Eigen  5.0.1-dev+284dcc12
 
Loading...
Searching...
No Matches
PacketMath.h
1// This file is part of Eigen, a lightweight C++ template library
2// for linear algebra.
3//
4// Copyright (C) 2016 Konstantinos Margaritis <markos@freevec.org>
5//
6// This Source Code Form is subject to the terms of the Mozilla
7// Public License v. 2.0. If a copy of the MPL was not distributed
8// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
9
10#ifndef EIGEN_PACKET_MATH_ZVECTOR_H
11#define EIGEN_PACKET_MATH_ZVECTOR_H
12
13// IWYU pragma: private
14#include "../../InternalHeaderCheck.h"
15
16namespace Eigen {
17
18namespace internal {
19
20#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
21#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 16
22#endif
23
24#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
25#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD
26#endif
27
28#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
29#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 32
30#endif
31
32typedef __vector int Packet4i;
33typedef __vector unsigned int Packet4ui;
34typedef __vector __bool int Packet4bi;
35typedef __vector short int Packet8i;
36typedef __vector unsigned char Packet16uc;
37typedef __vector double Packet2d;
38typedef __vector unsigned long long Packet2ul;
39typedef __vector long long Packet2l;
40
41// Z14 has builtin support for float vectors
42#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
43typedef __vector float Packet4f;
44#else
45typedef struct {
46 Packet2d v4f[2];
47} Packet4f;
48#endif
49
50typedef union {
51 numext::int32_t i[4];
52 numext::uint32_t ui[4];
53 numext::int64_t l[2];
54 numext::uint64_t ul[2];
55 double d[2];
56 float f[4];
57 Packet4i v4i;
58 Packet4ui v4ui;
59 Packet2l v2l;
60 Packet2ul v2ul;
61 Packet2d v2d;
62#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
63 Packet4f v4f;
64#endif
65} Packet;
66
67// We don't want to write the same code all the time, but we need to reuse the constants
68// and it doesn't really work to declare them global, so we define macros instead
69
70#define EIGEN_DECLARE_CONST_FAST_Packet4i(NAME, X) Packet4i p4i_##NAME = reinterpret_cast<Packet4i>(vec_splat_s32(X))
71
72#define EIGEN_DECLARE_CONST_FAST_Packet2d(NAME, X) Packet2d p2d_##NAME = reinterpret_cast<Packet2d>(vec_splat_s64(X))
73
74#define EIGEN_DECLARE_CONST_FAST_Packet2l(NAME, X) Packet2l p2l_##NAME = reinterpret_cast<Packet2l>(vec_splat_s64(X))
75
76#define EIGEN_DECLARE_CONST_Packet4i(NAME, X) Packet4i p4i_##NAME = pset1<Packet4i>(X)
77
78#define EIGEN_DECLARE_CONST_Packet2d(NAME, X) Packet2d p2d_##NAME = pset1<Packet2d>(X)
79
80#define EIGEN_DECLARE_CONST_Packet2l(NAME, X) Packet2l p2l_##NAME = pset1<Packet2l>(X)
81
82// These constants are endian-agnostic
83static EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); //{ 0, 0, 0, 0,}
84static EIGEN_DECLARE_CONST_FAST_Packet4i(ONE, 1); //{ 1, 1, 1, 1}
85
86static EIGEN_DECLARE_CONST_FAST_Packet2d(ZERO, 0);
87static EIGEN_DECLARE_CONST_FAST_Packet2l(ZERO, 0);
88static EIGEN_DECLARE_CONST_FAST_Packet2l(ONE, 1);
89
90static Packet2d p2d_ONE = {1.0, 1.0};
91static Packet2d p2d_ZERO_ = {numext::bit_cast<double>(0x8000000000000000ull),
92 numext::bit_cast<double>(0x8000000000000000ull)};
93
94#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
95#define EIGEN_DECLARE_CONST_FAST_Packet4f(NAME, X) Packet4f p4f_##NAME = reinterpret_cast<Packet4f>(vec_splat_s32(X))
96
97#define EIGEN_DECLARE_CONST_Packet4f(NAME, X) Packet4f p4f_##NAME = pset1<Packet4f>(X)
98
99#define EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME, X) \
100 const Packet4f p4f_##NAME = reinterpret_cast<Packet4f>(pset1<Packet4i>(X))
101
102static EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0); //{ 0.0, 0.0, 0.0, 0.0}
103static EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1, -1); //{ -1, -1, -1, -1}
104static Packet4f p4f_MZERO = {0x80000000, 0x80000000, 0x80000000, 0x80000000};
105#endif
106
107static Packet4i p4i_COUNTDOWN = {0, 1, 2, 3};
108static Packet4f p4f_COUNTDOWN = {0.0, 1.0, 2.0, 3.0};
109static Packet2d p2d_COUNTDOWN = reinterpret_cast<Packet2d>(
110 vec_sld(reinterpret_cast<Packet16uc>(p2d_ZERO), reinterpret_cast<Packet16uc>(p2d_ONE), 8));
111
112static Packet16uc p16uc_PSET64_HI = {0, 1, 2, 3, 4, 5, 6, 7, 0, 1, 2, 3, 4, 5, 6, 7};
113static Packet16uc p16uc_DUPLICATE32_HI = {0, 1, 2, 3, 0, 1, 2, 3, 4, 5, 6, 7, 4, 5, 6, 7};
114
115// Mask alignment
116#define EIGEN_MASK_ALIGNMENT 0xfffffffffffffff0
117
118#define EIGEN_ALIGNED_PTR(x) ((std::ptrdiff_t)(x) & EIGEN_MASK_ALIGNMENT)
119
120// Handle endianness properly while loading constants
121// Define global static constants:
122
123static Packet16uc p16uc_FORWARD = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15};
124static Packet16uc p16uc_REVERSE32 = {12, 13, 14, 15, 8, 9, 10, 11, 4, 5, 6, 7, 0, 1, 2, 3};
125static Packet16uc p16uc_REVERSE64 = {8, 9, 10, 11, 12, 13, 14, 15, 0, 1, 2, 3, 4, 5, 6, 7};
126
127static Packet16uc p16uc_PSET32_WODD =
128 vec_sld((Packet16uc)vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc)vec_splat((Packet4ui)p16uc_FORWARD, 2),
129 8); //{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 };
130static Packet16uc p16uc_PSET32_WEVEN = vec_sld(p16uc_DUPLICATE32_HI, (Packet16uc)vec_splat((Packet4ui)p16uc_FORWARD, 3),
131 8); //{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 };
132/*static Packet16uc p16uc_HALF64_0_16 = vec_sld((Packet16uc)p4i_ZERO, vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 3),
1338); //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16};
134
135static Packet16uc p16uc_PSET64_HI = (Packet16uc) vec_mergeh((Packet4ui)p16uc_PSET32_WODD,
136(Packet4ui)p16uc_PSET32_WEVEN); //{ 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 };*/
137static Packet16uc p16uc_PSET64_LO = (Packet16uc)vec_mergel(
