forked from Python-for-HPC/PyOMP
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathCHANGE_LOG
4667 lines (3927 loc) · 190 KB
/
CHANGE_LOG
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
Version 0.54.0
--------------
In development
NOTES:
* Minimum python version raised to 3.7.
Version 0.53.1 (25 March, 2021)
-------------------------------
This is a bugfix release for 0.53.0. It contains the following four
pull-requests which fix two critical regressions and two build failures
reported by the openSuSe team:
* PR #6826 Fix regression on gufunc serialization
* PR #6828 Fix regression in CUDA: Set stream in mapped and managed array
device_setup
* PR #6837 Ignore warnings from packaging module when testing import behaviour.
* PR #6851 set non-reported llvm timing values to 0.0
Authors:
* Ben Greiner
* Graham Markall
* Siu Kwan Lam
* Stuart Archibald
Version 0.53.0 (11 March, 2021)
-------------------------------
This release continues to add new features, bug fixes and stability improvements
to Numba.
Highlights of core changes:
* Support for Python 3.9 (Stuart Archibald).
* Function sub-typing (Lucio Fernandez-Arjona).
* Initial support for dynamic ``gufuncs`` (i.e. from ``@guvectorize``)
(Guilherme Leobas).
* Parallel Accelerator (``@njit(parallel=True)`` now supports Fortran ordered
arrays (Todd A. Anderson and Siu Kwan Lam).
Intel also kindly sponsored research and development that lead to two new
features:
* Exposing LLVM compilation pass timings for diagnostic purposes (Siu Kwan
Lam).
* An event system for broadcasting compiler events (Siu Kwan Lam).
Highlights of changes for the CUDA target:
* CUDA 11.2 onwards (versions of the toolkit using NVVM IR 1.6 / LLVM IR 7.0.1)
are now supported (Graham Markall).
* A fast cube root function is added (Michael Collison).
* Support for atomic ``xor``, increment, decrement, exchange, are added, and
compare-and-swap is extended to support 64-bit integers (Michael Collison).
* Addition of ``cuda.is_supported_version()`` to check if the CUDA runtime
version is supported (Graham Markall).
* The CUDA dispatcher now shares infrastructure with the CPU dispatcher,
improving launch times for lazily-compiled kernels (Graham Markall).
* The CUDA Array Interface is updated to version 3, with support for streams
added (Graham Markall).
* Tuples and ``namedtuples`` can now be passed to kernels (Graham Markall).
* Initial support for Cooperative Groups is added, with support for Grid Groups
and Grid Sync (Graham Markall and Nick White).
* Support for ``math.log2`` and ``math.remainder`` is added (Guilherme Leobas).
General deprecation notices:
* There are no new general deprecations.
CUDA target deprecation notices:
* CUDA support on macOS is deprecated with this release (it still works, it is
just unsupported).
* The ``argtypes``, ``restypes``, and ``bind`` keyword arguments to the
``cuda.jit`` decorator, deprecated since 0.51.0, are removed
* The ``Device.COMPUTE_CAPABILITY`` property, deprecated since 2014, has been
removed (use ``compute_capability`` instead).
* The ``to_host`` method of device arrays is removed (use ``copy_to_host``
instead).
General Enhancements:
* PR #4769: objmode complex type spelling (Siu Kwan Lam)
* PR #5579: Function subtyping (Lucio Fernandez-Arjona)
* PR #5659: Add support for parfors creating 'F'ortran layout Numpy arrays.
(Todd A. Anderson)
* PR #5936: Improve array analysis for user-defined data types. (Todd A.
Anderson)
* PR #5938: Initial support for dynamic gufuncs (Guilherme Leobas)
* PR #5958: Making typed.List a typing Generic (Lucio Fernandez-Arjona)
* PR #6334: Support attribute access from other modules (Farah Hariri)
* PR #6373: Allow Dispatchers to be cached (Eric Wieser)
* PR #6519: Avoid unnecessary ir.Del generation and removal (Ehsan Totoni)
* PR #6545: Refactoring ParforDiagnostics (Elena Totmenina)
* PR #6560: Add LLVM pass timer (Siu Kwan Lam)
* PR #6573: Improve ``__str__`` for typed.List when invoked from IPython shell
(Amin Sadeghi)
* PR #6575: Avoid temp variable assignments (Ehsan Totoni)
* PR #6578: Add support for numpy ``intersect1d`` and basic test cases
(``@caljrobe``)
* PR #6579: Python 3.9 support. (Stuart Archibald)
* PR #6580: Store partial typing errors in compiler state (Ehsan Totoni)
* PR #6626: A simple event system to broadcast compiler events (Siu Kwan Lam)
* PR #6635: Try to resolve dynamic getitems as static post unroll transform.
(Stuart Archibald)
* PR #6636: Adds llvm_lock event (Siu Kwan Lam)
* PR #6664: Adds tests for PR 5659 (Siu Kwan Lam)
* PR #6680: Allow getattr to work in objmode output type spec (Siu Kwan Lam)
Fixes:
* PR #6176: Remove references to deprecated numpy globals (Eric Wieser)
* PR #6374: Use Python 3 style OSError handling (Eric Wieser)
* PR #6402: Fix ``typed.Dict`` and ``typed.List`` crashing on parametrized types
(Andreas Sodeur)
* PR #6403: Add ``types.ListType.key`` (Andreas Sodeur)
* PR #6410: Fixes issue #6386 (Danny Weitekamp)
* PR #6425: Fix unicode join for issue #6405 (Teugea Ioan-Teodor)
* PR #6437: Don't pass reduction variables known in an outer parfor to inner
parfors when analyzing reductions. (Todd A. Anderson)
* PR #6453: Keep original variable names in metadata to improve diagnostics
(Ehsan Totoni)
* PR #6454: FIX: Fixes for literals (Eric Larson)
* PR #6463: Bump llvmlite to 0.36 series (Stuart Archibald)
* PR #6466: Remove the misspelling of finalize_dynamic_globals (Sergey
Pokhodenko)
* PR #6489: Improve the error message for unsupported Buffer in Buffer
situation. (Stuart Archibald)
* PR #6503: Add test to ensure Numba imports without warnings. (Stuart
Archibald)
* PR #6508: Defer requirements to setup.py (Siu Kwan Lam)
* PR #6521: Skip annotated jitclass test if typeguard is running. (Stuart
Archibald)
* PR #6524: Fix typed.List return value (Lucio Fernandez-Arjona)
* PR #6562: Correcting typo in numba sysinfo output (Nick Sutcliffe)
* PR #6574: Run parfor fusion if 2 or more parfors (Ehsan Totoni)
* PR #6582: Fix typed dict error with uninitialized padding bytes (Siu Kwan
Lam)
* PR #6584: Remove jitclass from ``__init__`` ``__all__``. (Stuart Archibald)
* PR #6586: Run closure inlining ahead of branch pruning in case of nonlocal
(Stuart Archibald)
* PR #6591: Fix inlineasm test failure. (Siu Kwan Lam)
* PR #6622: Fix 6534, handle unpack of assign-like tuples. (Stuart Archibald)
* PR #6652: Simplify PR-6334 (Siu Kwan Lam)
* PR #6653: Fix get_numba_envvar (Siu Kwan Lam)
* PR #6654: Fix #6632 support alternative dtype string spellings (Stuart
Archibald)
* PR #6685: Add Python 3.9 to classifiers. (Stuart Archibald)
* PR #6693: patch to compile _devicearray.cpp with c++11 (Valentin Haenel)
* PR #6716: Consider assignment lhs live if used in rhs (Fixes #6715) (Ehsan
Totoni)
* PR #6727: Avoid errors in array analysis for global tuples with non-int
(Ehsan Totoni)
* PR #6733: Fix segfault and errors in #6668 (Siu Kwan Lam)
* PR #6741: Enable SSA in IR inliner (Ehsan Totoni)
* PR #6763: use an alternative constraint for the conda packages (Valentin
Haenel)
* PR #6786: Fix gufunc kwargs support (Siu Kwan Lam)
CUDA Enhancements/Fixes:
* PR #5162: Specify synchronization semantics of CUDA Array Interface (Graham
Markall)
* PR #6245: CUDA Cooperative grid groups (Graham Markall and Nick White)
* PR #6333: Remove dead ``_Kernel.__call__`` (Graham Markall)
* PR #6343: CUDA: Add support for passing tuples and namedtuples to kernels
(Graham Markall)
* PR #6349: Refactor Dispatcher to remove unnecessary indirection (Graham
Markall)
* PR #6358: Add log2 and remainder implementations for cuda (Guilherme Leobas)
* PR #6376: Added a fixed seed in test_atomics.py for issue #6370 (Teugea
Ioan-Teodor)
* PR #6377: CUDA: Fix various issues in test suite (Graham Markall)
* PR #6409: Implement cuda atomic xor (Michael Collison)
* PR #6422: CUDA: Remove deprecated items, expect CUDA 11.1 (Graham Markall)
* PR #6427: Remove duplicate repeated definition of gufunc (Amit Kumar)
* PR #6432: CUDA: Use ``_dispatcher.Dispatcher`` as base Dispatcher class
(Graham Markall)
* PR #6447: CUDA: Add get_regs_per_thread method to Dispatcher (Graham Markall)
* PR #6499: CUDA atomic increment, decrement, exchange and compare and swap
(Michael Collison)
* PR #6510: CUDA: Make device array assignment synchronous where necessary
(Graham Markall)
* PR #6517: CUDA: Add NVVM test of all 8-bit characters (Graham Markall)
* PR #6567: Refactor llvm replacement code into separate function (Michael
Collison)
* PR #6642: Testhound/cuda cuberoot (Michael Collison)
* PR #6661: CUDA: Support NVVM70 / CUDA 11.2 (Graham Markall)
* PR #6663: Fix error caused by missing "-static" libraries defined for some
platforms (Siu Kwan Lam)
* PR #6666: CUDA: Add a function to query whether the runtime version is
supported. (Graham Markall)
* PR #6725: CUDA: Fix compile to PTX with debug for CUDA 11.2 (Graham Markall)
Documentation Updates:
* PR #5740: Add FAQ entry on how to create a MWR. (Stuart Archibald)
* PR #6346: DOC: add where to get dev builds from to FAQ (Eyal Trabelsi)
* PR #6418: docs: use https for homepage (``@imba-tjd``)
* PR #6430: CUDA docs: Add RNG example with 3D grid and strided loops (Graham
Markall)
* PR #6436: docs: remove typo in Deprecation Notices (Thibault Ballier)
* PR #6440: Add note about performance of typed containers from the interpreter.
(Stuart Archibald)
* PR #6457: Link to read the docs instead of numba homepage (Hannes Pahl)
* PR #6470: Adding PyCon Sweden 2020 talk on numba (Ankit Mahato)
* PR #6472: Document ``numba.extending.is_jitted`` (Stuart Archibald)
* PR #6495: Fix typo in literal list docs. (Stuart Archibald)
* PR #6501: Add doc entry on Numba's limited resources and how to help. (Stuart
Archibald)
* PR #6502: Add CODEOWNERS file. (Stuart Archibald)
* PR #6531: Update canonical URL. (Stuart Archibald)
* PR #6544: Minor typo / grammar fixes to 5 minute guide (Ollin Boer Bohan)
* PR #6599: docs: fix simple typo, consevatively -> conservatively (Tim Gates)
* PR #6609: Recommend miniforge instead of c4aarch64 (Isuru Fernando)
* PR #6671: Update environment creation example to python 3.8 (Lucio
Fernandez-Arjona)
* PR #6676: Update hardware and software versions in various docs. (Stuart
Archibald)
* PR #6682: Update deprecation notices for 0.53 (Stuart Archibald)
CI/Infrastructure Updates:
* PR #6458: Enable typeguard in CI (Siu Kwan Lam)
* PR #6500: Update bug and feature request templates. (Stuart Archibald)
* PR #6516: Fix RTD build by using conda. (Stuart Archibald)
* PR #6587: Add zenodo badge (Siu Kwan Lam)
Authors:
* Amin Sadeghi
* Amit Kumar
* Andreas Sodeur
* Ankit Mahato
* Chris Barnes
* Danny Weitekamp
* Ehsan Totoni (core dev)
* Eric Larson
* Eric Wieser
* Eyal Trabelsi
* Farah Hariri
* Graham Markall
* Guilherme Leobas
* Hannes Pahl
* Isuru Fernando
* Lucio Fernandez-Arjona
* Michael Collison
* Nick Sutcliffe
* Nick White
* Ollin Boer Bohan
* Sergey Pokhodenko
* Siu Kwan Lam (core dev)
* Stuart Archibald (core dev)
* Teugea Ioan-Teodor
* Thibault Ballier
* Tim Gates
* Todd A. Anderson (core dev)
* Valentin Haenel (core dev)
* ``@caljrobe``
* ``@imba-tjd``
Version 0.52.0 (30 November, 2020)
----------------------------------
This release focuses on performance improvements, but also adds some new
features and contains numerous bug fixes and stability improvements.
