Merge pull request #29 from projekter/empty-product

Fix #28: Empty product

Author: asistradition

sparse_dot_mkl/_dense_dense.py

@@ -48,7 +48,7 @@ def _dense_matmul(matrix_a, matrix_b, scalar=1., out=None, out_scalar=None):
     # The complex versions of these functions take void pointers instead of passed structs
     # So create a C struct if necessary to be passed by reference
     scalar = _mkl_scalar(scalar, complex_type, double_precision)
-    out_scalar = _mkl_scalar(out_scalar, complex_type, double_precision)
+    out_scalar = _mkl_scalar(0 if out is None else out_scalar, complex_type, double_precision)
 
     func(layout_a,
          111,
@@ -75,8 +75,13 @@ def _dense_dot_dense(matrix_a, matrix_b, cast=False, scalar=1., out=None, out_sc
     # Check for edge condition inputs which result in empty outputs
     if _empty_output_check(matrix_a, matrix_b):
         debug_print("Skipping multiplication because A (dot) B must yield an empty matrix")
-        final_dtype = np.float64 if matrix_a.dtype != matrix_b.dtype or matrix_a.dtype != np.float32 else np.float32
-        return _out_matrix((matrix_a.shape[0], matrix_b.shape[1]), final_dtype, out_arr=out)
+        output_arr = _out_matrix((matrix_a.shape[0], matrix_b.shape[1]),
+                                 _type_check(matrix_a, matrix_b, cast=cast, convert=False), out_arr=out)
+        if out is None or (out_scalar is not None and not out_scalar):
+            output_arr.fill(0)
+        elif out_scalar is not None:
+            output_arr *= out_scalar
+        return output_arr
 
     matrix_a, matrix_b = _type_check(matrix_a, matrix_b, cast=cast)
 

sparse_dot_mkl/_gram_matrix.py

@@ -93,7 +93,7 @@ def _gram_matrix_sparse(
 
 
 # Dict keyed by ('double_precision_bool', 'complex_bool')
-_mkl_skryd_funcs = {
+_mkl_syrkd_funcs = {
     (False, False): MKL._mkl_sparse_s_syrkd,
     (True, False): MKL._mkl_sparse_d_syrkd,
     (False, True): MKL._mkl_sparse_c_syrkd,
@@ -129,19 +129,29 @@ def _gram_matrix_sparse_to_dense(
     _order_mkl_handle(sp_ref_a)
 
     out_dtype = _output_dtypes[(double_prec, complex_type)]
-    func = _mkl_skryd_funcs[(double_prec, complex_type)]
+    func = _mkl_syrkd_funcs[(double_prec, complex_type)]
 
     out_dim = matrix_a.shape[0 if aat else 1]
 
     output_arr = _out_matrix(
         (out_dim, out_dim),
         out_dtype,
-        order="C", out_arr=out)
+        order="C",
+        out_arr=out,
+        initialize_zeros=True
+    )
     _, output_ld = _get_numpy_layout(output_arr)
 
     if _empty_output_check(matrix_a, matrix_a):
         _destroy_mkl_handle(sp_ref_a)
+        if out is None or (out_scalar is not None and not out_scalar):
+            output_arr.fill(0)
+        elif out_scalar is not None:
+            output_arr *= out_scalar
         return output_arr
+    
+    if out is None:
+        out_scalar = 0.
 
     scalar = _mkl_scalar(scalar, complex_type, double_prec)
     out_scalar = _mkl_scalar(out_scalar, complex_type, double_prec)
@@ -165,7 +175,7 @@ def _gram_matrix_sparse_to_dense(
     # matrix. This stupid thing only happens with specific flags
     # I could probably leave it but it's pretty annoying
 
-    if not aat and out is None and not complex_type:
+    if not aat and out is None:
         output_arr[np.tril_indices(output_arr.shape[0], k=-1)] = 0.0
 
     return output_arr
@@ -208,6 +218,13 @@ def _gram_matrix_dense_to_dense(
     :rtype: numpy.ndarray
     """
 
