CN112463532B - Method for constructing SNN workload automatic mapper and automatic mapper - Google Patents

Abstract

The invention relates to a method and an automatic mapper for constructing an SNN workload automatic mapper. Then go to the next step; if there is no change or no storage file, quantify the design and store the obtained parameters in the form of a file, and then go to the next step; write a special script for mapping, and provide a special function interface for mapping; The network parameters of the mapped SNN case are automatically collected to obtain network modeling parameters; the memory and time consumption predictions are performed on the obtained network modeling parameters; the obtained memory and time consumption prediction data are converted into mapping results. The present invention can ensure the stable and high-performance operation of the computing platform.

Description

Method and Automapper for Building SNN Workload Automapper

technical field

The present invention relates to the technical field of SNN workload, in particular to a method for constructing an SNN workload automatic mapper and the automatic mapper.

Background technique

In order to meet the computing requirements of a large-scale Spiking Neural Network (SNN), a brain-like computing system usually requires a large-scale parallel computing platform. Therefore, how to quickly determine a reasonable number of computing nodes for the SNN workload (that is, how to reasonably map the workload to the computing platform) to obtain the best performance, power consumption and other indicators has become one of the key issues to be solved by brain-like computing systems. one.

Aiming at the problem of SNN workload prediction and matching, Kunkel et al.'s work achieves the prediction of the storage consumption of the SNN network through modeling and manual analysis. The work of Schenck et al. realized the prediction of simulation time consumption by fitting and training the LM least squares optimization algorithm on the modeling and collected measured data sets. Although these methods effectively solve the problems of storage and simulation time prediction, their load models target a single case, are not universal, and have no advantage in solving the mapping problem. This way of fitting a model by manually collecting measurements is undoubtedly inefficient and expensive.

For the problem of how to realize large-scale simulation of SNN, the current mainstream method in the industry is to deploy large-scale distributed clusters to realize the dilution of storage and calculation of large-scale SNN network in a parallel manner. However, in actual operation, the following two load problems are encountered: First, regarding storage, when carrying large-scale SNN cases, based on the idea of distributed storage, the scale of the invested computing nodes will be too small, making the computing platform It is difficult to provide storage to support the storage required for the case, resulting in the inability of the computing platform to run stably; secondly, regarding performance, in the specific simulation, the performance of the computing platform will be affected by the node scale factor, that is, the balance between computing and communication. When the node scale is smaller/greater than the balance point, the computation time will be larger/smaller than the communication time, resulting in underutilization/waste of computing platform performance.

SUMMARY OF THE INVENTION

Therefore, the technical problem to be solved by the present invention is to overcome the storage and performance problems in the prior art, which will cause the workload of the SNN case to be seriously mismatched with the actual computing platform, resulting in the problem of inability to run stably and efficiently, thereby providing a solution for different The SNN case and different hardware platforms can quickly determine the node mapping that satisfies the minimum storage requirement and the balance of computing communication, and becomes a method and an automatic mapper for building an SNN workload automatic mapper to ensure the stable and high-performance operation of the platform.

In order to solve the above-mentioned technical problems, a method for constructing an SNN workload automatic mapper of the present invention includes: Step S1: judging whether the CPU operation frequency and memory of the computing platform have changed, and if there is a change, continue to judge whether there is a storage file, If there is a storage file, go to step S2; if there is no change, or there is no storage file, quantify the design and store the obtained parameters in the form of a file, and then enter step S2; Step S3: Automatically collect the network parameters of the SNN case to be mapped to obtain network modeling parameters; Step S4: Predict the memory and time consumption of the obtained network modeling parameters; Step S5: Convert the obtained memory and time consumption prediction data into mapping results.

In one embodiment of the present invention, the quantization design includes quantization of storage parameters and quantization of time parameters.

In an embodiment of the present invention, the method for quantifying the stored parameters is: quantifying and quantifying the parameters manually, and at the same time, using the memory tool provided by the NEST simulator to map and correct the parameters.

In one embodiment of the present invention, the quantification of the time parameter includes the time quantification of communication and the time quantification of device, neuron, and synapse models.

In one embodiment of the invention, the mapping specific script includes a prediction time module, a network building module and a start prediction module.

In one embodiment of the present invention, when the network parameters of the SNN case to be mapped are automatically collected, a series of creations and connections are performed between the prediction time module and the start prediction module to construct the network. The R of the neuron group _spike and T _ref are entered at creation time.

