vis: partly fix perf regression of Oct 2021
- drops usage of winding_storage_hybrid
- makes pstack_t trivially default initializable again (like the C
original)
- we weren't making use of the unlimited size of winding_storage_hybrid;
the original vis already has an allowance for if a winding exceeds
the fixed stack allocation size (24) and we were still applying that
limit
before:
cmake-build-release\tests\tests.exe -r=xml -sf=*benchmark.cc
Testing started at 1:53 PM ...
| ns/op | op/s | err% | total | benchmark
|--------------------:|--------------------:|--------:|----------:|:----------
| 45.15 | 22,149,002.06 | 0.1% | 0.01 | `create pstack_t`
| 47.04 | 21,256,429.80 | 0.1% | 0.01 | `create pstack_t + 1x AllocStackWinding`
| 48.28 | 20,710,594.32 | 0.2% | 0.01 | `create pstack_t + 2x AllocStackWinding`
| 128.52 | 7,780,603.30 | 0.3% | 0.01 | `setup + ClipStackWinding`
Benchmark from: https://www.quaddicted.com/tools/vis_benchmark
cmake-build-release\vis\vis.exe -nostate -nopercent gmsp3v2.bsp
153.31s elapsed
after:
cmake-build-release\tests\tests.exe -r=xml -sf=*benchmark.cc
Testing started at 1:46 PM ...
| ns/op | op/s | err% | total | benchmark
|--------------------:|--------------------:|--------:|----------:|:----------
| 1.64 | 610,528,630.06 | 0.3% | 0.01 | `create pstack_t`
| 3.29 | 303,895,535.45 | 0.6% | 0.01 | `create pstack_t + 1x AllocStackWinding`
| 5.20 | 192,172,148.84 | 0.2% | 0.01 | `create pstack_t + 2x AllocStackWinding`
| 66.84 | 14,960,818.66 | 1.0% | 0.01 | `setup + ClipStackWinding`
Benchmark from: https://www.quaddicted.com/tools/vis_benchmark
cmake-build-release\vis\vis.exe -nostate -nopercent gmsp3v2.bsp
141.42s elapsed
e9532346c7