A method and system for PPA prediction and source code annotation in integrated circuit RTL design

By constructing multi-graph representations and cross-graph correspondences, the problem of comprehensively characterizing timing, power consumption, and area results in the RTL design stage is solved, enabling direct prediction and source code-level localization in the RTL design stage, thereby improving the efficiency and accuracy of integrated circuit design.

CN122655657BActive Publication Date: 2026-11-10LANZHOU UNIV +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611152118.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-07-31
Publication Date
2026-11-10
Estimated Expiration
2046-07-31

AI Technical Summary

Technical Problem

In the existing digital integrated circuit design flow, the RTL design stage is difficult to fully reflect the syntax level, control structure, and bit-level logic relationship of the code. This results in an incomplete characterization of the key logic structure of timing, power consumption, and area results, and makes it difficult to locate problems, thus increasing the design iteration cost.

Method used

A multi-graph representation method based on historical RTL samples, including AST, SOG, AIG, AIMG, and XAG, is adopted to construct a unified intermediate representation, extract structural features, and establish a PPA prediction model through cross-graph correspondence. Timing, power consumption, and area prediction are performed directly in the RTL design stage, and the results are mapped to the source code location.

Benefits of technology

Obtaining timing, power, and area predictions directly during the RTL design phase shortens the design evaluation cycle, improves design efficiency and problem localization capabilities, reduces iteration costs, and enhances the consistency and interpretability of prediction results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122655657B_ABST
    Figure CN122655657B_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for PPA prediction and source code annotation in integrated circuit RTL design. First, a reference sample set is established to train a PPA prediction model. Then, the RTL code to be analyzed is preprocessed, a multi-graph representation is constructed, structural features are extracted from the multi-graph representation, and cross-graph correspondences are established. Based on the cross-graph correspondences, the extracted structural features are subjected to unified dimensional mapping and feature fusion to form a unified design representation. Next, the unified design representation of the RTL design to be analyzed is predicted based on the PPA prediction model to obtain timing, power consumption, and area prediction results. Finally, the prediction results are mapped and annotated at the source code level. This invention can map timing risks, power consumption risks, and area risks to corresponding locations in the RTL source code, improving the localizability and interpretability of the analysis results; shortening the design evaluation cycle; and improving the problem awareness and development efficiency in the chip front-end design stage.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of integrated circuit design technology, specifically relating to a method and system for PPA prediction and source code annotation in integrated circuit RTL design. Background Technology

[0002] In the integrated circuit design flow, the register-transfer level (RTL) is used to describe the functional behavior of the circuit, the data path structure, the control logic relationships, and the module hierarchy connections. RTL design affects system functional design and subsequent logic synthesis, timing analysis, power consumption analysis, and physical implementation. RTL code development is typically done using hardware description languages ​​such as Verilog, VHDL, or SystemVerilog. Because gate-level mapping and physical implementation are not yet complete in the RTL design phase, the design structure still has high adjustability. Therefore, this stage is also a crucial period for timing optimization, low-power optimization, and area control.

[0003] In existing digital integrated circuit design flows, the power consumption, performance, and area of ​​the design typically rely on subsequent EDA processes such as logic synthesis, static timing analysis, power consumption analysis, and area statistics. For example, after completing the RTL (Real-Time Limiting) code, synthesis and even multiple rounds of analysis are required to obtain metrics such as worst-case negative margin (WNS), total negative margin (TNS), overall design power consumption, module-level power distribution, overall design area, and module area contribution. When problems exist in the design, such as excessively deep combinational logic levels, cascaded multiplexed structures formed by multi-level conditional control, excessively heavy high-fan-out control signal loads, excessively large bit-width operational logic scales, or local high-flip sensitive areas, these issues are often only discovered in the synthesis report, timing report, or power consumption report, requiring a return to the RTL source code for location and modification. This feedback cycle is lengthy, especially in cases of large-scale designs, deep module hierarchies, and strong coupling between control and data paths, significantly increasing front-end development and iteration costs.

[0004] To shorten the design evaluation cycle, existing methods attempt to analyze or predict design quality during the RTL design phase. Typically, the RTL code is first converted into an intermediate representation, and then features are extracted and an evaluation model is built based on this representation. However, the following shortcomings remain: First, most methods use a single representation to model the RTL, making it difficult to simultaneously reflect the code's syntax hierarchy, control structure, bit-level logic relationships, and Boolean operation characteristics. This results in an incomplete characterization of key logic structures affecting timing, power consumption, and area results. Second, most methods focus on design-level or module-level output results, making it difficult to further pinpoint specific registers, signals, always blocks, assign statements, or conditional branches. It is also difficult to directly locate problem areas and modify the source code based on the analysis results. Summary of the Invention

[0005] To address the problems existing in the background technology, the purpose of this invention is to provide a PPA prediction and source code annotation method for integrated circuit RTL design, which provides feedback basis for chip front-end optimization and modification, reduces repeated iterations in subsequent synthesis and analysis, and improves the overall efficiency and convergence capability of integrated circuit front-end design.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for PPA prediction and source code annotation in integrated circuit RTL design includes the following steps: S1. Establish a reference sample set based on historical RTL samples and obtain reference PPA tags. Each historical RTL sample in the reference sample set includes the original RTL code, multiple structural representations generated from the original RTL code, and reference PPA tags corresponding to the design of the historical RTL sample. S2. Perform lexical analysis, syntax analysis, and hierarchical expansion on historical RTL samples in the reference sample set to extract module definitions, port lists, register declarations, net declarations, combinational logic descriptions, sequential logic descriptions, conditional statements, assignment statements, and module instantiation relationships, forming a unified intermediate representation. This intermediate representation is used for subsequent multi-graph representation construction, while retaining statement position information, signal name information, bit width information, and hierarchical relationship information from the source code to support subsequent prediction result mapping and source code annotation; it also records module name, signal name, register name, bit width information, bit-level index information, instance level information, and source code statement position information. S3. Based on the unified intermediate representation formed in step S2, construct multiple graph representations for the same RTL design, including Abstract Syntax Tree (AST), Simple Operator Graph (SOG), AND-Inverted Graph (AIG), AND-Inverted Multiplexed Graph (AIMG), and XOR-AND Graph (XAG). By constructing the above multiple graph representations, the same integrated circuit RTL design can be jointly represented from multiple perspectives such as syntax hierarchy, control structure, data path structure, bit-level connection relationship, and Boolean logic structure. S4. Based on the multi-graph representation constructed in step S3, extract the structural features reflecting the integrated circuit design, including design-level features, module-level features, node-level features, path-level features, and input cone-level features. Establish cross-graph correspondences between different graph representations based on the module name, signal name, register name, bit width information, bit-level index information, instance-level information, and source code statement location information recorded in step S2. This enables nodes or local structures in different graphs to be associated with the same RTL object, providing a foundation for subsequent feature fusion, prediction result mapping, and source code annotation. S5. Based on the cross-graph correspondence established in step S4, perform unified dimensional mapping on the structural features extracted in step S4, and form a unified design representation of historical RTL samples by splicing, weighting or fusion of the syntax hierarchical information in AST, bit-level connection information in SOG, Boolean logic structure in AIG, multiplexing structure in AIMG and XOR logic structure in XAG; thereby obtaining a more complete RTL design representation, so that the structural information scattered in different graph representations is unified and integrated, improving the completeness and consistency of subsequent timing, power consumption and area prediction, and providing a unified input for further source code-level result mapping; S6. Based on the reference PPA labels obtained in step S1, and using the unified design representation of the historical RTL samples formed in step S5 as training features, supervised training is performed to obtain the PPA prediction model; replace the historical RTL samples with the RTL design code to be analyzed and repeat steps S2-S5, using the unified design representation of the RTL design code to be analyzed as the prediction feature, inputting it into the PPA prediction model, and outputting the timing prediction result, power prediction result and area prediction result. S7. Based on the cross-graph correspondence of the RTL design code to be analyzed, map the timing prediction results, power prediction results and area prediction results obtained in step S6 to the corresponding positions in the RTL source code, and then add annotation information to the corresponding positions and output the results.

[0007] Furthermore, in step S1, the reference PPA tag is obtained by performing logic synthesis, static timing analysis, power consumption analysis, and area statistics on historical RTL sample designs. Specifically, by performing logic synthesis, the RTL code is mapped to a gate-level netlist; by performing static timing analysis, the worst-case negative margin (WNS), total negative margin (TNS), critical path information, and inter-register path timing information are obtained; by performing power consumption analysis, the overall power consumption of the design, module-level power consumption distribution, and information on local high-flip-sensitive areas are obtained; and by performing area statistics, the overall design area, module area contribution, and logic resource usage information are obtained based on the synthesis report or implementation report.

[0008] Furthermore, in step S3, the AST is used to characterize the syntax hierarchy, procedural block structure, conditional judgment relationship, and expression nesting relationship in the RTL code; SOG is used to characterize the connection relationship between bit-level simple logic operators, making the RTL representation closer to the synthesized bit-level logic network; AIG is used to characterize the Boolean logic structure based on AND gates and inverters; AIMG is used to describe multiplexing logic and its selection control relationship; XAG is used to characterize XOR logic and its combination relationship with basic logic operators.

[0009] Furthermore, in step S4, the design-level features include the total number of nodes, the total number of edges, the graph depth, and the number and proportion of different node types; the module-level features include the number of registers within the module, the number of logical nodes, the module hierarchy depth, and the number of control branches; the node-level features include node type, fan-in, fan-out, bit width, and the logic level; the path-level features include the candidate path length between registers, the combinational logic level, the statistics of different operation nodes on the critical path, the cascading depth of the multiplexing structure, and the size of the logic cone; the input cone-level features include the number of driving registers in the endpoint register input cone, the logic depth of the input cone, and the number of input cone nodes.

[0010] Furthermore, in step S4, the structural features also include toggle rate related features, module-level average toggle rate features, and high fan-out control signal features for power consumption analysis, and logic scale, resource complexity, large bit-width operation logic scale, and area contribution related features for area analysis.

[0011] Furthermore, in step S6, the timing prediction results include worst-case negative margin (WNS), total negative margin (TNS), and critical path or register-level risk information; the power consumption prediction results include overall design power consumption, module-level power consumption distribution, and information on highly sensitive flip-flop regions; and the area prediction results include overall design area and module area contribution information. Through this step, prediction information corresponding to subsequent logic synthesis, static timing analysis, power consumption analysis, and area statistics results can be obtained at the RTL stage, thereby identifying high-risk logic structures that may affect the implementation quality of integrated circuits in advance.

[0012] Furthermore, in step S7, the corresponding locations in the RTL source code include, but are not limited to, module definition locations, register and net declaration locations, always statement block locations, assign statement locations, conditional branch statement locations, and module instantiation statement locations, thereby converting design-level, module-level, and fine-grained prediction results into source code-level location information.

[0013] Furthermore, in step S7, the annotation information includes one or more of the following: timing risk, power consumption risk, area risk, risk prediction value, risk level, and optimization prompt information; for code regions with obvious risks, further optimization prompt information can be added.

[0014] Furthermore, in step S7, the output is an RTL source code file with annotation information or an analysis report containing the correspondence between source code location, risk type, and prediction results. This enables designers to directly locate modules, registers, signals, or statements with risks and modify and optimize the front-end RTL accordingly.

