Intelligent processor, data processing method and storage medium
An intelligent processor includes a first operation unit and a second operation unit. The first operation unit acquires first input data corresponding to a first operator, divides the first input data into multiple sets of first sub input data, and operates the first operator to perform an operation on the first sub input data to obtain first sub output data corresponding to each set of first input data. The second operation unit uses the first sub output data as second input data of the second operator, and operates the second operator to perform an operation on the second input data to obtain second output data corresponding to each set of second input data. Thus, the second operator need not wait until the first operator finishes all operations of the first input data, allowing two adjacent operator to achieve partial parallel calculation and improving operation efficiency.
1. An intelligent processor, comprising:
a first operation unit, acquiring first input data corresponding to a first operator, dividing the first input data into sets of first sub input data, and operating the first operator to perform an operation on each of the sets of first sub input data to obtain sets of first sub output data corresponding, respectively, to the sets of first sub input data; and
a second operation unit, using each of the sets of first sub output data as sets of second input data of a second operator, and operating the second operator to perform an operation on each of the sets of second input data to obtain a set of second output data corresponding to each of the sets of second input data.
2. The intelligent processor according to claim 1 , wherein the second operator and the first operator belong to a same network model and are adjacent.
3. The intelligent processor according to claim 1 , wherein the second operation unit uses the sets of first sub output data as the sets of second input data each time when the first operation unit obtains the sets of first sub output data.
4. The intelligent processor according to claim 1 , wherein the second operation unit uses the sets of first sub output data as the sets of second input data each time when a data amount of the sets of first sub output data reaches a predetermined data amount.
5. The intelligent processor according to claim 1 , further comprising:
a cache;
wherein, the first operation unit further writes the sets of first sub output data to the cache, and the second operation unit further reads the sets of first sub output data from the cache.
6. The intelligent processor according to claim 5 , wherein the first operation unit further determines a target division policy corresponding to the first input data according to type information of the first operator; and divides the first input data into the sets of first sub input data according to the target division policy.
7. The intelligent processor according to claim 6 , wherein when the target division policy is a first division policy, a first target data size of divided plurality of sets of first sub input data is determined according to processing capacity of the first operation unit and a temporary storage space of the cache.
8. The intelligent processor according to claim 6 , wherein when the target division policy is a second division policy, a second target data size of a divided plurality of first sub input data is determined according to operation logic of the first operator.
9. The intelligent processor according to claim 1 , being included in a chip.
10. A data processing method, applied to an intelligent processor, the intelligent processor comprising a first operation unit and a second operation unit, the data processing method comprising:
the first operation unit acquiring first input data corresponding to a first operator, and dividing the first input data into sets of first sub input data;
the first operation unit operating the first operator to perform an operation on each of the sets of first sub input data to obtain sets of first sub output data corresponding, respectively, to the sets of first sub input data;
the second operation unit using each of the sets of first sub output data as sets of second input data of a second operator; and
the second operation unit operating the second operator to perform an operation on each of the sets of second input data to obtain a set of second output data corresponding to each of the sets of second input data.
11. The data processing method according to claim 10 , wherein the second operator and the first operator belong to a same network model and are adjacent.
12. The data processing method according to claim 10 , wherein the second operation unit uses the sets of first sub output data as the sets of second input data each time when the first operation unit obtains the sets of first sub output data.
13. The data processing method according to claim 10 , wherein the second operation unit uses the sets of first sub output data as the sets of second input data each time when a data amount of the sets of first sub output data reaches a predetermined data amount.