Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

convert Complex abs into highway #8

Open
wants to merge 1 commit into
base: highway
Choose a base branch
from
+180 −4
Open

convert Complex abs into highway #8

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 4 additions & 3 deletions numpy/_core/meson.build
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -108,7 +108,8 @@ if use_highway
'src/highway/hwy/abort.cc',
'src/highway/hwy/per_target.cc',
'src/highway/hwy/targets.cc',
'src/highway/hwy/print.cc'
'src/highway/hwy/print.cc',
'src/highway/hwy/aligned_allocator.cc'
],
cpp_args: '-DTOOLCHAIN_MISS_ASM_HWCAP_H',
include_directories: ['src/highway'],
Expand Down Expand Up @@ -989,12 +990,12 @@ foreach gen_mtargets : [
],
[
'loops_unary_complex.dispatch.h',
src_file.process('src/umath/loops_unary_complex.dispatch.c.src'),
'src/umath/loops_unary_complex.dispatch.cpp',
[
AVX512F, [AVX2, FMA3], SSE2,
ASIMD, NEON,
VSX3, VSX2,
VXE, VX,
VXE, VX, RVV
]
],
[
Expand Down
1 change: 0 additions & 1 deletion numpy/_core/src/umath/loops_unary_complex.dispatch.c.src
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -90,7 +90,6 @@ NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(@TYPE@_absolute)
npy_intp len = dimensions[0];
npy_intp ssrc = steps[0] / sizeof(@ftype@);
npy_intp sdst = steps[1] / sizeof(@ftype@);

if (!is_mem_overlap(args[0], steps[0], args[1], steps[1], len) &&
npyv_loadable_stride_@sfx@(ssrc) && npyv_storable_stride_@sfx@(sdst)
&& steps[0] % sizeof(@ftype@) == 0
Expand Down
176 changes: 176 additions & 0 deletions numpy/_core/src/umath/loops_unary_complex.dispatch.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,176 @@
#define _UMATHMODULE
#define _MULTIARRAYMODULE
#define NPY_NO_DEPRECATED_API NPY_API_VERSION

#define PY_SSIZE_T_CLEAN
#include <Python.h>
#include <type_traits>
#include "numpy/ndarraytypes.h"
#include "numpy/npy_common.h"
#include "numpy/npy_math.h"
#include "numpy/utils.h"
#include "fast_loop_macros.h"
#include "loops_utils.h"
#include <hwy/highway.h>
#include <hwy/aligned_allocator.h>
#include <hwy/print-inl.h>

namespace hn = hwy::HWY_NAMESPACE;

inline float Chypot(float x, float y){
return npy_hypotf(x, y);
}

inline double Chypot(double x, double y){
return npy_hypot(x, y);
}

template <typename T>
HWY_ATTR void SuperCabsolute(char** args,
npy_intp const* dimensions,
npy_intp const* steps) {
npy_intp len = dimensions[0];
npy_intp ssrc = steps[0] / sizeof(T);
npy_intp sdst = steps[1] / sizeof(T);
if (!is_mem_overlap(args[0], steps[0], args[1], steps[1], len) &&
steps[0] % sizeof(T) == 0 && steps[1] % sizeof(T) == 0) {
const T* src = (T*)args[0];
T* dst = (T*)args[1];
const hn::ScalableTag<T> d;
const int vstep = hn::Lanes(d);
const int wstep = vstep * 2;
// const int hstep = vstep / 2;

using TI = hwy::MakeSigned<T>;
const hn::Rebind<TI, hn::ScalableTag<T>> di;
auto indices = hwy::AllocateAligned<TI>(vstep);
auto load_index = hn::Mul(hn::Iota(di, 0), hn::Set(di, ssrc));
auto store_index = hn::Mul(hn::Iota(di, 0), hn::Set(di, sdst));

using vec_f = hn::Vec<decltype(d)>;

// TODO: Optimize HWYComplexabsolute this by
// NPY_FINLINE npyv_@sfx@
// simd_cabsolute_@sfx@(npyv_@sfx@ re, npyv_@sfx@ im)
// {
// const npyv_@sfx@ inf = npyv_setall_@sfx@(@INF@);
// const npyv_@sfx@ nan = npyv_setall_@sfx@(@NAN@);

// re = npyv_abs_@sfx@(re);
// im = npyv_abs_@sfx@(im);
// /*
// * If real or imag = INF, then convert it to inf + j*inf
// * Handles: inf + j*nan, nan + j*inf
// */
// npyv_@bsfx@ re_infmask = npyv_cmpeq_@sfx@(re, inf);
// npyv_@bsfx@ im_infmask = npyv_cmpeq_@sfx@(im, inf);
// im = npyv_select_@sfx@(re_infmask, inf, im);
// re = npyv_select_@sfx@(im_infmask, inf, re);
// /*
// * If real or imag = NAN, then convert it to nan + j*nan
// * Handles: x + j*nan, nan + j*x
// */
// npyv_@bsfx@ re_nnanmask = npyv_notnan_@sfx@(re);
// npyv_@bsfx@ im_nnanmask = npyv_notnan_@sfx@(im);
// im = npyv_select_@sfx@(re_nnanmask, im, nan);
// re = npyv_select_@sfx@(im_nnanmask, re, nan);

// npyv_@sfx@ larger = npyv_max_@sfx@(re, im);
// npyv_@sfx@ smaller = npyv_min_@sfx@(im, re);
// /*
// * Calculate div_mask to prevent 0./0. and inf/inf operations in div
// */
// npyv_@bsfx@ zeromask = npyv_cmpeq_@sfx@(larger, npyv_zero_@sfx@());
// npyv_@bsfx@ infmask = npyv_cmpeq_@sfx@(smaller, inf);
// npyv_@bsfx@ div_mask = npyv_not_@bsfx@(npyv_or_@bsfx@(zeromask, infmask));

// npyv_@sfx@ ratio = npyv_ifdivz_@sfx@(div_mask, smaller, larger);
// npyv_@sfx@ hypot = npyv_sqrt_@sfx@(
// npyv_muladd_@sfx@(ratio, ratio, npyv_setall_@sfx@(1.0@c@)
// ));
// return npyv_mul_@sfx@(hypot, larger);
// }
auto HWYComplexabsolute = [d](vec_f re, vec_f im) {
const auto inf = hn::Set(d, std::is_floating_point<T>::value ? NPY_INFINITYF : NPY_INFINITY);
const auto nan = hn::Set(d, std::is_floating_point<T>::value ? NPY_NANF : NPY_NAN);
re = hn::Abs(re);
im = hn::Abs(im);

/*
* If real or imag = INF, then convert it to inf + j*inf
* Handles: inf + j*nan, nan + j*inf
*/
auto re_infmask = hn::IsInf(re);
auto im_infmask = hn::IsInf(im);
im = hn::IfThenElse(re_infmask, inf, im);
re = hn::IfThenElse(im_infmask, inf, re);
/*
* If real or imag = NAN, then convert it to nan + j*nan
* Handles: x + j*nan, nan + j*x
*/
auto re_nanmask = hn::IsNaN(re);
auto im_nanmask = hn::IsNaN(im);
im = hn::IfThenElse(re_nanmask, nan, im);
re = hn::IfThenElse(im_nanmask, nan, re);

auto larger = hn::Max(re, im);
auto smaller = hn::Min(re, im);
/*
* Calculate div_mask to prevent 0./0. and inf/inf operations in div
*/
auto zeromask = hn::Eq(larger, hn::Zero(d));
auto infmask = hn::IsInf(larger);
auto divmask = hn::Not(hn::Or(infmask, zeromask));
auto one = hn::Set(d, 1);
auto div = hn::Div(smaller, hn::IfThenElse(divmask, larger, one));
auto ratio = hn::IfThenElseZero(divmask, div);

auto hypot = hn::Sqrt(hn::MulAdd(ratio, ratio, hn::Set(d, 1.0)));
auto result = hn::Mul(hypot, larger);
return result;
};

if (ssrc == 2 && sdst == 1) {
for (; len >= vstep; len -= vstep, src += wstep, dst += vstep) {
vec_f re, im;
hn::LoadInterleaved2(d, src, re, im);
auto r = HWYComplexabsolute(re, im);
hn::StoreU(r, d, dst);
}
} else {
for (; len >= vstep;
len -= vstep, src += ssrc * vstep, dst += sdst * vstep) {
auto re = hn::GatherIndex(d, src, load_index);
auto im = hn::GatherIndex(d, src + 1, load_index);
auto r = HWYComplexabsolute(re, im);
hn::ScatterIndex(r, d, dst, store_index);
}
}
for (; len > 0; len -= vstep, src += ssrc * vstep, dst += sdst * vstep) {
auto re = hn::GatherIndexN(d, src, load_index, len);
auto im = hn::GatherIndexN(d, src + 1, load_index, len);
auto r = HWYComplexabsolute(re, im);
hn::ScatterIndexN(r, d, dst, store_index, len);
}
} else {
UNARY_LOOP {
const T re = ((T*)ip1)[0];
const T im = ((T*)ip1)[1];
*((T*)op1) = Chypot(re, im);
}
}
}

extern "C" {
NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(CFLOAT_absolute)
(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
{
SuperCabsolute<float>(args, dimensions, steps);
}

NPY_NO_EXPORT void NPY_CPU_DISPATCH_CURFX(CDOUBLE_absolute)
(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
{
SuperCabsolute<double>(args, dimensions, steps);
}
}
Loading