.

    If it is a queue, *cqa* specifies the queue in which the array carries out
    its computations by default. If a default queue (and thereby overloaded
    operators and many other niceties) are not desired, pass a
    :class:`Context`.

    *cqa* will at some point be renamed *cq*, so it should be considered
    'positional-only'. Arguments starting from 'order' should be considered
    keyword-only.

    *allocator* may be `None` or a callable that, upon being called with an
    argument of the number of bytes to be allocated, returns an
    :class:`pyopencl.Buffer` object. (A :class:`pyopencl.tools.MemoryPool`
    instance is one useful example of an object to pass here.)

    .. versionchanged:: 2011.1
        Renamed *context* to *cqa*, made it general-purpose.

        All arguments beyond *order* should be considered keyword-only.

    .. attribute :: data

        The :class:`pyopencl.MemoryObject` instance created for the memory that
        backs this :class:`Array`.

        .. versionchanged:: 2013.1

            If a non-zero :attr:`offset` has been specified for this array,
            this will fail with :exc:`ArrayHasOffsetError`.

    .. attribute :: base_data

        The :class:`pyopencl.MemoryObject` instance created for the memory that
        backs this :class:`Array`. Unlike :attr:`data`, the base address of
        *base_data* is allowed to be different from the beginning of the array.
        The actual beginning is the base address of *base_data* plus
        :attr:`offset` in units of :attr:`dtype`.

        Unlike :attr:`data`, retrieving :attr:`base_data` always succeeds.

        .. versionadded:: 2013.1

    .. attribute :: offset

        See :attr:`base_data`.

        .. versionadded:: 2013.1

    .. attribute :: shape

        The tuple of lengths of each dimension in the array.

    .. attribute :: dtype

        The :class:`numpy.dtype` of the items in the GPU array.

    .. attribute :: size

        The number of meaningful entries in the array. Can also be computed by
        multiplying up the numbers in :attr:`shape`.

    .. attribute :: nbytes

        The size of the entire array in bytes. Computed as :attr:`size` times
        ``dtype.itemsize``.

    .. attribute :: strides

        Tuple of bytes to step in each dimension when traversing an array.

    .. attribute :: flags

        Return an object with attributes `c_contiguous`, `f_contiguous` and
        `forc`, which may be used to query contiguity properties in analogy to
        :attr:`numpy.ndarray.flags`.

    .. rubric:: Methods

    .. automethod :: with_queue

    .. automethod :: __len__
    .. automethod :: reshape
    .. automethod :: ravel
    .. automethod :: view
    .. automethod :: set
    .. automethod :: get
    .. automethod :: copy

    .. automethod :: __str__
    .. automethod :: __repr__

    .. automethod :: mul_add
    .. automethod :: __add__
    .. automethod :: __sub__
    .. automethod :: __iadd__
    .. automethod :: __isub__
    .. automethod :: __neg__
    .. automethod :: __mul__
    .. automethod :: __div__
    .. automethod :: __rdiv__
    .. automethod :: __pow__

    .. automethod :: __abs__

    .. UNDOC reverse()

    .. automethod :: fill

    .. automethod :: astype

    .. autoattribute :: real
    .. autoattribute :: imag
    .. automethod :: conj

    .. automethod :: __getitem__
    .. automethod :: __setitem__

    .. automethod :: setitem

    .. automethod :: map_to_host

    .. rubric:: Comparisons, conditionals, any, all

    .. versionadded:: 2013.2

    Boolean arrays are stored as :class:`numpy.int8` because ``bool``
    has an unspecified size in the OpenCL spec.

    .. automethod :: __nonzero__

        Only works for device scalars. (i.e. "arrays" with ``shape == ()``.)

    .. automethod :: any
    .. automethod :: all

    .. automethod :: __eq__
    .. automethod :: __ne__
    .. automethod :: __lt__
    .. automethod :: __le__
    .. automethod :: __gt__
    .. automethod :: __ge__

    .. rubric:: Event management

    If an array is used from within an out-of-order queue, it needs to take
    care of its own operation ordering. The facilities in this section make
    this possible.

    .. versionadded:: 2014.1.1

    .. attribute:: events

        A list of :class:`pyopencl.Event` instances that the current content of
        this array depends on. User code may read, but should never modify this
        list directly. To update this list, instead use the following methods.

    .. automethod:: add_event
    .. automethod:: finish
    """

    __array_priority__ = 100

    def __init__(self, cqa, shape, dtype, order="C", allocator=None,
            data=None, offset=0, queue=None, strides=None, events=None):
        # {{{ backward compatibility

        from warnings import warn
        if queue is not None:
            warn("Passing the queue to the array through anything but the "
                    "first argument of the Array constructor is deprecated. "
                    "This will be continue to be accepted throughout the "
                    "2013.[0-6] versions of PyOpenCL.",
                    DeprecationWarning, 2)

        if isinstance(cqa, cl.CommandQueue):
            if queue is not None:
                raise TypeError("can't specify queue in 'cqa' and "
                        "'queue' arguments")
            queue = cqa

        elif isinstance(cqa, cl.Context):
            context = cqa

            if queue is not None:
                raise TypeError("may not pass a context and a queue "
                        "(just pass the queue)")
            if allocator is not None:
                # "is" would be wrong because two Python objects are allowed
                # to hold handles to the same context.

                # FIXME It would be nice to check this. But it would require
                # changing the allocator interface. Trust the user for now.

                #assert allocator.context == context
                pass

        else:
            # cqa is assumed to be an allocator
            warn("Passing an allocator for the 'cqa' parameter is deprecated. "
                    "This usage will be continue to be accepted throughout "
                    "the 2013.[0-6] versions of PyOpenCL.",
                    DeprecationWarning, stacklevel=2)
            if allocator is not None:
                raise TypeError("can't specify allocator in 'cqa' and "
                        "'allocator' arguments")

            allocator = cqa

        # Queue-less arrays do have a purpose in life.
        # They don't do very much, but at least they don't run kernels
        # in random queues.
        #
        # See also :meth:`with_queue`.

