Get column value by column label in Framework Core ASoA

[mytkom]

I will continue our discussion here @ktf, @jgrosseo, @saganatt.

@jgrosseo was correct; this solution shouldn't be implemented as it is. It was slower in this benchmark by 3-10 times for simple examples. I will leave this draft PR here and try to implement this alternative solution I was writing about in AliceO2Group/O2Physics#7371. That 20% I was speaking of during the presentation was the whole task performance difference locally on my laptop. These are my local benchmark results:

------------------------------------------------------------------------------------------------------------
Benchmark                                                  Time             CPU   Iterations UserCounters...
------------------------------------------------------------------------------------------------------------
BM_ASoADynamicColumnPresent/8                            237 ns          236 ns      2983783 bytes_per_second=258.465M/s
BM_ASoADynamicColumnPresent/64                           395 ns          395 ns      1717027 bytes_per_second=1.20857G/s
BM_ASoADynamicColumnPresent/512                         1559 ns         1555 ns       448023 bytes_per_second=2.45328G/s
BM_ASoADynamicColumnPresent/4096                       10961 ns        10933 ns        62651 bytes_per_second=2.79137G/s
BM_ASoADynamicColumnPresent/32768                      85382 ns        85164 ns         8024 bytes_per_second=2.86673G/s
BM_ASoADynamicColumnPresent/262144                    679514 ns       677289 ns          998 bytes_per_second=2.88374G/s
BM_ASoADynamicColumnPresentGetValueByLabel/8             419 ns          418 ns      1622625 bytes_per_second=145.876M/s
BM_ASoADynamicColumnPresentGetValueByLabel/64           1772 ns         1768 ns       392634 bytes_per_second=276.255M/s
BM_ASoADynamicColumnPresentGetValueByLabel/512         12671 ns        12641 ns        54994 bytes_per_second=309.019M/s
BM_ASoADynamicColumnPresentGetValueByLabel/4096        99301 ns        99056 ns         6893 bytes_per_second=315.478M/s
BM_ASoADynamicColumnPresentGetValueByLabel/32768      805034 ns       802857 ns          858 bytes_per_second=311.388M/s
BM_ASoADynamicColumnPresentGetValueByLabel/262144    6319410 ns      6301477 ns          107 bytes_per_second=317.386M/s
BM_ASoADynamicColumnCall/8                              25.3 ns         25.2 ns     28377608 bytes_per_second=2.36521G/s
BM_ASoADynamicColumnCall/64                              183 ns          182 ns      3722494 bytes_per_second=2.61903G/s
BM_ASoADynamicColumnCall/512                            1378 ns         1375 ns       478496 bytes_per_second=2.7751G/s
BM_ASoADynamicColumnCall/4096                          11039 ns        11012 ns        62893 bytes_per_second=2.77126G/s
BM_ASoADynamicColumnCall/32768                         88298 ns        88067 ns         7847 bytes_per_second=2.77221G/s
BM_ASoADynamicColumnCall/262144                       706708 ns       704322 ns          996 bytes_pelementaryer_second=2.77306G/s
BM_ASoADynamicColumnCallGetValueByLabel/8               67.6 ns         67.4 ns     10002511 bytes_per_second=905.098M/s
BM_ASoADynamicColumnCallGetValueByLabel/64               500 ns          498 ns      1345915 bytes_per_second=979.614M/s
BM_ASoADynamicColumnCallGetValueByLabel/512             3981 ns         3972 ns       175994 bytes_per_second=983.554M/s
BM_ASoADynamicColumnCallGetValueByLabel/4096           31831 ns        31757 ns        22014 bytes_per_second=984.025M/s
BM_ASoADynamicColumnCallGetValueByLabel/32768         255368 ns       254706 ns         2722 bytes_per_second=981.526M/s
BM_ASoADynamicColumnCallGetValueByLabel/262144       2029059 ns      2022533 ns          340 bytes_per_second=988.859M/s

[github-actions]

REQUEST FOR PRODUCTION RELEASES:
To request your PR to be included in production software, please add the corresponding labels called "async-" to your PR. Add the labels directly (if you have the permissions) or add a comment of the form (note that labels are separated by a ",")

+async-label <label1>, <label2>, !<label3> ...

This will add <label1> and <label2> and removes <label3>.

The following labels are available
async-2023-pbpb-apass3
async-2023-pbpb-apass4
async-2023-pp-apass4
async-2024-pp-apass1
async-2022-pp-apass7
async-2024-pp-cpass0

[ktf]

Thank you for your investigation. Much appreciated.

As a general note, please try to have a smaller contributions if possible, because that will for sure simplify and speedup the review process.

