Staged preview collaborative optimization method for collaborative layout of macro cells and standard cells

Abstract

The invention relates to the technical field of layout optimization in integrated circuit physical design, in particular to a staged preview collaborative optimization method for collaborative layout of macro cells and standard cells, and aims to solve the problems of poor collaboration, low overlap repair efficiency and high wiring congestion risk in mixed layout of the macro cells and the standard cells. According to the technical scheme, the method comprises the steps that firstly, a component neighborhood is constructed based on modularity and pairing weight, a neighborhood mode is dynamically switched through a differential evolution operator to generate a rehearsal population, and high-quality individuals are maintained through a double-end queue file; two-stage collaborative optimization is adopted, and adaptability optimization and typed overlap repair (only corresponding types of units are moved) are executed for macro units and standard units; the effect is evaluated through a combined cost function containing a line length and overlapping area weighted item, when any one condition of evaluation budget, overlapping standard reaching or too low improvement rate is met, termination is carried out, and local search optimization is disturbed through a tail end weight at the global optimal position. The layout quality and optimization efficiency can be improved, and the method is suitable for chip back-end layout design, power-assisted balance time sequence performance and wiring feasibility.

Description

Technical Field

[0001] This invention belongs to the field of physical design technology of very large-scale integrated circuits (VLSI), specifically involving a phased pre-evolutionary collaborative optimization method for the collaborative layout of macrocells and standard cells. Background Technology

[0002] The placement phase of chip physical design, as the first step in physical synthesis, fundamentally determines the performance ceiling for subsequent routing, timing closure, and power optimization. The core task of this phase is to find the optimal coordinate arrangement for two types of heterogeneous cells within a given planar boundary: one is macrocells (such as SRAM arrays, high-speed interface IPs, etc.), characterized by large area, regular shape, strict constraints, and small number; the other is standard cells (such as AND, OR, NOT gates, flip-flops, etc.), characterized by a huge number (up to tens of thousands), tiny size (usually micrometer-level), and complex and extensive coupling relationships. When these two types of objects are collaboratively laid out in a unified plane, four basic constraints must be met simultaneously: In terms of geometric feasibility, all modules must be strictly located within the chip outline, and no physical overlap should occur between any modules, otherwise the layout will be unmanufacturable; in terms of electrical reachability, the overall interconnection cost (usually measured by weighted half-circuit length HPWL) must be minimized to reduce routing difficulty and signal delay; in terms of scalability, the algorithm must maintain linear or quasi-linear computational efficiency in larger design spaces with hundreds to tens of thousands of modules; and in terms of connectivity, the output layout result should have a clear cell distribution and reasonable white space to facilitate subsequent congestion control, clock tree synthesis, and signal integrity optimization.

[0003] However, the existing mainstream methods mainly include simulated annealing heuristics, single-population differential evolution or genetic algorithms, and gradient or force-directed methods. However, these technologies have revealed the following five core shortcomings in practical engineering applications: First, the quality of initial solutions is low. The random initialization strategy leads to a significant waste of early evaluation costs in poor solution spaces that are clearly infeasible or far from the optimal solution region, thus lowering the overall search efficiency. Second, high-dimensional coupling is difficult to resolve. The complex coupling relationships between macrocells, between macrocells and standard cells, and between standard cells make it easy for full-dimensional perturbations to destroy the already formed high-quality substructures, leading to local optima. Third, overlap processing is costly. The overlap detection and repair mechanism, which is based entirely on pairwise traversal, exhibits quadratic computational complexity in large-scale scenarios, becoming a performance bottleneck. In addition, the operator function is singular. The same evolutionary operator must simultaneously undertake the dual tasks of global exploration and local convergence, which can easily lead to premature convergence of the population or oscillations in solution quality. Finally, there is a lack of typified strategies. Macrocells and standard cells have orders of magnitude differences in geometric scale, connection weights, and arrangement constraints, but they are treated uniformly without a targeted search mechanism, resulting in unbalanced resource allocation and low optimization efficiency. Therefore, the industry urgently needs a new layout method that can take into account prior solution domain construction, phased collaborative optimization and typed overlap processing, so as to systematically improve the quality and efficiency of large-scale hybrid cell layout and meet the physical design requirements under advanced process nodes. Summary of the Invention

[0004] Under the premise of satisfying planar boundary and module size constraints, this invention rapidly obtains a layout scheme with zero overlap of macrocells, low overlap of standard cells, and low interconnection cost, and provides an engineering process that can be scaled up to a larger scale and integrated with subsequent routing / congestion optimization. This invention proposes a phased pre-evolutionary collaborative optimization process, including the following steps:

[0005] S1 Component Neighborhood Construction and Pre-Evolution Population Generation

[0006] As the first step in this method, S1 aims to extract an elite population with structure-guided capabilities from the initial solution space, providing a high-quality search starting point for subsequent co-evolution.

[0007] S1.1 First, in the chip plane area The inside contains all Each module (including macro elements and standard elements) generates initial coordinates. And according to its geometric dimensions To perform boundary clamping, that is, to ensure and This satisfies the feasible region initialization under hard constraints.

[0008] S1.2 Prioritize modules based on their degree and pair weight. The degree of module is defined as the sum of the net weights associated with that module. The pairing weight is the cumulative value of the net weights between the two units. Based on the above metrics, this step constructs several "component neighborhoods" covering all modules. Each neighborhood contains Each module Dimension, that is ,in For a subset of module indexes, This type of neighborhood partitioning preserves the opportunity for cooperative evolution of highly connected units while reducing the dimensionality of sub-problems.

[0009] S1.3 is the core evolutionary stage, based on the differential evolutionary family operator, which operates between two evolutionary ranges, "full neighborhood" and "single neighborhood," according to a set probability. (Optimal 0.9) Dynamic switching. In full neighborhood mode, for each individual in the population... Execute the DE / rand / 1 mutation strategy, i.e., generate experimental vectors. , where scaling factor Adaptive adjustment, cross rate Random perturbation; in single-neighbor mode, a neighborhood is randomly selected. For in-depth development, adopt the DE / current-to-best / 2 strategy. Only active dimensions are updated, thus achieving a balance between exploration and development.

[0010] S1.4 via external archives To maintain a diverse pool of high-quality individuals, the archive employs a two-ended queue structure with a maximum capacity equal to the population size. When a new elite individual joins and the file is full, the first-place elimination strategy is implemented to ensure that the file always retains the latest and best solution domain information, forming a reliable reference context for subsequent co-evolution.

[0011] S2 Two-Phase Collaborative Optimization

[0012] S2 is the main optimization framework of this method, which adopts a phased strategy of prioritizing macrocells and following up with standard cells, and handles the differences between heterogeneous cells in a typological manner.

[0013] S2.1 Phase 1 (Macrocell Priority) Based on macrocell sets To optimize objects, by neighborhood Group by group in global context Optimize subvectors, that is, extract only the corresponding dimensions of this set. The process involves evolution, and the optimal sub-vector is written back to the globally optimal individual to avoid global perturbations from disrupting the already converged structure. This stage employs an enhanced SaNSDE operator as a sub-optimizer, utilizing adaptive parameters and local perturbations to accelerate macrocell position locking, and periodically triggering a dedicated macro-repair mechanism.

[0014] S2.2 Phase Two (Standard Unit Follow-up) uses a set of standard units. For the optimization target, the same "sub-component group optimization" method as in Phase 1 is adopted. Due to the larger scale of the sub-components, it is adapted to the characteristics of a large number of standard units and loose coupling. Phase 2 also adopts a standard unit-specific repair operator based on the optimization of the original S2.1.

[0015] S2.3 triggers typed overlap repair at a preset frequency (e.g., every 3 iterations) in both phases: only macrocells are moved in phase one, and only standard cells are moved in phase two. Mutual interference is avoided through solution space separation, ensuring the dual goals of macrostructure stability and standard cell flexibility.

[0016] S3 Typed Overlap Quick Fix

[0017] S3 is the geometric feasibility assurance module of this method. It designs exclusion strategies for macro elements and standard elements respectively to avoid structural crosstalk during the repair process.

[0018] S3.1 Macrocell overlap repair uses the overlap area as weight to calculate the vectorized repulsive force driving macrocell separation. For macrocell pairs... The overlapping area is defined as The repulsion direction is the vector connecting the centers. Decision made: Update step size to ,in The repulsion coefficient is used. The repair iteration terminates at the total overlap area. This ensures that there are no hard conflicts between macrocells.