        # }}}

        # invariant here: allocator, queue set

        # {{{ determine shape and strides
        dtype = np.dtype(dtype)

        try:
            s = 1
            for dim in shape:
                s *= dim
        except TypeError:
            import sys
            if sys.version_info >= (3,):
                admissible_types = (int, np.integer)
            else:
                admissible_types = (int, long, np.integer)

            if not isinstance(shape, admissible_types):
                raise TypeError("shape must either be iterable or "
                        "castable to an integer")
            s = shape
            shape = (shape,)

        if isinstance(s, np.integer):
            # bombs if s is a Python integer
            s = np.asscalar(s)

        if strides is None:
            strides = _make_strides(dtype.itemsize, shape, order)

        else:
            # FIXME: We should possibly perform some plausibility
            # checking on 'strides' here.

            strides = tuple(strides)

        # }}}

        if _dtype_is_object(dtype):
            raise TypeError("object arrays on the compute device are not allowed")

        self.queue = queue
        self.shape = shape
        self.dtype = dtype
        self.strides = strides
        if events is None:
            self.events = []
        else:
            self.events = events

        self.size = s
        alloc_nbytes = self.nbytes = self.dtype.itemsize * self.size

        self.allocator = allocator

        if data is None:
            if not alloc_nbytes:
                # Work around CL not allowing zero-sized buffers.
                alloc_nbytes = 1

            if allocator is None:
                # FIXME remove me when queues become required
                if queue is not None:
                    context = queue.context

                self.base_data = cl.Buffer(
                        context, cl.mem_flags.READ_WRITE, alloc_nbytes)
            else:
                self.base_data = self.allocator(alloc_nbytes)
        else:
            self.base_data = data

        self.offset = offset

    @property
    def context(self):
        return self.base_data.context

    @property
    def data(self):
        if self.offset:
            raise ArrayHasOffsetError()
        else:
            return self.base_data

    @property
    @memoize_method
    def flags(self):
        return _ArrayFlags(self)

    def _new_with_changes(self, data, offset, shape=None, dtype=None,
            strides=None, queue=_copy_queue):
        """
        :arg data: *None* means allocate a new array.
        """
        if shape is None:
            shape = self.shape
        if dtype is None:
            dtype = self.dtype
        if strides is None:
            strides = self.strides
        if queue is _copy_queue:
            queue = self.queue

        # If we're allocating new data, then there's not likely to be
        # a data dependency. Otherwise, the two arrays should probably
        # share the same events list.

        if data is None:
            events = None
        else:
            events = self.events

        if queue is not None:
            return Array(queue, shape, dtype, allocator=self.allocator,
                    strides=strides, data=data, offset=offset,
                    events=events)
        else:
            return Array(self.context, shape, dtype, queue=queue,
                    strides=strides, data=data, offset=offset,
                    events=events, allocator=self.allocator)

    def with_queue(self, queue):
        """Return a copy of *self* with the default queue set to *queue*.

        *None* is allowed as a value for *queue*.

        .. versionadded:: 2013.1
        """

        if queue is not None:
            assert queue.context == self.context

        return self._new_with_changes(self.base_data, self.offset,
                queue=queue)

    #@memoize_method FIXME: reenable
    def get_sizes(self, queue, kernel_specific_max_wg_size=None):
        if not self.flags.forc:
            raise NotImplementedError("cannot operate on non-contiguous array")
        return splay(queue, self.size,
                kernel_specific_max_wg_size=kernel_specific_max_wg_size)

    def set(self, ary, queue=None, async=False):
        """Transfer the contents the :class:`numpy.ndarray` object *ary*
        onto the device.

        *ary* must have the same dtype and size (not necessarily shape) as
        *self*.
        """

        assert ary.size == self.size
        assert ary.dtype == self.dtype

        if not ary.flags.forc:
            raise RuntimeError("cannot set from non-contiguous array")

            ary = ary.copy()

        if ary.strides != self.strides:
            from warnings import warn
            warn("Setting array from one with different "
                    "strides/storage order. This will cease to work "
                    "in 2013.x.",
                    stacklevel=2)

        if self.size:
            cl.enqueue_copy(queue or self.queue, self.base_data, ary,
                    device_offset=self.offset,
                    is_blocking=not async)

    def get(self, queue=None, ary=None, async=False):
        """Transfer the contents of *self* into *ary* or a newly allocated
        :mod:`numpy.ndarray`. If *ary* is given, it must have the right
        size (not necessarily shape) and dtype.
        """

        if ary is None:
            ary = np.empty(self.shape, self.dtype)

            ary = _as_strided(ary, strides=self.strides)
        else:
            if ary.size != self.size:
                raise TypeError("'ary' has non-matching size")
            if ary.dtype != self.dtype:
                raise TypeError("'ary' has non-matching type")

        assert self.flags.forc, "Array in get() must be contiguous"

        if self.size:
            cl.enqueue_copy(queue or self.queue, ary, self.base_data,
                    device_offset=self.offset,
                    is_blocking=not async)

        return ary

    def copy(self, queue=None):
        """.. versionadded:: 2013.1"""

        queue = queue or self.queue
        result = self._new_like_me()

        if self.nbytes:
            cl.enqueue_copy(queue, result.base_data, self.base_data,
                    src_offset=self.offset, byte_count=self.nbytes)

        return result

    def __str__(self):
        return str(self.get())

    def __repr__(self):
        return repr(self.get())

    def __hash__(self):
        raise TypeError("pyopencl arrays are not hashable.")

    # {{{ kernel invocation wrappers

    @staticmethod
    @elwise_kernel_runner
    def _axpbyz(out, afac, a, bfac, b, queue=None):
        """Compute ``out = selffac * self + otherfac*other``,
        where *other* is an array."""
        assert out.shape == a.shape
        assert out.shape == b.shape

        return elementwise.get_axpbyz_kernel(
                out.context, a.dtype, b.dtype, out.dtype)

    @staticmethod
    @elwise_kernel_runner
    def _axpbz(out, a, x, b, queue=None):
        """Compute ``z = a * x + b``, where *b* is a scalar."""
        a = np.array(a)
        b = np.array(b)
        assert out.shape == x.shape
        return elementwise.get_axpbz_kernel(out.context,
                a.dtype, x.dtype, b.dtype, out.dtype)

    @staticmethod
    @elwise_kernel_runner
    def _elwise_multiply(out, a, b, queue=None):
        assert out.shape == a.shape
        assert out.shape == b.shape
        return elementwise.get_multiply_kernel(
                a.context, a.dtype, b.dtype, out.dtype)

    @staticmethod
    @elwise_kernel_runner
    def _rdiv_scalar(out, ary, other, queue=None):
        other = np.array(other)
        assert out.shape == ary.shape
        return elementwise.get_rdivide_elwise_kernel(
                out.context, ary.dtype, other.dtype, out.dtype)

    @staticmethod
    @elwise_kernel_runner
    def _div(out, self, other, queue=None):
        """Divides an array by another array."""

        assert self.shape == other.shape

        return elementwise.get_divide_kernel(self.context,
                self.dtype, other.dtype, out.dtype)

    @staticmethod
    @elwise_kernel_runner
    def _fill(result, scalar):
        return elementwise.get_fill_kernel(result.context, result.dtype)

    @staticmethod
    @elwise_kernel_runner
    def _abs(result, arg):
        if arg.dtype.kind == "c":
            from pyopencl.elementwise import complex_dtype_to_name
            fname = "%s_abs" % complex_dtype_to_name(arg.dtype)
        elif arg.dtype.kind == "f":
            fname = "fabs"
        elif arg.dtype.kind in ["u", "i"]:
            fname = "abs"
        else:
            raise TypeError("unsupported dtype in _abs()")

        return elementwise.get_unary_func_kernel(
                arg.context, fname, arg.dtype, out_dtype=result.dtype)

    @staticmethod
    @elwise_kernel_runner
    def _real(result, arg):
        from pyopencl.elementwise import complex_dtype_to_name
        fname = "%s_real" % complex_dtype_to_name(arg.dtype)
        return elementwise.get_unary_func_kernel(
                arg.context, fname, arg.dtype, out_dtype=result.dtype)