In an embodiment of the present invention, for the related parameters such as the scale, in-and-out degree, etc. of neurons and synapses corresponding to the connection, there are two one-dimensional vectors: neuron type table v_n_type, synapse type table v_s_type and 3 A two-dimensional vector: create table v_n_create, out-degree table v_n_out and in-degree table v_n_in for storage.

In one embodiment of the present invention, when storing and predicting the time consumption of the obtained network modeling parameters, by designing each load model function, the modeling parameter data obtained through quantitative design and obtained is processed, so as to realize the scale of each node. forecast below.

In an embodiment of the present invention, the method for converting the obtained memory and time consumption prediction data into a mapping result is as follows: designing a mapping function, processing the obtained prediction data, obtaining mapping variables under the scale of each node, and converting the storage occupancy rate into The mapping threshold and optimal performance are used as mapping criteria, and the mapping result is derived through the specified function.

The invention provides an automatic mapper, comprising: a judgment module for judging whether the CPU operation frequency and memory of the computing platform have changed, if there is a change, continue to judge whether there is a storage file, and if there is a storage file, enter the writing module; If there is no change, or no file is stored, quantify the design and store the obtained parameters in the form of a file, and then enter the writing module; the writing module is used to write a special script for mapping and provides a special function interface for mapping; the acquisition module, It is used to automatically collect the network parameters of the SNN cases that need to be mapped to obtain network modeling parameters; the prediction module is used to predict the memory and time consumption of the obtained network modeling parameters; the mapping module is used to The obtained memory and time consumption prediction data are converted into mapping results.

The above-mentioned technical scheme of the present invention has the following advantages compared with the prior art:

The method for constructing an SNN workload automatic mapper and the automatic mapper described in the present invention design and implement a method for constructing an SNN workload automatic mapper (Workload Automatic Mapper for SNN, SWAM for short), and SWAM can cover different The load situation of the SNN network on different computing platforms; quickly and effectively deduce the reasonable mapping results of the entire workload on the computing platform to ensure the stable and high-performance operation of the computing platform; avoid the extremely time-consuming operation of the complete workload It has practical significance for the research and optimization of the brain-like architecture.

Description of drawings

In order to make the content of the present invention easier to understand clearly, the present invention will be described in further detail below according to specific embodiments of the present invention and in conjunction with the accompanying drawings, wherein

Fig. 1 is the method flow chart that the present invention constructs SNN workload automatic mapper;

Fig. 2 is the structural representation of the quantitative benchmark network of the present invention;

Fig. 3 is the quantization schematic diagram of storage and time parameter of the present invention;

Fig. 4 is the schematic diagram of collection, prediction and mapping of the present invention;

Fig. 5a is the composition schematic diagram of simulation time load model of the present invention;

Fig. 5b is the composition schematic diagram of memory and simulation time model of the present invention;

FIG. 6 is a schematic diagram of the subdivision of t _n in the present invention.

Detailed ways

Example 1

As shown in FIG. 1 and FIG. 2 , this embodiment provides a method for constructing an SNN workload automatic mapper, including: Step S1: Determine whether the CPU operation frequency and memory of the computing platform have changed, and if there are changes, continue to determine whether If there is a storage file, if there is a storage file, go to step S2; if there is no change, or there is no storage file, quantify the design and store the obtained parameters in the form of a file, and then enter step S2; Step S2: write a special script for mapping , providing a dedicated function interface for mapping; Step S3: Automatically collect network parameters of the SNN case to be mapped to obtain network modeling parameters; Step S4: Predict the memory and time consumption of the obtained network modeling parameters ; Step S5: Convert the obtained memory and time consumption prediction data into a mapping result.

In the method for constructing an SNN workload automatic mapper described in this embodiment, in the step S1, it is judged whether the CPU operation frequency and memory of the computing platform have changed, and if there is a change, continue to judge whether there is a storage file, and if there is a storage file, Then enter step S2; if there is no change, or there is no storage file, then quantify the design and store the obtained parameters in the form of a file, and then enter step S2, because the parameters obtained by the quantitative design have stability, it can be carried out in the form of a file. storage, and the obtained parameters can also be used as network modeling parameters; in the step S2, writing a mapping special script is the premise of the automatic design, and provides a mapping special function function interface for the automatic design; in the step S3, to The network parameters of the SNN cases to be mapped are automatically collected to obtain network modeling parameters. Since there is no need to create specific synapses to cause memory and time consumption, the time and space complexity of SWAM is greatly reduced; in the step S4 , predict the memory and time consumption of the obtained network modeling parameters. Due to the reusability of memory and time parameters and automatic design, the user only needs to perform quantization operations when using it for the first time, and then only needs to map the required mapping. The mapping work can be completed by writing a mapping special script for the case; in the step S5, the obtained memory and time consumption prediction data are converted into mapping results, and the reasonable mapping of the entire workload on the computing platform can be quickly and effectively deduced As a result, to ensure the stable and high-performance operation of the computing platform; and to avoid the extremely time-consuming operation and trial of the complete workload, which is of practical significance for the research and optimization of brain-like architectures.

The quantization design includes quantization of storage parameters and quantization of time parameters.

The values of the storage parameters are basically not affected by the computing platform and the SNN network, and are stable. The quantification method of the storage parameters is: through manual analysis and quantification, and at the same time, through the memory tool provided by the NEST emulator to map and correct the parameters. Specifically, the design idea of storage parameter quantization is derived from the work of Kunkel et al. The quantification method is specifically through manual analysis and quantification, and at the same time, the memory tool provided by the NEST simulator is used to map and correct the parameters.

The quantification of the time parameter includes the time quantification of communication and the time quantification of device, neuron, and synapse models.

For the quantization of time parameters, the time parameters have distinct platform characteristics, that is, the parameter value decreases with the enhancement of CPU performance. On the basis of the same platform, the value of the time parameter is basically stable. Therefore, for a specific computing platform, through the MPI quantization program and the benchmark quantization network, the time t _mpi (send _buffer , P) for a single communication and the time consumption for a single update of the model (t _s , t _poisson , t _spike , t _{n_ref} and t _{n_unref} ) are independently quantized.

Communication time quantification: The MPI quantization program follows the NEST communication mechanism, and quantifies the communication time consumption of the computing platform under different conditions by setting the amount of different sending tasks and the communication scale of different nodes.

Temporal quantification of device, neuron, and synapse models: Build benchmarked quantitative networks in python scripts, including various types of device, neuron, and synapse models. Taking the benchmark quantization network in Figure 2 as an example, the quantization network can realize the quantization of devices (PG and SD), neurons (gif_psc_exp, iaf_psc_alpha, iaf_psc_exp and iaf_psc_exp_ps), synapses (stdp_pl_synapse_hom_hpc, stdp_synapse and static, abbreviated as spshh, ss and static). ) to quantify the time consumed by a single update. PG connects four neuron populations, each of which is self-connected three times using spshh, ss, and static synapse types, respectively. The categorical quantification of t _spike , t _{n_ref} and t _{n_unref} parameters is achieved by flexibly adjusting the synaptic weights of PGs to neuron populations and the T _ref of neuron populations. d _t is set to 0.1, and the effect of different dd values on t _{n_ref} and t _{n_unref} is realized by changing the delay value in the connection.

In the NEST kernel, the time statistics function and the model update number statistics function respectively quantify the total time consumption and the total number of updates corresponding to various types of models in the benchmark network, and divide the two to obtain an average value. On this basis, several experiments were carried out to obtain the average value.

Since the storage parameters are stable and the time parameters of the same platform are also stable to a certain extent, the parameters obtained from the quantitative design are reusable and can be stored in the form of files as shown in Figure 3.

In the step S2, when writing a mapping special script, in order to facilitate the use of SWAM and ensure that the automatic design does not conflict with the use of the original NEST architecture, a mapping special script format is developed on the basis of retaining the original NEST network construction rules. , which provides a functional interface for automation design in the NEST core.

As shown in Figure 4, the mapping-specific script includes a prediction time module, a network building module, and a start prediction module.

In the prediction time module, when the simulation time information and the storage file address are input, it is also a sign to call the Auto_Collect module.

In the network construction module, since the network parameters to be collected are mainly distributed during the construction of the SNN network, the surface information of the creation (Create) and the connection (Connect) in the SNN network will be automatically collected in the NEST kernel by the network parameter collection module. At the same time, the pulse firing rate and refractory period information of the neuron group are also input during Create.