Highlights of core performance improvements include:
* Intel kindly sponsored research and development into producing a new reference
count pruning pass. This pass operates at the LLVM level and can prune a
number of common reference counting patterns. This will improve performance
for two primary reasons:
* There will be less pressure on the atomic locks used to do the reference
counting.
* Removal of reference counting operations permits more inlining and the
optimisation passes can in general do more with what is present.
(Siu Kwan Lam).
* Intel also sponsored work to improve the performance of the
``numba.typed.List`` container, particularly in the case of ``__getitem__``
and iteration (Stuart Archibald).
* Superword-level parallelism vectorization is now switched on and the
optimisation pipeline has been lightly analysed and tuned so as to be able to
vectorize more and more often (Stuart Archibald).
Highlights of core feature changes include:
* The ``inspect_cfg`` method on the JIT dispatcher object has been
significantly enhanced and now includes highlighted output and interleaved
line markers and Python source (Stuart Archibald).
* The BSD operating system is now unofficially supported (Stuart Archibald).
* Numerous features/functionality improvements to NumPy support, including
support for:
* ``np.asfarray`` (Guilherme Leobas)
* "subtyping" in record arrays (Lucio Fernandez-Arjona)
* ``np.split`` and ``np.array_split`` (Isaac Virshup)
* ``operator.contains`` with ``ndarray`` (``@mugoh``).
* ``np.asarray_chkfinite`` (Rishabh Varshney).
* NumPy 1.19 (Stuart Archibald).
* the ``ndarray`` allocators, ``empty``, ``ones`` and ``zeros``, accepting a
``dtype`` specified as a string literal (Stuart Archibald).
* Booleans are now supported as literal types (Alexey Kozlov).
* On the CUDA target:
* CUDA 9.0 is now the minimum supported version (Graham Markall).
* Support for Unified Memory has been added (Max Katz).
* Kernel launch overhead is reduced (Graham Markall).
* Cudasim support for mapped array, memcopies and memset has been added (Mike
Williams).
* Access has been wired in to all libdevice functions (Graham Markall).
* Additional CUDA atomic operations have been added (Michael Collison).
* Additional math library functions (``frexp``, ``ldexp``, ``isfinite``)
(Zhihao Yuan).
* Support for ``power`` on complex numbers (Graham Markall).
Deprecations to note:
There are no new deprecations. However, note that "compatibility" mode, which
was added some 40 releases ago to help transition from 0.11 to 0.12+, has been
removed! Also, the shim to permit the import of ``jitclass`` from Numba's top
level namespace has now been removed as per the deprecation schedule.
General Enhancements:
* PR #5418: Add np.asfarray impl (Guilherme Leobas)
* PR #5560: Record subtyping (Lucio Fernandez-Arjona)
* PR #5609: Jitclass Infer Spec from Type Annotations (Ethan Pronovost)
* PR #5699: Implement np.split and np.array_split (Isaac Virshup)
* PR #6015: Adding BooleanLiteral type (Alexey Kozlov)
* PR #6027: Support operators inlining in InlineOverloads (Alexey Kozlov)
* PR #6038: Closes #6037, fixing FreeBSD compilation (László Károlyi)
* PR #6086: Add more accessible version information (Stuart Archibald)
* PR #6157: Add pipeline_class argument to @cfunc as supported by @jit. (Arthur
Peters)
* PR #6262: Support dtype from str literal. (Stuart Archibald)
* PR #6271: Support ``ndarray`` contains (``@mugoh``)
* PR #6295: Enhance inspect_cfg (Stuart Archibald)
* PR #6304: Support NumPy 1.19 (Stuart Archibald)
* PR #6309: Add suitable file search path for BSDs. (Stuart Archibald)
* PR #6341: Re roll 6279 (Rishabh Varshney and Valentin Haenel)
Performance Enhancements:
* PR #6145: Patch to fingerprint namedtuples. (Stuart Archibald)
* PR #6202: Speed up str(int) (Stuart Archibald)
* PR #6261: Add np.ndarray.ptp() support. (Stuart Archibald)
* PR #6266: Use custom LLVM refcount pruning pass (Siu Kwan Lam)
* PR #6275: Switch on SLP vectorize. (Stuart Archibald)
* PR #6278: Improve typed list performance. (Stuart Archibald)
* PR #6335: Split optimisation passes. (Stuart Archibald)
* PR #6455: Fix refprune on obfuscated refs and stabilize optimisation WRT
wrappers. (Stuart Archibald)
Fixes:
* PR #5639: Make UnicodeType inherit from Hashable (Stuart Archibald)
* PR #6006: Resolves incorrectly hoisted list in parfor. (Todd A. Anderson)
* PR #6126: fix version_info if version can not be determined (Valentin Haenel)
* PR #6137: Remove references to Python 2's long (Eric Wieser)
* PR #6139: Use direct syntax instead of the ``add_metaclass`` decorator (Eric
Wieser)
* PR #6140: Replace calls to utils.iteritems(d) with d.items() (Eric Wieser)
* PR #6141: Fix #6130 objmode cache segfault (Siu Kwan Lam)
* PR #6156: Remove callers of ``reraise`` in favor of using ``with_traceback``
directly (Eric Wieser)
* PR #6162: Move charseq support out of init (Stuart Archibald)
* PR #6165: #5425 continued (Amos Bird and Stuart Archibald)
* PR #6166: Remove Python 2 compatibility from numba.core.utils (Eric Wieser)
* PR #6185: Better error message on NotDefinedError (Luiz Almeida)
* PR #6194: Remove recursion from traverse_types (Radu Popovici)
* PR #6200: Workaround #5973 (Stuart Archibald)
