Loop-oriented neural network compilation

A method of accelerating an execution of a neural network, the method comprising:

• receiving, by a compiler, a description of the neural network;
• and
• performing, by the compiler, operations during compilation comprising:
• identifying a plurality of operators based on the description of the neural network, wherein the plurality of operators includes a first operator and a second operator;
• generating a plurality of symbolic models associated with the plurality of operators, wherein the plurality of symbolic models includes a first symbolic model associated with the first operator and a second symbolic model associated with the second operator, and wherein generating the plurality of symbolic models includes, for each symbolic model of the plurality of symbolic models:
• identifying a nested loop associated with an operator of the plurality of operators, wherein the nested loop includes multiple iterative loops;
• defining a loop order for the symbolic model based on the nested loop;
• and
• defining a set of data dependencies for the symbolic model based on the nested loop;
• extracting a set of inter-operator dependencies based on the description of the neural network, wherein the set of inter-operator dependencies includes an indication that the first operator is performed prior to the second operator;
• analyzing the plurality of symbolic models and the set of inter-operator dependencies to identify combinable pairs of nested loops;
• determining that a first nested loop associated with the first operator is combinable with a second nested loop associated with the second operator based on the first symbolic model, the second symbolic model, and the set of inter-operator dependencies, wherein the first symbolic model includes a first loop order and a first set of data dependencies, and wherein the second symbolic model includes a second loop order and a second set of data dependencies;
• combining the first nested loop with the second nested loop to form a combined nested loop;
• and
• generating machine instructions for the neural network based on the combined nested loop.