11#ifndef EIGEN_COMPLEX32_ALTIVEC_H
12#define EIGEN_COMPLEX32_ALTIVEC_H
15#include "../../InternalHeaderCheck.h"
21inline Packet4ui p4ui_CONJ_XOR() {
22 return vec_mergeh((Packet4ui)p4i_ZERO, (Packet4ui)p4f_MZERO);
24#ifdef EIGEN_VECTORIZE_VSX
25#if defined(_BIG_ENDIAN)
26inline Packet2ul p2ul_CONJ_XOR1() {
27 return (Packet2ul)vec_sld((Packet4ui)p2d_MZERO, (Packet4ui)p2l_ZERO,
30inline Packet2ul p2ul_CONJ_XOR2() {
31 return (Packet2ul)vec_sld((Packet4ui)p2l_ZERO, (Packet4ui)p2d_MZERO,
35inline Packet2ul p2ul_CONJ_XOR1() {
36 return (Packet2ul)vec_sld((Packet4ui)p2l_ZERO, (Packet4ui)p2d_MZERO,
39inline Packet2ul p2ul_CONJ_XOR2() {
40 return (Packet2ul)vec_sld((Packet4ui)p2d_MZERO, (Packet4ui)p2l_ZERO,
48 EIGEN_STRONG_INLINE
explicit Packet2cf() {}
49 EIGEN_STRONG_INLINE
explicit Packet2cf(
const Packet4f& a) : v(a) {}
51 EIGEN_STRONG_INLINE Packet2cf pmul(
const Packet2cf& a,
const Packet2cf& b) {
55 v1 = vec_perm(a.v, a.v, p16uc_PSET32_WODD);
57 v2 = vec_perm(a.v, a.v, p16uc_PSET32_WEVEN);
59 v1 = vec_madd(v1, b.v, p4f_ZERO);
61 v2 = vec_madd(v2, b.v, p4f_ZERO);
62 v2 =
reinterpret_cast<Packet4f
>(pxor(v2,
reinterpret_cast<Packet4f
>(p4ui_CONJ_XOR())));
64 v2 = vec_perm(v2, v2, p16uc_COMPLEX32_REV);
66 return Packet2cf(padd<Packet4f>(v1, v2));
69 EIGEN_STRONG_INLINE Packet2cf& operator*=(
const Packet2cf& b) {
70 v = pmul(Packet2cf(*
this), b).v;
73 EIGEN_STRONG_INLINE Packet2cf operator*(
const Packet2cf& b)
const {
return Packet2cf(*
this) *= b; }
75 EIGEN_STRONG_INLINE Packet2cf& operator+=(
const Packet2cf& b) {
79 EIGEN_STRONG_INLINE Packet2cf operator+(
const Packet2cf& b)
const {
return Packet2cf(*
this) += b; }
80 EIGEN_STRONG_INLINE Packet2cf& operator-=(
const Packet2cf& b) {
84 EIGEN_STRONG_INLINE Packet2cf operator-(
const Packet2cf& b)
const {
return Packet2cf(*
this) -= b; }
85 EIGEN_STRONG_INLINE Packet2cf operator-(
void)
const {
return Packet2cf(-v); }
91struct packet_traits<std::complex<float> > : default_packet_traits {
92 typedef Packet2cf type;
93 typedef Packet2cf half;
94 typedef Packet4f as_real;
112#ifdef EIGEN_VECTORIZE_VSX
120struct unpacket_traits<Packet2cf> {
121 typedef std::complex<float> type;
126 masked_load_available =
false,
127 masked_store_available =
false
129 typedef Packet2cf half;
130 typedef Packet4f as_real;
134EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(
const std::complex<float>& from) {
136#ifdef EIGEN_VECTORIZE_VSX
141 __asm__(
"lxvdsx %x0,%y1" :
"=wa"(res.v) :
"Z"(from));
143 if ((std::ptrdiff_t(&from) % 16) == 0)
144 res.v = pload<Packet4f>((
const float*)&from);
146 res.v = ploadu<Packet4f>((
const float*)&from);
147 res.v = vec_perm(res.v, res.v, p16uc_PSET64_HI);
153EIGEN_STRONG_INLINE Packet2cf pload<Packet2cf>(
const std::complex<float>* from) {
154 return Packet2cf(pload<Packet4f>((
const float*)from));
157EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(
const std::complex<float>* from) {
158 return Packet2cf(ploadu<Packet4f>((
const float*)from));
161EIGEN_ALWAYS_INLINE Packet2cf pload_partial<Packet2cf>(
const std::complex<float>* from,
const Index n,
162 const Index offset) {
163 return Packet2cf(pload_partial<Packet4f>((
const float*)from, n * 2, offset * 2));
166EIGEN_ALWAYS_INLINE Packet2cf ploadu_partial<Packet2cf>(
const std::complex<float>* from,
const Index n,
167 const Index offset) {
168 return Packet2cf(ploadu_partial<Packet4f>((
const float*)from, n * 2, offset * 2));
171EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(
const std::complex<float>* from) {