+    if aat and np.iscomplexobj(matrix_a):
+        raise ValueError(
+            "transpose=True with dense complex data currently "
+            "fails with an Intel oneMKL ERROR: "
+            "Parameter 3 was incorrect on entry to cblas_csyrk"
+        )
+
     # Get dimensions
     n, k = matrix_a.shape if aat else matrix_a.shape[::-1]
 
@@ -223,15 +240,16 @@ def _gram_matrix_dense_to_dense(
         (n, n),
         out_dtype,
         order="C" if layout_a == LAYOUT_CODE_C else "F",
-        out_arr=out
+        out_arr=out,
+        initialize_zeros=True
     )
 
     # The complex versions of these functions take void pointers instead of
     # passed structs, so create a C struct if necessary to be passed by
     # reference
     scalar = _mkl_scalar(scalar, complex_type, double_precision)
     out_scalar = _mkl_scalar(out_scalar, complex_type, double_precision)
-
+    
     func(
         layout_a,
         MKL_UPPER,
@@ -243,7 +261,7 @@ def _gram_matrix_dense_to_dense(
         ld_a,
         out_scalar if not complex_type else _ctypes.byref(scalar),
         output_arr,
-        n,
+        n
     )
 
     return output_arr
@@ -279,6 +297,22 @@ def _gram_matrix(
     :rtype: scipy.sparse.csr_matrix, np.ndarray
     """
 
+    if _sps.issparse(matrix) and not (is_csr(matrix) or is_csc(matrix)):
+        raise ValueError(
+            "gram_matrix requires sparse matrix to be CSR or CSC format"
+        )
+    elif is_csc(matrix) and not cast:
+        raise ValueError(
+            "gram_matrix cannot use a CSC matrix unless cast=True"
+        )
+    elif out is not None and not dense:
+        raise ValueError(
+            "out argument cannot be used with sparse (dot) sparse "
+            "matrix multiplication"
+        )
+    elif out is not None and not isinstance(out, np.ndarray):
+        raise ValueError("out argument must be dense")
+
     # Check for edge condition inputs which result in empty outputs
     if _empty_output_check(matrix, matrix):
         debug_print(
@@ -290,25 +324,18 @@ def _gram_matrix(
             if transpose
             else (matrix.shape[0], matrix.shape[0])
         )
-        output_func = _sps.csr_matrix if _sps.isspmatrix(matrix) else np.zeros
-        return output_func(output_shape, dtype=matrix.dtype)
-
-    if np.iscomplexobj(matrix):
-        raise ValueError(
-            "gram_matrix_mkl does not support complex datatypes"
-        )
+        if out is None:
+            output_func = np.zeros if dense else _sps.csr_matrix
+            return output_func(output_shape, dtype=matrix.dtype)
+        elif out_scalar is not None and not out_scalar:
+            out.fill(0)
+        elif out_scalar is not None:
+            out *= out_scalar
+        return out
 
     matrix = _type_check(matrix, cast=cast)
 
-    if _sps.issparse(matrix) and not (is_csr(matrix) or is_csc(matrix)):
-        raise ValueError(
-            "gram_matrix requires sparse matrix to be CSR or CSC format"
-        )
-    elif is_csc(matrix) and not cast:
-        raise ValueError(
-            "gram_matrix cannot use a CSC matrix unless cast=True"
-        )
-    elif not _sps.issparse(matrix):
+    if not _sps.issparse(matrix):
         return _gram_matrix_dense_to_dense(
             matrix,
             aat=transpose,
@@ -322,11 +349,6 @@ def _gram_matrix(
             out=out,
             out_scalar=out_scalar
         )
-    elif out is not None:
-        raise ValueError(
-            "out argument cannot be used with sparse (dot) sparse "
-            "matrix multiplication"
-        )
     else:
         return _gram_matrix_sparse(
             matrix,

sparse_dot_mkl/_mkl_interface/_common.py

@@ -770,12 +770,14 @@ def _is_valid_dtype(matrix, complex_dtype=False, all_dtype=False):
         return matrix.dtype in NUMPY_FLOAT_DTYPES
 
