A clock tree construction method considering low-voltage clock skew fluctuation optimization
By optimizing the balanced grouping and buffer insertion of the clock tree using spectral clustering and genetic algorithms, the problem of clock tree deviation fluctuation under low voltage was solved, thereby improving the stability of the clock tree and the performance of synchronous digital circuits.
Patent Information
- Application Number
- CN202311157381.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-08
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2043-09-08
AI Technical Summary
At low voltage, existing clock tree synthesis algorithms fail to effectively optimize clock skew fluctuations, resulting in poor stability of the clock tree at low voltage, which limits the design and performance of synchronous digital circuits.
By combining spectral clustering and genetic algorithms, a balanced and symmetrical clock tree structure is constructed by balancing the group register nodes and optimizing the buffer insertion position and size, thereby reducing clock deviation fluctuations.
It effectively reduces clock skew fluctuations in the clock tree, improves the stability and robustness of the clock tree under low voltage, and enhances the design performance of synchronous digital circuits.
Smart Images

Figure CN117272878B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to integrated circuit design automation technology, and in particular discloses a clock tree establishment method that takes into account low voltage clock deviation fluctuations, belonging to the technical field of calculation, estimation or counting. Background Technology
[0002] As transistor sizes continue to shrink and chip integration becomes increasingly dense, the resulting heat dissipation issues severely limit chip development. Reducing the chip's operating voltage is one of the most effective methods for low-power chip design. However, at low voltages, the timing characteristics of clock buffers change significantly, primarily in terms of delay and delay ripple. Under the same conditions, compared to constant voltage, the delay and delay ripple of standard cells increase substantially at low voltages, greatly limiting the design and performance of synchronous digital circuits. Therefore, constructing a highly robust clock tree with minimal clock skew near the threshold voltage is crucial for synchronous digital circuit design.
[0003] The clock network mainly consists of buffers and interconnects, and the timing characteristics of both are crucial to the clock skew and fluctuation of the clock tree. Currently, most existing clock tree synthesis algorithms reduce clock skew by considering zero-skew interconnect routing and combining it with dynamic programming to insert buffers. However, they do not consider precise optimization for clock skew fluctuations, resulting in poor stability of the generated clock tree under low voltage conditions. Therefore, this application aims to propose a clock tree construction method optimized for clock skew fluctuations to overcome the above-mentioned shortcomings. Summary of the Invention
[0004] The purpose of this invention is to address the shortcomings of the aforementioned background technology by providing a clock tree establishment method that considers the optimization of low-voltage clock deviation fluctuations based on the timing characteristics of clock units and clock tree features under low voltage. This achieves the invention objective of constructing a highly robust clock tree with smaller clock deviation fluctuations under near-threshold voltage, and solves the technical problem of worsening clock deviation fluctuations in clock trees under near-threshold voltage.
[0005] To achieve the above objectives, the present invention adopts the following technical solution:
[0006] A clock tree establishment method considering low-voltage clock skew fluctuation optimization includes the following steps:
[0007] Step S1: Extract the timing and location information after the reference circuit layout, and merge the register nodes according to the register timing relationship;
[0008] Step S2: For the merged register nodes, the register nodes are evenly grouped according to their location information using a spectral clustering algorithm.
[0009] Step S3: Optimize the register equalization grouping obtained in step S2 and construct the local clock tree for each register equalization group;
[0010] Step S4: Based on the distribution of the drive buffer positions of the local clock tree of each register equalization group constructed in step S3, construct the initial top-level clock tree topology using a top-down symmetrical binary search method.
[0011] Step S5: Generate clock buffer insertion strategy based on genetic algorithm, and insert clock buffers into the initial top-level clock tree topology structure constructed in step S4 according to the clock buffer insertion strategy corresponding to the optimal value of fitness function.
[0012] Step S6: Optimize the initial topology of the top-level clock tree after inserting the clock buffer, and obtain the updated top-level clock tree topology.
[0013] Step S7: Select the clock buffer size based on the genetic algorithm, and optimize the size of the clock buffer in the updated top-level clock tree topology to the buffer size corresponding to the best value of the fitness function.
[0014] As a further optimization of the clock tree establishment method that considers low-voltage clock skew fluctuations, step S1, which involves merging register nodes based on register timing relationships, includes the following steps:
[0015] Step S101: Obtain the timing information of each data path after layout using EDA tool commands;
[0016] Step S102: Construct an undirected time-series weighted graph based on the extracted time-series information, with the time margin between register pairs as the weight;
[0017] Step S103: Cut the edges in the undirected temporal weight graph constructed in step S102 according to the weight threshold, and update the undirected temporal weight graph.
[0018] Step S104: Traverse all register nodes and, based on the updated undirected timing weight graph, merge the current register node and the register node with the smallest timing margin with respect to the current register node.
[0019] As a further optimization of the clock tree establishment method that considers low-voltage clock skew fluctuations, step S2, which uses a spectral clustering algorithm to perform balanced grouping of register nodes based on register node location information, includes the following steps:
[0020] Step S201: Determine the maximum group load capacitance constraint based on the output conversion time constraint and clock buffer size, and construct the clock tree routing topology in combination with the DME algorithm to evaluate the maximum load capacitance.
[0021] Step S202: Determine the number of groups k based on the maximum load capacitance and the maximum group load capacitance constraint obtained in step S201;
[0022] Step S203: For the merged register nodes in step S1, construct an undirected weighted graph containing n register nodes based on the Manhattan distance between the register nodes.
[0023] Step S204: For the undirected weighted graph constructed in step S203, a similarity matrix is constructed using the fully connected method;
[0024] Step S205: For the merged register nodes in step S1, calculate the degree matrix of all register nodes;
[0025] Step S206: Based on the similarity matrix and the degree matrix of all register nodes, calculate the normalized Laplacian matrix of the undirected temporal weight graph after merging nodes in step S1.
[0026] Step 207: Solve for the n-dimensional eigenvectors corresponding to the k smallest eigenvalues of the normalized Laplacian matrix;
[0027] Step 208: Cluster the elements in the n-dimensional feature vectors corresponding to the k smallest eigenvalues and convert them into integers. The n-dimensional feature vectors corresponding to the k smallest eigenvalues after clustering conversion form an n*k feature dimension matrix. Perform k-means algorithm-based partitioning on the n k-dimensional row vectors contained in the n*k feature dimension matrix to obtain n k-dimensional row vectors with only one element being 1 and the rest being 0. The n k-dimensional row vectors with only one element being 1 and the rest being 0 represent the balanced grouping result of the register nodes.
[0028] As a further optimization of the clock tree establishment method that considers low-voltage clock skew fluctuations, the specific method of step S3 is as follows:
[0029] First, the register nodes obtained in step S2 are grouped in a balanced manner. The DME algorithm is used to construct the local clock tree for each register group, the group capacitance of each group is calculated, and the groups are sorted in descending order according to the size of the group capacitance.
[0030] Next, traverse each node in the group with the largest group capacitance, select the register node that has a connection edge with the node in other groups as the boundary node, move the boundary node to the group to which the other register node connected by its connection edge belongs, and repeat the above traversal process until the number of iterations is reached or the maximum group capacitance no longer decreases.
[0031] Then, the DME algorithm is used to construct local clock trees within each group obtained in the previous step.
[0032] As a further optimization of the clock tree establishment method that considers low-voltage clock skew fluctuations, the specific method of step S4 is as follows:
[0033] First, calculate the center of the cluster composed of all register nodes, take the cluster center as the cluster merging node, and use the cluster center as the origin of the coordinate system to represent each sink node in the cluster in polar coordinates.
[0034] Next, traverse all register nodes, using each register node as an initial partitioning point, and evenly divide the register nodes into two groups, selecting the optimal partitioning strategy from the partitioning strategies that use each register node as an initial partitioning point.
[0035] Then, for each sub-cluster obtained after the optimal partitioning strategy, the center point of the sub-cluster is calculated, the center of the sub-cluster is used as the sub-cluster merging node, and the connection relationship between the center point of the sub-cluster and the center point of its sub-cluster is determined. This process is repeated until the connection relationship of the entire clock tree is determined, and the initial topology of the top-level clock tree is obtained.
[0036] As a further optimization of the clock tree establishment method that considers low-voltage clock deviation fluctuations, step S5, which generates a clock buffer insertion strategy based on a genetic algorithm, is as follows: depending on whether a clock buffer is inserted at each level of the clock tree, the individuals representing the clock buffer insertion strategy are genetically encoded. After initializing the population, the fitness function value is calculated based on the clock buffer insertion strategy represented by each individual. The population is iterated, and the individual with the best fitness function value is selected to propagate to the next generation.
[0037] As a further optimization of the clock tree establishment method that considers low-voltage clock skew fluctuations, step S6 specifically includes the following steps:
[0038] Step S601: Based on the Manhattan distance, merge the nodes of each level of the initial topology of the top-level clock tree from bottom to top, calculate the current level load capacitance after merging the nodes, and pass the current level load capacitance after merging the nodes to the next higher level node.
[0039] Step S602: During the upward merging process, if the current stage after merging does not have a clock buffer inserted, then continue with step S601; if the current stage after merging has a clock buffer inserted, then save the maximum load capacitance passed to the node of that stage.
[0040] Step S603: Calculate the difference between the capacitance of other nodes in the current level and the maximum load capacitance of the node in the current level of the top-level clock tree initial topology;
[0041] Step S604: Based on the difference calculated in step S603, use a serpentine routing method to balance the load capacitor and move the adjusted nodes closer to the center of all nodes in the current stage.
[0042] As a further optimization of the clock tree establishment method that considers low-voltage clock skew fluctuations, the specific method of step S7 is as follows:
[0043] First, for the top-level clock tree topology updated in step S6, save the load capacitance driven by each level of clock buffer.
[0044] Then, clock buffers with different drive strengths in the standard cell library are encoded, and the buffer size corresponding to the optimal value of the fitness function is selected based on a genetic algorithm.
[0045] As a further optimization of the clock tree construction method that considers low-voltage clock skew fluctuations, the genetic algorithm evaluates the fitness based on the clock skew fluctuations of the clock buffer insertion strategy. The fitness function is: in, and Let s represent the variance of the clock tree clock skew and the variance of the path delay, respectively. j Let N be the number of sink nodes, ρ be the correlation coefficient between the buffers, and σ be the insertion type of the j-th level buffer. j σ j+1 The delay ripples are for the j-th and j+1-th level buffers.
[0046] As a further optimization of the clock tree establishment method that considers low-voltage clock skew fluctuations, the delay fluctuation of the j-th stage buffer is related to the input transition time and the output load capacitance, and the load capacitance Cb of the j-th stage buffer is... j for Among them, C wire C is the total capacitance of the interconnects. buf L is the total input capacitance of the sink node or clock buffer. j Let C be the clock tree level of the j-th level buffer. i C is the line load capacitance of the i-th stage of the clock tree; in This is the input capacitor for the sink node or clock buffer.
[0047] The present invention adopts the above technical solution, and compared with the prior art, it has the following beneficial effects:
[0048] (1) The clock tree establishment method proposed in this invention addresses the high sensitivity of clock network process fluctuations under low voltage by constructing a balanced and symmetrical clock tree structure. First, a local clock tree is constructed, and then a symmetrical top-level clock tree is constructed based on the distribution of the root nodes of the local clock tree. This increases the common path part from the clock source to each register in the local clock tree and ensures that the path structure and buffer distribution from the clock source to each local clock tree are consistent, effectively reducing the clock deviation fluctuation of the clock tree.
[0049] (2) The clock tree establishment method proposed in this invention constructs a local clock tree based on register grouping with load capacitance balancing. The register grouping considers both timing and register location distribution. The coarsened register nodes are divided in a balanced manner using a spectral clustering algorithm. This reduces the wiring scale of the local clock tree while considering the timing margin between registers. In addition, the local clock adopts a non-buffered approach, which effectively reduces the delay fluctuations caused by buffer insertion.
[0050] (3) The clock tree establishment method proposed in this invention optimizes and adjusts the topology of the buffer insertion position and optimizes the buffer size, reduces the load capacitance of the driving buffer, reduces buffer delay and delay fluctuation, and thus effectively reduces clock tree clock deviation fluctuation. Attached Figure Description
[0051] Figure 1 This is a flowchart illustrating a clock tree synthesis algorithm optimized for low-voltage clock skew fluctuations, as provided in Example 1.
[0052] Figures 2 to 5 This is a schematic diagram illustrating balanced grouping based on the spectral clustering algorithm provided in Example 1. Figure 2 This is the register timing diagram after the benchmark circuit layout. Figure 3 This is a diagram illustrating the merging of register nodes. Figure 4 This is a schematic diagram of register grouping based on the spectral clustering algorithm. Figure 5 This is a schematic diagram of register grouping equalization optimization.
[0053] Figure 6 This is a schematic diagram of the process for constructing the top-level clock tree provided in Example 1.
[0054] Figure 7 This is a flowchart illustrating the genetic algorithm for selecting a buffer insertion scheme provided in Example 1.
[0055] Figure 8 This is a schematic diagram of the clock tree node merging process provided in Example 1.
[0056] Figure 9 This is a schematic diagram of the clock tree topology update process provided in Example 1. Detailed Implementation
[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0058] Example 1
[0059] See Figures 1 to 5 This embodiment provides a clock tree synthesis algorithm implementation optimized for low-voltage clock skew fluctuations. The algorithm implementation process is as follows: Figure 1 As shown, this specifically includes: local clock tree construction and top-level clock tree construction.
[0060] To reduce clock tree skew fluctuations and improve clock tree stability, two approaches are taken: firstly, increasing the common paths of the clock network to reduce clock tree insertion delay by evenly distributing registers into different groups; secondly, optimizing buffer sizes to reduce clock tree skew fluctuations by selecting appropriate buffer insertion positions and sizes.
[0061] Specifically, in this embodiment, the process of balancing register grouping includes register node merging, register grouping based on spectral clustering algorithm, and register balancing grouping optimization, as follows: Figures 2 to 5 As shown, the specific process for constructing a local clock tree includes steps S1 to S6.
[0062] Step S1: Extract timing and location information after the benchmark circuit layout, merge register nodes based on register timing relationships, and initially divide registers into different groups;
[0063] More specifically, in this embodiment, step S1, merging register nodes based on register timing relationships, specifically includes:
[0064] Step S101: Obtain the timing information of each data path after layout using EDA tool commands;
[0065] Step S102: Construct an undirected time-series weighted graph based on the extracted time-series information, with the time margin between register pairs as the weight;
[0066] Step S103: Based on the weight threshold, cut the edges with smaller weights in the undirected temporal weight graph constructed in step S102, and update the cut temporal relationship graph, as follows: Figure 2 As shown;
[0067] Step S104: Traverse all register nodes. Based on the undirected timing weight graph, select the register node with the smallest timing margin between itself and the current register node for merging, and represent it with a new node. The position of the merged node is represented by the center point between the two nodes. After merging nodes based on timing information, a coarsened timing graph is obtained. The thicker nodes after merging indicate that the weight corresponding to that node has increased. Figure 3 As shown.
[0068] Step S2: For the merged nodes in step S1, the registers are balanced and grouped based on the location information, and the capacitance is evaluated in combination with the DME algorithm. The balanced grouping process is implemented in combination with the spectral clustering algorithm.
[0069] More specifically, in this embodiment, step S2, which combines balanced grouping with the spectral clustering algorithm, specifically includes:
[0070] Step S201: Determine the maximum grouped load capacitance constraint C based on the output conversion time constraint and the clock buffer size. limit The clock tree routing topology is constructed using the DME algorithm, and the maximum load capacitance C is evaluated. max ;
[0071] Step S202, based on the maximum load capacitance C in step S201 max and maximum grouped load capacitance constraint C limit Determine the number of groups k = C max / C limit .
[0072] Step S203: For the nodes merged in step S1, construct an undirected weighted graph based on the Manhattan distance between the nodes. During the construction of the undirected weighted graph, set a position distance threshold and use the position distance threshold to prune the connection edges of all register nodes that are far away from other register nodes, generating a sparse weight matrix that retains only the close connection edges of all register nodes.
[0073] Step S204: For the undirected weighted graph constructed in step S203, a similarity matrix W is constructed using the fully connected method. The fully connected method preserves the connection relationships between all nodes in the weighted graph, and the elements w in the matrix... ij The calculation is as follows:
[0074]
[0075] Where i and j represent any two connected nodes; x i y i Represents the x and y coordinates of node i; g i Let be the weight of node i; σ be the function adjustment factor;
[0076] Step S205: Calculate and define the degree matrix of all nodes. The degree matrix is a diagonal matrix with diagonal elements as follows: Where n represents the number of nodes in the undirected weighted graph constructed in step S203, and all off-diagonal elements are 0;
[0077] Step S206: Calculate the normalized Laplacian matrix of the undirected temporal weight graph after merging nodes in step S1 based on the similarity matrix W and the degree matrix.
[0078] Step S207: Solve for the n-dimensional eigenvectors corresponding to the k smallest eigenvalues of the normalized Laplacian matrix. This set of eigenvectors corresponds to the optimal balanced grouping method.
[0079] Step S208: For the i-th n-dimensional feature vector h among the k n-dimensional feature column vectors in step S207 i ={h i0 ,h i1 ...h ij}, the elements h in the vector ij Clustering is converted to either 0 or 1. Here, i represents the i-th group, and h... i An indicator vector, such as h, indicates whether each node has been assigned to the i-th group. i medium element h i1 A value of 0 indicates that node 1 is not assigned to group i, h i1 A value of 1 indicates that it is assigned to group i. In step S207, the elements in the n-dimensional feature vectors obtained by matrix solving are generally decimals close to 0 or 1, requiring clustering to convert them into integers 0 and 1. The main process is as follows: First, a feature matrix is formed by k feature n-dimensional column vectors. The indicator vector indicating whether each node is assigned to k groups corresponds to a row vector in this feature matrix. Nodes with closer Manhattan distances have higher similarity vectors. The n k-dimensional row vectors are partitioned using the k-means algorithm, resulting in n k-dimensional row vectors with only one element being 1 and the rest being 0. This yields the grouping result for each register node, such as... Figure 4 As shown.
[0080] Step S3: Further optimize the grouping obtained in Step S2 by moving the boundary nodes of the largest capacitance group that are close to other groups to other groups. The result after adjusting the register grouping is as follows. Figure 5 As shown. Then, a local clock tree is constructed within each group, and a driving clock buffer is inserted at the root node of the local clock tree;
[0081] More specifically, in this embodiment, the group balance optimization process used in step S3 is as follows:
[0082] First, within the groups obtained in step S2, the DME algorithm is used to construct a local clock tree and calculate the group capacitance of each group. To reduce the maximum group capacitance, the groups are sorted in descending order according to their capacitance values.
[0083] Next, iterate through each node in the group with the largest group capacitance, iterate through each element in the sparse weight matrix generated in step S203, select the register node in the group with the largest current group capacitance that has a connection edge with the register node in other groups as the boundary node, move the boundary node to the group to which the other end of the register node connected by its connection edge belongs, and repeat the above process until a certain number of iterations or the maximum group capacitance no longer decreases.
[0084] Then, construct local clock trees within each group obtained in the previous step.
[0085] More specifically, in this embodiment, in step S3, the DME algorithm is used to construct local clock trees within each group obtained in the previous step. The specific process is as follows:
[0086] First, the node connection relationship is determined based on the minimum Manhattan segment distance; second, the Manhattan radius of node merging is determined by the constraint that the clock deviation is zero, thereby determining the merging range of each node from bottom to top; finally, the center point of the merging range of the root node is determined as the root node, and the position of each node is determined from top to bottom.
[0087] After spectral clustering, Figure 4 The registers shown are divided into three different groups, among which Figure 4 The group below is much larger than the other two groups; the boundary register nodes within the largest group that are closest to other groups are reassigned to the nearest group, resulting in a more balanced grouping outcome, such as... Figure 5 As shown.
[0088] Specifically, in this embodiment, the synthesis process of the top-level clock tree constructed based on the root node of the local clock tree is as follows: Figure 6 As shown, the specific process includes:
[0089] Step S4: Based on the distribution of the driving buffer positions generated in Step S3, the initial topology of the top-level clock tree is constructed using a top-down symmetrical bisection approach.
