2x perf regression

[noajshu]

I just noticed there is like a 2x perf regression on this surface code circuit. I am not sure if I'm doing something differently now than before.

git checkout c1b44c8975bc745647ce3d93310a4652e9343507
bazel build src:all && time ./bazel-bin/src/tesseract --sample-num-shots 1000 --circuit testdata/surfacecodes/r\=11\,d\=11\,p\=0.002\,noise\=si1000\,c\=surface_code_X\,q\=241\,gates\=cz.stim  --sample-seed 71734  --threads 64  --beam 20   --num-det-orders 21 --beam-climbing --pqlimit 1000000 --det-order-seed 2384257  --print-stats  --no-revisit-dets
num_shots = 1000 num_low_confidence = 0 num_errors = 0 total_time_seconds = 859.6755540000001

git checkout main
# With --no-det-order-bfs:
bazel build src:all && time ./bazel-bin/src/tesseract --sample-num-shots 1000 --circuit testdata/surfacecodes/r\=11\,d\=11\,p\=0.002\,noise\=si1000\,c\=surface_code_X\,q\=241\,gates\=cz.stim  --sample-seed 71734  --threads 64  --beam 20   --num-det-orders 21 --beam-climbing --pqlimit 1000000 --det-order-seed 2384257  --print-stats  --no-revisit-dets --no-det-order-bfs
num_shots = 1000 num_low_confidence = 0 num_errors = 0 total_time_seconds = 1741.893395


# With --det-order-bfs:
bazel build src:all && time ./bazel-bin/src/tesseract --sample-num-shots 1000 --circuit testdata/surfacecodes/r\=11\,d\=11\,p\=0.002\,noise\=si1000\,c\=surface_code_X\,q\=241\,gates\=cz.stim  --sample-seed 71734  --threads 64  --beam 20   --num-det-orders 21 --beam-climbing --pqlimit 1000000 --det-order-seed 2384257  --print-stats  --no-revisit-dets --det-order-bfs
num_shots = 1000 num_low_confidence = 0 num_errors = 0 total_time_seconds = 1854.489493000001

[draganaurosgrbic]

Before I leave, I want to leave these commands here to help you:

git checkout c1b44c8975bc745647ce3d93310a4652e9343507 (COMMIT FROM THE ISSUE)
bazel build src:all && time ./bazel-bin/src/tesseract --sample-num-shots 1000 --circuit testdata/surfacecodes/r\=11\,d\=11\,p\=0.002\,noise\=si1000\,c\=surface_code_X\,q\=241\,gates\=cz.stim  --sample-seed 71734  --threads 64  --beam 20   --num-det-orders 21 --beam-climbing --pqlimit 1000000 --det-order-seed 2384257  --print-stats  --no-revisit-dets
num_shots = 1000 num_low_confidence = 0 num_errors = 0 total_time_seconds = 6485.974051999997

git checkout 8d4a72eddf61d69eccfe07510ddd2c443d24c2d5 (MY FIRST COMMIT)
bazel build src:all && time ./bazel-bin/src/tesseract --sample-num-shots 1000 --circuit testdata/surfacecodes/r\=11\,d\=11\,p\=0.002\,noise\=si1000\,c\=surface_code_X\,q\=241\,gates\=cz.stim  --sample-seed 71734  --threads 64  --beam 20   --num-det-orders 21 --beam-climbing --pqlimit 1000000 --det-order-seed 2384257  --print-stats  --no-revisit-dets
num_shots = 1000 num_low_confidence = 0 num_errors = 0 total_time_seconds = 5715.616996

git checkout a77f5e71a6d3c23981dddb4c5d7fba9a986f3aeb (YOU ADDED VISUALIZATION, COMMIT RIGHT AFTER MY FIRST COMMIT)
bazel build src:all && time ./bazel-bin/src/tesseract --sample-num-shots 1000 --circuit testdata/surfacecodes/r\=11\,d\=11\,p\=0.002\,noise\=si1000\,c\=surface_code_X\,q\=241\,gates\=cz.stim  --sample-seed 71734  --threads 64  --beam 20   --num-det-orders 21 --beam-climbing --pqlimit 1000000 --det-order-seed 2384257  --print-stats  --no-revisit-dets
num_shots = 1000 num_low_confidence = 0 num_errors = 0 total_time_seconds = 21923.21425699999

git checkout main (MAIN BRANCH)
bazel build src:all && time ./bazel-bin/src/tesseract --sample-num-shots 1000 --circuit testdata/surfacecodes/r\=11\,d\=11\,p\=0.002\,noise\=si1000\,c\=surface_code_X\,q\=241\,gates\=cz.stim  --sample-seed 71734  --threads 64  --beam 20   --num-det-orders 21 --beam-climbing --pqlimit 1000000 --det-order-seed 2384257  --print-stats  --no-revisit-dets
num_shots = 1000 num_low_confidence = 0 num_errors = 0 total_time_seconds = 12668.841388

You should look at PR #26. That’s where the slowdown happens. Good luck!

[noajshu]

From gemini-cli:

✦ Success! The performance is back to 73.00 seconds.

This definitively isolates the performance regression to the changes in src/utils.cc.

The change was in the get_detector_coords function.

Older, faster version:

1 for (const stim::DemTarget& t : instruction.target_data) {
2 coord.push_back(t.val());
3 }

Newer, slower version:

1 for (const double& t : instruction.arg_data) {
2 coord.push_back(t);
3 }

[noajshu]

related: p-ranav/argparse#413