Minutes_2021_06_08
Valentin Haenel edited this page Jun 9, 2021
·
1 revision
Attendees: Siu, Graham, JimP, Stuart, ToddA, Val, Kalyan
NOTE: All communication is subject to the Numba Code of Conduct.
- Low level @overload update
- side effects of changes and bugs discovered
- target-context needs to be a singleton
- bugs with instruction-set changing config
- e.g. changes on SIMD-width
- other bugs related with stack data returned across function boundary (thus stack lifetime expired)
- side effects of changes and bugs discovered
- Last week mainline was broken too
- bug was caused by "flags" added to the dispatcher caching keys, but the equals methods were
id
based so caches were invalidated on every re-execution.
- bug was caused by "flags" added to the dispatcher caching keys, but the equals methods were
- JimP question on recursion:
- ndim=1, ndim>1, with ndim decreasing
-
map_tuple
https://github.com/numba/numba/pull/4813/files#diff-fea90b744bddcba2081c843dfb3fa990646969dd1996c3dfd989a083d778d7cdR557-R574 -
generated_jit
might be too old; might need to replacegenerated_jit
to useoverload
internally.
-
#7089 - CUDA: Overloading existing implementations of functions seems not to work
-
#7086 - Initialize NUMBA_DEFAULT_NUM_THREADS with a batch scheduler aware value
-
#7085 - Vectorize doesn't play well with
forceobj=True
- Use
numpy.vectorize
instead -
forceobj=True
should be disallowed with@vectorize
- Use
-
#7084 - Feature request: user-level statically defined tracepoints
- concern if static-trace works for JIT
- Graham testing out the idea but it is tricky due to the inlineasm trace-point requires switching ELF section
-
#7083 - CUDA_ERROR_OUT_OF_MEMORY with pinned_array
-
*** #7078 - Implement
numpy.broadcast_to
andnumpy.broadcast_arrays
- Could provide some pointers to existing code
-
*** #7077 - Support full broadcasting in
numpy.where
-
#7075 - Numba caches relative imports and not absolute
- #7079 - Significant performance regression on master
- #7076 - Numba throwing time error help needed urgently
- #7071 - Native installation of Numba using pip on M1 macOS
- #7088 - Initialize NUMBA_DEFAULT_NUM_THREADS with a batch scheduler aware value
- #7087 - Add note to docs about zero-initialization of variables.
- #7082 - Added Add/Sub between datetime64 array and timedelta64 scalar
- #7074 - Change compile unit language in DWARF to C++
- #7073 - [DO NOT MERGE] Temp/pr/7058
- #7072 - [DO NOT MERGE] Temp/enh/arrayallocapi