Microcosmic circuit model region decomposition modeling method and device

Abstract

The invention discloses a microscopic circuit model regional decomposition modeling method. The method comprises the following steps: dividing a microscopic equivalent circuit model of a large-scale electromagnetic problem into a plurality of sub-regions in a CPU (Central Processing Unit), and obtaining corresponding data matrixes; creating an order reduction matrix corresponding to the sub-regions in the GPU, and sequentially carrying out order reduction on the plurality of sub-regions until the order reduction of the plurality of sub-regions is finished: transmitting the data matrix of one of the sub-regions into the corresponding order reduction matrix to carry out model order reduction, and updating the data matrix in the reduced-order matrix, generating information data after the model of the sub-region is subjected to order reduction so as to transmit the correlation among the plurality of sub-regions, and updating the data matrix in the corresponding reduced-order matrix in combination with the information data when the model of the next sub-region is subjected to order reduction. According to the method, the large-scale electromagnetic problem is subjected to regional decomposition, the correlation between the sub-regions is reduced, and micro circuit model order reduction is performed on each sub-region, so that the problem that the capacity used by the large-scale electromagnetic problem in industrial application is far beyond the memory space of a CPU and a GPU is solved.

Description

Technical Field

[0001] The present invention relates to the field of micro circuit module order reduction, and in particular to a method and device for regional decomposition modeling of a micro circuit model. Background Art

[0002] Interconnect circuits are one of the key technologies of integrated circuits, which determine the data transmission speed and signal integrity performance of integrated circuits. As interconnect circuits continue to develop towards multi-layer stacking and high density, the model order required for interconnect circuit signal integrity analysis is growing rapidly. Some researchers have pioneered the use of microscopic circuit models based on physical equivalence, but the scalability still cannot meet the needs of large-scale circuit analysis. This project will conduct research on large-scale circuit regional decomposition modeling methods based on microscopic circuit models.

[0003] Mathematical equivalent reduction method based on circuit model Circuit models, such as Partial Element Equivalent Circuit (PEEC), use physically meaningful capacitors, inductors and resistors to describe electromagnetic fields, radiation, loss and other effects in integrated circuits. SPICE can be used to solve time domain and frequency domain responses, and are widely used in packaging, antennas, electromagnetic compatibility and other fields. In the early and middle stages of the transformation process, although the number of nodes decreases, the number of branches continues to increase, which leads to a significant increase in the model order. In addition, since only capacitor models can be processed, it is necessary to invert the large-scale potential coefficient matrix.

[0004] For large-scale integrated circuit analysis problems, the micro-circuit model avoids the bottleneck of using matrix inversion to obtain an accurate solution, effectively reducing the model simulation time. However, there are still the following scientific and technical problems: Scalability problems caused by limited GPU memory: When using GPU to accelerate the micro-circuit model algorithm, the amount of industrial-grade integrated circuit model data is much larger than the GPU memory, so each iteration requires a CPU-GPU data transfer. The CPU-GPU transmission speed is only one ten-thousandth of the GPU calculation speed, causing the GPU to wait for data for a long time and the acceleration algorithm to fail. Summary of the invention

[0005] The main purpose of the present invention is to overcome the above-mentioned defects in the prior art and to provide a method and device for regional decomposition modeling of a micro circuit model.

[0006] The present invention adopts the following technical solution:

[0007] A method for regional decomposition modeling of a micro circuit model, comprising the following steps:

[0008] In the CPU, a microscopic equivalent circuit model of a large-scale electromagnetic problem is divided into a plurality of sub-regions, and a data matrix corresponding to the sub-regions is obtained;

[0009] A reduced-order matrix corresponding to the sub-region is created in the GPU, and the order of several sub-regions is reduced in turn until the order reduction of several sub-regions is completed: the data matrix of one of the sub-regions is transferred to the corresponding reduced-order matrix for processing to achieve model order reduction of the sub-region, the data matrix in the reduced-order matrix is ​​updated, and information data is generated after the model reduction of the sub-region is completed to transmit the correlation between several sub-regions, and the data matrix in the corresponding reduced-order matrix is ​​updated in combination with the information data when the model of the next sub-region is reduced, and so on.

[0010] According to the characteristics of the microscopic equivalent circuit model in the three-dimensional rectangular coordinate system, a plurality of sub-areas are divided, and the data matrices corresponding to the sub-areas include: a coupling matrix PP between capacitors, a coupling matrix M between inductors between nodes and inductors between nodes, and a coupling matrix M between inductors on nodes and inductors between nodes. pm , the coupling matrix M between the inductance on the node and the inductance on the node pp , the label mapping matrix and the matrix M pmM , the matrix M pmM The coupling values ​​between the inductances between all nodes in the sub-region and the inductances on all nodes are stored, and the label mapping matrix stores the mapping relationship between the node labels before the sub-region is divided and the node labels after the division.

[0011] The microscopic equivalent circuit model is divided into several sub-regions, and the data matrix of the sub-region also includes a label matrix, which stores the starting label and the ending label of the node of each sub-region. When reducing the order of the sub-region, the data matrix of the corresponding sub-region is directly selected according to the label matrix.

[0012] A transition matrix is ​​also created in the GPU to store part of the data required for the sub-region model reduction. The model reduction of the sub-region includes:

[0013] The absorption condition value △ of each node in the sub-area is calculated using the MIN method, and the node with the smallest absorption condition value is absorbed first. After absorbing this node and updating all the above matrices, other nodes calculate the condition values ​​using the MIN method respectively, and iteratively find the next node with the smallest absorption condition value to be absorbed. The absorption condition value △ calculation formula is as follows:

[0014]

[0015] Among them, m II Represents the self-coupling value of the newly added inductor, m SS Represents the self-coupling value of the original inductor, m SIRepresents the mutual coupling value between the newly added inductor and the original inductor, P II Represents the self-coupling value of the newly added capacitor element, P SS Represents the self-coupling value of the original capacitor element, P SI Represents the mutual coupling value between the newly added capacitor element and the original capacitor element, Represents the maximum operating frequency of the microwave device;

[0016] Each time a node is absorbed, a loop is entered: first, according to the needs of each iteration, the coupling data values ​​of the node k to be absorbed and several nodes i connected to it are taken out from the data matrix and placed in the transition matrix X RM middle:

[0017]

[0018] PP i represents the capacitor autocoupling matrix, M i represents the inductor autocoupling matrix, M pmi Represents the coupling matrix of the inductance on several nodes i and the inductance between the other nodes in the sub-region, M ppi represents the inductance autocoupling matrix on the above-mentioned nodes i, M pmk Represents the inductance coupling matrix between the inductance at node k and the rest of the nodes in the sub-area.

