Current File : //proc/self/root/lib/python3/dist-packages/pythran/pythonic/numpy/correlate.hpp
#ifndef PYTHONIC_NUMPY_CORRELATE_HPP
#define PYTHONIC_NUMPY_CORRELATE_HPP
#include "pythonic/include/numpy/correlate.hpp"
#include "pythonic/numpy/dot.hpp"
#include "pythonic/numpy/conjugate.hpp"
#include "pythonic/numpy/asarray.hpp"
#include "pythonic/types/ndarray.hpp"
PYTHONIC_NS_BEGIN
namespace numpy
{
template <class A, class B>
types::ndarray<typename A::dtype, types::pshape<long>>
do_correlate(A const &inA, B const &inB, types::str const &type, int out_inc)
// out_inc is used to indicate the inputs were swapped, which means that the
// output must be time reversed and conjugated
{
auto shapeA = sutils::getshape(inA);
auto shapeB = sutils::getshape(inB);
long NA = shapeA[0];
long NB = shapeB[0];
using out_type =
typename __combined<typename A::dtype, typename B::dtype>::type;
// At this point, handling views would slow things down tremendously
auto inA_ = functor::asarray{}(inA);
auto inB_ = functor::asarray{}(inB);
auto outN = 0;
int iLeft;
if (type == "full") {
outN = NA + NB - 1;
iLeft = -NB + 1;
} else if (type == "valid") {
outN = NA - NB + 1;
iLeft = 0;
} else {
assert(type == "same" && "valid type");
outN = NA;
iLeft = -NB + 1 + (NB - 1) / 2;
}
// We need outN output values, no matter what.
int iRight = iLeft + outN;
// Allocate output array
types::ndarray<out_type, types::pshape<long>> out = {outN, out_type()};
out_type *out_ptr = (out_type *)out.buffer;
// if out_inc is -1, we reverse the output.
if (out_inc == -1)
out_ptr += outN - 1;
// For small correlations, numpy uses small_correlate, far more efficient.
// see numpy/core/src/multiarray/arraytypes.c.src
if (out_inc == 1) {
// Incomplete overlap left
for (int i = iLeft; i < 0; i++, out_ptr++) {
*out_ptr = numpy::dot(inA_(types::contiguous_slice(0, NB + i)),
inB_(types::contiguous_slice(-i, NB)));
}
// Complete overlap middle
for (int i = 0; i <= NA - NB; i++, out_ptr++) {
*out_ptr = numpy::dot(inA_(types::contiguous_slice(i, i + NB)),
inB_(types::contiguous_slice(0, NB)));
}
// Incomplete overlap right.
for (int i = NA - NB + 1; i < iRight; i++, out_ptr++) {
*out_ptr = numpy::dot(inA_(types::contiguous_slice(i, NA)),
inB_(types::contiguous_slice(0, NA - i)));
}
} else {
// Incomplete overlap left
for (int i = iLeft; i < 0; i++, out_ptr += out_inc) {
*out_ptr = wrapper::conjugate(
numpy::dot(inA_(types::contiguous_slice(0, NB + i)),
inB_(types::contiguous_slice(-i, NB))));
}
// Complete overlap middle
for (int i = 0; i <= NA - NB; i++, out_ptr += out_inc) {
*out_ptr = wrapper::conjugate(
numpy::dot(inA_(types::contiguous_slice(i, i + NB)),
inB_(types::contiguous_slice(0, NB))));
}
// Incomplete overlap right.
for (int i = NA - NB + 1; i < iRight; i++, out_ptr += out_inc) {
*out_ptr = wrapper::conjugate(
numpy::dot(inA_(types::contiguous_slice(i, NA)),
inB_(types::contiguous_slice(0, NA - i))));
}
}
return out;
}
template <class A, class B>
types::ndarray<typename A::dtype, types::pshape<long>>
correlate(A const &inA, B const &inB, types::str const &type)
{
long NA = inA.template shape<0>();
long NB = inB.template shape<0>();
// If inB is longer than inA, swap them, but time-reverse and conjugate the
// output (-1 flag)
if (NA > NB) {
auto inB_conj = functor::conjugate{}(inB);
return do_correlate(inA, inB_conj, type, 1);
} else {
auto inA_conj = functor::conjugate{}(inA);
return do_correlate(inB, inA_conj, type, -1);
}
}
NUMPY_EXPR_TO_NDARRAY0_IMPL(correlate)
}
PYTHONIC_NS_END
#endif
Mini Shell