* PR #6203: Make find_callname only lookup functions that are likely part of
NumPy. (Stuart Archibald)
* PR #6204: Fix unicode kind selection for getitem. (Stuart Archibald)
* PR #6206: Build all extension modules with -g -Wall -Werror on Linux x86,
provide -O0 flag option (Graham Markall)
* PR #6212: Fix for objmode recompilation issue (Alexey Kozlov)
* PR #6213: Fix #6177. Remove AOT dependency on the Numba package (Siu Kwan Lam)
* PR #6224: Add support for tuple concatenation to array analysis. (#5396
continued) (Todd A. Anderson)
* PR #6231: Remove compatibility mode (Graham Markall)
* PR #6254: Fix win-32 hashing bug (from Stuart Archibald) (Ray Donnelly)
* PR #6265: Fix #6260 (Stuart Archibald)
* PR #6267: speed up a couple of really slow unittests (Stuart Archibald)
* PR #6281: Remove numba.jitclass shim as per deprecation schedule. (Stuart
Archibald)
* PR #6294: Make return type propagate to all return variables (Andreas Sodeur)
* PR #6300: Un-skip tests that were skipped because of #4026. (Owen Anderson)
* PR #6307: Remove restrictions on SVML version due to bug in LLVM SVML CC
(Stuart Archibald)
* PR #6316: Make IR inliner tests not self mutating. (Stuart Archibald)
* PR #6318: PR #5892 continued (Todd A. Anderson, via Stuart Archibald)
* PR #6319: Permit switching off boundschecking when debug is on. (Stuart
Archibald)
* PR #6324: PR 6208 continued (Ivan Butygin and Stuart Archibald)
* PR #6337: Implements ``key`` on ``types.TypeRef`` (Andreas Sodeur)
* PR #6354: Bump llvmlite to 0.35. series. (Stuart Archibald)
* PR #6357: Fix enumerate invalid decref (Siu Kwan Lam)
* PR #6359: Fixes typed list indexing on 32bit (Stuart Archibald)
* PR #6378: Fix incorrect CPU override in vectorization test. (Stuart Archibald)
* PR #6379: Use O0 to enable inline and not affect loop-vectorization by later
O3... (Siu Kwan Lam)
* PR #6384: Fix failing tests to match on platform invariant int spelling.
(Stuart Archibald)
* PR #6390: Updates inspect_cfg (Stuart Archibald)
* PR #6396: Remove hard dependency on tbb package. (Stuart Archibald)
* PR #6408: Don't do array analysis for tuples that contain arrays. (Todd A.
Anderson)
* PR #6441: Fix ASCII flag in Unicode slicing (0.52.0rc2 regression) (Ehsan
Totoni)
* PR #6442: Fix array analysis regression in 0.52 RC2 for tuple of 1D arrays
(Ehsan Totoni)
* PR #6446: Fix #6444: pruner issues with reference stealing functions (Siu
Kwan Lam)
* PR #6450: Fix asfarray kwarg default handling. (Stuart Archibald)
* PR #6486: fix abstract base class import (Valentin Haenel)
* PR #6487: Restrict maximum version of python (Siu Kwan Lam)
* PR #6527: setup.py: fix py version guard (Chris Barnes)
CUDA Enhancements/Fixes:
* PR #5465: Remove macro expansion and replace uses with FE typing + BE lowering
(Graham Markall)
* PR #5741: CUDA: Add two-argument implementation of round() (Graham Markall)
* PR #5900: Enable CUDA Unified Memory (Max Katz)
* PR #6042: CUDA: Lower launch overhead by launching kernel directly (Graham
Markall)
* PR #6064: Lower math.frexp and math.ldexp in numba.cuda (Zhihao Yuan)
* PR #6066: Lower math.isfinite in numba.cuda (Zhihao Yuan)
* PR #6092: CUDA: Add mapped_array_like and pinned_array_like (Graham Markall)
* PR #6127: Fix race in reduction kernels on Volta, require CUDA 9, add syncwarp
with default mask (Graham Markall)
* PR #6129: Extend Cudasim to support most of the memory functionality. (Mike
Williams)
* PR #6150: CUDA: Turn on flake8 for cudadrv and fix errors (Graham Markall)
* PR #6152: CUDA: Provide wrappers for all libdevice functions, and fix typing
of math function (#4618) (Graham Markall)
* PR #6227: Raise exception when no supported architectures are found (Jacob
Tomlinson)
* PR #6244: CUDA Docs: Make workflow using simulator more explicit (Graham
Markall)
* PR #6248: Add support for CUDA atomic subtract operations (Michael Collison)
* PR #6289: Refactor atomic test cases to reduce code duplication (Michael
Collison)
* PR #6290: CUDA: Add support for complex power (Graham Markall)
* PR #6296: Fix flake8 violations in numba.cuda module (Graham Markall)
* PR #6297: Fix flake8 violations in numba.cuda.tests.cudapy module (Graham
Markall)
* PR #6298: Fix flake8 violations in numba.cuda.tests.cudadrv (Graham Markall)
* PR #6299: Fix flake8 violations in numba.cuda.simulator (Graham Markall)
* PR #6306: Fix flake8 in cuda atomic test from merge. (Stuart Archibald)
* PR #6325: Refactor code for atomic operations (Michael Collison)
* PR #6329: Flake8 fix for a CUDA test (Stuart Archibald)
* PR #6331: Explicitly state that NUMBA_ENABLE_CUDASIM needs to be set before
import (Graham Markall)
* PR #6340: CUDA: Fix #6339, performance regression launching specialized
kernels (Graham Markall)
* PR #6380: Only test managed allocations on Linux (Graham Markall)
Documentation Updates:
* PR #6090: doc: Add doc on direct creation of Numba typed-list (``@rht``)
* PR #6110: Update CONTRIBUTING.md (Stuart Archibald)
* PR #6128: CUDA Docs: Restore Dispatcher.forall() docs (Graham Markall)
* PR #6277: fix: cross2d wrong doc. reference (issue #6276) (``@jeertmans``)
* PR #6282: Remove docs on Python 2(.7) EOL. (Stuart Archibald)
* PR #6283: Add note on how public CI is impl and what users can do to help.
