Merge branch 'master' into update_tests

Author: vtavana

.github/workflows/conda-package.yml

@@ -13,7 +13,7 @@ env:
   # Follow oneAPI installation instruction for conda, since intel channel is not longer available
   # CHANNELS: '-c dppy/label/dev -c intel -c conda-forge --override-channels'
   CHANNELS: '-c dppy/label/dev -c https://software.repos.intel.com/python/conda/ -c conda-forge --override-channels'
-  CONDA_BUILD_VERSION: '24.11.1'
+  CONDA_BUILD_VERSION: '24.11.2'
   CONDA_INDEX_VERSION: '0.5.0'
   RERUN_TESTS_ON_FAILURE: 'true'
   RUN_TESTS_MAX_ATTEMPTS: 2

dpnp/fft/dpnp_utils_fft.py

@@ -78,11 +78,13 @@ def _commit_descriptor(a, forward, in_place, c2c, a_strides, index, batch_fft):
     shape = a_shape[index:]
     strides = (0,) + a_strides[index:]
     if c2c:  # c2c FFT
+        assert dpnp.issubdtype(a.dtype, dpnp.complexfloating)
         if a.dtype == dpnp.complex64:
             dsc = fi.Complex64Descriptor(shape)
         else:
             dsc = fi.Complex128Descriptor(shape)
     else:  # r2c/c2r FFT
+        assert dpnp.issubdtype(a.dtype, dpnp.inexact)
         if a.dtype in [dpnp.float32, dpnp.complex64]:
             dsc = fi.Real32Descriptor(shape)
         else:
@@ -262,12 +264,14 @@ def _copy_array(x, complex_input):
     in-place FFT can be performed.
     """
     dtype = x.dtype
+    copy_flag = False
     if numpy.min(x.strides) < 0:
         # negative stride is not allowed in OneMKL FFT
         # TODO: support for negative strides will be added in the future
         # versions of OneMKL, see discussion in MKLD-17597
         copy_flag = True
-    elif complex_input and not dpnp.issubdtype(dtype, dpnp.complexfloating):
+
+    if complex_input and not dpnp.issubdtype(dtype, dpnp.complexfloating):
         # c2c/c2r FFT, if input is not complex, convert to complex
         copy_flag = True
         if dtype in [dpnp.float16, dpnp.float32]:
@@ -279,8 +283,6 @@ def _copy_array(x, complex_input):
         # float32 or float64 depending on device capabilities
         copy_flag = True
         dtype = map_dtype_to_device(dpnp.float64, x.sycl_device)
-    else:
-        copy_flag = False
 
     if copy_flag:
         x_copy = dpnp.empty_like(x, dtype=dtype, order="C")

dpnp/linalg/dpnp_iface_linalg.py

@@ -1612,18 +1612,24 @@ def solve(a, b):
     ----------
     a : (..., M, M) {dpnp.ndarray, usm_ndarray}
         Coefficient matrix.
-    b : {(…, M,), (…, M, K)} {dpnp.ndarray, usm_ndarray}
+    b : {(M,), (..., M, K)} {dpnp.ndarray, usm_ndarray}
         Ordinate or "dependent variable" values.
 
     Returns
     -------
-    out : {(…, M,), (…, M, K)} dpnp.ndarray
+    out : {(..., M,), (..., M, K)} dpnp.ndarray
         Solution to the system `ax = b`. Returned shape is identical to `b`.
 
     See Also
     --------
     :obj:`dpnp.dot` : Returns the dot product of two arrays.
 