    @staticmethod
    @elwise_kernel_runner
    def _imag(result, arg):
        from pyopencl.elementwise import complex_dtype_to_name
        fname = "%s_imag" % complex_dtype_to_name(arg.dtype)
        return elementwise.get_unary_func_kernel(
                arg.context, fname, arg.dtype, out_dtype=result.dtype)

    @staticmethod
    @elwise_kernel_runner
    def _conj(result, arg):
        from pyopencl.elementwise import complex_dtype_to_name
        fname = "%s_conj" % complex_dtype_to_name(arg.dtype)
        return elementwise.get_unary_func_kernel(
                arg.context, fname, arg.dtype, out_dtype=result.dtype)

    @staticmethod
    @elwise_kernel_runner
    def _pow_scalar(result, ary, exponent):
        exponent = np.array(exponent)
        return elementwise.get_pow_kernel(result.context,
                ary.dtype, exponent.dtype, result.dtype,
                is_base_array=True, is_exp_array=False)

    @staticmethod
    @elwise_kernel_runner
    def _rpow_scalar(result, base, exponent):
        base = np.array(base)
        return elementwise.get_pow_kernel(result.context,
                base.dtype, exponent.dtype, result.dtype,
                is_base_array=False, is_exp_array=True)

    @staticmethod
    @elwise_kernel_runner
    def _pow_array(result, base, exponent):
        return elementwise.get_pow_kernel(
                result.context, base.dtype, exponent.dtype, result.dtype,
                is_base_array=True, is_exp_array=True)

    @staticmethod
    @elwise_kernel_runner
    def _reverse(result, ary):
        return elementwise.get_reverse_kernel(result.context, ary.dtype)

    @staticmethod
    @elwise_kernel_runner
    def _copy(dest, src):
        return elementwise.get_copy_kernel(
                dest.context, dest.dtype, src.dtype)

    def _new_like_me(self, dtype=None, queue=None):
        strides = None
        if dtype is None:
            dtype = self.dtype

        if dtype == self.dtype:
            strides = self.strides

        queue = queue or self.queue
        if queue is not None:
            return self.__class__(queue, self.shape, dtype,
                    allocator=self.allocator, strides=strides)
        elif self.allocator is not None:
            return self.__class__(self.allocator, self.shape, dtype,
                    strides=strides)
        else:
            return self.__class__(self.context, self.shape, dtype,
                    strides=strides)

    # }}}

    # {{{ operators

    def mul_add(self, selffac, other, otherfac, queue=None):
        """Return `selffac * self + otherfac*other`.
        """
        result = self._new_like_me(
                _get_common_dtype(self, other, queue or self.queue))
        result.add_event(
                self._axpbyz(result, selffac, self, otherfac, other))
        return result

    def __add__(self, other):
        """Add an array with an array or an array with a scalar."""

        if isinstance(other, Array):
            # add another vector
            result = self._new_like_me(
                    _get_common_dtype(self, other, self.queue))

            result.add_event(
                    self._axpbyz(result,
                        self.dtype.type(1), self,
                        other.dtype.type(1), other))

            return result
        else:
            # add a scalar
            if other == 0:
                return self.copy()
            else:
                common_dtype = _get_common_dtype(self, other, self.queue)
                result = self._new_like_me(common_dtype)
                result.add_event(
                        self._axpbz(result, self.dtype.type(1),
                            self, common_dtype.type(other)))
                return result

    __radd__ = __add__

    def __sub__(self, other):
        """Substract an array from an array or a scalar from an array."""

        if isinstance(other, Array):
            result = self._new_like_me(
                    _get_common_dtype(self, other, self.queue))
            result.add_event(
                    self._axpbyz(result,
                        self.dtype.type(1), self,
                        other.dtype.type(-1), other))

            return result
        else:
            # subtract a scalar
            if other == 0:
                return self.copy()
            else:
                result = self._new_like_me(
                        _get_common_dtype(self, other, self.queue))
                result.add_event(
                        self._axpbz(result, self.dtype.type(1), self, -other))
                return result

    def __rsub__(self, other):
        """Substracts an array by a scalar or an array::

           x = n - self
        """
        common_dtype = _get_common_dtype(self, other, self.queue)
        # other must be a scalar
        result = self._new_like_me(common_dtype)
        result.add_event(
                self._axpbz(result, self.dtype.type(-1), self,
                    common_dtype.type(other)))
        return result

    def __iadd__(self, other):
        if isinstance(other, Array):
            self.add_event(
                    self._axpbyz(self,
                        self.dtype.type(1), self,
                        other.dtype.type(1), other))
            return self
        else:
            self.add_event(
                    self._axpbz(self, self.dtype.type(1), self, other))
            return self

    def __isub__(self, other):
        if isinstance(other, Array):
            self.add_event(
                    self._axpbyz(self, self.dtype.type(1), self,
                        other.dtype.type(-1), other))
            return self
        else:
            self._axpbz(self, self.dtype.type(1), self, -other)
            return self

    def __neg__(self):
        result = self._new_like_me()
        result.add_event(self._axpbz(result, -1, self, 0))
        return result

    def __mul__(self, other):
        if isinstance(other, Array):
            result = self._new_like_me(
                    _get_common_dtype(self, other, self.queue))
            result.add_event(
                    self._elwise_multiply(result, self, other))
            return result
        else:
            common_dtype = _get_common_dtype(self, other, self.queue)
            result = self._new_like_me(common_dtype)
            result.add_event(
                    self._axpbz(result,
                        common_dtype.type(other), self, self.dtype.type(0)))
            return result

    def __rmul__(self, scalar):
        common_dtype = _get_common_dtype(self, scalar, self.queue)
        result = self._new_like_me(common_dtype)
        result.add_event(
                self._axpbz(result,
                    common_dtype.type(scalar), self, self.dtype.type(0)))
        return result

    def __imul__(self, other):
        if isinstance(other, Array):
            self.add_event(
                    self._elwise_multiply(self, self, other))
        else:
            # scalar
            self.add_event(
                    self._axpbz(self, other, self, self.dtype.type(0)))

        return self

    def __div__(self, other):
        """Divides an array by an array or a scalar, i.e. ``self / other``.
        """
        if isinstance(other, Array):
            result = self._new_like_me(
                    _get_common_dtype(self, other, self.queue))
            result.add_event(self._div(result, self, other))
        else:
            if other == 1:
                return self.copy()
            else:
                # create a new array for the result
                common_dtype = _get_common_dtype(self, other, self.queue)
                result = self._new_like_me(common_dtype)
                result.add_event(
                        self._axpbz(result,
                            common_dtype.type(1/other), self, self.dtype.type(0)))

        return result

    __truediv__ = __div__

    def __rdiv__(self, other):
        """Divides an array by a scalar or an array, i.e. ``other / self``.
        """

        if isinstance(other, Array):
            result = self._new_like_me(
                    _get_common_dtype(self, other, self.queue))
            result.add_event(other._div(result, self))
        else:
            # create a new array for the result
            common_dtype = _get_common_dtype(self, other, self.queue)
            result = self._new_like_me(common_dtype)
            result.add_event(
                    self._rdiv_scalar(result, self, common_dtype.type(other)))

        return result

    __rtruediv__ = __rdiv__

    def fill(self, value, queue=None, wait_for=None):
        """Fill the array with *scalar*.

        :returns: *self*.
        """
        self.add_event(
                self._fill(self, value, queue=queue, wait_for=wait_for))

        return self

    def __len__(self):
        """Returns the size of the leading dimension of *self*."""
        if len(self.shape):
            return self.shape[0]
        else:
            return TypeError("scalar has no len()")

    def __abs__(self):
        """Return a `Array` of the absolute values of the elements
        of *self*.
        """

        result = self._new_like_me(self.dtype.type(0).real.dtype)
        result.add_event(self._abs(result, self))
        return result

    def __pow__(self, other):
        """Exponentiation by a scalar or elementwise by another
        :class:`Array`.
        """

        if isinstance(other, Array):
            assert self.shape == other.shape

            result = self._new_like_me(
                    _get_common_dtype(self, other, self.queue))
            result.add_event(
                    self._pow_array(result, self, other))
        else:
            result = self._new_like_me(
                    _get_common_dtype(self, other, self.queue))
            result.add_event(self._pow_scalar(result, self, other))

        return result

    def __rpow__(self, other):
        # other must be a scalar
        common_dtype = _get_common_dtype(self, other, self.queue)
        result = self._new_like_me(common_dtype)
        result.add_event(
                self._rpow_scalar(result, common_dtype.type(other), self))
        return result

    # }}}

    def reverse(self, queue=None):
        """Return this array in reversed order. The array is treated
        as one-dimensional.
        """

        result = self._new_like_me()
        result.add_event(
                self._reverse(result, self))
        return result

    def astype(self, dtype, queue=None):
        """Return a copy of *self*, cast to *dtype*."""
        if dtype == self.dtype:
            return self.copy()

        result = self._new_like_me(dtype=dtype)
        result.add_event(self._copy(result, self, queue=queue))
        return result

    # {{{ rich comparisons, any, all

    def __nonzero__(self):
        if self.shape == ():
            return bool(self.get())
        else:
            raise ValueError("The truth value of an array with "
                    "more than one element is ambiguous. Use a.any() or a.all()")

    def any(self, queue=None, wait_for=None):
        from pyopencl.reduction import get_any_kernel
        krnl = get_any_kernel(self.context, self.dtype)
        return krnl(self, queue=queue, wait_for=wait_for)

    def all(self, queue=None, wait_for=None):
        from pyopencl.reduction import get_all_kernel
        krnl = get_all_kernel(self.context, self.dtype)
        return krnl(self, queue=queue, wait_for=wait_for)

    @staticmethod
    @elwise_kernel_runner
    def _scalar_comparison(out, a, b, queue=None, op=None):
        return elementwise.get_array_scalar_comparison_kernel(
                out.context, op, a.dtype)

    @staticmethod
    @elwise_kernel_runner
    def _array_comparison(out, a, b, queue=None, op=None):
        if a.shape != b.shape:
            raise ValueError("shapes of comparison arguments do not match")
        return elementwise.get_array_comparison_kernel(
                out.context, op, a.dtype, b.dtype)

    def __eq__(self, other):
        if isinstance(other, Array):
            result = self._new_like_me(np.int8)
            result.add_event(
                    self._array_comparison(result, self, other, op="=="))
            return result
        else:
            result = self._new_like_me(np.int8)
            result.add_event(
                    self._scalar_comparison(result, self, other, op="=="))
            return result

    def __ne__(self, other):
        if isinstance(other, Array):
            result = self._new_like_me(np.int8)
            result.add_event(
                    self._array_comparison(result, self, other, op="!="))
            return result
        else:
            result = self._new_like_me(np.int8)
            result.add_event(
                    self._scalar_comparison(result, self, other, op="!="))
            return result

    def __le__(self, other):
        if isinstance(other, Array):
            result = self._new_like_me(np.int8)
            result.add_event(
                    self._array_comparison(result, self, other, op="<="))
            return result
        else:
            result = self._new_like_me(np.int8)
            self._scalar_comparison(result, self, other, op="<=")
            return result

    def __ge__(self, other):
        if isinstance(other, Array):
            result = self._new_like_me(np.int8)
            result.add_event(
                    self._array_comparison(result, self, other, op=">="))
            return result
        else:
            result = self._new_like_me(np.int8)
            result.add_event(
                    self._scalar_comparison(result, self, other, op=">="))
            return result

    def __lt__(self, other):
        if isinstance(other, Array):
            result = self._new_like_me(np.int8)
            result.add_event(
                    self._array_comparison(result, self, other, op="<"))
            return result
        else:
            result = self._new_like_me(np.int8)
            result.add_event(
                    self._scalar_comparison(result, self, other, op="<"))
            return result

    def __gt__(self, other):
        if isinstance(other, Array):
            result = self._new_like_me(np.int8)
            result.add_event(
                    self._array_comparison(result, self, other, op=">"))
            return result
        else:
            result = self._new_like_me(np.int8)
            result.add_event(
                    self._scalar_comparison(result, self, other, op=">"))
            return result

    # }}}

    # {{{ complex-valued business

    def real(self):
        if self.dtype.kind == "c":
            result = self._new_like_me(self.dtype.type(0).real.dtype)
            result.add_event(
                    self._real(result, self))
            return result
        else:
            return self
    real = property(real, doc=".. versionadded:: 2012.1")

    def imag(self):
        if self.dtype.kind == "c":
            result = self._new_like_me(self.dtype.type(0).real.dtype)
            result.add_event(
                    self._imag(result, self))
            return result
        else:
            return zeros_like(self)
    imag = property(imag, doc=".. versionadded:: 2012.1")

    def conj(self):
        """.. versionadded:: 2012.1"""
        if self.dtype.kind == "c":
            result = self._new_like_me()
            result.add_event(self._conj(result, self))
            return result
        else:
            return self

    # }}}

    # {{{ event management

    def add_event(self, evt):
        """Add *evt* to :attr:`events`. If :attr:`events` is too long, this method
        may implicitly wait for a subset of :attr:`events` and clear them from the
        list.
        """
        n_wait = 4

        self.events.append(evt)

        if len(self.events) > 3*n_wait:
            wait_events = self.events[:n_wait]
            cl.wait_for_events(wait_events)
            del self.events[:n_wait]

    def finish(self):
        """Wait for the entire contents of :attr:`events`, clear it."""

        if self.events:
            cl.wait_for_events(self.events)
            del self.events[:]

    # }}}

    # {{{ views

    def reshape(self, *shape, **kwargs):
        """Returns an array containing the same data with a new shape."""

        order = kwargs.pop("order", "C")
        if kwargs:
            raise TypeError("unexpected keyword arguments: %s"
                    % kwargs.keys())

        if order not in "CF":
            raise ValueError("order must be either 'C' or 'F'")

