CN114912594A - Neuron release control method, many-core system, processing core and medium - Google Patents
Neuron release control method, many-core system, processing core and medium Download PDFInfo
- Publication number
- CN114912594A CN114912594A CN202210528170.1A CN202210528170A CN114912594A CN 114912594 A CN114912594 A CN 114912594A CN 202210528170 A CN202210528170 A CN 202210528170A CN 114912594 A CN114912594 A CN 114912594A
- Authority
- CN
- China
- Prior art keywords
- synchronous
- neuron
- neurons
- issuing
- state information
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F15/00—Digital computers in general; Data processing equipment in general
- G06F15/16—Combinations of two or more digital computers each having at least an arithmetic unit, a program unit and a register, e.g. for a simultaneous processing of several programs
- G06F15/163—Interprocessor communication
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Biophysics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Computer Hardware Design (AREA)
- Artificial Intelligence (AREA)
- Neurology (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Hardware Redundancy (AREA)
Abstract
The disclosure provides a neuron release control method, a many-core system, a processing core and a medium. The method is applied to processing synchronous neurons in a core, and comprises the following steps: responding to first release information sent by a preceding neuron of the synchronous neuron, and performing release operation to obtain parameter state information of the synchronous neuron; synchronously issuing with synchronous neurons outside the processing core based on the parameter state information. According to the embodiment of the disclosure, the transmission delay can be reduced, and the processing efficiency of the many-core system is improved.
Description
Technical Field
The present disclosure relates to the field of computer technologies, and in particular, to a neuron firing control method, a many-core system, a processing core, and a computer-readable medium.
Background
High-performance brain-like computing and simulation techniques are important for the research of brain science, in which a large number of neurons are required to co-operate to complete the basic functions of the human brain, and a large-scale brain simulation system is required to accurately simulate the brain functions.
At the present stage, when a system performs brain simulation, multiple devices may need to be simulated together, cross-machine communication needs to be performed between different devices, but the data volume of cross-machine transmission is large, the transmission delay is high, and the system efficiency is low.
Disclosure of Invention
The present disclosure provides a neuron firing control method, a many-core system, a processing core, and a computer readable medium.
In a first aspect, the present disclosure provides a neuron firing control method, applied to processing synchronous neurons in a core, the method including: responding to first release information sent by a preceding neuron of the synchronous neuron, and performing release operation to obtain parameter state information of the synchronous neuron; synchronously issuing with synchronous neurons outside the processing core based on the parameter state information.
In some embodiments, synchronously firing with a synchronization neuron outside the processing core based on parameter state information includes: according to the parameter state information, when a first preset issuing condition is met, pre-issuing is carried out, and synchronous neurons outside the processing core are controlled to issue.
In some embodiments, according to the parameter state information, when a first preset issuing condition is satisfied, performing pre-issuing, and controlling a synchronous neuron outside the processing core to issue, includes:
sending second issuing information to a subsequent neuron of the synchronous neuron; and sending second issuing information to the synchronous neurons outside the processing core so that the synchronous neurons outside the processing core issue according to the second issuing information.
In some embodiments, synchronously firing with out-of-core synchronization neurons based on parameter state information includes: sending parameter state information to a synchronous control unit outside the processing cores, so that the synchronous control unit performs issuing operation according to the parameter state information sent by a synchronous neuron connected with the synchronous control unit, and issuing a synchronous control instruction to the synchronous neuron connected with the synchronous control unit when a second preset issuing condition is met, wherein each two synchronous neurons belong to different processing cores;
and responding to a synchronous control instruction issued by the synchronous control unit, and issuing the synchronous control instruction after the synchronous control instruction is issued.
In some embodiments, the synchronization control unit is a synchronization control neuron, which is any one of the synchronization neurons.
In some embodiments, the parameter status information comprises: at least one of a membrane potential parameter, a weight parameter, and a firing threshold parameter.
In a second aspect, the present disclosure provides a neuron firing control method, applied to processing synchronous neurons in a core, the method including: and responding second issuing information sent by the synchronous neurons outside the processing core to issue, wherein the second issuing information is sent by the synchronous neurons outside the processing core when the synchronous neurons meet the first preset issuing condition and issue before.
In a third aspect, the present disclosure provides a neuron release control method applied to a synchronization control unit in a processing core, the method including:
performing issuing operation according to parameter state information sent by a synchronous neuron connected with the synchronous control unit; when a second preset issuing condition is met, issuing a synchronous control instruction to a synchronous neuron connected with the synchronous control unit so as to control the synchronous neuron connected with the synchronous control unit to issue; wherein, every two synchronous neurons belong to different processing cores; the parameter state information is obtained by the synchronous neuron through issuing operation according to the first issuing information sent by the preceding neuron.
In a fourth aspect, the present disclosure provides a many-core system, comprising: a plurality of processing cores, wherein at least some of the processing cores have synchronization neurons disposed therein; the synchronous neuron is used for responding to first release information sent by a preceding neuron of the synchronous neuron and performing release operation to obtain parameter state information of the synchronous neuron; synchronously issuing with synchronous neurons outside the processing core based on the parameter state information.
In a fifth aspect, the present disclosure provides a processing core having a program stored thereon, wherein the processing core executes the program to implement the neuron issuance control method according to the above-described embodiment.
