New implementation of dpnp.outer (#1436)

* Add a new implementation of dpnp.outer

* Update dpnp.outer implementation

---------

Co-authored-by: Anton <[email protected]>

Author: vlad-perevezentsev

dpnp/dpnp_algo/dpnp_algo_linearalgebra.pxi

@@ -40,7 +40,6 @@ __all__ += [
     "dpnp_inner",
     "dpnp_kron",
     "dpnp_matmul",
-    "dpnp_outer"
 ]
 
 
@@ -378,27 +377,3 @@ cpdef utils.dpnp_descriptor dpnp_matmul(utils.dpnp_descriptor in_array1, utils.d
     c_dpctl.DPCTLEvent_Delete(event_ref)
 
     return result
-
-
-cpdef utils.dpnp_descriptor dpnp_outer(utils.dpnp_descriptor array1, utils.dpnp_descriptor array2):
-    cdef shape_type_c result_shape = (array1.size, array2.size)
-    result_type = numpy.promote_types(array1.dtype, array1.dtype)
-
-    result_sycl_device, result_usm_type, result_sycl_queue = utils.get_common_usm_allocation(array1, array2)
-
-    cdef utils.dpnp_descriptor result = utils_py.create_output_descriptor_py(result_shape,
-                                                                             result_type,
-                                                                             None,
-                                                                             device=result_sycl_device,
-                                                                             usm_type=result_usm_type,
-                                                                             sycl_queue=result_sycl_queue)
-
-    result_flatiter = result.get_pyobj().flat
-    array1_flatiter = array1.get_pyobj().flat
-    array2_flatiter = array2.get_pyobj().flat
-
-    for idx1 in range(array1.size):
-        for idx2 in range(array2.size):
-            result_flatiter[idx1 * array2.size + idx2] = array1_flatiter[idx1] * array2_flatiter[idx2]
-
-    return result

dpnp/dpnp_iface_linearalgebra.py

@@ -73,7 +73,7 @@ def dot(x1, x2, out=None, **kwargs):
     y : dpnp.ndarray
         Returns the dot product of `x1` and `x2`.
         If `out` is given, then it is returned.
-    
+
     Limitations
     -----------
     Parameters `x1` and `x2` are supported as either scalar, :class:`dpnp.ndarray`
@@ -298,16 +298,16 @@ def matmul(x1, x2, out=None, **kwargs):
     return call_origin(numpy.matmul, x1, x2, out=out, **kwargs)
 
 
-def outer(x1, x2, **kwargs):
+def outer(x1, x2, out=None):
     """
     Returns the outer product of two arrays.
 
     For full documentation refer to :obj:`numpy.outer`.
 
     Limitations
     -----------
-        Parameters ``x1`` and ``x2`` are supported as :obj:`dpnp.ndarray`.
-        Keyword arguments ``kwargs`` are currently unsupported.
+        Parameters `x1` and `x2` are supported as either scalar, :class:`dpnp.ndarray`
+        or :class:`dpctl.tensor.usm_ndarray`, but both `x1` and `x2` can not be scalars at the same time.
         Otherwise the functions will be executed sequentially on CPU.
         Input array data types are limited by supported DPNP :ref:`Data types`.
 
@@ -323,21 +323,26 @@ def outer(x1, x2, **kwargs):
     >>> b = np.array([1, 2, 3])
     >>> result = np.outer(a, b)
     >>> [x for x in result]
-    [1, 2, 3, 1, 2, 3, 1, 2, 3]
+    array([[1, 2, 3],
+           [1, 2, 3],
+           [1, 2, 3]])
 
     """
+    x1_is_scalar = dpnp.isscalar(x1)
+    x2_is_scalar = dpnp.isscalar(x2)
 
-    if not kwargs:
-        if isinstance(x1, dpnp_array) and isinstance(x2, dpnp_array):
-            ravel = lambda x: x.flatten() if x.ndim > 1 else x
-            return ravel(x1)[:, None] * ravel(x2)[None, :]
-
-        x1_desc = dpnp.get_dpnp_descriptor(x1, copy_when_nondefault_queue=False)
-        x2_desc = dpnp.get_dpnp_descriptor(x2, copy_when_nondefault_queue=False)
-        if x1_desc and x2_desc:
-            return dpnp_outer(x1_desc, x2_desc).get_pyobj()
+    if x1_is_scalar and x2_is_scalar:
+        pass
+    elif not (x1_is_scalar or dpnp.is_supported_array_type(x1)):
+        pass
+    elif not (x2_is_scalar or dpnp.is_supported_array_type(x2)):
+        pass
+    else:
+        x1_in = x1 if x1_is_scalar else (x1.reshape(-1) if x1.ndim > 1 else x1)[:, None]
+        x2_in = x2 if x2_is_scalar else (x2.reshape(-1) if x2.ndim > 1 else x2)[None, :]
+        return dpnp.multiply(x1_in, x2_in, out=out)
 
-    return call_origin(numpy.outer, x1, x2, **kwargs)
+    return call_origin(numpy.outer, x1, x2, out=out)
 
 
 def tensordot(x1, x2, axes=2):

tests/skipped_tests_gpu.tbl

@@ -210,7 +210,6 @@ tests/third_party/cupy/linalg_tests/test_einsum.py::TestEinSumError::test_too_ma
 tests/third_party/cupy/linalg_tests/test_einsum.py::TestListArgEinSumError::test_dim_mismatch3
 tests/third_party/cupy/linalg_tests/test_einsum.py::TestListArgEinSumError::test_too_many_dims3
 
-tests/third_party/cupy/linalg_tests/test_product.py::TestProduct::test_reversed_outer
 tests/third_party/cupy/linalg_tests/test_product.py::TestProduct::test_reversed_vdot
 tests/third_party/cupy/manipulation_tests/test_basic.py::TestCopytoFromScalar_param_7_{dst_shape=(0,), src=3.2}::test_copyto_where
 tests/third_party/cupy/manipulation_tests/test_basic.py::TestCopytoFromScalar_param_8_{dst_shape=(0,), src=0}::test_copyto_where

tests/test_outer.py

@@ -3,6 +3,7 @@
 
 import dpnp as dp
 import numpy as np
+import pytest
 
 from numpy.testing import assert_raises
 
@@ -40,23 +41,21 @@ def test_the_same_matrix(self, xp, dtype):
 
 class TestScalarOuter(unittest.TestCase):
 
-    @unittest.skip("A scalar isn't currently supported as input")
     @testing.for_all_dtypes()
-    @testing.numpy_cupy_allclose()
+    @testing.numpy_cupy_allclose(type_check=False)
     def test_first_is_scalar(self, xp, dtype):
         scalar = xp.int64(4)
         a = xp.arange(5**3, dtype=dtype).reshape(5, 5, 5)
         return xp.outer(scalar, a)
 
-    @unittest.skip("A scalar isn't currently supported as input")
     @testing.for_all_dtypes()
-    @testing.numpy_cupy_allclose()
+    @testing.numpy_cupy_allclose(type_check=False)
     def test_second_is_scalar(self, xp, dtype):
         scalar = xp.int32(7)
         a = xp.arange(5**3, dtype=dtype).reshape(5, 5, 5)
         return xp.outer(a, scalar)
 
-    @unittest.skip("A scalar isn't currently supported as input")
+    @pytest.mark.usefixtures("allow_fall_back_on_numpy")
     @testing.numpy_cupy_array_equal()
     def test_both_inputs_as_scalar(self, xp):
         a = xp.int64(4)