Multi-layer global routing method, device and equipment based on edge resource continuous flow modeling

Abstract

This application provides a multi-layer global routing method, apparatus, and device based on edge resource continuous flow modeling, effectively solving the problems of inaccurate congestion prediction, low redistribution efficiency, and poor boot file quality in existing global routing methods. The method includes: constructing a multi-layer global routing cell mesh diagram and generating a pin-sensing potential energy field using three-dimensional coordinates; constructing a multi-objective optimization function for optimizing horizontal and vertical edge flow variables, and solving for the target horizontal and vertical edge flow variables; reconstructing the cell-level occupancy confidence field of each global routing cell on each metal layer based on the target horizontal and vertical edge flow tensors, calculating the layer score of each network to be routed on each metal layer, and determining the active layer set; determining a boot region generation strategy with a conversion mechanism based on the pin type, and executing the boot region generation strategy in the active layer set to obtain the routing boot file.

Description

Technical Field

[0001] This application relates to the field of global routing technology for integrated circuits, and more specifically, to a multi-layer global routing method, apparatus, and device based on edge resource continuous flow modeling. Background Technology

[0002] Current integrated circuit multilayer global routing commonly employs discrete GCell mesh modeling, abstracting the routing problem into searching for Steiner trees or shortest paths in the multilayer graph. This technique optimizes GCell cell occupancy or discrete paths, indirectly alleviating resource contention through iterative rip-up and reroute operations and congestion penalty.

[0003] However, real-world routing resource bottlenecks are concentrated in the boundary channels (i.e., horizontal / vertical edges) between adjacent GCells. Existing methods lack direct, continuous, and differentiable modeling of edge capacity, leading to inaccurate congestion prediction, slow redistribution convergence, and recurring local congestion. Furthermore, layer allocation often relies on pre-defined rules such as "short lines use lower layers, long lines use higher layers," making it difficult to dynamically respond to real-time differences in inter-layer congestion, via cost changes, and network geometry such as pin density and orientation matching during optimization. In addition, while continuous optimization methods can generate softwire density fields, they often result in missing pin coverage, broken path connectivity, and fragmented guide regions due to the lack of coupling between pin position constraints and downstream tool format requirements, failing to stably output rectangular routeguide files that conform to detailed router input specifications. Ultimately, this leads to inaccurate congestion prediction, low redistribution efficiency, and poor guide file quality in global routing.

[0004] Therefore, there is an urgent need for a novel global routing method that uses GCell edge flow as the native variable, integrates pin-aware potential energy guidance, supports adaptive selection of congestion drive layer, and has a verifiable conversion mechanism for boot files. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide a multi-layer global routing method, apparatus and equipment based on edge resource continuous flow modeling, which effectively solves the problems of inaccurate congestion prediction, low rerouting efficiency and poor boot file quality in existing global routing methods.

[0006] In a first aspect, embodiments of this application provide a multi-layer global routing method based on edge resource continuous flow modeling, the method comprising: Based on the obtained layout exchange file and design exchange file of the integrated circuit board, a multi-layer global routing cell mesh diagram is constructed, and a pin-sensing potential energy field is generated by using the three-dimensional coordinates of all pins of each network to be routed in the multi-layer global routing cell mesh diagram. For each network to be routed, set horizontal and vertical side current variables to construct a multi-objective optimization function for optimizing the horizontal and vertical side current variables with the pin-sensing potential field, and solve for the optimized target horizontal and vertical side current variables. Based on the target horizontal edge flow tensor and the target vertical edge flow tensor, the cell-level occupancy confidence field of each global routing cell in the multi-layer global routing cell mesh diagram is reconstructed on each metal layer. Based on the cell-level occupancy confidence field and the calculated real-time congestion level of each metal layer, the layer score of each network to be routed on each metal layer is calculated, and the set of active layers is determined. The pin type of each network to be routed is determined, and a guide region generation strategy with a conversion mechanism is determined based on the pin type. The guide region generation strategy is then executed in the set of active layers to obtain rectangular routing guide regions and pin access guide regions for all networks to be routed, so as to obtain a routing guide file that can be used downstream and realize multi-layer global routing.

[0007] In conjunction with the first aspect, this application provides a first possible implementation of the first aspect, wherein executing the guide region generation strategy to obtain rectangular routing guide regions and explicit pin access guide regions for all networks to be routed includes: A target shortest path search method is pre-configured, guided by the pin-sensing potential energy field and constrained by the preferred wiring direction of the metal layer and the cost of interlayer vias. If the current wiring network is a low-pin network, then a discrete path is obtained by directly executing the pre-configured target shortest path search method based on the set of active layers; The rectangular routing guide area and explicit pin access guide area for each network to be routed are obtained based on the discrete path.

[0008] In conjunction with the first aspect, embodiments of this application provide a second possible implementation of the first aspect, wherein, if the current wiring network is not a low-pin network, it includes: If the current wiring network is a multi-pin network, then construct a rectangular minimum spanning tree topology and decompose it into multiple two-pin connection segments; For each two-pin connection segment, the target shortest path search method is performed within its active layer set to obtain discrete paths.

[0009] In conjunction with the first aspect, this application provides a third possible implementation of the first aspect, wherein obtaining the rectangular routing guide region and pin access guide region for each network to be routed based on discrete paths includes: The discrete path is transformed into an initial rectangular routing guide region aligned with the global routing cell grid, and adjacent initial rectangular routing guide regions on the same layer are merged to obtain a rectangular routing guide region. For each pin of a global routing cell in a network to be routed, explicitly generate a pin access guide region that covers the entire area of ​​that global routing cell.

[0010] In conjunction with the first aspect, this application provides a fourth possible implementation of the first aspect, wherein, based on the cell-level occupancy confidence field and the calculated real-time congestion level of each metal layer, the layer score of each network to be routed on each metal layer is calculated, and the set of active layers is determined, including: Based on the unit-level occupancy confidence field, spatial dimensions are summed to obtain the confidence quality, and the confidence quality is used to calculate the layer score. By comparing the layer score of each metal layer with the relative threshold of the layer score, the comparison results are obtained to determine the active layers, and all active layers are integrated to obtain the active layer set.

[0011] In conjunction with the first aspect, this application provides a fifth possible implementation of the first aspect, wherein calculating the confidence quality to obtain a layer score includes: Obtain and compare the average congestion level of the underlying metal stack with the corresponding preset congestion threshold to obtain the comparison result; Based on the comparison results, the layer score weights of the upper metal layers are dynamically adjusted to calculate the layer score based on the adjusted layer score weights, the degree of orientation matching, the via distance, and the degree of congestion.

[0012] In conjunction with the first aspect, embodiments of this application provide a sixth possible implementation of the first aspect, wherein a multi-objective optimization function for optimizing horizontal and vertical sideflow variables is constructed using the pin-sensing potential energy field, including: Based on the pin-sensing potential energy field, edge capacity, via cost, and preferred wiring direction of the metal layer, a geometric guidance term, an edge capacity overflow term, an inter-layer switching regularization term, and a preferred direction penalty term are constructed respectively. By integrating the geometric guidance term, edge capacity overflow term, inter-layer switching regularization term, and preferred direction penalty term, a multi-objective optimization function is obtained for continuous gradient optimization.

[0013] In conjunction with the first aspect, this application provides a seventh possible implementation of the first aspect, wherein generating a pin-sensing potential energy field using the three-dimensional coordinates of all pins of each network to be routed in a multilayer global routing cell mesh diagram includes: Map all pins of each network to be routed to the multilayer global routing unit mesh diagram to obtain its corresponding three-dimensional coordinate set; The potential energy value of the target node is calculated based on the Manhattan distance between the three-dimensional coordinate set and the target node coordinates, as well as the via cost, to generate a pin-sensing potential energy field.

[0014] Secondly, embodiments of this application provide a multi-layer global routing apparatus based on edge resource continuous flow modeling, the apparatus comprising: The acquisition module is used to construct a multi-layer global routing cell mesh diagram based on the obtained layout exchange file and design exchange file of the integrated circuit board, and generate a pin sensing potential energy field by using the three-dimensional coordinates of all pins of each network to be routed in the multi-layer global routing cell mesh diagram. The configuration module is used to set the horizontal and vertical side current variables for each network to be wired, so as to construct a multi-objective optimization function for optimizing the horizontal and vertical side current variables with the pin sensing potential field, and solve for the optimized target horizontal and vertical side current variables. The reconstruction module is used to reconstruct the cell-level occupancy confidence field of each global routing unit on each metal layer of the multi-layer global routing unit mesh graph based on the target horizontal edge flow tensor and the target vertical edge flow tensor. Based on the cell-level occupancy confidence field and the calculated real-time congestion degree of each metal layer, the module calculates the layer score of each network to be routed on each metal layer and determines the set of active layers. The execution module is used to determine the pin type of each network to be routed, determine the guide region generation strategy with conversion mechanism based on the pin type, and execute the guide region generation strategy in the active layer set to obtain the rectangular routing guide region and pin access guide region of all networks to be routed, so as to obtain the routing guide file for downstream use and realize multi-layer global routing.

[0015] Thirdly, embodiments of this application provide an electronic device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of any of the following multi-layer global routing methods based on edge resource continuous flow modeling.

[0016] This application provides a multi-layer global routing method based on continuous flow modeling of edge resources, which achieves significant technical improvement through four-stage collaborative design: (1) Edge flow modeling + potential energy field guidance: Using horizontal / vertical edge flow tensors as optimization variables, and combining the three-dimensional coordinates of pins to construct a potential energy field, the routing tendency naturally focuses on the pin neighborhood and reasonable corridors, improving geometric rationality from the source; (2) Congestion direct control + multi-objective differentiable optimization: The edge capacity overflow term directly constrains the total edge flow exceeding the limit, avoiding the oscillating convergence of traditional heuristic redistribution, and the congestion relief is more stable; (3) Layer score-driven activation layer selection: Based on the confidence field, direction matching degree, via distance and layer congestion status, the layer score is dynamically calculated, breaking through the fixed rule restriction, and realizing adaptive resource scheduling of "low layer to keep compact and high layer to relieve congestion"; (4) According to the pin type of each network to be routed, a guiding region generation strategy with conversion mechanism is determined, and the corresponding guiding region generation strategy is executed, thereby solving the problems of inaccurate congestion prediction, low redistribution efficiency and poor guiding file quality in the existing global routing method. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart of a multi-layer global routing method based on edge resource continuous flow modeling provided in an embodiment of this application is shown; Figure 2 This paper shows the layout diagram of the GCD chip provided in the embodiment of this application before global routing; Figure 3 This application shows a unit density thermal diagram of GCD provided in an embodiment of the present application; Figure 4 A cross-sectional view of the metal laminate provided in an embodiment of this application is shown; Figure 5 A pin count diagram provided in an embodiment of this application is shown; Figure 6 A wiring comparison diagram provided in an embodiment of this application is shown; Figure 7 A comparison diagram of the guide density distribution provided in the embodiments of this application is shown; Figure 8 This paper shows a structural block diagram of a multi-layer global routing device based on edge resource continuous flow modeling according to an embodiment of this application; Figure 9 A structural block diagram of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0020] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0021] It should be noted that the term "comprising" will be used in the embodiments of this application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0022] Existing global routing methods focus on discrete paths or GCell cell occupancy, failing to directly model GCell boundary resources that determine routing feasibility. Layer allocation relies on fixed rules, lacking congestion-aware dynamism. Continuous optimization results are difficult to reliably convert into rectangular routing boot files that satisfy pin coverage, path connectivity, and compatibility with downstream tools. These shortcomings lead to inaccurate congestion prediction, low rerouting efficiency, and poor boot file quality.

[0023] Based on this, embodiments of this application provide a multi-layer global routing method, apparatus, and device based on edge resource continuous flow modeling, which are described below through embodiments.

[0024] Example 1 To facilitate understanding of this embodiment, a multi-layer global routing method based on edge resource continuous flow modeling, disclosed in this application embodiment, will first be described in detail. For example... Figure 1The diagram shows a flowchart of a multi-layer global routing method based on edge resource continuous flow modeling. This application provides a multi-layer global routing method based on edge resource continuous flow modeling, the method comprising: S101. Construct a multi-layer global routing cell mesh diagram based on the obtained layout exchange file and design exchange file of the integrated circuit board, and generate a pin sensing potential energy field by using the three-dimensional coordinates of all pins of each network to be routed in the multi-layer global routing cell mesh diagram. S102. Set horizontal and vertical side current variables for each network to be routed, and construct a multi-objective optimization function for optimizing the horizontal and vertical side current variables with the pin sensing potential field, and solve for the optimized target horizontal and vertical side current variables. S103. Based on the target horizontal edge flow tensor and the target vertical edge flow tensor, reconstruct the cell-level occupancy confidence field of each global cabling unit in the multi-layer global cabling unit mesh diagram on each metal layer. Based on the cell-level occupancy confidence field and the calculated real-time congestion level of each metal layer, calculate the layer score of each network to be cabled on each metal layer and determine the set of active layers. S104. Determine the pin type of each network to be routed, determine the guide region generation strategy with conversion mechanism based on the pin type, and execute the guide region generation strategy in the active layer set to obtain the rectangular routing guide region and pin access guide region of all networks to be routed, so as to obtain a routing guide file for downstream use and to realize multi-layer global routing.