[0019] S3.2 Standard element overlap repair uses a similar mechanism, but the target is limited to the set of standard elements. The repulsion coefficient was adjusted to The number of iterations can be increased appropriately (e.g., 80-120 times) to avoid affecting the converged structure of macrocells and achieve hierarchical repair.

[0020] S4 Fast Overlap Detection and Fractal Accumulation

[0021] S4 is the core of this method's efficiency, dynamically selecting the detection strategy based on the module size to balance computational accuracy and overhead.

[0022] S4.1 When the module size (threshold) When the optimal value is 450, matrix broadcasting is used to calculate the overlapping region of the upper triangular pairs of all modules at once. Specifically, a coordinate matrix is ​​constructed. With size matrix Calculated through broadcast mechanism Similarly, calculate The overlapping mask is Subsequently, based on the macrocell flag vector With standard unit flag vector Perform fractal accumulation: Macro-macro overlap area is The calculations for standard units overlapping across classes are similar, resulting in a total complexity of O(n). This allows for full utilization of SIMD acceleration.

[0023] S4.2 When the size exceeds the threshold At that time, a regular grid bucket and nine-square neighborhood enumeration strategy is adopted: the plane is divided into Each module is mapped to its own bucket and neighboring buckets within a grid. Overlapping computations are only performed on modules within the same bucket or neighboring buckets, reducing the complexity to [missing value]. ,in To reduce costs and increase efficiency by averaging the number of modules in the bucket.

[0024] S5 End-Weight Perturbation Local Search

[0025] S5 serves as the convergence accelerator for this method, performing a low-cost reinforcement search in the local neighborhood of the global optimum. Specifically, in the global optimum individual... At this point, a set of dimensions is randomly selected as the "hot zone". A hot zone corresponds to a component neighborhood. Construct differential test vectors The weight vector Set 1 for the hot zone dimension and 0 for the rest, i.e. , , This represents element-wise product. This mechanism only amplifies hot zone perturbations and has no effect on non-hot zones; it is selectively accepted after typed repair: if the cost function... Then, the global optimum is updated. This step greatly strengthens the "current solution → better solution" direction, enabling precise allocation of computing resources.

[0026] S6 Evaluation and Termination

[0027] S6 is the termination decision module of this method, which comprehensively ensures output quality through multiple dimensions of indicators. The evaluation function adopts a combined cost function of weighted HPWL and fractal overlap penalty:

[0028]

[0029] in , As a penalty weight, , , The fractal overlap area is calculated for S4. Termination conditions include: 1) The number of function evaluations reaches the preset budget. (e.g., 30,000 times); 2) Macrocell overlap area (Engineering Zero Standard); 3) Gain Threshold Determination: If the cost improvement rate of the most recent 50 generations is less than 0.1%, the process will be terminated early. At the same time, the phase indicators (HPWL, overlap, cost) are recorded to the historical array for subsequent performance analysis and visualization.

[0030] The present invention has the following advantages and effects compared with the prior art:

[0031] (1) This invention proposes a pre-evolutionary population screening method. Before the main loop of collaborative optimization, the initial population is subjected to multiple rounds of iterative evolution, and elite individuals are dynamically retained by external archives with a double-ended queue structure. Its function is to quickly move the search starting point from a random region to a feasible and low-cost high-quality solution domain, thereby shortening the convergence time of macrocells with zero overlap by about 55% and significantly reducing the waste of early ineffective evaluation.

[0032] (2) This invention proposes a two-stage collaborative optimization method of "sub-component-context", which separates macro-units and standard units into two stages for independent scheduling. In the first stage, the macro-unit position is locked with priority at the group size, and in the second stage, the standard unit is optimized with a larger group size. Both stages use the globally optimal individual as the context anchor point. Its function is to avoid the destruction of the converged structure by full-dimensional disturbance through dimensional decoupling and information isolation, thereby improving the stability of the high-dimensional optimization process by 40% and effectively suppressing premature convergence.

[0033] (3) This invention proposes a typed overlap fast repair method, which designs independent vectorized repulsion functions for macrocells and standard cells respectively. The macrocell repair configuration has a high repulsion coefficient and few iterations, while the standard cell has the opposite. The repair scope is strictly limited to avoid mutual interference between heterogeneous cells. At the same time, combined with the grid bucket hierarchical detection strategy, the computational complexity is reduced from Down to Support 10 5 Engineering deployment at the unit level.