138 (Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN); //{ 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 };
139/*static Packet16uc p16uc_TRANSPOSE64_HI = vec_add(p16uc_PSET64_HI, p16uc_HALF64_0_16); //{ 0,1,2,3, 4,5,6,7,
14016,17,18,19, 20,21,22,23}; static Packet16uc p16uc_TRANSPOSE64_LO = vec_add(p16uc_PSET64_LO, p16uc_HALF64_0_16); //{
1418,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31};*/
142static Packet16uc p16uc_TRANSPOSE64_HI = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
143static Packet16uc p16uc_TRANSPOSE64_LO = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
144
145static Packet16uc p16uc_COMPLEX32_REV =
146 vec_sld(p16uc_REVERSE32, p16uc_REVERSE32, 8); //{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 };
147
148static Packet16uc p16uc_COMPLEX32_REV2 =
149 vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8); //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
150
151#if EIGEN_HAS_BUILTIN(__builtin_prefetch) || EIGEN_COMP_GNUC
152#define EIGEN_ZVECTOR_PREFETCH(ADDR) __builtin_prefetch(ADDR);
153#else
154#define EIGEN_ZVECTOR_PREFETCH(ADDR) asm(" pfd [%[addr]]\n" ::[addr] "r"(ADDR) : "cc");
155#endif
156
157template <>
158struct packet_traits<int> : default_packet_traits {
159 typedef Packet4i type;
160 typedef Packet4i half;
161 enum {
162 Vectorizable = 1,
163 AlignedOnScalar = 1,
164 size = 4,
165
166 HasAdd = 1,
167 HasSub = 1,
168 HasMul = 1,
169 HasDiv = 1,
170 HasBlend = 1
171 };
172};
173
174template <>
175struct packet_traits<float> : default_packet_traits {
176 typedef Packet4f type;
177 typedef Packet4f half;
178 enum {
179 Vectorizable = 1,
180 AlignedOnScalar = 1,
181 size = 4,
182
183 HasCmp = 1,
184 HasAdd = 1,
185 HasSub = 1,
186 HasMul = 1,
187 HasDiv = 1,
188 HasMin = 1,
189 HasMax = 1,
190 HasAbs = 1,
191 HasSin = 0,
192 HasCos = 0,
193 HasLog = 0,
194 HasExp = 1,
195 HasSqrt = 1,
196 HasRsqrt = 1,
197 HasTanh = 1,
198 HasErf = 1,
199 HasNegate = 1,
200 HasBlend = 1
201 };
202};
203
204template <>
205struct packet_traits<double> : default_packet_traits {
206 typedef Packet2d type;
207 typedef Packet2d half;
208 enum {
209 Vectorizable = 1,
210 AlignedOnScalar = 1,
211 size = 2,
212
213 HasAdd = 1,
214 HasSub = 1,
215 HasMul = 1,
216 HasDiv = 1,
217 HasMin = 1,
218 HasMax = 1,
219 HasAbs = 1,
220 HasSin = 0,
221 HasCos = 0,
222 HasLog = 0,
223 HasExp = 1,
224 HasSqrt = 1,
225 HasRsqrt = 1,
226 HasNegate = 1,
227 HasBlend = 1
228 };
229};
230
231template <>
232struct unpacket_traits<Packet4i> {
233 typedef int type;
234 enum {
235 size = 4,
236 alignment = Aligned16,
237 vectorizable = true,
238 masked_load_available = false,
239 masked_store_available = false
240 };
241 typedef Packet4i half;
242};
243template <>
244struct unpacket_traits<Packet4f> {
245 typedef float type;
246 enum {
247 size = 4,
248 alignment = Aligned16,
249 vectorizable = true,
250 masked_load_available = false,
251 masked_store_available = false
252 };
253 typedef Packet4f half;
254 typedef Packet4i integer_packet;
255};
256template <>
257struct unpacket_traits<Packet2d> {
258 typedef double type;
259 enum {
260 size = 2,
261 alignment = Aligned16,
262 vectorizable = true,
263 masked_load_available = false,
264 masked_store_available = false
265 };
266 typedef Packet2d half;
267 typedef Packet2l integer_packet;
268};
269
270/* Forward declaration */
271EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4f, 4>& kernel);
272
273inline std::ostream& operator<<(std::ostream& s, const Packet4i& v) {
274 Packet vt;
275 vt.v4i = v;
276 s << vt.i[0] << ", " << vt.i[1] << ", " << vt.i[2] << ", " << vt.i[3];
277 return s;
278}
279
280inline std::ostream& operator<<(std::ostream& s, const Packet4ui& v) {
281 Packet vt;
282 vt.v4ui = v;
283 s << vt.ui[0] << ", " << vt.ui[1] << ", " << vt.ui[2] << ", " << vt.ui[3];
284 return s;
285}
286
287inline std::ostream& operator<<(std::ostream& s, const Packet2l& v) {
288 Packet vt;
289 vt.v2l = v;
290 s << vt.l[0] << ", " << vt.l[1];
291 return s;
292}
293
294inline std::ostream& operator<<(std::ostream& s, const Packet2ul& v) {
295 Packet vt;
296 vt.v2ul = v;
297 s << vt.ul[0] << ", " << vt.ul[1];
298 return s;
299}
300
301inline std::ostream& operator<<(std::ostream& s, const Packet2d& v) {
302 Packet vt;
303 vt.v2d = v;
304 s << vt.d[0] << ", " << vt.d[1];
305 return s;
306}
307
308#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
309inline std::ostream& operator<<(std::ostream& s, const Packet4f& v) {
310 Packet vt;
311 vt.v4f = v;
312 s << vt.f[0] << ", " << vt.f[1] << ", " << vt.f[2] << ", " << vt.f[3];
313 return s;
314}
315#endif
316
317template <>
318EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) {
319 EIGEN_DEBUG_ALIGNED_LOAD
320 return vec_xl(0, from);
321}
322
323template <>
324EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from) {
325 EIGEN_DEBUG_ALIGNED_LOAD
326 return vec_xl(0, from);
327}
328
329template <>
330EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet4i& from) {
331 EIGEN_DEBUG_ALIGNED_STORE
332 vec_xst(from, 0, to);
333}
334
335template <>
336EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet2d& from) {
337 EIGEN_DEBUG_ALIGNED_STORE
338 vec_xst(from, 0, to);
339}
340
341template <>
342EIGEN_STRONG_INLINE Packet4f pfrexp<Packet4f>(const Packet4f& a, Packet4f& exponent) {
343 return pfrexp_generic(a, exponent);
344}
345
346template <>
347EIGEN_STRONG_INLINE Packet2d pfrexp<Packet2d>(const Packet2d& a, Packet2d& exponent) {
348 return pfrexp_generic(a, exponent);
349}
350
351template <>
352EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) {
353 return vec_splats(from);
354}
355template <>
356EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) {
357 return vec_splats(from);
358}
359
360template <>
361EIGEN_STRONG_INLINE void pbroadcast4<Packet4i>(const int* a, Packet4i& a0, Packet4i& a1, Packet4i& a2, Packet4i& a3) {