[0025] In step S101, this application uses an integrated circuit board with a greatest common divisor (GCD) circuit design generated based on the OpenROAD RTL-to-GDS flow as input. This design is the physical design after clock tree synthesis and before global routing, with a process node of SkyWater 130nm and a design name of gcd. Figure 2 As shown, Figure 4 The diagram shows six metal wiring layers: li1, met1, met2, met3, met4, and met5. Figure 4 The symbols following li1, met1, met2, met3, met4, and met5 indicate directional preference, showing that the metal layers alternate between horizontal and vertical orientations. Table 1 illustrates the meaning of the six metal wiring layers, with a design area of ​​approximately [missing information]. It contains 479 signal networks to be routed. Based on the obtained Library Exchange Format (LEF) and Design Exchange Format (DEF) files of the integrated circuit board, the key data contained in the LEF and DEF files are analyzed, including chip physical boundaries, metal layer stack information, preferred routing directions for each metal layer (e.g., met1 is preferred horizontally, met2 is preferred vertically), routing track spacing, via type and size, standard cell layout position, pin coordinates and net connection relationships. Based on the chip design area and preset GCell size of the integrated circuit board, a three-dimensional multilayer global routing cell mesh (GCell Graph) covering the entire integrated circuit board is constructed. Its vertices are represented as triples (r, c, l), where r and c are the index or number of GCell in the row and column directions, respectively, and l is the metal layer number (e.g., li1, met1…met5). Each GCell boundary is explicitly modeled as two types of directed edges: horizontal edges (connecting (r,c,l) and (r,c+1,l)) and vertical edges (connecting (r,c,l) and (r+1,c,l)). This application further calculates the capacity of the horizontal edges based on the preferred wiring direction and wiring track spacing of each metal layer. and vertical edge capacity The wiring capacity of different metal layers varies with the wiring track spacing and preferred wiring direction. For example, the horizontal edge capacity of layer met1 is significantly higher than its vertical edge capacity, while layer met2 is the opposite, with higher vertical edge capacity. Higher metal layers such as met4 and met5 have relatively lower capacity but can be used to alleviate local congestion or carry long-distance connections. The capacity of each edge is automatically calculated by the process layer parameters based on the following formula: ; in This refers to the track spacing on this level. For the corresponding dimension of GCell, and These represent the width and spacing of the conductors in that layer, respectively. This application uniformly divides the integrated circuit board into a 12×12 GCell grid, as shown below. Figure 3 As shown, this is the network to be routed. Each network has a side length of 6.9 μm (6900 DBU) and contains 6 metal routing layers. Each layer has a strict orientation preference (alternating horizontal / vertical), and adjacent layers are connected by vias. All pins of each network to be routed are precisely mapped to the corresponding GCell's three-dimensional coordinates to form a pin set. ,in, Indicates the network to be wired The system defines the GCell row, column, and metal layer where a specific pin is located. For single-pin networks or networks requiring only pin access, the system treats them as self-loop networks, ensuring that a pin access guide region covering the pin location is generated during the subsequent routing guide file output stage, and a pin set is formed based on three-dimensional coordinates. Generate a potential energy field for pin sensing.

[0026] Table 1. Explanation of the meaning of the six metal wiring layers.

[0027] In a specific implementation of step S101, one embodiment is as follows: A pin-sensing potential energy field is generated using the three-dimensional coordinates of all pins of each network to be routed in the multilayer global routing cell mesh diagram, including: S1011. Map all pins of each network to be routed to the multilayer global routing unit mesh diagram to obtain its corresponding three-dimensional coordinate set; S1012. Calculate the potential energy value of the target node based on the Manhattan distance between the three-dimensional coordinate set and the target node coordinates and the via cost, so as to generate a pin-sensing potential energy field.

[0028] In steps S1011-S1012, to achieve pin-aware fine-grained routing guidance, this application first maps all pins associated with each network to be routed to a pre-constructed multi-layer global routing unit mesh diagram based on their physical location and the information of the metal layer to which they belong, forming a set of three-dimensional coordinates corresponding to the network to be routed. Each coordinate contains the X and Y plane positions and the corresponding metal layer index in the Z direction. Indicates the network to be wired The GCell row, column, and metal layer where a specific pin is located. For each network n to be routed, its pin set... Using the source point, potential energy propagation is performed on the 3D GCell graph: Dijkstra's algorithm is used for edge-weighted diffusion, with weights comprehensively considering Manhattan distance, cross-layer via cost, and edge capacity surplus; alternatively, a more efficient closed approximation method is used, constructing the initial potential energy using a weighted reciprocal distance function, followed by multiple rounds of Gaussian smoothing to suppress noise. The resulting pin-sensing potential energy field is represented as: ; Where d opt The shortest path length between two pins is normalized so that the value near the pin approaches 1 and the value far from the pin approaches 0. The pin-sensing potential field characterizes GCell(r,c,l) and the network to be routed. The geometric affinity between pins indicates that the higher the value, the closer the cell is to the pin or within a reasonable wiring corridor, thus providing unbiased, differentiable, pin-driven geometric guidance signals for subsequent continuous sideflow optimization.

[0029] Since the potential energy fields of each network to be wired are independent of each other, multiple sets of networks to be wired can be divided into multiple subsets and parallel computing can be performed using multi-core central processing units to improve the efficiency of potential energy field construction.

[0030] In step S102, this application sets horizontal edge flow variables and vertical edge flow variables for each network to be wired, respectively. ,in, For each network to be wired This represents the horizontal edge flow from cell (r,c) to cell (r,c+1) in the l-th layer. For each network to be wired This represents the corresponding vertical edge flow variable. The flow rate is normalized to [0,1], and multiplied by the capacity to correspond to the actual number of occupied tracks, thus representing the degree of continuous routing occupancy on the horizontal GCell boundary and the vertical GCell boundary, respectively. Furthermore, the horizontal and vertical edge flow variables are combined with the pin-sensing potential field to construct a multi-objective optimization function for optimizing the horizontal and vertical edge flow variables. This application argues that global routing is mathematically a large-scale nonlinear continuous optimization problem with mixed constraints, and its essence can be characterized from the following three dimensions: (1) Problem structure: Continuous relaxation of multi-commodity flow, each network to be wired i Treated as a "commodity," it needs to flow from the source pin to the target pin on the GCell graph G=(V,E). Classical Integer Multi-Commodity Flow (IMCF) requires a discrete tree structure for the path, which is an NP-hard problem. This application relaxes it to a continuous edge flow variable: ; Thus, the discrete path selection problem is transformed into a density allocation problem in a continuous space, and the problem complexity is reduced from combinatorial explosion to non-convex continuous optimization, making gradient methods applicable. (2) Constraint Structure: A hybrid implementation of soft constraint penalties and hard projection, as shown in Table 2, treats the problem as containing three types of constraints with different properties, and adopts different implementation strategies for each: Table 2 Constraint-Policy Table

[0031] This hybrid hard and soft structure is necessary: ​​if all constraints are reduced to penalty terms, the condition number of the objective function will deteriorate drastically; if projection is used entirely, nonlinear capacity constraints cannot provide effective gradient guidance.

