Minutes_2021_06_22
Valentin Haenel edited this page Jun 22, 2021
·
1 revision
Attendees: Graham Markall, Guilherme, Hadia, Stuart, Valentin Haenel, Siu Kwan Lam, Brandon Willard, Jim Pivarski
NOTE: All communication is subject to the Numba Code of Conduct.
- 0.54.0 RC 1 by the end of June
- Must have:
- remaining target-overload PRs
- remove ROC
- dwarf improvement
- LLVM 11
- some CUDA PRs are very close
- possibly others
- Must have:
- faster compile PR vs debug info quality
-
#6720 alters stack slot assignment in LLVM; reduce
alloca
- Debug info relies on local in stack slot
- Elided stack slot leads to missing debug info for some locals
- Solutions:
- skip the stack slot eliding when debug=True.
- changes to a different LLVM debug metadata.
- Summary:
- try 1 as workaround for the upcoming release.
-
#6720 alters stack slot assignment in LLVM; reduce
- #7131 - Array.astype not supported with string literals
- #7130 - Performance gains halt after 8 (out of 16) cores
- ** #7124 - Float Literal Support
- Jim
- knows of useful cases in np.quantile and np.percentile.
- interesting float values also include things like
+inf
,-inf
,nan
,0
, etc...
- Stuart's concerns:
- overspecialization
- loop unswitch will push the
if
checks outside of loops. - large impact (compile time) of many code
- Jim
- #7123 - Add chrome-tracing support in event API
-
#7120 - segfault regression case (happens in 0.52.0, 0.53.0 & 0.53.1, works OK in 0.51.2)
- immutability needs to propagte into elements of a container
- #7117 - NUMBA_DISABLE_JIT causes an infinite loop in Structref instance creation.
- #7133 - Numbers don't add up?
- #7128 - The example for jitclass does not appear to work
- #7127 - Name error - numba doesn't find a function
- #7126 - Silent crash for parallel typed list appending
- #7134 - Move minimum LLVM version to 11.
- #7132 - gh #7131 Support for astype with literal strings
- #7129 - Add support for axis keyword argument to np.argmin()
- #7125 - Declare minimum sphinx version required for docs as 4.
- #7121 - Continue work on numpy subclasses
-
#7119 - Add support for
np.broadcast_to
- #7116 - Fix various issues with dwarf emission:
- #7115 - use np.asarray to convert arg to array