CI: Continuous Benchmarking about rv32emu

[jserv]

And one more thing, may I know if @qwe661234 is working on this issue? If not, I could help to write the CI script if the above concern is OK.

Please move forward. Before submitting pull request(s), do write down your plans and communicate here.

[jserv]

If this CI run is enabled for PRs, the benchmark charts will update with each pull request. Since we must first confirm a performance improvement before updating the benchmark charts, this might not be what we desire. In other words, we need to discuss about how the PR is doing before updating the benchmark chart. Or, is it acceptable to update the benchmark chart with each PR?

If the results generated by the CI are deterministic, we can certainly treat them as references for incremental development. However, due to our current lack of confidence, it is premature to define a clear strategy at this stage.

[jserv]

1. Run the benchmark workflow and update the GitHub pages whenever "pushing on master branch" and "opening of a pull request and synchronizing the pull request(if any) on master branch" are performed when editing the "emulate.c" and "rv32_template.c" files because these 2 files have the potential to directly impact the benchmark. Under this constraint, we can prohibit excessive benchmark workflow with unrelated files, such as those in docs/ . Do you have any suggestions for other files that might impact the benchmark?

We shall track decode.c as well.

3. Main benchmark will be “Dhrystone” and “CoreMark” in tests/ for now because I can easily modify the script to generate the JSON format file which required by the benchmark action.

Agree.

[ChinYikMing]

Support pull requests. Instead of updating GitHub pages, add a comment to the pull request to explain benchmark results.

Quote from github-action-benchmark's future work.

If this CI run is enabled for PRs, the benchmark charts will update with each pull request. Since we must first confirm a performance improvement before updating the benchmark charts, this might not be what we desire. In other words, we need to discuss about how the PR is doing before updating the benchmark chart. Or, is it acceptable to update the benchmark chart with each PR?

And one more thing, may I know if @qwe661234 is working on this issue? If not, I could help to write the CI script if the above concern is OK.

[ChinYikMing]

My plan is as follows:

1. Run the benchmark workflow and update the GitHub pages whenever "pushing on master branch" and "opening of a pull request and synchronizing the pull request(if any) on master branch" are performed when editing the "emulate.c" and "rv32_template.c" files because these 2 files have the potential to directly impact the benchmark. Under this constraint, we can prohibit excessive benchmark workflow with unrelated files, such as those in docs/ . Do you have any suggestions for other files that might impact the benchmark?

2. The benchmark comparison between the most recent commit and the previous commit will be noted, allowing the maintainer and committer to debate the benchmark comparison. An example of a benchmark comparison can be seen in the benchmark action README: link

3. Main benchmark will be “Dhrystone” and “CoreMark” in tests/ for now because I can easily modify the script to generate the JSON format file which required by the benchmark action.

4. The x-axis and y-axis unit of benchmark graph will be commit SHA and unit of benchmark respectively. Let’s take “Dhrystone” be an example: the y-axis will be average MIPS over 10 runs. Example benchmark graph taken from the benchmark action README: link

[ChinYikMing]

2. The benchmark comparison between the most recent commit and the previous commit will be noted, allowing the maintainer and committer to debate the benchmark comparison. An example of a benchmark comparison can be seen in the benchmark action README: link

We can set the threshold to trigger the generation of benchmark comparison, for example 0.5x worse than before, default is 2x worse than before.