[0032] (3) Objective function: A special structure of non-convex multi-objective weighted sum, the four costs have different mathematical properties: J wl It is a linear function of (H,V), has a global optimum, and guides the optimizer to concentrate traffic near the pin; J ovf It is an element-wise convex function composed of softplus, but the historical congestion term conghist(e) is dynamically updated with iteration, introducing non-stationarity; J via It involves layer-by-layer max operations, and is Lipschitz continuous but not smooth overall, with subgradients present; J dir The sum of linear and quadratic terms results in a strong gradient signal near zero (linearity dominates), while the quadratic term prevents significant violations when far from zero. The overall objective function is non-convex. The Adam optimizer provides local optima under an adaptive first-order method, and NCR scheduling dynamically adjusts... Gradually changing the optimized landscape is equivalent to applying progressively stronger Lagrangian relaxation to the capacity constraints, helping to escape early low-congestion local minima. The core challenges of global routing are: scale (GCD design involves approximately 12 × 11 × 6 × 479 ≈ 3.8 million horizontal edge variables, requiring efficient tensor parallelism); non-stationarity (historical congestion terms cause the objective function to exhibit different geometric structures at different iteration stages, making it difficult to analyze using standard convergence theory); and coupling of projection and gradient (hard projection and gradient updates are performed alternately, belonging to a variant of the Projected Gradient Method, whose convergence depends on the compatibility of the step size and projection operator). Therefore, this application designs a multi-objective optimization function suitable for gradient optimization based on the above, and solves the multi-objective optimization function using a gradient optimization algorithm to obtain the optimized target horizontal edge flow variables and target vertical edge flow variables. The target horizontal edge flow variable and the target vertical edge flow variable are not the final discrete routing path, nor the final routing guide file, but rather represent the soft routing tendency of each network on the global routing diagram. In a specific implementation of step S102, one embodiment involves constructing a multi-objective optimization function for optimizing horizontal and vertical sideflow variables using the pin-sensing potential energy field, including: S1021. Based on the pin-sensing potential energy field, edge capacity, via cost, and preferred wiring direction of the metal layer, respectively construct the geometric guidance term, edge capacity overflow term, inter-layer switching regularization term, and preferred direction penalty term; S1022. Integrate the geometric guidance term, edge capacity overflow term, inter-layer switching regularization term, and preferred direction penalty term to obtain a multi-objective optimization function for continuous gradient optimization.

[0033] In steps S1021-S1022, the multi-objective optimization function finally designed in this application, including the geometric guidance term, edge capacity overflow term, inter-layer switching regularization term, and preferred direction penalty term, is expressed as follows: ; J wl To minimize line length cost, driven by distance field, the geometric guidance term is constructed based on the pin-sensing potential energy field. The edge cost is derived from the distance field. A distance field value of dist=1 indicates that the pin is close to the pin (low cost), and dist=0 indicates that the pin is far away (high cost). ; The physical meaning is that the optimizer is "attracted" by the distance field to the shortest path of the pin connection, which is equivalent to automatically guiding the network to route along the Steiner shortest tree. J ovf It is the overflow cost, subject to soft capacity constraints, and is an edge capacity overflow term constructed based on edge capacity. ; where the effective capacity c eff Considering historical congestion (FastRoute style): Physically, softplus is a differentiable approximation of ReLU. This allows congestion penalties to be backpropagated in a gradient; historical congestion terms make repeatedly congested edges "more and more expensive," simulating iterative rip-up and reroute behavior; J via It is a layer jump penalty, which is an inter-layer switching regularization term constructed based on via cost. ; where the grid density Defined as three-dimensional coordinates Maximum flow of four adjacent edges: The physical meaning is that the greater the density difference between adjacent layers, the more likely a layer jump (via) occurs in the network to be routed at this point. This item encourages paths to remain on the same layer as much as possible to reduce the number of vias. J dir It is the cost of directional violation, which is a penalty term for the preferred directional routing constructed based on the preferred routing direction of the metal layer. The physical meaning is: to provide a strong gradient signal near zero, the quadratic term prevents large-scale directional violations, and the resultant force compels each layer to comply with the directional preference.

[0034] The constraints of the multi-objective optimization function are: ; ; ; As shown in Table 3, the meanings of each letter are explained above, where the overflow weight wovf Dynamically adjusted during the iteration process, it is divided into two phases: warmup and stabilization: ; The meaning of the phases is as follows: Warmup phase (t<Twarmup): with low overflow weight, the optimizer prioritizes finding the shortest path tree and ignores slight congestion; Stabilization phase (t≥Twarmup): with full weight, it makes every effort to eliminate congestion violations and approach a feasible solution.

[0035] Table 3 Table for explaining the meaning of each letter

[0036] In step S103, the present application fuses the target horizontal edge flow tensor and the target vertical edge flow tensor. These tensors have soft distribution, contain noise, and may have multiple coexisting channels, which are not convenient for direct use in topology decision-making. Therefore, combined with the unit topology structure and the direction preference characteristics of the metal layers, the unit-level occupancy field of each global routing unit on each metal layer is reconstructed layer by layer , which takes the maximum value of the flow of four adjacent edges: ; The present application also performs convex combination of the two pieces of information of the pin-aware potential field and the unit-level occupancy field to form the unit-level occupancy confidence field of each metal routing layer for each to-be-routed network, and truncates it to [0,1]: ; is the mixing weight between the pin-aware potential field and the unit-level field, which is determined adaptively for each to-be-routed network: , the more pins npins a network has, the more complex its topology is, the smaller it is (with a lower limit of 0.60), that is, the more trust is placed in the unit-level occupancy field rather than the pin-aware potential field; the clock network has a fixed preference for the backbone (0.75) to keep the routing paths regular and symmetrical. Its physical meaning is that the pin-aware potential field ensures that the unit-level occupancy field has a clean backbone and is not misled by optimization noise, while the unit-level occupancy field injects the real routing direction after congestion avoidance. The fusion of the two obtains a grid-by-grid confidence field that is both regular and consistent with the optimization result. This confidence field quantitatively reflects the possibility distribution that the global routing unit is actually occupied by the routing path on the corresponding metal layer. Based on the reconstructed unit-level occupancy confidence field, the local occupancy density of each metal layer (the metal layer in the present application is the metal routing layer) in the key area is counted, and dynamic normalization is performed to obtain the real-time congestion degree of each metal layer. For each to-be-routed network, the present application also weights and aggregates the occupancy confidence values of each unit on each metal layer along the path it passes through according to its geometric span, pin distribution characteristics and electrical constraints, and performs collaborative modeling with the real-time congestion degree of the corresponding metal layer to generate a layer score of the network on each metal layer; the layer score comprehensively reflects the feasibility and robustness of completing high-quality routing for the network on a specific metal layer. Finally, all metal layers with layer scores higher than the preset threshold form an active layer set The relative threshold allows the "number of activated layers" to adapt to the network: networks that clearly dominate a certain layer select only a few layers, while networks with dispersed activation select more layers. Then, two minimum protection constraints are applied: H / V orthogonal coverage: if... If there is no H layer, fill in the H layer with the highest score; the same applies to the V layer. Ensure there is at least one horizontal layer and one vertical layer so that the routing can make orthogonal turns; Pin connection: always incorporate the 0th layer (li1). This ensures that the pins are accessible to the set, and the final pin position will be attracted to it. The most recent active layer is selected, and subsequent RMST + Dijkstra guide extraction only routes within this set of active layers. This set of active layers can be directly used for layer allocation decisions in subsequent detailed routing stages, significantly improving routing convergence and the balance of metal layer resource utilization.

