cva6/perf-model

CVA6 cycle-accurate performance model

This repository contains a cycle-accurate performance model of CVA6 control-path.

It was developed to explore microarchitecture changes in CVA6 before implementing them.

To cite this model, please head to the end of this document.

Getting started

Generate an RVFI trace

To generate an RVFI trace, follow the instructions in the CVA6 repository to run a simulation. The RVFI trace will be in verif/sim/out_<date>/<simulator>/<test-name>.log.

Running the model

python3 model.py verif/sim/out_<date>/<simulator>/<test-name>.log

The annotated trace is generated in an annotated.log file in the current directory.

It prints a lot of debug information so that you can see all the events (note: it slows down model execution). To disable these prints, modify the instantiation with Model(debug=False) in the main function.

At the end of the simulation, the model prints statistics with the print_stats call in the main function. These statistics can be performed on the timed part, which is filtered with the filter_timed_part function. Feel free to modify the filter_timed_part function to suit your needs.

Exploring design space

In model.py, the main function runs the model with arguments which override default values. Generic parameters are available in Model.__init__. You can add new parameters to explore here.

To perform exploration, run the model in a loop, like issue_commit_graph does. The display_scores function is meant to print a 3D plot if you have matplotlib. issue_commit_graph prints the scores so that you can store it and display the figure without re-running the model.

Comparing the model and the RTL

Adapt RVFI trace generation

The regular expression expects the cycle number to be in the RVFI trace. The value is not used by the model but it is used to compare the model and CVA6.

To emit cycle number in RVFI trace, modify corev_apu/tb/rvfi_tracer.sv in CVA6 repository as below.

-        $fwrite(f, "core   0: 0x%h (0x%h) DASM(%h)\n",
-          pc64, rvfi_i[i].insn, rvfi_i[i].insn);
+        $fwrite(f, "core   0: 0x%h (0x%h) @%d DASM(%h)\n",
+          pc64, rvfi_i[i].insn, cycles, rvfi_i[i].insn);

Calculate instruction duration

Run cycle_diff.py for each script.

python3 perf-model/cycle_diff.py annotated.log
mv traceout.log model.log
python3 perf-model/cycle_diff.py verif/sim/out_<date>/<simulator>/<test-name>.log

Note: the cycles_diff.py script filters the "timed" part of the program. To do this, it tries to find csrr minstret instructions. Feel free to modify the script to suit your needs.

List the differences

diff -y traceout.log model.log | less

Files

Name	Description
cycle_diff.py	Calculates duration of each instruction in an RVFI trace
isa.py	Module to create Python objects from RISC-V instructions
model.py	The CVA6 performance model

Citing

@inproceedings{cf24,
   author = {Allart, C\^{o}me and Coulon, Jean-Roch and Sintzoff, Andr\'{e} and Potin, Olivier and Rigaud, Jean-Baptiste},
   title = {Using a Performance Model to Implement a Superscalar CVA6},
   year = {2024},
   isbn = {9798400704925},
   publisher = {Association for Computing Machinery},
   url = {https://doi.org/10.1145/3637543.3652871},
   doi = {10.1145/3637543.3652871},
   abstract = {A performance model of CVA6 RISC-V processor is built to evaluate performance-related modifications before implementing them in RTL. Its accuracy is 99.2\% on CoreMark. This model is used to evaluate a superscalar feature for CVA6. During design phase, the model helped detecting and fixing performance bugs. The superscalar feature resulted in a CVA6 performance improvement of 40\% on CoreMark.},
   booktitle = {Proceedings of the 21st ACM International Conference on Computing Frontiers: Workshops and Special Sessions},
   pages = {43–46},
   numpages = {4},
   keywords = {CVA6, Cycle-Based Model, Multi-Issue, Performance, RISC-V, Superscalar},
   location = {Ischia, Italy},
   series = {CF '24 Companion}
}