update docs about GPU arrays (#2704)

Author: david-cortes-intel

doc/sources/index.rst

@@ -102,17 +102,21 @@ Note: executing on GPU has `additional system software requirements <https://www
          .. tab:: With GPU arrays
             .. code-block:: python
 
+               import os
+               os.environ["SCIPY_ARRAY_API"] = "1"
                import numpy as np
                import dpnp
                from sklearnex import patch_sklearn
                patch_sklearn()
+               from sklearn import config_context
 
                from sklearn.cluster import DBSCAN
 
                X = np.array([[1., 2.], [2., 2.], [2., 3.],
                              [8., 7.], [8., 8.], [25., 80.]], dtype=np.float32)
                X = dpnp.array(X, device="gpu")
-               clustering = DBSCAN(eps=3, min_samples=2).fit(X)
+               with config_context(array_api_dispatch=True)
+                   clustering = DBSCAN(eps=3, min_samples=2).fit(X)
 
    .. tab:: Without patching
       .. tabs::
@@ -131,14 +135,18 @@ Note: executing on GPU has `additional system software requirements <https://www
          .. tab:: With GPU arrays
             .. code-block:: python
 
+               import os
+               os.environ["SCIPY_ARRAY_API"] = "1"
                import numpy as np
                import dpnp
+               from sklearnex import config_context
                from sklearnex.cluster import DBSCAN
 
                X = np.array([[1., 2.], [2., 2.], [2., 3.],
                              [8., 7.], [8., 8.], [25., 80.]], dtype=np.float32)
                X = dpnp.array(X, device="gpu")
-               clustering = DBSCAN(eps=3, min_samples=2).fit(X)
+               with config_context(array_api_dispatch=True)
+                   clustering = DBSCAN(eps=3, min_samples=2).fit(X)
 
 
 See :ref:`oneapi_gpu` for other ways of executing on GPU.

doc/sources/oneapi-gpu.rst

@@ -128,6 +128,12 @@ call :code:`sklearnex.get_config()`.
   located, and the result will be returned as :code:`usm_ndarray` to the same
   device.
 
+  .. important::
+    In order to enable zero-copy operations on GPU arrays, it's necessary to enable
+    :ref:`array API support <array_api>` for scikit-learn. Otherwise, if passing a GPU
+    array and array API support is not enabled, GPU arrays will first be transferred to
+    host and then back to GPU.
+
   .. note::
     All the input data for an algorithm must reside on the same device.
 
@@ -154,7 +160,7 @@ A full example of how to patch your code with Intel CPU/GPU optimizations:
    X = np.array([[1., 2.], [2., 2.], [2., 3.],
                  [8., 7.], [8., 8.], [25., 80.]], dtype=np.float32)
    with config_context(target_offload="gpu:0"):
-      clustering = DBSCAN(eps=3, min_samples=2).fit(X)
+       clustering = DBSCAN(eps=3, min_samples=2).fit(X)
 
 
 .. note:: Current offloading behavior restricts fitting and predictions (a.k.a. inference) of any models to be