[0037] In a specific implementation of step S103, one embodiment is as follows: based on the cell-level occupancy confidence field and the calculated real-time congestion level of each metal layer, the layer score of each network to be routed on each metal layer is calculated, and the set of active layers is determined, including: S1031. Based on the unit-level occupancy confidence field, sum the spatial dimensions to obtain the confidence quality, and calculate the confidence quality to obtain the layer score; S1032. Compare the layer score of each metal layer with the relative threshold of the layer score to obtain the comparison result, determine the active layer, and integrate all the active layers to obtain the active layer set.

[0038] In steps S1031-S1032, this application uses the aggregated amount obtained by summing the cell-level occupancy confidence field of each metal wiring layer of each network to be routed along the spatial dimension as the confidence quality, which measures which layers the confidence of a network is "concentrated" on. Two aggregation methods are used for layer selection, one of which is: quality and proportion of each layer, used to sum the entire layout layer by layer and then normalize it to a distribution. ; in The density quality reflects the distribution of network confidence across layers and is the source of the "density quality" term in the layer score. Another approach is to sum the corridor quality layer-by-layer only within the network corridors (row and column envelopes where dist > 0). Its normalized value The focus is on the area actually traversed by the network to be cabled, treated as a local component of the congestion term in the layer score (mixed with global layer congestion at a 6:4 ratio). The layer score is then calculated based on the following formula: ; The meanings of the letters are explained in Table 4, where density and mass are listed below. To measure the confidence level of the network to be routed, how much of it falls within layer l, the quality of the cell-level occupancy confidence field is normalized across all layers: Direction matching Assume the network corridor (the envelope of dist > 0) extends horizontally. Vertical extension grid, horizontal proportion ,but: This means that horizontally-dominated networks prefer layer H, while vertically-dominated networks prefer layer V; a lower bound of 0.15 ensures that no layer will be completely rejected by the direction term. (Via cost) From layer l to the lowest pin layer Normalized distance: In other words, the farther away from the pin layer, the more vias are needed for connection. This naturally suppresses layers far from the pins, causing simple 2-pin short networks to tend to remain at the bottom layer, without any hard constraints. Upper limit. Congestion. The global congestion rate for each layer is cong. layer (l) (The demand / capacity of this layer, taking the larger one in the H / V direction) is mixed with the congestion within the network corridor in a 4:6 ratio, and then normalized: ;in This is the normalized value of the confidence quality of this layer within the network corridor. Once the bottom layer is saturated, its... As scores rise and fall, scores at higher levels rise relatively, leading to a natural process of moving upwards to avoid congestion, rather than being forced by switching on or off.

[0039] Table 4. Explanation of the meaning of letters

[0040] In a specific implementation of step S1031, one embodiment involves calculating the confidence quality to obtain a layer score, including: S10311. Obtain and compare the average congestion level of the underlying metal stack and the corresponding preset congestion threshold to obtain the comparison result; S10312. Based on the comparison results, dynamically adjust the layer score weights of the upper metal layer to calculate the layer score based on the adjusted layer score weights, the degree of orientation matching, the via distance, and the degree of congestion.

[0041] In steps S10311-S10312, when the scenario explicitly requires evacuation to higher floors (upper floors), this application... layer Scene, =0.35) Obtain and compare the average congestion level of the bottom metal stack (usually M1–M3) composed of multiple bottom metal layers. And based on the congestion threshold calculated according to the congestion level of each metal wiring layer, a comparison result is obtained, where the average congestion level The saturation level of basic cabling resources is accurately reflected by statistically integrating the confidence field of each unit level on the underlying metal layer and normalizing it. If the comparison results show that the congestion level of the underlying layer exceeds the threshold, the system determines that the underlying resources are becoming strained. At this time, the weight coefficient of the upper metal layer (such as M4 and above) in the layer score is automatically increased to encourage cabling to migrate to the upper layer. Conversely, if the underlying resources are sufficient, the weight of the upper layer is appropriately reduced, and the use of low-cost, low-latency underlying cabling resources is prioritized. This is expressed as: ; in =0.35, and the higher the layer number, the greater the bonus, thus accelerating the overall upward movement of the cabling. And... When =0 (default), this mechanism is disabled. This weight adjustment is not a fixed, abrupt switch, but rather a smooth, continuous adjustment based on the degree of congestion deviation from the threshold, ensuring strategy stability. The adjusted weights are applied to subsequent layer score calculations, working in conjunction with three indicators: direction matching degree, via distance, and current metal layer congestion level, to generate the final layer score. Direction matching degree measures the fit between the network routing and the main routing direction of the metal layer; via distance assesses the physical overhead and reliability impact introduced by cross-layer connections; and congestion level focuses on the local resource pressure of the target metal layer itself. This closed-loop feedback mechanism ensures that the layer score always responds to the actual distribution of routing resources, significantly enhancing the adaptability and robustness of global routing decisions.

