Version history¶

0.7.0 (in development)¶

Removed ArrayMetadataLike, use the strict ArrayMetadata instead. (PR_20)
Removed None variant for compiler_options and constant_arrays, use an empty list instead. (PR_20)
Module discovery is now fully automatic and does not need __process_modules__() to be defined in user classes. (PR_21)
Snippet.from_string() now takes an iterable of argument names. (PR_23)
Dynamic API imports (cuda_api, opencl_api, any_api) removed in favor of API static methods. (PR_25)

local_mem keyword parameter of kernel calls renamed to cu_dynamic_local_mem. (PR_17)
Renamed no_async keyword parameter to sync. (PR_18)

Made ArrayMetadata public. (PR_18)
metadata attribute to Array. (PR_18)
ArrayMetadata.buffer_size, span, min_offset, first_element_offset, and get_sub_region(); Array.minimum_subregion(). (PR_18)

Minimal Python version bumped to 3.10, and numpy to 2.0. (PR_12)
Refactored alignment logic to avoid inconsistent behavior and redundant calls. (PR_13)
Array does not support empty shape anymore. (PR_13)
dtypes.flatten_dtype returns FieldInfo objects. (PR_13)

Array.from_host() is now synchronous if the device is given as the first parameter. (PR_13)
Mako tracebacks are shown render error. (PR_13)

device_idx parameters are gone; now high level functions take BoundDevice or BoundMultiDevice arguments to indicate which devices to use; these objects include the corresponding contexts as well, so they don’t have to be passed separately.
Now API adapters only use device indices in a sense of “device index in the platform”; context adapters keep internal objects in dictionaries indexed by these indices, instead of in lists.
py.test plugin extracted into a separate package (pytest-grunnur).

Arrays don’t hold queues any more; they are passed explicitly to get() or set().
Prepared kernels don’t hold queues any more; they are passed on call.
Queue now stands for a single-device queue only; multi-device queues are extracted into MultiQueue.

Package build fixed.

Initial version