[0034] (4) This invention proposes an end-group granular weight perturbation method. At the end of the evaluation budget, a component neighborhood is randomly selected as a hot zone. The weight vector is constructed to amplify the differential perturbation of the hot zone dimension only, and low-cost local repair is selectively performed. Its function is to further reduce HPWL by about 3%~7% at the budget edge with minimal overhead of a single DE operation, thereby maximizing marginal benefits. Attached Figure Description

[0035] Figure 1 is an overall flowchart of the phased pre-evolutionary collaborative optimization method for the collaborative layout of macrocells and standard cells according to the present invention.

[0036] Figure 2 is a detailed flowchart of component neighborhood construction and pre-evolution population generation.

[0037] Figure 3 shows the module scheduling logic diagram for two-stage collaborative optimization.

[0038] Figure 4 is a schematic diagram of the force-directed principle for rapid repair of typified overlap.

[0039] Figure 5 is a schematic diagram of the hierarchical strategy for fast overlap detection.

[0040] Figure 6 is a schematic diagram of the local search process with end weight perturbation. Detailed Implementation

[0041] Figure 1 The invention clearly presents the six core stages from initial solution generation to final layout output and their logical connections, with each stage forming a complete optimization loop through data interaction. The specific implementation of the invention is described in detail below with reference to the accompanying drawings and practical application scenarios. Other drawings are intended to aid in understanding the core logic and execution flow of the technical solution.

[0042] Assume the following placement and routing task, where the chip planar dimensions are 150.0μm × 150.0μm (boundary range [0, 150.0] × [0, 150.0]); module configuration: 5 macrocells (size 30.0-45.0μm), 300 standard cells (size 2.0-5.0μm), with each module associated with an average of 3 nets; core algorithm parameters: population size 60, pre-evolution budget accounting for 20% of the total evaluation iterations, macrocell group size 2, standard cell group size 10, differential evolution operator scaling factor. Cross rate Termination conditions: The maximum number of function evaluations is 30,000, the macrocell overlap area is ≤10⁻³μm², or the cost improvement rate for 50 consecutive generations is <0.1%. The specific implementation steps are as follows:

[0043] (I) S1 Component Neighborhood Construction and Pre-evolutionary Population Generation

[0044] The core objective of this step is to improve the quality of the initial solution, providing a high-quality search starting point for subsequent co-optimization. The detailed process is attached. Figure 2 As shown.

[0045] S1.1 Feasible Region Initialization: Randomly generate the initial coordinates of all modules within the chip plane [0,150.0]×[0,150.0]. Combined with the module's own dimensions Implement boundary clamping to ensure and This prevents the initial position from exceeding the chip boundary.

[0046] S1.2 Module Priority Ranking: Calculate two key indicators, module degree and pairing weights Based on the weighted sum of these two metrics, all modules are ranked, prioritizing the retention of modules with close connections. Component neighborhood partitioning: All modules are divided into several "component neighborhoods" according to the ranking result, with each neighborhood containing 8 modules. Coordinate dimensions (i.e., each module corresponds to 2 dimensions) (As a subset of module indexes), this partitioning method preserves the opportunity for collaborative evolution of highly connected modules while reducing the dimensionality of subproblems.

[0047] S1.3 Dual-mode pre-evolution: Differential evolution operators are used to dynamically switch between "full neighborhood" and "single neighborhood" modes with a 90% probability. In the full neighborhood mode, DE / rand / 1 mutation is performed on all dimensions of each individual in the population, generating the experimental vector. In single-neighborhood mode, a component neighborhood is randomly selected, and DE / current-to-best / 2 mutation is performed only on the dimensions within that neighborhood, thus generating the experimental vector. The two modes achieve a balance between global exploration and local development, respectively.

[0048] S1.4 Elite Population Preservation: External Archives via a Double-Ended Queue Structure Preserve high-quality individuals, with the file size matching the population size (60 individuals). When a new elite individual joins and the file is full, eliminate the worst individual at the top of the queue to ensure that the file always retains the latest and best solution domain information. Finally, output the file as the initial population for subsequent collaborative optimization.