(Stuart Archibald)
* PR #6292: Document support for structured array attribute access
(Graham Markall)
* PR #6310: Declare unofficial \*BSD support (Stuart Archibald)
* PR #6342: Fix docs on literally usage. (Stuart Archibald)
* PR #6348: doc: fix typo in jitclass.rst ("initilising" -> "initialising")
(``@muxator``)
* PR #6362: Move llvmlite support in README to 0.35 (Stuart Archibald)
* PR #6363: Note that reference counted types are not permitted in set().
(Stuart Archibald)
* PR #6364: Move deprecation schedules for 0.52 (Stuart Archibald)
CI/Infrastructure Updates:
* PR #6252: Show channel URLs (Siu Kwan Lam)
* PR #6338: Direct user questions to Discourse instead of the Google Group.
(Stan Seibert)
* PR #6474: Add skip on PPC64LE for tests causing SIGABRT in LLVM. (Stuart
Archibald)
Authors:
* Alexey Kozlov
* Amos Bird
* Andreas Sodeur
* Arthur Peters
* Chris Barnes
* Ehsan Totoni (core dev)
* Eric Wieser
* Ethan Pronovost
* Graham Markall
* Guilherme Leobas
* Isaac Virshup
* Ivan Butygin
* Jacob Tomlinson
* Luiz Almeida
* László Károlyi
* Lucio Fernandez-Arjona
* Max Katz
* Michael Collison
* Mike Williams
* Owen Anderson
* Radu Popovici
* Ray Donnelly
* Rishabh Varshney
* Siu Kwan Lam (core dev)
* Stan Seibert (core dev)
* Stuart Archibald (core dev)
* Todd A. Anderson (core dev)
* Valentin Haenel (core dev)
* Zhihao Yuan
* ``@jeertmans``
* ``@mugoh``
* ``@muxator``
* ``@rht``
Version 0.51.2 (September 2, 2020)
----------------------------------
This is a bugfix release for 0.51.1. It fixes a critical performance bug in the
CFG back edge computation algorithm that leads to exponential time complexity
arising in compilation for use cases with certain pathological properties.
* PR #6195: PR 6187 Continue. Don't visit already checked successors
Authors:
* Graham Markall
* Siu Kwan Lam (core dev)
Version 0.51.1 (August 26, 2020)
--------------------------------
This is a bugfix release for 0.51.0, it fixes a critical bug in caching, another
critical bug in the CUDA target initialisation sequence and also fixes some
compile time performance regressions:
* PR #6141: Fix #6130 objmode cache segfault
* PR #6146: Fix compilation slowdown due to controlflow analysis
* PR #6147: CUDA: Don't make a runtime call on import
* PR #6153: Fix for #6151. Make UnicodeCharSeq into str for comparison.
* PR #6168: Fix Issue #6167: Failure in test_cuda_submodules
Authors:
* Graham Markall
* Siu Kwan Lam (core dev)
* Stuart Archibald (core dev)
Version 0.51.0 (August 12, 2020)
--------------------------------
This release continues to add new features to Numba and also contains a
significant number of bug fixes and stability improvements.
Highlights of core feature changes include:
* The compilation chain is now based on LLVM 10 (Valentin Haenel).
* Numba has internally switched to prefer non-literal types over literal ones so
as to reduce function over-specialisation, this with view of speeding up
compile times (Siu Kwan Lam).
* On the CUDA target: Support for CUDA Toolkit 11, Ampere, and Compute
Capability 8.0; Printing of ``SASS`` code for kernels; Callbacks to Python
functions can be inserted into CUDA streams, and streams are async awaitable;
Atomic ``nanmin`` and ``nanmax`` functions are added; Fixes for various
miscompilations and segfaults. (mostly Graham Markall; call backs on
streams by Peter Würtz).
Intel also kindly sponsored research and development that lead to some exciting
new features:
* Support for heterogeneous immutable lists and heterogeneous immutable string
key dictionaries. Also optional initial/construction value capturing for all
lists and dictionaries containing literal values (Stuart Archibald).
* A new pass-by-reference mutable structure extension type ``StructRef`` (Siu
Kwan Lam).
* Object mode blocks are now cacheable, with the side effect of numerous bug
fixes and performance improvements in caching. This also permits caching of
functions defined in closures (Siu Kwan Lam).
Deprecations to note:
To align with other targets, the ``argtypes`` and ``restypes`` kwargs to
``@cuda.jit`` are now deprecated, the ``bind`` kwarg is also deprecated.
Further the ``target`` kwarg to the ``numba.jit`` decorator family is
deprecated.
General Enhancements:
* PR #5463: Add str(int) impl
* PR #5526: Impl. np.asarray(literal)
* PR #5619: Add support for multi-output ufuncs
* PR #5711: Division with timedelta input
* PR #5763: Support minlength argument to np.bincount
* PR #5779: Return zero array from np.dot when the arguments are empty.
* PR #5796: Add implementation for np.positive
* PR #5849: Setitem for records when index is StringLiteral, including literal
unroll
* PR #5856: Add support for conversion of inplace_binop to parfor.
* PR #5893: Allocate 1D iteration space one at a time for more even
distribution.
* PR #5922: Reduce objmode and unpickling overhead
* PR #5944: re-enable OpenMP in wheels
* PR #5946: Implement literal dictionaries and lists.