In a sixth aspect, the present disclosure provides a computer-readable medium, on which a computer program is stored, wherein the computer program, when executed by a processor, implements the steps in the neuron firing control method as in the above embodiments.
The neuron release control method, the many-core system, the processing core and the computer readable medium provided by the disclosure can disassemble neurons into synchronous neurons which are in different processing cores and have an association relationship, and the synchronous neurons can perform synchronous release, so that a cross-machine data transmission mechanism with neuron granularity is established, transmission delay among the neurons can be reduced, and the processing efficiency of the many-core system is improved.
It should be understood that the statements in this section do not necessarily identify key or critical features of the embodiments of the present disclosure, nor do they limit the scope of the present disclosure. Other features of the present disclosure will become apparent from the following description.
Drawings
The accompanying drawings are included to provide a further understanding of the disclosure and are incorporated in and constitute a part of this specification, illustrate embodiments of the disclosure and together with the description serve to explain the principles of the disclosure and not to limit the disclosure. The above and other features and advantages will become more apparent to those skilled in the art by describing in detail exemplary embodiments thereof with reference to the attached drawings, in which:
FIG. 1 is a schematic diagram of a key neuron provided by an embodiment of the present disclosure;
fig. 2 is a flowchart of a neuron firing control method according to an embodiment of the present disclosure;
FIG. 3 is a schematic diagram of a synchronization neuron provided by an embodiment of the present disclosure;
fig. 4 is a flowchart of a neuron firing control method according to an embodiment of the present disclosure;
fig. 5 is a flowchart of a neuron firing control method according to an embodiment of the present disclosure;
FIG. 6 is a schematic diagram of a synchronization neuron and a synchronization control unit according to an embodiment of the present disclosure;
fig. 7a and 7b are schematic diagrams of a neuron firing control method provided by an embodiment of the present disclosure; fig. 8 is a schematic structural diagram of a many-core system according to an embodiment of the present disclosure.
Detailed Description
To facilitate a better understanding of the technical aspects of the present disclosure, exemplary embodiments of the present disclosure are described below in conjunction with the accompanying drawings, wherein various details of the embodiments of the present disclosure are included to facilitate an understanding, and they should be considered as merely exemplary. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the present disclosure. Also, descriptions of well-known functions and constructions are omitted in the following description for clarity and conciseness.
Embodiments of the present disclosure and features of embodiments may be combined with each other without conflict.
As used herein, the term "and/or" includes any and all combinations of one or more of the associated listed items.
The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and/or "comprising," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof. The terms "connected" or "coupled" and the like are not restricted to physical or mechanical connections, but may include electrical connections, whether direct or indirect.
Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and the present disclosure, and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.
When the brain simulation system carries out brain simulation, a plurality of devices may be required to carry out common simulation, cross-machine communication is required to be carried out among different devices, but the cross-machine transmission data volume is large, the transmission delay is high, and the system efficiency is low.
Especially for some key node neurons (which may be called key neurons), as shown in fig. 1, the key neuron has a connection relationship with many neurons (e.g., neurons in different processing cores or different machines), and issues and receives more neurons. For example, a key neuron located in machine 1 in fig. 1 may have a connection relationship with a plurality of neurons in machines 1 and 2, and may be capable of receiving information from a preceding neuron located in machines 1 and 2 and transmitting information to a succeeding neuron located in machines 1 and 2.
The more key neurons of the key nodes, the better the brain simulation effect. However, the more the number of the key node neurons is, especially under the condition that the clustering is more in a many-core system for brain simulation, the problem of large delay of cross-machine transmission of the brain simulation is more difficult to overcome.
According to the embodiment of the disclosure, the key neuron can be decomposed into two or more neurons with an association relationship, and the neurons are in different processing cores, even different machines and devices, and can be synchronously issued. Therefore, a cross-machine data transmission mechanism with neuron granularity can be established, transmission delay among neurons is reduced, and system efficiency is improved.
The method and the device can be applied to processing cores of a many-core system. The processing cores of the many-core system, which may also be referred to as "cores", or "functional cores", are the smallest units of the many-core system that can be independently scheduled and have complete computing power, and each processing core has its own storage computing resource. The same many-core system comprises a plurality of processing cores, and the processing cores are in communication connection through a network on chip or a bus on chip. One or more processing cores may be included in a device or machine, among others.
In some embodiments, each processing core may be loaded with one or more neurons that may perform a firing operation based on information from a previous stage (e.g., a preceding neuron or input layer, etc.) and send firing information to a subsequent stage (e.g., a succeeding neuron or output layer, etc.) when a firing condition is met (e.g., a firing threshold is reached).
In some embodiments, the neurons loaded in the many-core system may be neurons in a neural network, the neural network is used for implementing preset processing tasks, such as image processing tasks, voice processing tasks, text processing tasks, video processing tasks, and the like, and the disclosure does not limit the specific types of processing tasks performed by the neural network.
Fig. 2 is a flowchart of a neuron firing control method according to an embodiment of the present disclosure. In some embodiments, the method is applied to synchronized neurons within a processing core of a many-core system. As shown in fig. 2, the method includes:
step S11, responding to the first release information sent by the preceding neuron of the synchronous neuron, and performing release operation to obtain the parameter state information of the synchronous neuron;
step S12, synchronously issuing with the synchronous neurons outside the processing core based on the parameter status information.
