Allow no-copy construction from SciPy COO arrays. (#822)

* Hold reference to converted scipy.sparse.coo_*.

* Allow comparison against NumPy dtypes.

sparse/mlir_backend/_conversions.py

@@ -78,8 +78,8 @@ def _from_scipy(arr: ScipySparseArray, copy: bool | None = None) -> Array:
 
             return from_constituent_arrays(format=csx_format, arrays=(indptr, indices, data), shape=arr.shape)
         case "coo":
-            if copy is not None and not copy:
-                raise RuntimeError(f"`scipy.sparse.{type(arr.__name__)}` cannot be zero-copy converted.")
+            from ._common import _hold_ref
+
             row, col = arr.row, arr.col
             if row.dtype != col.dtype:
                 raise RuntimeError(f"`row` and `col` dtypes must be the same: {row.dtype} != {col.dtype}.")
@@ -89,10 +89,8 @@ def _from_scipy(arr: ScipySparseArray, copy: bool | None = None) -> Array:
             data = arr.data
             if copy:
                 data = data.copy()
-
-            # TODO: Make them own the data until https://github.com/llvm/llvm-project/issues/116012 is fixed.
-            row = row.copy()
-            col = col.copy()
+                row = row.copy()
+                col = col.copy()
 
             coo_format = (
                 Coo()
@@ -103,7 +101,10 @@ def _from_scipy(arr: ScipySparseArray, copy: bool | None = None) -> Array:
                 .build()
             )
 
-            return from_constituent_arrays(format=coo_format, arrays=(pos, row, col, data), shape=arr.shape)
+            ret = from_constituent_arrays(format=coo_format, arrays=(pos, row, col, data), shape=arr.shape)
+            if not copy:
+                _hold_ref(ret, arr)
+            return ret
         case _:
             raise NotImplementedError(f"No conversion implemented for `scipy.sparse.{type(arr.__name__)}`.")
 

sparse/mlir_backend/_dtypes.py

@@ -33,6 +33,11 @@ def np_dtype(self) -> np.dtype:
     def to_ctype(self):
         return rt.as_ctype(self.np_dtype)
 
+    def __eq__(self, value):
+        if np.isdtype(value) or isinstance(value, str):
+            value = asdtype(value)
+        return super().__eq__(value)
+
 
 @dataclasses.dataclass(eq=True, frozen=True, kw_only=True)
 class IeeeRealFloatingDType(DType):

sparse/mlir_backend/tests/test_simple.py

@@ -301,7 +301,7 @@ def test_coo_3d_format(dtype):
 @parametrize_dtypes
 def test_sparse_vector_format(dtype):
     if sparse.asdtype(dtype) in {sparse.complex64, sparse.complex128}:
-        pytest.xfail("Heisenbug")
+        pytest.xfail("The sparse_vector format returns incorrect results for complex dtypes.")
     format = sparse.formats.Coo().with_ndim(1).with_dtype(dtype).build()
 
     SHAPE = (10,)
@@ -465,8 +465,7 @@ def test_asformat(rng, src_fmt, dst_fmt):
 
     expected = sps_arr.asformat(dst_fmt)
 
-    copy = None if dst_fmt == "coo" else False
-    actual_fmt = sparse.asarray(expected, copy=copy).format
+    actual_fmt = sparse.asarray(expected, copy=False).format
     actual = sp_arr.asformat(actual_fmt)
     actual_sps = sparse.to_scipy(actual)