waveshaper: Avoid allocations while resampling

[uklotzde]

Fixes #143.

[github-actions]

Benchmark result:


bench_ctor
  Instructions:             5063057 (+0.232196%)
  L1 Accesses:              7559138 (+0.208008%)
  L2 Accesses:                54122 (+0.024026%)
  RAM Accesses:               61585 (+0.042236%)
  Estimated Cycles:         9985223 (+0.167186%)

bench_sine
  Instructions:            78716964 (+0.234685%)
  L1 Accesses:            115623207 (+0.268041%)
  L2 Accesses:               291029 (-5.125639%)
  RAM Accesses:               62681 (+0.033514%)
  Estimated Cycles:       119272187 (+0.194226%)

bench_sine_gain
  Instructions:            83527005 (+0.325033%)
  L1 Accesses:            122886134 (+0.365048%)
  L2 Accesses:               308987 (-4.551745%)
  RAM Accesses:               62835 (+0.030247%)
  Estimated Cycles:       126630294 (+0.296185%)

bench_sine_gain_delay
  Instructions:           157632228 (+0.226942%)
  L1 Accesses:            224092252 (+0.261607%)
  L2 Accesses:               724083 (-1.844409%)
  RAM Accesses:               64518 (+0.023255%)
  Estimated Cycles:       229970797 (+0.225408%)

bench_buffer_src
  Instructions:            18613888 (+0.937175%)
  L1 Accesses:             26917457 (+1.053067%)
  L2 Accesses:                93433 (-6.894731%)
  RAM Accesses:               96123 (+0.022893%)
  Estimated Cycles:        30748927 (+0.808727%)

bench_buffer_src_delay
  Instructions:            91698570 (+0.283399%)
  L1 Accesses:            126801446 (+0.317627%)
  L2 Accesses:               219921 (+5.257087%)
  RAM Accesses:               96311 (+0.017654%)
  Estimated Cycles:       131271936 (+0.349345%)

bench_buffer_src_iir
  Instructions:            42590408 (+3.444808%)
  L1 Accesses:             61836734 (+3.076097%)
  L2 Accesses:                99856 (-11.30219%)
  RAM Accesses:               96220 (+0.022869%)
  Estimated Cycles:        65703714 (+2.788657%)

bench_buffer_src_biquad
  Instructions:            39283874 (+1.336580%)
  L1 Accesses:             55241000 (+1.538339%)
  L2 Accesses:               196774 (-3.816563%)
  RAM Accesses:               96327 (+0.021805%)
  Estimated Cycles:        59596315 (+1.358240%)

bench_stereo_positional
  Instructions:            47271739 (+3.236776%)
  L1 Accesses:             69839904 (+3.432274%)
  L2 Accesses:               779452 (+8.991705%)
  RAM Accesses:               96492 (+0.023842%)
  Estimated Cycles:        77114384 (+3.544672%)

bench_stereo_panning_automation
  Instructions:            32791310 (+0.785521%)
  L1 Accesses:             48782168 (+0.921552%)
  L2 Accesses:               169025 (-18.87566%)
  RAM Accesses:               96248 (+0.027021%)
  Estimated Cycles:        52995973 (+0.473433%)

bench_analyser_node
  Instructions:            39165878 (+0.434069%)
  L1 Accesses:             54834449 (+0.511961%)
  L2 Accesses:               208391 (-5.544728%)
  RAM Accesses:               96821 (+0.027894%)
  Estimated Cycles:        59265139 (+0.371034%)

[orottier]

Thanks a lot for you contribution!

I need a bit more time to review and verify it in full, but I left a few first remarks. Could you have a look?

If you wish, you could add a benchmark to benches/my_benchmark.rs to see if it actually outperforms the previous version (it will run in the CI and post a comment on your PR).

[uklotzde]

If you wish, you could add a benchmark to benches/my_benchmark.rs to see if it actually outperforms the previous version (it will run in the CI and post a comment on your PR).

Benchmarking is not helpful here. The allocation might only cause issues very rarely but in fatal ways. Same applies to mutexes. Lock-free code is often less performant on average than locking.

[github-actions]

Benchmark result:


bench_ctor
  Instructions:             5063063 (+0.232196%)
  L1 Accesses:              7559159 (+0.208247%)
  L2 Accesses:                54111 (+0.011090%)
  RAM Accesses:               61585 (+0.024363%)
  Estimated Cycles:         9985189 (+0.163147%)

bench_sine
  Instructions:            78716970 (+0.234677%)
  L1 Accesses:            115623937 (+0.256801%)
  L2 Accesses:               290303 (-0.956985%)
  RAM Accesses:               62687 (+0.036704%)
  Estimated Cycles:       119269497 (+0.237794%)

bench_sine_gain
  Instructions:            83527015 (+0.325036%)
  L1 Accesses:            122870414 (+0.338513%)
  L2 Accesses:               324714 (+5.763142%)
  RAM Accesses:               62842 (+0.039798%)
  Estimated Cycles:       126693454 (+0.399299%)

bench_sine_gain_delay
  Instructions:           157632304 (+0.226945%)
  L1 Accesses:            224043689 (+0.248001%)
  L2 Accesses:               772720 (+2.228814%)
  RAM Accesses:               64524 (+0.032557%)
  Estimated Cycles:       230165629 (+0.278498%)

bench_buffer_src
  Instructions:            18613906 (+0.937261%)
  L1 Accesses:             26916452 (+1.029204%)
  L2 Accesses:                94470 (-0.630069%)
  RAM Accesses:               96114 (+0.018731%)
  Estimated Cycles:        30752792 (+0.891829%)

bench_buffer_src_delay
  Instructions:            91698592 (+0.283438%)
  L1 Accesses:            126817169 (+0.319560%)
  L2 Accesses:               204233 (+4.367178%)
  RAM Accesses:               96306 (+0.014539%)
  Estimated Cycles:       131209044 (+0.341985%)

bench_buffer_src_iir
  Instructions:            42590424 (+3.444786%)
  L1 Accesses:             61828948 (+3.040860%)
  L2 Accesses:               107666 (+8.044155%)
  RAM Accesses:               96218 (+0.018711%)
  Estimated Cycles:        65734908 (+2.920573%)

bench_buffer_src_biquad
  Instructions:            39283866 (+1.336538%)
  L1 Accesses:             55252702 (+1.505409%)
  L2 Accesses:               185080 (+5.506784%)
  RAM Accesses:               96319 (+0.014537%)
  Estimated Cycles:        59549267 (+1.479580%)

bench_stereo_positional
  Instructions:            47271739 (+3.236754%)
  L1 Accesses:             69842817 (+3.464888%)
  L2 Accesses:               776544 (+5.849398%)
  RAM Accesses:               96486 (+0.013475%)
  Estimated Cycles:        77102547 (+3.425892%)

bench_stereo_panning_automation
  Instructions:            32791352 (+0.785724%)
  L1 Accesses:             48768116 (+0.826245%)
  L2 Accesses:               183131 (+3.715219%)
  RAM Accesses:               96244 (+0.019745%)
  Estimated Cycles:        53052311 (+0.823097%)

bench_analyser_node
  Instructions:            39165877 (+0.434074%)
  L1 Accesses:             54839178 (+0.487415%)
  L2 Accesses:               203669 (+0.534588%)
  RAM Accesses:               96816 (+0.014463%)
  Estimated Cycles:        59246083 (+0.461054%)

[orottier]

Lock-free code is often less performant on average than locking.

Fair point! Although I'm reasonably sure process_into_buffer is both lock-free and more performant :)

I have merged your updated code. Thanks again