For example, for any neuron, its predecessor neuron refers to the neuron whose output is connected to its own input; accordingly, a successor neuron refers to a neuron whose input is connected to its output.
In some embodiments, two or more neurons having an associative relationship may be referred to as a set of synchronized neurons. Different ones of a group of synchronization neurons are in different processing cores, and may also be in different devices or machines.
For example, an intermediate neuron having a connection relationship with multiple neurons in different processing cores or even different machines may be decomposed into a set of synchronous neurons having an association relationship. It should be understood that there may be other situations for a group of synchronous neurons having an association relationship, and those skilled in the art can set the association relationship between the neurons according to the actual situation, and the disclosure is not limited thereto.
In some embodiments, each synchronization neuron may have at least one preceding neuron and/or at least one succeeding neuron, both of which are in its own processing core (or in its own machine). The preceding neurons and the succeeding neurons in the other processing cores (or other machines) are connected to the synchronization neurons in the corresponding processing core (or corresponding machine). In this way, each synchronization neuron receives information of a preceding neuron within the core (or within the machine) and issues information to a succeeding neuron within the core (or within the machine), so that the reception and the issuance of information do not have to wait for a transmission delay between cores (or between machines), thereby being capable of increasing the speed of reception and issuance.
In some embodiments, in response to a case that one of the synchronization neurons a has a plurality of preceding neurons and no succeeding neuron in the processing core to which the synchronization neuron a belongs, and another of the synchronization neurons B has a plurality of succeeding neurons and no preceding neuron in the processing core to which the synchronization neuron B belongs, the preceding neurons of the synchronization neuron a, the synchronization neuron B, and the succeeding neurons of the synchronization neuron B constitute a neuron cluster across the cores, and the synchronization neuron a and the synchronization neuron B are a fixed set of synchronization neurons, which is equivalent to one intermediate neuron in the neural network.
In some embodiments, a group of synchronization neurons may each perform a firing operation, and information may be aggregated and synchronized among the group of synchronization neurons. This scheme may be referred to as a summarization scheme, i.e., summarize information for a set of synchronized neurons. Fig. 3 is a schematic diagram of a synchronization neuron provided by an embodiment of the present disclosure. As shown in fig. 3, two synchronous neurons having an association relationship each perform a firing operation and can perform information synchronization. Through the processing mode of information summarization, each synchronous neuron in the same group can be operated respectively, and can be issued before issuing when the neuron needs to be issued, so that the issuing speed is increased; meanwhile, the information of each synchronous neuron is collected and then sent when the information needs to be sent after being collected, so that the processing precision of the system is improved.
In some embodiments, a group of synchronization neurons may have a firing operation performed by one synchronization neuron, and other synchronization neurons may directly synchronize the firing operation result of the synchronization neuron, and may also fire if the synchronization neuron fires. This scheme may be referred to as a shadow scheme, i.e. other sync neurons act as shadow neurons for this sync neuron. By means of the processing mode of shadow issuing, only part of synchronous neurons in the same group operate, and other synchronous neurons only issue, so that the number of neurons participating in operation can be reduced, and the processing efficiency of the system is improved.
It should be understood that the person skilled in the art can set the synchronization mode among a group of synchronous neurons according to practical situations, and the disclosure is not limited thereto.
In some embodiments, at any time step during the operation of the system, for any one of the neurons, in the case of receiving the first release information (possibly including one or more pieces of release information) sent by the neuron preceding the neuron, in step S11, a release operation may be performed in response to the first release information, that is, the membrane potential state of the neuron itself is calculated according to the first release information and its current neuron parameters (including the weight, the release threshold, and the like), so as to obtain the parameter state information of the neuron.
In some embodiments, the parameter status information may include: at least one of a membrane potential parameter, a weight parameter, and a firing threshold parameter. The parameter state information can also comprise a release identifier used for directly indicating whether to release; in some embodiments, the weight parameter and the release threshold parameter may be adaptively adjusted, and the parameter status information may further include dynamic change information and a change identifier corresponding to each parameter.
In some embodiments, the parameter state information may further include identification information of the synchronization neuron, the identification information pointing to at least one of a processing core and a device to which the neuron belongs, and pointing to the corresponding neuron, and the identification information may include various forms, such as a tag and a flag, which are not limited by the present disclosure.
The present disclosure does not limit the specific content included in the parameter status information.
In some embodiments, in step S12, firing may be synchronized with a synchronization neuron outside the processing core based on the parameter state information.
In some embodiments, in the case that a group of synchronization neurons adopts a shadow scheme, the synchronization neurons may determine whether a preset issuing condition (for example, reaching an issuing threshold) is satisfied according to the parameter status information in step S12; if the preset issuing condition is met, directly issuing (called pre-issuing), and controlling other synchronous neurons outside the processing core to issue, for example, sending corresponding control information; and if the preset issuing condition is not met, not sending information to other synchronous neurons outside the processing core.
The synchronous neuron performs operation in a shadow issuing processing mode, and performs pre-issuing when the synchronous neuron needs to issue, so that the issuing speed is increased; other synchronous neurons in the same group are issued only based on the control information, so that the number of neurons participating in operation can be reduced, and the processing efficiency of the system is improved.
