mirror of
https://github.com/pybind/pybind11.git
synced 2024-11-14 17:43:53 +00:00
202 lines
8.3 KiB
C++
202 lines
8.3 KiB
C++
/*
|
|
pybind/numpy.h: Basic NumPy support, auto-vectorization support
|
|
|
|
Copyright (c) 2015 Wenzel Jakob <wenzel@inf.ethz.ch>
|
|
|
|
All rights reserved. Use of this source code is governed by a
|
|
BSD-style license that can be found in the LICENSE file.
|
|
*/
|
|
|
|
#pragma once
|
|
|
|
#include <pybind/pybind.h>
|
|
#if defined(_MSC_VER)
|
|
#pragma warning(push)
|
|
#pragma warning(disable: 4127) // warning C4127: Conditional expression is constant
|
|
#endif
|
|
|
|
NAMESPACE_BEGIN(pybind)
|
|
|
|
class array : public buffer {
|
|
protected:
|
|
struct API {
|
|
enum Entries {
|
|
API_PyArray_Type = 2,
|
|
API_PyArray_DescrFromType = 45,
|
|
API_PyArray_FromAny = 69,
|
|
API_PyArray_NewCopy = 85,
|
|
API_PyArray_NewFromDescr = 94,
|
|
API_NPY_C_CONTIGUOUS = 0x0001,
|
|
API_NPY_F_CONTIGUOUS = 0x0002,
|
|
API_NPY_NPY_ARRAY_FORCECAST = 0x0010,
|
|
API_NPY_ENSURE_ARRAY = 0x0040
|
|
};
|
|
|
|
static API lookup() {
|
|
PyObject *numpy = PyImport_ImportModule("numpy.core.multiarray");
|
|
PyObject *capsule = numpy ? PyObject_GetAttrString(numpy, "_ARRAY_API") : nullptr;
|
|
void **api_ptr = (void **) (capsule ? PyCapsule_GetPointer(capsule, NULL) : nullptr);
|
|
Py_XDECREF(capsule);
|
|
Py_XDECREF(numpy);
|
|
if (api_ptr == nullptr)
|
|
throw std::runtime_error("Could not acquire pointer to NumPy API!");
|
|
API api;
|
|
api.PyArray_Type = (decltype(api.PyArray_Type)) api_ptr[API_PyArray_Type];
|
|
api.PyArray_DescrFromType = (decltype(api.PyArray_DescrFromType)) api_ptr[API_PyArray_DescrFromType];
|
|
api.PyArray_FromAny = (decltype(api.PyArray_FromAny)) api_ptr[API_PyArray_FromAny];
|
|
api.PyArray_NewCopy = (decltype(api.PyArray_NewCopy)) api_ptr[API_PyArray_NewCopy];
|
|
api.PyArray_NewFromDescr = (decltype(api.PyArray_NewFromDescr)) api_ptr[API_PyArray_NewFromDescr];
|
|
return api;
|
|
}
|
|
|
|
bool PyArray_Check(PyObject *obj) const { return (bool) PyObject_TypeCheck(obj, PyArray_Type); }
|
|
|
|
PyObject *(*PyArray_DescrFromType)(int);
|
|
PyObject *(*PyArray_NewFromDescr)
|
|
(PyTypeObject *, PyObject *, int, Py_intptr_t *,
|
|
Py_intptr_t *, void *, int, PyObject *);
|
|
PyObject *(*PyArray_NewCopy)(PyObject *, int);
|
|
PyTypeObject *PyArray_Type;
|
|
PyObject *(*PyArray_FromAny) (PyObject *, PyObject *, int, int, int, PyObject *);
|
|
};
|
|
public:
|
|
PYBIND_OBJECT_DEFAULT(array, buffer, lookup_api().PyArray_Check)
|
|
|
|
template <typename Type> array(size_t size, const Type *ptr) {
|
|
API& api = lookup_api();
|
|
PyObject *descr = api.PyArray_DescrFromType(
|
|
(int) format_descriptor<Type>::value()[0]);
|
|
if (descr == nullptr)
|
|
throw std::runtime_error("NumPy: unsupported buffer format!");
|
|
Py_intptr_t shape = (Py_intptr_t) size;
|
|
PyObject *tmp = api.PyArray_NewFromDescr(
|
|
api.PyArray_Type, descr, 1, &shape, nullptr, (void *) ptr, 0, nullptr);
|
|
if (tmp == nullptr)
|
|
throw std::runtime_error("NumPy: unable to create array!");
|
|
m_ptr = api.PyArray_NewCopy(tmp, -1 /* any order */);
|
|
Py_DECREF(tmp);
|
|
if (m_ptr == nullptr)
|
|
throw std::runtime_error("NumPy: unable to copy array!");
|
|
}
|
|
|
|
array(const buffer_info &info) {
|
|
API& api = lookup_api();
|
|
if (info.format.size() != 1)
|
|
throw std::runtime_error("Unsupported buffer format!");
|
|
PyObject *descr = api.PyArray_DescrFromType(info.format[0]);
|
|
if (descr == nullptr)
|
|
throw std::runtime_error("NumPy: unsupported buffer format '" + info.format + "'!");
|
|
PyObject *tmp = api.PyArray_NewFromDescr(
|
|
api.PyArray_Type, descr, info.ndim, (Py_intptr_t *) &info.shape[0],
|
|
(Py_intptr_t *) &info.strides[0], info.ptr, 0, nullptr);
|
|
if (tmp == nullptr)
|
|
throw std::runtime_error("NumPy: unable to create array!");
|
|
m_ptr = api.PyArray_NewCopy(tmp, -1 /* any order */);
|
|
Py_DECREF(tmp);
|
|
if (m_ptr == nullptr)
|
|
throw std::runtime_error("NumPy: unable to copy array!");
|
|
}
|
|
|
|
protected:
|
|
static API &lookup_api() {
|
|
static API api = API::lookup();
|
|
return api;
|
|
}
|
|
};
|
|
|
|
template <typename T> class array_dtype : public array {
|
|
public:
|
|
PYBIND_OBJECT_CVT(array_dtype, array, is_non_null, m_ptr = ensure(m_ptr));
|
|
array_dtype() : array() { }
|
|
static bool is_non_null(PyObject *ptr) { return ptr != nullptr; }
|
|
static PyObject *ensure(PyObject *ptr) {
|
|
API &api = lookup_api();
|
|
PyObject *descr = api.PyArray_DescrFromType(format_descriptor<T>::value()[0]);
|
|
return api.PyArray_FromAny(ptr, descr, 0, 0,
|
|
API::API_NPY_C_CONTIGUOUS | API::API_NPY_ENSURE_ARRAY |
|
|
API::API_NPY_NPY_ARRAY_FORCECAST, nullptr);
|
|
}
|
|
};
|
|
|
|
NAMESPACE_BEGIN(detail)
|
|
PYBIND_TYPE_CASTER_PYTYPE(array)
|
|
PYBIND_TYPE_CASTER_PYTYPE(array_dtype<int8_t>) PYBIND_TYPE_CASTER_PYTYPE(array_dtype<uint8_t>)
|
|
PYBIND_TYPE_CASTER_PYTYPE(array_dtype<int16_t>) PYBIND_TYPE_CASTER_PYTYPE(array_dtype<uint16_t>)
|
|
PYBIND_TYPE_CASTER_PYTYPE(array_dtype<int32_t>) PYBIND_TYPE_CASTER_PYTYPE(array_dtype<uint32_t>)
|
|
PYBIND_TYPE_CASTER_PYTYPE(array_dtype<int64_t>) PYBIND_TYPE_CASTER_PYTYPE(array_dtype<uint64_t>)
|
|
PYBIND_TYPE_CASTER_PYTYPE(array_dtype<float>) PYBIND_TYPE_CASTER_PYTYPE(array_dtype<double>)
|
|
NAMESPACE_END(detail)
|
|
|
|
template <typename func_type, typename return_type, typename... args_type, size_t... Index>
|
|
std::function<object(array_dtype<args_type>...)>
|
|
vectorize(func_type &&f, return_type (*) (args_type ...),
|
|
detail::index_sequence<Index...>) {
|
|
|
|
return [f](array_dtype<args_type>... args) -> array {
|
|
/* Request buffers from all parameters */
|
|
const size_t N = sizeof...(args_type);
|
|
std::array<buffer_info, N> buffers {{ args.request()... }};
|
|
|
|
/* Determine dimensions parameters of output array */
|
|
int ndim = 0; size_t count = 0;
|
|
std::vector<size_t> shape;
|
|
for (size_t i=0; i<N; ++i) {
|
|
if (buffers[i].count > count) {
|
|
ndim = buffers[i].ndim;
|
|
shape = buffers[i].shape;
|
|
count = buffers[i].count;
|
|
}
|
|
}
|
|
std::vector<size_t> strides(ndim);
|
|
if (ndim > 0) {
|
|
strides[ndim-1] = sizeof(return_type);
|
|
for (int i=ndim-1; i>0; --i)
|
|
strides[i-1] = strides[i] * shape[i];
|
|
}
|
|
|
|
/* Check if the parameters are actually compatible */
|
|
for (size_t i=0; i<N; ++i) {
|
|
if (buffers[i].count != 1 && (buffers[i].ndim != ndim || buffers[i].shape != shape))
|
|
throw std::runtime_error("pybind::vectorize: incompatible size/dimension of inputs!");
|
|
}
|
|
|
|
/* Call the function */
|
|
std::vector<return_type> result(count);
|
|
for (size_t i=0; i<count; ++i)
|
|
result[i] = f((buffers[Index].count == 1
|
|
? *((args_type *) buffers[Index].ptr)
|
|
: ((args_type *) buffers[Index].ptr)[i])...);
|
|
|
|
if (count == 1)
|
|
return cast(result[0]);
|
|
|
|
/* Return the result */
|
|
return array(buffer_info(result.data(), sizeof(return_type),
|
|
format_descriptor<return_type>::value(),
|
|
ndim, shape, strides));
|
|
};
|
|
}
|
|
|
|
template <typename func_type, typename return_type, typename... args_type>
|
|
std::function<object(array_dtype<args_type>...)>
|
|
vectorize(func_type &&f, return_type (*f_) (args_type ...) = nullptr) {
|
|
return vectorize(f, f_, typename detail::make_index_sequence<sizeof...(args_type)>::type());
|
|
}
|
|
|
|
template <typename return_type, typename... args_type>
|
|
std::function<object(array_dtype<args_type>...)> vectorize(return_type (*f) (args_type ...)) {
|
|
return vectorize(f, f);
|
|
}
|
|
|
|
template <typename func> auto vectorize(func &&f) -> decltype(
|
|
vectorize(std::forward<func>(f), (typename detail::remove_class<decltype(&std::remove_reference<func>::type::operator())>::type *) nullptr)) {
|
|
return vectorize(std::forward<func>(f), (typename detail::remove_class<decltype(
|
|
&std::remove_reference<func>::type::operator())>::type *) nullptr);
|
|
}
|
|
|
|
NAMESPACE_END(pybind)
|
|
|
|
#if defined(_MSC_VER)
|
|
#pragma warning(pop)
|
|
#endif
|