Skip to content

Conversation

@gvreuls
Copy link
Contributor

@gvreuls gvreuls commented Aug 6, 2020

This patch consists of two parts:

  1. Extend the UCI bench command so it can run benchmarks with either the
    classical or the NNUE evaluation by adding a sixth "USE NNUE" option
    that must be "true" or "false", the default is to use classical
    evaluation ("false").

  2. Extend the PGO build process to a) download the NNUE net and b) run
    the bench twice, once in classical and once in NNUE mode.

No functional change.

This patch consists of two parts:

1) Extend the UCI bench command so it can run benchmarks with either the
classical or the NNUE evaluation by adding a sixth "USE NNUE" option
that must be "true" or "false", the default is to use classical
evaluation ("false").

2) Extend the PGO build process to a) download the NNUE net and b) run
the bench twice, once in classical and once in NNUE mode.

No functional change.
@gvreuls
Copy link
Contributor Author

gvreuls commented Aug 7, 2020

I somehow totally missed mstembera's #2879 Closing, this wasn't intended as competition.

@gvreuls gvreuls closed this Aug 7, 2020
@mstembera
Copy link
Contributor

Does the second bench run
$(PGOBENCH) false > /dev/null
$(PGOBENCH) true > /dev/null
append to the results of the first run or overwrite them?

@gvreuls
Copy link
Contributor Author

gvreuls commented Aug 7, 2020

@mstembera The second bench run appends to the *.gcda files created by the first run.

gcov-dump evaluate_nnue.gcda after the first bench run produces:

evaluate_nnue.gcda:data:magic gcda':version A93R'
evaluate_nnue.gcda:stamp 3316773654
evaluate_nnue.gcda: a1000000: 2:OBJECT_SUMMARY runs=1, sum_max=26963616
...

after the second run it produces:

evaluate_nnue.gcda:data:magic gcda':version A93R'
evaluate_nnue.gcda:stamp 3316773654
evaluate_nnue.gcda: a1000000: 2:OBJECT_SUMMARY runs=2, sum_max=422975776
...

@mstembera
Copy link
Contributor

@gvreuls You may want to keep this open. I like the simplicity of running the bench twice. @vondele also has a PR #2902 related to this and he may be able to come up w/ a better final solution after taking a look at this.

@gvreuls
Copy link
Contributor Author

gvreuls commented Aug 7, 2020

Okay, reopened.

@gvreuls gvreuls reopened this Aug 7, 2020
@gvreuls
Copy link
Contributor Author

gvreuls commented Aug 7, 2020

Closed in favor of #2931

@gvreuls gvreuls closed this Aug 7, 2020
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants