Lockstep co-simulation

[jason23g]

Hello guys,

By reading the README files for the verification and simulations, I understand that the co-simulation of CVA6 is being achieved by comparing the log trace of one RTL simulators such as verilator or vcs with the log trace of the Spike simulator. However, I view in another closed issue that for CV32A65X there is the possibility for lockstep co-simulation between vcs and spike, but I can not find where this is happening in the repo. Could you clarify if it is implemented the lockstep co-simulation and where they are the corresponding scripts?

Thanks in advance for your time.

[JeanRochCoulon]

The Spike Tandem is implemented, this is a co-simulation of RTL and Spike to check the RTL behaviour. This is implemented for cv32a65x and not already deployed on other CVA6 configurations. To enable it, you should set SPIKE_TANDEM=1 when executing a simulation with DV_SIMULATOR=vcs-uvm (or vcs-testhaness or veri-testharness).

[jason23g]

Thanks for your answer, indeed set the parameters as you indicate , i run a co-simulation of RTL and spike, however in the log info shows that the TANDEM YAML report is being incomplete, is this something that should happen or not ?

[jason23g]

UPDATE !!

SETUP :
The setup of the whole system has been made following the README file of the repository. After system setup i set the env variable SPIKE_TANDEM = 1, UVM_VERBOSITY=UVM_MEDIUM and the DV_SIMULATOR= veri-testharness. Then i have created a custom script to run only tests for the configuration CVA32A65X.

PROBLEM :

The first strange thing that i observe is the message "Incomplete TANDEM YAML report". After a thorough inspection of the source code of cva6.py, the sv files used for the simulation and the generated log files with expansion .log.iss i observe that even if the final messages are saying *** SUCESS *** , the function rvfi_gen_report in the file uvmc_rvfi_scoreboard_utils.sv is never being executed and so on it is never producing a yaml report. The reason for that is both the reference_model_t and t_core instance have al the fields for the whole simulation time equal to zero and the corresponding instruction equal to c.unimp

QUESTION :
Is this a bug or did i miss anything in the process setting up the repository.

logs.zip
custom_test.zip

[JeanRochCoulon]

We meet difficulties to make functional Verilator in Tandem mode. Verification flow using VCS is more robust.

[valentinThomazic]

Related to #2526, please take a look at my answer there.