Task processing method and device based on neural network
Provided is a task processing method based on a neural network, including: acquiring input data, wherein the input data is intended to trigger thread tasks, and is source input data or cache exchange data; generating processing result data by scheduling at least two corresponding module threads in parallel based on at least two triggered thread tasks to process the input data, wherein the at least two module threads respectively correspond to at least two network modules in the neural network; and outputting the processing result data to a cache as the cache exchange data of module threads except the at least two module threads, or outputting the processing result data as a processing result of the source input data.
1. A task processing method based on a neural network, applied in a neural network-related application of a device, comprising:
acquiring input data, wherein the input data is intended to trigger thread tasks, and is source input data or cache exchange data;
generating processing result data by scheduling at least two corresponding module threads in parallel based on at least two triggered thread tasks to process the input data, wherein the at least two module threads respectively correspond to at least two network modules in the neural network;
outputting the processing result data to a cache, wherein the processing result data is used as the cache exchange data of module threads except the at least two module threads, or outputting the processing result data, wherein the processing result data is used as a processing result of the source input data;
wherein the at least two module threads at least comprise a start module thread and an end module thread;
wherein generating the processing result data by scheduling at least two corresponding module threads in parallel based on at least two triggered thread tasks to process the input data comprises:
scheduling the start module thread based on the triggered thread task to process input data to the start module thread; and
scheduling the end module thread based on the triggered thread task to process input data to the end module thread; and
executing tasks of the scheduled start module thread on a first core of a multi-core processor and executing tasks of the scheduled end module thread on a different core of the multi-core processor.
2. The method according to claim 1 , wherein prior to acquiring the input data, the method further comprises:
acquiring the at least two network modules by dividing a plurality of network layers in the neural network.
3. The method according to claim 2 , wherein acquiring the at least two network modules by dividing the plurality of network layers in the neural network comprises:
determining a number of channels between adjacent network layers in the neural network;
identifying a former network layer in the adjacent network layers as an input layer for inputting a network module and identifying a latter network layer in the adjacent network layers as an output layer for outputting the network module in response to determining that one channel is present between the adjacent network layers; and
generating the at least two network modules based on the output layer for outputting the network module and the input layer for inputting the network module.
4. The method according to claim 1 , wherein a difference value between processing times of the at least two network modules is less than a predetermined threshold.
5. The method according to claim 1 , wherein
scheduling the start module thread based on the triggered thread task to process input data to the start module thread comprises: calling the start module thread to pre-process the source input data, performing task processing based on a start network module of the neural network corresponding to the start module thread, and outputting the processing result data to the cache as the cache exchange data; and
scheduling the end module thread based on the triggered thread task to process input data to the end module thread comprises: calling the end module thread to acquire the corresponding cache exchange data from the cache as the input data, performing task processing based on an end network module of the neural network corresponding to the end module thread, and post-processing and outputting processing result data as the processing result of the source input data.
6. The method according to claim 1 , wherein the at least two module threads further comprise at least one intermediate module thread; and scheduling the at least one intermediate module thread based on the triggered thread task to process the input data comprises:
calling the at least one intermediate module thread to acquire the corresponding cache exchange data from the cache as input data to the intermediate module thread, performing task processing based on an intermediate network module of the neural network corresponding to the intermediate module thread, and outputting the processing result data to the cache as the cache exchange data.
7. The method according to claim 1 , wherein the source input data is image frame data, the pre-processing is image pre-processing, the post-processing is image post-processing, and the neural network is a convolutional neural network.
8. The method according to claim 1 , wherein acquiring the input data comprises:
determining the data written in a target cache space as the input data in response to monitoring that data is written in a target cache space correspondingly read by the thread tasks, wherein the module threads are bound to the thread tasks in one-to-one correspondence, and the thread tasks are bound to the target cache space.
9. The method according to claim 1 , wherein the module threads are bound to processor cores in one-to-one correspondence.
