numba-dpex
numba-dpex copied to clipboard
IntelPython
Data Parallel Extension for Numba
- [x] operator.add() (+) - [ ] operator.and_() (&) - [x] operator.eq() (==) - [x] operator.floordiv() (//) - [x] operator.ge() (>=) - [x] operator.gt() (>) - [x] operator.iadd() (+=) -...
- parfor nodes in @jit functions - check with additional issues, like dpnp library calls
For some reason "level0" fails with "SyclProgramCompilationError" when debug is activated (`export NUMBA_DPPY_DEBUG=1`). PR with tests: https://github.com/IntelPython/numba-dppy/pull/297
I'm running on Gen9 and dppy 17.4 and have sporadic inaccurate results relative to numpy for the following code ``` import argparse import math import time import dpctl import numba...
Reproducer: ```python import dpnp as np import numba_dpex as ndpx from numba import literally @ndpx.kernel def lit(a, n): literally(n) local_a_0 = ndpx.private.array(n, dtype=a.dtype) lid = ndpx.get_local_id(0) if lid == 0:...
Numba keyword argument `fastmath` is not supported in `@dppy.kernel`. It can cause performance degradation in some cases. ```python import dpctl import numba_dppy import numpy as np import numpy.random as rnd...
I am not sure that examples should be a part of the package. We could install if but it is mostly for developers (not users) and easiest way to work...
- [x] Have you provided a meaningful PR description? - [x] Have you added a test, reproducer or referred to an issue with a reproducer? - [x] Have you tested...
If we look at `dpnp` and `numpy`, they support zero-length array creation (i.e. `shape` is 0) ```python >>> import numpy as np >>> a = np.empty(0) >>> a array([], dtype=float64)...
Below is a valid sycl code where we are allowed to multiply two sycl::range objects. ```c++ sycl::range dimBlock(1, 1, BLOCK_SIZE); sycl::range dimGrid(1, 1, blockCols); cgh.parallel_for(sycl::nd_range(dimGrid * dimBlock, dimBlock), [=](sycl::nd_item item_ct1)...