 
-def _type_check(matrix_a, matrix_b=None, cast=False, allow_complex=True):
+def _type_check(matrix_a, matrix_b=None, cast=False, allow_complex=True, convert=True):
     """
     Make sure that both matrices are single precision floats or both are
     double precision floats.
     If not, convert to double precision floats if cast is True,
     or raise an error if cast is False
+    If convert is set to False, the resulting data type is returned without
+    any conversion happening.
     """
 
     _n_complex = _np.iscomplexobj(matrix_a) + _np.iscomplexobj(matrix_b)
@@ -785,17 +787,17 @@ def _type_check(matrix_a, matrix_b=None, cast=False, allow_complex=True):
 
     # If there's no matrix B and matrix A is valid dtype, return it
     if matrix_b is None and _is_valid_dtype(matrix_a, all_dtype=True):
-        return matrix_a
+        return matrix_a if convert else matrix_a.dtype
 
     # If matrix A is complex but not csingle or cdouble, and cast is True,
     # convert it to a cdouble
     elif matrix_b is None and cast and _n_complex == 1:
-        return _cast_to(matrix_a, _np.cdouble)
+        return _cast_to(matrix_a, _np.cdouble) if convert else _np.cdouble
 
     # If matrix A is real but not float32 or float64, and cast is True,
     # convert it to a float64
     elif matrix_b is None and cast:
-        return _cast_to(matrix_a, _np.float64)
+        return _cast_to(matrix_a, _np.float64) if convert else _np.float64
 
     # Raise an error - the dtype is invalid and cast is False
     elif matrix_b is None:
@@ -809,7 +811,7 @@ def _type_check(matrix_a, matrix_b=None, cast=False, allow_complex=True):
         _is_valid_dtype(matrix_a, all_dtype=True) and
         matrix_a.dtype == matrix_b.dtype
     ):
-        return matrix_a, matrix_b
+        return (matrix_a, matrix_b) if convert else matrix_a.dtype
 
     # If neither matrix is complex and cast is True, convert to float64s
     # and return them
@@ -818,7 +820,7 @@ def _type_check(matrix_a, matrix_b=None, cast=False, allow_complex=True):
             f"Recasting matrix data types {matrix_a.dtype} and "
             f"{matrix_b.dtype} to np.float64"
         )
-        return _cast_to(matrix_a, _np.float64), _cast_to(matrix_b, _np.float64)
+        return (_cast_to(matrix_a, _np.float64), _cast_to(matrix_b, _np.float64)) if convert else _np.float64
 
     # If both matrices are complex and cast is True, convert to cdoubles
     # and return them
@@ -827,7 +829,7 @@ def _type_check(matrix_a, matrix_b=None, cast=False, allow_complex=True):
             f"Recasting matrix data types {matrix_a.dtype} and "
             f"{matrix_b.dtype} to _np.cdouble"
         )
-        return _cast_to(matrix_a, _np.cdouble), _cast_to(matrix_b, _np.cdouble)
+        return (_cast_to(matrix_a, _np.cdouble), _cast_to(matrix_b, _np.cdouble)) if convert else _np.cdouble
 
     # Cast reals and complex matrices together
     elif (
@@ -838,7 +840,7 @@ def _type_check(matrix_a, matrix_b=None, cast=False, allow_complex=True):
         debug_print(
             f"Recasting matrix data type {matrix_b.dtype} to {matrix_a.dtype}"
         )
-        return matrix_a, _cast_to(matrix_b, matrix_a.dtype)
+        return (matrix_a, _cast_to(matrix_b, matrix_a.dtype)) if convert else matrix_a.dtype
 
     elif (
         cast and
@@ -848,14 +850,14 @@ def _type_check(matrix_a, matrix_b=None, cast=False, allow_complex=True):
         debug_print(
             f"Recasting matrix data type {matrix_a.dtype} to {matrix_b.dtype}"
         )
-        return _cast_to(matrix_a, matrix_b.dtype), matrix_b
+        return (_cast_to(matrix_a, matrix_b.dtype), matrix_b) if convert else matrix_b.dtype
 
     elif cast and _n_complex == 1:
         debug_print(
             f"Recasting matrix data type {matrix_a.dtype} and {matrix_b.dtype}"
             f" to np.cdouble"
         )
-        return _cast_to(matrix_a, _np.cdouble), _cast_to(matrix_b, _np.cdouble)
+        return (_cast_to(matrix_a, _np.cdouble), _cast_to(matrix_b, _np.cdouble)) if convert else _np.cdouble
 
     # If cast is False, can't cast anything together
     elif not cast:
@@ -882,9 +884,16 @@ def _mkl_scalar(scalar, complex_type, double_precision):
         return float(scalar)
 
 
-def _out_matrix(shape, dtype, order="C", out_arr=None, out_t=False):
+def _out_matrix(
+    shape,
+    dtype,
+    order="C",
+    out_arr=None,
+    out_t=False,
+    initialize_zeros=False
+):
     """
-    Create an all-zero matrix or check to make sure that
+    Create an undefined matrix or check to make sure that
     the provided output array matches
 
     :param shape: Required output shape
@@ -904,8 +913,10 @@ def _out_matrix(shape, dtype, order="C", out_arr=None, out_t=False):
     out_t = False if out_t is None else out_t
 
     # If there's no output array allocate a new array and return it
-    if out_arr is None:
+    if out_arr is None and initialize_zeros:
         return _np.zeros(shape, dtype=dtype, order=order)
+    elif out_arr is None:
+        return _np.ndarray(shape, dtype=dtype, order=order)
 
     # Check and make sure the order is correct
     # Note 1d arrays have both flags set

sparse_dot_mkl/_sparse_dense.py

@@ -106,7 +106,7 @@ def _sparse_dense_matmul(
 
         # Create a C struct if necessary to be passed
         scalar = _mkl_scalar(scalar, cplx, dbl)
-        out_scalar = _mkl_scalar(out_scalar, cplx, dbl)
+        out_scalar = _mkl_scalar(0 if out is None else out_scalar, cplx, dbl)
 
         ret_val = func(
             11 if transpose else 10,
@@ -165,14 +165,15 @@ def _sparse_dot_dense(
         debug_print(
             "Skipping multiplication because A (dot) B must yield empty matrix"
         )
-        final_dtype = (
-            np.float64
-            if matrix_a.dtype != matrix_b.dtype or matrix_a.dtype != np.float32
-            else np.float32
-        )
-        return _out_matrix(
-            (matrix_a.shape[0], matrix_b.shape[1]), final_dtype, out_arr=out
+        output_arr = _out_matrix(
+            (matrix_a.shape[0], matrix_b.shape[1]),
+            _type_check(matrix_a, matrix_b, cast=cast, convert=False), out_arr=out
         )
+        if out is None or (out_scalar is not None and not out_scalar):
+            output_arr.fill(0)
+        elif out_scalar is not None:
+            output_arr *= out_scalar
+        return output_arr
 
     matrix_a, matrix_b = _type_check(matrix_a, matrix_b, cast=cast)
 

sparse_dot_mkl/_sparse_sparse.py

@@ -169,16 +169,19 @@ def _sparse_dot_sparse(
     # Check for edge condition inputs which result in empty outputs
     if _empty_output_check(matrix_a, matrix_b):
 
+        final_dtype = _type_check(matrix_a, matrix_b, cast=cast, convert=False)
         if dense:
-            return _out_matrix(
+            output_arr = _out_matrix(
                 (matrix_a.shape[0], matrix_b.shape[1]),
-                matrix_a.dtype,
+                final_dtype,
                 out_arr=out
             )
+            output_arr.fill(0)
+            return output_arr
         else:
             return default_output(
                 (matrix_a.shape[0], matrix_b.shape[1]),
-                dtype=matrix_a.dtype
+                dtype=final_dtype
             )
 
     # Check dtypes

sparse_dot_mkl/_sparse_sypr.py

@@ -69,8 +69,8 @@ def _sypr_sparse_A_dense_B(
         matrix_b,
         layout_b,
         ld_b,
-        float(out_scalar) if a_scalar is not None else 1.0,
-        float(out_scalar) if out_scalar is not None else 1.0,
+        float(a_scalar) if a_scalar is not None else 1.0,
+        0.0 if out is None else (float(out_scalar) if out_scalar is not None else 1.0),
         output_arr,
         output_layout,
         output_ld,