* PR #5956: Update numba_sysinfo.py
* PR #5978: Add structref as a mutable struct that is pass-by-ref
* PR #5980: Deprecate target kwarg for numba.jit.
* PR #6058: Add prefer_literal option to overload API
Fixes:
* PR #5674: Fix #3955. Allow `with objmode` to be cached
* PR #5724: Initialize process lock lazily to prevent multiprocessing issue
* PR #5783: Make np.divide and np.remainder code more similar
* PR #5808: Fix 5665 Block jit(nopython=True, forceobj=True) and suppress
njit(forceobj=True)
* PR #5834: Fix the is operator on Ellipsis
* PR #5838: Ensure ``Dispatcher.__eq__`` always returns a bool
* PR #5841: cleanup: Use PythonAPI.bool_from_bool in more places
* PR #5862: Do not leak loop iteration variables into the numba.np.npyimpl
namespace
* PR #5869: Update repomap
* PR #5879: Fix erroneous input mutation in linalg routines
* PR #5882: Type check function in jit decorator
* PR #5925: Use np.inf and -np.inf for max and min float values respectively.
* PR #5935: Fix default arguments with multiprocessing
* PR #5952: Fix "Internal error ... local variable 'errstr' referenced before
assignment during BoundFunction(...)"
* PR #5962: Fix SVML tests with LLVM 10 and AVX512
* PR #5972: fix flake8 for numba/runtests.py
* PR #5995: Update setup.py with new llvmlite versions
* PR #5996: Set lower bound for llvmlite to 0.33
* PR #6004: Fix problem in branch pruning with LiteralStrKeyDict
* PR #6017: Fixing up numba_do_raise
* PR #6028: Fix #6023
* PR #6031: Continue 5821
* PR #6035: Fix overspecialize of literal
* PR #6046: Fixes statement reordering bug in maximize fusion step.
* PR #6056: Fix issue on invalid inlining of non-empty build_list by
inline_arraycall
* PR #6057: fix aarch64/python_3.8 failure on master
* PR #6070: Fix overspecialized containers
* PR #6071: Remove f-strings in setup.py
* PR #6072: Fix for #6005
* PR #6073: Fixes invalid C prototype in helper function.
* PR #6078: Duplicate NumPy's PyArray_DescrCheck macro
* PR #6081: Fix issue with cross drive use and relpath.
* PR #6083: Fix bug in initial value unify.
* PR #6087: remove invalid sanity check from randrange tests
* PR #6089: Fix invalid reference to TypingError
* PR #6097: Add function code and closure bytes into cache key
* PR #6099: Restrict upper limit of TBB version due to ABI changes.
* PR #6101: Restrict lower limit of icc_rt version due to assumed SVML bug.
* PR #6107: Fix and test #6095
* PR #6109: Fixes an issue reported in #6094
* PR #6111: Decouple LiteralList and LiteralStrKeyDict from tuple
* PR #6116: Fix #6102. Problem with non-unique label.
CUDA Enhancements/Fixes:
* PR #5359: Remove special-casing of 0d arrays
* PR #5709: CUDA: Refactoring of cuda.jit and kernel / dispatcher abstractions
* PR #5732: CUDA Docs: document ``forall`` method of kernels
* PR #5745: CUDA stream callbacks and async awaitable streams
* PR #5761: Add implmentation for int types for isnan and isinf for CUDA
* PR #5819: Add support for CUDA 11 and Ampere / CC 8.0
* PR #5826: CUDA: Add function to get SASS for kernels
* PR #5846: CUDA: Allow disabling NVVM optimizations, and fix debug issues
* PR #5851: CUDA EMM enhancements - add default get_ipc_handle implementation,
skip a test conditionally
* PR #5852: CUDA: Fix ``cuda.test()``
* PR #5857: CUDA docs: Add notes on resetting the EMM plugin
* PR #5859: CUDA: Fix reduce docs and style improvements
* PR #6016: Fixes change of list spelling in a cuda test.
* PR #6020: CUDA: Fix #5820, adding atomic nanmin / nanmax
* PR #6030: CUDA: Don't optimize IR before sending it to NVVM
* PR #6052: Fix dtype for atomic_add_double testsuite
* PR #6080: CUDA: Prevent auto-upgrade of atomic intrinsics
* PR #6123: Fix #6121
Documentation Updates:
* PR #5782: Host docs on Read the Docs
* PR #5830: doc: Mention that caching uses pickle
* PR #5963: Fix broken link to numpy ufunc signature docs
* PR #5975: restructure communication section
* PR #5981: Document bounds-checking behavior in python deviations page
* PR #5993: Docs for structref
* PR #6008: Small fix so bullet points are rendered by sphinx
* PR #6013: emphasize cuda kernel functions are asynchronous
* PR #6036: Update deprecation doc from numba.errors to numba.core.errors
* PR #6062: Change references to numba.pydata.org to https
CI updates:
* PR #5850: Updates the "New Issue" behaviour to better redirect users.
* PR #5940: Add discourse badge
* PR #5960: Setting mypy on CI
Enhancements from user contributed PRs (with thanks!):
* Aisha Tammy added the ability to switch off TBB support at compile time in
#5821 (continued in #6031 by Stuart Archibald).
* Alexander Stiebing fixed a reference before assignment bug in #5952.
* Alexey Kozlov fixed a bug in tuple getitem for literals in #6028.
* Andrew Eckart updated the repomap in #5869, added support for Read the Docs
in #5782, fixed a bug in the ``np.dot`` implementation to correctly handle
empty arrays in #5779 and added support for ``minlength`` to ``np.bincount``
in #5763.
* ``@bitsisbits`` updated ``numba_sysinfo.py`` to handle HSA agents correctly in
#5956.
* Daichi Suzuo Fixed a bug in the threading backend initialisation sequence such
that it is now correctly a lazy lock in #5724.
* Eric Wieser contributed a number of patches, particularly in enhancing and
improving the ``ufunc`` capabilities:
* #5359: Remove special-casing of 0d arrays
* #5834: Fix the is operator on Ellipsis
* #5619: Add support for multi-output ufuncs
* #5841: cleanup: Use PythonAPI.bool_from_bool in more places
* #5862: Do not leak loop iteration variables into the numba.np.npyimpl
namespace
* #5838: Ensure ``Dispatcher.__eq__`` always returns a bool
* #5830: doc: Mention that caching uses pickle
* #5783: Make np.divide and np.remainder code more similar
* Ethan Pronovost added a guard to prevent the common mistake of applying a jit
decorator to the same function twice in #5881.
* Graham Markall contributed many patches to the CUDA target, as follows:
* #6052: Fix dtype for atomic_add_double tests
* #6030: CUDA: Don't optimize IR before sending it to NVVM
* #5846: CUDA: Allow disabling NVVM optimizations, and fix debug issues
* #5826: CUDA: Add function to get SASS for kernels
* #5851: CUDA EMM enhancements - add default get_ipc_handle implementation,
skip a test conditionally
* #5709: CUDA: Refactoring of cuda.jit and kernel / dispatcher abstractions
* #5819: Add support for CUDA 11 and Ampere / CC 8.0
* #6020: CUDA: Fix #5820, adding atomic nanmin / nanmax
* #5857: CUDA docs: Add notes on resetting the EMM plugin
* #5859: CUDA: Fix reduce docs and style improvements
* #5852: CUDA: Fix ``cuda.test()``
* #5732: CUDA Docs: document ``forall`` method of kernels
* Guilherme Leobas added support for ``str(int)`` in #5463 and
``np.asarray(literal value)``` in #5526.
* Hameer Abbasi deprecated the ``target`` kwarg for ``numba.jit`` in #5980.
* Hannes Pahl added a badge to the Numba github page linking to the new
discourse forum in #5940 and also fixed a bug that permitted illegal
combinations of flags to be passed into ``@jit`` in #5808.
* Kayran Schmidt emphasized that CUDA kernel functions are asynchronous in the
documentation in #6013.
* Leonardo Uieda fixed a broken link to the NumPy ufunc signature docs in #5963.
* Lucio Fernandez-Arjona added mypy to CI and started adding type annotations to
the code base in #5960, also fixed a (de)serialization problem on the
dispatcher in #5935, improved the undefined variable error message in #5876,
added support for division with timedelta input in #5711 and implemented
``setitem`` for records when the index is a ``StringLiteral`` in #5849.
* Ludovic Tiako documented Numba's bounds-checking behavior in the python
deviations page in #5981.
* Matt Roeschke changed all ``http`` references ``https`` in #6062.
* ``@niteya-shah`` implemented ``isnan`` and ``isinf`` for integer types on the
CUDA target in #5761 and implemented ``np.positive`` in #5796.
* Peter Würtz added CUDA stream callbacks and async awaitable streams in #5745.
* ``@rht`` fixed an invalid import referred to in the deprecation documentation
in #6036.
* Sergey Pokhodenko updated the SVML tests for LLVM 10 in #5962.
* Shyam Saladi fixed a Sphinx rendering bug in #6008.
Authors:
* Aisha Tammy
* Alexander Stiebing
* Alexey Kozlov
* Andrew Eckart
* ``@bitsisbits``
* Daichi Suzuo
* Eric Wieser
* Ethan Pronovost
* Graham Markall
* Guilherme Leobas
* Hameer Abbasi
* Hannes Pahl
* Kayran Schmidt
* Kozlov, Alexey
* Leonardo Uieda
* Lucio Fernandez-Arjona
* Ludovic Tiako
* Matt Roeschke
* ``@niteya-shah``
* Peter Würtz
* Sergey Pokhodenko
* Shyam Saladi
* ``@rht``
* Siu Kwan Lam (core dev)
* Stuart Archibald (core dev)
* Todd A. Anderson (core dev)
* Valentin Haenel (core dev)
Version 0.50.1 (Jun 24, 2020)
-----------------------------
This is a bugfix release for 0.50.0, it fixes a critical bug in error reporting
and a number of other smaller issues:
* PR #5861: Added except for possible Windows get_terminal_size exception
* PR #5876: Improve undefined variable error message
* PR #5884: Update the deprecation notices for 0.50.1
* PR #5889: Fixes literally not forcing re-dispatch for inline='always'
* PR #5912: Fix bad attr access on certain typing templates breaking exceptions.
* PR #5918: Fix cuda test due to #5876
Authors:
* ``@pepping_dore``
* Lucio Fernandez-Arjona
* Siu Kwan Lam (core dev)
* Stuart Archibald (core dev)
Version 0.50.0 (Jun 10, 2020)
-----------------------------
This is a more usual release in comparison to the others that have been made in
the last six months. It comprises the result of a number of maintenance tasks
along with some new features and a lot of bug fixes.
Highlights of core feature changes include:
* The compilation chain is now based on LLVM 9.
* The error handling and reporting system has been improved to reduce the size
of error messages, and also improve quality and specificity.
* The CUDA target has more stream constructors available and a new function for
compiling to PTX without linking and loading the code to a device. Further,
the macro-based system for describing CUDA threads and blocks has been
replaced with standard typing and lowering implementations, for improved
debugging and extensibility.
IMPORTANT: The backwards compatibility shim, that was present in 0.49.x to
accommodate the refactoring of Numba's internals, has been removed. If a module
is imported from a moved location an ``ImportError`` will occur.
General Enhancements:
* PR #5060: Enables np.sum for timedelta64
* PR #5225: Adjust interpreter to make conditionals predicates via bool() call.
* PR #5506: Jitclass static methods
* PR #5580: Revert shim
* PR #5591: Fix #5525 Add figure for total memory to ``numba -s`` output.
* PR #5616: Simplify the ufunc kernel registration
* PR #5617: Remove /examples from the Numba repo.
* PR #5673: Fix inliners to run all passes on IR and clean up correctly.
* PR #5700: Make it easier to understand type inference: add SSA dump, use for
``DEBUG_TYPEINFER``
* PR #5702: Fixes for LLVM 9
* PR #5722: Improve error messages.