        # TODO: add more error-checking, perhaps

        if isinstance(shape[0], tuple) or isinstance(shape[0], list):
            shape = tuple(shape[0])

        if -1 in shape:
            shape = list(shape)
            idx = shape.index(-1)
            size = -reduce(lambda x, y: x * y, shape, 1)
            shape[idx] = self.size // size
            if any(s < 0 for s in shape):
                raise ValueError("can only specify one unknown dimension")
            shape = tuple(shape)

        if shape == self.shape:
            return self._new_with_changes(
                    data=self.base_data, offset=self.offset, shape=shape,
                    strides=self.strides)

        import operator
        size = reduce(operator.mul, shape, 1)
        if size != self.size:
            raise ValueError("total size of new array must be unchanged")

        # {{{ determine reshaped strides

        # copied and translated from
        # https://github.com/numpy/numpy/blob/4083883228d61a3b571dec640185b5a5d983bf59/numpy/core/src/multiarray/shape.c  # noqa

        newdims = shape
        newnd = len(newdims)

        # Remove axes with dimension 1 from the old array. They have no effect
        # but would need special cases since their strides do not matter.

        olddims = []
        oldstrides = []
        for oi in range(len(self.shape)):
            s = self.shape[oi]
            if s != 1:
                olddims.append(s)
                oldstrides.append(self.strides[oi])

        oldnd = len(olddims)

        newstrides = [-1]*len(newdims)

        # oi to oj and ni to nj give the axis ranges currently worked with
        oi = 0
        oj = 1
        ni = 0
        nj = 1
        while ni < newnd and oi < oldnd:
            np = newdims[ni]
            op = olddims[oi]

            while np != op:
                if np < op:
                    # Misses trailing 1s, these are handled later
                    np *= newdims[nj]
                    nj += 1
                else:
                    op *= olddims[oj]
                    oj += 1

            # Check whether the original axes can be combined
            for ok in range(oi, oj-1):
                if order == "F":
                    if oldstrides[ok+1] != olddims[ok]*oldstrides[ok]:
                        raise ValueError("cannot reshape without copy")
                else:
                    # C order
                    if (oldstrides[ok] != olddims[ok+1]*oldstrides[ok+1]):
                        raise ValueError("cannot reshape without copy")

            # Calculate new strides for all axes currently worked with
            if order == "F":
                newstrides[ni] = oldstrides[oi]
                for nk in xrange(ni+1, nj):
                    newstrides[nk] = newstrides[nk - 1]*newdims[nk - 1]
            else:
                # C order
                newstrides[nj - 1] = oldstrides[oj - 1]
                for nk in range(nj-1, ni, -1):
                    newstrides[nk - 1] = newstrides[nk]*newdims[nk]

            ni = nj
            nj += 1

            oi = oj
            oj += 1

        # Set strides corresponding to trailing 1s of the new shape.
        if ni >= 1:
            last_stride = newstrides[ni - 1]
        else:
            last_stride = self.dtype.itemsize

        if order == "F":
            last_stride *= newdims[ni - 1]

        for nk in range(ni, len(shape)):
            newstrides[nk] = last_stride

        # }}}

        return self._new_with_changes(
                data=self.base_data, offset=self.offset, shape=shape,
                strides=tuple(newstrides))

    def ravel(self):
        """Returns flattened array containing the same data."""
        return self.reshape(self.size)

    def view(self, dtype=None):
        """Returns view of array with the same data. If *dtype* is different
        from current dtype, the actual bytes of memory will be reinterpreted.
        """

        if dtype is None:
            dtype = self.dtype

        old_itemsize = self.dtype.itemsize
        itemsize = np.dtype(dtype).itemsize

        from pytools import argmin2
        min_stride_axis = argmin2(
                (axis, abs(stride))
                for axis, stride in enumerate(self.strides))

        if self.shape[min_stride_axis] * old_itemsize % itemsize != 0:
            raise ValueError("new type not compatible with array")

        new_shape = (
                self.shape[:min_stride_axis]
                + (self.shape[min_stride_axis] * old_itemsize // itemsize,)
                + self.shape[min_stride_axis+1:])
        new_strides = (
                self.strides[:min_stride_axis]
                + (self.strides[min_stride_axis] * itemsize // old_itemsize,)
                + self.strides[min_stride_axis+1:])

        return self._new_with_changes(
                self.base_data, self.offset,
                shape=new_shape, dtype=dtype,
                strides=new_strides)

    # }}}

    def map_to_host(self, queue=None, flags=None, is_blocking=True, wait_for=None):
        """If *is_blocking*, return a :class:`numpy.ndarray` corresponding to the
        same memory as *self*.

        If *is_blocking* is not true, return a tuple ``(ary, evt)``, where
        *ary* is the above-mentioned array.

        The host array is obtained using :func:`pyopencl.enqueue_map_buffer`.
        See there for further details.

        :arg flags: A combination of :class:`pyopencl.map_flags`.
            Defaults to read-write.

        .. versionadded :: 2013.2
        """

        if flags is None:
            flags = cl.map_flags.READ | cl.map_flags.WRITE

        ary, evt = cl.enqueue_map_buffer(
                queue or self.queue, self.base_data, flags, self.offset,
                self.shape, self.dtype, strides=self.strides, wait_for=wait_for,
                is_blocking=is_blocking)

        if is_blocking:
            return ary
        else:
            return ary, evt

    # {{{ getitem/setitem

    def __getitem__(self, index):
        """
        .. versionadded:: 2013.1
        """

        if isinstance(index, Array):
            if index.dtype.kind != "i":
                raise TypeError(
                        "fancy indexing is only allowed with integers")
            if len(index.shape) != 1:
                raise NotImplementedError(
                        "multidimensional fancy indexing is not supported")
            if len(self.shape) != 1:
                raise NotImplementedError(
                        "fancy indexing into a multi-d array is not supported")

            return take(self, index)

        if not isinstance(index, tuple):
            index = (index,)

        new_shape = []
        new_offset = self.offset
        new_strides = []

        seen_ellipsis = False

        index_axis = 0
        array_axis = 0
        while index_axis < len(index):
            index_entry = index[index_axis]

            if array_axis > len(self.shape):
                raise IndexError("too many axes in index")

            if isinstance(index_entry, slice):
                start, stop, idx_stride = index_entry.indices(
                        self.shape[array_axis])

                array_stride = self.strides[array_axis]

                new_shape.append((stop-start)//idx_stride)
                new_strides.append(idx_stride*array_stride)
                new_offset += array_stride*start

                index_axis += 1
                array_axis += 1

            elif isinstance(index_entry, (int, np.integer)):
                array_shape = self.shape[array_axis]
                if index_entry < 0:
                    index_entry += array_shape

                if not (0 <= index_entry < array_shape):
                    raise IndexError(
                            "subindex in axis %d out of range" % index_axis)

                new_offset += self.strides[array_axis]*index_entry

                index_axis += 1
                array_axis += 1

            elif index_entry is Ellipsis:
                index_axis += 1

                remaining_index_count = len(index) - index_axis
                new_array_axis = len(self.shape) - remaining_index_count
                if new_array_axis < array_axis:
                    raise IndexError("invalid use of ellipsis in index")
                while array_axis < new_array_axis:
                    new_shape.append(self.shape[array_axis])
                    new_strides.append(self.strides[array_axis])
                    array_axis += 1

                if seen_ellipsis:
                    raise IndexError(
                            "more than one ellipsis not allowed in index")
                seen_ellipsis = True

            else:
                raise IndexError("invalid subindex in axis %d" % index_axis)

        while array_axis < len(self.shape):
            new_shape.append(self.shape[array_axis])
            new_strides.append(self.strides[array_axis])

            array_axis += 1

        return self._new_with_changes(
                self.base_data, offset=new_offset,
                shape=tuple(new_shape),
                strides=tuple(new_strides))

    def setitem(self, subscript, value, queue=None, wait_for=None):
        """Like :meth:`__setitem__`, but with the ability to specify
        a *queue* and *wait_for*.

        .. versionadded:: 2013.1

        .. versionchanged:: 2013.2

            Added *wait_for*.
        """

        if isinstance(subscript, Array):
            if subscript.dtype.kind != "i":
                raise TypeError(
                        "fancy indexing is only allowed with integers")
            if len(subscript.shape) != 1:
                raise NotImplementedError(
                        "multidimensional fancy indexing is not supported")
            if len(self.shape) != 1:
                raise NotImplementedError(
                        "fancy indexing into a multi-d array is supported")

            multi_put([value], subscript, out=[self], queue=self.queue,
                    wait_for=wait_for)
            return

        queue = queue or self.queue or value.queue

        subarray = self[subscript]

        if isinstance(value, np.ndarray):
            if subarray.shape == value.shape and subarray.strides == value.strides:
                self.add_event(
                        cl.enqueue_copy(queue, subarray.base_data,
                            value, device_offset=subarray.offset, wait_for=wait_for))
                return
            else:
                value = to_device(queue, value, self.allocator)

        if isinstance(value, Array):
            if len(subarray.shape) != len(value.shape):
                raise NotImplementedError("broadcasting is not "
                        "supported in __setitem__")
            if subarray.shape != value.shape:
                raise ValueError("cannot assign between arrays of "
                        "differing shapes")
            if subarray.strides != value.strides:
                raise ValueError("cannot assign between arrays of "
                        "differing strides")

            self.add_event(
                    self._copy(subarray, value, queue=queue, wait_for=wait_for))

        else:
            # Let's assume it's a scalar
            subarray.fill(value, queue=queue, wait_for=wait_for)

    def __setitem__(self, subscript, value):
        """Set the slice of *self* identified *subscript* to *value*.

        *value* is allowed to be:

        * A :class:`Array` of the same :attr:`shape` and (for now) :attr:`strides`,
          but with potentially different :attr:`dtype`.
        * A :class:`numpy.ndarray` of the same :attr:`shape` and (for now)
          :attr:`strides`, but with potentially different :attr:`dtype`.
        * A scalar.

        Non-scalar broadcasting is not currently supported.

        .. versionadded:: 2013.1
        """
        self.setitem(subscript, value)

    # }}}

# }}}


def as_strided(ary, shape=None, strides=None):
    """Make an :class:`Array` from the given array with the given
    shape and strides.
    """

    # undocumented for the moment

    shape = shape or ary.shape
    strides = strides or ary.strides

    return Array(ary.queue, shape, ary.dtype, allocator=ary.allocator,
            data=ary.data, strides=strides)

# }}}


# {{{ creation helpers

def to_device(queue, ary, allocator=None, async=False):
    """Return a :class:`Array` that is an exact copy of the
    :class:`numpy.ndarray` instance *ary*.

    See :class:`Array` for the meaning of *allocator*.

    .. versionchanged:: 2011.1
        *context* argument was deprecated.
    """

    if _dtype_is_object(ary.dtype):
        raise RuntimeError("to_device does not work on object arrays.")

    result = Array(queue, ary.shape, ary.dtype,
                    allocator=allocator, strides=ary.strides)
    result.set(ary, async=async)
    return result


empty = Array


def zeros(queue, shape, dtype, order="C", allocator=None):
    """Same as :func:`empty`, but the :class:`Array` is zero-initialized before
    being returned.

    .. versionchanged:: 2011.1
        *context* argument was deprecated.
    """

    result = Array(queue, shape, dtype,
            order=order, allocator=allocator)
    zero = np.zeros((), dtype)
    result.fill(zero)
    return result


def empty_like(ary):
    """Make a new, uninitialized :class:`Array` having the same properties
    as *other_ary*.
    """

    return ary._new_with_changes(data=None, offset=0)


def zeros_like(ary):
    """Make a new, zero-initialized :class:`Array` having the same properties
    as *other_ary*.
    """

    result = empty_like(ary)
    zero = np.zeros((), ary.dtype)
    result.fill(zero)
    return result


@elwise_kernel_runner
def _arange_knl(result, start, step):
    return elementwise.get_arange_kernel(
            result.context, result.dtype)


def arange(queue, *args, **kwargs):
    """Create a :class:`Array` filled with numbers spaced `step` apart,
    starting from `start` and ending at `stop`.

    For floating point arguments, the length of the result is
    `ceil((stop - start)/step)`.  This rule may result in the last
    element of the result being greater than `stop`.

    *dtype*, if not specified, is taken as the largest common type
    of *start*, *stop* and *step*.

    .. versionchanged:: 2011.1
        *context* argument was deprecated.

    .. versionchanged:: 2011.2
        *allocator* keyword argument was added.
    """

    # argument processing -----------------------------------------------------

    # Yuck. Thanks, numpy developers. ;)
    from pytools import Record

    class Info(Record):
        pass

    explicit_dtype = False

    inf = Info()
    inf.start = None
    inf.stop = None
    inf.step = None
    inf.dtype = None
    inf.allocator = None
    inf.wait_for = []

    if isinstance(args[-1], np.dtype):
        inf.dtype = args[-1]
        args = args[:-1]
        explicit_dtype = True

    argc = len(args)
    if argc == 0:
        raise ValueError("stop argument required")
    elif argc == 1:
        inf.stop = args[0]
    elif argc == 2:
        inf.start = args[0]
        inf.stop = args[1]
    elif argc == 3:
        inf.start = args[0]
        inf.stop = args[1]
        inf.step = args[2]
    else:
        raise ValueError("too many arguments")

    admissible_names = ["start", "stop", "step", "dtype", "allocator"]
    for k, v in kwargs.iteritems():
        if k in admissible_names:
            if getattr(inf, k) is None:
                setattr(inf, k, v)
                if k == "dtype":
                    explicit_dtype = True
            else:
                raise ValueError(
                        "may not specify '%s' by position and keyword" % k)
        else:
            raise ValueError("unexpected keyword argument '%s'" % k)

    if inf.start is None:
        inf.start = 0
    if inf.step is None:
        inf.step = 1
    if inf.dtype is None:
        inf.dtype = np.array([inf.start, inf.stop, inf.step]).dtype