[0019] Each time a node is absorbed, the data matrix of the sub-region is compressed, specifically: the last row and the last column of data in the data matrix corresponding to the sub-region are moved to the data position corresponding to the absorbed node.

[0020] After the model order reduction of the sub-region is completed, the information data generated is as follows:

[0021]

[0022] Specifically, x pk 、x mk 、x pmk They represent the update information generated by the i-th group of sub-regions absorbing the k-th node, n represents the total number of nodes i connected around this node k, including the coupling data between the capacitor and all capacitors, the mutual coupling data of the inductance between nodes, and the coupling data between the inductance on the node and the inductance between nodes; the update information X generated by the order reduction of each sub-region i A matrix X is formed, and the matrix X is used to update the data matrices corresponding to all the sub-regions, where N represents the total number of sub-regions that need to be absorbed.

[0023] When the model of the next sub-region is reduced in order, the data matrix in the reduced-order matrix corresponding to the sub-region is updated in combination with the information data, as follows:

[0024] PP=PP+c 0 x p T ·x p

[0025] M=M+c 1 x m T ·x m

[0026] Mpp=Mpp+c 2 x pm T ·x p +c 3 x pm T ·x pm

[0027] Mpm=Mpm+c 4 x pm T ·x m +c 5 x p T ·x m ;

[0028] Among them, c 0 、c 1 、c 2 、c 3 、c 4 、c 5 is the constant when the microcircuit model is transformed, x p Represents the mutual coupling information between capacitor elements, x m Represents the mutual coupling information between inductor components, x pm Indicates the mutual coupling information between the inductance on the PpI branch and the inductance between nodes.

[0029] The model reduction of the sub-region includes merging parallel branches. When the parallel branches are merged, all mutual impedances in the circuit need to be updated. The parallel circuit elements and the related mutual coupling are listed as follows:

[0030] The self-impedance of the parallel branch S is:

[0031]

[0032] The mutual impedance between the parallel branch S and the remaining inductance branch or other PpI branches is:

[0033] z' RS =[(y SS +y SI ) RS+(y II +y SI ) RI ] / yt;

[0034] z' OS =[(y SS +y SI ) OS +(y II +y SI ) OI ] / yt;

[0035] where y SS ,y II They represent the self-coupling admittance of one of the two parallel branches, y SI represents the mutual coupling admittance of the two parallel branches, z RS represents the coupling impedance between one of the two PPI branches and the other common branches that do not participate in the combined PPI branch, z OS represents the coupling impedance of one branch to other PpI branches, z RI It represents the coupling impedance between the other branch of the two parallel branches and other common branches that do not participate in the combined PpI branch, z OI represents the coupling impedance between the other branch of the two parallel branches and the other PpI branches, z' RS It represents the coupling impedance between the two PPI branches after merging and other common branches that do not participate in the merging of PPI branches, z' OS Indicates the coupling impedance with other PpI branches after the two PpI branches are merged.

[0036] The mutual impedance between the other PpI branches and the remaining inductance branches is:

[0037]

[0038] Among them, Z' OO represents the self-coupling after the other PpI branches are merged, Z OO represents the self-coupling of other PpI branches before merging, Z' OR It represents the coupling between other PpI branches and common branches after merging PpI branches, Z OR represents the coupling between other PpI branches and common branches before merging the PpI branches, Z' RR represents the self-coupling of other common branches after merging the PpI branch, Z RR Represents the self-coupling of other common branches before merging the PpI branch.

[0039] A micro circuit model regional decomposition modeling device, comprising:

[0040] A sub-region division module divides the microscopic equivalent circuit model of a large-scale electromagnetic problem into several sub-regions in the CPU to obtain a data matrix corresponding to the sub-regions;

[0041] An order reduction module is used to create an order reduction matrix corresponding to the sub-region in the GPU, and to reduce the order of several sub-regions in turn until the order reduction of several sub-regions is completed: a data matrix of one of the sub-regions is transferred into the corresponding order reduction matrix for processing to achieve model order reduction of the sub-region;

[0042] The data update module generates information data after the model is reduced to transmit the correlation between the sub-regions. When the model of the next sub-region is reduced, the information data is combined to update the data matrix in the corresponding reduced-order matrix. After the update is completed, the nodes in another sub-region are continuously absorbed, and so on.

[0043] It can be seen from the above description of the present invention that, compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. The regional decomposition modeling method used in the present invention is very suitable for the model degradation of large-scale micro circuit models. The large-scale micro circuit model is decomposed into multiple strip sub-regions so that the size of its data matrix meets the limited memory size of the GPU, which greatly enhances the scalability of the micro circuit model degradation algorithm.

[0045] 2. By utilizing the symmetric relationship and low correlation between decomposed regions, the corresponding data of the sub-regions in several vectors in the sub-region data can be used to iteratively update the data in another matrix without using the full region data to calculate and update, which greatly reduces the number of data transmissions between the CPU and GPU and the total execution time of the algorithm. This is conducive to the use of high-performance GPU parallel computing technology in micro-circuit models to solve and analyze the problem of large-scale interconnect packaging circuit reduction.