In the start prediction module, while inputting parameters, it is also a sign that the network parameter collection is completed and the prediction step and the mapping step are called. The input parameters include the node scale of the platform, the number of threads on a single node, the hardware memory provided by a single node, and the mapping threshold for storage occupancy.

The automation design is embedded in the NEST kernel as a whole, and consists of network parameter collection automation (Auto_Collect), prediction automation (Auto_Forecast) and mapping automation (Auto_Map).

The network parameter collection automation mainly collects related parameters of the SNN network. Specifically, in the step S3, when automatically collecting the network parameters of the SNN case to be mapped, a series of creations and connections are performed between the prediction time module and the start prediction module to construct the network, and the R of the neuron group _spike and T _ref are entered at creation time.

As shown in Figure 4, for the related parameters such as the scale and in-out degree of neurons and synapses corresponding to the creation and connection, there are two one-dimensional vectors: neuron type table v_n_type, synapse type table v_s_type and three two-dimensional vectors. Vector: Create table v_n_create, out-degree table v_n_out and in-degree table v_n_in for storage.

For v_n_type and v_s_type, data is pushed onto the stack whenever a new neuron type or synapse type is created (Create) or connected (Connect). The data to be pushed into the stack is specifically related data obtained by querying the storage file according to the model type numbers mod_id and dd, such as t _s , t _poisson , t _spike , t _{n_ref} and t _{n_unref} .

For v_n_create or v_n_out, v_n_in, the first dimension needs to be consistent with v_n_type or v_s_type respectively. Every time you create, v_n_create is pushed onto the stack. The data pushed into the stack is specifically N _n (i), R _spike and T _ref . Each time Connect, the v_n_out/v_n_in is pushed to the stack according to the neuron GID of the source/target group. The data pushed into the stack is the in-out degree of the source/target group of this connection, and the in-out degree is calculated in a similar way.

In the step S4, when storing and predicting the time consumption of the obtained network modeling parameters, by designing each load model function, the modeling parameter data obtained by quantitative design and obtained is processed to realize the prediction under the scale of each node. .

The prediction idea of the prediction automation module is embodied in the further instantiation of the typical SNN load model, including the storage load model, the calculation time load model and the communication time load model.

In terms of storage load model, the storage M of SNN is mainly composed of the storage M of _neurons and the storage M _synapse of synapses. The calculation formula of M is as follows:

M=M _neuron +M _synapse

=N _n *m _n +N _s *m _s (1)

Among them, M _neuron is determined by the number of neurons _{N n} and the memory mn consumed by the storage of a single neuron. While M _synapse is determined by the number of synapses N _s and the memory consumed by a single synapse storage m _s , N _s is determined by the connection topology and N _n .

As shown in Figure 5a, on the basis of formula (1), adding the initial cost M _o of the NEST system, the total storage M of each process becomes composed of three parts: M _neuron , M _synapse and M _o .

M _o is a serial initialization storage overhead that includes basic data structures required for NEST to operate, such as event scheduling and communication. For a fixed computing platform, the M _o value is a fixed value. The instantiation of the M _neuron and M _synapse formulas now includes the storage overhead of the neuron and synaptic infrastructure. The formula is as follows:

和

分别为每个进程为网络中所有、非本地和本地的每个神经元提供神经元型基础设施的存储开销。同理，

和

分别为每个进程为网络中所有、非本地目标和本地目标的每个神经元的提供突触型基础设施存储开销；

则为在本地没有目标神经元的神经元个数；Avg_in为目标神经元对应的入度；i代表着不同的神经元和设备模型类型；z则代表着不同的突触类型；j_post则代表着连接关系中的目标神经元群。Among them, N _n /P is the number of neurons allocated to each process, and N _D is the number of devices that each thread T needs to replicate and create.

and

Provides the storage overhead of neuron-type infrastructure for all, non-local, and local neurons in the network, per process, respectively. Similarly,

and

Provide synaptic infrastructure storage overhead for each neuron of all, non-local targets and local targets in the network separately for each process;

is the number of neurons without target neurons locally; Avg _in is the in-degree corresponding to the target neuron; i represents different neuron and device model types; z represents different synaptic types; j _post is Represents the target group of neurons in the connection.