[mytkom]

Previously, I used std::function as the return type for the getter, but the performance was pretty bad. I changed O2 to debug mode and used perf + speedscope to analyze it. I obtained a much better performance, almost equal to an explicit getter call (less than 10% execution time difference). Then I switched back to release mode for compilation, and these are the results:

-------------------------------------------------------------------------------------------------------------
Benchmark                                                   Time             CPU   Iterations UserCounters...
-------------------------------------------------------------------------------------------------------------
BM_ASoADynamicColumnPresent/8                             236 ns          235 ns      2951784 bytes_per_second=259.24M/s
BM_ASoADynamicColumnPresent/64                            397 ns          396 ns      1765157 bytes_per_second=1.2049G/s
BM_ASoADynamicColumnPresent/512                          1557 ns         1553 ns       454082 bytes_per_second=2.45591G/s
BM_ASoADynamicColumnPresent/4096                        10783 ns        10758 ns        63512 bytes_per_second=2.83666G/s
BM_ASoADynamicColumnPresent/32768                       85087 ns        84863 ns         8263 bytes_per_second=2.87686G/s
BM_ASoADynamicColumnPresent/262144                     683696 ns       681478 ns         1030 bytes_per_second=2.86601G/s
BM_ASoADynamicColumnPresentGetGetterByLabel/8             277 ns          276 ns      2529198 bytes_per_second=221.199M/s
BM_ASoADynamicColumnPresentGetGetterByLabel/64            658 ns          656 ns      1077150 bytes_per_second=744.385M/s
BM_ASoADynamicColumnPresentGetGetterByLabel/512          3490 ns         3482 ns       199110 bytes_per_second=1121.89M/s
BM_ASoADynamicColumnPresentGetGetterByLabel/4096        27125 ns        27054 ns        26808 bytes_per_second=1.12801G/s
BM_ASoADynamicColumnPresentGetGetterByLabel/32768      208795 ns       208255 ns         3366 bytes_per_second=1.17232G/s
BM_ASoADynamicColumnPresentGetGetterByLabel/262144    1668853 ns      1663740 ns          423 bytes_per_second=1.17394G/s
BM_ASoADynamicColumnCall/8                               24.8 ns         24.7 ns     28070062 bytes_per_second=2.41155G/s
BM_ASoADynamicColumnCall/64                               180 ns          180 ns      3630937 bytes_per_second=2.65376G/s
BM_ASoADynamicColumnCall/512                             1377 ns         1374 ns       497025 bytes_per_second=2.77621G/s
BM_ASoADynamicColumnCall/4096                           10961 ns        10935 ns        63683 bytes_per_second=2.79084G/s
BM_ASoADynamicColumnCall/32768                          87853 ns        87642 ns         8035 bytes_per_second=2.78565G/s
BM_ASoADynamicColumnCall/262144                        703888 ns       701646 ns          982 bytes_per_second=2.78363G/s
BM_ASoADynamicColumnCallGetGetterByLabel/8               40.0 ns         39.9 ns     17621092 bytes_per_second=1.4949G/s
BM_ASoADynamicColumnCallGetGetterByLabel/64               299 ns          298 ns      2335159 bytes_per_second=1.60043G/s
BM_ASoADynamicColumnCallGetGetterByLabel/512             2400 ns         2394 ns       300728 bytes_per_second=1.59373G/s
BM_ASoADynamicColumnCallGetGetterByLabel/4096           18681 ns        18635 ns        36845 bytes_per_second=1.63765G/s
BM_ASoADynamicColumnCallGetGetterByLabel/32768         148504 ns       148136 ns         4728 bytes_per_second=1.64809G/s
BM_ASoADynamicColumnCallGetGetterByLabel/262144       1189493 ns      1185832 ns          571 bytes_per_second=1.64705G/s

I am currently stuck with this solution; I would like to know if using the get() explicitly column struct member method is possible. I was trying to implement such an approach, but there is a type problem; the member method pointer type is ex. float(*ColumnType)(const RowItType&), and I would know ColumnType only in runtime after string comparison, so I can't think of an efficient implementation.

I would like to know if you think such a performance difference would be acceptable and if you can think of a more efficient solution.

[mytkom]

@ktf @jgrosseo,
May you have a look at the latest changes?

[ktf]

@mytkom thank you for your work. See inline comments. I might have more once those are resolved.

[mytkom]

@ktf
I have finally made the changes requested in your last review. Could you take a look at the current implementation?

[mytkom]

@ktf
I simplified this using nullptr as you advised; now I wonder if this strncmp in my implementation is not making it overcomplicated for such a comparison. Could you take a look and give me your opinion on this? Maybe using string + strcmp for str.data() instead of string_view would be more readable and performance drawback should not be so heavy.