[0015] A system for implementing the above-mentioned PPA prediction and source code annotation method for integrated circuit RTL design includes: The RTL code acquisition unit is used to receive historical RTL samples of integrated circuits and RTL design code to be analyzed. The reference sample construction and reference EDA analysis unit is used to build a reference sample set from the acquired historical RTL samples according to the reference EDA analysis conditions. The reference EDA analysis conditions include clock cycle constraints, input / output delay constraints, load constraints, drive constraints, process corner conditions, and target standard cell library information. It also performs logic synthesis, static timing analysis, power consumption analysis, and area statistics on the historical RTL samples to obtain reference PPA tags. The RTL preprocessing unit performs lexical analysis, syntax analysis, and hierarchical expansion on the historical RTL samples and the RTL design code to be analyzed received by the RTL code acquisition unit. It extracts module definitions, port lists, register declarations, net declarations, combinational logic descriptions, sequential logic descriptions, conditional statements, assignment statements, and module instantiation relationships to form a unified intermediate representation of the historical RTL samples and the RTL design code to be analyzed. It also records module names, signal names, register names, bit width information, bit-level index information, instance level information, and source code statement location information. The multi-graph representation construction unit is used to receive the unified intermediate representation of historical RTL samples and RTL design code to be analyzed formed by the RTL preprocessing unit, and construct multi-graph representations such as AST, SOG, AIG, AIMG and XAG for the RTL design in the historical RTL samples and RTL design code to be analyzed according to the unified intermediate representation. The feature extraction and cross-graph alignment unit is used to extract structural features from the historical RTL samples constructed by the multi-graph representation construction unit and the multi-graph representation of the RTL design code to be analyzed, including design-level features, module-level features, node-level features, path-level features and input cone-level features. It also establishes cross-graph correspondences between different graph representations based on the module name, signal name, register name, bit width information, bit-level index information, instance-level information and source code statement position information recorded by the RTL preprocessing unit. The feature fusion unit is used to perform unified dimensional mapping on the structural features of historical RTL samples and the RTL design code to be analyzed based on the cross-graph correspondence established by the feature extraction and cross-graph alignment units, and form the corresponding unified design representation by splicing, weighting or fusion. The PPA prediction unit receives the unified design representation formed by the feature fusion unit, uses the unified design representation of historical RTL samples as training features, and performs supervised training based on reference samples and reference PPA labels provided by the reference EDA analysis unit to obtain a PPA prediction model including a timing prediction sub-model, a power prediction sub-model, an area prediction sub-model, and a design-level calibration process. The PPA prediction model is used to predict the RTL design code to be analyzed, using the unified design representation of the RTL design code to be analyzed as the prediction feature, inputting it into the PPA prediction model, and outputting timing prediction results, power prediction results, and area prediction results. The result mapping unit is used to map the prediction results output by the PPA prediction unit to the corresponding positions in the RTL source code based on the cross-graph correspondence established by the feature extraction and cross-graph alignment units. The source code annotation output unit is used to add annotation information to the corresponding positions in the RTL source code mapped by the result mapping unit, and output the RTL code file or analysis report with annotation information.

[0016] Compared with the shortcomings and deficiencies of existing technologies, the present invention has the following beneficial effects: 1. This invention can directly obtain prediction results related to timing, power consumption and area during the RTL design stage of integrated circuits, thereby moving problems that would normally only be exposed after logic synthesis, static timing analysis, power consumption analysis or area statistics to the front-end design stage, shortening the design evaluation cycle, and improving the problem perception capability and development efficiency of the chip front-end design stage. 2. By introducing reference PPA information, this invention establishes a correspondence between the prediction results in the RTL stage and the subsequent logic synthesis, static timing analysis, power consumption analysis, and area statistics results, thereby improving the consistency between the PPA evaluation results and the actual integrated circuit implementation results; 3. By constructing multiple structural representations such as AST, SOG, AIG, AIMG and XAG, this invention can jointly characterize the same RTL design from multiple perspectives, which can more comprehensively identify the key logic structures that affect timing, power consumption and area results. This is beneficial to improving the ability to analyze critical path risks, power hotspots and logic regions with high area occupancy in complex digital circuits. 4. By establishing a correspondence between graph structures and RTL source code objects, this invention can map timing risks, power risks, and area risks to corresponding locations in the RTL source code, improving the localizability and interpretability of the analysis results. Designers no longer need to rely solely on synthesis or timing reports and manually review the source code; they can directly locate high-risk logic regions in the RTL code, reducing the difficulty of troubleshooting and improving the efficiency of source code-level location. 5. The output of this invention adds annotation information to the corresponding source code locations and can further provide optimization suggestions, thereby improving the targeting of front-end RTL modifications and directly serving RTL modifications and front-end optimization, rather than just remaining at the result output level. Attached Figure Description

[0017] Figure 1 This is a flowchart of the PPA prediction and source code annotation method for integrated circuit RTL design provided in this embodiment of the invention; Figure 2 This is a diagram of the PPA prediction and source code annotation system for integrated circuit RTL design provided in this embodiment of the invention; Figure 3 A schematic diagram illustrating the construction of multiple graph representations for the same RTL design provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of source code-level result mapping and annotation provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] This invention provides a method for PPA prediction and source code annotation in integrated circuit RTL design, used in the front-end RTL design stage of a chip, such as... Figure 1 As shown, the main process includes: Dataset construction: Establish a reference sample set based on historical RTL samples and obtain reference PPA labels; RTL code is preprocessed to form a unified intermediate representation; Multi-graph representation construction: Constructing multi-graph representations such as AST, SOG, AIG, AIMG, and XAG for RTL code; Feature extraction and cross-graph alignment: Extract structural features from multi-graph representations, establish cross-graph correspondences between different graph representations, and complete cross-graph alignment; Multi-image feature fusion maps structural features to a unified dimension and forms a unified design representation through splicing, weighting, or fusion. Model training: Supervised training of the model is performed using a unified design representation of historical RTL samples to establish a PPA prediction model; PPA prediction uses the PPA prediction model to predict the RTL design code to be analyzed and outputs the prediction results. Result mapping maps the prediction results to the corresponding locations in the RTL source code; Source code annotation and output: Add annotation information to the corresponding positions in the RTL source code that maps the prediction results, and output the RTL code file or analysis report with annotation information. This allows designers to receive front-end analysis feedback related to timing, power consumption, and area without relying on a complete back-end implementation process, and provides a basis for subsequent RTL modifications.

[0020] PPA prediction and source code annotation system for integrated circuit RTL design, such as Figure 2 As shown, it includes: The RTL code acquisition unit is used to receive historical RTL samples of integrated circuits and RTL design code to be analyzed. The reference sample construction and reference EDA analysis unit is used to build a reference sample set from the acquired historical RTL samples according to the reference EDA analysis conditions. The reference EDA analysis conditions include clock cycle constraints, input / output delay constraints, load constraints, drive constraints, process corner conditions, and target standard cell library information. It also performs logic synthesis, static timing analysis, power consumption analysis, and area statistics on the historical RTL samples to obtain reference PPA tags. The reference PPA tags include WNS, TNS, path or register-level timing information, overall design power consumption, module-level power consumption, overall design area, and module area contribution.

[0021] The RTL preprocessing unit is used to preprocess the historical RTL samples and the RTL design code to be analyzed received by the RTL code acquisition unit. The preprocessing includes lexical analysis, syntax analysis, and hierarchical expansion; extracting module definitions, port lists, register declarations, net declarations, combinational logic descriptions, sequential logic descriptions, conditional statements, assignment statements, and module instantiation relationships to form a unified intermediate representation of the historical RTL samples and the RTL design code to be analyzed; and recording module names, signal names, register names, bit width information, bit-level index information, instance level information, and source code statement location information. The multi-graph representation building unit is used to receive the unified intermediate representation of historical RTL samples and RTL design code to be analyzed formed by the RTL preprocessing unit, and to construct multi-graph representations such as AST, SOG, AIG, AIMG and XAG for the RTL design in the historical RTL samples and RTL design code to be analyzed according to the unified intermediate representation. The feature extraction and cross-graph alignment unit is used to extract structural features such as design-level features, module-level features, node-level features, path-level features, and input cone-level features from the historical RTL samples constructed by the multi-graph representation construction unit and the RTL design code to be analyzed. It also establishes cross-graph correspondences between different graph representations based on the module name, signal name, register name, bit width information, bit-level index information, instance-level information, and source code statement position information recorded by the RTL preprocessing unit. The feature fusion unit is used to perform unified dimensional mapping on the structural features of the historical RTL samples and the RTL design code to be analyzed, which are output by the feature extraction and cross-graph alignment unit, based on the cross-graph correspondence. The unified design representation of the historical RTL samples and the unified design representation of the RTL design code to be analyzed are formed by splicing, weighting or fusion, respectively. The PPA prediction unit receives unified design representations of historical RTL samples and the unified design representation of the RTL design code to be analyzed, generated by the feature fusion unit. It uses the unified design representations of historical RTL samples as training features and, based on reference samples and reference PPA labels provided by the reference EDA analysis unit, performs supervised training to obtain a PPA prediction model including a timing prediction sub-model, a power prediction sub-model, an area prediction sub-model, and a design-level calibration process. The PPA prediction model is used to predict the RTL design code to be analyzed. The unified design representation of the RTL design code is used as the prediction feature input to the trained PPA prediction model, which outputs timing prediction results, power prediction results, and area prediction results. Specifically, the timing prediction sub-model outputs WNS, TNS, and critical path or register-level timing risks; the power prediction sub-model outputs the overall design power consumption, module-level power consumption distribution, and high flip-sensitive regions; and the area prediction sub-model outputs the overall design area and module area contribution information. Based on the predicted values, preset thresholds, or ranking results of the PPA prediction model, corresponding timing risk levels, power risk levels, and area risk levels are generated for subsequent source code annotation. The result mapping unit is used to map the prediction results output by the PPA prediction unit to the corresponding positions in the RTL source code based on the cross-graph correspondence established by the feature extraction and cross-graph alignment units. The source code annotation output unit is used to add annotation information (inline annotation) to the corresponding position in the RTL source code mapped by the result mapping unit. The annotation information includes timing risk level, power risk level and area risk level, and outputs an RTL code file or analysis report with annotation information.

[0022] The PPA prediction and source code annotation method is implemented based on the above system, which will be explained in detail below.

[0023] 1. Establish a reference sample set A reference sample set is established based on historical RTL samples. Each historical RTL sample in the reference sample set includes the original RTL code, various structural representations generated from the original RTL code, and a reference PPA tag corresponding to the historical RTL sample design. The original RTL code can come from historical chip front-end design projects, open-source digital circuit design samples, or existing integrated circuit RTL design sample libraries, and can be in Verilog, VHDL, or SystemVerilog format.

[0024] For each historical RTL sample, reference EDA analysis conditions are first set according to the target integrated circuit implementation requirements. These conditions include clock cycle constraints, input / output delay constraints, load constraints, drive constraints, process corner conditions, and target standard cell library information. These constraints ensure that subsequent logic synthesis, static timing analysis, power consumption analysis, and area statistics results remain consistent with the target integrated circuit implementation conditions.

[0025] Reference PPA tags are obtained by performing a reference EDA analysis process on the corresponding RTL design. This process includes logic synthesis, static timing analysis, power analysis, and area statistics. Logic synthesis can be performed using DesignCompiler, Genus, or other synthesis tools, mapping the RTL code to a gate-level netlist. Static timing analysis can be performed using PrimeTime, Tempus, or other timing analysis tools to obtain worst-case negative margin (WNS), total negative margin (TNS), critical path information, and inter-register path timing information. Power analysis can be performed using PrimeTime PX, Voltus, or other power analysis tools to obtain overall design power consumption, module-level power distribution, and information on local high-flip-sensitivity regions. The switching activity information required for power analysis can be obtained from SAIF or VCD files generated by RTL simulation or gate-level simulation. In the absence of actual excitation vectors, vector-independent switching activity estimation methods provided by EDA tools can be used to obtain approximate switching activity information. The resulting power consumption results can include dynamic power consumption, static power consumption, module-level power distribution, and information on local high-flip-sensitivity regions. Area statistics are obtained from the comprehensive report or implementation report, including information on the overall design area, module area contribution, and logical resource usage.

[0026] For reference samples requiring fine-grained prediction and labeling, further extraction of critical path start and end registers, path margin distribution, and register-level timing risk information can be obtained from the timing analysis report; module-level power hotspot information and high-flip-sensitive area information can be obtained from the power analysis report; and module-level area contribution information can be obtained from the area statistics report. These reference PPA labels serve as the basis for subsequent PPA prediction model construction, parameter determination, and result calibration.

[0027] For historical RTL samples, the same RTL code preprocessing, multi-graph representation construction, feature extraction and cross-graph alignment, and feature fusion processes as the RTL design code to be analyzed are performed to form a unified design representation. This unified design representation from historical RTL samples serves as training features, which, together with reference PPA labels obtained through reference EDA analysis, constitute supervised training data to obtain the PPA prediction model. The PPA prediction model includes timing prediction sub-models, power prediction sub-models, area prediction sub-models, and a design-level calibration process. After the PPA prediction model is established, the RTL design code to be analyzed undergoes RTL code preprocessing, multi-graph representation construction, feature extraction and cross-graph alignment, and feature fusion processes to form a unified design representation, which serves as the prediction feature. The PPA prediction model is then used for prediction.

[0028] The established reference sample set not only obtains reference PPA labels consistent with the integrated circuit implementation process, but also ensures that the supervision information used by the subsequent prediction model corresponds with the actual logic synthesis, static timing analysis, power consumption analysis and area statistics results, thereby improving the applicability and reliability of the present invention in the integrated circuit RTL front-end design scenario.

[0029] 2. RTL code preprocessing Lexical analysis, syntax analysis, and hierarchical expansion are performed on the RTL design code to be analyzed, extracting module definitions, port lists, register declarations, net declarations, combinational logic descriptions, sequential logic descriptions, conditional statements, assignment statements, and module instantiation relationships to form a unified intermediate representation.

[0030] RTL design code preprocessing can be implemented using Pyverilog, Yosys, or other RTL parsing tools. For example, for Verilog code, Pyverilog can be used to perform lexical analysis, syntax analysis, and the generation of an abstract syntax tree, extracting syntactic information such as module definitions, port information, signal declarations, always blocks, assign statements, and conditional branches. Yosys can be used to perform hierarchical expansion, bit-level logic expansion, and internal intermediate representation transformation of the RTL code, thereby obtaining structural information that more closely approximates the subsequent logic synthesis form, and providing support for the construction of bit-level or Boolean logic structure representations such as SOG, AIG, AIMG, and XAG. For VHDL or SystemVerilog RTL code, similar processing can be performed using their corresponding parsing tools.

[0031] During RTL code preprocessing, the module name, signal name, register name, bit width information, bit-level index information, instance level information, and source code statement location information are recorded simultaneously. The source code statement location information includes the module definition location, register and net declaration location, always block location, assign statement location, conditional branch statement location, and module instantiation statement location. By recording the above source code information, timing prediction results, power prediction results, and area prediction results can be accurately mapped back to specific objects in the RTL design code after PPA prediction is completed.

[0032] RTL code preprocessing yields a unified intermediate representation that includes module hierarchy relationships, inter-module connection relationships, register transfer relationships, signal driving relationships, control logic relationships, and data path relationships, providing a foundation for subsequent multi-graph representation construction, feature extraction, PPA prediction, and source code-level annotation.

[0033] 3. Construction of multi-graph representation Based on the unified intermediate representation obtained above, multiple graph representations of the same RTL design are constructed as AST, SOG, AIG, AIMG, and XAG, such as... Figure 3 As shown, the symbols in the figure are used to represent the node types and connection relationships of RTL design in different graphical representations. Among them, R1 and R2 represent different registers or signal identifiers in RTL design, R1[0], R2[0], etc. represent single-bit nodes after bit-level expansion of the corresponding registers or signals, and the numbers in square brackets represent the bit index of the single-bit node in the corresponding register or signal; AND, OR, XOR, NOT, MUX, ADD, and MUL represent basic logic operator nodes such as AND, OR, XOR, NOT, 2-to-1 multiplexing, addition, and multiplication, respectively; the arrows indicate the signal driving direction or data dependency relationship between nodes.

[0034] An Abstract Syntax Tree (AST) can be obtained by organizing and pruning the Abstract Syntax Tree generated by Pyverilog or other syntax parsing tools. It is used to represent the syntax hierarchy, procedural block structure, conditional relationships, and expression nesting relationships in RTL code. SOG (Simple Bit-Level Logic Operator) is used to represent the connection relationships between simple bit-level logic operators, making the RTL representation closer to the synthesized bit-level logic network. SOG can be constructed by Yosys or other RTL conversion tools after hierarchical expansion and bit-level logic expansion. Specifically, multi-bit operations in the original RTL code are broken down into single-bit logic operators, registers, constants, and basic logic operations are represented as nodes in the graph, and the connection relationships between nodes are established according to signal driving relationships, thus forming a bit-level SOG. SOG is closer to the synthesized bit-level logic network and is beneficial for reflecting the inter-register paths, local logic cones, and bit-level data path structures. AIG is used to characterize Boolean logic structures based on AND gates and inverters; AIG reflects the complexity of Boolean logic implementation in RTL design by converting logical operations into a form composed of AND gates and inverters. AIMG is used to describe multiplexing logic and its selection control relationships. AIMG highlights the multi-level condition judgment and multiplexing structure cascade features by transforming conditional selection relationships and multiplexing logic into a graph form that combines inversion and selection structures. XAG is used to characterize XOR logic and its combination with basic logic operators. By retaining XOR logic and its combination with basic logic operations separately, XAG highlights the structural features of XOR, comparison, and some data path logic.

[0035] AST is better suited to reflect control logic, conditional branches, and nested expression levels; SOG is better suited to reflect bit-level connections and basic logic operation connection structures; AIG is better suited to reflect gate-level logic implementation complexity; AIMG is better suited to identify the cascading characteristics of multiplexing structures; XAG is better suited to reflect the implementation complexity of XOR, comparison, and some data path logic. Through these various structural representations, the same RTL design can be jointly characterized from different perspectives, including control structure, data path structure, bit-level connection structure, and Boolean logic structure.

[0036] 4. Feature extraction and cross-graph alignment After constructing the multi-graph representation, key structural features related to timing, power consumption, and area are further extracted from it, including design-level features, module-level features, node-level features, path-level features, and input cone-level features. Design-level features include the total number of nodes, the total number of edges, the graph depth, and the number and proportion of different node types. Module-level features include the number of registers within a module, the number of logical nodes, the module hierarchy depth, and the number of control branches. Node-level features include node type, fan-in, fan-out, bit width, and the logical level it occupies. Path-level features include the candidate path length between registers, the combinational logic level, critical path-related statistics, the cascade depth of multiplexing structures, and the logic cone complexity.

[0037] For timing-related analysis, the focus can be on extracting the logical hierarchy features of register-to-register paths, the timing-sensitive structural features formed by multi-level control logic, and the features of long paths across modules. For example, for a candidate path from the start register to the end register, its path length feature can be defined. for: ; in, Indicates the number of logical nodes in the path. Indicates the first The approximate delay or hierarchical contribution value corresponding to each logical node. For the same endpoint register, the maximum value among its candidate paths can be used. As a critical path approximation feature, namely: ; in, Indicates all terminating in the register The set of candidate paths. The path length features and critical path approximation features mentioned above can be used to reflect the potential timing pressure on the register endpoints.

[0038] For power consumption-related analysis, further features such as toggle rate, module-level average toggle rate, and high fan-out control signal characteristics can be extracted. For the first... Each node, its flip rate characteristics It can be represented as: ; in, Represents a node The number of flips within the statistics window. Represents a node The corresponding number of clock cycles. Based on the node toggle rate, a module-level average toggle rate feature can also be defined. for: ; in, This indicates the number of nodes participating in the statistics within the module. This module-level average toggle rate characteristic can be used to reflect the strength of the module's dynamic activity and serves as an important basis for power consumption risk assessment.