In the computing time load model, the simulation computing time time _compute of SNN is mainly composed of neuron computing time time _neuron synapse computing time time _synapse . The time _compute computing formula is as follows:

time _compute =time _neuron +time _synapse

=N _n *t _n *Td _t +N _spike *Avg _out *t _s (4)

Among them, since the neuron update mode is time-triggered, Td _t =T _sim /d _t determines the total number of neuron updates in the simulation time. T _sim and d _t are the simulation time and resolution, respectively, and t _n is the time taken for a single update of the neuron. Since the synaptic update mode is event-triggered, N _spike *Avg _out determines the number of synaptic updates, and t _s is the time taken for a single synapse update.

As shown in Figure 5b, on the basis of formula (4), the Poisson device calculation time time _poisson is added to the time _compute calculation formula. time _compute becomes composed of three parts: time _poisson , time _neuron and time _synapse .

Poisson device update is a time-triggered update mode. Each Poisson update needs to go through all synapses, but whether the corresponding synapse is updated is determined by the random number generated by the pulse frequency set by Poisson. The time consumption mainly depends on the random number judgment after each post-synaptic pass. The time _poisson calculation formula is as follows:

Among them, N _poisson is the number of Poisson devices, and t _poisson is the unit time required for a single update of Poisson devices. j _pre represents the source device group in the connection relationship.

On the basis of inheriting the time-triggered update mode of formula (4), the neuron calculation time time _neuron considers the influence of the following two factors on the time t _n of a single neuron update, one is the different neuron types i, The second is the number of updates dd in a single time slice. The calculation formula of time _neuron is as follows:

Wherein, t _n in formula (4) is subdivided into three parts: t _{n_unref} , t _{n_ref} and t _spike . As shown in Fig. 6, t _{n_ref} is the time taken to calculate some basic formulas of neurons in the refractory period. t _{n_unref} is the time taken to calculate the voltage formula on the basis of t _{n_ref} for neurons that are not in the refractory period and have no pulses. t _spike is the neuron that generates the spike plus the time spent processing the spike on the basis of t _{n_unref} . The above analysis shows that the value of t _spike is the largest, followed by t _{n_unref} , and the smallest value of t _{n_ref} . Therefore, if the number of pulse excitations is more or less, t _n will be too large or too small, which will lead to deviations in the model prediction. Experiments show that further subdivision operations can fit the neuron computational time load model more efficiently.

Correspondingly, the number of neuron updates td is subdivided into _{cn_unref} , _{cn_ref} and N _spike . c _{n_ref} and c _{n_unref} are the update times of neurons in the refractory period and not in the refractory period and without spikes in the whole simulation stage, and N _spike is the number of spike excitations. The calculation formula is as follows:

Synapse calculation time time _synapse on the basis of inheriting the event-triggered update mode of formula (4), adding the adaptability of the formula, such as the effect of different synapse types z on _ts , according to the neuron group j _pre and synapse Type z divides the out-degree Avg _out , and each neuron group j _pre has its own number of spike firings N _spike . The time _synapse calculation formula is as follows:

It is worth noting that due to the task distribution method of NEST, the load balancing among the VPs is well guaranteed. Therefore, when modeling, this rule is also reflected in the formula of each load model.

In the communication time load model, the communication time time _mpi is mainly the time consumed by the pulse information transmitted by each node through MPI in the whole simulation process, which is the inevitable time cost of NEST to realize parallel computing. Therefore, in addition to time _compute , time _mpi is also a part of the time _sim of the entire simulation, as shown in Figure 5b.

There are many factors that affect the MPI communication time, such as the network bandwidth and the number of communicating nodes. This paper mainly considers the influence of the communication delay caused by the size of the sent information, the number of communication processes and the size of the bandwidth. The formula for calculating time _mpi is as follows:

time _mpi = t _mpi (send _buffer , P)*Td+t _bandwidth (9)

Among them, t _mpi (send _buffer , P) is the communication time consumed by a single transmission of information of the size of the send _buffer under the scale of P. t _bandwidth is the time effect of bandwidth limitation.

In the step S5, the method of converting the obtained memory and time consumption prediction data into a mapping result is as follows: designing a mapping function, processing the obtained prediction data, obtaining the mapping variables under the scale of each node, and storing the mapping threshold of the occupancy rate. And the optimal performance is used as the mapping standard, and the mapping result is deduced by the specified function.

Mapping idea: In terms of storage mapping standard, the storage occupancy _Rate (P) is used as the mapping variable. The formula for calculating Rate _m (P) is as follows:

Rate _m (P)=M(P)/M _hardware (10)

Among them, M(P) is the memory evenly allocated under the scale of P nodes, and M _hardware is the storage provided by the hardware of a single node. The mapping threshold TMM (Threshold for Mapping Memory) of the storage occupancy rate is used as the mapping standard. Take the node scale P _stable with storage occupancy rate TMM as the mapping result. The node scale of the cluster is above P _stable to ensure the stable operation of the platform. TMM is subject to user input.

In terms of performance mapping standard, the time _sim (P) that is amortized under the scale of P nodes is used as the mapping variable. The minimum mapping variable time _sim is used as the mapping criterion. Taking the node scale P _least (time _sim is the smallest node scale) with the best performance of the computing platform as the mapping result, the computing platform carrying the specific SNN case can run efficiently. The specific formula is as follows:

time _sim (P _least )=min{time _sim (P)} (11)

In the specific design, as shown in the mapping automation in Figure 4, the mapping function is designed according to the mapping idea of formulas (10) and (11), and the prediction data obtained by the Auto_Forecast module is processed by the Rate _m and time _sim functions. mapping variables (Rate _m (P) and time _sim (P)). Taking the mapping threshold of storage occupancy (TMM) and the optimal performance (minimum time) as the mapping criteria, the mapping results (P _stable and P _least ) are derived through the TMM and time_min functions.

In the connection stage of the automatic collection module of SWAM, the present invention only needs to collect macro information such as source groups, target groups and connection methods to meet the modeling parameter conditions, without creating specific synapses to cause memory and time consumption, greatly reducing the time and space complexity of SWAM. Compared with complicated and tedious manual node exploration or manual parameter collection, SWAM has an absolute time cost advantage, and the average time cost is reduced by 879 times.

Embodiment 2

Based on the same inventive concept, this embodiment provides an automatic mapper, and the principle of solving the problem is similar to the method for constructing the SNN workload automatic mapper, and the repetition will not be repeated.

This embodiment provides an automatic mapper, including:

The judgment module is used to judge whether the CPU operation frequency and memory of the computing platform have changed. If there is a change, continue to judge whether there is a storage file. If there is a storage file, enter the writing module; if there is no change or no storage file, then quantify Design and store the obtained parameters in the form of files, and then enter the writing module;

Write a module for writing mapping-specific scripts and providing mapping-specific function interfaces;

The acquisition module is used to automatically collect the network parameters of the SNN case to be mapped to obtain the network modeling parameters;

The prediction module is used to predict the memory and time consumption of the obtained network modeling parameters;

The mapping module is used to convert the obtained memory and time consumption prediction data into mapping results.

As will be appreciated by those skilled in the art, the embodiments of the present application may be provided as a method, a system, or a computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.

The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It will be understood that each flow and/or block in the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to the processor of a general purpose computer, special purpose computer, embedded processor or other programmable data processing device to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing device produce Means for implementing the functions specified in a flow or flow of a flowchart and/or a block or blocks of a block diagram.

These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory result in an article of manufacture comprising instruction means, the instructions The apparatus implements the functions specified in the flow or flow of the flowcharts and/or the block or blocks of the block diagrams.

These computer program instructions can also be loaded on a computer or other programmable data processing device to cause a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process such that The instructions provide steps for implementing the functions specified in the flow or blocks of the flowcharts and/or the block or blocks of the block diagrams.

Obviously, the above-mentioned embodiments are only examples for clear description, and are not intended to limit the implementation manner. For those of ordinary skill in the art, other different forms of changes or modifications can also be made on the basis of the above description. There is no need and cannot be exhaustive of all implementations here. And the obvious changes or changes derived from this are still within the protection scope of the present invention.

Claims (8)

1. a method for constructing SNN workload automatic mapper, is characterized in that, comprises: Step S1: Determine whether the CPU operation frequency and memory of the computing platform have changed. If there is a change, continue to judge whether there is a storage file. If there is a storage file, go to Step S2; The obtained parameters are stored in the form of files, and then enter step S2, and the quantization design includes the quantization of storage parameters and the quantization of time parameters; Step S2: write a mapping-specific script to provide a mapping-specific function function interface, and the mapping-specific script includes a prediction time module, a network construction module, and a start prediction module; Step S3: Automatically collect network parameters of the SNN case to be mapped to obtain network modeling parameters; Step S4: Predict the consumption of memory and time for the obtained network modeling parameters; Step S5: Convert the obtained memory and time consumption prediction data into a mapping result, and the method for converting the obtained memory and time consumption prediction data into a mapping result is: designing a mapping function, processing the obtained prediction data, and obtaining each node For the mapping variables under the scale, the mapping threshold and optimal performance of the storage _occupancy rate are used as the mapping criteria, and the mapping result is deduced through the specified function. variable, Rate _m (P) is calculated as follows: Rate _m (P)=M(P)/M _hardware Among them, M(P) is the memory evenly allocated under the scale of P nodes, M _hardware is the storage provided by the hardware of a single node, the mapping threshold TMM of the storage occupancy rate is used as the mapping standard, and the node size P _stable with the storage occupancy rate of TMM is used as the mapping standard. As a result of the mapping, if the node scale of the cluster is above P _stable , the platform can run stably, and TMM is subject to user input. 2. the method for constructing SNN workload automatic mapper according to claim 1, is characterized in that: the quantization method of described storage parameter is: by manual analysis and quantization, the memory tool that NEST emulator provides simultaneously surveys and corrects parameter. 3 . The method for constructing an SNN workload automapper according to claim 1 , wherein the quantification of the time parameter includes the time quantification of communication and the time quantification of device, neuron, and synapse models. 4 . 4. the method for constructing SNN workload automatic mapper according to claim 1 is characterized in that: when the network parameter of the SNN case of required mapping is automatically collected, between the prediction time module and the start prediction module, carry out A series of creations and connections to build the network, R _spike and T _ref of the neuron population are input at creation time. 5. The method for constructing an SNN workload automatic mapper according to claim 4, wherein: for the related parameters such as the scale, in-out degree, etc. of the neuron and synapse corresponding to the connection, the two one-dimensional vectors are composed of two one-dimensional vectors. : neuron type table v_n_type, synapse type table v_s_type and 3 two-dimensional vectors: create table v_n_create, out-degree table v_n_out and in-degree table v_n_in for storage. 6. the method for constructing SNN workload automatic mapper according to claim 1, is characterized in that: when carrying out storage and time consumption prediction to the network modeling parameter that obtains, by designing each load model function, handle by quantitative design and the obtained modeling parameter data to achieve predictions at various node scales. 7. the method for constructing SNN workload automatic mapper according to claim 1, is characterized in that: the method that the consumption prediction data of obtained memory and time is converted into mapping result is: design mapping function, process and obtain prediction data , get the mapping variables under the scale of each node, take the mapping threshold of storage occupancy and the optimal performance as the mapping criteria, and deduce the mapping result through the specified function. 8. An automatic mapper, comprising: The judgment module is used to judge whether the CPU operation frequency and memory of the computing platform have changed. If there is a change, continue to judge whether there is a storage file. If there is a storage file, enter the writing module; if there is no change or no storage file, then quantify Design and store the obtained parameters in the form of files, and then enter the writing module, and the quantization design includes the quantification of storage parameters and the quantification of time parameters; A writing module for writing a mapping-specific script, providing a mapping-specific function function interface, and the mapping-specific script includes a prediction time module, a network building module, and a start prediction module; The acquisition module is used to automatically collect the network parameters of the SNN case to be mapped to obtain the network modeling parameters; The prediction module is used to predict the memory and time consumption of the obtained network modeling parameters; The mapping module is used to convert the obtained memory and time consumption prediction data into mapping results, and the method for converting the obtained memory and time consumption prediction data into mapping results is: designing a mapping function, processing the obtained prediction data, and obtaining For the mapping variables under the scale of each node, the mapping threshold of storage _occupancy and the optimal performance are used as mapping criteria, and the mapping result is deduced through a specified function. As a mapping variable, Rate _m (P) is calculated as follows: Rate _m (P)=M(P)/M _hardware Among them, M(P) is the memory evenly allocated under the scale of P nodes, M _hardware is the storage provided by the hardware of a single node, the mapping threshold TMM of the storage occupancy rate is used as the mapping standard, and the node size P _stable with the storage occupancy rate of TMM is used as the mapping standard. As a result of the mapping, if the node scale of the cluster is above P _stable , the platform can run stably, and TMM is subject to user input.

Images