172 return pset1<Packet2cf>(*from);
176EIGEN_STRONG_INLINE
void pstore<std::complex<float> >(std::complex<float>* to,
const Packet2cf& from) {
177 pstore((
float*)to, from.v);
180EIGEN_STRONG_INLINE
void pstoreu<std::complex<float> >(std::complex<float>* to,
const Packet2cf& from) {
181 pstoreu((
float*)to, from.v);
184EIGEN_ALWAYS_INLINE
void pstore_partial<std::complex<float> >(std::complex<float>* to,
const Packet2cf& from,
186 pstore_partial((
float*)to, from.v, n * 2, offset * 2);
189EIGEN_ALWAYS_INLINE
void pstoreu_partial<std::complex<float> >(std::complex<float>* to,
const Packet2cf& from,
191 pstoreu_partial((
float*)to, from.v, n * 2, offset * 2);
194EIGEN_STRONG_INLINE Packet2cf pload2(
const std::complex<float>& from0,
const std::complex<float>& from1) {
196#ifdef EIGEN_VECTORIZE_VSX
198 __asm__(
"lxsdx %x0,%y1" :
"=wa"(res0) :
"Z"(from0));
199 __asm__(
"lxsdx %x0,%y1" :
"=wa"(res1) :
"Z"(from1));
201 __asm__(
"xxpermdi %x0, %x1, %x2, 0" :
"=wa"(res0) :
"wa"(res0),
"wa"(res1));
203 __asm__(
"xxpermdi %x0, %x2, %x1, 0" :
"=wa"(res0) :
"wa"(res0),
"wa"(res1));
206 *
reinterpret_cast<std::complex<float>*
>(&res0) = from0;
207 *
reinterpret_cast<std::complex<float>*
>(&res1) = from1;
208 res0 = vec_perm(res0, res1, p16uc_TRANSPOSE64_HI);
210 return Packet2cf(res0);
214EIGEN_ALWAYS_INLINE Packet2cf pload_ignore<Packet2cf>(
const std::complex<float>* from) {
216 res.v = pload_ignore<Packet4f>(
reinterpret_cast<const float*
>(from));
220template <
typename Scalar,
typename Packet>
221EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet pgather_complex_size2(
const Scalar* from,
Index stride,
223 eigen_internal_assert(n <= unpacket_traits<Packet>::size &&
"number of elements will gather past end of packet");
224 EIGEN_ALIGN16 Scalar af[2];
225 for (
Index i = 0; i < n; i++) {
226 af[i] = from[i * stride];
228 return pload_ignore<Packet>(af);
231EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet2cf pgather<std::complex<float>, Packet2cf>(
const std::complex<float>* from,
233 return pgather_complex_size2<std::complex<float>, Packet2cf>(from, stride);
236EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet2cf
237pgather_partial<std::complex<float>, Packet2cf>(
const std::complex<float>* from,
Index stride,
const Index n) {
238 return pgather_complex_size2<std::complex<float>, Packet2cf>(from, stride, n);
240template <
typename Scalar,
typename Packet>
241EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
void pscatter_complex_size2(Scalar* to,
const Packet& from,
Index stride,
243 eigen_internal_assert(n <= unpacket_traits<Packet>::size &&
"number of elements will scatter past end of packet");
244 EIGEN_ALIGN16 Scalar af[2];
245 pstore<Scalar>((Scalar*)af, from);
246 for (
Index i = 0; i < n; i++) {
247 to[i * stride] = af[i];
251EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to,
252 const Packet2cf& from,
254 pscatter_complex_size2<std::complex<float>, Packet2cf>(to, from, stride);
257EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
void pscatter_partial<std::complex<float>, Packet2cf>(std::complex<float>* to,
258 const Packet2cf& from,
261 pscatter_complex_size2<std::complex<float>, Packet2cf>(to, from, stride, n);
265EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(
const Packet2cf& a,
const Packet2cf& b) {
266 return Packet2cf(a.v + b.v);
269EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(
const Packet2cf& a,
const Packet2cf& b) {
270 return Packet2cf(a.v - b.v);