362 a3 = pload<Packet4i>(a);
363 a0 = vec_splat(a3, 0);
364 a1 = vec_splat(a3, 1);
365 a2 = vec_splat(a3, 2);
366 a3 = vec_splat(a3, 3);
367}
368
369template <>
370EIGEN_STRONG_INLINE void pbroadcast4<Packet2d>(const double* a, Packet2d& a0, Packet2d& a1, Packet2d& a2,
371 Packet2d& a3) {
372 a1 = pload<Packet2d>(a);
373 a0 = vec_splat(a1, 0);
374 a1 = vec_splat(a1, 1);
375 a3 = pload<Packet2d>(a + 2);
376 a2 = vec_splat(a3, 0);
377 a3 = vec_splat(a3, 1);
378}
379
380template <>
381EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, Index stride) {
382 EIGEN_ALIGN16 int ai[4];
383 ai[0] = from[0 * stride];
384 ai[1] = from[1 * stride];
385 ai[2] = from[2 * stride];
386 ai[3] = from[3 * stride];
387 return pload<Packet4i>(ai);
388}
389
390template <>
391EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride) {
392 EIGEN_ALIGN16 double af[2];
393 af[0] = from[0 * stride];
394 af[1] = from[1 * stride];
395 return pload<Packet2d>(af);
396}
397
398template <>
399EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, Index stride) {
400 EIGEN_ALIGN16 int ai[4];
401 pstore<int>((int*)ai, from);
402 to[0 * stride] = ai[0];
403 to[1 * stride] = ai[1];
404 to[2 * stride] = ai[2];
405 to[3 * stride] = ai[3];
406}
407
408template <>
409EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride) {
410 EIGEN_ALIGN16 double af[2];
411 pstore<double>(af, from);
412 to[0 * stride] = af[0];
413 to[1 * stride] = af[1];
414}
415
416template <>
417EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) {
418 return (a + b);
419}
420template <>
421EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) {
422 return (a + b);
423}
424
425template <>
426EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) {
427 return (a - b);
428}
429template <>
430EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) {
431 return (a - b);
432}
433
434template <>
435EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) {
436 return (a * b);
437}
438template <>
439EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) {
440 return (a * b);
441}
442
443template <>
444EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& a, const Packet4i& b) {
445 return (a / b);
446}
447template <>
448EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) {
449 return (a / b);
450}
451
452template <>
453EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) {
454 return (-a);
455}
456template <>
457EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) {
458 return (-a);
459}
460
461template <>
462EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) {
463 return a;
464}
465template <>
466EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) {
467 return a;
468}
469
470template <>
471EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) {
472 return padd<Packet4i>(pmul<Packet4i>(a, b), c);
473}
474template <>
475EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) {
476 return vec_madd(a, b, c);
477}
478
479template <>
480EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a) {
481 return padd<Packet4i>(pset1<Packet4i>(a), p4i_COUNTDOWN);
482}
483template <>
484EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) {
485 return padd<Packet2d>(pset1<Packet2d>(a), p2d_COUNTDOWN);
486}
487
488template <>
489EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) {
490 return vec_min(a, b);
491}
492template <>
493EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) {
494 return vec_min(a, b);
495}
496
497template <>
498EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) {
499 return vec_max(a, b);
500}
501template <>
502EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) {
503 return vec_max(a, b);
504}
505
506template <>
507EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) {
508 return vec_and(a, b);
509}
510template <>
511EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) {
512 return vec_and(a, b);
513}
514
515template <>
516EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) {
517 return vec_or(a, b);
518}
519template <>
520EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) {
521 return vec_or(a, b);
522}
523
524template <>
525EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) {
526 return vec_xor(a, b);
527}
528template <>
529EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) {
530 return vec_xor(a, b);
531}
532
533template <>
534EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) {
535 return pand<Packet4i>(a, vec_nor(b, b));
536}
537template <>
538EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) {
539 return vec_and(a, vec_nor(b, b));
540}
541
542template <>
543EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) {
544 /* Uses non-default rounding for vec_round */
545 return __builtin_s390_vfidb(a, 0, 1);
546}
547template <>
548EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const Packet2d& a) {
549 return vec_ceil(a);
550}
551template <>
552EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) {
553 return vec_floor(a);
554}
555
556template <>
557EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) {
558 return pload<Packet4i>(from);
559}
560template <>
561EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) {
562 return pload<Packet2d>(from);
563}
564
565template <>
566EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int* from) {
567 Packet4i p = pload<Packet4i>(from);
568 return vec_perm(p, p, p16uc_DUPLICATE32_HI);
569}
570
571template <>
572EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double* from) {
573 Packet2d p = pload<Packet2d>(from);
574 return vec_perm(p, p, p16uc_PSET64_HI);
575}
576
577template <>
578EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from) {
579 pstore<int>(to, from);
580}
581template <>
582EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) {
583 pstore<double>(to, from);
