clone code from github

cupy/_core/_routines_linalg.pxd

+from cupy._core._carray cimport shape_t
+from cupy._core.core cimport _ndarray_base
+
+
+cpdef compute_type_to_str(compute_type)
+
+cpdef get_compute_type(dtype)
+
+cpdef _ndarray_base dot(_ndarray_base a, _ndarray_base b, _ndarray_base out=*)
+
+cpdef _ndarray_base tensordot_core(
+    _ndarray_base a, _ndarray_base b, _ndarray_base out, Py_ssize_t n,
+    Py_ssize_t m, Py_ssize_t k, const shape_t& ret_shape)
+
+cpdef _ndarray_base matmul(
+    _ndarray_base a, _ndarray_base b, _ndarray_base out=*)
+
+
+cpdef enum:
+    COMPUTE_TYPE_TBD = 0
+    COMPUTE_TYPE_DEFAULT = 1   # default
+    COMPUTE_TYPE_PEDANTIC = 2  # disable algorithmic optimizations
+    COMPUTE_TYPE_FP16 = 3      # allow converting inputs to FP16
+    COMPUTE_TYPE_FP32 = 4      # allow converting inputs to FP32
+    COMPUTE_TYPE_FP64 = 5      # allow converting inputs to FP64
+    COMPUTE_TYPE_BF16 = 6      # allow converting inputs to BF16
+    COMPUTE_TYPE_TF32 = 7      # allow converting inputs to TF32

cupy/_core/_routines_logic.pxd

+from cupy._core.core cimport _ndarray_base
+
+
+cdef _ndarray_base _ndarray_all(_ndarray_base self, axis, out, keepdims)
+cdef _ndarray_base _ndarray_any(_ndarray_base self, axis, out, keepdims)
+cdef _ndarray_base _ndarray_greater(_ndarray_base self, other)
+cdef _ndarray_base _ndarray_greater_equal(_ndarray_base self, other)
+cdef _ndarray_base _ndarray_less(_ndarray_base self, other)
+cdef _ndarray_base _ndarray_less_equal(_ndarray_base self, other)
+cdef _ndarray_base _ndarray_equal(_ndarray_base self, other)
+cdef _ndarray_base _ndarray_not_equal(_ndarray_base self, other)

cupy/_core/_routines_manipulation.pxd

+from libcpp cimport vector
+
+from cupy._core._carray cimport shape_t
+from cupy._core._carray cimport strides_t
+from cupy._core.core cimport _ndarray_base
+
+
+cdef class broadcast:
+    cdef:
+        readonly tuple values
+        readonly tuple shape
+        readonly Py_ssize_t size
+        readonly Py_ssize_t nd
+
+
+cdef _ndarray_shape_setter(_ndarray_base self, newshape)
+cdef _ndarray_base _ndarray_reshape(_ndarray_base self, tuple shape, order)
+cdef _ndarray_base _ndarray_transpose(_ndarray_base self, tuple axes)
+cdef _ndarray_base _ndarray_swapaxes(
+    _ndarray_base self, Py_ssize_t axis1, Py_ssize_t axis2)
+cdef _ndarray_base _ndarray_flatten(_ndarray_base self, order)
+cdef _ndarray_base _ndarray_ravel(_ndarray_base self, order)
+cdef _ndarray_base _ndarray_squeeze(_ndarray_base self, axis)
+cdef _ndarray_base _ndarray_repeat(_ndarray_base self, repeats, axis)
+
+cpdef _ndarray_base _expand_dims(_ndarray_base a, tuple axis)
+cpdef _ndarray_base moveaxis(_ndarray_base a, source, destination)
+cpdef _ndarray_base _move_single_axis(
+    _ndarray_base a, Py_ssize_t source, Py_ssize_t destination)
+cpdef _ndarray_base rollaxis(
+    _ndarray_base a, Py_ssize_t axis, Py_ssize_t start=*)
+cpdef _ndarray_base broadcast_to(_ndarray_base array, shape)
+cpdef _ndarray_base _reshape(_ndarray_base self, const shape_t &shape_spec)
+cpdef _ndarray_base _T(_ndarray_base self)
+cpdef _ndarray_base _transpose(
+    _ndarray_base self, const vector.vector[Py_ssize_t] &axes)
+cpdef _ndarray_base _concatenate(
+    list arrays, Py_ssize_t axis, tuple shape, _ndarray_base out, str casting)
+cpdef _ndarray_base concatenate_method(
+    tup, int axis, _ndarray_base out=*, dtype=*, casting=*)

