You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The numba-dpex test suite parameterized all unit tests on a set of filter strings. The result is a bulk of tests get skipped when unit tests are run a system that does not support all devices specified as filter strings.
For example, on a recent run of the public CI in github only ~600 tests ran out of ~2300. the rest got skipped.
The test suite needs to be redesigned:
Remove filter string parameter and instead always use the device returned by dpctl.select_default_device.
Control the default device using dpc++'s environment flags
Reorganize tests into directories to help run subset of test suite when needed. The test directory structure should be the same as the source directory layout.
Remove init.py inside test directories as we do not want tests to be importable modules.
Refactor all tests that use NumPy arrays as arguments to kernels (later for njit once njit supports compute follows data)
Fix deprecation and other warnings
Each test must have a docstring explaining what the test is doing.
Add more Numba IR and LLVM IR level tests. Currently, most tests are end-to-end tests. We need more IR level tests especially as we start focusing on IR transformations and optimizations.
The numba-dpex test suite parameterized all unit tests on a set of filter strings. The result is a bulk of tests get skipped when unit tests are run a system that does not support all devices specified as filter strings.
For example, on a recent run of the public CI in github only ~600 tests ran out of ~2300. the rest got skipped.
The test suite needs to be redesigned:
dpctl.select_default_device
.Remove init.py inside test directories as we do not want tests to be importable modules.The text was updated successfully, but these errors were encountered: