Optimize downsample1d when data is shared between layers

holoviews/operation/downsample.py

@@ -232,18 +232,27 @@ class downsample1d(ResampleOperation1D):
         Whether to add the neighbor points to the range before downsampling.
         By default this is only enabled for the viewport algorithm.""")
 
-    def _process(self, element, key=None):
+    def _process(self, element, key=None, shared_data=None):
         if isinstance(element, (Overlay, NdOverlay)):
-            _process = partial(self._process, key=key)
+            # Shared data is so we only slice the given data once
+            kwargs = {'key': key, 'shared_data': {}}
+            _process = partial(self._process, **kwargs)
             if isinstance(element, Overlay):
                 elements = [v.map(_process) for v in element]
             else:
                 elements = {k: v.map(_process) for k, v in element.items()}
             return element.clone(elements)
 
         if self.p.x_range:
-            mask = self._compute_mask(element)
-            element = element[mask]
+            key = (id(element.data), str(element.kdims[0]))
+            if shared_data is not None and key in shared_data:
+                element = element.clone(shared_data[key])
+            else:
+                mask = self._compute_mask(element)
+                element = element[mask]
+                if shared_data is not None:
+                    shared_data[key] = element.data
+
         if len(element) <= self.p.width:
             return element
         xs, ys = (element.dimension_values(i) for i in range(2))

holoviews/tests/operation/test_downsample.py

@@ -1,4 +1,5 @@
 import numpy as np
+import pandas as pd
 import pytest
 
 import holoviews as hv
@@ -10,7 +11,8 @@
     tsdownsample = None
 
 algorithms = _ALGORITHMS.copy()
-algorithms.pop('viewport', None)  # viewport return slice(len(data)) no matter the width
+algorithms.pop("viewport", None)  # viewport return slice(len(data)) no matter the width
+
 
 @pytest.mark.parametrize("plottype", ["overlay", "ndoverlay"])
 def test_downsample1d_multi(plottype):
@@ -28,6 +30,46 @@ def test_downsample1d_multi(plottype):
             assert value.size == downsample1d.width
 
 
+def test_downsample1d_shared_data():
+    runs = [0]
+
+    class mocksample(downsample1d):
+        def _compute_mask(self, element):
+            # Use _compute_mask as this should only be called once
+            # and then it should be cloned.
+            runs[0] += 1
+            return super()._compute_mask(element)
+
+    N = 1000
+    df = pd.DataFrame({c: range(N) for c in "xyz"})
+    figure = hv.Overlay([hv.Curve(df, kdims="x", vdims=c) for c in "yz"])
+
+    # We set x_range to trigger _compute_mask
+    mocksample(figure, dynamic=False, x_range=(0, 500))
+    assert runs[0] == 1
+
+
+# Should be fixed when https://github.com/holoviz/holoviews/pull/6061 is merged
+@pytest.mark.xfail(reason="This will make a copy of the data")
+def test_downsample1d_shared_data_index():
+    runs = [0]
+
+    class mocksample(downsample1d):
+        def _compute_mask(self, element):
+            # Use _compute_mask as this should only be called once
+            # and then it should be cloned.
+            runs[0] += 1
+            return super()._compute_mask(element)
+
+    N = 1000
+    df = pd.DataFrame({c: range(N) for c in "xyz"})
+    figure = hv.Overlay([hv.Curve(df, kdims="index", vdims=c) for c in "xyz"])
+
+    # We set x_range to trigger _compute_mask
+    mocksample(figure, dynamic=False, x_range=(0, 500))
+    assert runs[0] == 1
+
+
 @pytest.mark.parametrize("algorithm", algorithms.values(), ids=algorithms)
 def test_downsample_algorithm(algorithm, unimport):
     unimport("tsdownsample")