In some embodiments, in a case that a group of synchronous neurons adopts a summary scheme, the synchronous neurons may send their own parameter state information to a synchronous control unit (e.g., one of the group of synchronous neurons, or an individual neuron), and the synchronous control unit performs a distribution operation on the parameter state information of each synchronous neuron of the group, that is, performs membrane potential accumulation and distribution judgment according to membrane potential information fed back by the synchronous neurons; when the release condition is determined to be met, controlling the release of each synchronous neuron; whether the membrane potential is released or not, the synchronous control unit can control each synchronous neuron to update state information, such as undistributed and accumulated membrane potential, and membrane potential reset after being released, so as to perform operations such as integral operation and release judgment in the next time step.
Through the processing mode of information summarization, each synchronous neuron in the same group can be operated respectively, and can be issued before issuing when the neuron needs to be issued, so that the issuing speed is increased; meanwhile, the information of each synchronous neuron is collected, the state information is updated after the collection, and the state information is issued after the state information is issued when the state information needs to be issued, so that the processing precision of the system is improved.
According to the embodiment of the disclosure, the key neurons can be disassembled into the synchronous neurons which are in different processing cores and have the incidence relation, and the synchronous neurons can be synchronously issued, so that the transmission delay between the neurons in the same processing core or the same machine can be reduced, and the processing efficiency of a many-core system is improved.
The following is a description of a neuron firing control method according to an embodiment of the present disclosure.
As previously described, two or more neurons having an associative relationship may be referred to as a set of synchronized neurons. Before the processing of steps S11-S12 is performed, an association relationship between the neurons in synchronization may be established.
In some embodiments, before step S11, the neuron issuance control method according to the embodiment of the present disclosure may further include: and sending the identification information of the synchronous neurons to other synchronous neurons to establish an association relationship with the other synchronous neurons. The identification information points to at least one of a processing core and a device to which the neuron belongs, and points to the corresponding neuron, and the identification information includes various forms such as a tag and a flag bit. The present disclosure is not so limited.
In some embodiments, the synchronous neurons may mutually identify identification information to establish an association relationship between the synchronous neurons, such that a group of synchronous neurons can be equivalent to one intermediate neuron in a neural network.
In some embodiments, a group of synchronization neurons may use a shadow scheme, that is, one synchronization neuron performs a firing operation, and the other synchronization neurons directly synchronize the firing operation results of the synchronization neuron.
For a synchronous neuron performing a firing operation, in any time step, the membrane potential state of the synchronous neuron is calculated in step S11 according to the first firing information sent by the preceding neuron of the synchronous neuron and its current neuron parameters (including weight, firing threshold, and the like), so as to obtain parameter state information of the synchronous neuron. Further, synchronization is performed in step S12.
In some embodiments, step S12 may include:
according to the parameter state information, when a first preset issuing condition is met, pre-issuing is carried out, and synchronous neurons outside the processing core are controlled to issue.
That is, the synchronization neuron may determine whether a first preset issuing condition is satisfied (e.g., an issuing threshold is reached) according to the parameter status information; if the first preset issuing condition is met, directly issuing (called pre-issuing), and controlling other synchronous neurons outside the processing core to issue, for example, sending corresponding control information; and if the first preset issuing condition is not met, not sending information to other synchronous neurons outside the processing core. The present disclosure does not limit the specific contents of the first preset dispensing condition.
In this way, the issuing operation can be centralized in one neuron, and the processing efficiency is improved.
In some embodiments, the step of performing pre-firing and controlling the synchronous neurons outside the processing core to fire when the first preset fire condition is met according to the parameter state information may include:
sending second issuing information to a subsequent neuron of the synchronous neuron; and sending second issuing information to the synchronous neurons outside the processing core so that the synchronous neurons outside the processing core issue according to the second issuing information.
That is, the process of pre-issue is: and sending second issuing information to subsequent neurons of the synchronous neuron so that each subsequent neuron performs corresponding processing according to the second issuing information. In this way, the information issuing speed can be increased, so that the subsequent neurons can receive the issued information earlier to perform corresponding processing.
In some embodiments, the sending of the second firing information to the subsequent neuron of the synchronous neuron itself may be sending the second firing information to all or part of the subsequent neuron, which is not limited by the present disclosure.
In some embodiments, the process of controlling the firing of the synchronous neurons outside the processing core is: and sending second issuing information to the synchronous neurons outside the processing core so that the synchronous neurons outside the processing core issue according to the second issuing information. In this way, the number of neurons participating in the calculation can be reduced, and the processing efficiency of the system can be improved.
Fig. 4 is a flowchart of a neuron firing control method according to an embodiment of the present disclosure. The application is used for processing synchronous neurons in a core, and as shown in fig. 4, the method comprises the following steps:
step S31, issuing in response to second issue information sent by the external synchronization neuron, where the second issue information is sent by the external synchronization neuron when it meets the first preset issue condition and issues before.