cupy/_core/_routines_math.pxd

+from cupy._core.core cimport _ndarray_base
+
+
+cdef _ndarray_base _ndarray_conj(_ndarray_base self)
+cdef _ndarray_base _ndarray_real_getter(_ndarray_base self)
+cdef _ndarray_base _ndarray_real_setter(_ndarray_base self, value)
+cdef _ndarray_base _ndarray_imag_getter(_ndarray_base self)
+cdef _ndarray_base _ndarray_imag_setter(_ndarray_base self, value)
+cdef _ndarray_base _ndarray_prod(
+    _ndarray_base self, axis, dtype, out, keepdims)
+cdef _ndarray_base _ndarray_sum(_ndarray_base self, axis, dtype, out, keepdims)
+cdef _ndarray_base _ndarray_cumsum(_ndarray_base self, axis, dtype, out)
+cdef _ndarray_base _ndarray_cumprod(_ndarray_base self, axis, dtype, out)
+cdef _ndarray_base _ndarray_clip(_ndarray_base self, a_min, a_max, out)
+
+cpdef _ndarray_base _nansum(_ndarray_base a, axis, dtype, out, keepdims)
+cpdef _ndarray_base _nanprod(_ndarray_base a, axis, dtype, out, keepdims)
+
+cpdef enum scan_op:
+    SCAN_SUM = 0
+    SCAN_PROD = 1
+
+cdef _ndarray_base scan(_ndarray_base a, op, dtype=*, _ndarray_base out=*,
+                        incomplete=*, chunk_size=*)
+cdef object _sum_auto_dtype
+cdef object _add
+cdef object _conj
+cdef object _angle
+cdef object _positive
+cdef object _negative
+cdef object _multiply
+cdef object _divide
+cdef object _power
+cdef object _subtract
+cdef object _true_divide
+cdef object _floor_divide
+cdef object _remainder
+cdef object _absolute
+cdef object _sqrt

cupy/_core/_routines_sorting.pxd

+from cupy._core.core cimport _ndarray_base
+
+
+cdef _ndarray_sort(_ndarray_base self, int axis)
+cdef _ndarray_base _ndarray_argsort(_ndarray_base self, axis)
+cdef _ndarray_partition(_ndarray_base self, kth, int axis)
+cdef _ndarray_base _ndarray_argpartition(self, kth, axis)

cupy/_core/_routines_statistics.pxd

+from cupy._core.core cimport _ndarray_base
+
+
+# TODO(niboshi): Move {nan,}arg{min,max} to sorting
+
+
+cdef _ndarray_base _ndarray_max(_ndarray_base self, axis, out, dtype, keepdims)
+cdef _ndarray_base _ndarray_min(_ndarray_base self, axis, out, dtype, keepdims)
+cdef _ndarray_base _ndarray_ptp(_ndarray_base self, axis, out, keepdims)
+cdef _ndarray_base _ndarray_argmax(
+    _ndarray_base self, axis, out, dtype, keepdims)
+cdef _ndarray_base _ndarray_argmin(
+    _ndarray_base self, axis, out, dtype, keepdims)
+cdef _ndarray_base _ndarray_mean(
+    _ndarray_base self, axis, dtype, out, keepdims)
+cdef _ndarray_base _ndarray_var(
+    _ndarray_base self, axis, dtype, out, ddof, keepdims)
+cdef _ndarray_base _ndarray_std(
+    _ndarray_base self, axis, dtype, out, ddof, keepdims)
+
+cpdef _ndarray_base _median(
+    _ndarray_base a, axis, out, overwrite_input, keepdims)
+
+cpdef _ndarray_base _nanmean(_ndarray_base a, axis, dtype, out, keepdims)
+cpdef _ndarray_base _nanvar(_ndarray_base a, axis, dtype, out, ddof, keepdims)
+cpdef _ndarray_base _nanstd(_ndarray_base a, axis, dtype, out, ddof, keepdims)
+
+
+cpdef _ndarray_base _nanargmin(_ndarray_base a, axis, out, dtype, keepdims)
+cpdef _ndarray_base _nanargmax(_ndarray_base a, axis, out, dtype, keepdims)

cupy/_core/_ufuncs.py

+from cupy._core._kernel import create_ufunc
+
+
+elementwise_copy = create_ufunc(
+    'cupy_copy',
+    ('?->?', 'b->b', 'B->B', 'h->h', 'H->H', 'i->i', 'I->I', 'l->l', 'L->L',
+     'q->q', 'Q->Q', 'e->e', 'f->f', 'd->d', 'F->F', 'D->D'),
+    'out0 = in0',
+    default_casting='unsafe')