Rewrite interp to use apply_ufunc
#9881
Conversation
| @@ -4127,18 +4119,6 @@ def interp( | |||
|
|
|||
| coords = either_dict_or_kwargs(coords, coords_kwargs, "interp") | |||
| indexers = dict(self._validate_interp_indexers(coords)) | |||
|
|
|||
| if coords: | |||
There was a problem hiding this comment.
Handled by vectorize=True. This is possibly a perf regression with numpy arrays, but a massive improvement with chunked arrays.
There was a problem hiding this comment.
For posterity the bad thing about this approach is that it can greatly expand the number of core dimensions for the problem, limiting the potential for parallelism.
Consider the problem in #6799 (comment). In the following, dimension names are listed out in [].
da[time, q, lat, lon].interp(q=bar[lat,lon]) gets rewritten to da[time,q,lat,lon].interp(q=bar[lat, lon], lat=lat[lat], lon=lon[lon]) which thanks to our automatic rechunking, makes dask merge chunks in lat, lon too, for no benefit.
| def _chunked_aware_interpnd(var, *coords, interp_func, interp_kwargs, localize=True): | ||
| """Wrapper for `_interpnd` through `blockwise` for chunked arrays. | ||
| def _interpnd( |
There was a problem hiding this comment.
I merged in two functions to reduce indirection and make it easier to read.
| exclude_dims=all_in_core_dims, | ||
| dask="parallelized", | ||
| kwargs=dict(interp_func=func, interp_kwargs=kwargs), | ||
| dask_gufunc_kwargs=dict(output_sizes=output_sizes, allow_rechunk=True), |
There was a problem hiding this comment.
allow_rechunk=True matches the current behaviour where we rechunk along all core dimensions to a single chunk.
dcherian
force-pushed
the
redo-blockwise-interp
branch
from
245697e to
a5e1854
Compare
dcherian
force-pushed
the
redo-blockwise-interp
branch
from
652a239 to
586f638
Compare
for more information, see https://pre-commit.ci
|
Merging on thursday if there are no comments. IMO this is a big win for maintainability. |
| @@ -566,29 +577,30 @@ def _get_valid_fill_mask(arr, dim, limit): | |||
| ) <= limit | |||
|
|
|||
|
|
|||
| def _localize(var, indexes_coords): | |||
| def _localize(obj: T, indexes_coords: SourceDest) -> tuple[T, SourceDest]: | |||
There was a problem hiding this comment.
Probably should use T_Xarray instead of a plain T to get rid of the type ignore at return.
There was a problem hiding this comment.
That doesn't have Variable, so I'd have to make a new T_DatasetOrVariable or a protocol with .isel perhaps?
Co-authored-by: Michael Niklas <mick.niklas@gmail.com>
apply_ufuncinstead ofblockwisedirectly.vectorize=Trueto get sane dask graphs for vectorized interpolation to chunked arrays (interpperformance with chunked dimensions #6799 (comment))interpperformance with chunked dimensions #6799 (comment)whats-new.rstapi.rstcc @ks905383 your vectorized interpolation example now has this graph:
instead of this quadratic monstrosity
