Minutes_2019_08_08
Siu Kwan Lam edited this page Aug 13, 2019
·
2 revisions
Attendees: Aaron, Stuart, Pearu, Todd, Stan, Siu, Val, James
-
Need to build a new llvmdev at some point soon... anyone want anything else?
- Polly needs adding (*)
- gcov needs adding
- Random stuff for grail quest needs adding
- patches not accepted into mainline
- OrcJit-only patches (* maybe moving to OrcJit)
- compiler-rt needs resolving
- adding in the ref count pruner at LLVM level needs adding
- that label length patch needs removing(?)
- LLVM 8.0.1 patch release (*)
footnote:
(*)
required -
JIT modules
- PR: https://github.com/numba/numba/pull/4331
- stuart to investigate importhook
-
Performance of calling methods on jitclass objects: https://github.com/numba/numba/issues/2166
- Request from Pandas developers (they see 15x slowdown compared to calling a function)
- Related to insufficient reference count pruning
-
Label length patch - stay or go?
- https://reviews.llvm.org/D41296#1607582
- hold until LLVM developers decide their direction
-
boundschecking: https://github.com/numba/numba/issues/730
- Siu: put some pointers to where to modify
- in numba/cgutils.py: get_item_pointer2
def get_item_pointer2(builder, data, shape, strides, layout, inds, wraparound=False): if wraparound:
- Siu: pointer to how to raise exception in builder land
- Siu: put some pointers to where to modify
-
Pearu wants to work on unicode in arrays #4018
-
#4406 -
numba.errors.NotDefinedError: Failed in object mode pipeline (step: analyzing bytecode) Variable '$6.2'
is not defined.- likely not a good application for Numba
-
#4405 - CUDA functions can be compiled but not called with Tuple and UniTuple arguments
- should be easy to fix?
-
#4404 - Support keyword arguments with guvectorize
- blocked because NumPy gufuncs objects don't take kwargs
-
#4403 - Cannot lower constant bytes.
- opened PR #4408 to fix
- **** #4402 - meta issue: fp16 support
- Is this also of interest for recent CPU targets?
-
#4401 - Named Tuple in objmode context
- Limitiation of object mode block that should be lifted
-
#4400 - numba.errors.UnsupportedError: Failed in nopython mode pipeline (step: analyzing bytecode) Use of unsupported opcode (SETUP_EXCEPT) found
- unsupported try except usage
- Could we have a better error message?
-
#4399 - Numpy corrcoef performance is better without Numba
-
np.cov
needs a review
-
-
#4398 - CUDA simulator mode does not recognize customized numpy dtype array
- open PR to fix (#4410)
-
#4396 - Turn off function name numbering for easier caching for CUDA
- instead, let exact function name be specified
- **** #4395 - fp16 GPU support
- See above
-
#4392 - Issue lowering
ir.StaticSetItem
- Can't figure out what the issue is from user-provided information
- **** aside: does ir.StaticSetItem even need to exist now that we have literal types?
- Static{G,S}etItem is ignoring
-
#4389 - Ahead-of-time compilation for cuda.jit
- Will be helped by being able to set explicit name for generated function
- Will need docs and/or better ABI for passing NumPy arrays
-
#4387 - Jitting linked list takes an unexpectedly long time
- How does this even compile??
- We should raise an error message
-
#4385 - Default argument gives "this should not have happened" error
- We should forbid usage of mutable types as defaults
- this specific case is failing due to reflected list
-
#4382 - Default argument with typed list?
- see above
-
#4394 - NotImplementedError:
(<class 'numba.ir.StaticSetItem'>, heatmap.1[(slice(None, None, None), slice(None, None, None), -1)] = $151.17)
- #4393 - report a bug
- #4391 - Hello. I am ru
- #4390 - CUDA kernel slowdown after several thousand kernel iterations
- #4384 - [duplicate bug, please delete] Default argument can cause segmentation fault
- #4381 - Numba 0.45.1 checklist
- #4410 - Fix #4111. cudasim mishandling recarray
- **** #4409 - Add exception chaining for better error context
- #4408 - Add lowering for constant bytes.
- **** #4407 - Restore the "free" conda channel for NumPy 1.10 support.
- Should discuss whether we should drop some NumPy's from testing?
- for reference, dask's min numpy is 1.13
- #4397 - Add note about np.sum
- **** #4388 - Update Python 2.7 EOL statement
- !!!! more discussion next week !!!!
- #4386 - Implement np.count_nonzero
- #4383 - Change log update for 0.45.1
- Finish pending rewrite passes
- Python 3.8 support
- Document best practices for constructing new compiler pipelines
- Define autodiscovery system for Numba extensions (like numba_scipy or HPAT) that don't need direct user import
- Allow opt-in dispatching of functions by literal value
- Making caching aware of transitive dependencies
- Define declarative typing system for @overload (to be used in future releases)
- Numba Runtime C API for extensions to register reference-counted memory with the runtime.
- Start using new CI system in parallel with existing one
- Priority bug fixes:
- Low performance of JIT method calls (requested by Pandas devs)
- Others TBD