🕳️ handle gaps in x

[jvdd]

The current implementation crashed when there are gaps in the x (i.e., an equidistant x-bin that does not contain any datapoints).

To overcome this issue we now handle gaps

• in the searchsorted functionality (aka the equidistant binning) by returning None when the bin is empty
  -> the output type of the Iterator is thus now Option<(usize, usize)>
• in the M4 and MinMax generic code we check if the bin_size is <= 4 for M4 and 2 for minmax and then add all the indices
  -> the othter case is the default implementation (as was before)

Code updates:

• handle gaps for MinMax
• test this
• handle gaps for M4
• test this
• add Python tests

TODO:

• check performance impact

Benches on the gaps branch ⬇️

minmax_scalx_50M_f32    time:   [47.923 ms 48.396 ms 49.009 ms]
minmax_simdx_50M_f32    time:   [18.414 ms 18.586 ms 18.924 ms]
minmax_scalx_p_50M_f32  time:   [5.0812 ms 5.1216 ms 5.1748 ms]
minmax_simdx_p_50M_f32  time:   [4.0378 ms 4.0979 ms 4.1694 ms]

m4_scalx_50M_f32        time:   [46.977 ms 47.279 ms 47.669 ms]
m4_simdx_50M_f32        time:   [18.329 ms 18.441 ms 18.659 ms]
m4_scalx_p_50M_f32      time:   [7.3339 ms 7.4294 ms 7.5635 ms]
m4_simdx_p_50M_f32      time:   [4.8553 ms 4.8932 ms 4.9480 ms]

mlttb_scalx_50M_f32     time:   [66.573 ms 66.716 ms 66.881 ms]
mlttb_simdx_50M_f32     time:   [39.012 ms 39.743 ms 40.736 ms]
mlttb_scalx_p_50M_f32   time:   [13.433 ms 13.916 ms 14.530 ms]
mlttb_simdx_p_50M_f32   time:   [12.676 ms 12.772 ms 12.875 ms]

Benches on the main branch ⬇️

minmax_scalx_50M_f32    time:   [50.004 ms 51.015 ms 52.223 ms]
minmax_simdx_50M_f32    time:   [22.209 ms 22.425 ms 22.749 ms]
minmax_scalx_p_50M_f32  time:   [4.7367 ms 4.7549 ms 4.7757 ms]
minmax_simdx_p_50M_f32  time:   [3.8919 ms 3.9095 ms 3.9282 ms]

m4_scalx_50M_f32        time:   [49.127 ms 49.451 ms 49.871 ms]
m4_simdx_50M_f32        time:   [18.848 ms 18.865 ms 18.883 ms]
m4_scalx_p_50M_f32      time:   [7.2405 ms 7.2528 ms 7.2660 ms]
m4_simdx_p_50M_f32      time:   [4.9203 ms 4.9338 ms 4.9486 ms]

mlttb_scalx_50M_f32     time:   [66.978 ms 67.068 ms 67.170 ms]
mlttb_simdx_50M_f32     time:   [42.044 ms 43.382 ms 45.161 ms]
mlttb_scalx_p_50M_f32   time:   [13.511 ms 14.040 ms 14.689 ms]
mlttb_simdx_p_50M_f32   time:   [13.285 ms 13.666 ms 14.202 ms]

=> no real statistical differences :)

---

Some remarks 🤔
➡️ The most beautiful & flexible code update would be to alter the get_equidistant_bins to not return an (start, end) when the start & end are the same value (i.e., an empty bin).
-> I might revert the first commit & look into smth like that performed this in PR #25

[jvdd]

LGTM 🙃