[0039] For area-related analysis, features related to logical scale, resource complexity, large bit-width operation logical scale, and area contribution can also be extracted.

[0040] After completing the above feature extraction, based on the module name, signal name, register name, bit width information, bit-level index information, instance level information, and source code statement position information recorded during RTL code preprocessing, a correspondence between nodes, paths, or local structures in different graphs and the same RTL object is established. That is, although the same RTL signal, register bit, assign statement, always, or conditional branch may have different forms in AST, SOG, AIG, AIMG, and XAG, they can be aligned through the above RTL object information. This completes the establishment of cross-graph correspondence between different graph representations, enabling nodes or local structures in different graphs to be associated with the same RTL object, providing a foundation for subsequent feature fusion, prediction result mapping, and source code annotation. After cross-graph alignment, corresponding features are extracted from different graphs: AST mainly extracts syntactic hierarchy, procedural block structure, conditional branches, and expression nesting features; SOG mainly extracts bit-level connectivity, number of simple operators, path length, and input cone structure; AIG mainly extracts the Boolean logic complexity composed of AND gates / inverters; AIMG mainly extracts multiplexing structures and control selection relationships; and XAG mainly extracts XOR logic and related combinational structures. Then, the features obtained from different graphs are uniformly mapped to a single or concatenable feature space.

[0041] For historical RTL samples, cross-graph correspondences are used to align structural features extracted from different graph representations to the same RTL object and establish correspondences with reference PPA labels or fine-grained reference PPA labels, thus enabling supervised training of the PPA prediction model. For the RTL design code to be analyzed, cross-graph correspondences are used to form a unified design representation of the RTL design code to be analyzed, which serves as prediction features for PPA prediction. After PPA prediction is completed, the timing prediction results, power prediction results, and area prediction results are mapped to the corresponding RTL source code locations.

[0042] 5. Multi-image feature fusion The structural features extracted from the different graph representations are mapped to a unified dimension, and the syntactic hierarchical information in the AST, the bit-level connection information in the SOG, the Boolean logic structure in the AIG, the multiplexing structure in the AIMG, and the XOR logic structure in the XAG are syntactically integrated to form a unified design representation. This results in a more complete RTL design representation, which integrates the structural information scattered in different graph representations, improves the completeness and consistency of subsequent timing, power consumption and area predictions, and provides a unified input for further source code-level result mapping.

[0043] For example, features extracted from different graph representations can be mapped to a unified feature space and then weighted and combined to form a unified fused feature; alternatively, features from multiple graphs can be spliced ​​together and then a unified mapping method can be used to form a feature representation acceptable to subsequent models.

[0044] 6. PPA Forecast Based on the established PPA prediction model, predictions are made for the RTL design to be analyzed. The unified design representation formed by the RTL design to be analyzed is used as the prediction feature and input into the PPA prediction model. The outputs timing prediction results, power prediction results, and area prediction results. PPA prediction does not only output a single result for the overall design, but establishes corresponding evaluation processes for timing, power, and area respectively. This allows the prediction results to reflect the overall design quality of the integrated circuit and further locate modules, registers, paths, or local logic regions.

[0045] (1) Time series prediction The prediction results include worst-case negative margin (WNS), total negative margin (TNS), and critical path or register-level risk information.

[0046] First, based on the extracted path-level features and input cone-level features, path delay is evaluated for register-to-register candidate paths. For each register endpoint, a slowest candidate path and several randomly sampled paths are extracted, and the arrival times of these paths are predicted. Finally, the endpoint arrival time prediction result is taken as the value with the largest predicted arrival time among all candidate paths for that endpoint. For register signals composed of multiple bits, the bit with the largest arrival time can be taken as the maximum arrival time prediction result for that signal. Regarding the prediction model implementation, path-level or endpoint-level time-series prediction models can employ random forest or XGBoost models to perform regression prediction of path delays; alternatively, a Transformer plus Multilayer Perceptron (MLP) path sequence model can be used to encode the sequence of operation nodes in the path and output the predicted path delay value.

[0047] Then, design-level calibration is performed on the overall WNS and TNS by incorporating design-level features. The design-level calibration model can be an XGBoost regression model, which outputs the overall WNS and TNS prediction results for the design.

[0048] When fine-grained risk ranking is required, the register endpoints can be further ranked for criticality, outputting register-level timing risk levels. Specifically, a learned ranking model can be used to rank register endpoints according to risk priority.

[0049] (2) Power consumption prediction The prediction results include overall power consumption, module-level power consumption distribution, and information on highly flip-sensitive areas.

[0050] In power consumption prediction, in addition to relying on structural features, toggle rate features directly related to dynamic activities can be introduced. Toggle rate features can be obtained from SAIF information generated by RTL simulation or from vector-independent switching activity information provided by synthesis tools. First, node toggle rates are extracted at the node level. Then, at the module level, features such as the average module toggle rate, the number of high-fan-out control signals within the module, the distribution of node types within the module, the number and cascading depth of multiplexed structures, the size of logic cones, and the number of registers / combinational logic are statistically analyzed. These features reflect both the strength of the module's dynamic activities and the structural reasons for power consumption hotspots.

[0051] In terms of model implementation, a module-level power consumption prediction model is first established to predict the module-level power consumption of each module. Then, the module-level prediction results and design-level structural characteristics are summarized and calibrated to obtain the overall design power consumption prediction results. The module-level power consumption prediction model can be implemented using an XGBoost regression model, or a graph convolutional neural network (GCN) can be used to perform graph-level regression on the corresponding subgraphs of the modules to output the module-level power consumption prediction values.

[0052] (3) Area prediction The prediction results include information on the overall design area and the area contribution of each module.

[0053] The area prediction method preferably combines combinational logic area prediction with sequential logic area statistics. First, the combinational logic area is predicted based on the extracted design-level and module-level structural features. The combinational logic area prediction model can employ an XGBoost regression model. The sequential logic area can be directly calculated based on the number of registers, register bit width, and corresponding sequential unit mapping rules. The final overall design area is obtained by adding the predicted combinational logic area value and the statistical value of the sequential logic area. The module-level area contribution can be further obtained based on the module-level structural features and the scale of sequential / combinatorial logic within the module. Finally, the overall design area prediction result and the module area contribution prediction result are obtained, providing a basis for subsequent area hotspot region identification and source code annotation.