[0049] (II) S2 Two-stage Cooperative Optimization

[0050] This step employs a phased strategy of "macrocell priority, standard cell follow-up" to address the layout differences of heterogeneous cells. Its scheduling logic is shown in the attached figure. Figure 3 As shown.

[0051] S2.1 Phase 1: Macrocell Priority Optimization. Optimize 5 macrocells according to the attached... Figure 3 The process on the left is as follows: First, extract the sub-vectors corresponding to the macrounits from the globally optimal individual. ( For the macrocell component neighborhood), sub-optimization is performed using the enhanced SaNSDE operator, through adaptive adjustment. and The parameters accelerate macrocell location convergence; a dedicated macrocell repair mechanism is triggered every 3 iterations (see S3), which moves only macrocells to eliminate overlap and avoids global disturbances that could damage the already formed high-quality structure. This stage consumes 45% of the total number of evaluations (approximately 13,500) and completes macrocell location locking.

[0052] S2.2 Phase Two: Standard Cell Follow-up Optimization. Optimization is performed on 300 standard cells, and the execution logic is as follows: Figure 3 As shown on the right: the sub-component size is expanded to 10 (to accommodate the large number of standard cells and their loose coupling); similarly, a dedicated standard cell repair mechanism is triggered every 3 iterations, moving only the standard cells without interfering with the positions of converged macrocells, ensuring both macrostructure stability and standard cell flexibility. This stage consumes the remaining evaluation attempts until the total budget of 30,000 is reached.

[0053] (III) S3 Typed Overlap Fast Repair

[0054] This step addresses the characteristic differences between macro-elements and standard elements by designing an independent overlap repair mechanism to avoid structural interference during the repair process. Its force-directing principle is detailed in the attached figure. Figure 4 As shown.

[0055] S3.1 Macrocell Overlap Repair: See Appendix Figure 4 The diagram on the left shows macrocell repair. First, the macrocell pairs are calculated. overlapping area

[0056] ;

[0057] Then based on the center connection vector of the two macrocells Determine the direction of repulsion, according to Calculate the movement step size (0.35 is the macrocell repulsion coefficient); iteratively perform this repair operation until the total overlap area is <10⁻ 6 μm², ensuring no hard conflicts between macrocells.

[0058] S3.2 Standard cell overlap repair: The same force-directed mechanism as macrocells is adopted, but the rejection coefficient of the standard cell repair logic is adjusted to 0.20, the number of iterations is increased to 80-120, and the repair range is strictly limited to the standard cell set to avoid moving macrocells and ensure that the converged macrostructure is not affected.

[0059] (iv) S4 Fast Overlap Detection and Classification Accumulation

[0060] This step dynamically selects the detection strategy based on scale, balancing computational accuracy and efficiency. Its hierarchical switching logic is shown in the appendix. Figure 5 As shown.

[0061] S4.1 Small-scale module detection (N≤450): Matrix broadcasting method is used, as shown in the attached diagram. Figure 5 As shown in the process above, a coordinate matrix is ​​constructed. Size matrix The x-direction overlap of all module pairs is calculated at once using a broadcast mechanism. Overlapping with the y-direction Based on overlapping masks Calculate the overlapping area and accumulate it separately for three categories: macro-macro, standard-standard, and macro-standard.

[0062] S4.2 Large-scale module detection (N>450): Employs a regular grid bucket + nine-square neighborhood enumeration strategy, as shown in the attached diagram. Figure 5 As shown in the flowchart below, by dividing the chip plane into 15×15 grids (grid size 10.0μm), each module is mapped to its own grid and adjacent 3x3 grids. Overlapping calculations are only performed on module pairs within the same or neighboring buckets, reducing complexity from... Down to ( (This is the average number of modules in the bucket), thus achieving cost reduction and efficiency improvement.

[0063] (v) Local search with S5 end weight perturbation

[0064] This step performs reinforcement search in the local neighborhood of the global optimum to further improve the optimization quality with extremely low overhead. Its operational logic is shown in the appendix. Figure 6 As shown.

[0065] S5.1 Select and activate the current hot zone neighborhood: in the globally optimal individual In the module distribution, a neighborhood of a randomly selected component is designated as a "hot zone". This hot zone corresponds to the coordinate dimensions of a set of highly connected modules. Simultaneously, activating this hot zone initializes the basic structure of the experimental vectors, ensuring that subsequent perturbations are based on a stable starting point of the current optimal solution, thus avoiding global offsets.

