goodboy
commented
1 year ago As follow up after the huge rework in #420 and it's derivative child PRs (ending in https://github.com/pikers/piker/pull/455), the 2 main spots that pertain to this are:
piker.data._pathops.ohlc_flatten()which is currently unused- and
.path_arrays_from_ohlc()which is used in the rendering subsys for OHLC graphics downsampling
Currently our OHLC sampling (bars) graphics convert to a line curve when the user zooms out enough that you can't visually distinguish a bar's shape (uppx is way too high to represent them per horizonal pixel).
To convert to this line we currently use a structured array flattener routine which uses
np.recfunctions.structure_to_unstructured():ohlc_to_line().In theory we can actually just use our m4 implementation to post-process either this flattened output or write an OHLC aware version which iterates the
['high', 'low']fields of the input array when computing max/min values. In other words theyinput could be pre-processed from another (numba) routine which flattens only the high/low values from the OHLC struct array and then feeds them into this same function. The caveat will be making sure thexinput is sized to match 🤔This is definitely not a high priority task but may be interesting to tinker with especially in the context of overloading
ds_m4()for other types of aggregation schemes (eg. as we still need with vlm in #326).