Conversion of circuit description to an abstract model of the circuit
View Patent ↗A system and method is disclosed for converting an existing circuit description from a lower level description, such as RTL, to a higher-level description, such as TLM, while raising the abstraction level. By changing the abstraction level, the conversion is not simply a code conversion from one language to another, but a process of learning the circuit using neural networks and representing the circuit using a system of equations that approximate the circuit behavior, particularly with respect to timing aspects. A higher level of abstraction eliminates much of the particular implementation details, and allows easier and faster design exploration, analysis, and test, before implementation. In one aspect, a model description of the circuit, protocol information relating to the circuit, and simulation data associated with the lower level description of the circuit are used to generate an abstract model of the circuit that approximates the circuit behavior.
1. An apparatus for converting a description of a circuit into an abstract model of the circuit, the apparatus comprising:
a causality engine coupled to at least one database comprising protocol information relating to the circuit, a model description of the circuit, and simulation data associated with the circuit, the causality engine configured to determine deterministic behavior between input and output signals of the simulation data; and
a neural network coupled to the causality engine and configured to generate an abstract model of the circuit approximating a behavior of the circuit based on the determination of deterministic repetitive behavior by the causality engine;
wherein, to generate an abstract model of the circuit, the neural network is configured to:
generate a system of weighted equations representing the behavior of the circuit;
apply input patterns to the system of equations to generate actual output values;
calculate an error using a difference between the actual output values and desired values; and
modify the weightings in the system of equations based on the calculated error.
2. The apparatus of claim 1 , further comprising a table generator coupled between the causality engine and the neural network, the table generator configured to receive the protocol information, the model description of the circuit, and the simulation data and to generate tables used by the neural network.
3. The apparatus of claim 1 , further comprising a model checker coupled to the neural network, the model checker for testing the abstract model of the circuit.
4. The apparatus of claim 1 , wherein the causality engine is located on a client computer and the neural network is located on a server computer.
5. The apparatus of claim 1 , wherein the causality engine is further configured to define a set of causality characters, each causality character representing an association between an event and an output, and to reduce the set of causality characters through statistical analysis.
6. The apparatus of claim 5 , wherein the causality engine is further configured to reduce the set of causality characters using a genetic optimization algorithm.
7. The apparatus of claim 1 , wherein, to generate an abstract model of the circuit, the neural network is further configured to generate fork tables and latency tables, the fork tables comprising state information and path information associated with the circuit and the latency tables comprising a latency period associated with a simulated event.
8. The apparatus of claim 1 , wherein the model description of the circuit is in a register transfer level and the abstract model of the circuit is at a transaction-level model, the transaction-level model having a different level of abstraction than the register transfer level.