Significantly faster cohorts detection.

[dcherian]

Closes #271

I was iterating over array.blocks to figure out the shape of each chunk.

When indexing this object, it creates a dask array per chunk, which is slow for many reasons

• dask array construction, which is useless
• calling np.array() on a chunks tuple repeatedly in a loop (surprisingly slow!)

Replace with a function that calculates the chunk shape. On the arco era5 with 93044 time chunks, this is a speedup from infinite time to 840ms.

TIME = 92044
da = xr.DataArray(
    dask.array.ones((TIME, 721, 1440), chunks=(1, -1, -1)),
    dims=("time", "lat", "lon"),
    coords=dict(time=pd.date_range("1959-01-01", freq="6H", periods=TIME)),
)
%time xarray_reduce(da, da.time.dt.day, method="cohorts", func="any")

[dcherian]

The profile looks like

   225     92045   36723000.0    399.0      5.7      for idx, blockindex in enumerate(np.ndindex(array.numblocks)):
   226     92044  169811000.0   1844.9     26.5          chunkshape = get_chunk_shape(array_chunks, blockindex)
   227     92044  142355000.0   1546.6     22.2          blocks[idx] = np.full(chunkshape, idx)
   228         1  189328000.0    2e+08     29.5      which_chunk = np.block(blocks.reshape(shape).tolist()).reshape(-1)

I strongly suspect we can do better. The tolist is copying blocks which should be unnecessary.

[dcherian]

Down to

546 ms ± 3.68 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

[dcherian]

I strongly suspect we can do better.

Need a way to assign to a nested list. This is why I chose the numpy object array route in the first place :)

Punting to later, since this is already a massive improvement.

[dcherian]

Change	Before [9f82e19]	After [fa93406]	Ratio	Benchmark (Parameter)
-	192±1ms	172±4ms	0.9	cohorts.ERA5DayOfYear.time_graph_construct
-	186±3ms	158±2ms	0.85	cohorts.ERA5DayOfYearRechunked.time_graph_construct
-	95.5±0.9ms	68.6±0.9ms	0.72	cohorts.ERA5DayOfYearRechunked.time_find_group_cohorts
-	48.0±0.1ms	25.0±0.1ms	0.52	cohorts.ERA5DayOfYear.time_find_group_cohorts
-	44.5±2ms	18.6±0.6ms	0.42	cohorts.ERA5MonthHourRechunked.time_graph_construct
-	42.3±2ms	17.1±0.2ms	0.4	cohorts.ERA5MonthHour.time_graph_construct
-	9.52±0.04ms	3.72±0.01ms	0.39	cohorts.PerfectMonthly.time_graph_construct
-	9.60±0.2ms	3.72±0.02ms	0.39	cohorts.PerfectMonthlyRechunked.time_graph_construct
-	73.5±0.9ms	25.4±0.5ms	0.35	cohorts.time_cohorts_era5_single
-	31.2±0.2ms	7.58±0.1ms	0.24	cohorts.ERA5MonthHour.time_find_group_cohorts
-	34.1±0.1ms	7.80±0.2ms	0.23	cohorts.ERA5MonthHourRechunked.time_find_group_cohorts
-	6.87±0.03ms	1.02±0.04ms	0.15	cohorts.PerfectMonthlyRechunked.time_find_group_cohorts
-	6.95±0.1ms	1.00±0.01ms	0.14	cohorts.PerfectMonthly.time_find_group_cohorts

[dcherian]

Benchmarks seem to be broken after the numbagg PR. I'll fix in a new branch.