[python-package] do not copy column-major numpy arrays when creating Dataset (#6721)

* do not copy column-major numpy arrays when creating Dataset

* fix logic

* lint

* code review

* update test

* move dataset test to basic

* increase features

* assert single layout

---------
Co-authored-by: Nikita Titov <nekit94-08@mail.ru>

python-package/lightgbm/basic.py

@@ -188,6 +188,23 @@ def _get_sample_count(total_nrow: int, params: str) -> int:
     return sample_cnt.value
 
 
+def _np2d_to_np1d(mat: np.ndarray) -> Tuple[np.ndarray, int]:
+    if mat.dtype in (np.float32, np.float64):
+        dtype = mat.dtype
+    else:
+        dtype = np.float32
+    if mat.flags["F_CONTIGUOUS"]:
+        order = "F"
+        layout = _C_API_IS_COL_MAJOR
+    else:
+        order = "C"
+        layout = _C_API_IS_ROW_MAJOR
+    # ensure dtype and order, copies if either do not match
+    data = np.asarray(mat, dtype=dtype, order=order)
+    # flatten array without copying
+    return data.ravel(order=order), layout
+
+
 class _MissingType(Enum):
     NONE = "None"
     NAN = "NaN"
@@ -684,7 +701,8 @@ _C_API_DTYPE_FLOAT64 = 1
 _C_API_DTYPE_INT32 = 2
 _C_API_DTYPE_INT64 = 3
 
-"""Matrix is row major in Python"""
+"""Macro definition of data order in matrix"""
+_C_API_IS_COL_MAJOR = 0
 _C_API_IS_ROW_MAJOR = 1
 
 """Macro definition of prediction type in C API of LightGBM"""
@@ -2297,11 +2315,7 @@ class Dataset:
             raise ValueError("Input numpy.ndarray must be 2 dimensional")
 
         self._handle = ctypes.c_void_p()
-        if mat.dtype == np.float32 or mat.dtype == np.float64:
-            data = np.asarray(mat.reshape(mat.size), dtype=mat.dtype)
-        else:  # change non-float data to float data, need to copy
-            data = np.asarray(mat.reshape(mat.size), dtype=np.float32)
-
+        data, layout = _np2d_to_np1d(mat)
         ptr_data, type_ptr_data, _ = _c_float_array(data)
         _safe_call(
             _LIB.LGBM_DatasetCreateFromMat(
@@ -2309,7 +2323,7 @@ class Dataset:
                 ctypes.c_int(type_ptr_data),
                 ctypes.c_int32(mat.shape[0]),
                 ctypes.c_int32(mat.shape[1]),
-                ctypes.c_int(_C_API_IS_ROW_MAJOR),
+                ctypes.c_int(layout),
                 _c_str(params_str),
                 ref_dataset,
                 ctypes.byref(self._handle),

tests/python_package_test/test_basic.py

@@ -947,3 +947,39 @@ def test_max_depth_warning_is_raised_if_max_depth_gte_5_and_num_leaves_omitted(c
         "in params. Alternatively, pass (max_depth=-1) and just use 'num_leaves' to constrain model complexity."
     )
     assert expected_warning in capsys.readouterr().out
+
+
+@pytest.mark.parametrize("order", ["C", "F"])
+@pytest.mark.parametrize("dtype", ["float32", "int64"])
+def test_no_copy_in_dataset_from_numpy_2d(rng, order, dtype):
+    X = rng.random(size=(100, 3))
+    X = np.require(X, dtype=dtype, requirements=order)
+    X1d, layout = lgb.basic._np2d_to_np1d(X)
+    if order == "F":
+        assert layout == lgb.basic._C_API_IS_COL_MAJOR
+    else:
+        assert layout == lgb.basic._C_API_IS_ROW_MAJOR
+    if dtype == "float32":
+        assert np.shares_memory(X, X1d)
+    else:
+        # makes a copy
+        assert not np.shares_memory(X, X1d)
+
+
+def test_equal_datasets_from_row_major_and_col_major_data(tmp_path):
+    # row-major dataset
+    X_row, y = make_blobs(n_samples=1_000, n_features=3, centers=2)
+    assert X_row.flags["C_CONTIGUOUS"] and not X_row.flags["F_CONTIGUOUS"]
+    ds_row = lgb.Dataset(X_row, y)
+    ds_row_path = tmp_path / "ds_row.txt"
+    ds_row._dump_text(ds_row_path)
+
+    # col-major dataset
+    X_col = np.asfortranarray(X_row)
+    assert X_col.flags["F_CONTIGUOUS"] and not X_col.flags["C_CONTIGUOUS"]
+    ds_col = lgb.Dataset(X_col, y)
+    ds_col_path = tmp_path / "ds_col.txt"
+    ds_col._dump_text(ds_col_path)
+
+    # check datasets are equal
+    assert filecmp.cmp(ds_row_path, ds_col_path)