Deep learning optimization compiling system

The invention provides a deep learning optimization compiling system, which comprises an analyzer for segmenting a complete neural network calculation graph into a plurality of sub-graphs and distributing specific search time according to the importance degree of the sub-graphs; the optimizer is used for extracting high-level features in each sub-graph operator and carrying out coarse optimization on each sub-graph operator to form a basic code structure; the interpreter randomly initializes the size of a data block and compilation strategies of some for cycles, and obtains a complete representation of a computational graph at the same time; and the evaluator is used for evaluating the performance of the codes according to a cost model based on low energy consumption preference, obtaining a group of high-score implementations as selected, obtaining the actual performance through a compiler runtime module, and selecting the implementation with the optimal actual performance as output. The method has the follow