584}
585
586template <>
587EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) {
588 EIGEN_ZVECTOR_PREFETCH(addr);
589}
590template <>
591EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) {
592 EIGEN_ZVECTOR_PREFETCH(addr);
593}
594
595template <int N>
596EIGEN_STRONG_INLINE Packet2l parithmetic_shift_right(const Packet2l& a) {
597 return Packet2l { parithmetic_shift_right<N>(a[0]), parithmetic_shift_right<N>(a[1]) };
598}
599template <int N>
600EIGEN_STRONG_INLINE Packet4i parithmetic_shift_right(const Packet4i& a) {
601 return Packet4i {
602 parithmetic_shift_right<N>(a[0]),
603 parithmetic_shift_right<N>(a[1]),
604 parithmetic_shift_right<N>(a[2]),
605 parithmetic_shift_right<N>(a[3]) };
606}
607
608template <int N>
609EIGEN_STRONG_INLINE Packet2l plogical_shift_right(const Packet2l& a) {
610 return Packet2l { plogical_shift_right<N>(a[0]), plogical_shift_right<N>(a[1]) };
611}
612template <int N>
613EIGEN_STRONG_INLINE Packet4i plogical_shift_right(const Packet4i& a) {
614 return Packet4i {
615 plogical_shift_right<N>(a[0]),
616 plogical_shift_right<N>(a[1]),
617 plogical_shift_right<N>(a[2]),
618 plogical_shift_right<N>(a[3]) };
619}
620
621template <int N>
622EIGEN_STRONG_INLINE Packet2l plogical_shift_left(const Packet2l& a) {
623 return Packet2l { plogical_shift_left<N>(a[0]), plogical_shift_left<N>(a[1]) };
624}
625template <int N>
626EIGEN_STRONG_INLINE Packet4i plogical_shift_left(const Packet4i& a) {
627 return Packet4i {
628 plogical_shift_left<N>(a[0]),
629 plogical_shift_left<N>(a[1]),
630 plogical_shift_left<N>(a[2]),
631 plogical_shift_left<N>(a[3]) };
632}
633
634template <>
635EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) {
636 EIGEN_ALIGN16 int x[4];
637 pstore(x, a);
638 return x[0];
639}
640template <>
641EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) {
642 EIGEN_ALIGN16 double x[2];
643 pstore(x, a);
644 return x[0];
645}
646
647template <>
648EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) {
649 return reinterpret_cast<Packet4i>(
650 vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE32));
651}
652
653template <>
654EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) {
655 return reinterpret_cast<Packet2d>(
656 vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE64));
657}
658
659template <>
660EIGEN_STRONG_INLINE Packet4i pabs<Packet4i>(const Packet4i& a) {
661 return vec_abs(a);
662}
663template <>
664EIGEN_STRONG_INLINE Packet2d pabs<Packet2d>(const Packet2d& a) {
665 return vec_abs(a);
666}
667
668template <>
669EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a) {
670 Packet4i b, sum;
671 b = vec_sld(a, a, 8);
672 sum = padd<Packet4i>(a, b);
673 b = vec_sld(sum, sum, 4);
674 sum = padd<Packet4i>(sum, b);
675 return pfirst(sum);
676}
677
678template <>
679EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) {
680 Packet2d b, sum;
681 b = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8));
682 sum = padd<Packet2d>(a, b);
683 return pfirst(sum);
684}
685
686// Other reduction functions:
687// mul
688template <>
689EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a) {
690 EIGEN_ALIGN16 int aux[4];
691 pstore(aux, a);
692 return aux[0] * aux[1] * aux[2] * aux[3];
693}
694
695template <>
696EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) {
697 return pfirst(
698 pmul(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8))));
699}
700
701// min
702template <>
703EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a) {
704 Packet4i b, res;
705 b = pmin<Packet4i>(a, vec_sld(a, a, 8));
706 res = pmin<Packet4i>(b, vec_sld(b, b, 4));
707 return pfirst(res);
708}
709
710template <>
711EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a) {
712 return pfirst(pmin<Packet2d>(
713 a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8))));
714}
715
716// max
717template <>
718EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a) {
719 Packet4i b, res;
720 b = pmax<Packet4i>(a, vec_sld(a, a, 8));
721 res = pmax<Packet4i>(b, vec_sld(b, b, 4));
722 return pfirst(res);
723}
724
725// max
726template <>
727EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a) {
728 return pfirst(pmax<Packet2d>(
729 a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8))));
730}
731
732EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4i, 4>& kernel) {
733 Packet4i t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]);
734 Packet4i t1 = vec_mergel(kernel.packet[0], kernel.packet[2]);
735 Packet4i t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]);
736 Packet4i t3 = vec_mergel(kernel.packet[1], kernel.packet[3]);
737 kernel.packet[0] = vec_mergeh(t0, t2);
738 kernel.packet[1] = vec_mergel(t0, t2);
739 kernel.packet[2] = vec_mergeh(t1, t3);
740 kernel.packet[3] = vec_mergel(t1, t3);
741}
742
743EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet2d, 2>& kernel) {
744 Packet2d t0 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_HI);
745 Packet2d t1 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_LO);
746 kernel.packet[0] = t0;
747 kernel.packet[1] = t1;
748}
749
750template <>
751EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket,
752 const Packet4i& elsePacket) {
753 Packet4ui select = {ifPacket.select[0], ifPacket.select[1], ifPacket.select[2], ifPacket.select[3]};
754 Packet4ui mask = vec_cmpeq(select, reinterpret_cast<Packet4ui>(p4i_ONE));
755 return vec_sel(elsePacket, thenPacket, mask);
756}
757
758template <>
759EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket,
760 const Packet2d& elsePacket) {
761 Packet2ul select = {ifPacket.select[0], ifPacket.select[1]};
762 Packet2ul mask = vec_cmpeq(select, reinterpret_cast<Packet2ul>(p2l_ONE));
763 return vec_sel(elsePacket, thenPacket, mask);
764}
765
766/* z13 has no vector float support so we emulate that with double
767 z14 has proper vector float support.