+    Notes
+    -----
+    The `b` array is only treated as a shape (M,) column vector if it is
+    exactly 1-dimensional. In all other instances it is treated as a stack
+    of (M, K) matrices.
+
     Examples
     --------
     >>> import dpnp as dp
@@ -1644,14 +1650,38 @@ def solve(a, b):
     assert_stacked_2d(a)
     assert_stacked_square(a)
 
-    if not (
-        a.ndim in [b.ndim, b.ndim + 1] and a.shape[:-1] == b.shape[: a.ndim - 1]
-    ):
-        raise dpnp.linalg.LinAlgError(
-            "a must have (..., M, M) shape and b must have (..., M) "
-            "or (..., M, K)"
+    a_shape = a.shape
+    b_shape = b.shape
+    b_ndim = b.ndim
+
+    # compatible with numpy>=2.0
+    if b_ndim == 0:
+        raise ValueError("b must have at least one dimension")
+    if b_ndim == 1:
+        if a_shape[-1] != b.size:
+            raise ValueError(
+                "a must have (..., M, M) shape and b must have (M,) "
+                "for one-dimensional b"
+            )
+        b = dpnp.broadcast_to(b, a_shape[:-1])
+        return dpnp_solve(a, b)
+
+    if a_shape[-1] != b_shape[-2]:
+        raise ValueError(
+            "a must have (..., M, M) shape and b must have (..., M, K) shape"
         )
 
+    # Use dpnp.broadcast_shapes() to align the resulting batch shapes
+    broadcasted_batch_shape = dpnp.broadcast_shapes(a_shape[:-2], b_shape[:-2])
+
+    a_broadcasted_shape = broadcasted_batch_shape + a_shape[-2:]
+    b_broadcasted_shape = broadcasted_batch_shape + b_shape[-2:]
+
+    if a_shape != a_broadcasted_shape:
+        a = dpnp.broadcast_to(a, a_broadcasted_shape)
+    if b_shape != b_broadcasted_shape:
+        b = dpnp.broadcast_to(b, b_broadcasted_shape)
+
     return dpnp_solve(a, b)
 
 

dpnp/tests/test_fft.py

@@ -378,6 +378,16 @@ def test_fft_validate_out(self):
         out = dpnp.empty((10,), dtype=dpnp.float32)
         assert_raises(TypeError, dpnp.fft.fft, a, out=out)
 
+    @pytest.mark.parametrize(
+        "dtype", get_all_dtypes(no_none=True, no_bool=True)
+    )
+    def test_negative_stride(self, dtype):
+        a = dpnp.arange(10, dtype=dtype)
+        result = dpnp.fft.fft(a[::-1])
+        expected = numpy.fft.fft(a.asnumpy()[::-1])
+
+        assert_dtype_allclose(result, expected, check_only_type_kind=True)
+
 
 class TestFft2:
     def setup_method(self):

dpnp/tests/test_histogram.py

@@ -113,6 +113,7 @@ def test_digitize_error(self):
         bins_np = dpnp.asnumpy(bins_dp)
         with pytest.raises(TypeError):
             dpnp.digitize(x_np, bins_dp)
+        with pytest.raises(TypeError):
             dpnp.digitize(x_dp, bins_np)
 
         # bins ndim < 1

dpnp/tests/test_linalg.py

@@ -2502,6 +2502,36 @@ def test_solve(self, dtype):
 
         assert_allclose(expected, result, rtol=1e-06)
 
+    @testing.with_requires("numpy>=2.0")
+    @pytest.mark.parametrize("dtype", get_float_complex_dtypes())
+    @pytest.mark.parametrize(
+        "a_shape, b_shape",
+        [
+            ((4, 4), (2, 2, 4, 3)),
+            ((2, 5, 5), (1, 5, 3)),
+            ((2, 4, 4), (2, 2, 4, 2)),
+            ((3, 2, 2), (3, 1, 2, 1)),
+            ((2, 2, 2, 2, 2), (2,)),
+            ((2, 2, 2, 2, 2), (2, 3)),
+        ],
+    )
+    def test_solve_broadcast(self, a_shape, b_shape, dtype):
+        # Set seed_value=81 to prevent
+        # random generation of the input singular matrix
+        a_np = generate_random_numpy_array(a_shape, dtype, seed_value=81)
+
+        # Set seed_value=76 to prevent
+        # random generation of the input singular matrix
+        b_np = generate_random_numpy_array(b_shape, dtype, seed_value=76)
+
+        a_dp = inp.array(a_np)
+        b_dp = inp.array(b_np)
+
+        expected = numpy.linalg.solve(a_np, b_np)
+        result = inp.linalg.solve(a_dp, b_dp)
+
+        assert_dtype_allclose(result, expected)
+
     @pytest.mark.parametrize("dtype", get_all_dtypes(no_bool=True))
     def test_solve_nrhs_greater_n(self, dtype):
         # Test checking the case when nrhs > n for
@@ -2607,6 +2637,10 @@ def test_solve_errors(self):
             dpnp.linalg.LinAlgError, dpnp.linalg.solve, a_dp_ndim_1, b_dp
         )
 