[0066] S5.2 Constructing a weighted population for hot zone perturbations: Constructing a weight vector The dimension corresponding to the hot zone is set to 1, and the other dimensions are set to 0; a perturbation population is generated based on this weight vector, including multiple differential experimental vector variants (such as...). , (For element-wise multiplication), only the differential perturbation in the hot zone dimension is amplified, and it has no effect on the non-hot zone module.

[0067] S5.3 Differential Evolution Optimization Local Search: A local optimization search is performed within the hot zone using differential evolution operators, with adaptive parameter adjustments (such as scaling factors). and cross rate Generate and evaluate individuals in the perturbation population to achieve a balance between fine exploration and development of hot zone coordinates, ensuring that the search covers a wider neighborhood space and avoids premature convergence.

[0068] S5.4 Attempt to repair legalization: Perform standard unit-specific repair mechanisms on the experimental vectors generated by differential evolution, including typified overlap repair and boundary clamping, limited to the overlap and connection adjustment of units within the hot zone, to eliminate potential geometric conflicts and constraint violations, ensuring individual legalization while minimizing interference with the overall structure.

[0069] S5.5 Selectively accept the best individual: Calculate the cost function value of the best individual after repair; if Then, the best solution is selected and updated to the globally optimal solution, achieving precise iteration from "current solution → better solution". This step emphasizes the targeted selection, ensuring that the updated solution is not only optimized on HPWL, but also stable within the overall feasible region, ready to be output to the termination decision module.

[0070] (vi) S6 Evaluation and Termination

[0071] The layout quality is quantitatively evaluated using a combined cost function: HPWL is the weighted half-period length (measuring the cost of connecting lines). , , These are the overlapping areas of macro-macro, standard-standard, and macro-standard cells, respectively (as a penalty).

[0072] When any of the following termination conditions are met, output the final layout scheme and record metrics such as HPWL, overlap area, and computation time:

[0073] The function has been evaluated up to the preset budget of 30,000 times;

[0074] Macrocell overlap area μm² (engineering zero standard);

[0075] The cost improvement rate of the last 50 generations is less than 0.1% (converging and stable).

Claims

1. A phased pre-evolutionary collaborative optimization method for the collaborative layout of macrocells and standard cells, characterized in that, Includes the following steps: S1 Component Neighborhood Construction and Pre-Evolution Population Generation Extracting elite populations with structure-guided capabilities from the initial solution space provides a high-quality search starting point for subsequent co-evolution; S2 Two-Phase Collaborative Optimization A phased strategy of prioritizing macrocells and following up with standard cells is adopted to handle heterogeneous cell differences in a typological manner. S3 Typed Overlap Quick Fix Rejection strategies were designed for macrocells and standard cells respectively to avoid structural crosstalk during the repair process; S4 Fast Overlap Detection and Fractal Accumulation Dynamically select detection strategies based on module size to balance computational accuracy and overhead; S5 End-Weight Perturbation Local Search Perform a low-cost reinforcement search in the local neighborhood of the global optimum; S6 Evaluation and Termination To ensure output quality, a multi-dimensional set of indicators is used, and the evaluation function employs a combined cost function of weighted HPWL and fractal overlap penalty.

2. The phased pre-evolutionary collaborative optimization method for macrocell and standard cell collaborative layout according to claim 1, characterized in that, The S1 component neighborhood construction and pre-evolution population generation include: S1.1 Generate initial coordinates for all modules within the chip plane and perform boundary clamping based on their geometric dimensions, thus ensuring... and ,in Let i be the coordinates of module i. For its size, This refers to the chip's planar dimensions; S1.2 Prioritize modules based on modularity and pairing weights, and construct several component neighborhoods covering all modules. Each neighborhood contains 2K dimensions of K modules, i.e. Where K is the size of the module index subset; S1.3 Based on the differential evolution family operator, the evolution is carried out by dynamically switching between two evolution ranges, the full neighborhood and the single neighborhood, according to a set probability; S1.4 Maintain a diverse range of high-quality individuals through external archives, which employ a double-ended queue structure. When a new elite individual joins and the archive is full, a head-of-the-queue elimination strategy is implemented.

