Minutes_2019_07_09
Siu Kwan Lam edited this page Jul 30, 2019
·
1 revision
Attendees: Todd, Stuart, Stan, Siu
- next release?! Bugs++
- What is in flight?
- Waiting for CI:
- #4291
- Final fixes to PRs:
- #4211
- #4151
- Waiting for Anaconda team to fix linux-32 bug (should be simple)
- Testing new ARM64 build environment
- POWER8 tolerance error
- Waiting for CI:
-
#4294 - issue with prange writing to preallocated list entries
- this may have worked before
- #4293 - prange leaky?
-
#4292 - Define a standard way to provide overloads for a module
- good feature request
- but is blocked by other features
- we want to be more explicit until we work out how the magic works
- **** #4290 - Investigate PGO in the LLVM backend
- how does this work in clang/llvm?
- need to do experiments to see how much benefit there is
- https://llvm.org/devmtg/2013-11/slides/Carruth-PGO.pdf
-
#4289 - JIT dask task fused blocks
- What is the hook to detect when this is happening?
-
#4288 - Write N~=20? examples of doing commonly done things with Numba (CPU and GPU)
- Yes
- Write a book (Numerical Recipes + GPU Gems 1-17)
- Add every example from this book to numba-examples
-
#4287 - Port the error rewrites and highlighting to GPU stack
- Good idea
- How much can we share between CPU and GPU
-
#4286 - Type highlighting in
inspect_types
- Yeah, that would be great.
- Should also type integer constants as smallest type
-
#4285 - Should Numba check consistency of cudatoolkit and CUDA driver?
- Sure, 🤷
-
#4284 - Add FAQ entry about scipy.special
- ufuncs are hard
-
#4283 - Implement a radix sort in cuda device function?
- need to figure out how to bitcast something on GPU
- seems to be a Numba bug
-
#4281 - cannot import name _typeconv
- need more info from user
-
#4280 - Getting ValueError: cannot compute fingerprint of empty list with a list of tuples
- should we handle empty list as a special type?
-
#4279 - armv7 automatic parallelization
- 32-bit is the blocker
- **** #4278 - New @jit option to get mixed nopython/forceobj (old default behavior)
-
#4277 - LoweringError when passing a tuple to numpy with under @njit(parallel=True) mode
- duplicate
-
#4276 - Recommended workaround to reflected list in a function using prange ?
- to close
-
#4274 - temp_directory cleanup and revisit its impl
- concurrent use. per-PID-unique and per-testcase-unique named subdirectory.
- manual cleanup.
- per-datetime directory
-
#4271 - Lowering error with Numpy.fromiter
- generator expression not supported.
- PR #4282 to clean up error
- #4269 - Unsupported functionality in the code Numba
- #4268 - coverage.py support for Numba
-
#4267 - prange hook for those of us who like status bar
- sub-issue: Objmode in prange
- #4266 - Add numba.cuda.jit examples
-
#4265 - Typing behavior with Integer and IntegerLiteral
- FIXME: zero unused out-of-bound location.
-
#4264 - np.sum(X, Y) causes LoweringError
- FIXME: stricten it
- **** #4261 - Numba
nogil
+ dask threading backend results in no speed up (computation is slower)- atomic refct is the bottleneck
- Needs discussion.
- May need better refct pruning (possibly as custom LLVM passes)
- TLS
- #4275 - error!
- #4272 - Call Cuda device Function in a module
- #4270 - Windows 10 numba overwriting input values
- #4282 - Detect and raise unsupported on generator expressions
- **** #4273 - pin to MKL BLAS for testing to get consistent results
- linux32: maybe conda bug
- openblas skylake bug
- **** #4263 - [WIP]lower() and islower() method support for unicode
- Val is helping this user.
- #4262 - Resolves #4251.
- Type List
- Caching improvements requirement gathering
- transitive dependency
- bug related to storing runtime-only info
- may punt on jitclasses
- may punt on issues of (g)ufunc
- future: pre-filled/package-able cache
- open questions:
- numba version pinning(?)
- regen cache for different numba versions
- regen at post-install
- open questions: