void
randv<int>(vsip::Vector<int> v)
{
vsip::Rand<float> rgen(1, true);
v = rgen.randu(v.size()) * v.size();
}
vsip::Vector<T, B>
test_ramp(vsip::Vector<T, B> view, T a, T b)
{
for (vsip::index_type i=0; i<view.size(); ++i)
view.put(i, a + T(i)*b);
return view;
}
void
conv(
vsip::symmetry_type sym,
vsip::support_region_type sup,
vsip::const_Vector<T, Block1> coeff,
vsip::const_Vector<T, Block2> in,
vsip::Vector<T, Block3> out,
vsip::length_type D)
{
using vsip::index_type;
using vsip::length_type;
using vsip::stride_type;
using vsip::Vector;
using vsip::const_Vector;
using vsip::Domain;
using vsip::unbiased;
using vsip::impl::convert_to_local;
using vsip::impl::Working_view_holder;
typedef typename vsip::impl::scalar_of<T>::type scalar_type;
Working_view_holder<const_Vector<T, Block1> > w_coeff(coeff);
Working_view_holder<const_Vector<T, Block2> > w_in(in);
Working_view_holder<Vector<T, Block3> > w_out(out);
Vector<T> kernel = kernel_from_coeff(sym, w_coeff.view);
length_type M = kernel.size(0);
length_type N = in.size(0);
length_type P = out.size(0);
stride_type shift = conv_expected_shift(sup, M);
// expected_P == conv_output_size(sup, M, N, D) == P;
assert(conv_output_size(sup, M, N, D) == P);
Vector<T> sub(M);
// Check result
for (index_type i=0; i<P; ++i)
{
sub = T();
index_type pos = i*D + shift;
if (pos+1 < M)
sub(Domain<1>(0, 1, pos+1)) = w_in.view(Domain<1>(pos, -1, pos+1));
else if (pos >= N)
{
index_type start = pos - N + 1;
sub(Domain<1>(start, 1, M-start)) = w_in.view(Domain<1>(N-1, -1, M-start));
}
else
sub = w_in.view(Domain<1>(pos, -1, M));
w_out.view(i) = dot(kernel, sub);
}
}
void
corr(
vsip::bias_type bias,
vsip::support_region_type sup,
vsip::const_Vector<T, Block1> ref,
vsip::const_Vector<T, Block2> in,
vsip::Vector<T, Block3> out)
{
using vsip::index_type;
using vsip::length_type;
using vsip::stride_type;
using vsip::Vector;
using vsip::Domain;
using vsip::unbiased;
typedef typename vsip::impl::Scalar_of<T>::type scalar_type;
length_type M = ref.size(0);
length_type N = in.size(0);
length_type P = out.size(0);
length_type expected_P = corr_output_size(sup, M, N);
stride_type shift = expected_shift(sup, M);
assert(expected_P == P);
Vector<T> sub(M);
// compute correlation
for (index_type i=0; i<P; ++i)
{
sub = T();
stride_type pos = static_cast<stride_type>(i) + shift;
scalar_type scale;
if (pos < 0)
{
sub(Domain<1>(-pos, 1, M + pos)) = in(Domain<1>(0, 1, M+pos));
scale = scalar_type(M + pos);
}
else if (pos + M > N)
{
sub(Domain<1>(0, 1, N-pos)) = in(Domain<1>(pos, 1, N-pos));
scale = scalar_type(N - pos);
}
else
{
sub = in(Domain<1>(pos, 1, M));
scale = scalar_type(M);
}
#if VSIP_IMPL_CORR_CORRECT_SAME_SUPPORT_SCALING
#else
if (sup == vsip::support_same)
{
if (i < (M/2)) scale = i + (M+1)/2; // i + ceil(M/2)
else if (i < N - (M/2)) scale = M; // M
else scale = N - 1 + (M+1)/2 - i; // N-1+ceil(M/2)-i
}
#endif
T val = dot(ref, impl_conj(sub));
if (bias == vsip::unbiased)
val /= scale;
out(i) = val;
}
}