That is to say, for a synchronous neuron which does not perform a firing operation in the shadow scheme, the synchronous neuron which performs the firing operation in the same group of synchronous neurons will send second firing information to the synchronous neuron during firing; if the second issuing information is received, the synchronous neuron sends the second issuing information to the subsequent neurons, so that each subsequent neuron performs corresponding processing according to the second issuing information, and the process can be called as post-issuing; if the second issue information is not received, no processing is performed in the synchronization neuron.
By the method, the whole issuing process of the group of synchronous neurons adopting the shadow scheme can be realized, the synchronous neurons executing the operation can be quickly issued, and the issuing speed of the information is improved; and other synchronous neurons in the same group are controlled to be issued, so that the number of neurons participating in operation is reduced, and the processing efficiency of the system is improved.
In some embodiments, a group of synchronization neurons may also adopt an aggregation scheme, that is, a group of synchronization neurons may each perform a firing operation, and information aggregation and synchronization may be performed between the group of synchronization neurons.
In the summary scheme, a synchronization control unit may be provided for performing information summary and synchronization control of a group of synchronization neurons. The synchronous control unit is a synchronous control neuron, and the synchronous control neuron is any one of synchronous neurons. That is, any one of a group of synchronization neurons can be used as a synchronization control unit, thereby reducing the number of neurons occupied by a group of synchronization neurons.
In some embodiments, a synchronization control unit of an individual neuron may also be provided, and the specific arrangement manner of the synchronization control unit is not limited by the present disclosure.
In some embodiments, in the aggregation scheme, for one of a group of synchronized neurons, an issuing operation may be performed in step S11 in response to the first issuing information sent by its preceding neuron, to obtain parameter status information of the synchronized neuron. Further, synchronization is performed in step S12.
In some embodiments, step S12 may include:
sending parameter state information to a synchronous control unit outside the processing cores, so that the synchronous control unit performs issuing operation according to the parameter state information sent by a synchronous neuron connected with the synchronous control unit, and issuing a synchronous control instruction to the synchronous neuron connected with the synchronous control unit when a second preset issuing condition is met, wherein each two synchronous neurons belong to different processing cores;
and responding to a synchronous control instruction issued by the synchronous control unit, and issuing the synchronous control instruction after the synchronous control instruction is issued.
For example, each two synchronous neurons belong to different processing cores, the synchronous control unit is located in another processing core except the processing core where the synchronous neuron is located, and the synchronous neurons can send parameter state information to the synchronous control unit outside the processing core, so that the synchronous control unit can realize information summarization.
In some embodiments, when receiving parameter state information sent by a synchronous neuron connected to the synchronous control unit, the synchronous control unit may perform an issuing operation according to the parameter state information to obtain aggregated state information (e.g., membrane potential state); and if the summarized state information meets a second preset issuing condition, issuing a synchronous control instruction to synchronous neurons connected with the synchronous control unit so as to issue each synchronous neuron.
In some embodiments, when receiving a synchronization control instruction issued by a synchronization control unit, a synchronization neuron may send corresponding issue information to its own subsequent neuron, which is referred to as post-issue. By means of the processing modes of gathering and synchronous issuing of the multiple neurons, the synchronous neurons in the same group can perform operation respectively, issue before the synchronous neurons need to be issued, and issue speed is improved; meanwhile, the information of each synchronous neuron is collected through the synchronous control unit, and the synchronous neurons are controlled to issue after the information is collected and needs to be issued, so that the processing precision of the system is improved.
In some embodiments, for the synchronization control unit, if the summarized state information does not satisfy the second preset issuing condition, the summarized state information may be issued to a synchronization neuron connected to the synchronization control unit. When receiving the issued state information, the synchronization neuron may update its own state, for example, update a membrane potential state.
By the mode, the state synchronization among all the synchronous neurons is realized, so that the next time step is processed continuously based on the state information of the current time step, the processing accuracy of all the synchronous neurons is improved, and the processing precision of a system is improved.
In some embodiments, for synchronous neurons, in step S12, if the parameter status information has satisfied the firing condition, the firing may be directly performed to its successor neurons, which may be referred to as a pre-firing. And meanwhile, parameter state information is sent to the synchronous control unit for the synchronous control unit to gather. By the method, the self-issuing speed of the synchronous neurons can be increased, so that the subsequent neurons can quickly receive issuing information for subsequent processing. Moreover, the mode does not influence the aggregation and synchronization among all synchronous neurons.
In this case, in the same time step, if the synchronization control unit that has issued before completion receives the synchronization control instruction of the synchronization control unit, it is not necessary to issue again, thereby avoiding repeated issuing of information.
Fig. 5 is a flowchart of a neuron firing control method according to an embodiment of the present disclosure. The synchronization control unit applied in the processing core, as shown in fig. 5, includes:
step S41, according to the parameter state information sent by the synchronous neuron connected with the synchronous control unit, the issuing operation is carried out;
step S42, when a second preset issuing condition is satisfied, issuing a synchronous control instruction to the synchronous neuron connected with the synchronous control unit to control the synchronous neuron connected with the synchronous control unit to issue;
wherein, every two synchronous neurons belong to different processing cores; the parameter state information is obtained by the synchronous neuron through issuing operation according to the first issuing information sent by the preceding neuron.
That is, when the synchronization control unit receives the parameter state information transmitted from the synchronization neuron connected to the synchronization control unit, the distribution operation is performed based on the parameter state information in step S41, and the status information (for example, membrane potential status) after the distribution operation is obtained.
