Graham Markall
Graham Markall
It looks like Numba doesn't handle it very gracefully when `CUDA_HOME` is incorrectly set (see traceback in report for #9216). We can improve on the feedback to the user: *...
Creating a copy of a device array is not trivial, and should be. A couple of current workarounds are: ```python # Variant 1 @numba.vectorize(['float32(float32)'], target='cuda') def copy(x): return x #...
CUDA: allow the user to specify that a kernel will have low occupancy in the `@cuda.jit` decorator
Some kernels will need to be called with low occupancy in the context of a workflow, and there isn't always a sensible way to avoid this. It would be nice...
The following: ```python from numba import cuda @cuda.jit(device=True) def f(): raise ValueError('Error') @cuda.jit(debug=True, opt=False) def kernel(): f() kernel[1, 1]() ``` results in an error like: ``` Traceback (most recent call...
Merge of `main` into #986 - I am not permitted to push to that branch, so testing here.
The `_FunctionCompiler` implementation has various complexities that I think are no longer required now that `@generated_jit` is removed and other changes have been made since the original implementation. This PR...
Testing Numba 0.61.2 locally with NumPy 2.3 results in the following failures: ``` ====================================================================== ERROR: test_global_record (numba.tests.test_globals.TestGlobals.test_global_record) ---------------------------------------------------------------------- Traceback (most recent call last): File "/home/gmarkall/miniforge3/envs/numba-np-23/lib/python3.12/site-packages/numba/tests/test_globals.py", line 215, in test_global_record res...
I just spent an hour trying to understand my own explanation of this mechanism, along with spelunking the Numba history and various versions of the docs for `weakref.finalize`. Clearly my...