Add sprite sorting benchmark

[janisozaur]

See #8442

[IntelOrca]

Why can't the benchmark just be run on the actual game?

[janisozaur]

It probably could, that also gave me the idea that I could extract paint session data from a park passed in an argument, which solves the long compilation time and the need to extract it beforehand.

I'll work on that.

[janisozaur]

I purposely implemented the benchmark as a single file initially so it is easier to experiment with. Right now there is no verification of the order of sorted sprites is valid (same as some predefined order), but with screenshot command perhaps I wouldn't need to implement that.

[ZehMatt]

I think it would be slightly better to move the sorter out into its own and let the benchmark use that code, otherwise this adds duplicate code.

[IntelOrca]

The problem is that that test input data will become out of date very quickly particularly if the structs get changed. It would be much better (easier to maintain) if it just benchmarked via calling the paint function with a loaded saved game.

[ZehMatt]

I agree with @IntelOrca here, it would probably be better to have a benchmark like I described here: #6339 (comment)

[janisozaur]

This was significantly improved, it will now load the park, automatically extract the needed data and time the sprite sorting function.

This is not done yet, but should be actually usable now.

Things to improve:

• The park to load is hardcoded to data.sv6
• Make sure to actually load park only once
• Consider integration with main binary
• Current approach is to pass a pointer to std::vector<paint_struct> to viewport paint functions. Perhaps a better approach would be to pass a function object/lambda?
• Consider adding some baseline
• Remove the test/tests/SpriteBenchmarkData.inc, now obsolete
• Hook up CI
• Update the comment
• Implement some benchmark performance improvements
• Make a call to state.setItemsProcessed() to get nice performance labels

To launch: compile, select some park, name it data.sv6 in current directory, start the test. It can take a looong time to finish, I suggest running it this way:

$ ./test_sprite_sort_benchmark --benchmark_min_time=0.001
<some spam here>
----------------------------------------------------------------
Benchmark                         Time           CPU Iterations
----------------------------------------------------------------
BM_paint_session_arrange      14417 ns      11830 ns        114

[ZehMatt]

I have somewhat mixed feelings about this benchmark. If we decide to go for a better painting structure the whole code may become obsolete and I'm not sure you can actually improve the current sorting method to something that has a measurable increase, so measuring the current sorter is somewhat pointless, I would rather know how much time was spent to a build the entire frame and how long each component took. I'm just worried that the sorting code will either be dropped entirely or that no one comes up with a much faster version.

[janisozaur]

We can remove the benchmark code if its structure changes to the point where it can longer be supported. Before that time comes though I'd like to merge this, I would like to try out some things with the sprite sorter. Additionally, it would be easier to implement more benchmarks once this lands for other parts of code.

[janisozaur]

Thing still needed from the list above and others:

• Hook up CI? Would likely require objects, but it builds now, which satisfies me.
• Make sure to actually load park only once
• Correct parsing of arguments for benchmark library
• Un-hardcode the park being loaded.

[IntelOrca]

You can also update https://github.com/OpenRCT2/OpenRCT2/wiki/Benchmarking-&-stress-testing-OpenRCT2 afterwards.

[janisozaur]

This should be considered done now.

Make sure google benchmark is installed, build, then:

$ ./openrct2 benchspritesort ~/.config/OpenRCT2/save/dome-roof-on_zoom1.sv6 --benchmark_min_time=0.001
INFO[../src/openrct2/cmdline/BenchSpriteSort.cpp:74 (extract_paint_session)]: Starting...
INFO[../src/openrct2/cmdline/BenchSpriteSort.cpp:128 (extract_paint_session)]: Obtaining sprite data...
INFO[../src/openrct2/cmdline/BenchSpriteSort.cpp:134 (extract_paint_session)]: Got 242 paint sessions.
2019-01-21 23:08:31
Run on (8 X 4400 MHz CPU s)
CPU Caches:
  L1 Data 32K (x4)
  L1 Instruction 32K (x4)
  L2 Unified 256K (x4)
  L3 Unified 8192K (x1)
-----------------------------------------------------------------------------------------------------
Benchmark                                                              Time           CPU Iterations
-----------------------------------------------------------------------------------------------------
baseline                                                             437 ns        376 ns       2680   2.53875M items/s
/home/janisozaur/.config/OpenRCT2/save/dome-roof-on_zoom1.sv6       8450 ns       6784 ns        194   34.0183M items/s