[0090] More specifically, in this embodiment, the initial clock tree topology construction process in step S4 includes:
[0091] First, calculate the center of the cluster composed of all register nodes, and use this cluster center as the cluster merging node. With the cluster center as the origin, represent each sink node in the cluster in polar coordinates. Traverse all register nodes, using each register node as the initial partitioning point, and evenly divide the register nodes into two groups. Select the optimal partitioning strategy from the partitioning strategies that use each register node as the initial partitioning point. For each sub-cluster obtained after partitioning with the optimal partitioning strategy, calculate the center point of the sub-cluster, use the sub-cluster center as the sub-cluster merging node, and determine the connection relationship between the sub-cluster center point and the center point of its respective cluster. Continue in this manner to obtain the connection relationship of the entire clock tree and determine the initial topology.
[0092] Step S5: Insert a clock buffer into the initial topology constructed in step S4. The insertion of the clock buffer is implemented based on a genetic algorithm, and a clock deviation fluctuation evaluation model of the clock tree is constructed as the fitness function to select the buffer insertion strategy with the smallest clock deviation fluctuation.
[0093] More specifically, in this embodiment, the clock buffer insertion strategy in step S5 is as follows:
[0094] A method is used to insert buffers of the same size simultaneously into nodes at the same level. Each level of the clock tree can choose whether or not to insert a buffer. If a buffer is inserted at level i, the corresponding position i in the gene chain is encoded as 1; otherwise, it is encoded as 0. Gene encoding is completed in this manner, and the population is initialized. The fitness of each insertion scheme is evaluated based on its clock deviation fluctuation. The population is iterated, and the insertion scheme with the smaller clock deviation fluctuation is selected and propagated to the next generation.
[0095] For each buffer insertion scheme, the clock skew fluctuation assessment model is as follows:
[0096]
[0097]
[0098] in, and These represent the variances of clock tree clock skew and path delay, respectively; s j The insertion type of the j-th level buffer; N is the number of sink nodes; ρ is the correlation coefficient between buffers; σ j σ j+1 The delay ripple of the j-th and j+1-th stage buffers is related to the input slew and the output load capacitance. Since the input slew is also related to the output load capacitance, calculating the load capacitance of the buffer is crucial. The load capacitance Cb of the j-th stage buffer is... j The calculation method is as follows:
[0099] Cb j =Cwire +C buf
[0100]
[0101]
[0102] Among them, C wire C is the total capacitance of the interconnects. buf L is the total input capacitance of the sink node or buffer. j Let C be the clock tree level of the j-th level buffer. i C is the line load capacitance of the i-th stage of the clock tree; in The input capacitor for the sink node or buffer.
[0103] Step S6: After determining the buffer insertion position, balance the load capacitance of the inserted buffer using serpentine routing, and update the clock tree topology, such as... Figure 8 , Figure 9 As shown.
[0104] More specifically, in this embodiment, the specific process of updating the clock tree topology in step S6 is as follows:
[0105] Step S601: Merge nodes at each level from bottom to top based on Manhattan distance, calculate the load capacitance of the merged nodes and pass it to the next higher level node;
[0106] Step S602: During the upward merging process, if no buffer is inserted in the merged level, continue with step S601; if the merged level is marked as having a buffer inserted, save the maximum load capacitance passed to the node at that level.
[0107] Step S603: Calculate the difference between the capacitance of other nodes in this level of the clock tree and the maximum load capacitance;
[0108] Step S604: For the difference in step S603, use a serpentine routing method to balance the load capacitance, and adjust the position of the nodes after routing, bringing them closer to the center of all nodes in this stage, such as... Figure 9 As shown.
[0109] Figure 8 This represents the initial merging range of the left and right topology nodes. The winding length is the Manhattan radius. The Manhattan rectangles of the left and right nodes are redefined. Then, a Manhattan rectangle is drawn with the center point between the left and right merging ranges. The part that overlaps with the Manhattan rectangles of the left and right nodes is determined as the merging range of the nodes after winding.
[0110] Step S7: Based on the updated and optimized clock tree, call the clock buffers with different drive strengths in the buffer cell library, and combine the genetic algorithm to optimize the buffer size.
[0111] More specifically, in this embodiment, the buffer size optimization process in step S7 is as follows:
[0112] For the clock tree topology updated in step S6, save the drive load capacitance of each level of buffer; then, encode the clock buffers with different drive strengths in the standard cell library as 1, 2, ..., and so on. As described in step S5, use a genetic algorithm to select the clock buffer size with the smallest clock deviation fluctuation.