[0042] This application pre-defines six optimization scenarios for different design goals, each with a corresponding weight setting strategy, as shown in Table 5: Table 5 Weight-Strategy Table

[0043] In step S104, after obtaining the set of active layers, this application determines the pin type of each network to be routed based on the number of pins in each network. This includes determining whether the pin type is a multi-pin type or a few-pin type, defined by a pin count greater than 2, and ≤2 being a few-pin type. Figure 5The pin diagram shown shows the network fan-out (pin count) distribution on the left. Of the 479 networks to be routed, 267 are 2-pin networks, representing a low-pin type. The right diagram shows the top 15 networks with the most pins, with a maximum fan-out of 20, representing a high-pin type. This corresponds to the Steiner tree modeling requirements, and a guide region generation strategy with a conversion mechanism is determined. This strategy is then executed within the active layer set to obtain rectangular routing guide regions and pin access guide regions for all networks to be routed. All guide regions are output in a standard geometric description format, including layer identifiers, coordinate boundaries, access priorities, and constraint attributes. These are then integrated into a unified routing guide file. This file can be directly delivered to downstream detailed routing tools, supporting collaborative routing planning across metal layers and significantly improving the structural controllability and implementation consistency of multi-layer global routing.

[0044] In a specific implementation of step S104, one embodiment is as follows: The guide region generation strategy is executed to obtain rectangular routing guide regions and explicit pin access guide regions for all networks to be routed, including: S1041. A target shortest path search method is pre-configured, guided by the pin-sensing potential energy field and constrained by the preferred wiring direction of the metal layer and the cost of interlayer vias. S1042. If the current wiring network is a low-pin network, then a discrete path is obtained by directly executing the pre-configured target shortest path search method based on the set of active layers. S1043. Obtain the rectangular routing guide area and explicit pin access guide area for each network to be routed based on the discrete path.

[0045] In steps S1041-S1042, this application pre-configures a target shortest path search method, the core of which is to use the pin-sensing potential energy field as a global guiding force: this potential energy field starts from each pin position and decays anisotropically, with its intensity decreasing smoothly as the distance increases, and integrates cell density and local obstacle information to form an attractive distribution that has physical significance for the routing direction; at the same time, the preferred routing direction of the metal layer (e.g., even-numbered layers prioritize horizontal, odd-numbered layers prioritize vertical) and the cost of interlayer vias are used as hard constraints to ensure that the discrete path achieves directional compliance and cross-layer economy under the premise of meeting process rules. For the network to be routed that is identified as a low-pin type (i.e., containing only two pins or pins with adjacent heights), the target shortest path search method is directly started within the range of metal layers covered by its active layer set, without cross-layer redundant exploration, but instead performing a restricted shortest path search layer by layer, prioritizing the metal layer with the optimal potential gradient, high directional matching degree and the fewest via introductions as the main routing layer, and outputting a discrete path with a clear layer affiliation and coordinate sequence. Based on this discrete path, a preset safety margin is further extended outward along the path centerline to generate a rectangular routing guidance area covering the entire path. Simultaneously, at the start and end points of the path, explicit pin access guidance areas are generated based on pin shape, orientation, and metal layer direction, precisely defining the geometric window and accessible edges for pin connections. All guidance areas are uniformly organized into a structured routing guidance file, possessing complete semantics such as layer identifiers, coordinate boundaries, and access constraints. This allows for seamless integration with downstream detailed routing engines, supporting efficient and deterministic multi-layer global routing implementation.

[0046] In a specific implementation of step S1042, one embodiment is as follows: if the current wiring network is not a low-pin network, it includes: S10421. If the current wiring network is a multi-pin network, construct a rectangular minimum spanning tree topology and decompose it into multiple two-pin connection segments. S10422. For each two-pin connection segment, perform the target shortest path search method within its active layer set to obtain discrete paths.

[0047] In steps S10421-S10422, when determining that the current wiring network is a multi-pin network (i.e., connecting three or more physical pins, and the pins are significantly dispersed on the layout plane), this application first constructs its rectangular minimum spanning tree topology: this structure takes the geometric center of each pin as the vertex, connects edges based on Manhattan distance, and embeds metal layer orientation constraints and local congestion avoidance mechanisms to ensure that the generated tree not only satisfies topological connectivity and acyclicity, but also naturally adapts to the back-end wiring process rules; subsequently, the spanning tree is decomposed into multiple two-pin connection segments according to standard graph theory methods. Each connection segment represents two pin nodes connected by an edge in the tree. All connection segments together cover all pins and maintain the logical connectivity of the original network. For each two-pin connection segment, within the available metal layers defined by its corresponding active layer set, a pre-configured target shortest path search method is invoked to solve the path problem. This search uses the pin-sensing potential energy field as a global guide, allowing paths to naturally converge in the pin region. It strictly follows the preferred routing direction of each metal layer to avoid illegal oblique routing. When necessary to cross layers, it comprehensively evaluates the rationality of via locations, inter-layer alignment accuracy, and manufacturing risks, and selects the optimal via introduction strategy. Ultimately, each connection segment outputs a discrete path with clear layer affiliation, complete coordinate sequence, compliant direction, and controllable vias. This segmented processing preserves the overall routing intent of multi-pin networks while avoiding the challenges of high-dimensional joint optimization, providing a scalable and verifiable path foundation for generating a unified routing guidance file downstream. Based on the discrete paths, the rectangular routing guidance region and explicit pin access guidance region for each network to be routed are obtained.

[0048] In a specific implementation of step S1043, one embodiment is as follows: The rectangular routing guide region and pin access guide region for each network to be routed are obtained based on discrete paths, including: S10431. The discrete path is converted into an initial rectangular routing guide region aligned with the global routing unit grid, and adjacent initial rectangular routing guide regions on the same layer are merged to obtain a rectangular routing guide region. S10432. For each pin of a global routing unit in a network to be routed, explicitly generate a pin access guide area that covers the entire area of ​​the global routing unit.

[0049] In steps S10431-S10432, for each discrete path obtained through path search for the network to be routed, this application first converts it into an initial rectangular routing guide region that is strictly aligned with the global routing unit grid. This conversion process maps the discrete path point sequence to the coordinates of the center grid of the global routing unit it passes through, and expands the preset width along the path direction to form a set of continuous, non-rotating rectangular units with sides parallel to the coordinate axes, ensuring full compatibility with the grid-based modeling system of the back-end physical design process. Subsequently, a geometric merging operation is performed on multiple initial rectangular routing guide regions belonging to the same metal layer and spatially adjacent (sharing edges or corners) to eliminate redundant boundaries and generate a compact rectangular routing guide region covering the continuous routing channel, significantly reducing the processing complexity of downstream tools and improving resource allocation efficiency. Based on this, for each pin of each network to be routed, an explicit pin access guide region is generated for the global routing cell to which it falls. This region completely covers the entire rectangular area of ​​the corresponding global routing cell and marks the accessible sides (e.g., left / right / top / bottom) and priority information of the pin within the cell, ensuring accurate identification of the pin's physical location and legal connection window during the detailed routing phase. All rectangular routing guide regions and pin access guide regions are organized in a standard structured format, including layer identifiers, mesh coordinates, size attributes, and constraint types, together forming a complete routing guide file that can be directly called by subsequent physical implementation processes, supporting highly consistent and low-conflict multi-layer global routing implementation.