10. A task processing device based on a neural network, comprising: a multi-core processor and a memory storing at least one instruction therein;
wherein the instruction, when executed by the multi-core processor, causes the device to execute a task processing method comprising:
acquiring input data, wherein the input data is intended to trigger thread tasks, and is source input data or cache exchange data;
generating processing result data by scheduling at least two corresponding module threads in parallel based on at least two triggered thread tasks to process the input data, wherein the at least two module threads respectively correspond to at least two network modules in the neural network;
outputting the processing result data to a cache, wherein the processing result data is used as the cache exchange data of module threads except the at least two module threads, or outputting the processing result data, wherein the processing result data is used as a processing result of the source input data;
wherein the at least two module threads at least comprise a start module thread and an end module thread;
wherein generating the processing result data by scheduling at least two corresponding module threads in parallel based on at least two triggered thread tasks to process the input data comprises:
scheduling the start module thread based on the triggered thread task to process input data to the start module thread; and
scheduling the end module thread based on the triggered thread task to process input data to the end module thread; and
executing tasks of the scheduled start module thread on a first core of the multi-core processor and executing tasks of the scheduled end module thread on a different core of the multi-core processor.
11. The device according to claim 10 , wherein prior to acquiring the input data, the method further comprises:
acquiring the at least two network modules by dividing a plurality of network layers in the neural network.
12. The device according to claim 11 , wherein acquiring the at least two network modules by dividing the plurality of network layers in the neural network comprises:
determining a number of channels between adjacent network layers in the neural network;
identifying a former network layer in the adjacent network layers as an input layer for inputting a network module and identifying a latter network layer in the adjacent network layers as an output layer for outputting the network module in response to determining that one channel is present between the adjacent network layers; and
generating the at least two network modules based on the output layer for outputting the network module and the input layer for inputting the network module.
13. The device according to claim 10 , wherein a difference value between processing times of the at least two network modules is less than a predetermined threshold.
14. The device according to claim 10 , wherein scheduling the start module thread based on the triggered thread task to process the input data to the start module thread comprises: calling the start module thread to pre-process the source input data, performing task processing based on a start network module of the neural network corresponding to the start module thread, and outputting the processing result data to the cache as the cache exchange data; and
scheduling the end module thread based on the triggered thread task to process the input data to the end module thread comprises: calling the end module thread to acquire the corresponding cache exchange data from the cache as the input data, performing task processing based on an end network module of the neural network corresponding to the end module thread, and post-processing and outputting processing result data as the processing result of the source input data.
15. The device according to claim 10 , wherein the at least two module threads further comprise at least one intermediate module thread; and scheduling the at least one intermediate module thread based on the triggered thread task to process the input data comprises:
calling the at least one intermediate module thread to acquire the corresponding cache exchange data from the cache as input data to the intermediate module thread, performing task processing based on an intermediate network module of the neural network corresponding to the intermediate module thread, and outputting the processing result data to the cache as the cache exchange data.
16. The device according to claim 10 , wherein the source input data is image frame data, the pre-processing is image pre-processing, the post-processing is image post-processing, and the neural network is a convolutional neural network.
17. The device according to claim 10 , wherein acquiring the input data comprises:
determining data written in a target cache space as the input data in response to monitoring that data is written in a target cache space correspondingly read by the thread tasks, wherein the module threads are bound to the thread tasks in one-to-one correspondence, and the thread tasks are bound to the target cache space.
18. A non-transitory computer-readable storage medium, storing at least one instruction therein, wherein the at least one instruction, when executed by a multi-core processor of a device, causes the device to execute a task processing method based on a neural network comprising:
acquiring input data, wherein the input data is intended to trigger thread tasks, and is source input data or cache exchange data;
generating processing result data by scheduling at least two corresponding module threads in parallel based on at least two triggered thread tasks to process the input data, wherein the at least two module threads respectively correspond to at least two network modules in the neural network;
outputting the processing result data to a cache, wherein the processing result data is used as the cache exchange data of module threads except the at least two module threads, or outputting the processing result data, wherein the processing result data is used as a processing result of the source input data;
wherein the at least two module threads at least comprise a start module thread and an end module thread;
wherein generating the processing result data by scheduling at least two corresponding module threads in parallel based on at least two triggered thread tasks to process the input data comprises:
scheduling the start module thread based on the triggered thread task to process input data to the start module thread; and
scheduling the end module thread based on the triggered thread task to process input data to the end module thread; and
executing tasks of the scheduled start module thread on a first core of the multi-core processor and executing tasks of the scheduled end module thread on a different core of the multi-core processor.