fix #705: made LArray.divnot0 faster when the divisor array has many axes and many zeros (#708)

fix #705: made LArray.divnot0 faster when the divisor array has many axes and many zeros

Author: gdementen

doc/source/changes/version_0_30.rst.inc

@@ -157,4 +157,4 @@ Miscellaneous improvements
 Fixes
 -----
 
-* fixed something (closes :issue:`1`).
+* fixed LArray.divnot0 being slow when the divisor has many axes and many zeros (closes :issue:`705`).

larray/core/array.py

@@ -5420,10 +5420,16 @@ def divnot0(self, other):
             else:
                 return self / other
         else:
-            with np.errstate(divide='ignore', invalid='ignore'):
-                res = self / other
-            res[other == 0] = 0
-            return res
+            (self, other), res_axes = make_numpy_broadcastable((self, other))
+            otherdata = other.data
+            other_eq0 = otherdata == 0
+            # numpy array division gets slower the more zeros you have in other, so we change it before the division
+            # happens. This is obviously slower than doing nothing if we have very few zeros but I think it's a win
+            # on average given that other is likely to contain zeros when using divnot0.
+            otherdata = np.where(other_eq0, 1, otherdata)
+            res_data = self.data / otherdata
+            res_data[np.broadcast_to(other_eq0, res_data.shape)] = 0.0
+            return LArray(res_data, res_axes)
 
     # XXX: rename/change to "add_axes" ?
     # TODO: add a flag copy=True to force a new array.