    # actual functionality ----------------------------------------------------
    dtype = np.dtype(inf.dtype)
    start = dtype.type(inf.start)
    step = dtype.type(inf.step)
    stop = dtype.type(inf.stop)
    wait_for = inf.wait_for

    if not explicit_dtype:
        raise TypeError("arange requires a dtype argument")

    from math import ceil
    size = int(ceil((stop-start)/step))

    result = Array(queue, (size,), dtype, allocator=inf.allocator)
    result.add_event(
            _arange_knl(result, start, step, queue=queue, wait_for=wait_for))
    return result

# }}}


# {{{ take/put/concatenate/diff

@elwise_kernel_runner
def _take(result, ary, indices):
    return elementwise.get_take_kernel(
            result.context, result.dtype, indices.dtype)


def take(a, indices, out=None, queue=None, wait_for=None):
    """Return the :class:`Array` ``[a[indices[0]], ..., a[indices[n]]]``.
    For the moment, *a* must be a type that can be bound to a texture.
    """

    queue = queue or a.queue
    if out is None:
        out = Array(queue, indices.shape, a.dtype, allocator=a.allocator)

    assert len(indices.shape) == 1
    out.add_event(
            _take(out, a, indices, queue=queue, wait_for=wait_for))
    return out


def multi_take(arrays, indices, out=None, queue=None):
    if not len(arrays):
        return []

    assert len(indices.shape) == 1

    from pytools import single_valued
    a_dtype = single_valued(a.dtype for a in arrays)
    a_allocator = arrays[0].dtype
    context = indices.context
    queue = queue or indices.queue

    vec_count = len(arrays)

    if out is None:
        out = [Array(context, queue, indices.shape, a_dtype,
            allocator=a_allocator)
                for i in range(vec_count)]
    else:
        if len(out) != len(arrays):
            raise ValueError("out and arrays must have the same length")

    chunk_size = _builtin_min(vec_count, 10)

    def make_func_for_chunk_size(chunk_size):
        knl = elementwise.get_take_kernel(
                indices.context, a_dtype, indices.dtype,
                vec_count=chunk_size)
        knl.set_block_shape(*indices._block)
        return knl

    knl = make_func_for_chunk_size(chunk_size)

    for start_i in range(0, len(arrays), chunk_size):
        chunk_slice = slice(start_i, start_i+chunk_size)

        if start_i + chunk_size > vec_count:
            knl = make_func_for_chunk_size(vec_count-start_i)

        gs, ls = indices.get_sizes(queue,
                knl.get_work_group_info(
                    cl.kernel_work_group_info.WORK_GROUP_SIZE,
                    queue.device))

        knl(queue, gs, ls,
                indices.data,
                *([o.data for o in out[chunk_slice]]
                    + [i.data for i in arrays[chunk_slice]]
                    + [indices.size]))

    return out


def multi_take_put(arrays, dest_indices, src_indices, dest_shape=None,
        out=None, queue=None, src_offsets=None):
    if not len(arrays):
        return []

    from pytools import single_valued
    a_dtype = single_valued(a.dtype for a in arrays)
    a_allocator = arrays[0].allocator
    context = src_indices.context
    queue = queue or src_indices.queue

    vec_count = len(arrays)

    if out is None:
        out = [Array(queue, dest_shape, a_dtype, allocator=a_allocator)
                for i in range(vec_count)]
    else:
        if a_dtype != single_valued(o.dtype for o in out):
            raise TypeError("arrays and out must have the same dtype")
        if len(out) != vec_count:
            raise ValueError("out and arrays must have the same length")

    if src_indices.dtype != dest_indices.dtype:
        raise TypeError(
                "src_indices and dest_indices must have the same dtype")

    if len(src_indices.shape) != 1:
        raise ValueError("src_indices must be 1D")

    if src_indices.shape != dest_indices.shape:
        raise ValueError(
                "src_indices and dest_indices must have the same shape")

    if src_offsets is None:
        src_offsets_list = []
    else:
        src_offsets_list = src_offsets
        if len(src_offsets) != vec_count:
            raise ValueError(
                    "src_indices and src_offsets must have the same length")

    max_chunk_size = 10

    chunk_size = _builtin_min(vec_count, max_chunk_size)

    def make_func_for_chunk_size(chunk_size):
        return elementwise.get_take_put_kernel(context,
                a_dtype, src_indices.dtype,
                with_offsets=src_offsets is not None,
                vec_count=chunk_size)

    knl = make_func_for_chunk_size(chunk_size)

    for start_i in range(0, len(arrays), chunk_size):
        chunk_slice = slice(start_i, start_i+chunk_size)

        if start_i + chunk_size > vec_count:
            knl = make_func_for_chunk_size(vec_count-start_i)

        gs, ls = src_indices.get_sizes(queue,
                knl.get_work_group_info(
                    cl.kernel_work_group_info.WORK_GROUP_SIZE,
                    queue.device))

        from pytools import flatten
        knl(queue, gs, ls,
                *([o.data for o in out[chunk_slice]]
                    + [dest_indices.base_data,
                        dest_indices.offset,
                        src_indices.base_data,
                        src_indices.offset]
                    + list(flatten(
                        (i.base_data, i.offset)
                        for i in arrays[chunk_slice]))
                    + src_offsets_list[chunk_slice]
                    + [src_indices.size]))

    return out


def multi_put(arrays, dest_indices, dest_shape=None, out=None, queue=None,
        wait_for=None):
    if not len(arrays):
        return []

    from pytools import single_valued
    a_dtype = single_valued(a.dtype for a in arrays)
    a_allocator = arrays[0].allocator
    context = dest_indices.context
    queue = queue or dest_indices.queue

    vec_count = len(arrays)

    if out is None:
        out = [Array(queue, dest_shape, a_dtype,
            allocator=a_allocator, queue=queue)
            for i in range(vec_count)]
    else:
        if a_dtype != single_valued(o.dtype for o in out):
            raise TypeError("arrays and out must have the same dtype")
        if len(out) != vec_count:
            raise ValueError("out and arrays must have the same length")

    if len(dest_indices.shape) != 1:
        raise ValueError("dest_indices must be 1D")

    chunk_size = _builtin_min(vec_count, 10)

    def make_func_for_chunk_size(chunk_size):
        knl = elementwise.get_put_kernel(
                context,
                a_dtype, dest_indices.dtype, vec_count=chunk_size)
        return knl

    knl = make_func_for_chunk_size(chunk_size)

    for start_i in range(0, len(arrays), chunk_size):
        chunk_slice = slice(start_i, start_i+chunk_size)

        if start_i + chunk_size > vec_count:
            knl = make_func_for_chunk_size(vec_count-start_i)

        gs, ls = dest_indices.get_sizes(queue,
                knl.get_work_group_info(
                    cl.kernel_work_group_info.WORK_GROUP_SIZE,
                    queue.device))

        from pytools import flatten
        evt = knl(queue, gs, ls,
                *(
                    list(flatten(
                        (o.base_data, o.offset)
                        for o in out[chunk_slice]))
                    + [dest_indices.base_data, dest_indices.offset]
                    + list(flatten(
                        (i.base_data, i.offset)
                        for i in arrays[chunk_slice]))
                    + [dest_indices.size]),
                **dict(wait_for=wait_for))