273EIGEN_STRONG_INLINE Packet2cf pnegate(
const Packet2cf& a) {
274 return Packet2cf(pnegate(a.v));
277EIGEN_STRONG_INLINE Packet2cf pconj(
const Packet2cf& a) {
278 return Packet2cf(pxor<Packet4f>(a.v,
reinterpret_cast<Packet4f
>(p4ui_CONJ_XOR())));
282EIGEN_STRONG_INLINE Packet2cf pand<Packet2cf>(
const Packet2cf& a,
const Packet2cf& b) {
283 return Packet2cf(pand<Packet4f>(a.v, b.v));
286EIGEN_STRONG_INLINE Packet2cf por<Packet2cf>(
const Packet2cf& a,
const Packet2cf& b) {
287 return Packet2cf(por<Packet4f>(a.v, b.v));
290EIGEN_STRONG_INLINE Packet2cf pxor<Packet2cf>(
const Packet2cf& a,
const Packet2cf& b) {
291 return Packet2cf(pxor<Packet4f>(a.v, b.v));
294EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(
const Packet2cf& a,
const Packet2cf& b) {
295 return Packet2cf(pandnot<Packet4f>(a.v, b.v));
299EIGEN_STRONG_INLINE
void prefetch<std::complex<float> >(
const std::complex<float>* addr) {
300 EIGEN_PPC_PREFETCH(addr);
304EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(
const Packet2cf& a) {
305 EIGEN_ALIGN16 std::complex<float> res[2];
306 pstore((
float*)&res, a.v);
312EIGEN_STRONG_INLINE Packet2cf preverse(
const Packet2cf& a) {
314 rev_a = vec_sld(a.v, a.v, 8);
315 return Packet2cf(rev_a);
319EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(
const Packet2cf& a) {
321 b = vec_sld(a.v, a.v, 8);
322 b = padd<Packet4f>(a.v, b);
323 return pfirst<Packet2cf>(Packet2cf(b));
327EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(
const Packet2cf& a) {
330 b = vec_sld(a.v, a.v, 8);
331 prod = pmul<Packet2cf>(a, Packet2cf(b));
333 return pfirst<Packet2cf>(prod);
336EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf, Packet4f)
339EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(
const Packet2cf& a,
const Packet2cf& b) {
340 return pdiv_complex(a, b);
344EIGEN_STRONG_INLINE Packet2cf pcplxflip<Packet2cf>(
const Packet2cf& x) {
345 return Packet2cf(vec_perm(x.v, x.v, p16uc_COMPLEX32_REV));
348EIGEN_STRONG_INLINE
void ptranspose(PacketBlock<Packet2cf, 2>& kernel) {
349#ifdef EIGEN_VECTORIZE_VSX
350 Packet4f tmp =
reinterpret_cast<Packet4f
>(
351 vec_mergeh(
reinterpret_cast<Packet2d
>(kernel.packet[0].v),
reinterpret_cast<Packet2d
>(kernel.packet[1].v)));
352 kernel.packet[1].v =
reinterpret_cast<Packet4f
>(
353 vec_mergel(
reinterpret_cast<Packet2d
>(kernel.packet[0].v),
reinterpret_cast<Packet2d
>(kernel.packet[1].v)));
355 Packet4f tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI);
356 kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO);
358 kernel.packet[0].v = tmp;
362EIGEN_STRONG_INLINE Packet2cf pcmp_eq(
const Packet2cf& a,
const Packet2cf& b) {
363 Packet4f eq =
reinterpret_cast<Packet4f
>(vec_cmpeq(a.v, b.v));
364 return Packet2cf(vec_and(eq, vec_perm(eq, eq, p16uc_COMPLEX32_REV)));
367#ifdef EIGEN_VECTORIZE_VSX
369EIGEN_STRONG_INLINE Packet2cf pblend(
const Selector<2>& ifPacket,
const Packet2cf& thenPacket,
370 const Packet2cf& elsePacket) {
372 result.v =
reinterpret_cast<Packet4f
>(
373 pblend<Packet2d>(ifPacket,
reinterpret_cast<Packet2d
>(thenPacket.v),
reinterpret_cast<Packet2d
>(elsePacket.v)));
379EIGEN_STRONG_INLINE Packet2cf psqrt<Packet2cf>(
const Packet2cf& a) {
380 return psqrt_complex<Packet2cf>(a);
384EIGEN_STRONG_INLINE Packet2cf plog<Packet2cf>(
const Packet2cf& a) {
385 return plog_complex<Packet2cf>(a);
389EIGEN_STRONG_INLINE Packet2cf pexp<Packet2cf>(
const Packet2cf& a) {
390 return pexp_complex<Packet2cf>(a);
394#ifdef EIGEN_VECTORIZE_VSX
396 EIGEN_STRONG_INLINE Packet1cd() {}
397 EIGEN_STRONG_INLINE
explicit Packet1cd(
const Packet2d& a) : v(a) {}
399 EIGEN_STRONG_INLINE Packet1cd pmul(
const Packet1cd& a,
const Packet1cd& b) {
400 Packet2d a_re, a_im, v1, v2;
403 a_re = vec_perm(a.v, a.v, p16uc_PSET64_HI);
405 a_im = vec_perm(a.v, a.v, p16uc_PSET64_LO);
407 v1 = vec_madd(a_re, b.v, p2d_ZERO);
409 v2 = vec_madd(a_im, b.v, p2d_ZERO);
410 v2 =
reinterpret_cast<Packet2d
>(vec_sld(
reinterpret_cast<Packet4ui
>(v2),
reinterpret_cast<Packet4ui
>(v2), 8));
411 v2 = pxor(v2,
reinterpret_cast<Packet2d
>(p2ul_CONJ_XOR1()));
413 return Packet1cd(padd<Packet2d>(v1, v2));
416 EIGEN_STRONG_INLINE Packet1cd& operator*=(
const Packet1cd& b) {
417 v = pmul(Packet1cd(*
this), b).v;
420 EIGEN_STRONG_INLINE Packet1cd operator*(
const Packet1cd& b)
const {
return Packet1cd(*
this) *= b; }
422 EIGEN_STRONG_INLINE Packet1cd& operator+=(
const Packet1cd& b) {
426 EIGEN_STRONG_INLINE Packet1cd operator+(
const Packet1cd& b)
const {
return Packet1cd(*
this) += b; }
427 EIGEN_STRONG_INLINE Packet1cd& operator-=(
const Packet1cd& b) {
431 EIGEN_STRONG_INLINE Packet1cd operator-(
const Packet1cd& b)
const {
return Packet1cd(*
this) -= b; }
432 EIGEN_STRONG_INLINE Packet1cd operator-(
void)
const {
return Packet1cd(-v); }
438struct packet_traits<std::complex<double> > : default_packet_traits {
439 typedef Packet1cd type;
440 typedef Packet1cd half;
441 typedef Packet2d as_real;
463struct unpacket_traits<Packet1cd> {
464 typedef std::complex<double> type;
469 masked_load_available =
false,
470 masked_store_available =
false
472 typedef Packet1cd half;
473 typedef Packet2d as_real;
477EIGEN_STRONG_INLINE Packet1cd pload<Packet1cd>(
const std::complex<double>* from) {
478 return Packet1cd(pload<Packet2d>((
const double*)from));
481EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(
const std::complex<double>* from) {
482 return Packet1cd(ploadu<Packet2d>((
const double*)from));
485EIGEN_ALWAYS_INLINE Packet1cd pload_partial<Packet1cd>(
const std::complex<double>* from,
const Index n,
486 const Index offset) {
487 return Packet1cd(pload_partial<Packet2d>((
const double*)from, n * 2, offset * 2));
490EIGEN_ALWAYS_INLINE Packet1cd ploadu_partial<Packet1cd>(
const std::complex<double>* from,
const Index n,
491 const Index offset) {
492 return Packet1cd(ploadu_partial<Packet2d>((
const double*)from, n * 2, offset * 2));
495EIGEN_STRONG_INLINE
void pstore<std::complex<double> >(std::complex<double>* to,
const Packet1cd& from) {
496 pstore((
double*)to, from.v);
499EIGEN_STRONG_INLINE
void pstoreu<std::complex<double> >(std::complex<double>* to,
const Packet1cd& from) {
500 pstoreu((
double*)to, from.v);
503EIGEN_ALWAYS_INLINE
void pstore_partial<std::complex<double> >(std::complex<double>* to,
const Packet1cd& from,
505 pstore_partial((
double*)to, from.v, n * 2, offset * 2);
508EIGEN_ALWAYS_INLINE
void pstoreu_partial<std::complex<double> >(std::complex<double>* to,
const Packet1cd& from,
510 pstoreu_partial((
double*)to, from.v, n * 2, offset * 2);
514EIGEN_STRONG_INLINE Packet1cd
515pset1<Packet1cd>(
const std::complex<double>& from) {
516 return ploadu<Packet1cd>(&from);
520EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet1cd
521pgather<std::complex<double>, Packet1cd>(
const std::complex<double>* from,
Index) {
522 return pload<Packet1cd>(from);
525EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet1cd
526pgather_partial<std::complex<double>, Packet1cd>(
const std::complex<double>* from,
Index,
const Index) {
527 return pload<Packet1cd>(from);
530EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
void pscatter<std::complex<double>, Packet1cd>(std::complex<double>* to,
531 const Packet1cd& from,
Index) {
532 pstore<std::complex<double> >(to, from);
535EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
void pscatter_partial<std::complex<double>, Packet1cd>(std::complex<double>* to,
536 const Packet1cd& from,
538 pstore<std::complex<double> >(to, from);
542EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(
const Packet1cd& a,
const Packet1cd& b) {
543 return Packet1cd(a.v + b.v);
546EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(
const Packet1cd& a,
const Packet1cd& b) {
547 return Packet1cd(a.v - b.v);
550EIGEN_STRONG_INLINE Packet1cd pnegate(
const Packet1cd& a) {
551 return Packet1cd(pnegate(Packet2d(a.v)));
554EIGEN_STRONG_INLINE Packet1cd pconj(
const Packet1cd& a) {
555 return Packet1cd(pxor(a.v,
reinterpret_cast<Packet2d
>(p2ul_CONJ_XOR2())));
559EIGEN_STRONG_INLINE Packet1cd pand<Packet1cd>(
const Packet1cd& a,
const Packet1cd& b) {
560 return Packet1cd(pand(a.v, b.v));
563EIGEN_STRONG_INLINE Packet1cd por<Packet1cd>(
const Packet1cd& a,
const Packet1cd& b) {
564 return Packet1cd(por(a.v, b.v));
567EIGEN_STRONG_INLINE Packet1cd pxor<Packet1cd>(
const Packet1cd& a,
const Packet1cd& b) {
568 return Packet1cd(pxor(a.v, b.v));
571EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(
const Packet1cd& a,
const Packet1cd& b) {
572 return Packet1cd(pandnot(a.v, b.v));
576EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(
const std::complex<double>* from) {
577 return pset1<Packet1cd>(*from);
581EIGEN_STRONG_INLINE
void prefetch<std::complex<double> >(
const std::complex<double>* addr) {
582 EIGEN_PPC_PREFETCH(addr);
586EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(
const Packet1cd& a) {
587 EIGEN_ALIGN16 std::complex<double> res[1];
588 pstore<std::complex<double> >(res, a);
594EIGEN_STRONG_INLINE Packet1cd preverse(
const Packet1cd& a) {
599EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(
const Packet1cd& a) {
604EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(
const Packet1cd& a) {
608EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd, Packet2d)
611EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(
const Packet1cd& a,
const Packet1cd& b) {
612 return pdiv_complex(a, b);
615EIGEN_STRONG_INLINE Packet1cd pcplxflip (
const Packet1cd& x) {
616 return Packet1cd(preverse(Packet2d(x.v)));
619EIGEN_STRONG_INLINE
void ptranspose(PacketBlock<Packet1cd, 2>& kernel) {
620 Packet2d tmp = vec_mergeh(kernel.packet[0].v, kernel.packet[1].v);
621 kernel.packet[1].v = vec_mergel(kernel.packet[0].v, kernel.packet[1].v);
622 kernel.packet[0].v = tmp;
626EIGEN_STRONG_INLINE Packet1cd pcmp_eq(
const Packet1cd& a,
const Packet1cd& b) {
629 Packet2d eq =
reinterpret_cast<Packet2d
>(vec_cmpeq(a.v, b.v));
632 Packet2d eq_swapped =
633 reinterpret_cast<Packet2d
>(vec_sld(
reinterpret_cast<Packet4ui
>(eq),
reinterpret_cast<Packet4ui
>(eq), 8));
635 return Packet1cd(vec_and(eq, eq_swapped));
639EIGEN_STRONG_INLINE Packet1cd psqrt<Packet1cd>(
const Packet1cd& a) {
640 return psqrt_complex<Packet1cd>(a);
644EIGEN_STRONG_INLINE Packet1cd plog<Packet1cd>(
const Packet1cd& a) {
645 return plog_complex<Packet1cd>(a);
@ Aligned16
Definition Constants.h:237
Namespace containing all symbols from the Eigen library.
Definition B01_Experimental.dox:1
EIGEN_DEFAULT_DENSE_INDEX_TYPE Index
The Index type as used for the API.
Definition Meta.h:82