In some embodiments, if the summarized state information satisfies the second preset issuing condition, a synchronization control instruction is issued to the synchronization neurons connected to the synchronization control unit in step S42 to cause the respective synchronization neurons to issue. If the summarized state information does not meet the second preset issuing condition, the summarized state information can be issued to the synchronous neurons connected with the synchronous control unit, so that each synchronous neuron can synchronize the state of the neuron. Therefore, each synchronous neuron can continue to process based on the state information of the current time step in the next time step, the accuracy of the state information of each synchronous neuron is improved, and the processing precision of the system is improved
In some embodiments, when the synchronization neuron receives a synchronization control instruction issued by the synchronization control unit and does not perform pre-issuance, the synchronization neuron may send corresponding issuance information to its successor neuron, that is, post-issuance.
In some embodiments, when the synchronization neuron receives the synchronization control instruction issued by the synchronization control unit and performs pre-issue, the synchronization neuron may update only its state (e.g., a membrane potential reset after issue) without performing post-issue, so as to perform processing (integration operation, issue determination) at the next time step and avoid repeatedly issuing information.
In some embodiments, when receiving the state information sent by the synchronization control unit, the synchronization neuron may update its own state, for example, update the state of the non-released and accumulated membrane potential, so as to perform processing (integration operation, release determination) at the next time step.
By means of the processing modes of gathering and synchronous issuing of the multiple neurons, the synchronous neurons in the same group can perform operation respectively, issue before the synchronous neurons need to be issued, and issue speed is improved; meanwhile, the information of each synchronous neuron is collected through the synchronous control unit, and the synchronous neurons are controlled to issue after the information is collected and needs to be issued, so that the processing precision of the system is improved.
Fig. 6 is a schematic diagram of a synchronization neuron and a synchronization control unit according to an embodiment of the present disclosure. As shown in fig. 6, a synchronization neuron a51 in the machine 1, a synchronization neuron a52 in the machine 2, and a synchronization control unit B51 in the machine 3. Sync neuron a51 and sync neuron a52 have respective preceding neurons and following neurons (not shown). The synchronization neuron is capable of receiving information of the preceding neuron (the arrow pointing to the synchronization neuron) and issuing information to the following neuron (the outward arrow pointing from the synchronization neuron).
In an example, the synchronous neuron a51 performs a firing operation when receiving first firing information sent by a preceding neuron, to obtain parameter state information of the synchronous neuron a 51; if the parameter state information has satisfied the firing condition of the synchronization neuron A51, the preceding firing can be directly performed to the subsequent neuron of itself, so that the subsequent neuron can quickly receive the firing information for subsequent processing.
In an example, the parameter status information is sent to synchronization control unit B51 regardless of whether synchronization neuron a51 was previously fired; when the synchronization control unit B51 receives the parameter state information of the synchronization neuron a51 and the synchronization neuron a52, it performs issuing operation according to the two parameter state information to obtain the aggregated state information.
In an example, if the aggregated state information satisfies the issuance condition of the synchronization control unit B51 (which may be different from the issuance conditions of a51 and a 52), a synchronization control instruction is issued to a51 and a52 to cause a51 and a52 to issue. If the aggregated state information does not satisfy the issuance condition, the aggregated state information may be issued to a51 and a52 to synchronize the respective synchronous neurons with their own states.
In an example, the synchronization neuron a51 may post-fire to its successor neuron in the case where it receives a synchronization control instruction issued by the synchronization control unit B51 and it does not pre-fire itself.
In an example, in the case where the synchronization neuron a51 receives the synchronization control instruction issued by the synchronization control unit B51 and has performed pre-firing itself, it may update only the state of itself (e.g., the membrane potential reset after firing) without performing post-firing, so that the next time step is processed and the repetitive firing of information is avoided.
In an example, the sync neuron a51, upon receiving the state information issued by the sync control unit B51, may update its own state, e.g., update the non-firing, accumulated membrane potential state, for processing at the next time step.
In an example, sync neuron a52 processes in a similar manner as sync neuron a 51.
In this way, a summary and synchronization process for a set of synchronized neurons is achieved.
Fig. 7a and 7b are schematic diagrams of a neuron firing control method according to an embodiment of the present disclosure. As shown in fig. 7a, the structure of the fully-connected layer is shown, wherein neurons a1, a2, A3 and a4 are respectively the successor neurons of neurons B1, B2 and B3, and neurons C1, C2, C3 and C4 are respectively the successor neurons of neurons B1, B2 and B3. The whole processing process has large processing amount, cannot be realized in the same machine, and has large transmission delay across machines, so that the processing efficiency is low.
As shown in fig. 7B, neurons B1, B2, B3 may be split into B1-1, B1-2, B2-1, B2-2, B3-1, B3-2, neurons a1, a2, A3, a4 and neurons B1-1, B2-1, B3-1, respectively, to be executed in machine 1; neurons B1-2, B2-2, B3-2 and neurons C1, C2, C3, C4 are implemented in machine 2. Neurons a1, a2, A3, a4, B1-1, B1-2, B2-1, B2-2, B3-1, B3-2, C1, C2, C3, C4 constitute an inter-machine neuron cluster.