[0050] The routing boot file generated in this application was compared with a reference routing boot file. The comparison results show that, as follows: Figure 6 As shown, the routing guide file (my_route.guide) generated in this application covers 479 networks and generates 1256 rectangular routing guide regions; the reference routing guide file (route.guide) covers 487 networks and generates 3095 rectangular routing guide regions. It is evident that the number of rectangular routing guide regions generated in this application is significantly less than that in the reference file, indicating a more compact generation result.

[0051] From the perspective of metal layer usage, the routing guidance areas generated in this application are mainly distributed on three low- or mid-low-level metal layers: li1, met1, and met2. Specifically, li1 contains 801 rectangles, met1 contains 402 rectangles, and met2 contains 53 rectangles. In contrast, the reference routing guidance file has rectangles distributed on li1, met1, met2, met3, and met4, with 1141, 1316, 525, 87, and 26 rectangles respectively. This result indicates that this application tends to generate more compact low-level routing guidance areas in this design, while still being able to select higher metal layer resources based on congestion and layer scoring mechanisms. Looking at the proportion of rectangles in each network, the median proportion of rectangles generated in this application is approximately 0.43, meaning that most networks use fewer rectangles than the reference routing guide file. Bounding box overlap statistics show that most networks have inclusion, being included, or partially overlapping relationships with the reference routing guide bounding box. There are 218 networks with inclusion relationships, 163 networks with being included relationships, and 87 networks with partial overlap relationships. These results indicate that the routing guide region generated by this invention has a high spatial correlation with the reference routing guide region, while also exhibiting good compactness.

[0052] From the visualization results, such as Figure 7 As shown, the routing guidance region generated in this application forms a continuous rectangular coverage area across the entire chip and exhibits a clear layer distribution on different metal layers. The surface height reflects the GCell guidance density, allowing for a direct comparison of the congestion distribution across layers. Compared to the reference routing guidance file, this invention generates fewer rectangles with a more concentrated distribution, making it suitable as input for downstream detailed routers.

[0053] The above embodiments are only used to illustrate specific implementations of the present invention. The present invention is not limited to GCD design, SkyWater 130nm process, or chip design tool flow; any integrated circuit design that can be represented as a multi-layer GCell routing diagram and has edge capacity constraints can use the method of the present invention to perform global routing.

[0054] Example 2 This application also provides a multi-layer global routing device based on edge resource continuous flow modeling, such as... Figure 8 The diagram shows a block diagram of a multi-layer global routing device based on edge resource continuous flow modeling. The function implemented by this device corresponds to the steps described above in executing a multi-layer global routing method based on edge resource continuous flow modeling on a terminal device. This device can be understood as a server component including a processor. The multi-layer global routing device based on edge resource continuous flow modeling described in this application includes: The acquisition module 801 is used to construct a multi-layer global routing cell mesh diagram based on the layout exchange file and design exchange file of the acquired integrated circuit board, and generate a pin sensing potential energy field by using the three-dimensional coordinates of all pins of each network to be routed in the multi-layer global routing cell mesh diagram. The setting module 802 is used to set the horizontal side current variables and vertical side current variables for each network to be wired, so as to construct a multi-objective optimization function for optimizing the horizontal side current variables and vertical side current variables with the pin sensing potential field, and solve for the optimized target horizontal side current variables and target vertical side current variables. The reconstruction module 803 is used to reconstruct the cell-level occupancy confidence field of each global routing unit on each metal layer of the multi-layer global routing unit mesh graph based on the target horizontal edge flow tensor and the target vertical edge flow tensor, calculate the layer score of each network to be routed on each metal layer based on the cell-level occupancy confidence field and the calculated real-time congestion degree of each metal layer, and determine the set of active layers. The execution module 804 is used to determine the pin type of each network to be routed, determine the guide region generation strategy with conversion mechanism based on the pin type, and execute the guide region generation strategy in the active layer set to obtain the rectangular routing guide region and pin access guide region of all networks to be routed, so as to obtain a routing guide file for downstream use and realizing multi-layer global routing.

[0055] In one feasible implementation, the execution module includes: A target shortest path search method is pre-configured, guided by the pin-sensing potential energy field and constrained by the preferred wiring direction of the metal layer and the cost of interlayer vias. If the current wiring network is a low-pin network, then a discrete path is obtained by directly executing the pre-configured target shortest path search method based on the set of active layers; The rectangular routing guide area and explicit pin access guide area for each network to be routed are obtained based on the discrete path.

[0056] In one feasible implementation, the execution module further includes: If the current wiring network is a multi-pin network, then construct a rectangular minimum spanning tree topology and decompose it into multiple two-pin connection segments; For each two-pin connection segment, the target shortest path search method is performed within its active layer set to obtain discrete paths.

[0057] In one feasible implementation, the execution module also includes: The discrete path is transformed into an initial rectangular routing guide region aligned with the global routing cell grid, and adjacent initial rectangular routing guide regions on the same layer are merged to obtain a rectangular routing guide region. For each pin of a global routing cell in a network to be routed, explicitly generate a pin access guide region that covers the entire area of ​​that global routing cell.

[0058] In one feasible implementation, the reconstruction module includes: Based on the unit-level occupancy confidence field, spatial dimensions are summed to obtain the confidence quality, and the layer score is calculated from the confidence quality. By comparing the layer score of each metal layer with the relative threshold of the layer score, the comparison results are obtained to determine the active layers, and all active layers are integrated to obtain the active layer set.

[0059] In one feasible implementation, the reconstruction module further includes: Obtain and compare the average congestion level of the underlying metal stack with the corresponding preset congestion threshold to obtain the comparison result; Based on the comparison results, the layer score weights of the upper metal layers are dynamically adjusted to calculate the layer score based on the adjusted layer score weights, the degree of orientation matching, the via distance, and the degree of congestion.

[0060] In one feasible implementation, the setting module includes: Based on the pin-sensing potential energy field, edge capacity, via cost, and preferred wiring direction of the metal layer, a geometric guidance term, an edge capacity overflow term, an inter-layer switching regularization term, and a preferred direction penalty term are constructed respectively. By integrating the geometric guidance term, edge capacity overflow term, inter-layer switching regularization term, and preferred direction penalty term, a multi-objective optimization function is obtained for continuous gradient optimization.

[0061] In one feasible implementation, the acquisition module includes: Map all pins of each network to be routed to the multilayer global routing unit mesh diagram to obtain its corresponding three-dimensional coordinate set; The potential energy value of the target node is calculated based on the Manhattan distance between the three-dimensional coordinate set and the target node coordinates, as well as the via cost, to generate a pin-sensing potential energy field.

