11#ifndef EIGEN_COMPLEX_LSX_H
12#define EIGEN_COMPLEX_LSX_H
15#include "../../InternalHeaderCheck.h"
23 EIGEN_STRONG_INLINE Packet2cf() {}
24 EIGEN_STRONG_INLINE
explicit Packet2cf(
const __m128& a) : v(a) {}
29struct packet_traits<std::complex<float> > : default_packet_traits {
30 typedef Packet2cf type;
31 typedef Packet2cf half;
54struct unpacket_traits<Packet2cf> {
55 typedef std::complex<float> type;
56 typedef Packet2cf half;
57 typedef Packet4f as_real;
62 masked_load_available =
false,
63 masked_store_available =
false
68EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(
const Packet2cf& a,
const Packet2cf& b) {
69 return Packet2cf(__lsx_vfadd_s(a.v, b.v));
72EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(
const Packet2cf& a,
const Packet2cf& b) {
73 return Packet2cf(__lsx_vfsub_s(a.v, b.v));
77EIGEN_STRONG_INLINE Packet2cf pnegate(
const Packet2cf& a) {
78 const uint32_t b[4] = {0x80000000u, 0x80000000u, 0x80000000u, 0x80000000u};
79 Packet4i mask = (Packet4i)__lsx_vld(b, 0);
81 res.v = (Packet4f)__lsx_vxor_v((__m128i)a.v, mask);
86EIGEN_STRONG_INLINE Packet2cf pconj(
const Packet2cf& a) {
87 const uint32_t b[4] = {0x00000000u, 0x80000000u, 0x00000000u, 0x80000000u};
88 Packet4i mask = (__m128i)__lsx_vld(b, 0);
90 res.v = (Packet4f)__lsx_vxor_v((__m128i)a.v, mask);
95EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(
const Packet2cf& a,
const Packet2cf& b) {
96 Packet4f part0_tmp = (Packet4f)__lsx_vfmul_s(a.v, b.v);
97 Packet4f part0 = __lsx_vfsub_s(part0_tmp, (__m128)__lsx_vshuf4i_w(part0_tmp, 0x31));
98 Packet4f part1_tmp = __lsx_vfmul_s((__m128)__lsx_vshuf4i_w(a.v, 0xb1), b.v);
99 Packet4f part1 = __lsx_vfadd_s(part1_tmp, (__m128)__lsx_vshuf4i_w(part1_tmp, 0x31));
101 res.v = (Packet4f)__lsx_vpackev_w((__m128i)part1, (__m128i)part0);
106EIGEN_STRONG_INLINE Packet2cf ptrue<Packet2cf>(
const Packet2cf& a) {
107 return Packet2cf(ptrue(Packet4f(a.v)));
111EIGEN_STRONG_INLINE Packet2cf pand<Packet2cf>(
const Packet2cf& a,
const Packet2cf& b) {
113 res.v = (Packet4f)__lsx_vand_v((__m128i)a.v, (__m128i)b.v);
118EIGEN_STRONG_INLINE Packet2cf por<Packet2cf>(
const Packet2cf& a,
const Packet2cf& b) {
120 res.v = (Packet4f)__lsx_vor_v((__m128i)a.v, (__m128i)b.v);
125EIGEN_STRONG_INLINE Packet2cf pxor<Packet2cf>(
const Packet2cf& a,
const Packet2cf& b) {
127 res.v = (Packet4f)__lsx_vxor_v((__m128i)a.v, (__m128i)b.v);
132EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(
const Packet2cf& a,
const Packet2cf& b) {
134 res.v = (Packet4f)__lsx_vandn_v((__m128i)b.v, (__m128i)a.v);
139EIGEN_STRONG_INLINE Packet2cf pload<Packet2cf>(
const std::complex<float>* from) {
140 EIGEN_DEBUG_ALIGNED_LOAD
return Packet2cf(pload<Packet4f>(&numext::real_ref(*from)));
144EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(
const std::complex<float>* from) {
145 EIGEN_DEBUG_UNALIGNED_LOAD
return Packet2cf(ploadu<Packet4f>(&numext::real_ref(*from)));
149EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(
const std::complex<float>& from) {
150 float f0 = from.real(), f1 = from.imag();
151 Packet4f re = {f0, f0, f0, f0};
152 Packet4f im = {f1, f1, f1, f1};
153 return Packet2cf((Packet4f)__lsx_vilvl_w((__m128i)im, (__m128i)re));
157EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(
const std::complex<float>* from) {
158 return pset1<Packet2cf>(*from);