In an example, neurons B1-1, B1-2, B2-1, B2-2, B3-1, B3-2 may form three groups of neurons in synchronization (B1-1, B1-2), (B2-1, B2-2) and (B3-1, B3-2) using a shadow scheme. For the synchronous neuron B1-1, the issuing information of the preceding neurons A1, A2, A3 and A4 is received, issuing operation is carried out, and the issuing information is sent to the synchronous neuron B1-2 when the issuing condition is met; after receiving the issue information, the synchronous neuron B1-2 issues information to the subsequent neurons C1, C2, C3, and C4.
In this way, neuron firing synchronization across machines is achieved.
According to the neuron release control method disclosed by the embodiment of the disclosure, key neurons can be decomposed into synchronous neurons which are in different processing cores and have an incidence relation, the synchronous neurons can be synchronously released by adopting a shadow scheme or a summary scheme and the like, and a cross-machine data transmission mechanism of neuron granularity is established, so that transmission delay is reduced, the synchronism among the neurons and the flexibility of data transmission are improved, and the processing efficiency of a many-core system is improved.
Fig. 8 is a schematic structural diagram of a many-core system according to an embodiment of the present disclosure. As shown in fig. 8, there is also provided, in accordance with an embodiment of the present disclosure, a many-core system, including:
a plurality of processing cores 71, wherein at least some of the processing cores have synchronization neurons (not shown) disposed therein; the synchronous neuron is used for responding to first release information sent by a preceding neuron of the synchronous neuron and performing release operation to obtain parameter state information of the synchronous neuron; synchronously issuing with synchronous neurons outside the processing core based on the parameter state information.
In some embodiments, as shown in fig. 8, the many-core system further includes a network-on-chip 72, wherein the plurality of processing cores 71 are each connected to the network-on-chip 72, and the network-on-chip 72 is configured to interact data between the plurality of processing cores and external data.
One or more instructions are stored in the one or more processing cores 71, and the one or more instructions are executed by the one or more processing cores 71, so that the synchronization neuron or the synchronization control unit in the one or more processing cores 71 can execute the neuron release control method.
According to an embodiment of the present disclosure, a processing core is also provided. The processing core stores a program thereon, wherein the processing core executes the program to implement the steps in the neuron release control method according to any one of the above embodiments.
According to an embodiment of the present disclosure, a computer-readable medium is also provided. The computer readable medium has stored thereon a computer program, wherein the computer program realizes the steps in the neuron firing control method according to any one of the above embodiments when executed by a processor.
It will be understood by those of ordinary skill in the art that all or some of the steps of the methods, systems, functional modules/units in the devices disclosed above may be implemented as software, firmware, hardware, and suitable combinations thereof. In a hardware implementation, the division between functional modules/units mentioned in the above description does not necessarily correspond to the division of physical components; for example, one physical component may have multiple functions, or one function or step may be performed by several physical components in cooperation. Some or all of the physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application specific integrated circuit. Such software may be distributed on computer readable media, which may include computer storage media (or non-transitory media) and communication media (or transitory media). The term computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data, as is well known to those of ordinary skill in the art. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, Digital Versatile Disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can accessed by a computer. In addition, communication media typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism and includes any information delivery media as known to those skilled in the art.
Example embodiments have been disclosed herein, and although specific terms are employed, they are used and should be interpreted in a generic and descriptive sense only and not for purposes of limitation. In some instances, features, characteristics and/or elements described in connection with a particular embodiment may be used alone or in combination with features, characteristics and/or elements described in connection with other embodiments, unless expressly stated otherwise, as would be apparent to one skilled in the art. Accordingly, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the scope of the disclosure as set forth in the appended claims.
Claims (11)
1. A neuron release control method is applied to synchronous neurons in a processing core, and comprises the following steps:
responding to first release information sent by a preceding neuron of the synchronous neurons, and performing release operation to obtain parameter state information of the synchronous neurons;
and synchronously issuing with synchronous neurons outside the processing core based on the parameter state information.
2. The neuron firing control method according to claim 1, wherein the synchronous firing with the synchronous neurons outside the processing core based on the parameter state information includes:
and according to the parameter state information, when a first preset issuing condition is met, performing pre-issuing, and controlling synchronous neurons outside the processing core to issue.
3. The neuron release control method according to claim 2, wherein the performing pre-release and controlling the synchronous neurons outside the processing core to release when a first preset release condition is satisfied according to the parameter state information comprises:
sending second issuing information to a subsequent neuron of the synchronous neuron;
and sending the second issuing information to the synchronous neurons outside the processing core so that the synchronous neurons outside the processing core issue according to the second issuing information.
4. The neuron firing control method according to claim 1, wherein the synchronous firing with the out-of-core synchronous neurons based on the parameter state information includes:
sending the parameter state information to a synchronous control unit outside the processing core, so that the synchronous control unit performs issuing operation according to the parameter state information sent by a synchronous neuron connected with the synchronous control unit, and issuing a synchronous control instruction to the synchronous neuron connected with the synchronous control unit when a second preset issuing condition is met, wherein two synchronous neurons belong to different processing cores;
and responding to the synchronous control instruction issued by the synchronous control unit, and issuing the synchronous control instruction after the synchronous control instruction is issued.