[0046] 3. The order reduction of the micro-circuit model starts from the last sub-region and iterates from back to front. First, the last sub-region is updated, and then the previous sub-region is updated. This way of updating nodes can keep the data storage location of the matrix unchanged to the greatest extent, which can greatly reduce the complexity of verifying the correctness of the calculation results and improve the reusability of the program. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is the overall flow chart of the algorithm implementation of the present invention;

[0048] Figure 2 is the schematic diagram of the parallel plate capacitor used as the numerical calculation example;

[0049] FIG3 is a schematic diagram of a parallel plate capacitor microcircuit modeling region by region;

[0050] Figure 4 It is a schematic diagram of the sub-region segmentation of the corresponding inductive and capacitive data matrices after the parallel plate capacitor microcircuit model is modeled during the algorithm implementation, as well as a schematic diagram of the data transmission relationship between the CPU and GPU for the related inductive and capacitive data matrices;

[0051] Figure 5 It is a schematic diagram of combined parallel capacitors;

[0052] Figure 6 It is a schematic diagram of the process of optimizing the data matrix data storage location during the node absorption process of regional decomposition;

[0053] 7 is a schematic diagram of the process of optimizing the data matrix for storing the newly generated element data in the process of regional decomposition node absorption;

[0054] Figure 8 It uses sub-region data to update information, and one data transmission can iteratively update other sub-region data schematics;

[0055] Fig. 9 It is a schematic diagram of the location of the updated information data vector and the coefficient vector reintegrated and saved when the parallel capacitors are merged;

[0056] Fig.10 It is a schematic diagram of the updated information data vector and the coefficient vector reintegrated and saved position when the parallel inductance is merged;

[0057] Fig.11 The iterative absorption node is a flag matrix indicating whether there is a merged parallel inductance and the row number stored;

[0058] Fig.12 It is a data curve calculated by using the overall method and the regional method in the present invention after absorbing one node;

[0059] Fig.13 It is the data difference curve calculated by the overall method and the regional method in the present invention after absorbing one node;

[0060] Fig.14 It is a data curve calculated by using the overall method and the regional method in the present invention after absorbing multiple nodes;

[0061] Fig.15 It is a data difference curve calculated by using the overall method and the regional method in the present invention after absorbing multiple nodes;

[0062] Fig.16 It is a data curve calculated by using the overall method and the regional method in the present invention after absorbing two nodes and merging a pair of inductors;

[0063] Fig.17 It is a data difference curve calculated by using the overall method after absorbing two nodes and merging a pair of inductors and the data calculated by the regional method in the present invention.

[0064] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. DETAILED DESCRIPTION

[0065] The present invention is further described below through specific implementation modes.

[0066] The purpose of the present invention is to realize the regional modeling of large-scale micro circuit models, and to provide a strip region decomposition method to reduce the use of GPU memory space. By utilizing the low correlation between sub-regions after regional decomposition, a very small number of data vectors are used to iteratively update the data of other sub-regions, ensuring that the degradation results of the sub-region model are consistent with the way in which all data are transferred into the GPU, and there is no loss of precision and error.

[0067] A method for regional decomposition modeling of a micro circuit model of the present invention, see Figure 1 This is the overall flow chart of the micro-circuit model reduction. The overall process consists of three main parts. The first part is to process the micro-circuit model and its data matrix in different regions. The second part is the main part of the micro-circuit model reduction. The third part is to update the data matrix. The details are as follows:

[0068] S1 divides the microscopic equivalent circuit model of the large-scale electromagnetic problem into several sub-regions in the CPU, and obtains the data matrix corresponding to the sub-region, that is, each sub-region has a corresponding data matrix.

[0069] According to the characteristics of the microscopic equivalent circuit model in the three-dimensional rectangular coordinate system, several sub-areas are divided. The data matrices corresponding to the sub-areas include: the coupling matrix PP between capacitors, the coupling matrix M between the inductances between nodes and the inductances between nodes, and the coupling matrix M between the inductances on the nodes and the inductances between the nodes. pm , the coupling matrix M between the inductance on the node and the inductance on the node pp , the label mapping matrix and the matrix M pmM , the matrix M pmM The coupling values ​​between the inductances between all nodes in the sub-region and the inductances on all nodes are stored, and the label mapping matrix stores the mapping relationship between the node labels before the sub-region is divided and the node labels after the division.

[0070] Taking a parallel plate capacitor as an example, a parallel plate capacitor is divided into two upper and lower plates, with a total of 32 nodes. Each node has an inductor and a capacitor under it, and an inductor is connected between nodes, with a total of 50 inductors. In a standard three-dimensional rectangular coordinate system, there are three coordinates, x, y, and z. The placement of a parallel plate capacitor in the rectangular coordinate system can be divided into two sub-regions about the x-axis. The part with x<0 is a sub-region named Domain0, indicating sub-region 0, and the part with x>0 is divided into another sub-region named Domain1, indicating sub-region 1.

[0071] See also Figure 2 , Figure 2 For the numerical calculation example used, the size of the capacitor is 4×4 mm, the spacing between the upper and lower layers is 0.2 mm, and the medium is air medium epr=3.

[0072] Please refer to Figure 3, which shows the micro circuit model generated by the above parallel plate capacitor. Figure 3 (a) shows the PEEC circuit model of the upper plate, and Figure 3 (b) shows the PEEC circuit model of the lower plate. The node number range of the upper plate is 0-15, and the node number range of the lower plate is 16-31. There is an inductor and a capacitor element connected under each node, and an inductor element is connected between every two nodes. The inductor element number range corresponds to two sub-areas, 0-24 and 25-49 respectively.

[0073] According to the characteristics of the microscopic circuit model in the three-dimensional rectangular coordinate system, the node area is divided into multiple sub-areas and then a label mapping matrix x containing the mapping relationship between the new and old nodes is generated. P_N2O 、x P_O2N 、x M_N2O 、x M_O2N , x P_N2O Represents the mapping matrix from the new label to the old label of the node where the capacitor is located, x P_O2N Represents the mapping matrix from the old label to the new label of the node where the capacitor is located, x M_N2O Represents the mapping matrix from the new number of the node where the inductor is located to the old number, x M_O2N Represents the mapping matrix from the old label to the new label of the node where the inductor is located. Use the label mapping matrix to update other data matrices, such as the node label matrix X at the two endpoints of the inductor element NODE , assuming there are m nodes:

[0074]

[0075] Among them, n i0 Indicates the node numbers at both ends of the i-th inductor, and the labels of the two ends of the edge passing through it are X NODE_SIDE :

[0076]

[0077] Among them, snd i0 Indicates the labels at both ends of the side node through which the i-th inductor passes. Generate an inductive coupling matrix with row labels for all nodes and column labels for the inductor components in the sub-region. P_N2O 、x P_O2N 、x M_N2O 、x M_O2N Match the data matrices PP, M, M respectively pp 、M pm 、M pmM , thus dividing the overall data matrix into two data matrices corresponding to the sub-regions, and the start and end labels of each sub-region are separately stored in a label matrix. When using related data, the data matrix corresponding to the sub-region can be directly selected according to the label matrix. The nodes 0-15 and the components with inductance components labeled 0-24 of the numerical example are divided into sub-region 0, and the nodes 16-31 and the components with inductance components labeled 25-49 are divided into sub-region 1.

[0078] S2 creates a reduced-order matrix corresponding to the sub-region in the GPU, and reduces the order of several sub-regions in turn until the reduction of several sub-regions is completed: the data matrix of one of the sub-regions is transferred to the corresponding reduced-order matrix for processing to achieve model reduction of the sub-region, the data matrix in the reduced-order matrix is ​​updated, and information data is generated after the model reduction of the sub-region is completed to convey the correlation between several sub-regions, and the data matrix in the corresponding reduced-order matrix is ​​updated in combination with the information data when the model of the next sub-region is reduced, and so on.

[0079] This step includes initializing the PEEC circuit model data matrix, with the purpose of transferring each sub-region data to the GPU global memory (Global Momery) for standby. The absorption condition value △ of each node in the sub-region is calculated using the MIN method, and the node with the smallest absorption condition value is absorbed first. After absorbing this node and updating all the above matrices, the other nodes calculate the condition values ​​using the MIN method respectively, and iterate to find the next node with the smallest absorption condition value to be absorbed. The absorption condition value △ calculation formula is as follows:

[0080]

[0081] Among them, m II Represents the self-coupling value of the newly added inductor, m SS Represents the self-coupling value of the original inductor, m SI Represents the mutual coupling value between the newly added inductor and the original inductor, P II Represents the self-coupling value of the newly added capacitor element, P SSRepresents the self-coupling value of the original capacitor element, P SI Represents the mutual coupling value between the newly added capacitor element and the original capacitor element, Represents the maximum operating frequency of the microwave device;

[0082] Each time a node is absorbed, a loop is entered: first, according to the needs of each iteration, the relevant coupling data values ​​of the node k to be absorbed and several nodes i connected to it are taken out from the data matrix and placed in the transition matrix X RM middle:

[0083]

[0084] PP i represents the capacitor autocoupling matrix, M i represents the inductor autocoupling matrix, M pmi Represents the coupling matrix of the inductance on several nodes i and the inductance between the other nodes in the sub-region, M ppi represents the inductance autocoupling matrix on the above-mentioned nodes i, M pmk Represents the inductance coupling matrix between the inductance at node k and the rest of the nodes in the sub-area.

[0085] The model reduction of the sub-domain includes merging parallel branches, refer to Figure 5 , (a) is the impedance before merging the parallel branches, (b) is the impedance after merging the parallel branches, and (c) is the impedance after merging the parallel capacitor branches.

[0086] Consider the parallel branches S and I, which can be a pair of PpI branches (capacitor and inductor elements in series below the node) or a pair of inductor branches, merged into branch S. The elements in the two branches are not only coupled to each other, but also to the rest of the circuit elements. The PpI branch and the inductor branch in the rest of the circuit are distinguished by the subscripts 0 and R, respectively. When the parallel branches are merged, all the mutual impedances in the circuit need to be updated.

[0087] For the sake of simplicity, the parallel circuit elements and the related mutual coupling can be directly listed as follows: When the parallel branches are merged, all the mutual impedances in the circuit need to be updated, so the parallel circuit elements and the related mutual coupling are listed as follows:

[0088] The self-impedance of the parallel branch S is:

[0089]

[0090] The mutual impedance between the parallel branch S and the remaining inductance branch or other PpI branches is:

[0091] z' RS =[(y SS +y SI )RS +(y II +y SI ) RI ] / yt;

[0092] z' OS =[(y SS +y SI ) OS +(y II +y SI ) OI ] / yt;

[0093] where y SS ,y II They represent the self-coupling admittance of one of the two parallel branches, y SI represents the mutual coupling admittance of the two parallel branches, z RS represents the coupling impedance between one of the two PPI branches and the other common branches that do not participate in the combined PPI branch, z OS represents the coupling impedance of one branch to other PpI branches, z RI It represents the coupling impedance between the other branch of the two parallel branches and other common branches that do not participate in the combined PpI branch, z OI represents the coupling impedance between the other branch of the two parallel branches and the other PpI branches, z' RS It represents the coupling impedance between the two PPI branches after merging and other common branches that do not participate in the merging of PPI branches, z' OS Indicates the coupling impedance with other PpI branches after the two PpI branches are merged.

[0094] The mutual impedance between the other PpI branches and the remaining inductance branches is:

[0095]

[0096] Among them, Z' OO represents the self-coupling after the other PpI branches are merged, Z OO represents the self-coupling of other PpI branches before merging, Z' OR It represents the coupling between other PpI branches and common branches after merging PpI branches, Z OR represents the coupling between other PpI branches and common branches before merging the PpI branches, Z' RR represents the self-coupling of other common branches after merging the PpI branch, Z RR Represents the self-coupling of other common branches before merging the PpI branch.

[0097] In the process of sub-region reduction of the present invention, each time a node is absorbed, the data matrix of the sub-region is

[0098] PP, M, M pp 、M pm 、M pmM Compression is performed: the last row and column of data in the data matrix corresponding to the sub-region are moved to the data position corresponding to the absorbed node, further saving GPU storage space. Importantly, absorbing the nodes of sub-region 0 first will cause data processing and verification to be extremely cumbersome and the program reusability to be poor. Therefore, the verification time cost is greatly reduced by absorbing the nodes of sub-region 1 first.

[0099] The third part of the flowchart is to use the generated information data to transmit the correlation between data blocks after iteratively absorbing all the nodes in sub-region 1 (Domain1), and then convert the data matrices PP, M, M corresponding to sub-region 0 (Domain0) into pp 、M pm 、M pmM Passed from the CPU to the GPU, the data is updated using the X matrix:

[0100]

[0101] Specifically, the X matrix stores the impact of each node update on the inductance and capacitance components of all nodes in the global region, that is, information data. The X matrix is ​​divided into multiple related information data corresponding to the sub-regions, including storage PP, M, M pp 、M pm 、M pmM The relevant information data corresponding to the nodes in the sub-region 0 (Domain0) is required to store the relevant update information by opening up a data matrix space of M×5 rows and k_link_num_i×PPmatsize_subD columns to store the relevant update information. Each iteration update (Domain0) takes out the corresponding node information and the corresponding coupling data between the inductive and inductive elements at different positions and the coupling data between the capacitive and capacitive elements from the corresponding row, where k_link_num_i represents the number of other nodes connected around the i-th node when absorbing it. Taking the coupling value of the capacitive element under the node and the capacitive element under other nodes as an example, PPmatsize_subD represents the number of couplings between the i-th node and the nodes in the sub-region 0 (Domain0) plus additional space. The i-th row number represents the absorption of the i-th node, and the column number represents the coupling value between the k_link_num nodes connected around this node and all the nodes in the sub-region 0 (Domain0).

[0102] When the model of the next sub-region is reduced in order, the data matrix in the corresponding reduced-order matrix is ​​updated in combination with the information data. Specifically, taking the coupling between the capacitor element under a node and the capacitor elements under other nodes as an example, after extracting the essence of the node absorption algorithm, it can be found that the data matrix in the corresponding reduced-order matrix is ​​updated in combination with the information data and generated using the following formula:

[0103] PP=PP+c 0 x p T ·x p

[0104] M=M+c 1 x m T ·x m

[0105] Mpp=Mpp+c 2 x pm T ·x p +c 3 x pm T ·x pm

[0106] Mpm=Mpm+c 4 x pm T ·x m +c 5 x p T ·x m ;

[0107] Among them, c 0 、c 1 、c 2 、c 3 、c 4 、c 5 are all constants when the microscopic circuit model is transformed, x p Represents the mutual coupling information between capacitor elements, x m Represents the mutual coupling information between inductor components, x pm Indicates the mutual coupling information between the inductance of the PpI branch and the inductance between nodes. Figure 5 , and They represent the autocoupling values ​​of the two capacitive elements before merging the parallel branches, Represents the mutual coupling value of two capacitive elements before merging the parallel branches.

[0108] The following describes the specific method of area division in the specific process of node absorption of the present invention in steps.

[0109] See also Figure 4 , taking the data matrix of capacitor elements as an example, the left side represents the operations performed in the CPU. Assuming that a total of N nodes need to be absorbed, the global area is now divided into two sub-areas, namely sub-area 0 and sub-area 1. First, the nodes in sub-area 1 are absorbed, and then the nodes in sub-area 0 are absorbed. After the nodes in sub-area 1 are absorbed, information data will be generated, and then this information data will be used to update sub-area 0.

[0110] Figure 4 The figure shows an important part of domain decomposition modeling, which is how to transfer the data matrix from the CPU to the GPU after label matching. 401 represents the data matrix corresponding to sub-domain 0 (sub-Domain0), 402 represents the data matrix corresponding to sub-domain 1 (sub-Domain1), and 403 is the sub-domain 1 that needs to absorb nodes. Assume that there are M nodes to be absorbed in sub-domain 1. In the process of implementing the sub-domain algorithm, the mutual influence between the nodes in the global domain needs to be considered. Specifically, the capacitor element needs to consider not only the coupling between it and other nodes in the sub-domain, but also the coupling between it and all nodes contained in sub-domain 0 (sub-Domain0). In this way, after all the nodes are absorbed, the change information of each node in the global domain is taken into account, and then the coupling values ​​between the surrounding elements generated by the nodes absorbed by this sub-domain in each round of iteration and all elements of sub-domain 0 are stored in the matrix update_PreDomain_vector_mat of information data that has been allocated space in the GPU in advance. After absorbing sub-domain 1, the update information generated by this sub-domain 1 can be used to realize data transmission between the CPU and the GPU only once, and all update information is generated in the GPU. Then, all data matrices corresponding to sub-domain 0, including GPU_PPmat_subD, GPU_Mmat_subD, etc., are transferred from the CPU to the GPU. According to the order of eliminating nodes, according to the row number of update_PreDomain_vector_mat, nodes and matrices are updated one by one.

[0111] See also Figure 6, This figure shows the optimization of the storage method of matrix elements in the process of node elimination in different regions in GPU. Specifically, 601 represents the autocoupling value and mutual coupling value of node k to be absorbed, the node_k_subD row represents the mutual coupling value between node k and all nodes in the global region, and the row number stored in the matrix needs to be specified in the corresponding position of sub-region 1, because the starting position of sub-region 1 should be the next position of the end position of sub-region 0, 602 represents the last column and the last row of sub-region 1, indicating the coupling value between the node with the last number in sub-region 1 and other nodes, end_Ni_subD is the node number of the last row of sub-region 1, indicating the capacitance element coupling value between the node and all nodes in the global region, and the end_Ni column represents the node number of the last row of sub-region 1, which is also the node number of the last column of the global region, indicating the capacitance element coupling value between the node and all nodes in sub-region 1. By analogy, the next iteration absorbs the node at the second position, absorbing the k1 node, then it is necessary to fill the data of the end_Ni_subD-1 row and end_Ni column to the row and column where the k1 node is located, and 603 represents the coupling value of the next absorption node with all the capacitor elements of sub-region 1 and the coupling value with the global region node. With the increase of absorption nodes, this data matrix can be optimized very little, plus a small amount of data matrix for transmitting information, which greatly saves the repeated transmission of data between the CPU and GPU, and greatly reduces the GPU memory resources occupied.

[0112] When performing node elimination and updating information, the completeness of the information must be ensured, one part is Mpmmat_subD, and the other part is MpmMmat_subD. The Mpmmat_subD matrix stores the coupling value of the inductance element on the node and the inductance element between the two nodes in sub-area 1, with the inductance element on the node as the core.

[0113] Please refer to FIG. 7 , FIG. 7 (a) shows that the coupling values ​​between the capacitor elements under the nodes in the sub-region circuit model and the capacitor elements under other nodes are stored in the matrix GPU_PPmat_subD. The matrix stores the coupling values ​​between the capacitor elements on all nodes in the global region and the capacitor elements between all nodes in the sub-region 1 (Domain1). In this step, new capacitive element data will also be generated. Region 701 indicates the location where the coupling values ​​between the original capacitor elements are stored, region 702 indicates the location where the coupling values ​​between the newly generated capacitor elements and the capacitor elements on all nodes in the sub-region 1 (Domain1) are stored, region 703 indicates the location where the coupling values ​​between the newly generated capacitor elements and the capacitor elements on all nodes in the global region are stored, and region 704 indicates the location where the coupling values ​​between the newly generated capacitor elements and themselves and the coupling values ​​between each other are stored.

[0114] FIG7(b) is a data storage matrix GPU_MpmMmat_subD for the coupling data between the inductance elements below the nodes and the inductance elements between the nodes in the sub-region circuit model. This matrix is ​​generated to store the coupling values ​​between the inductance elements on all nodes in the global region and the inductance elements between all nodes in sub-region 1 (Domain1). New inductive and capacitive element data are also generated in this step. Region 705 indicates the location where the original data is stored, region 706 indicates the location where the coupling values ​​between the newly generated inductance elements and the inductance elements on all nodes in the entire region are stored, region 707 indicates the location where the coupling values ​​between the newly generated inductance elements and the inductance elements on all nodes within the sub-region are stored, and region 708 indicates the location where the coupling values ​​between the newly generated inductance elements and themselves and between each other are stored.

[0115] See also Figure 8 After the node absorption, the present invention stores all the updated information data in a matrix, and updates the sub-region 0 data matrix in the GPU. Specifically, according to the above-mentioned sub-region division method, the data matrix is ​​divided into two parts, Domain0 and Domain1, and several data matrices related to Domain1 are transferred to the GPU for standby, and the node absorption of Domain1 is performed, and the related data matrices inside Domain1 are updated. At the same time, a data vector (vector) for updating Domain0 is generated, and these data vectors are collected and arranged according to certain rules, so that the corresponding data can be quickly found when the Domain0 data matrix is ​​updated in the future. At the same time, the number of data transmissions between the CPU and the GPU is saved. Only one data transmission is required to use the parallel computing performance advantage of the GPU to reduce the circuit model. The matrix for storing data vectors is update_PreDoamin_vector_mat, and the coefficient matrix for storing correlation coefficients is

[0116] update_PreDoamin_coeff_mat, after absorbing the nodes of Domain1, update the data of Domain0, pass the relevant data matrix of Domain0 to the GPU, and use the above update_PreDoamin_vector_mat and

[0117] The data stored in update_PreDoamin_coeff_mat updates the data in Domain0, and finally the updated Domain1 and Domain0 data are transferred out of the GPU.

[0118] The following describes the update_PreDoamin_vector_mat and update_PreDoamin_coeff_mat matrix designs. Fig. 9 , Pc, Pcm, Mc represent the capacitive element vectors under the nodes in the updated nodes respectively. The coefficient matrix distinguishes nodes according to rows. Several rows are stored for absorbing several nodes. Pc, Pcm, Mc related to each node in the iteration are grouped together. The row label is node. When reading data, you only need to read the row label, and then you can find the relevant data for node absorption and update. For the column, because each update requires k_link_num connected nodes, the column is divided into k_link_num groups. Each group represents the data coupled between one of the nodes connected to the node and the remaining nodes. The size is maxLC, the maximum row label is M×3, and the maximum column label is k_link_num×maxLC. Below is

[0119] update_PreDoamin_coeff_mat matrix, which stores the coefficient constants needed. The rows represent the two sets of data stored, coeffmat1 and coeffmat2, which are the coefficient constants needed by different data matrices. Two rows form a node's data. If M nodes are absorbed, it will be M×2 rows. The columns are the k_link_num data corresponding to each coeffmat.

[0120] See also Fig.10 When merging parallel inductors, we also need to create a coefficient matrix and a matrix to store information vectors. However, unlike merging parallel capacitors on nodes, we only need to use two vectors, Pc and Mc, so the maximum row number is M×2, and the maximum column number is also k_link_num×maxLC.

[0121] The rows of the coefficient matrix represent the rounds in which inductance absorption is required. Inductance absorption is not necessarily required in every iteration. Depending on the changes in the circuit structure, there may be one or more pairs of inductance absorption. The columns of each row represent the constant coefficients corresponding to each pair of inductances in this round of absorption, for a total of ind_num_comb.

[0122] See also Fig.11, whether it is necessary to merge the parallel inductance components after the node absorbs the parallel capacitance components. Since the circuit structure is constantly changing during the node absorption process, it is impossible to determine when we need to merge the parallel inductance components, but the capacitance components need to be merged every time. Therefore, according to each change in the circuit structure, if the circuit structure after absorbing the capacitive node produces the situation that the parallel inductance components need to be merged as shown in the figure, such as 1101 in the figure, then the flag bit is set to 0. In the next round of node absorption, if the situation of merging inductance components is encountered again, it is set to 1 as shown in 1102, and so on. At the same time, in the next round of node absorption, the corresponding data can be found in the merged inductance component update information data and coefficient data matrix according to this value. The design iteration update sequence flag matrix is ​​as follows:

[0123]

[0124] The storage updates the location information of the inductor and capacitor corresponding to the node.

[0125] This embodiment performs a regional algorithm simulation on the parallel class capacitor, compares the node absorption data after regional division with the original mode data, and presents the data curve and the difference curve on the Figure 12-Figure 17 .

[0126] Fig.12 This is the updated numerical result of the parallel plate capacitor numerical example after absorbing one node. There are 256 data points in total, and it can be seen that the data results calculated by the original method of overall calculation are completely consistent with the data calculated by the regional method used in the present invention.

[0127] Fig.13 After absorbing one point, we can see the difference between the data results calculated by the original method of overall calculation and the data results calculated by the regional method used in the present invention. We can see that the difference values ​​of 256 points are all 0.

[0128] Fig.14 The numerical results of the parallel plate capacitor numerical calculation example after absorbing two nodes are updated. There are 256 data points in total, and it can be seen that the data results calculated by the original method of overall calculation and the data calculated by the regional method used in the present invention are completely consistent.

[0129] Fig.15 It is the difference between the data result calculated by the original method of overall calculation after absorbing 2 points and the data result calculated by the regional method used in the present invention. It can be seen that the difference values ​​of 256 points are all 0.

[0130] Fig.16This is the numerical result of the parallel plate capacitor numerical example after absorbing 2 nodes and merging a pair of parallel inductors. There are 625 data points in total, and it can be seen that the data calculated by the original method of overall calculation is completely consistent with the data calculated by the regional method used in the present invention.

[0131] Fig.17 For the numerical example of parallel plate capacitance, after absorbing 2 nodes and merging a pair of parallel inductors, the difference between the data results calculated using the original method of overall calculation and the data results calculated using the regional calculation method used in the present invention shows that the difference values ​​of 625 points are all 0.

[0132] Based on this, the present invention also proposes a micro circuit model regional decomposition modeling device, which is used to execute the above-mentioned micro circuit model regional decomposition modeling method, comprising:

[0133] The sub-region division module divides the microscopic equivalent circuit model of the large-scale electromagnetic problem into several sub-regions in the CPU to obtain the data matrix corresponding to the sub-regions.

[0134] The order reduction module creates a reduced-order matrix corresponding to the sub-region in the GPU, and reduces the order of several sub-regions in turn until the order reduction of several sub-regions is completed: the data matrix of one of the sub-regions is transferred to the corresponding reduced-order matrix for processing to achieve model order reduction of the sub-region.

[0135] Update data module. After the model is reduced in order, information data is generated to transmit the correlation between several sub-regions. When the model of the next sub-region is reduced in order, the information data is combined to update the data matrix in the corresponding reduced-order matrix. After the update is completed, the nodes in another sub-region are continuously absorbed, and so on.

[0136] The sub-region division module of the present invention is used to execute the above step S1, and the order reduction module and the data updating module are used to execute the above step S2.

[0137] The present invention performs regional decomposition modeling for industrial-grade large-scale PEEC circuits, divides the microscopic circuit model into multiple regions suitable for GPU memory resource constraints, performs circuit model reduction by region, utilizes the data vector generated after each region is reduced due to the low correlation between regions, and iteratively updates the data of other regions according to the number and position information of the absorbed nodes in the model reduction process, utilizes the characteristics of GPU parallel computing to accelerate the reduction of the microscopic circuit model, solves the problem that large-scale electromagnetic field massive data cannot be quickly reduced in GPUs with limited memory resources, and optimizes the matrix of storage data in the process of absorbing nodes. On this basis, the essence of updating nodes is studied, and it is found that only one data matrix needs to be input each iteration and several nodes therein are absorbed, and information data is generated and output to transmit the correlation between the data matrices, and then the next data block is input and updated.

[0138] It should be noted that, although several modules or units of the device for action execution are mentioned in the above detailed description, this division is not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. On the contrary, the features and functions of one module or unit described above can be further divided into multiple modules or units to be embodied.

[0139] Through the description of the above implementation, it is easy for those skilled in the art to understand that the example implementation described here can be implemented by software, or by software combined with necessary hardware. Therefore, the technical solution according to the implementation of the present disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, including several instructions to enable a computing device (which can be a personal computer, a server, a touch terminal, or a network device, etc.) to execute the method according to the implementation of the present disclosure.

[0140] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. The present disclosure is intended to cover any variations, uses or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary technical means in the art that are not disclosed in the present disclosure.

[0141] The above is only a specific implementation of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial changes to the present invention using this concept shall be deemed as an infringement of the protection scope of the present invention.

Claims

1. A method for regional decomposition modeling of a micro circuit model, characterized in that: These include: In the CPU, a microscopic equivalent circuit model of a large-scale electromagnetic problem is divided into a plurality of sub-regions, and a data matrix corresponding to the sub-regions is obtained; A reduced-order matrix corresponding to the sub-region is created in the GPU, and the order of several sub-regions is reduced in turn until the order reduction of several sub-regions is completed: the data matrix of one of the sub-regions is transferred to the corresponding reduced-order matrix for processing to achieve model order reduction of the sub-region, the data matrix in the reduced-order matrix is ​​updated, and information data is generated after the model reduction of the sub-region is completed to transmit the correlation between several sub-regions, and the data matrix in the corresponding reduced-order matrix is ​​updated in combination with the information data when the model of the next sub-region is reduced, and so on.

2. A method for regional decomposition modeling of a micro circuit model as claimed in claim 1, characterized in that: According to the characteristics of the microscopic equivalent circuit model in the three-dimensional rectangular coordinate system, a plurality of sub-areas are divided, and the data matrices corresponding to the sub-areas include: a coupling matrix PP between capacitors, a coupling matrix M between inductors between nodes and inductors between nodes, and a coupling matrix M between inductors on nodes and inductors between nodes. pm , the coupling matrix M between the inductance on the node and the inductance on the node pp , the label mapping matrix and the matrix M pmM , the matrix M pmM The coupling values ​​between the inductances between all nodes in the sub-region and the inductances on all nodes are stored, and the label mapping matrix stores the mapping relationship between the node labels before the sub-region is divided and the node labels after the division.

3. A method for regional decomposition modeling of a micro circuit model as claimed in claim 1, characterized in that: The microscopic equivalent circuit model is divided into several sub-regions, and the data matrix of the sub-region also includes a label matrix, which stores the starting label and the ending label of the node of each sub-region. When reducing the order of the sub-region, the data matrix of the corresponding sub-region is directly selected according to the label matrix.

4. A method for regional decomposition modeling of a micro circuit model as claimed in claim 2, characterized in that: A transition matrix is ​​also created in the GPU to store part of the data required for the sub-region model reduction. The model reduction of the sub-region includes: The absorption condition value △ of each node in the sub-area is calculated using the MIN method, and the node with the smallest absorption condition value is absorbed first. After absorbing this node and updating all the above matrices, other nodes calculate the condition values ​​using the MIN method respectively, and iteratively find the next node with the smallest absorption condition value to be absorbed. The absorption condition value △ calculation formula is as follows: Among them, m II Represents the self-coupling value of the newly added inductor, m SS Represents the self-coupling value of the original inductor, m SI Represents the mutual coupling value between the newly added inductor and the original inductor, P II Represents the self-coupling value of the newly added capacitor element, P SS Represents the self-coupling value of the original capacitor element, P SI Represents the mutual coupling value between the newly added capacitor element and the original capacitor element, Represents the maximum operating frequency of the microwave device; Each time a node is absorbed, a loop is entered: first, according to the needs of each iteration, the coupling data values ​​of the node k to be absorbed and several nodes i connected to it are taken out from the data matrix and placed in the transition matrix X RM middle: PP i represents the capacitor autocoupling matrix, M i represents the inductor autocoupling matrix, M pmi Represents the coupling matrix of the inductance on several nodes i and the inductance between the other nodes in the sub-region, M ppi represents the inductance autocoupling matrix on the above-mentioned nodes i, M pmk Represents the inductance coupling matrix between the inductance at node k and the rest of the nodes in the sub-area.

5. A method for regional decomposition modeling of a micro circuit model as claimed in claim 4, characterized in that: Each time a node is absorbed, the data matrix of the sub-region is compressed, specifically: the last row and the last column of data in the data matrix corresponding to the sub-region are moved to the data position corresponding to the absorbed node.

6. A method for regional decomposition modeling of a micro circuit model as claimed in claim 2, characterized in that: After the model order reduction of the sub-region is completed, the information data generated is as follows: Specifically, x pk 、x mk 、x pmk They represent the update information generated by the i-th group of sub-regions absorbing the k-th node, n represents the total number of nodes i connected around this node k, including the coupling data between the capacitor and all capacitors, the mutual coupling data of the inductance between nodes, and the coupling data between the inductance on the node and the inductance between nodes; the update information X generated by the order reduction of each sub-region i A matrix X is formed, and the matrix X is used to update the data matrices corresponding to all the sub-regions, where N represents the total number of sub-regions that need to be absorbed.

7. A method for regional decomposition modeling of a micro circuit model as claimed in claim 6, characterized in that: When the model of the next sub-region is reduced in order, the data matrix in the reduced-order matrix corresponding to the sub-region is updated in combination with the information data, as follows: PP=PP+c0x p T ·x p M=M+c1x m T ·x m Mpp=Mpp+c2x pm T ·x p +c3x pm T ·x pm Mpm=Mpm+c4x pm T ·x m +c5x p T ·x m ; Wherein, c0, c1, c2, c3, c4, c5 are constants during the transformation of the microcircuit model, and x p Represents the mutual coupling information between capacitor elements, x m Represents the mutual coupling information between inductor components, x pm Indicates the mutual coupling information between the inductance on the PpI branch and the inductance between nodes.

8. A method for regional decomposition modeling of a micro circuit model as claimed in claim 6, characterized in that: The model reduction of the sub-region includes merging parallel branches. When the parallel branches are merged, all mutual impedances in the circuit need to be updated. The parallel circuit elements and the related mutual coupling are listed as follows: The self-impedance of the parallel branch S is: The mutual impedance between the parallel branch S and the remaining inductance branch or other PpI branches is: z' RS =[(y SS +y SI )z RS +(y II +y SI )z RI ] / yt; z' OS =[(y SS +y SI )z OS +(y II +y SI )z OI ] / yt; where y SS ,y II They represent the self-coupling admittance of one of the two parallel branches, y SI represents the mutual coupling admittance of the two parallel branches, z RS represents the coupling impedance between one of the two PPI branches and the other common branches that do not participate in the combined PPI branch, z OS represents the coupling impedance of one branch to other PpI branches, z RI It represents the coupling impedance between the other branch of the two parallel branches and other common branches that do not participate in the combined PpI branch, z OI represents the coupling impedance between the other branch of the two parallel branches and the other PpI branches, z' RS It represents the coupling impedance between the two PPI branches after merging and other common branches that do not participate in the merging of PPI branches, z' OS Indicates the coupling impedance with other PpI branches after the two PpI branches are merged. The mutual impedance between the other PpI branches and the remaining inductance branches is: Among them, Z' OO represents the self-coupling after the other PpI branches are merged, Z OO represents the self-coupling of other PpI branches before merging, Z' OR It represents the coupling between other PpI branches and common branches after merging PpI branches, Z OR represents the coupling between other PpI branches and common branches before merging the PpI branches, Z' RR represents the self-coupling of other common branches after merging the PpI branch, Z RR Represents the self-coupling of other common branches before merging the PpI branch.

9. A device for regional decomposition modeling of a micro circuit model, characterized in that: include: A sub-region division module divides the microscopic equivalent circuit model of a large-scale electromagnetic problem into several sub-regions in the CPU to obtain a data matrix corresponding to the sub-regions; An order reduction module is used to create an order reduction matrix corresponding to the sub-region in the GPU, and to reduce the order of several sub-regions in turn until the order reduction of several sub-regions is completed: a data matrix of one of the sub-regions is transferred into the corresponding order reduction matrix for processing to achieve model order reduction of the sub-region; The data update module generates information data after the model is reduced to transmit the correlation between the sub-regions. When the model of the next sub-region is reduced, the information data is combined to update the data matrix in the corresponding reduced-order matrix. After the update is completed, the nodes in another sub-region are continuously absorbed, and so on.

Images