Method for optimizing the size of a data subset of a processing space for improved execution performance

A method for optimizing the parallel processing of data on a hardware platform, the hardware platform comprising:

• at least one computing unit comprising a plurality of processing units executing a plurality of work items in parallel in a processing space, the processing space being a space in which a same sequence of operations is carried out on data, the data being broken into data subsets, the processing space being broken down into processing subsets, each processing subset performing the same processing on a data subset;
• wherein the method comprises the following steps:
• obtaining a maximum number of work items executable in parallel by a computing unit of the hardware platform,
• determining sizes of data subsets of the processing space through a same sequence of operations to obtain determined sizes,
• optimizing the determined size of at least one data subset of the processing space, the optimization comprising a step for:
• selecting, for the at least one data subset, at least one positive integer number strictly greater than the determined sire of the selected data subset, the positive integer being a multiple of numbers smaller than the determined size of the selected data subset,
• generating, for the at least one data subset selected in the selection step, a computation subset having, as size, the positive integer selected from the at least one data subset and comprising all data of the at least one data subset, the computation subset replacing the at least one data subset selected in the processing space, to obtain a new processing space,
• the optimization being implemented so as to obtain an improved execution performance measurement for the new processing space relative to the execution performance measurement obtained for the processing space in which the selected data subset has the determined sire according to at least one chosen criterion, the measurement being associated with the execution of a program including program code instructions implementing a processing partition corresponding to the partitioning of the new processing space into processing subsets, and assigning at least one executable work item, executing said sequence of operations, to each data subset, the total number of executable items per data subset being less than or equal to said maximum number of work items.