+        # b.ndim == 0
+        b_dp_ndim_0 = inp.array(2)
+        assert_raises(ValueError, inp.linalg.solve, a_dp, b_dp_ndim_0)
+
 
 class TestSlogdet:
     @pytest.mark.parametrize("dtype", get_all_dtypes(no_bool=True))

dpnp/tests/test_sycl_queue.py

@@ -2392,40 +2392,34 @@ def test_where(device):
     ids=[device.filter_string for device in valid_devices],
 )
 @pytest.mark.parametrize(
-    "matrix, vector",
+    "matrix, rhs",
     [
         ([[1, 2], [3, 5]], numpy.empty((2, 0))),
         ([[1, 2], [3, 5]], [1, 2]),
         (
             [
-                [[1, 1, 1], [0, 2, 5], [2, 5, -1]],
-                [[3, -1, 1], [1, 2, 3], [2, 3, 1]],
-                [[1, 4, 1], [1, 2, -2], [4, 1, 2]],
+                [[1, 1], [0, 2]],
+                [[3, -1], [1, 2]],
+            ],
+            [
+                [[6, -4], [9, -6]],
+                [[15, 1], [15, 1]],
             ],
-            [[6, -4, 27], [9, -6, 15], [15, 1, 11]],
         ),
     ],
     ids=[
-        "2D_Matrix_Empty_Vector",
-        "2D_Matrix_1D_Vector",
-        "3D_Matrix_and_Vectors",
+        "2D_Matrix_Empty_RHS",
+        "2D_Matrix_1D_RHS",
+        "3D_Matrix_and_3D_RHS",
     ],
 )
-def test_solve(matrix, vector, device):
+def test_solve(matrix, rhs, device):
     a_np = numpy.array(matrix)
-    b_np = numpy.array(vector)
+    b_np = numpy.array(rhs)
 
     a_dp = dpnp.array(a_np, device=device)
     b_dp = dpnp.array(b_np, device=device)
 
-    # In numpy 2.0 the broadcast ambiguity has been removed and now
-    # b is treaded as a single vector if and only if it is 1-dimensional;
-    # for other cases this signature must be followed
-    # (..., m, m), (..., m, n) -> (..., m, n)
-    # https://github.com/numpy/numpy/pull/25914
-    if a_dp.ndim > 2 and numpy.lib.NumpyVersion(numpy.__version__) >= "2.0.0":
-        pytest.skip("SAT-6928")
-
     result = dpnp.linalg.solve(a_dp, b_dp)
     expected = numpy.linalg.solve(a_np, b_np)
     assert_dtype_allclose(result, expected)

