implement backward compatibility for sparse matrix multiplication for…

… older xarray versions

pyproject.toml

@@ -1,5 +1,5 @@
 [build-system]
-requires = ["setuptools>=61.0","setuptools-scm>=8","numpy<=1.26","Cython>=3","wheel","pytest","scipy"]
+requires = ["setuptools>=61.0","setuptools-scm>=8","numpy<2","Cython>=3","wheel","pytest","scipy"]
 build-backend = "setuptools.build_meta"
 [project]
 name = "shxarray"
@@ -20,7 +20,11 @@ classifiers = [
     "Development Status :: 1 - Planning"
 ]
 dependencies = [ "pandas >= 2.0", "pyaml >= 23.9.0", "scipy", "xarray >= 2023.1.0",
-"numpy<=1.26","numba", "sparse","importlib_metadata","requests","openpyxl"]
+"numpy<2","numba", "sparse","importlib_metadata","requests","openpyxl"]
+
+[project.optional-dependencies]
+#you need dask in combination with when using older xarray versions
+dask=["dask>=2022.9.2"]
 
 [tool.setuptools_scm]
 # empty for now

src/shxarray/kernels/anisokernel.py

@@ -9,6 +9,9 @@
 from shxarray.core.sh_indexing import SHindexBase
 import sparse
 
+from packaging import version
+
+
 
 
 
@@ -21,6 +24,13 @@ def __init__(self,dsobj,name="aniso",truncate=True):
         self._dskernel=dsobj
         self.name=name 
         self.truncate=truncate
+        self.useDask=version.parse(xr.__version__) < version.parse('2023.11.0')
+        if self.useDask:
+            from dask.array.core import einsum_lookup
+            #Register the einsum functions which are needed to do the sparse dot functions (for earlier versions of xarray)
+            einsum_lookup.register(sparse.COO,AnisoKernel.daskeinsumReplace)
+            #convert to xarray with dask structure
+            self._dskernel=self._dskernel.chunk()
 
     @property
     def nmax(self):
@@ -38,7 +48,10 @@ def __call__(self,dain:xr.DataArray):
         daout=xr.dot(self._dskernel.mat,dain,dims=[SHindexBase.name]) 
         #rename nm and convert to dense array
         daout=daout.sh.toggle_nm()
-        daout=xr.DataArray(daout.data.todense(),coords=daout.coords,name=self.name)
+        if self.useDask:
+            daout=xr.DataArray(daout.compute().data,coords=daout.coords,name=self.name)
+        else:
+            daout=xr.DataArray(daout.data.todense(),coords=daout.coords,name=self.name)
 
         if not self.truncate and self.nmin > 0:
             #also add the unfiltered lower degree coefficients back to the results
@@ -60,4 +73,23 @@ def position(self,lon,lat):
         ynmdata=ynmdata/normv
 
         return self.__call__(ynmdata)
+
+    @staticmethod
+    def daskeinsumReplace(subscripts, *operands, out=None, dtype=None, order='K', casting='safe', optimize=False):
+        """Mimics the interface of https://numpy.org/doc/stable/reference/generated/numpy.einsum.html, but uses the sparse.COO dot function"""
+        if subscripts == "ab,cb->ac":
+            return operands[0].dot(operands[1].T)
+        elif subscripts == "ab,ca->bc":
+            return operands[0].T.dot(operands[1].T)
+        elif subscripts == "ab,bc->ac":
+            return operands[0].dot(operands[1])
+        elif subscripts == "ab,b->a":
+            return operands[0].dot(operands[1])
+        elif subscripts == "ab,a->b":
+            return operands[0].T.dot(operands[1])
+        elif subscripts == "ab,ac->bc":
+            return operands[0].T.dot(operands[1])
+
+        else:
+            raise NotImplementedError(f"Don't know (yet) how to handle this einsum: {subscripts} with sparse.dot operations")
 

tests/test_filters.py

@@ -74,8 +74,9 @@ def test_ddkmult(shinput,shoutput):
 
     dacheck=shinm.sh.filter('DDK2')
     dadiff=(dacheck-shoutm)/shoutm
-    maxdiff=max(abs(dadiff.max()),abs(dadiff.min())) 
-    assert(maxdiff < tol)
+    maxdiff=max(abs(dadiff.max()),abs(dadiff.min()))
+    reltol=3e-11
+    assert(maxdiff < reltol)
 
 @pytest.fixture
 def shgausstest():