[0113] Any aspects of this invention not described in detail are well-known to those skilled in the art.
[0114] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.
Claims
1. A clock tree establishment method considering low-voltage clock skew fluctuation optimization, characterized in that, Includes the following steps: Step S1: Extract the timing and location information after the reference circuit layout, and merge the register nodes according to the register timing relationship; Step S2: For the merged register nodes, the register nodes are evenly grouped according to their location information using a spectral clustering algorithm. Step S3: Optimize the register equalization grouping obtained in step S2 and construct the local clock tree for each register equalization group; Step S4: Based on the distribution of the drive buffer positions of the local clock tree of each register equalization group constructed in step S3, construct the initial top-level clock tree topology using a top-down symmetrical binary search method. Step S5: Generate clock buffer insertion strategy based on genetic algorithm, and insert clock buffers into the initial top-level clock tree topology structure constructed in step S4 according to the clock buffer insertion strategy corresponding to the optimal value of fitness function. Step S6: Optimize the initial topology of the top-level clock tree after inserting the clock buffer, and obtain the updated top-level clock tree topology. Step S7: Select the clock buffer size based on the genetic algorithm, and optimize the size of the clock buffer in the updated top-level clock tree topology to the buffer size corresponding to the best value of the fitness function.
2. The clock tree establishment method considering low-voltage clock skew fluctuation optimization according to claim 1, characterized in that, The specific method for merging register nodes according to register timing relationships in step S1 includes the following steps: Step S101: Obtain the timing information of each data path after layout using EDA tool commands; Step S102: Construct an undirected time-series weighted graph based on the extracted time-series information, with the time margin between register pairs as the weight; Step S103: Cut the edges in the undirected temporal weight graph constructed in step S102 according to the weight threshold, and update the undirected temporal weight graph. Step S104: Traverse all register nodes and, based on the updated undirected timing weight graph, merge the current register node and the register node with the smallest timing margin with respect to the current register node.
3. The clock tree establishment method considering low-voltage clock skew fluctuation optimization according to claim 2, characterized in that, The specific method for performing balanced grouping of register nodes based on register node location information using the spectral clustering algorithm in step S2 includes the following steps: Step S201: Determine the maximum group load capacitance constraint based on the output conversion time constraint and clock buffer size, and construct the clock tree routing topology in combination with the DME algorithm to evaluate the maximum load capacitance. Step S202: Determine the number of groups k based on the maximum load capacitance and the maximum group load capacitance constraint obtained in step S201; Step S203: For the merged register nodes in step S1, construct an undirected weighted graph containing n register nodes based on the Manhattan distance between the register nodes. Step S204: For the undirected weighted graph constructed in step S203, a similarity matrix is constructed using the fully connected method; Step S205: For the merged register nodes in step S1, calculate the degree matrix of all register nodes; Step S206: Based on the similarity matrix and the degree matrix of all register nodes, calculate the normalized Laplacian matrix of the undirected temporal weight graph after merging nodes in step S1. Step 207: Solve for the n-dimensional eigenvectors corresponding to the k smallest eigenvalues of the normalized Laplacian matrix; Step 208: Cluster the elements in the n-dimensional feature vectors corresponding to the k smallest eigenvalues and convert them into integers. The n-dimensional feature vectors corresponding to the k smallest eigenvalues after clustering conversion form an n*k feature dimension matrix. Perform k-means algorithm-based partitioning on the n k-dimensional row vectors contained in the n*k feature dimension matrix to obtain n k-dimensional row vectors with only one element being 1 and the rest being 0. The n k-dimensional row vectors with only one element being 1 and the rest being 0 represent the balanced grouping result of the register nodes.
4. The clock tree establishment method considering low-voltage clock skew fluctuation optimization according to claim 3, characterized in that, The specific method for step S3 is as follows: First, the register nodes obtained in step S2 are grouped in a balanced manner. The DME algorithm is used to construct the local clock tree for each register group, the group capacitance of each group is calculated, and the groups are sorted in descending order according to the size of the group capacitance. Next, traverse each node in the group with the largest group capacitance, select the register node that has a connection edge with the node in other groups as the boundary node, move the boundary node to the group to which the other register node connected by its connection edge belongs, and repeat the above traversal process until the number of iterations is reached or the maximum group capacitance no longer decreases. Then, the DME algorithm is used to construct local clock trees within each group obtained in the previous step.
5. The clock tree establishment method considering low-voltage clock skew fluctuation optimization according to claim 4, characterized in that, The specific method for step S4 is as follows: First, calculate the center of the cluster composed of all register nodes, take the cluster center as the cluster merging node, and use the cluster center as the origin of the coordinate system to represent each sink node in the cluster in polar coordinates. Next, traverse all register nodes, using each register node as an initial partitioning point, and evenly divide the register nodes into two groups, selecting the optimal partitioning strategy from the partitioning strategies that use each register node as an initial partitioning point. Then, for each sub-cluster obtained after the optimal partitioning strategy, the center point of the sub-cluster is calculated, the center of the sub-cluster is used as the sub-cluster merging node, and the connection relationship between the center point of the sub-cluster and the center point of its sub-cluster is determined. This process is repeated until the connection relationship of the entire clock tree is determined, and the initial topology of the top-level clock tree is obtained.
6. The clock tree establishment method considering low-voltage clock skew fluctuation optimization according to claim 5, characterized in that, The specific method for generating the clock buffer insertion strategy based on the genetic algorithm in step S5 is as follows: according to whether the clock buffer is inserted at each level of the clock tree, the individuals representing the clock buffer insertion strategy are genetically encoded. After initializing the population, the fitness function value is calculated based on the clock buffer insertion strategy represented by each individual. The population is iterated, and the individual with the best fitness function value is selected to be propagated to the next generation.
7. The clock tree establishment method considering low-voltage clock skew fluctuation optimization according to claim 6, characterized in that, The specific method of step S6 includes the following steps: Step S601: Based on the Manhattan distance, merge the nodes of each level of the initial topology of the top-level clock tree from bottom to top, calculate the current level load capacitance after merging the nodes, and pass the current level load capacitance after merging the nodes to the next higher level node. Step S602: During the upward merging process, if the current stage after merging does not have a clock buffer inserted, then continue with step S601; if the current stage after merging has a clock buffer inserted, then save the maximum load capacitance passed to the node of that stage. Step S603: Calculate the difference between the capacitance of other nodes in the current level and the maximum load capacitance of the node in the current level of the top-level clock tree initial topology; Step S604: Based on the difference calculated in step S603, use a serpentine routing method to balance the load capacitor and move the adjusted nodes closer to the center of all nodes in the current stage.
8. The clock tree establishment method considering low-voltage clock skew fluctuation optimization according to claim 7, characterized in that, The specific method for step S7 is as follows: First, for the top-level clock tree topology updated in step S6, save the load capacitance driven by each level of clock buffer. Then, clock buffers with different drive strengths in the standard cell library are encoded, and the buffer size corresponding to the optimal value of the fitness function is selected based on a genetic algorithm.
9. The clock tree establishment method considering low-voltage clock skew fluctuation optimization according to claim 8, characterized in that, The genetic algorithm evaluates fitness based on the clock skew fluctuation of the clock buffer insertion strategy, and the fitness function is: in, and Let s represent the variance of the clock tree clock skew and the variance of the path delay, respectively. j Let N be the number of sink nodes, ρ be the correlation coefficient between the buffers, and σ be the insertion type of the j-th level buffer. j σ j+1 The delay ripples are for the j-th and j+1-th level buffers.
10. The clock tree establishment method considering low-voltage clock skew fluctuation optimization according to claim 9, characterized in that, The delay ripple of the j-th stage buffer is related to the input switching time and the output load capacitance. The load capacitance Cb of the j-th stage buffer is... j For Cb j =C wire +C buf , Among them, C wire C is the total capacitance of the interconnects. buf L is the total input capacitance of the sink node or clock buffer. j Let C be the clock tree level of the j-th level buffer. i C is the line load capacitance of the i-th stage of the clock tree; in This is the input capacitor for the sink node or clock buffer.
Citation Information
Patent Citations
Method and device for constructing clock network
CN103259526A
Clock-domain-crossing clock tree building method, clock-domain-crossing clock tree building system, integrated circuit and manufacturing method
CN104077427A
