Merge pull request #1778 from IntelPython/take-along-axis

Author: oleksandr-pavlyk

docs/doc_sources/api_reference/dpctl/tensor.indexing_functions.rst

@@ -15,3 +15,4 @@ by either integral arrays of indices or boolean mask arrays.
     place
     put
     take
+    take_along_axis

dpctl/tensor/__init__.py

@@ -60,7 +60,14 @@
 )
 from dpctl.tensor._device import Device
 from dpctl.tensor._dlpack import from_dlpack
-from dpctl.tensor._indexing_functions import extract, nonzero, place, put, take
+from dpctl.tensor._indexing_functions import (
+    extract,
+    nonzero,
+    place,
+    put,
+    take,
+    take_along_axis,
+)
 from dpctl.tensor._linear_algebra_functions import (
     matmul,
     matrix_transpose,
@@ -376,4 +383,5 @@
     "nextafter",
     "diff",
     "count_nonzero",
+    "take_along_axis",
 ]

dpctl/tensor/_copy_utils.py

@@ -795,13 +795,18 @@ def _nonzero_impl(ary):
     return res
 
 
-def _take_multi_index(ary, inds, p):
+def _take_multi_index(ary, inds, p, mode=0):
     if not isinstance(ary, dpt.usm_ndarray):
         raise TypeError(
             f"Expecting type dpctl.tensor.usm_ndarray, got {type(ary)}"
         )
     ary_nd = ary.ndim
     p = normalize_axis_index(operator.index(p), ary_nd)
+    mode = operator.index(mode)
+    if mode not in [0, 1]:
+        raise ValueError(
+            "Invalid value for mode keyword, only 0 or 1 is supported"
+        )
     queues_ = [
         ary.sycl_queue,
     ]
@@ -860,7 +865,7 @@ def _take_multi_index(ary, inds, p):
         ind=inds,
         dst=res,
         axis_start=p,
-        mode=0,
+        mode=mode,
         sycl_queue=exec_q,
         depends=dep_ev,
     )

dpctl/tensor/_indexing_functions.py

@@ -21,7 +21,7 @@
 import dpctl.tensor._tensor_impl as ti
 import dpctl.utils
 
-from ._copy_utils import _extract_impl, _nonzero_impl
+from ._copy_utils import _extract_impl, _nonzero_impl, _take_multi_index
 from ._numpy_helper import normalize_axis_index
 
 
@@ -423,3 +423,82 @@ def nonzero(arr):
     if arr.ndim == 0:
         raise ValueError("Array of positive rank is expected")
     return _nonzero_impl(arr)
+
+
+def _range(sh_i, i, nd, q, usm_t, dt):
+    ind = dpt.arange(sh_i, dtype=dt, usm_type=usm_t, sycl_queue=q)
+    ind.shape = tuple(sh_i if i == j else 1 for j in range(nd))
+    return ind
+
+
+def take_along_axis(x, indices, /, *, axis=-1, mode="wrap"):
+    """
+    Returns elements from an array at the one-dimensional indices specified
+    by ``indices`` along a provided ``axis``.
+
+    Args:
+        x (usm_ndarray):
+            input array. Must be compatible with ``indices``, except for the
+            axis (dimension) specified by ``axis``.
+        indices (usm_ndarray):
+            array indices. Must have the same rank (i.e., number of dimensions)
+            as ``x``.
+        axis: int
+            axis along which to select values. If ``axis`` is negative, the
+            function determines the axis along which to select values by
+            counting from the last dimension. Default: ``-1``.
+        mode (str, optional):
+            How out-of-bounds indices will be handled. Possible values
+            are:
+
+            - ``"wrap"``: clamps indices to (``-n <= i < n``), then wraps
+              negative indices.
+            - ``"clip"``: clips indices to (``0 <= i < n``).
+
+            Default: ``"wrap"``.
+
+    Returns:
+        usm_ndarray:
+            an array having the same data type as ``x``. The returned array has
+            the same rank (i.e., number of dimensions) as ``x`` and a shape
+            determined according to :ref:`broadcasting`, except for the axis
+            (dimension) specified by ``axis`` whose size must equal the size
+            of the corresponding axis (dimension) in ``indices``.
+
+    Note:
+        Treatment of the out-of-bound indices in ``indices`` array is controlled
+        by the value of ``mode`` keyword.
+    """
+    if not isinstance(x, dpt.usm_ndarray):
+        raise TypeError(f"Expected dpctl.tensor.usm_ndarray, got {type(x)}")
+    if not isinstance(indices, dpt.usm_ndarray):
+        raise TypeError(
+            f"Expected dpctl.tensor.usm_ndarray, got {type(indices)}"
+        )
+    x_nd = x.ndim
+    if x_nd != indices.ndim:
+        raise ValueError(
+            "Number of dimensions in the first and the second "
+            "argument arrays must be equal"
+        )
+    pp = normalize_axis_index(operator.index(axis), x_nd)
+    out_usm_type = dpctl.utils.get_coerced_usm_type(
+        (x.usm_type, indices.usm_type)
+    )
+    exec_q = dpctl.utils.get_execution_queue((x.sycl_queue, indices.sycl_queue))
+    if exec_q is None:
+        raise dpctl.utils.ExecutionPlacementError(
+            "Execution placement can not be unambiguously inferred "
+            "from input arguments. "
+        )
+    mode_i = _get_indexing_mode(mode)
+    indexes_dt = ti.default_device_index_type(exec_q.sycl_device)
+    _ind = tuple(
+        (
+            indices
+            if i == pp
+            else _range(x.shape[i], i, x_nd, exec_q, out_usm_type, indexes_dt)
+        )
+        for i in range(x_nd)
+    )
+    return _take_multi_index(x, _ind, 0, mode=mode_i)

