Circuit for executing stateful neural network
Some embodiments provide a neural network inference circuit for executing a neural network that includes multiple nodes that use state data from previous executions of the neural network. The neural network inference circuit includes (i) a set of computation circuits configured to execute the nodes of the neural network and (ii) a set of memories configured to implement a set of one or more registers to store, while executing the neural network for a particular input, state data generated during at least two executions of the network for previous inputs. The state data is for use by the set of computation circuits when executing a set of the nodes of the neural network for the particular input.
1 . A neural network inference circuit that executes a neural network comprising a plurality of layers, the neural network inference circuit comprising:
a set of computation circuits configured to execute the plurality of layers of the neural network; and
a set of non-transitory machine-readable memories configured to store state data generated by a first set of layers of the neural network while executing the neural network for a first input, the state data for retrieval and use by a second set of layers of the neural network while executing the neural network for each of a plurality of inputs after the first input,
wherein pointer data indicating that the state data was generated for the first input, is modified during subsequent executions of the neural network to indicate a number of executions that have occurred since the state data was generated.
2 . The neural network inference circuit of claim 1 , wherein a plurality of nodes are arranged in the first set of layers, and wherein the plurality of nodes partially overlaps with nodes arranged in the second set of layers.
3 . The neural network inference circuit of claim 1 , wherein a plurality of nodes are arranged in a first set of layers, and wherein the plurality of nodes overlap with nodes arranged in the second set of layers.
4 . The neural network inference circuit of claim 1 , wherein the first set of layers is completely separate from the second set of layers.
5 . The neural network inference circuit of claim 1 , wherein the set of non-transitory machine-readable memories are configured as a set of shift registers.
6 . The neural network inference circuit of claim 5 , wherein state data generated by the first set of layers during execution of the neural network for the first input is stored in a first set of memory locations of the set of shift registers.
7 . The neural network inference circuit of claim 6 , wherein the state data generated by the first set of layers during execution of the neural network for the first input is shifted to different sets of memory locations after execution of the neural network for each of a set of subsequent inputs.
8 . The neural network inference circuit of claim 6 , wherein the state data generated by the first set of layers during execution of the neural network for the first input is shifted to different sets of memory locations prior to execution of the neural network for each of a set of subsequent inputs.
9 . The neural network inference circuit of claim 6 , wherein the state data generated by the first set of layers during execution of the neural network for the first input is shifted to different sets of memory locations during execution of the neural network for each of a set of subsequent inputs.
10 . The neural network inference circuit of claim 6 , wherein:
the state data generated by the first set of layers during execution of the neural network for the first input is not moved from the first set of memory locations.
11 . The neural network inference circuit of claim 1 , wherein the set of non-transitory machine-readable memories are further configured to store (i) weight data for executing the neural network and (ii) intermediate activation values while executing the neural network.
12 . The neural network inference circuit of claim 1 , wherein:
the first input is a first frame of video from a video stream; and
the plurality of inputs after the first input are subsequent frames of video from the video stream.
13 . The neural network inference circuit of claim 1 further comprising a plurality of cores, wherein:
each core comprises (i) a subset of the set of computation circuits and (ii) a subset of the set of non-transitory machine-readable memories; and
each respective layer of the neural network is executed by the subset of the set of computation circuits belonging to a respective subset of the plurality of cores of neural network inference circuit.
14 . The neural network inference circuit of claim 13 , wherein the subset of the set of non-transitory machine-readable memories belonging to the respective subset of the plurality of cores that execute the second set of layers store the state data generated by the first set of layers.
15 . The neural network inference circuit of claim 14 , wherein the subset of the set of computation circuits belonging to the respective subset of the plurality of cores that execute the first set of layers generate the state data.
16 . The neural network inference circuit of claim 15 , wherein the respective subset of the plurality of cores that execute the first set of layers and the respective subset of the plurality of cores that execute the second set of layers are different groups of cores.
17 . The neural network inference circuit of claim 1 , wherein a particular layer of the first set of layers generates state data based on input from a plurality of previous layers of the neural network.
18 . The neural network inference circuit of claim 17 , wherein the plurality of previous layers that provide input to the particular layer comprises at least one layer of the second set of layers.
19 . The neural network inference circuit of claim 1 , wherein the set of non-transitory machine-readable memories store state data from a particular number of previous executions, and wherein the set of computation circuits use a subset of the state data for a particular execution of the neural network.
20 . A method for using a neural network inference circuit to execute a neural network comprising a plurality of layers, the method comprising:
executing, using a set of computation circuits, the plurality of layers of the neural network, wherein executing the plurality of layers of the neural network comprises:
generating, for a first input by a first set of layers of the neural network, state data;
storing the state data in a set of non-transitory machine-readable memories for retrieval and use by a second set of layers of the neural network while executing the neural network for each of a plurality of inputs after the first input,
generating pointer data that indicates that the state data was generated for the first input, and
modifying, during subsequent executions of the neural network, the pointer data to indicate a number of executions that have occurred since the state data was generated.