[0054] (4) Prediction results output Through timing prediction, power consumption prediction, and area prediction, the final output of the PPA prediction model can be obtained as follows: a. RTL design of overall WNS, TNS, total power consumption, and total area; b. Module-level power consumption distribution and module area contribution; c. Critical path, register-level timing risk information, and information on locally high-flip sensitive areas.

[0055] By predicting timing, power consumption, and area, predictive information corresponding to subsequent logic synthesis, static timing analysis, power consumption analysis, and area statistics can be obtained during the integrated circuit RTL design stage. This allows for the early identification of high-risk logic structures that may affect chip implementation quality and provides a foundation for subsequent result mapping and source code-level annotation.

[0056] 7. Source code-level result mapping and annotation like Figure 4 As shown, firstly, based on the established cross-graph correspondence, the timing prediction results, power prediction results, and area prediction results output by the PPA prediction model are mapped to corresponding locations in the RTL source code, including but not limited to module definition locations, register and net declaration locations, always statement block locations, assign statement locations, conditional branch statement locations, and module instantiation statement locations. Specifically, module-level area contribution can be mapped to module definition locations or module instantiation statement locations; register-level timing risks can be mapped to relevant register declaration locations, register driver statement locations, or their respective always statement block locations; highly flip-sensitive regions or local power hotspots can be mapped to signal declaration locations, assignment statement locations, or control logic locations. Through this step, design-level, module-level, and fine-grained prediction results can be transformed into source code-level location information. Figure 4 In the RTL source code and Two code sections are designated for adding PPA prediction annotation information; source-level PPA annotation examples are added in each of these two sections. This indicates that power consumption risk, area risk, high flip-sensitive region or area contribution label has been added to the first source code object; The text indicates that timing risk, critical path risk, or register-level risk annotations have been added to the second source code object. These two annotations are used to illustrate the annotation results for different source code objects or different risk types; the actual annotation content is not limited to what is shown in the diagram, but is dynamically generated based on the specific RTL code object and its corresponding PPA prediction results.

[0057] Then, add annotation information to the corresponding locations in the RTL source code and output the results. The annotation information includes one or more of the following: timing risk, power consumption risk, area risk, risk prediction value, risk level, and optimization hints; for code areas with obvious risks, further optimization hints can be added.

[0058] Comprehensive risk score of a source code object It can be obtained by a weighted combination of timing risk, power consumption risk, and area risk, and is expressed as: ; in, , and These represent timing risk, power consumption risk, and area risk, respectively. α , β and γ This represents the weighting coefficient for the corresponding risk. All weighting coefficients are non-negative and satisfy the following conditions: α + β + γ =1. The weighting coefficients are determined based on the optimization priorities of timing, power consumption, and area in the target integrated circuit design, the deviation of each PPA index from the preset constraints, and the impact of different risks on the design implementation results. A parameter optimization method can be used: based on different combinations of weighting coefficients, a global optimization is performed on the multi-objective design space of the target integrated circuit design, including timing, power consumption, and area, and the weighting coefficients corresponding to the global optimum are selected. When timing convergence is the primary optimization objective of the target integrated circuit design, the weighting coefficients corresponding to timing risks can be increased. α When the target design has high requirements for low power consumption, the weighting coefficient corresponding to power consumption risk can be increased. β When chip area or logic resource usage is the primary constraint, the weighting coefficient corresponding to area risk can be increased. γ When there are no significant differences in optimization priorities regarding timing, power consumption, and area, it is possible to... α , β and γ Set to equal or approximately equal values. Further risk level or optimization priority information is generated based on the comprehensive risk score to guide designers in prioritizing high-risk code areas.

[0059] The final output is an RTL source code file with annotations or an analysis report containing the correspondence between source code location, risk type, and prediction results. This enables designers to directly locate risky modules, registers, signals, or statements, and to modify and optimize the front-end RTL accordingly.

[0060] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for PPA prediction and source code annotation in integrated circuit RTL design, characterized in that, Includes the following steps: S1. Establish a reference sample set based on historical RTL samples and obtain reference PPA tags. Each historical RTL sample in the reference sample set includes the original RTL code, multiple structural representations generated from the original RTL code, and reference PPA tags corresponding to the design of the historical RTL sample. S2. Perform lexical analysis, syntax analysis, and hierarchical expansion on historical RTL samples in the reference sample set to extract module definitions, port lists, register declarations, net declarations, combinational logic descriptions, sequential logic descriptions, conditional statements, assignment statements, and module instantiation relationships to form a unified intermediate representation; at the same time, record module names, signal names, register names, bit width information, bit-level index information, instance level information, and source code statement location information. S3. Based on the unified intermediate representation formed in step S2, construct multiple graph representations for the same RTL design, including abstract syntax tree, simple operator graph, AND-inverse graph, AND-inverse multiplexing graph, and XOR-AND graph. S4. Based on the multi-graph representation constructed in step S3, extract the structural features that reflect the integrated circuit design, including design-level features, module-level features, node-level features, path-level features, and input cone-level features. Establish cross-graph correspondence between different graph representations based on the module name, signal name, register name, bit width information, bit-level index information, instance-level information, and source code statement position information recorded in step S2. S5. Based on the cross-graph correspondence established in step S4, perform unified dimensional mapping on the structural features extracted in step S4, and form a unified design representation of historical RTL samples by splicing, weighting or fusion of the syntax hierarchy information in the abstract syntax tree, the bit-level connection information in the simple operator graph, the Boolean logic structure in the inverse graph, the multiplexing structure in the inverse multiplexing graph and the XOR logic structure in the XOR graph. S6. Based on the reference PPA labels obtained in step S1, and using the unified design representation formed in step S5 as training features, supervised training is performed to obtain the PPA prediction model; replace the historical RTL samples with the RTL design code to be analyzed and repeat steps S2-S5, using the unified design representation of the RTL design code to be analyzed as prediction features, inputting it into the PPA prediction model, and outputting the timing prediction result, power consumption prediction result and area prediction result. S7. Based on the cross-graph correspondence of the RTL design code to be analyzed, map the timing prediction results, power prediction results and area prediction results obtained in step S6 to the corresponding positions in the RTL source code, and then add annotation information to the corresponding positions and output the results.