[0062] Example 3 This application also provides an electronic device, such as Figure 9 As shown, it includes: a processor 901, a memory 902, and a bus 903. The memory 902 stores machine-readable instructions that can be executed by the processor 901. When the electronic device is running, the processor 901 and the memory 902 communicate through the bus 903. When the machine-readable instructions are executed by the processor 901, they perform any of the steps of the multi-layer global routing method based on edge resource continuous flow modeling described above.

[0063] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection can be through some communication interfaces; the indirect coupling or communication connection of devices or modules can be electrical, mechanical, or other forms.

[0064] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0065] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0066] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a platform server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0067] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multi-layer global routing method based on edge resource continuous flow modeling, characterized in that, The method includes: Based on the obtained layout exchange file and design exchange file of the integrated circuit board, a multi-layer global routing cell mesh diagram is constructed, and a pin-sensing potential energy field is generated by using the three-dimensional coordinates of all pins of each network to be routed in the multi-layer global routing cell mesh diagram. For each network to be routed, set horizontal and vertical side current variables to construct a multi-objective optimization function for optimizing the horizontal and vertical side current variables with the pin-sensing potential field, and solve for the optimized target horizontal and vertical side current variables. Based on the target horizontal edge flow tensor and the target vertical edge flow tensor, the cell-level occupancy confidence field of each global routing cell in the multi-layer global routing cell mesh diagram is reconstructed on each metal layer. Based on the cell-level occupancy confidence field and the calculated real-time congestion level of each metal layer, the layer score of each network to be routed on each metal layer is calculated, and the set of active layers is determined. The pin type of each network to be routed is determined, and a guide region generation strategy with a conversion mechanism is determined based on the pin type. The guide region generation strategy is then executed in the set of active layers to obtain rectangular routing guide regions and pin access guide regions for all networks to be routed, so as to obtain a routing guide file that can be used downstream and realize multi-layer global routing.

2. The method according to claim 1, characterized in that, The routing guide region generation strategy is executed to obtain rectangular routing guide regions and explicit pin access guide regions for all networks to be routed, including: A target shortest path search method is pre-configured, guided by the pin-sensing potential energy field and constrained by the preferred wiring direction of the metal layer and the cost of interlayer vias. If the current wiring network is a low-pin network, then a discrete path is obtained by directly executing the pre-configured target shortest path search method based on the set of active layers; The rectangular routing guide area and explicit pin access guide area for each network to be routed are obtained based on the discrete path.

3. The method according to claim 2, characterized in that, If the current routing network is not a low-pin network, it includes: If the current wiring network is a multi-pin network, then construct a rectangular minimum spanning tree topology and decompose it into multiple two-pin connection segments; For each two-pin connection segment, the target shortest path search method is performed within its active layer set to obtain discrete paths.

4. The method according to claim 2, characterized in that, Based on discrete paths, the rectangular routing guide area and pin access guide area for each network to be routed are obtained, including: The discrete path is transformed into an initial rectangular routing guide region aligned with the global routing cell grid, and adjacent initial rectangular routing guide regions on the same layer are merged to obtain a rectangular routing guide region. For each pin of a global routing cell in a network to be routed, explicitly generate a pin access guide region that covers the entire area of ​​that global routing cell.

5. The method according to claim 1, characterized in that, Based on the cell-level occupancy confidence field and the calculated real-time congestion level of each metal layer, the layer score of each network to be routed on each metal layer is calculated, and the set of active layers is determined, including: Based on the unit-level occupancy confidence field, spatial dimensions are summed to obtain the confidence quality, and the confidence quality is used to calculate the layer score. By comparing the layer score of each metal layer with the relative threshold of the layer score, the comparison results are obtained to determine the active layers, and all active layers are integrated to obtain the active layer set.

6. The method according to claim 5, characterized in that, The confidence quality is calculated to obtain the layer score, including: Obtain and compare the average congestion level of the underlying metal stack with the corresponding preset congestion threshold to obtain the comparison result; Based on the comparison results, the layer score weights of the upper metal layers are dynamically adjusted to calculate the layer score based on the adjusted layer score weights, the degree of orientation matching, the via distance, and the degree of congestion.

7. The method according to claim 1, characterized in that, A multi-objective optimization function for optimizing horizontal and vertical sideflow variables is constructed using the pin-sensing potential field, including: Based on the pin-sensing potential energy field, edge capacity, via cost, and preferred wiring direction of the metal layer, a geometric guidance term, an edge capacity overflow term, an inter-layer switching regularization term, and a preferred direction penalty term are constructed respectively. By integrating the geometric guidance term, edge capacity overflow term, inter-layer switching regularization term, and preferred direction penalty term, a multi-objective optimization function is obtained for continuous gradient optimization.

8. The method according to claim 1, characterized in that, A pin-sensing potential field is generated using the three-dimensional coordinates of all pins in the multilayer global routing cell mesh diagram for each network to be routed, including: Map all pins of each network to be routed to the multilayer global routing unit mesh diagram to obtain its corresponding three-dimensional coordinate set; The potential energy value of the target node is calculated based on the Manhattan distance between the three-dimensional coordinate set and the target node coordinates, as well as the via cost, to generate a pin-sensing potential energy field.

9. A multi-layer global routing device based on edge resource continuous flow modeling, characterized in that, The device includes: The acquisition module is used to construct a multi-layer global routing cell mesh diagram based on the obtained layout exchange file and design exchange file of the integrated circuit board, and generate a pin sensing potential energy field by using the three-dimensional coordinates of all pins of each network to be routed in the multi-layer global routing cell mesh diagram. The configuration module is used to set the horizontal and vertical side current variables for each network to be wired, so as to construct a multi-objective optimization function for optimizing the horizontal and vertical side current variables with the pin sensing potential field, and solve for the optimized target horizontal and vertical side current variables. The reconstruction module is used to reconstruct the cell-level occupancy confidence field of each global routing unit on each metal layer of the multi-layer global routing unit mesh graph based on the target horizontal edge flow tensor and the target vertical edge flow tensor. Based on the cell-level occupancy confidence field and the calculated real-time congestion degree of each metal layer, the module calculates the layer score of each network to be routed on each metal layer and determines the set of active layers. The execution module is used to determine the pin type of each network to be routed, determine the guide region generation strategy with conversion mechanism based on the pin type, and execute the guide region generation strategy in the active layer set to obtain the rectangular routing guide region and pin access guide region of all networks to be routed, so as to obtain the routing guide file for downstream use and realize multi-layer global routing.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of a multi-layer global routing method based on edge resource continuous flow modeling as described in any one of claims 1-8.

Images