Thread safety

The package does not get extensively tested with multiple threads. Some parts of it are likely to be thread-safe, but at present there are no guarantees. If you have a multi-threaded use case in mind, you should first review the code to check that it will be safe, and then file an issue to ask the maintainer to add a guarantee that it will remain safe.

In general, it is not safe to perform two operations on an object in parallel if at least one of them modifies the object (data race). One exception is the autotuning cache: it is safe to run autotuning functions in parallel, although it is not recommended as the same autotuning might end up running multiple times, and running autotuning benchmarks in parallel will mess up their timing.

Due to limitations in OpenCL, it is not thread-safe to make multiple calls to the same kernel at the same time. Note that it is only the Python function call that needs to be serialized; the execution of the kernels on the device can proceed in parallel. It also only applies to a single Python object; two kernel objects created by two different calls to AbstractProgram.get_kernel() can be used in parallel. It’s thus recommended that this function is called again for each operation created from an operation template.

It is also not safe to make two calls to the same Operation at the same time as there are internal structures that may be modified by a call. Operations should be cheap to manufacture from templates, so each thread should have its own instance.