Skip to content

GitLab

P
pypocketfft

Planned maintenance on Wednesday, 2021-01-20, 17:00-18:00. Expect some interruptions during that time

Commit 6ef09ab7 authored by Martin Reinecke's avatar Martin Reinecke
Browse files
Options

Merge branch 'no-ortho' into 'sincos' 

Remove redundant "ortho" parameter from python interface

See merge request !14
parents 8bbee24d 8b4a762a
Show whitespace changes
Inline Side-by-side
Showing with 49 additions and 84 deletions
pypocketfft.cc
View file @ 6ef09ab7
......@@ -227,7 +227,7 @@ py::array r2r_fftpack(const py::array &in, const py::object &axes_,
}
template<typename T> py::array dct_internal(const py::array &in,
const py::object &axes_, int type, int inorm, bool ortho, py::object &out_,
const py::object &axes_, int type, int inorm, py::object &out_,
size_t nthreads)
{
auto axes = makeaxes(in, axes_);
......@@ -239,10 +239,9 @@ template<typename T> py::array dct_internal(const py::array &in,
auto d_out=reinterpret_cast<T *>(res.mutable_data());
{
py::gil_scoped_release release;
// override
if (ortho) inorm=1;
T fct = (type==1) ? norm_fct<T>(inorm, dims, axes, 2, -1)
: norm_fct<T>(inorm, dims, axes, 2);
bool ortho = inorm == 1;
pocketfft::dct(dims, s_in, s_out, axes, type, d_in, d_out, fct, ortho,
nthreads);
}
......@@ -250,15 +249,15 @@ template<typename T> py::array dct_internal(const py::array &in,
}
py::array dct(const py::array &in, int type, const py::object &axes_,
int inorm, bool ortho, py::object &out_, size_t nthreads)
int inorm, py::object &out_, size_t nthreads)
{
if ((type<1) || (type>4)) throw invalid_argument("invalid DCT type");
DISPATCH(in, f64, f32, flong, dct_internal, (in, axes_, type, inorm, ortho,
out_, nthreads))
DISPATCH(in, f64, f32, flong, dct_internal, (in, axes_, type, inorm, out_,
nthreads))
}
template<typename T> py::array dst_internal(const py::array &in,
const py::object &axes_, int type, int inorm, bool ortho, py::object &out_,
const py::object &axes_, int type, int inorm, py::object &out_,
size_t nthreads)
{
auto axes = makeaxes(in, axes_);
......@@ -270,10 +269,9 @@ template<typename T> py::array dst_internal(const py::array &in,
auto d_out=reinterpret_cast<T *>(res.mutable_data());
{
py::gil_scoped_release release;
// override
if (ortho) inorm=1;
T fct = (type==1) ? norm_fct<T>(inorm, dims, axes, 2, 1)
: norm_fct<T>(inorm, dims, axes, 2);
bool ortho = inorm == 1;
pocketfft::dst(dims, s_in, s_out, axes, type, d_in, d_out, fct, ortho,
nthreads);
}
......@@ -281,10 +279,10 @@ template<typename T> py::array dst_internal(const py::array &in,
}
py::array dst(const py::array &in, int type, const py::object &axes_,
int inorm, bool ortho, py::object &out_, size_t nthreads)
int inorm, py::object &out_, size_t nthreads)
{
if ((type<1) || (type>4)) throw invalid_argument("invalid DST type");
DISPATCH(in, f64, f32, flong, dst_internal, (in, axes_, type, inorm, ortho,
DISPATCH(in, f64, f32, flong, dst_internal, (in, axes_, type, inorm,
out_, nthreads))
}
......@@ -599,11 +597,18 @@ axes : list of integers
inorm : int
Normalization type
0 : no normalization
1 : divide by sqrt(N)
1 : make transform orthogonal and divide by sqrt(N)
2 : divide by N
where N is the product of n_i for every transformed axis i.
n_i is 2*(<axis_length>-1 for type 1 and 2*<axis length>
for types 2, 3, 4.
Making the transform orthogonal involves the following additional steps
for every 1D sub-transform:
Type 1 : multiply first and last input value by sqrt(2)
divide first and last output value by sqrt(2)