162EIGEN_STRONG_INLINE
void pstore<std::complex<float> >(std::complex<float>* to,
const Packet2cf& from) {
163 EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), Packet4f(from.v));
167EIGEN_STRONG_INLINE
void pstoreu<std::complex<float> >(std::complex<float>* to,
const Packet2cf& from) {
168 EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), Packet4f(from.v));
172EIGEN_DEVICE_FUNC
inline Packet2cf pgather<std::complex<float>, Packet2cf>(
const std::complex<float>* from,
175 __m128i tmp = __lsx_vldrepl_d(from, 0);
176 __m128i tmp1 = __lsx_vldrepl_d(from + stride, 0);
177 tmp = __lsx_vilvl_d(tmp1, tmp);
183EIGEN_DEVICE_FUNC
inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to,
const Packet2cf& from,
185 __lsx_vstelm_d((__m128i)from.v, to, 0, 0);
186 __lsx_vstelm_d((__m128i)from.v, to + stride, 0, 1);
190EIGEN_STRONG_INLINE
void prefetch<std::complex<float> >(
const std::complex<float>* addr) {
191 __builtin_prefetch(addr);
195EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(
const Packet2cf& a) {
196 EIGEN_ALIGN16 std::complex<float> res[2];
197 __lsx_vst(a.v, res, 0);
202EIGEN_STRONG_INLINE Packet2cf preverse(
const Packet2cf& a) {
204 res.v = (Packet4f)__lsx_vshuf4i_w(a.v, 0x4e);
209EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(
const Packet2cf& a) {
210 return pfirst(Packet2cf(__lsx_vfadd_s(a.v, vec4f_movehl(a.v, a.v))));
214EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(
const Packet2cf& a) {
215 return pfirst(pmul(a, Packet2cf(vec4f_movehl(a.v, a.v))));
218EIGEN_STRONG_INLINE Packet2cf pcplxflip (
const Packet2cf& x) {
219 return Packet2cf(vec4f_swizzle1(x.v, 1, 0, 3, 2));
222EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf, Packet4f)
225EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(
const Packet2cf& a,
const Packet2cf& b) {
226 return pdiv_complex(a, b);
230EIGEN_STRONG_INLINE Packet2cf plog<Packet2cf>(
const Packet2cf& a) {
231 return plog_complex(a);
235EIGEN_STRONG_INLINE Packet2cf pzero(
const Packet2cf& ) {
236 __m128 v = {0.0f, 0.0f, 0.0f, 0.0f};
241EIGEN_STRONG_INLINE Packet2cf pmadd<Packet2cf>(
const Packet2cf& a,
const Packet2cf& b,
const Packet2cf& c) {
242 Packet2cf result, t0, t1, t2;
244 t0.v = (__m128)__lsx_vpackev_w((__m128i)a.v, (__m128i)a.v);
245 t2.v = __lsx_vfmadd_s(t0.v, b.v, c.v);
246 result.v = __lsx_vfadd_s(t2.v, t1.v);
247 t1.v = __lsx_vfsub_s(t1.v, a.v);
248 t1.v = (__m128)__lsx_vpackod_w((__m128i)a.v, (__m128i)t1.v);
249 t2.v = (__m128)__lsx_vshuf4i_w((__m128i)b.v, 0xb1);
250 result.v = __lsx_vfmadd_s(t1.v, t2.v, result.v);
255EIGEN_STRONG_INLINE Packet2cf pexp<Packet2cf>(
const Packet2cf& a) {
256 return pexp_complex(a);
261 EIGEN_STRONG_INLINE Packet1cd() {}
262 EIGEN_STRONG_INLINE
explicit Packet1cd(
const __m128d& a) : v(a) {}
267struct packet_traits<std::complex<double> > : default_packet_traits {
268 typedef Packet1cd type;
269 typedef Packet1cd half;
291struct unpacket_traits<Packet1cd> {
292 typedef std::complex<double> type;
293 typedef Packet1cd half;
294 typedef Packet2d as_real;
299 masked_load_available =
false,
300 masked_store_available =
false
305EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(
const Packet1cd& a,
const Packet1cd& b) {
306 return Packet1cd(__lsx_vfadd_d(a.v, b.v));
309EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(
const Packet1cd& a,
const Packet1cd& b) {
310 return Packet1cd(__lsx_vfsub_d(a.v, b.v));
313EIGEN_STRONG_INLINE Packet1cd pnegate(
const Packet1cd& a) {
314 return Packet1cd(pnegate(Packet2d(a.v)));
318EIGEN_STRONG_INLINE Packet1cd pconj(
const Packet1cd& a) {
319 const uint64_t tmp[2] = {0x0000000000000000u, 0x8000000000000000u};
320 __m128i mask = __lsx_vld(tmp, 0);
322 res.v = (Packet2d)__lsx_vxor_v((__m128i)a.v, mask);
327EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(
const Packet1cd& a,
const Packet1cd& b) {
328 Packet2d tmp_real = __lsx_vfmul_d(a.v, b.v);
329 Packet2d
real = __lsx_vfsub_d(tmp_real, preverse(tmp_real));
331 Packet2d tmp_imag = __lsx_vfmul_d(preverse(a.v), b.v);
332 Packet2d
imag = (__m128d)__lsx_vfadd_d((__m128d)tmp_imag, preverse(tmp_imag));
334 res.v = (__m128d)__lsx_vilvl_d((__m128i)
imag, (__m128i)
real);
339EIGEN_STRONG_INLINE Packet1cd ptrue<Packet1cd>(
const Packet1cd& a) {
340 return Packet1cd(ptrue(Packet2d(a.v)));
343EIGEN_STRONG_INLINE Packet1cd pand<Packet1cd>(
const Packet1cd& a,
const Packet1cd& b) {
345 res.v = (Packet2d)__lsx_vand_v((__m128i)a.v, (__m128i)b.v);
349EIGEN_STRONG_INLINE Packet1cd por<Packet1cd>(
const Packet1cd& a,
const Packet1cd& b) {
351 res.v = (Packet2d)__lsx_vor_v((__m128i)a.v, (__m128i)b.v);
355EIGEN_STRONG_INLINE Packet1cd pxor<Packet1cd>(
const Packet1cd& a,
const Packet1cd& b) {
357 res.v = (Packet2d)__lsx_vxor_v((__m128i)a.v, (__m128i)b.v);
361EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(
const Packet1cd& a,
const Packet1cd& b) {
363 res.v = (Packet2d)__lsx_vandn_v((__m128i)b.v, (__m128i)a.v);
369EIGEN_STRONG_INLINE Packet1cd pload<Packet1cd>(
const std::complex<double>* from) {
370 EIGEN_DEBUG_ALIGNED_LOAD
return Packet1cd(pload<Packet2d>((
const double*)from));
373EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(
const std::complex<double>* from) {
374 EIGEN_DEBUG_UNALIGNED_LOAD
return Packet1cd(ploadu<Packet2d>((
const double*)from));
377EIGEN_STRONG_INLINE Packet1cd
378pset1<Packet1cd>(
const std::complex<double>& from) {
379 return ploadu<Packet1cd>(&from);
383EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(
const std::complex<double>* from) {
384 return pset1<Packet1cd>(*from);
389EIGEN_STRONG_INLINE
void pstore<std::complex<double> >(std::complex<double>* to,
const Packet1cd& from) {
390 EIGEN_DEBUG_ALIGNED_STORE pstore((
double*)to, Packet2d(from.v));
393EIGEN_STRONG_INLINE
void pstoreu<std::complex<double> >(std::complex<double>* to,
const Packet1cd& from) {
394 EIGEN_DEBUG_UNALIGNED_STORE pstoreu((
double*)to, Packet2d(from.v));
398EIGEN_STRONG_INLINE
void prefetch<std::complex<double> >(
const std::complex<double>* addr) {
399 __builtin_prefetch(addr);
403EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(
const Packet1cd& a) {
404 EIGEN_ALIGN16
double res[2];
405 __lsx_vst(a.v, res, 0);
406 return std::complex<double>(res[0], res[1]);
410EIGEN_STRONG_INLINE Packet1cd preverse(
const Packet1cd& a) {
415EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(
const Packet1cd& a) {
420EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(
const Packet1cd& a) {
424EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd, Packet2d)
427EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(
const Packet1cd& a,
const Packet1cd& b) {
428 return pdiv_complex(a, b);
431EIGEN_STRONG_INLINE Packet1cd pcplxflip (
const Packet1cd& x) {
432 return Packet1cd(preverse(Packet2d(x.v)));
435EIGEN_DEVICE_FUNC
inline void ptranspose(PacketBlock<Packet2cf, 2>& kernel) {
436 Packet4f tmp1 = (Packet4f)__lsx_vilvl_w((__m128i)kernel.packet[1].v, (__m128i)kernel.packet[0].v);
437 Packet4f tmp2 = (Packet4f)__lsx_vilvh_w((__m128i)kernel.packet[1].v, (__m128i)kernel.packet[0].v);
438 kernel.packet[0].v = (Packet4f)__lsx_vshuf4i_w(tmp1, 0xd8);
439 kernel.packet[1].v = (Packet4f)__lsx_vshuf4i_w(tmp2, 0xd8);
443EIGEN_STRONG_INLINE Packet2cf pcmp_eq(
const Packet2cf& a,
const Packet2cf& b) {
444 Packet4f eq = (Packet4f)__lsx_vfcmp_ceq_s(a.v, b.v);
445 return Packet2cf(pand<Packet4f>(eq, vec4f_swizzle1(eq, 1, 0, 3, 2)));
449EIGEN_STRONG_INLINE Packet1cd pcmp_eq(
const Packet1cd& a,
const Packet1cd& b) {
450 Packet2d eq = (Packet2d)__lsx_vfcmp_ceq_d(a.v, b.v);
451 return Packet1cd(pand<Packet2d>(eq, preverse(eq)));
455EIGEN_DEVICE_FUNC
inline Packet2cf pselect(
const Packet2cf& mask,
const Packet2cf& a,
const Packet2cf& b) {
457 res.v = (Packet4f)__lsx_vbitsel_v((__m128i)b.v, (__m128i)a.v, (__m128i)mask.v);
462EIGEN_STRONG_INLINE Packet1cd psqrt<Packet1cd>(
const Packet1cd& a) {
463 return psqrt_complex<Packet1cd>(a);
467EIGEN_STRONG_INLINE Packet2cf psqrt<Packet2cf>(
const Packet2cf& a) {
468 return psqrt_complex<Packet2cf>(a);
472EIGEN_STRONG_INLINE Packet1cd plog<Packet1cd>(
const Packet1cd& a) {
473 return plog_complex(a);
477EIGEN_STRONG_INLINE Packet1cd pzero<Packet1cd>(
const Packet1cd& ) {
478 __m128d v = {0.0, 0.0};
483EIGEN_STRONG_INLINE Packet1cd pmadd<Packet1cd>(
const Packet1cd& a,
const Packet1cd& b,
const Packet1cd& c) {
484 Packet1cd result, t0, t1, t2;
486 t0.v = (__m128d)__lsx_vpackev_d((__m128i)a.v, (__m128i)a.v);
487 t2.v = __lsx_vfmadd_d(t0.v, b.v, c.v);
488 result.v = __lsx_vfadd_d(t2.v, t1.v);
489 t1.v = __lsx_vfsub_d(t1.v, a.v);
490 t1.v = (__m128d)__lsx_vpackod_d((__m128i)a.v, (__m128i)t1.v);
491 t2.v = (__m128d)__lsx_vshuf4i_d((__m128i)t2.v, (__m128i)b.v, 0xb);
492 result.v = __lsx_vfmadd_d(t1.v, t2.v, result.v);
497EIGEN_DEVICE_FUNC
inline Packet1cd pgather<std::complex<double>, Packet1cd>(
const std::complex<double>* from,
500 __m128i tmp = __lsx_vld((
void*)from, 0);
501 res.v = (__m128d)tmp;
506EIGEN_DEVICE_FUNC
inline void pscatter<std::complex<double>, Packet1cd>(std::complex<double>* to,
const Packet1cd& from,
508 __lsx_vst((__m128i)from.v, (
void*)to, 0);
511EIGEN_STRONG_INLINE
void ptranspose(PacketBlock<Packet1cd, 2>& kernel) {
512 Packet2d tmp = (__m128d)__lsx_vilvl_d((__m128i)kernel.packet[1].v, (__m128i)kernel.packet[0].v);
513 kernel.packet[1].v = (__m128d)__lsx_vilvh_d((__m128i)kernel.packet[1].v, (__m128i)kernel.packet[0].v);
514 kernel.packet[0].v = tmp;
@ Aligned16
Definition Constants.h:237
Namespace containing all symbols from the Eigen library.
Definition B01_Experimental.dox:1
const Eigen::CwiseUnaryOp< Eigen::internal::scalar_real_op< typename Derived::Scalar >, const Derived > real(const Eigen::ArrayBase< Derived > &x)
EIGEN_DEFAULT_DENSE_INDEX_TYPE Index
The Index type as used for the API.
Definition Meta.h:82
const Eigen::CwiseUnaryOp< Eigen::internal::scalar_imag_op< typename Derived::Scalar >, const Derived > imag(const Eigen::ArrayBase< Derived > &x)