[enhancement] refactor onedal to_table (#2151)

* carryover from #2126

* formatting

* Update test_memory_usage.py

* Update _data_conversion.py

* Update _data_conversion.py

* Update _data_conversion.py

* Update test_data.py

* Update onedal/datatypes/_data_conversion.py

Co-authored-by: Victoriya Fedotova <viktoria.nn@gmail.com>

* Update _data_conversion.py

* Update _data_conversion.py

* Update onedal/datatypes/_data_conversion.py

Co-authored-by: Samir Nasibli <samir.nasibli@intel.com>

* Update _data_conversion.py

* Update _data_conversion.py

---------

Co-authored-by: Victoriya Fedotova <viktoria.nn@gmail.com>
Co-authored-by: Samir Nasibli <samir.nasibli@intel.com>

onedal/datatypes/_data_conversion.py

@@ -18,28 +18,67 @@
 
 import numpy as np
 
-from daal4py.sklearn._utils import make2d
 from onedal import _backend, _is_dpc_backend
 
-from ..utils import _is_csr
-
 
 def _apply_and_pass(func, *args, **kwargs):
     if len(args) == 1:
         return func(args[0], **kwargs)
     return tuple(map(lambda arg: func(arg, **kwargs), args))
 
 
-if _is_dpc_backend:
+def _convert_one_to_table(arg):
+    # All inputs for table conversion must be array-like or sparse, not scalars
+    return _backend.to_table(np.atleast_2d(arg) if np.isscalar(arg) else arg)
+
+
+def to_table(*args):
+    """Create oneDAL tables from scalars and/or arrays.
+
+    Note: this implementation can be used with contiguous scipy.sparse, numpy
+    ndarrays, DPCTL/DPNP usm_ndarrays and scalars. Tables will use pointers to the
+    original array data. Scalars will be copies. Arrays may be modified in-
+    place by oneDAL during computation. This works for data located on CPU and
+    SYCL-enabled Intel GPUs. Each array may only be of a single datatype (i.e.
+    each must be homogeneous).
+
+    Parameters
+    ----------
+    *args : {scalar, numpy array, sycl_usm_ndarray, csr_matrix, or csr_array}
+        arg1, arg2... The arrays should be given as arguments.
 
-    from ..utils._dpep_helpers import dpctl_available, dpnp_available
+    Returns
+    -------
+    tables: {oneDAL homogeneous tables}
+    """
+    return _apply_and_pass(_convert_one_to_table, *args)
 
-    if dpctl_available:
-        import dpctl.tensor as dpt
 
-    if dpnp_available:
+if _is_dpc_backend:
+
+    try:
+        # try/catch is used here instead of dpep_helpers because
+        # of circular import issues of _data_conversion.py and
+        # utils/validation.py. This is a temporary fix until the
+        # issue with dpnp is addressed, at which point this can
+        # be removed entirely.
         import dpnp
 
+        def _table_to_array(table, xp=None):
+            # By default DPNP ndarray created with a copy.
+            # TODO:
+            # investigate why dpnp.array(table, copy=False) doesn't work.
+            # Work around with using dpctl.tensor.asarray.
+            if xp == dpnp:
+                return dpnp.array(dpnp.dpctl.tensor.asarray(table), copy=False)
+            else:
+                return xp.asarray(table)
+
+    except ImportError:
+
+        def _table_to_array(table, xp=None):
+            return xp.asarray(table)
+
     from ..common._policy import _HostInteropPolicy
 
     def _convert_to_supported(policy, *data):
@@ -87,26 +126,9 @@ def convert_one_from_table(table, sycl_queue=None, sua_iface=None, xp=None):
                     _backend.from_table(table), usm_type="device", sycl_queue=sycl_queue
                 )
             else:
-                xp_name = xp.__name__
-                if dpnp_available and xp_name == "dpnp":
-                    # By default DPNP ndarray created with a copy.
-                    # TODO:
-                    # investigate why dpnp.array(table, copy=False) doesn't work.
-                    # Work around with using dpctl.tensor.asarray.
-                    return dpnp.array(dpt.asarray(table), copy=False)
-                else:
-                    return xp.asarray(table)
-        return _backend.from_table(table)
-
-    def convert_one_to_table(arg, sua_iface=None):
-        # Note: currently only oneDAL homogen tables are supported and the
-        # contiuginity of the input array should be checked in advance.
-        if sua_iface:
-            return _backend.sua_iface_to_table(arg)
+                return _table_to_array(table, xp=xp)
 
-        if not _is_csr(arg):
-            arg = make2d(arg)
-        return _backend.to_table(arg)
+        return _backend.from_table(table)
 
 else:
 
@@ -125,22 +147,8 @@ def convert_one_from_table(table, sycl_queue=None, sua_iface=None, xp=None):
             )
         return _backend.from_table(table)
 