        # FIXME should wait on incoming events

        for o in out[chunk_slice]:
            o.add_event(evt)

    return out


def concatenate(arrays, axis=0, queue=None, allocator=None):
    """
    .. versionadded:: 2013.1
    """
    # {{{ find properties of result array

    shape = None

    for i_ary, ary in enumerate(arrays):
        queue = queue or ary.queue
        allocator = allocator or ary.allocator

        if shape is None:
            # first array
            shape = list(ary.shape)
        else:
            if len(ary.shape) != len(shape):
                raise ValueError("%d'th array has different number of axes "
                        "(shold have %d, has %d)"
                        % (i_ary, len(ary.shape), len(shape)))

            ary_shape_list = list(ary.shape)
            if (ary_shape_list[:axis] != shape[:axis]
                    or ary_shape_list[axis+1:] != shape[axis+1:]):
                raise ValueError("%d'th array has residual not matching "
                        "other arrays" % i_ary)

            shape[axis] += ary.shape[axis]

    # }}}

    shape = tuple(shape)
    dtype = np.find_common_type([ary.dtype for ary in arrays], [])
    result = empty(queue, shape, dtype, allocator=allocator)

    full_slice = (slice(None),) * len(shape)

    base_idx = 0
    for ary in arrays:
        my_len = ary.shape[axis]
        result.setitem(
                full_slice[:axis]
                + (slice(base_idx, base_idx+my_len),)
                + full_slice[axis+1:],
                ary)

        base_idx += my_len

    return result


@elwise_kernel_runner
def _diff(result, array):
    return elementwise.get_diff_kernel(array.context, array.dtype)


def diff(array, queue=None, allocator=None):
    """
    .. versionadded:: 2013.2
    """

    if len(array.shape) != 1:
        raise ValueError("multi-D arrays are not supported")

    n, = array.shape

    queue = queue or array.queue
    allocator = allocator or array.allocator

    result = empty(queue, (n-1,), array.dtype, allocator=allocator)
    _diff(result, array, queue=queue)
    return result


def hstack(arrays, queue=None):
    from pyopencl.array import empty

    if len(arrays) == 0:
        return empty(queue, (), dtype=np.float64)

    if queue is None:
        for ary in arrays:
            if ary.queue is not None:
                queue = ary.queue
                break

    from pytools import all_equal, single_valued
    if not all_equal(len(ary.shape) for ary in arrays):
        raise ValueError("arguments must all have the same number of axes")

    lead_shape = single_valued(ary.shape[:-1] for ary in arrays)

    w = _builtin_sum([ary.shape[-1] for ary in arrays])
    result = empty(queue, lead_shape+(w,), arrays[0].dtype)
    index = 0
    for ary in arrays:
        result[..., index:index+ary.shape[-1]] = ary
        index += ary.shape[-1]

    return result

# }}}


# {{{ conditionals

@elwise_kernel_runner
def _if_positive(result, criterion, then_, else_):
    return elementwise.get_if_positive_kernel(
            result.context, criterion.dtype, then_.dtype)


def if_positive(criterion, then_, else_, out=None, queue=None):
    """Return an array like *then_*, which, for the element at index *i*,
    contains *then_[i]* if *criterion[i]>0*, else *else_[i]*.
    """

    if not (criterion.shape == then_.shape == else_.shape):
        raise ValueError("shapes do not match")

    if not (then_.dtype == else_.dtype):
        raise ValueError("dtypes do not match")

    if out is None:
        out = empty_like(then_)
    _if_positive(out, criterion, then_, else_, queue=queue)
    return out


def maximum(a, b, out=None, queue=None):
    """Return the elementwise maximum of *a* and *b*."""

    # silly, but functional
    return if_positive(a.mul_add(1, b, -1, queue=queue), a, b,
            queue=queue, out=out)


def minimum(a, b, out=None, queue=None):
    """Return the elementwise minimum of *a* and *b*."""
    # silly, but functional
    return if_positive(a.mul_add(1, b, -1, queue=queue), b, a,
            queue=queue, out=out)

# }}}


# {{{ reductions
_builtin_sum = sum
_builtin_min = min
_builtin_max = max


def sum(a, dtype=None, queue=None):
    """
    .. versionadded:: 2011.1
    """
    from pyopencl.reduction import get_sum_kernel
    krnl = get_sum_kernel(a.context, dtype, a.dtype)
    return krnl(a, queue=queue)


def dot(a, b, dtype=None, queue=None):
    """
    .. versionadded:: 2011.1
    """
    from pyopencl.reduction import get_dot_kernel
    krnl = get_dot_kernel(a.context, dtype, a.dtype, b.dtype)
    return krnl(a, b, queue=queue)


def vdot(a, b, dtype=None, queue=None):
    """Like :func:`numpy.vdot`.

    .. versionadded:: 2013.1
    """
    from pyopencl.reduction import get_dot_kernel
    krnl = get_dot_kernel(a.context, dtype, a.dtype, b.dtype,
            conjugate_first=True)
    return krnl(a, b, queue=queue)


def subset_dot(subset, a, b, dtype=None, queue=None):
    """
    .. versionadded:: 2011.1
    """
    from pyopencl.reduction import get_subset_dot_kernel
    krnl = get_subset_dot_kernel(
            a.context, dtype, subset.dtype, a.dtype, b.dtype)
    return krnl(subset, a, b, queue=queue)


def _make_minmax_kernel(what):
    def f(a, queue=None):
        from pyopencl.reduction import get_minmax_kernel
        krnl = get_minmax_kernel(a.context, what, a.dtype)
        return krnl(a,  queue=queue)

    return f

min = _make_minmax_kernel("min")
min.__doc__ = """
    .. versionadded:: 2011.1
    """

max = _make_minmax_kernel("max")
max.__doc__ = """
    .. versionadded:: 2011.1
    """


def _make_subset_minmax_kernel(what):
    def f(subset, a, queue=None):
        from pyopencl.reduction import get_subset_minmax_kernel
        krnl = get_subset_minmax_kernel(a.context, what, a.dtype, subset.dtype)
        return krnl(subset, a,  queue=queue)

    return f

subset_min = _make_subset_minmax_kernel("min")
subset_min.__doc__ = """.. versionadded:: 2011.1"""
subset_max = _make_subset_minmax_kernel("max")
subset_max.__doc__ = """.. versionadded:: 2011.1"""

# }}}


# {{{ scans

def cumsum(a, output_dtype=None, queue=None,
        wait_for=None, return_event=False):
    # undocumented for now

    """
    .. versionadded:: 2013.1
    """

    if output_dtype is None:
        output_dtype = a.dtype

    result = a._new_like_me(output_dtype)

    from pyopencl.scan import get_cumsum_kernel
    krnl = get_cumsum_kernel(a.context, a.dtype, output_dtype)
    evt = krnl(a, result, queue=queue, wait_for=wait_for)

    if return_event:
        return evt, result
    else:
        return result

# }}}

# vim: foldmethod=marker