3. The phased pre-evolutionary collaborative optimization method for macrocell and standard cell collaborative layout according to claim 1, characterized in that, In S1.2, the modularity is defined as the sum of the net weights associated with that unit, and the calculation formula is as follows: ;in Let i be the modularity of module i. The net weights between modules i and j; The pairing weight is the cumulative value of the net weights between two units, calculated using the following formula: ,in Let nets(i,j) be the set of nets connecting modules i and j, where nets(i,j) is the pairing weights for modules i and j. For wire mesh The weight.

4. The phased pre-evolutionary collaborative optimization method for the collaborative layout of macrocells and standard cells according to claim 1, characterized in that, In the full neighborhood mode described in S1.3, the DE / rand / 1 mutation strategy is executed for each individual in the population to generate an experimental vector. The calculation formula is as follows: ,in For the test vector, , , For randomly selected individuals, F is the scaling factor; In single-neighborhood mode, the DE / current-to-best / 2 strategy is adopted, updating only the active dimension. The calculation formula is as follows: ,in For the current individual, For the optimal individual, F is the scaling factor.

5. The phased pre-evolutionary collaborative optimization method for the collaborative layout of macrocells and standard cells according to claim 1, characterized in that, The S2 two-stage collaborative optimization includes: S2.1 Phase 1 (Macrocell Priority): Taking the macrocell set as the optimization object, optimize the sub-vectors in the global context group by group according to the neighborhood, use the enhanced SaNSDE operator as the sub-optimizer, and periodically trigger the dedicated macro repair mechanism; S2.2 Phase Two (Standard Cell Follow-up): The standard cell set is used as the optimization object. The same sub-components as in Phase One are used for optimization group by group, and a standard cell-specific repair operator is used. S2.3 In both phases, typed overlap repair is triggered at a preset frequency: in phase one, only macro cells are moved, and in phase two, only standard cells are moved.

6. The staged pre-evolutionary collaborative optimization method for macrocell and standard cell collaborative layout according to claim 1, characterized in that, The S3-typed overlap fast repair includes: S3.1 Macrocell overlap repair: Using the overlap area as weight, calculate the vectorized repulsive force to drive macrocell separation, and the repair iteration terminates when the total overlap area is lower than the threshold. S3.2 Standard cell overlap repair: A similar mechanism is used, but the target is limited to the standard cell set, the repulsion coefficient is adjusted to a lower value, and the number of iterations is increased.

7. The phased pre-evolutionary collaborative optimization method for the collaborative layout of macrocells and standard cells according to claim 1, characterized in that, The overlap area of ​​macrocell pairs in S3.1 is defined as follows: ,in Let i be the overlapping area of ​​modules i and j. , The coordinates and dimensions of module i; the repulsion direction is the vector connecting the centers. Decision: Update step size as follows: Where k is the repulsion coefficient, These are the coordinates of the module's center.

8. The phased pre-evolutionary collaborative optimization method for macrocell and standard cell collaborative layout according to claim 1, characterized in that, The S4 fast overlap detection and classification accumulation includes: S4.1 When the module size is less than or equal to the threshold, matrix broadcasting is used to calculate the overlapping area of ​​the upper triangular region of all modules at once, and the classification is accumulated based on the macro unit flag vector and the standard unit flag vector; S4.2 When the size exceeds the threshold, a regular grid bucket and nine-square neighborhood enumeration strategy is adopted to divide the plane into a grid. Each module is mapped to its own bucket and neighboring buckets. Only the modules in the same bucket or neighboring buckets are calculated to overlap.

9. The phased pre-evolutionary collaborative optimization method for the collaborative layout of macrocells and standard cells according to claim 1, characterized in that, In step S4.1, coordinate matrices X and Y and dimension matrices W and H are constructed, and the overlap in the x-direction is calculated through a broadcast mechanism: Similarly, calculate The overlapping mask is The overlapping area is ,in Indicates transpose; macro-macro overlap area is: , where M is the macro-unit marker vector, and ⊙ represents element-wise multiplication.

10. A staged pre-evolutionary collaborative optimization method for the collaborative layout of macrocells and standard cells according to claim 1, characterized in that, The S5 end-weight perturbation local search includes: at the globally optimal individual, randomly select a set of dimensions as hot zones, construct differential test vectors, where the weight vectors are set to 1 on the hot zone dimensions and 0 on the rest, and then selectively accept the best after typified repair.

Images