Siu Kwan Lam comments

Results 276 comments of


                                            Siu Kwan Lam

Use get_parfor_writes to detect illegal array access that prevents fusion.

towncrier is broken. don't worry about the error.

Use get_parfor_writes to detect illegal array access that prevents fusion.

(CI failed due to conda timeout. i've restarted failing builds)

Use `omp_set_max_active_levels()` instead of `omp_set_nested()`

Restarted buildfarm to refresh the test failures status: `numba_smoketest_cpu_yaml_160`

Add support for compiler-rt inclusion

Developer Note: This PR depends on merging LLVM14 changes.

Add support for compiler-rt inclusion

From OOB discussion, the OSX build failure is depending on https://github.com/conda/conda-build/issues/4787 to successfully build compiler-rt

Add support for compiler-rt inclusion

@esc, please consider this and related PR for inclusion of the upcoming llvmlite release.

guvectorize does not check contiguity, breaks indexing

Numba needs to guard against the use of Array type with layout other than `'A'` in ufunc and gufunc. NumPy is assuming strided array.

Integrate ZLUDA for AMD CUDA

I am unfamiliar with ZLUDA. Is there an example that does what @adworacz described? If that is the case, is this just a documentation change to tell users to try...

Some array methods much slower with non-C-order arrays

Developer Note: - `clip` is using `ndindex` which is assuming C ordered. It should use `nditer`. https://github.com/numba/numba/blob/fed996d3177b64fabf62a526fd329851819dc2da/numba/np/arrayobj.py#L2295 - `astype` is relying on `array.copy` which is also not considering memory layout....

Elapsed running time of parallel jit for the logistic regression documentation example is dozens of seconds

Could this be the overhead cost for parfor region? I'm seeing the speed difference just by using the numpy-parfor-pass e.g. ```python parallel = dict( comprehension = False, reduction = False,...