pyncarray

pyncarray

Functions

Return	Name	Description
NCA_PYNCARRAY_API void	set_eager
NCA_PYNCARRAY_API bool	is_eager
bool	is_nc_array inline
bool	needs_device_vm inline
void	_build_node_helper
py::object	eval_python_expr
void	register_expr_class
void	register_common_array_methods	Helper function to attach common methods to a Python binding for an array specialization.
auto	host_dispatch inline	This is a duplicate dispatch to that in ncarry/array_impl.hh but host only. The original needs to be host/device, so we get a lot of spurious warnings.
ncarray::DType	pydtype_to_dtype inline	Cast a pybind11 array dtype to a ncarray::DType.
auto	pyarray_to_view inline	Construct a view over a NumPy array.
ncarray::ArrayImpl< LayoutT, RefT >	pyarray_to_ref inline	Construct a reference type array from an array of arrays.
ncarray::ArrayImpl< LayoutT, RefT >	pylist_to_ref inline	Construct a reference type array from a list of arrays.
py::array	ncarr_to_numpy inline	Construct a NumPy array from an input nc/soarray type object.
ncarray::Slice	pyslice_to_slice inline
std::vector< ncarray::IndexItem >	pytuple_to_indices inline

---

set_eager

NCA_PYNCARRAY_API void set_eager(bool eager)

---

is_eager

NCA_PYNCARRAY_API bool is_eager()

---

is_nc_array

inline

template<class LPolicy, class MemType> inline bool is_nc_array(const py::handle & val)

---

needs_device_vm

inline

inline bool needs_device_vm(const py::handle & node)

---

_build_node_helper

template<typename MemType> void _build_node_helper(ncarray::ExprMVNode< MemType > & self, py::object other)

---

eval_python_expr

template<typename MemType> py::object eval_python_expr(const py::object & pyexpr)

---

register_expr_class

template<typename MemType> void register_expr_class(py::module_ & m, const std::string & name)

---

register_common_array_methods

template<typename ArrayT> void register_common_array_methods(py::classh< ArrayT > & arr_cl)

Helper function to attach common methods to a Python binding for an array specialization.

All array specializations generally have the same functions in C++ and in their Python bindings (plus/minus some speciality features). This function just attaches those all.

Parameters

• ArrayT The kind of array being used. This is a full array specifier, including storage+layout specifier.

Parameters

• arr_cl The class of the Python binding.

---

host_dispatch

inline

template<typename Visitor> inline auto host_dispatch(ncarray::DType type, Visitor && visitor)

This is a duplicate dispatch to that in ncarry/array_impl.hh but host only. The original needs to be host/device, so we get a lot of spurious warnings.

We do NOT want to just silence the warning (e.g. pragma diag_suppress 20014) for nvcc as we want to catch other instances of host/device called from host.

---

pydtype_to_dtype

inline

inline ncarray::DType pydtype_to_dtype(py::dtype dtype)

Cast a pybind11 array dtype to a ncarray::DType.

Parameters

• dtype The pybind11 datatype.

Returns

The ncarray datatype.

---

pyarray_to_view

inline

template<class ViewType> inline auto pyarray_to_view(const py::array & arr, ssize_t target_ndim = -1)

Construct a view over a NumPy array.

Parameters

• ViewType The kind of view to construct. E.g. NCArrayView, SOArrayView.

Parameters

• arr An input NumPy array to convert.

• target_ndim The target dimensionality of the output with padding. TODO: This mostly serves as a hack until broadcasting is suppported. Because of the implicit dimensionality mismatch between NCArrays* and NumPy….

Returns

A view of ViewType over the data from the NumPy array.

---

pyarray_to_ref

inline

template<class LayoutT, class RefT = ncarray::RefPolicy> inline ncarray::ArrayImpl< LayoutT, RefT > pyarray_to_ref(const py::array & arr, const bool read_only_ = false)

Construct a reference type array from an array of arrays.

NOTE: This function assumes the arrays in the array all have the same shape and strides.

Parameters

• LayoutT The layout specifier for the returned reference type. E.g., NCOffsetsPolicy or SOArrayPolicy (PEP3118).

• RefT The reference storage specifier. E.g. RefPolicy or DevRefPolicy.

Parameters

• arr The input array of arrays.

• read_only_ Whether this reference should be read only.

Returns

The constructed reference type.

---

pylist_to_ref

inline

template<class LayoutT, class RefT = ncarray::RefPolicy> inline ncarray::ArrayImpl< LayoutT, RefT > pylist_to_ref(const py::list & list, const bool read_only_ = false)

Construct a reference type array from a list of arrays.

NOTE: This function assumes the arrays in the list all have the same shape and strides.

Parameters

• LayoutT The layout specifier for the returned reference type. E.g., NCOffsetsPolicy or SOArrayPolicy (PEP3118).

• RefT The reference storage specifier. E.g. RefPolicy or DevRefPolicy.

Parameters

• list The input list of arrays.

• read_only_ Whether this reference should be read only.

Returns

The constructed reference type.

---

ncarr_to_numpy

inline

template<class ArrayT> inline py::array ncarr_to_numpy(const ArrayT & ncarr, std::optional< ncarray::DType > dtype = std::nullopt)

Construct a NumPy array from an input nc/soarray type object.

Parameters

• ArrayT The kind of array being used. This is a full array specifier, including storage+layout specifier.

Parameters

• ncarr The input array to convert.

• dtype The datatype for the converted NumPy array. I.e., if provided a cast will be performed. Otherwise, the input array dtype will be used.

Returns

A NumPy array of specified dtype over the input data.

---

pyslice_to_slice

inline

inline ncarray::Slice pyslice_to_slice(ssize_t axis_shape, py::slice slice)

---

pytuple_to_indices

inline

template<class ArrayT> inline std::vector< ncarray::IndexItem > pytuple_to_indices(const ArrayT & self, py::tuple tup)

Variables

Return	Name	Description
NCA_PYNCARRAY_API bool	g_eager_eval	Global toggle for whether Python will eagerly convert Expr objects to arrays.

---

g_eager_eval

NCA_PYNCARRAY_API bool g_eager_eval = true

Global toggle for whether Python will eagerly convert Expr objects to arrays.