dpnp/tests/test_usm_type.py

@@ -1285,37 +1285,39 @@ def test_fftshift(self, func, usm_type):
     "usm_type_matrix", list_of_usm_types, ids=list_of_usm_types
 )
 @pytest.mark.parametrize(
-    "usm_type_vector", list_of_usm_types, ids=list_of_usm_types
+    "usm_type_rhs", list_of_usm_types, ids=list_of_usm_types
 )
 @pytest.mark.parametrize(
-    "matrix, vector",
+    "matrix, rhs",
     [
-        ([[1, 2], [3, 5]], dp.empty((2, 0))),
+        ([[1, 2], [3, 5]], numpy.empty((2, 0))),
         ([[1, 2], [3, 5]], [1, 2]),
         (
             [
-                [[1, 1, 1], [0, 2, 5], [2, 5, -1]],
-                [[3, -1, 1], [1, 2, 3], [2, 3, 1]],
-                [[1, 4, 1], [1, 2, -2], [4, 1, 2]],
+                [[1, 1], [0, 2]],
+                [[3, -1], [1, 2]],
+            ],
+            [
+                [[6, -4], [9, -6]],
+                [[15, 1], [15, 1]],
             ],
-            [[6, -4, 27], [9, -6, 15], [15, 1, 11]],
         ),
     ],
     ids=[
-        "2D_Matrix_Empty_Vector",
-        "2D_Matrix_1D_Vector",
-        "3D_Matrix_and_Vectors",
+        "2D_Matrix_Empty_RHS",
+        "2D_Matrix_1D_RHS",
+        "3D_Matrix_and_3D_RHS",
     ],
 )
-def test_solve(matrix, vector, usm_type_matrix, usm_type_vector):
+def test_solve(matrix, rhs, usm_type_matrix, usm_type_rhs):
     x = dp.array(matrix, usm_type=usm_type_matrix)
-    y = dp.array(vector, usm_type=usm_type_vector)
+    y = dp.array(rhs, usm_type=usm_type_rhs)
     z = dp.linalg.solve(x, y)
 
     assert x.usm_type == usm_type_matrix
-    assert y.usm_type == usm_type_vector
+    assert y.usm_type == usm_type_rhs
     assert z.usm_type == du.get_coerced_usm_type(
-        [usm_type_matrix, usm_type_vector]
+        [usm_type_matrix, usm_type_rhs]
     )
 
 

dpnp/tests/third_party/cupy/linalg_tests/test_solve.py

@@ -47,6 +47,7 @@ def check_x(self, a_shape, b_shape, xp, dtype):
         testing.assert_array_equal(b_copy, b)
         return result
 
+    @testing.with_requires("numpy>=2.0")
     def test_solve(self):
         self.check_x((4, 4), (4,))
         self.check_x((5, 5), (5, 2))
@@ -55,15 +56,9 @@ def test_solve(self):
         self.check_x((0, 0), (0,))
         self.check_x((0, 0), (0, 2))
         self.check_x((0, 2, 2), (0, 2, 3))
-        # In numpy 2.0 the broadcast ambiguity has been removed and now
-        # b is treaded as a single vector if and only if it is 1-dimensional;
-        # for other cases this signature must be followed
-        # (..., m, m), (..., m, n) -> (..., m, n)
-        # https://github.com/numpy/numpy/pull/25914
-        if numpy.lib.NumpyVersion(numpy.__version__) < "2.0.0":
-            self.check_x((2, 4, 4), (2, 4))
-            self.check_x((2, 3, 2, 2), (2, 3, 2))
-            self.check_x((0, 2, 2), (0, 2))
+        # Allowed since numpy 2
+        self.check_x((2, 3, 3), (3,))
+        self.check_x((2, 5, 3, 3), (3,))
 
     def check_shape(self, a_shape, b_shape, error_types):
         for xp, error_type in error_types.items():
@@ -82,6 +77,7 @@ def test_solve_singular_empty(self, xp):
         # LinAlgError("Singular matrix") is not raised
         return xp.linalg.solve(a, b)
 
+    @testing.with_requires("numpy>=2.0")
     def test_invalid_shape(self):
         linalg_errors = {
             numpy: numpy.linalg.LinAlgError,
@@ -96,11 +92,12 @@ def test_invalid_shape(self):
         self.check_shape((3, 3), (2,), value_errors)
         self.check_shape((3, 3), (2, 2), value_errors)
         self.check_shape((3, 3, 4), (3,), linalg_errors)
-        # Since numpy >= 2.0, this case does not raise an error
-        if numpy.lib.NumpyVersion(numpy.__version__) < "2.0.0":
-            self.check_shape((2, 3, 3), (3,), value_errors)
         self.check_shape((3, 3), (0,), value_errors)
         self.check_shape((0, 3, 4), (3,), linalg_errors)
+        # Not allowed since numpy 2.0
+        self.check_shape((0, 2, 2), (0, 2), value_errors)
+        self.check_shape((2, 4, 4), (2, 4), value_errors)
+        self.check_shape((2, 3, 2, 2), (2, 3, 2), value_errors)
 
 
 @testing.parameterize(