Franck Charras
Franck Charras
Large numpy arrays stored in big-endian format cannot be serialized, leading to errors with Parallel
`inplace = True` to byteswap on a memmap might be bad for performance however. Continuing investigating a better fix after discussing with @lesteve . Edit: actually it doesn't even work...
Large numpy arrays stored in big-endian format cannot be serialized, leading to errors with Parallel
So the bug was in fact likely introduced there https://github.com/joblib/joblib/commit/0fa2cb91f920a8c1733689e0af5da4e6879d689c The changes to endianess in this PR were aimed at fixing behavior of `joblib.dump` and `joblib.load` and it was overshighted...
I can't reproduce in my environment. Could you please provide the full traceback you have with your minimal reproducer and give more details on your environment (OS/kernel version/cpu architecture) ?...
Fixes https://github.com/joblib/joblib/issues/1559
@buddhist-uni thank you for the PR. Before merging it requires adding at least a test to `Memory` test suite at `joblib/test/test_memory.py`, also a usage example in `examples/memory_basic_usage.py` would be a...
Closed by https://github.com/joblib/joblib/pull/1549
Have you considered using native [multiprocessing.shared_memory](https://docs.python.org/3/library/multiprocessing.shared_memory.html) ? > I thought this would be a first class use case Joblib is mostly targeted for simpler use cases of embarassingly parallels jobs...
hi @diptorupd thank you for this explanation, maybe loop unrolling might have been applied by the `igc` then. Debug features such as the `"--debug-only=loop-vectorize"` option but for the `igc` would...
@diptorupd is there a simple way to hack into `numba_dpex` code so that it enables things such as `loop unrolling` currently, at least at the compiler level ? maybe by...
If I may suggest something there: a useful addition could be to expose a keyword that automatically adjust each dimension of the global size to the nearest greatest multiple of...