Changelog

1.10.0

• Support numpy 2.x.

• Support pytest 8.

• Drop support for Python versions older than 3.8, and update the code to use modern syntax.

• Autoformat the code with ruff, and use ruff for linting.

1.9.0

• Make Fft work with CUDA 12.9. Note that the work_area slot may now be absent in some cases.

• Bring readthedocs configuration up to date.

1.8.1

• Use np.prod() instead of deprecated np.product().

• Improve the exception display when accessing an unknown slot.

1.8.0

• Add KATSDPSIGPROC_TUNE_DB environment variable to override the location of the tuning database.

• Remove use of deprecated numba.generated_jit() decorator.

1.7.0

• Add Operation.ensure_bound().

• Remove dependency on deprecated pkg_resources and use importlib.resources instead.

• Fix autotuner() decorator for autotune functions with no parameters.

• Update type annotations to allow os.PathLike in the extra_dirs argument to build().

1.6.1

• Fix a race condition that sometimes caused events passed to async_wait_for_events() to be garbage-collected in a worker thread, and thus leak if the CUDA context could not be made current.

1.6.0

• Add bind argument to IOSlotBase.allocate().

• Add IOSlotBase.allocate_host().

• Update tests to avoid triggering deprecation warnings.

1.5.0

• Add KATSDPSIGPROC_TUNE_MATCH environment variable to control use of partial matches from the autotuning database.

1.4.3

• Switching testing from nose to pytest.

• Some packaging modernisation.

• Fix some links to other packages in the docs.

• Fix the return annotation on katsdpsigproc.resource.wait_until().

• Fix a spurious warning about a future exception not being consumed if an exception is thrown in the scope of a ResourceAllocation context manager.

• Change return type annotations on the abstract base classes to work better with mypy.

1.4.2

• Change some return type annotations on abstract base classes to work around mypy limitations.

1.4.1

• Destroy cuFFT plan when FftTemplate is garbage collected. This lack was (for unknown reasons) causing segmentation faults when repeatedly creating and destroying contexts without cleaning up the plans.

• Fix a potential failure of FftTemplate on 32-bit systems due to an incorrect type signature.

• Fix documentation on readthedocs by updating sphinxcontrib-tikz.

1.4

• Add a module for FFTs (using cuFFT).

1.3

• Add a pytest plugin.

1.2

• Add a user guide.

• Update the dependencies.

• Fix the type annotations to work with numpy 1.20.

• Fix deprecated usage of np.bool.

• Fix handling of Context.compile when no extra flags are passed.

• Fix enqueue_zero_buffer being run on the default CUDA stream.

• Fix sequencing of SVMArray get and set operations with other commands in the provided command queue.

1.1

• Use BLOCKING_SYNC for CUDA events, to avoid spinning on the CPU.

• Drop support for Python 3.5.

• Add type annotations for many classes.

• Make some classes abstract base classes (they were already semantically, but now it’s indicated using abc.

• Introduce some abstract base classes to underly the CUDA and OpenCL implementations.

1.0

This is the first versioned release.