2. The PPA prediction and source code annotation method as described in claim 1, characterized in that, In step S1, the reference PPA tag is obtained by performing logic synthesis, static timing analysis, power consumption analysis, and area statistics on historical RTL sample designs. Specifically, logic synthesis maps the RTL code to a gate-level netlist; static timing analysis obtains worst-case margin, total margin, critical path information, and inter-register path timing information; power consumption analysis obtains overall design power consumption, module-level power consumption distribution, and information on local high-flip-sensitive areas; and area statistics are performed based on the synthesis report or implementation report to obtain overall design area, module area contribution, and logic resource usage information.

3. The PPA prediction and source code annotation method as described in claim 1, characterized in that, In step S3, the abstract syntax tree is used to represent the syntax hierarchy, procedural block structure, conditional judgment relationship, and expression nesting relationship in RTL code; the simple operator graph is used to represent the connection relationship between bit-level simple logic operators; the AND-inverting graph is used to depict the Boolean logic structure based on AND gates and inverters; the AND-inverting multiplexing graph is used to describe multiplexing logic and its selection control relationship; and the XOR-AND graph is used to represent XOR logic and its combination relationship with basic logic operators.

4. The PPA prediction and source code annotation method as described in claim 1, characterized in that, In step S4, the design-level features include the total number of nodes, the total number of edges, the graph depth, and the number and proportion of different node types; the module-level features include the number of registers within the module, the number of logical nodes, the module hierarchy depth, and the number of control branches; the node-level features include node type, fan-in, fan-out, bit width, and the logic level; the path-level features include the candidate path length between registers, the combinational logic level, the statistics of different operation nodes on the critical path, the cascading depth of the multiplexing structure, and the size of the logic cone; the input cone-level features include the number of driver registers in the endpoint register input cone, the logic depth of the input cone, and the number of input cone nodes.

5. The PPA prediction and source code annotation method as described in claim 1 or 4, characterized in that, In step S4, the extracted structural features also include toggle rate related features, module-level average toggle rate features, and high fan-out control signal features for power consumption analysis, and logic scale, resource complexity, large bit-width operation logic scale, and area contribution related features for area analysis.

6. The PPA prediction and source code annotation method as described in claim 1, characterized in that, In step S6, the timing prediction results include worst-case negative margin, total negative margin, and critical path or register-level risk information; the power consumption prediction results include overall design power consumption, module-level power consumption distribution, and information on highly flip-flop sensitive areas; and the area prediction results include overall design area and module area contribution information.

7. The PPA prediction and source code annotation method as described in claim 1, characterized in that, In step S7, the corresponding locations in the RTL source code include, but are not limited to, module definition locations, register and net declaration locations, always statement block locations, assign statement locations, conditional branch statement locations, and module instantiation statement locations.

8. The PPA prediction and source code annotation method as described in claim 1, characterized in that, In step S7, the annotation information includes one or more of the following: timing risk, power consumption risk, area risk, risk prediction value, risk level, and optimization prompt information; for code regions with risks, additional optimization prompt information is provided.

9. The PPA prediction and source code annotation method as described in claim 1, 7, or 8, characterized in that, In step S7, the output result is an RTL source code file with annotation information or an analysis report containing the correspondence between source code location, risk type and prediction result.

10. A system for implementing the PPA prediction and source code annotation method according to any one of claims 1-9, comprising: The RTL code acquisition unit is used to receive historical RTL samples of integrated circuits and RTL design code to be analyzed. The reference sample construction and reference EDA analysis unit is used to build a reference sample set from the acquired historical RTL samples according to the reference EDA analysis conditions, and to perform logic synthesis, static timing analysis, power consumption analysis and area statistics on the historical RTL samples to obtain reference PPA tags. The RTL preprocessing unit performs lexical analysis, syntax analysis, and hierarchical expansion on the historical RTL samples and RTL design code received by the RTL code acquisition unit. It extracts module definitions, port lists, register declarations, net declarations, combinational logic descriptions, sequential logic descriptions, conditional statements, assignment statements, and module instantiation relationships to form a unified intermediate representation of the historical RTL samples and the RTL design code to be analyzed. It also records module names, signal names, register names, bit width information, bit-level index information, instance level information, and source code statement location information. The multi-graph representation construction unit is used to receive the unified intermediate representation of historical RTL samples and RTL design code to be analyzed formed by the RTL preprocessing unit, and to construct abstract syntax trees, simple operator graphs, AND-inverse graphs, AND-inverse multiplexing graphs and XOR-AND graphs for the RTL designs in the historical RTL samples and RTL design code to be analyzed, respectively, based on the unified intermediate representation. The feature extraction and cross-graph alignment unit is used to extract design-level features, module-level features, node-level features, path-level features, and input cone-level features from the historical RTL samples constructed by the multi-graph representation construction unit and the multi-graph representation of the RTL design code to be analyzed. Based on the signal name, register name, bit width information, bit-level index information, instance-level information, and source code statement position information recorded by the RTL preprocessing unit, it establishes cross-graph correspondence between different graph representations. The feature fusion unit is used to perform unified dimensional mapping on the structural features of historical RTL samples and the RTL design code to be analyzed based on the cross-graph correspondence established by the feature extraction and cross-graph alignment units, and form the corresponding unified design representation by splicing, weighting or fusion. The PPA prediction unit receives the unified design representation of historical RTL samples formed by the feature fusion unit, and uses the unified design representation of historical RTL samples as training features. Based on the reference samples, it constructs and trains the reference PPA with reference labels provided by the reference EDA analysis unit to obtain the PPA prediction model. The PPA prediction model is then used to predict the RTL design code to be analyzed. The unified design representation of the RTL design code to be analyzed is used as the prediction feature input to the PPA prediction model, and the output is the timing prediction result, power prediction result, and area prediction result. The PPA prediction model includes a timing prediction sub-model, a power prediction sub-model, an area prediction sub-model, and a design-level calibration process. The result mapping unit is used to map the prediction results output by the PPA prediction unit to the corresponding positions in the RTL source code based on the cross-graph correspondence established by the feature extraction and cross-graph alignment units. The source code annotation output unit is used to add annotation information to the corresponding positions in the RTL source code mapped by the result mapping unit, and output the RTL code file or analysis report with annotation information.

Citation Information

Patent Citations

  • Multi-model fusion memory macro-cell design quality prediction system and method

    CN121659892A

  • Chip logic synthesis method based on element reinforcement learning and multi-mode circuit representation

    CN121997854A