-    def convert_one_to_table(arg, sua_iface=None):
-        if sua_iface:
-            raise RuntimeError(
-                "SYCL usm array conversion to table requires the DPC backend"
-            )
-
-        if not _is_csr(arg):
-            arg = make2d(arg)
-        return _backend.to_table(arg)
-
 
 def from_table(*args, sycl_queue=None, sua_iface=None, xp=None):
     return _apply_and_pass(
         convert_one_from_table, *args, sycl_queue=sycl_queue, sua_iface=sua_iface, xp=xp
     )
-
-
-def to_table(*args, sua_iface=None):
-    return _apply_and_pass(convert_one_to_table, *args, sua_iface=sua_iface)

onedal/datatypes/table.cpp

@@ -78,6 +78,12 @@ ONEDAL_PY_INIT_MODULE(table) {
 #endif // ONEDAL_DATA_PARALLEL
 
     m.def("to_table", [](py::object obj) {
+        #ifdef ONEDAL_DATA_PARALLEL
+        if (py::hasattr(obj, "__sycl_usm_array_interface__")) {
+            return convert_from_sua_iface(obj);
+        }
+        #endif // ONEDAL_DATA_PARALLEL
+
         auto* obj_ptr = obj.ptr();
         return convert_to_table(obj_ptr);
     });
@@ -86,12 +92,6 @@ ONEDAL_PY_INIT_MODULE(table) {
         auto* obj_ptr = convert_to_pyobject(t);
         return obj_ptr;
     });
-
-#ifdef ONEDAL_DATA_PARALLEL
-    m.def("sua_iface_to_table", [](py::object obj) {
-        return convert_from_sua_iface(obj);
-    });
-#endif // ONEDAL_DATA_PARALLEL
 }
 
 } // namespace oneapi::dal::python

onedal/datatypes/tests/test_data.py

@@ -68,7 +68,7 @@ def fit(self, X, y=None):
                 X = xp.astype(X, dtype=xp.float64)
             dtype = get_dtype(X)
             params = bs_DBSCAN._get_onedal_params(dtype)
-            X_table = to_table(X, sua_iface=sua_iface)
+            X_table = to_table(X)
             # TODO:
             # check other candidates for the dummy base oneDAL func.
             # oneDAL backend func is needed to check result table checks.
@@ -251,7 +251,7 @@ def test_input_sua_iface_zero_copy(dataframe, queue, order, dtype):
 
     sua_iface, X_dp_namespace, _ = _get_sycl_namespace(X_dp)
 
-    X_table = to_table(X_dp, sua_iface=sua_iface)
+    X_table = to_table(X_dp)
     _assert_sua_iface_fields(X_dp, X_table)
 
     X_dp_from_table = from_table(
@@ -339,7 +339,7 @@ def test_sua_iface_interop_invalid_shape(dataframe, queue, data_shape):
         "Unable to convert from SUA interface: only 1D & 2D tensors are allowed"
     )
     with pytest.raises(ValueError, match=expected_err_msg):
-        to_table(X, sua_iface=sua_iface)
+        to_table(X)
 
 
 @pytest.mark.skipif(
@@ -368,7 +368,7 @@ def test_sua_iface_interop_unsupported_dtypes(dataframe, queue, dtype):
 
     expected_err_msg = "Unable to convert from SUA interface: unknown data type"
     with pytest.raises(ValueError, match=expected_err_msg):
-        to_table(X, sua_iface=sua_iface)
+        to_table(X)
 
 
 @pytest.mark.parametrize(
@@ -393,7 +393,7 @@ def test_to_table_non_contiguous_input(dataframe, queue):
     else:
         expected_err_msg = "Numpy input Could not convert Python object to onedal table."
     with pytest.raises(ValueError, match=expected_err_msg):
-        to_table(X, sua_iface=sua_iface)
+        to_table(X)
 
 
 @pytest.mark.skipif(
@@ -411,4 +411,4 @@ def test_sua_iface_interop_if_no_dpc_backend(dataframe, queue, dtype):
 
     expected_err_msg = "SYCL usm array conversion to table requires the DPC backend"
     with pytest.raises(RuntimeError, match=expected_err_msg):
-        to_table(X, sua_iface=sua_iface)
+        to_table(X)

sklearnex/tests/test_memory_usage.py

@@ -142,8 +142,8 @@ class DummyEstimatorWithTableConversions(BaseEstimator):
 
         def fit(self, X, y=None):
             sua_iface, xp, _ = _get_sycl_namespace(X)
-            X_table = to_table(X, sua_iface=sua_iface)
-            y_table = to_table(y, sua_iface=sua_iface)
+            X_table = to_table(X)
+            y_table = to_table(y)
             # The presence of the fitted attributes (ending with a trailing
             # underscore) is required for the correct check. The cleanup of
             # the memory will occur at the estimator instance deletion.
@@ -160,7 +160,7 @@ def predict(self, X):
             # fitted attributes (ending with a trailing underscore).
             check_is_fitted(self)
             sua_iface, xp, _ = _get_sycl_namespace(X)
-            X_table = to_table(X, sua_iface=sua_iface)
+            X_table = to_table(X)
             returned_X = from_table(
                 X_table, sua_iface=sua_iface, sycl_queue=X.sycl_queue, xp=xp
             )