768*/
769#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ < 12)
770/* Helper function to simulate a vec_splat_packet4f
771 */
772template <int element>
773EIGEN_STRONG_INLINE Packet4f vec_splat_packet4f(const Packet4f& from) {
774 Packet4f splat;
775 switch (element) {
776 case 0:
777 splat.v4f[0] = vec_splat(from.v4f[0], 0);
778 splat.v4f[1] = splat.v4f[0];
779 break;
780 case 1:
781 splat.v4f[0] = vec_splat(from.v4f[0], 1);
782 splat.v4f[1] = splat.v4f[0];
783 break;
784 case 2:
785 splat.v4f[0] = vec_splat(from.v4f[1], 0);
786 splat.v4f[1] = splat.v4f[0];
787 break;
788 case 3:
789 splat.v4f[0] = vec_splat(from.v4f[1], 1);
790 splat.v4f[1] = splat.v4f[0];
791 break;
792 }
793 return splat;
794}
795
796template <>
797EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) {
798 // FIXME: No intrinsic yet
799 EIGEN_DEBUG_ALIGNED_LOAD
800 Packet4f vfrom;
801 vfrom.v4f[0] = vec_ld2f(&from[0]);
802 vfrom.v4f[1] = vec_ld2f(&from[2]);
803 return vfrom;
804}
805
806template <>
807EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) {
808 // FIXME: No intrinsic yet
809 EIGEN_DEBUG_ALIGNED_STORE
810 vec_st2f(from.v4f[0], &to[0]);
811 vec_st2f(from.v4f[1], &to[2]);
812}
813
814template <>
815EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) {
816 Packet4f to;
817 to.v4f[0] = pset1<Packet2d>(static_cast<const double&>(from));
818 to.v4f[1] = to.v4f[0];
819 return to;
820}
821
822template <>
823EIGEN_STRONG_INLINE void pbroadcast4<Packet4f>(const float* a, Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3) {
824 a3 = pload<Packet4f>(a);
825 a0 = vec_splat_packet4f<0>(a3);
826 a1 = vec_splat_packet4f<1>(a3);
827 a2 = vec_splat_packet4f<2>(a3);
828 a3 = vec_splat_packet4f<3>(a3);
829}
830
831template <>
832EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride) {
833 EIGEN_ALIGN16 float ai[4];
834 ai[0] = from[0 * stride];
835 ai[1] = from[1 * stride];
836 ai[2] = from[2 * stride];
837 ai[3] = from[3 * stride];
838 return pload<Packet4f>(ai);
839}
840
841template <>
842EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride) {
843 EIGEN_ALIGN16 float ai[4];
844 pstore<float>((float*)ai, from);
845 to[0 * stride] = ai[0];
846 to[1 * stride] = ai[1];
847 to[2 * stride] = ai[2];
848 to[3 * stride] = ai[3];
849}
850
851template <>
852EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) {
853 Packet4f c;
854 c.v4f[0] = a.v4f[0] + b.v4f[0];
855 c.v4f[1] = a.v4f[1] + b.v4f[1];
856 return c;
857}
858
859template <>
860EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) {
861 Packet4f c;
862 c.v4f[0] = a.v4f[0] - b.v4f[0];
863 c.v4f[1] = a.v4f[1] - b.v4f[1];
864 return c;
865}
866
867template <>
868EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) {
869 Packet4f c;
870 c.v4f[0] = a.v4f[0] * b.v4f[0];
871 c.v4f[1] = a.v4f[1] * b.v4f[1];
872 return c;
873}
874
875template <>
876EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) {
877 Packet4f c;
878 c.v4f[0] = a.v4f[0] / b.v4f[0];
879 c.v4f[1] = a.v4f[1] / b.v4f[1];
880 return c;
881}
882
883template <>
884EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) {
885 Packet4f c;
886 c.v4f[0] = -a.v4f[0];
887 c.v4f[1] = -a.v4f[1];
888 return c;
889}
890
891template <>
892EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) {
893 Packet4f res;
894 res.v4f[0] = vec_madd(a.v4f[0], b.v4f[0], c.v4f[0]);
895 res.v4f[1] = vec_madd(a.v4f[1], b.v4f[1], c.v4f[1]);
896 return res;
897}
898
899template <>
900EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) {
901 Packet4f res;
902 res.v4f[0] = pmin(a.v4f[0], b.v4f[0]);
903 res.v4f[1] = pmin(a.v4f[1], b.v4f[1]);
904 return res;
905}
906
907template <>
908EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) {
909 Packet4f res;
910 res.v4f[0] = pmax(a.v4f[0], b.v4f[0]);
911 res.v4f[1] = pmax(a.v4f[1], b.v4f[1]);
912 return res;
913}
914
915template <>
916EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) {
917 Packet4f res;
918 res.v4f[0] = pand(a.v4f[0], b.v4f[0]);
919 res.v4f[1] = pand(a.v4f[1], b.v4f[1]);
920 return res;
921}
922
923template <>
924EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) {
925 Packet4f res;
926 res.v4f[0] = por(a.v4f[0], b.v4f[0]);
927 res.v4f[1] = por(a.v4f[1], b.v4f[1]);
928 return res;
929}
930
931template <>
932EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) {
933 Packet4f res;
934 res.v4f[0] = pxor(a.v4f[0], b.v4f[0]);
935 res.v4f[1] = pxor(a.v4f[1], b.v4f[1]);
936 return res;
937}
938
939template <>
940EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) {
941 Packet4f res;
942 res.v4f[0] = pandnot(a.v4f[0], b.v4f[0]);
943 res.v4f[1] = pandnot(a.v4f[1], b.v4f[1]);
944 return res;
945}
946
947template <>
948EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) {
949 Packet4f res;
950 res.v4f[0] = generic_round(a.v4f[0]);
951 res.v4f[1] = generic_round(a.v4f[1]);
952 return res;
953}
954
955template <>
956EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const Packet4f& a) {
957 Packet4f res;
958 res.v4f[0] = vec_ceil(a.v4f[0]);
959 res.v4f[1] = vec_ceil(a.v4f[1]);
960 return res;
961}
962
963template <>
964EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) {
965 Packet4f res;
966 res.v4f[0] = vec_floor(a.v4f[0]);
967 res.v4f[1] = vec_floor(a.v4f[1]);
968 return res;
969}
970
971template <>
972EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from) {
973 Packet4f p = pload<Packet4f>(from);
974 p.v4f[1] = vec_splat(p.v4f[0], 1);
975 p.v4f[0] = vec_splat(p.v4f[0], 0);
976 return p;
977}
978
979template <>
980EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) {
981 EIGEN_ALIGN16 float x[2];
982 vec_st2f(a.v4f[0], &x[0]);
983 return x[0];
984}
985
986template <>
987EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) {
988 Packet4f rev;
989 rev.v4f[0] = preverse<Packet2d>(a.v4f[1]);
990 rev.v4f[1] = preverse<Packet2d>(a.v4f[0]);
991 return rev;
992}
993
994template <>
995EIGEN_STRONG_INLINE Packet4f pabs<Packet4f>(const Packet4f& a) {
996 Packet4f res;
997 res.v4f[0] = pabs(a.v4f[0]);
998 res.v4f[1] = pabs(a.v4f[1]);
999 return res;
1000}
1001
1002template <>
1003EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) {
1004 Packet2d sum;
1005 sum = padd<Packet2d>(a.v4f[0], a.v4f[1]);
1006 double first = predux<Packet2d>(sum);
1007 return static_cast<float>(first);
1008}
1009
1010template <>
1011EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a) {
1012 // Return predux_mul<Packet2d> of the subvectors product
1013 return static_cast<float>(pfirst(predux_mul(pmul(a.v4f[0], a.v4f[1]))));
1014}
1015
1016template <>
1017EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a) {
1018 Packet2d b, res;
1019 b = pmin<Packet2d>(a.v4f[0], a.v4f[1]);
1020 res = pmin<Packet2d>(
1021 b, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(b), reinterpret_cast<Packet4i>(b), 8)));
1022 return static_cast<float>(pfirst(res));
1023}
1024
1025template <>
1026EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a) {
1027 Packet2d b, res;
1028 b = pmax<Packet2d>(a.v4f[0], a.v4f[1]);
1029 res = pmax<Packet2d>(
1030 b, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(b), reinterpret_cast<Packet4i>(b), 8)));
1031 return static_cast<float>(pfirst(res));
1032}
1033
1034/* Split the Packet4f PacketBlock into 4 Packet2d PacketBlocks and transpose each one
1035 */
1036EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4f, 4>& kernel) {
1037 PacketBlock<Packet2d, 2> t0, t1, t2, t3;
1038 // copy top-left 2x2 Packet2d block
1039 t0.packet[0] = kernel.packet[0].v4f[0];
1040 t0.packet[1] = kernel.packet[1].v4f[0];
1041
1042 // copy top-right 2x2 Packet2d block
1043 t1.packet[0] = kernel.packet[0].v4f[1];
1044 t1.packet[1] = kernel.packet[1].v4f[1];
1045
1046 // copy bottom-left 2x2 Packet2d block
1047 t2.packet[0] = kernel.packet[2].v4f[0];
1048 t2.packet[1] = kernel.packet[3].v4f[0];
1049
1050 // copy bottom-right 2x2 Packet2d block
1051 t3.packet[0] = kernel.packet[2].v4f[1];
1052 t3.packet[1] = kernel.packet[3].v4f[1];
1053
1054 // Transpose all 2x2 blocks
1055 ptranspose(t0);
1056 ptranspose(t1);
1057 ptranspose(t2);
1058 ptranspose(t3);
1059
1060 // Copy back transposed blocks, but exchange t1 and t2 due to transposition
1061 kernel.packet[0].v4f[0] = t0.packet[0];
1062 kernel.packet[0].v4f[1] = t2.packet[0];
1063 kernel.packet[1].v4f[0] = t0.packet[1];
1064 kernel.packet[1].v4f[1] = t2.packet[1];
1065 kernel.packet[2].v4f[0] = t1.packet[0];
1066 kernel.packet[2].v4f[1] = t3.packet[0];
1067 kernel.packet[3].v4f[0] = t1.packet[1];
1068 kernel.packet[3].v4f[1] = t3.packet[1];
1069}
1070
1071template <>
1072EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket,
1073 const Packet4f& elsePacket) {
1074 Packet2ul select_hi = {ifPacket.select[0], ifPacket.select[1]};
1075 Packet2ul select_lo = {ifPacket.select[2], ifPacket.select[3]};
1076 Packet2ul mask_hi = vec_cmpeq(select_hi, reinterpret_cast<Packet2ul>(p2l_ONE));
1077 Packet2ul mask_lo = vec_cmpeq(select_lo, reinterpret_cast<Packet2ul>(p2l_ONE));
1078 Packet4f result;
1079 result.v4f[0] = vec_sel(elsePacket.v4f[0], thenPacket.v4f[0], mask_hi);
1080 result.v4f[1] = vec_sel(elsePacket.v4f[1], thenPacket.v4f[1], mask_lo);
1081 return result;
1082}
1083
1084template <>
1085Packet4f EIGEN_STRONG_INLINE pcmp_le<Packet4f>(const Packet4f& a, const Packet4f& b) {
1086 Packet4f res;
1087 res.v4f[0] = pcmp_le(a.v4f[0], b.v4f[0]);
1088 res.v4f[1] = pcmp_le(a.v4f[1], b.v4f[1]);
1089 return res;
1090}
1091
1092template <>
1093Packet4f EIGEN_STRONG_INLINE pcmp_lt<Packet4f>(const Packet4f& a, const Packet4f& b) {
1094 Packet4f res;
1095 res.v4f[0] = pcmp_lt(a.v4f[0], b.v4f[0]);
1096 res.v4f[1] = pcmp_lt(a.v4f[1], b.v4f[1]);
1097 return res;
1098}
1099
1100template <>
1101Packet4f EIGEN_STRONG_INLINE pcmp_eq<Packet4f>(const Packet4f& a, const Packet4f& b) {
1102 Packet4f res;
1103 res.v4f[0] = pcmp_eq(a.v4f[0], b.v4f[0]);
1104 res.v4f[1] = pcmp_eq(a.v4f[1], b.v4f[1]);
1105 return res;
1106}
1107
1108#else
1109template <>
1110EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) {
1111 EIGEN_DEBUG_ALIGNED_LOAD
1112 return vec_xl(0, from);
1113}
1114
1115template <>
1116EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) {
1117 EIGEN_DEBUG_ALIGNED_STORE
1118 vec_xst(from, 0, to);
1119}
1120
1121template <>
1122EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) {
1123 return vec_splats(from);
1124}
1125
1126template <>
1127EIGEN_STRONG_INLINE void pbroadcast4<Packet4f>(const float* a, Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3) {
1128 a3 = pload<Packet4f>(a);
1129 a0 = vec_splat(a3, 0);
1130 a1 = vec_splat(a3, 1);
1131 a2 = vec_splat(a3, 2);
1132 a3 = vec_splat(a3, 3);
1133}
1134
1135template <>
1136EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride) {
1137 EIGEN_ALIGN16 float af[4];
1138 af[0] = from[0 * stride];
1139 af[1] = from[1 * stride];
1140 af[2] = from[2 * stride];
1141 af[3] = from[3 * stride];
1142 return pload<Packet4f>(af);
1143}
1144
1145template <>
1146EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride) {
1147 EIGEN_ALIGN16 float af[4];
1148 pstore<float>((float*)af, from);
1149 to[0 * stride] = af[0];
1150 to[1 * stride] = af[1];
1151 to[2 * stride] = af[2];
1152 to[3 * stride] = af[3];
1153}
1154
1155template <>
1156EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) {
1157 return (a + b);
1158}
1159template <>
1160EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) {
1161 return (a - b);
1162}
1163template <>
1164EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) {
1165 return (a * b);
1166}
1167template <>
1168EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) {
1169 return (a / b);
1170}
1171template <>
1172EIGEN_STRONG_INLINE Packet4f pnegate<Packet4f>(const Packet4f& a) {
1173 return (-a);
1174}
1175template <>
1176EIGEN_STRONG_INLINE Packet4f pconj<Packet4f>(const Packet4f& a) {
1177 return a;
1178}
1179template <>
1180EIGEN_STRONG_INLINE Packet4f pmadd<Packet4f>(const Packet4f& a, const Packet4f& b, const Packet4f& c) {
1181 return vec_madd(a, b, c);
1182}
1183template <>
1184EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) {
1185 return vec_min(a, b);
1186}
1187template <>
1188EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) {
1189 return vec_max(a, b);
1190}
1191template <>
1192EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) {
1193 return vec_and(a, b);
1194}
1195template <>
1196EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) {
1197 return vec_or(a, b);
1198}
1199template <>
1200EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) {
1201 return vec_xor(a, b);
1202}
1203template <>
1204EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) {
1205 return vec_and(a, vec_nor(b, b));
1206}
1207template <>
1208EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) {
1209 /* Uses non-default rounding for vec_round */
1210 return __builtin_s390_vfisb(a, 0, 1);
1211}
1212template <>
1213EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const Packet4f& a) {
1214 return vec_ceil(a);
1215}
1216template <>
1217EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) {
1218 return vec_floor(a);
1219}
1220template <>
1221EIGEN_STRONG_INLINE Packet4f pabs<Packet4f>(const Packet4f& a) {
1222 return vec_abs(a);
1223}
1224template <>
1225EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) {
1226 EIGEN_ALIGN16 float x[4];
1227 pstore(x, a);
1228 return x[0];
1229}
1230
1231template <>
1232EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from) {
1233 Packet4f p = pload<Packet4f>(from);
1234 return vec_perm(p, p, p16uc_DUPLICATE32_HI);
1235}
1236
1237template <>
1238EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) {
1239 return reinterpret_cast<Packet4f>(
1240 vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE32));
1241}
1242
1243template <>
1244EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) {
1245 Packet4f b, sum;
1246 b = vec_sld(a, a, 8);
1247 sum = padd<Packet4f>(a, b);
1248 b = vec_sld(sum, sum, 4);
1249 sum = padd<Packet4f>(sum, b);
1250 return pfirst(sum);
1251}
1252
1253// Other reduction functions:
1254// mul
1255template <>
1256EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a) {
1257 Packet4f prod;
1258 prod = pmul(a, vec_sld(a, a, 8));
1259 return pfirst(pmul(prod, vec_sld(prod, prod, 4)));
1260}
1261
1262// min
1263template <>
1264EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a) {
1265 Packet4f b, res;
1266 b = pmin<Packet4f>(a, vec_sld(a, a, 8));
1267 res = pmin<Packet4f>(b, vec_sld(b, b, 4));
1268 return pfirst(res);
1269}
1270
1271// max
1272template <>
1273EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a) {
1274 Packet4f b, res;
1275 b = pmax<Packet4f>(a, vec_sld(a, a, 8));
1276 res = pmax<Packet4f>(b, vec_sld(b, b, 4));
1277 return pfirst(res);
1278}
1279
1280EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4f, 4>& kernel) {
1281 Packet4f t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]);
1282 Packet4f t1 = vec_mergel(kernel.packet[0], kernel.packet[2]);
1283 Packet4f t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]);
1284 Packet4f t3 = vec_mergel(kernel.packet[1], kernel.packet[3]);
1285 kernel.packet[0] = vec_mergeh(t0, t2);
1286 kernel.packet[1] = vec_mergel(t0, t2);
1287 kernel.packet[2] = vec_mergeh(t1, t3);
1288 kernel.packet[3] = vec_mergel(t1, t3);
1289}
1290
1291template <>
1292EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket,
1293 const Packet4f& elsePacket) {
1294 Packet4ui select = {ifPacket.select[0], ifPacket.select[1], ifPacket.select[2], ifPacket.select[3]};
1295 Packet4ui mask = vec_cmpeq(select, reinterpret_cast<Packet4ui>(p4i_ONE));
1296 return vec_sel(elsePacket, thenPacket, mask);
1297}
1298
1299#endif
1300
1301template <>
1302EIGEN_STRONG_INLINE Packet4f pldexp<Packet4f>(const Packet4f& a, const Packet4f& exponent) {
1303 return pldexp_generic(a, exponent);
1304}
1305
1306template <>
1307EIGEN_STRONG_INLINE Packet2d pldexp<Packet2d>(const Packet2d& a, const Packet2d& exponent) {
1308 // Clamp exponent to [-2099, 2099]
1309 const Packet2d max_exponent = pset1<Packet2d>(2099.0);
1310 const Packet2l e = pcast<Packet2d, Packet2l>(pmin(pmax(exponent, pnegate(max_exponent)), max_exponent));
1311
1312 // Split 2^e into four factors and multiply:
1313 const Packet2l bias = {1023, 1023};
1314 Packet2l b = plogical_shift_right<2>(e); // floor(e/4)
1315 Packet2d c = reinterpret_cast<Packet2d>(plogical_shift_left<52>(b + bias));
1316 Packet2d out = pmul(pmul(pmul(a, c), c), c); // a * 2^(3b)
1317 b = psub(psub(psub(e, b), b), b); // e - 3b
1318 c = reinterpret_cast<Packet2d>(plogical_shift_left<52>(b + bias)); // 2^(e - 3b)
1319 out = pmul(out, c); // a * 2^e
1320 return out;
1321}
1322
1323template <>
1324EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) {
1325 EIGEN_ZVECTOR_PREFETCH(addr);
1326}
1327template <>
1328EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) {
1329 return pload<Packet4f>(from);
1330}
1331template <>
1332EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) {
1333 pstore<float>(to, from);
1334}
1335template <>
1336EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a) {
1337 return padd<Packet4f>(pset1<Packet4f>(a), p4f_COUNTDOWN);
1338}
1339
1340#if !defined(vec_float) || !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ < 13)
1341#pragma GCC warning \
1342 "float->int and int->float conversion is simulated. compile for z15 for improved performance"
1343template <>
1344struct cast_impl<Packet4i, Packet4f> {
1345 EIGEN_DEVICE_FUNC static inline Packet4f run(const Packet4i& a) {
1346 return Packet4f{float(a[0]), float(a[1]), float(a[2]), float(a[3]) };
1347 }
1348};
1349
1350template <>
1351struct cast_impl<Packet4f, Packet4i> {
1352 EIGEN_DEVICE_FUNC static inline Packet4i run(const Packet4f& a) {
1353 return Packet4i{int(a[0]), int(a[1]), int(a[2]), int(a[3]) };
1354 }
1355};
1356
1357template <>
1358struct cast_impl<Packet2l, Packet2d> {
1359 EIGEN_DEVICE_FUNC static inline Packet2d run(const Packet2l& a) {
1360 return Packet2d{double(a[0]), double(a[1]) };
1361 }
1362};
1363
1364template <>
1365struct cast_impl<Packet2d, Packet2l> {
1366 EIGEN_DEVICE_FUNC static inline Packet2l run(const Packet2d& a) {
1367 return Packet2l{(long long)(a[0]), (long long)(a[1]) };
1368 }
1369};
1370#else
1371template <>
1372struct cast_impl<Packet4i, Packet4f> {
1373 EIGEN_DEVICE_FUNC static inline Packet4f run(const Packet4i& a) {
1374 return vec_float(a);
1375 }
1376};
1377
1378template <>
1379struct cast_impl<Packet4f, Packet4i> {
1380 EIGEN_DEVICE_FUNC static inline Packet4i run(const Packet4f& a) {
1381 return vec_signed(a);
1382 }
1383};
1384
1385template <>
1386struct cast_impl<Packet2l, Packet2d> {
1387 EIGEN_DEVICE_FUNC static inline Packet2d run(const Packet2l& a) {
1388 return vec_double(a);
1389 }
1390};
1391
1392template <>
1393struct cast_impl<Packet2d, Packet2l> {
1394 EIGEN_DEVICE_FUNC static inline Packet2l run(const Packet2d& a) {
1395 return vec_signed(a);
1396 }
1397};
1398#endif
1399
1400template <>
1401EIGEN_STRONG_INLINE Packet4f pset1frombits<Packet4f>(uint32_t from) {
1402 return pset1<Packet4f>(Eigen::numext::bit_cast<float>(from));
1403}
1404template <>
1405EIGEN_STRONG_INLINE Packet2d pset1frombits<Packet2d>(uint64_t from) {
1406 return pset1<Packet2d>(Eigen::numext::bit_cast<double>(from));
1407}
1408
1409} // end namespace internal
1410
1411} // end namespace Eigen
1412
1413#endif // EIGEN_PACKET_MATH_ZVECTOR_H
Eigen::Aligned16
@ Aligned16
Definition Constants.h:237
Eigen
Namespace containing all symbols from the Eigen library.
Definition B01_Experimental.dox:1
Eigen::Index
EIGEN_DEFAULT_DENSE_INDEX_TYPE Index
The Index type as used for the API.
Definition Meta.h:82