* PR #5758: Support NumPy 1.18
Fixes:
* PR #5390: add error handling for lookup_module
* PR #5464: Jitclass drops annotations to avoid error
* PR #5478: Fix #5471. Issue with omitted type not recognized as literal value.
* PR #5517: Fix numba.typed.List extend for singleton and empty iterable
* PR #5549: Check type getitem
* PR #5568: Add skip to entrypoint test on windows
* PR #5581: Revert #5568
* PR #5602: Fix segfault caused by pop from numba.typed.List
* PR #5645: Fix SSA redundant CFG computation
* PR #5686: Fix issue with SSA not minimal
* PR #5689: Fix bug in unified_function_type (issue 5685)
* PR #5694: Skip part of slice array analysis if any part is not analyzable.
* PR #5697: Fix usedef issue with parfor loopnest variables.
* PR #5705: A fix for cases where SSA looks like a reduction variable.
* PR #5714: Fix bug in test
* PR #5717: Initialise Numba extensions ahead of any compilation starting.
* PR #5721: Fix array iterator layout.
* PR #5738: Unbreak master on buildfarm
* PR #5757: Force LLVM to use ZMM registers for vectorization.
* PR #5764: fix flake8 errors
* PR #5768: Interval example: fix import
* PR #5781: Moving record array examples to a test module
* PR #5791: Fix up no cgroups problem
* PR #5795: Restore refct removal pass and make it strict
* PR #5807: Skip failing test on POWER8 due to PPC CTR Loop problem.
* PR #5812: Fix side issue from #5792, @overload inliner cached IR being
mutated.
* PR #5815: Pin llvmlite to 0.33
* PR #5833: Fixes the source location appearing incorrectly in error messages.
CUDA Enhancements/Fixes:
* PR #5347: CUDA: Provide more stream constructors
* PR #5388: CUDA: Fix OOB write in test_round{f4,f8}
* PR #5437: Fix #5429: Exception using ``.get_ipc_handle(...)`` on array from
``as_cuda_array(...)``
* PR #5481: CUDA: Replace macros with typing and lowering implementations
* PR #5556: CUDA: Make atomic semantics match Python / NumPy, and fix #5458
* PR #5558: CUDA: Only release primary ctx if retained
* PR #5561: CUDA: Add function for compiling to PTX (+ other small fixes)
* PR #5573: CUDA: Skip tests under cuda-memcheck that hang it
* PR #5578: Implement math.modf for CUDA target
* PR #5704: CUDA Eager compilation: Fix max_registers kwarg
* PR #5718: CUDA lib path tests: unset CUDA_PATH when CUDA_HOME unset
* PR #5800: Fix LLVM 9 IR for NVVM
* PR #5803: CUDA Update expected error messages to fix #5797
Documentation Updates:
* PR #5546: DOC: Add documentation about cost model to inlining notes.
* PR #5653: Update doc with respect to try-finally case
Enhancements from user contributed PRs (with thanks!):
* Elias Kuthe fixed in issue with imports in the Interval example in #5768
* Eric Wieser Simplified the ufunc kernel registration mechanism in #5616
* Ethan Pronovost patched a problem with ``__annotations__`` in ``jitclass`` in
#5464, fixed a bug that lead to infinite loops in Numba's ``Type.__getitem__``
in #5549, fixed a bug in ``np.arange`` testing in #5714 and added support for
``@staticmethod`` to ``jitclass`` in #5506.
* Gabriele Gemmi implemented ``math.modf`` for the CUDA target in #5578
* Graham Markall contributed many patches, largely to the CUDA target, as
follows:
* #5347: CUDA: Provide more stream constructors
* #5388: CUDA: Fix OOB write in test_round{f4,f8}
* #5437: Fix #5429: Exception using ``.get_ipc_handle(...)`` on array from
``as_cuda_array(...)``
* #5481: CUDA: Replace macros with typing and lowering implementations
* #5556: CUDA: Make atomic semantics match Python / NumPy, and fix #5458
* #5558: CUDA: Only release primary ctx if retained
* #5561: CUDA: Add function for compiling to PTX (+ other small fixes)
* #5573: CUDA: Skip tests under cuda-memcheck that hang it
* #5648: Unset the memory manager after EMM Plugin tests
* #5700: Make it easier to understand type inference: add SSA dump, use for
``DEBUG_TYPEINFER``
* #5704: CUDA Eager compilation: Fix max_registers kwarg
* #5718: CUDA lib path tests: unset CUDA_PATH when CUDA_HOME unset
* #5800: Fix LLVM 9 IR for NVVM
* #5803: CUDA Update expected error messages to fix #5797
* Guilherme Leobas updated the documentation surrounding try-finally in #5653
* Hameer Abbasi added documentation about the cost model to the notes on
inlining in #5546
* Jacques Gaudin rewrote ``numba -s`` to produce and consume a dictionary of
output about the current system in #5591
* James Bourbeau Updated min/argmin and max/argmax to handle non-leading nans
(via #5758)
* Lucio Fernandez-Arjona moved the record array examples to a test module in
#5781 and added ``np.timedelta64`` handling to ``np.sum`` in #5060
* Pearu Peterson Fixed a bug in unified_function_type in #5689
* Sergey Pokhodenko fixed an issue impacting LLVM 10 regarding vectorization
widths on Intel SkyLake processors in #5757
* Shan Sikdar added error handling for ``lookup_module`` in #5390
* @toddrme2178 add CI testing for NumPy 1.18 (via #5758)
Authors:
* Elias Kuthe
* Eric Wieser
* Ethan Pronovost
* Gabriele Gemmi
* Graham Markall
* Guilherme Leobas
* Hameer Abbasi
* Jacques Gaudin
* James Bourbeau
* Lucio Fernandez-Arjona
* Pearu Peterson
* Sergey Pokhodenko
* Shan Sikdar
* Siu Kwan Lam (core dev)
* Stuart Archibald (core dev)
* Todd A. Anderson (core dev)
* ``@toddrme2178``
* Valentin Haenel (core dev)
Version 0.49.1 (May 7, 2020)