dpctl/tests/test_usm_ndarray_indexing.py

@@ -1535,5 +1535,151 @@ def test_advanced_integer_indexing_cast_indices():
     inds1 = dpt.astype(inds0, "u4")
     inds2 = dpt.astype(inds0, "u8")
     x = dpt.ones((3, 4, 5, 6), dtype="i4")
+    # test getitem
     with pytest.raises(ValueError):
         x[inds0, inds1, inds2, ...]
+    # test setitem
+    with pytest.raises(ValueError):
+        x[inds0, inds1, inds2, ...] = 1
+
+
+def test_take_along_axis():
+    get_queue_or_skip()
+
+    n0, n1, n2 = 3, 5, 7
+    x = dpt.reshape(dpt.arange(n0 * n1 * n2), (n0, n1, n2))
+    ind_dt = dpt.__array_namespace_info__().default_dtypes(
+        device=x.sycl_device
+    )["indexing"]
+    ind0 = dpt.ones((1, n1, n2), dtype=ind_dt)
+    ind1 = dpt.ones((n0, 1, n2), dtype=ind_dt)
+    ind2 = dpt.ones((n0, n1, 1), dtype=ind_dt)
+
+    y0 = dpt.take_along_axis(x, ind0, axis=0)
+    assert y0.shape == ind0.shape
+    y1 = dpt.take_along_axis(x, ind1, axis=1)
+    assert y1.shape == ind1.shape
+    y2 = dpt.take_along_axis(x, ind2, axis=2)
+    assert y2.shape == ind2.shape
+
+
+def test_take_along_axis_validation():
+    # type check on the first argument
+    with pytest.raises(TypeError):
+        dpt.take_along_axis(tuple(), list())
+    get_queue_or_skip()
+    n1, n2 = 2, 5
+    x = dpt.ones(n1 * n2)
+    # type check on the second argument
+    with pytest.raises(TypeError):
+        dpt.take_along_axis(x, list())
+    x_dev = x.sycl_device
+    info_ = dpt.__array_namespace_info__()
+    def_dtypes = info_.default_dtypes(device=x_dev)
+    ind_dt = def_dtypes["indexing"]
+    ind = dpt.zeros(1, dtype=ind_dt)
+    # axis valudation
+    with pytest.raises(ValueError):
+        dpt.take_along_axis(x, ind, axis=1)
+    # mode validation
+    with pytest.raises(ValueError):
+        dpt.take_along_axis(x, ind, axis=0, mode="invalid")
+    # same array-ranks validation
+    with pytest.raises(ValueError):
+        dpt.take_along_axis(dpt.reshape(x, (n1, n2)), ind)
+    # check compute-follows-data
+    q2 = dpctl.SyclQueue(x_dev, property="enable_profiling")
+    ind2 = dpt.zeros(1, dtype=ind_dt, sycl_queue=q2)
+    with pytest.raises(ExecutionPlacementError):
+        dpt.take_along_axis(x, ind2)
+
+
+def check__extract_impl_validation(fn):
+    x = dpt.ones(10)
+    ind = dpt.ones(10, dtype="?")
+    with pytest.raises(TypeError):
+        fn(list(), ind)
+    with pytest.raises(TypeError):
+        fn(x, list())
+    q2 = dpctl.SyclQueue(x.sycl_device, property="enable_profiling")
+    ind2 = dpt.ones(10, dtype="?", sycl_queue=q2)
+    with pytest.raises(ExecutionPlacementError):
+        fn(x, ind2)
+    with pytest.raises(ValueError):
+        fn(x, ind, 1)
+
+
+def check__nonzero_impl_validation(fn):
+    with pytest.raises(TypeError):
+        fn(list())
+
+
+def check__take_multi_index(fn):
+    x = dpt.ones(10)
+    x_dev = x.sycl_device
+    info_ = dpt.__array_namespace_info__()
+    def_dtypes = info_.default_dtypes(device=x_dev)
+    ind_dt = def_dtypes["indexing"]
+    ind = dpt.arange(10, dtype=ind_dt)
+    with pytest.raises(TypeError):
+        fn(list(), tuple(), 1)
+    with pytest.raises(ValueError):
+        fn(x, (ind,), 0, mode=2)
+    with pytest.raises(ValueError):
+        fn(x, (None,), 1)
+    with pytest.raises(IndexError):
+        fn(x, (x,), 1)
+    q2 = dpctl.SyclQueue(x.sycl_device, property="enable_profiling")
+    ind2 = dpt.arange(10, dtype=ind_dt, sycl_queue=q2)
+    with pytest.raises(ExecutionPlacementError):
+        fn(x, (ind2,), 0)
+    m = dpt.ones((10, 10))
+    ind_1 = dpt.arange(10, dtype="i8")
+    ind_2 = dpt.arange(10, dtype="u8")
+    with pytest.raises(ValueError):
+        fn(m, (ind_1, ind_2), 0)
+
+
+def check__place_impl_validation(fn):
+    with pytest.raises(TypeError):
+        fn(list(), list(), list())
+    x = dpt.ones(10)
+    with pytest.raises(TypeError):
+        fn(x, list(), list())
+    q2 = dpctl.SyclQueue(x.sycl_device, property="enable_profiling")
+    mask2 = dpt.ones(10, dtype="?", sycl_queue=q2)
+    with pytest.raises(ExecutionPlacementError):
+        fn(x, mask2, 1)
+    x2 = dpt.ones((5, 5))
+    mask2 = dpt.ones((5, 5), dtype="?")
+    with pytest.raises(ValueError):
+        fn(x2, mask2, x2, axis=1)
+
+
+def check__put_multi_index_validation(fn):
+    with pytest.raises(TypeError):
+        fn(list(), list(), 0, list())
+    x = dpt.ones(10)
+    inds = dpt.arange(10, dtype="i8")
+    vals = dpt.zeros(10)
+    # test inds which is not a tuple/list
+    fn(x, inds, 0, vals)
+    x2 = dpt.ones((5, 5))
+    ind1 = dpt.arange(5, dtype="i8")
+    ind2 = dpt.arange(5, dtype="u8")
+    with pytest.raises(ValueError):
+        fn(x2, (ind1, ind2), 0, x2)
+    with pytest.raises(TypeError):
+        fn(x2, (ind1, list()), 0, x2)
+
+
+def test__copy_utils():
+    import dpctl.tensor._copy_utils as cu
+
+    get_queue_or_skip()
+
+    check__extract_impl_validation(cu._extract_impl)
+    check__nonzero_impl_validation(cu._nonzero_impl)
+    check__take_multi_index(cu._take_multi_index)
+    check__place_impl_validation(cu._place_impl)
+    check__put_multi_index_validation(cu._put_multi_index)

dpctl/tests/test_usm_ndarray_sorting.py

@@ -177,12 +177,24 @@ def test_argsort_axis0():
     x = dpt.reshape(xf, (n, m))
     idx = dpt.argsort(x, axis=0)
 
-    conseq_idx = dpt.arange(m, dtype=idx.dtype)
-    s = x[idx, conseq_idx[dpt.newaxis, :]]
+    s = dpt.take_along_axis(x, idx, axis=0)
 
     assert dpt.all(s[:-1, :] <= s[1:, :])
 
 
+def test_argsort_axis1():
+    get_queue_or_skip()
+
+    n, m = 200, 30
+    xf = dpt.arange(n * m, 0, step=-1, dtype="i4")
+    x = dpt.reshape(xf, (n, m))
+    idx = dpt.argsort(x, axis=1)
+
+    s = dpt.take_along_axis(x, idx, axis=1)
+
+    assert dpt.all(s[:, :-1] <= s[:, 1:])
+
+
 def test_sort_strided():
     get_queue_or_skip()
 
@@ -199,8 +211,9 @@ def test_argsort_strided():
     x_orig = dpt.arange(100, dtype="i4")
     x_flipped = dpt.flip(x_orig, axis=0)
     idx = dpt.argsort(x_flipped)
+    s = dpt.take_along_axis(x_flipped, idx, axis=0)
 
-    assert dpt.all(x_flipped[idx] == x_orig)
+    assert dpt.all(s == x_orig)
 
 
 def test_sort_0d_array():