Type 2 : divide first output value by sqrt(2)
Type 3 : multiply first input value by sqrt(2)
Type 4 : nothing
out : numpy.ndarray (same shape and data type as `a`)
May be identical to `a`, but if it isn't, it must not overlap with `a`.
If None, a new array is allocated to store the output.
......@@ -631,11 +636,17 @@ axes : list of integers
inorm : int
Normalization type
0 : no normalization
1 : divide by sqrt(N)
1 : make transform orthogonal and divide by sqrt(N)
2 : divide by N
where N is the product of n_i for every transformed axis i.
n_i is 2*(<axis_length>+1 for type 1 and 2*<axis length>
for types 2, 3, 4.
Making the transform orthogonal involves the following additional steps
for every 1D sub-transform:
Type 1 : nothing
Type 2 : divide first output value by sqrt(2)
Type 3 : multiply first input value by sqrt(2)
Type 4 : nothing
out : numpy.ndarray (same shape and data type as `a`)
May be identical to `a`, but if it isn't, it must not overlap with `a`.
If None, a new array is allocated to store the output.
......@@ -669,7 +680,7 @@ PYBIND11_MODULE(pypocketfft, m)
m.def("genuine_hartley", genuine_hartley, genuine_hartley_DS, "a"_a,
"axes"_a=None, "inorm"_a=0, "out"_a=None, "nthreads"_a=1);
m.def("dct", dct, dct_DS, "a"_a, "type"_a, "axes"_a=None, "inorm"_a=0,
"ortho"_a=false, "out"_a=None, "nthreads"_a=1);
"out"_a=None, "nthreads"_a=1);
m.def("dst", dst, dst_DS, "a"_a, "type"_a, "axes"_a=None, "inorm"_a=0,
"ortho"_a=false, "out"_a=None, "nthreads"_a=1);
"out"_a=None, "nthreads"_a=1);
}
test.py
View file @ 6ef09ab7
......@@ -55,36 +55,27 @@ def irfft_scipy(a, axis, inorm=0, out=None, nthreads=1):
forward=False, inorm=inorm, out=out,
nthreads=nthreads)
tol = {np.float32: 6e-7, np.float64: 1.5e-15, np.longfloat: 1e-18}
ctype = {np.float32: np.complex64, np.float64: np.complex128, np.longfloat: np.longcomplex}
@pmp("len", len1D)
@pmp("inorm", [0, 1, 2])
def test1D(len, inorm):
@pmp("dtype", [np.float32, np.float64, np.longfloat])
def test1D(len, inorm, dtype):
a = np.random.rand(len)-0.5 + 1j*np.random.rand(len)-0.5j
b = a.astype(np.complex64)
c = a.astype(np.complex256)
_assert_close(a, ifftn(fftn(c, inorm=inorm), inorm=2-inorm), 1e-18)
assert_(_l2error(a, ifftn(fftn(a, inorm=inorm), inorm=2-inorm)) < 1.5e-15)
a = a.astype(ctype[dtype])
eps = tol[dtype]
assert_(_l2error(a, ifftn(fftn(a, inorm=inorm), inorm=2-inorm)) < eps)
assert_(_l2error(a.real, ifftn(fftn(a.real, inorm=inorm), inorm=2-inorm))
< 1.5e-15)
< eps)
assert_(_l2error(a.real, fftn(ifftn(a.real, inorm=inorm), inorm=2-inorm))
< 1.5e-15)
< eps)
assert_(_l2error(a.real, irfftn(rfftn(a.real, inorm=inorm),
inorm=2-inorm, lastsize=len)) < 1.5e-15)
inorm=2-inorm, lastsize=len)) < eps)
tmp = a.copy()
assert_(ifftn(fftn(tmp, out=tmp, inorm=inorm), out=tmp, inorm=2-inorm)
is tmp)
assert_(_l2error(tmp, a) < 1.5e-15)
assert_(_l2error(b, ifftn(fftn(b, inorm=inorm), inorm=2-inorm)) < 6e-7)
assert_(_l2error(b.real, ifftn(fftn(b.real, inorm=inorm), inorm=2-inorm))
< 6e-7)
assert_(_l2error(b.real, fftn(ifftn(b.real, inorm=inorm), inorm=2-inorm))
< 6e-7)
assert_(_l2error(b.real, irfftn(rfftn(b.real, inorm=inorm), lastsize=len,
inorm=2-inorm)) < 6e-7)
tmp = b.copy()
assert_(ifftn(fftn(tmp, out=tmp, inorm=inorm), out=tmp, inorm=2-inorm)
is tmp)
assert_(_l2error(tmp, b) < 6e-7)
assert_(_l2error(tmp, a) < eps)
@pmp("shp", shapes)
......@@ -206,54 +197,17 @@ def test_genuine_hartley_2D(shp, axes):
@pmp("len", len1D)
@pmp("inorm", [0, 1, 2])
@pmp("type", [1, 2, 3])
def testdcst1D(len, inorm, type):
a = np.random.rand(len)-0.5
b = a.astype(np.float32)
c = a.astype(np.float128)
itp = (0, 1, 3, 2, 4)
itype = itp[type]
if type != 1 or len > 1:
_assert_close(a, pypocketfft.dct(pypocketfft.dct(c, inorm=inorm, type=type), inorm=2-inorm, type=itype), 2e-18)
_assert_close(a, pypocketfft.dct(pypocketfft.dct(a, inorm=inorm, type=type), inorm=2-inorm, type=itype), 1.5e-15)
_assert_close(b, pypocketfft.dct(pypocketfft.dct(b, inorm=inorm, type=type), inorm=2-inorm, type=itype), 6e-7)
_assert_close(a, pypocketfft.dst(pypocketfft.dst(c, inorm=inorm, type=type), inorm=2-inorm, type=itype), 2e-18)
_assert_close(a, pypocketfft.dst(pypocketfft.dst(a, inorm=inorm, type=type), inorm=2-inorm, type=itype), 1.5e-15)
_assert_close(b, pypocketfft.dst(pypocketfft.dst(b, inorm=inorm, type=type), inorm=2-inorm, type=itype), 6e-7)
@pmp("len", len1D)
@pmp("type", [1, 2, 3])
def testdcst1Dortho(len, type):
a = np.random.rand(len)-0.5
b = a.astype(np.float32)
c = a.astype(np.float128)
itp = (0, 1, 3, 2, 4)
itype = itp[type]
if type != 1 or len > 1:
_assert_close(a, pypocketfft.dct(pypocketfft.dct(c, ortho=True, type=type), ortho=True, type=itype), 2e-18)
_assert_close(a, pypocketfft.dct(pypocketfft.dct(a, ortho=True, type=type), ortho=True, type=itype), 1.5e-15)
_assert_close(b, pypocketfft.dct(pypocketfft.dct(b, ortho=True, type=type), ortho=True, type=itype), 6e-7)
if type != 1:
_assert_close(a, pypocketfft.dst(pypocketfft.dst(c, ortho=True, type=type), ortho=True, type=itype), 2e-18)
_assert_close(a, pypocketfft.dst(pypocketfft.dst(a, ortho=True, type=type), ortho=True, type=itype), 1.5e-15)
_assert_close(b, pypocketfft.dst(pypocketfft.dst(b, ortho=True, type=type), ortho=True, type=itype), 6e-7)
# TEMPORARY: separate test for DCT/DST IV, since they are less accurate
@pmp("len", len1D)
@pmp("inorm", [0, 1, 2])
@pmp("type", [4])
def testdcst1D4(len, inorm, type):
a = np.random.rand(len)-0.5
b = a.astype(np.float32)
c = a.astype(np.float128)
@pmp("inorm", [0, 1]) # inorm==2 not needed, tested via inverse
@pmp("type", [1, 2, 3, 4])
@pmp("dtype", [np.float32, np.float64, np.longfloat])
def testdcst1D(len, inorm, type, dtype):
a = (np.random.rand(len)-0.5).astype(dtype)
eps = tol[dtype]
itp = (0, 1, 3, 2, 4)
if type==4 and len%2 == 1: # relaxed accuracies for odd-length type 4 transforms
special_tol = {np.float32: 4e-5, np.float64: 6e-14, np.longfloat: 4e-17}
eps = special_tol[dtype]
itype = itp[type]
if type != 1 or len > 1:
_assert_close(a, pypocketfft.dct(pypocketfft.dct(c, inorm=inorm, type=type), inorm=2-inorm, type=itype), 2e-16)
_assert_close(a, pypocketfft.dct(pypocketfft.dct(a, inorm=inorm, type=type), inorm=2-inorm, type=itype), 1.5e-13)
_assert_close(b, pypocketfft.dct(pypocketfft.dct(b, inorm=inorm, type=type), inorm=2-inorm, type=itype), 6e-5)
_assert_close(a, pypocketfft.dst(pypocketfft.dst(c, inorm=inorm, type=type), inorm=2-inorm, type=itype), 2e-16)
_assert_close(a, pypocketfft.dst(pypocketfft.dst(a, inorm=inorm, type=type), inorm=2-inorm, type=itype), 1.5e-13)
_assert_close(b, pypocketfft.dst(pypocketfft.dst(b, inorm=inorm, type=type), inorm=2-inorm, type=itype), 6e-5)
if type != 1 or len > 1: # there are no length-1 type 1 DCTs
_assert_close(a, pypocketfft.dct(pypocketfft.dct(a, inorm=inorm, type=type), inorm=2-inorm, type=itype), eps)
_assert_close(a, pypocketfft.dst(pypocketfft.dst(a, inorm=inorm, type=type), inorm=2-inorm, type=itype), eps)
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment