Fixed ----- - Fixed bug when dividing work over multiple devices. The batch size was provided incorrectly after changes in v0.9.5.
Other ----- - Renamed the lib_ clfunctions to _spf functions. This formalizes the specialization a little bit better. - Moved the Bracket method outside the Powell routine, making it a standalone routine now.
0.9.5
===================
Added ----- - Adds switch to the minimize function to disable local reduction if so desired. - Adds a utility function for easy splitting work over the work items.
Changed ------- - Let the powell algorithm use local reduction. - CL function will now process all provided work instances in one batch instead of splitting it in batches.
Other ----- - Moved repository to private github account.
0.9.4
===================
Added ----- - Added the method 'get_subset' to allow getting kernel data for a subset of the instances.
Changed ------- - Changed the Array kernel data type by removing the offset_str in favor of a simple boolean switch.
Other ----- - Bugfix in convert_data_to_dtype, in the case a vector type was already in the correct dtype. - Improved error reporting in the function evaluator.
0.9.3
===================
Added ----- - Adds include guards to the Struct kernel data object.
0.9.2
===================
Added ----- - Makes CL functions directly callable. Calls are forwarded to the evaluate method.
0.9.1
===================
Added ----- - Adds OpenCL approximations to the Gamma CDF and Gamma PPF functions.
Other ----- - Removed atomic functions. They were unsupport on some platforms. - Made the ML routines more robust for objective functions that return NaN.