5. The neuron firing control method according to claim 4, wherein the synchronization control unit is a synchronization control neuron that is any one of synchronization neurons.
6. The neuron firing control method according to claim 1, wherein the parameter state information includes: at least one of a membrane potential parameter, a weight parameter, and a firing threshold parameter.
7. A neuron release control method is applied to synchronous neurons in a processing core, and comprises the following steps:
and responding to second issuing information sent by the synchronous neurons outside the processing core for issuing, wherein the second issuing information is sent by the synchronous neurons outside the processing core when the synchronous neurons meet a first preset issuing condition and issue before the synchronous neurons meet the first preset issuing condition.
8. A neuron release control method is applied to a synchronous control unit in a processing core, and comprises the following steps:
performing issuing operation according to parameter state information sent by a synchronous neuron connected with the synchronous control unit;
when a second preset issuing condition is met, issuing a synchronous control instruction to a synchronous neuron connected with the synchronous control unit so as to control the synchronous neuron connected with the synchronous control unit to issue;
wherein, every two synchronous neurons belong to different processing cores; the parameter state information is obtained by the synchronous neuron through issuing operation according to the first issuing information sent by the preceding neuron.
9. A many-core system, comprising:
a plurality of processing cores, wherein at least some of the processing cores have synchronization neurons disposed therein;
the synchronous neuron is used for responding to first release information sent by a preceding neuron of the synchronous neuron and performing release operation to obtain parameter state information of the synchronous neuron; and synchronously issuing with synchronous neurons outside the processing core based on the parameter state information.
10. A processing core having a program stored thereon, wherein the processing core executes the program to implement the neuron firing control method according to any one of claims 1-8.
11. A computer-readable medium, on which a computer program is stored, wherein the computer program, when being executed by a processor, implements the neuron firing control method according to any one of claims 1-8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210528170.1A CN114912594A (en) | 2022-05-16 | 2022-05-16 | Neuron release control method, many-core system, processing core and medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210528170.1A CN114912594A (en) | 2022-05-16 | 2022-05-16 | Neuron release control method, many-core system, processing core and medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114912594A true CN114912594A (en) | 2022-08-16 |
Family
ID=82766970
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210528170.1A Pending CN114912594A (en) | 2022-05-16 | 2022-05-16 | Neuron release control method, many-core system, processing core and medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114912594A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117194051A (en) * | 2023-11-01 | 2023-12-08 | 北京灵汐科技有限公司 | Brain simulation processing method and device, electronic equipment and computer readable storage medium |
-
2022
- 2022-05-16 CN CN202210528170.1A patent/CN114912594A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117194051A (en) * | 2023-11-01 | 2023-12-08 | 北京灵汐科技有限公司 | Brain simulation processing method and device, electronic equipment and computer readable storage medium |
| CN117194051B (en) * | 2023-11-01 | 2024-01-23 | 北京灵汐科技有限公司 | Brain simulation processing method and device, electronic equipment and computer readable storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11372688B2 (en) | Resource scheduling method, scheduling server, cloud computing system, and storage medium | |
| TWI803663B (en) | A computing device and computing method | |
| WO2018140885A1 (en) | Memory side acceleration for deep learning parameter updates | |
| CN104317650A (en) | Map/Reduce type mass data processing platform-orientated job scheduling method | |
| US9037524B2 (en) | Neural working memory device | |
| CN112380020A (en) | Computing power resource allocation method, device, equipment and storage medium | |
| CN110708369B (en) | File deployment method and device for equipment nodes, scheduling server and storage medium | |
| KR20210156243A (en) | Training methods of deep-running frameworks, devices and storage media | |
| CN114912594A (en) | Neuron release control method, many-core system, processing core and medium | |
| JP2020515130A (en) | Programmable logic circuits for controlling electrical installations, particularly nuclear installations, and associated control devices and methods | |
| WO2022111456A1 (en) | Core sharing method and apparatus based on many-core system, electronic device, and medium | |
| EP2759927B1 (en) | Apparatus and method for sharing function logic between functional units, and reconfigurable processor thereof | |
| CN114691372A (en) | Group intelligent control method of multimedia end edge cloud system | |
| CN113347238A (en) | Message partitioning method, system, device and storage medium based on block chain | |
| CN116663639B (en) | Gradient data synchronization method, system, device and medium | |
| CN116700933B (en) | Heterogeneous computing power federation-oriented multi-cluster job scheduling system and method | |
| EP3841530B1 (en) | Distributed ai training topology based on flexible cable connection | |
| CN111049900B (en) | Internet of things flow calculation scheduling method and device and electronic equipment | |
| CN102387137B (en) | Implementation method and system of intelligent operation logic of a plurality of network devices | |
| CN110704382B (en) | File deployment method, device, server and storage medium | |
| CN110275772A (en) | A kind of data processing method and its relevant device | |
| JP2012038275A (en) | Transaction calculation simulation system, method, and program | |
| CN112242919B (en) | Fault file processing method and system | |
| CN114546926A (en) | Core cluster synchronization, control method, data processing method, core, device, and medium | |
| CN105991413B (en) | The processing method and processing device of message mailbox failure under user terminal and its AMP system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |