Éric Dupuis

- Yes, the visible contour lines. - Sounds good - Yes, the idea I had for a new method could set the last 4 bits to 0. In the case...

In the notebook `a_tutorial.ipynb` accompanying your paper @milankl, there was this method: ```julia keepbits_var = argmax(cumsum(bitinf_var)/sum(bitinf_var) .>= 0.99) - 9 # subtract 9 to count mantissa bits ``` (for "var"...

Haha, that's true, but better be safe than sorry, thanks for the tip! It's a very nice topic theoretically with exciting benefits for data management, so I'm glad I can...

Ah, got it, thanks! The example dataset `eraint_uvz` also has `int16` variables. You could use `dTdx` and `dTdy` from `air_temperature_gradient` instead which are `float32` (there might be other issues I'm...

@rsignell-usgs I didn't focus on this aspect lately (still looking at data with lossy compression somewhere in the pipeline as discussed above). The approach "inspecting *all* bits and keep 99%...

## Small example, 4-5x speed up ```python from xclim.testing import open_dataset import numpy as np from fastnanquantile.xrcompat import xr_apply_nanquantile from xclim.sdba import nbutils as nbu ds = open_dataset("sdba/CanESM2_1950-2100.nc") tx =...

`nbutils` could be improved a lot. We (@aulemahal, @SarahG-579462) were already doing good progress before I stumbled upon `fastnanquantile`. The bottleneck in `nbutils` can be reduced a lot by avoiding...

```python # alternative functions for nbutils @njit def _wrapper_sortquantile(arr,q): out = np.empty((arr.shape[0], q.size), dtype=arr.dtype) for index in range(out.shape[0]): out[index] = _sortquantile(arr[index], q) return out def _quantile2(arr, q, nreduce): if arr.ndim...

Replacing `stack` with `flatten` on temporal dims was already a very big improvement. My next idea was to try to avoid the spatial stacking of dimensions, I was thinking of...

I've adapted Sarah's PR with the functions above ```python %%timeit xr_apply_nanquantile(tx,dim ="time", q=q).compute() >>> 2.23 ms ± 6.08 µs per loop (mean ± std. dev. of 7 runs, 100 loops...