• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2008-2016 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_ALTIVEC_H
11 #define EIGEN_PACKET_MATH_ALTIVEC_H
12 
13 namespace Eigen {
14 
15 namespace internal {
16 
17 #ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
18 #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 4
19 #endif
20 
21 #ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
22 #define EIGEN_HAS_SINGLE_INSTRUCTION_MADD
23 #endif
24 
25 #ifndef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
26 #define EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
27 #endif
28 
29 // NOTE Altivec has 32 registers, but Eigen only accepts a value of 8 or 16
30 #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
31 #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS  32
32 #endif
33 
34 typedef __vector float          Packet4f;
35 typedef __vector int            Packet4i;
36 typedef __vector unsigned int   Packet4ui;
37 typedef __vector __bool int     Packet4bi;
38 typedef __vector short int      Packet8i;
39 typedef __vector unsigned char  Packet16uc;
40 
41 // We don't want to write the same code all the time, but we need to reuse the constants
42 // and it doesn't really work to declare them global, so we define macros instead
43 
44 #define _EIGEN_DECLARE_CONST_FAST_Packet4f(NAME,X) \
45   Packet4f p4f_##NAME = reinterpret_cast<Packet4f>(vec_splat_s32(X))
46 
47 #define _EIGEN_DECLARE_CONST_FAST_Packet4i(NAME,X) \
48   Packet4i p4i_##NAME = vec_splat_s32(X)
49 
50 #define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
51   Packet4f p4f_##NAME = pset1<Packet4f>(X)
52 
53 #define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
54   Packet4i p4i_##NAME = pset1<Packet4i>(X)
55 
56 #define _EIGEN_DECLARE_CONST_Packet2d(NAME,X) \
57   Packet2d p2d_##NAME = pset1<Packet2d>(X)
58 
59 #define _EIGEN_DECLARE_CONST_Packet2l(NAME,X) \
60   Packet2l p2l_##NAME = pset1<Packet2l>(X)
61 
62 #define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
63   const Packet4f p4f_##NAME = reinterpret_cast<Packet4f>(pset1<Packet4i>(X))
64 
65 #define DST_CHAN 1
66 #define DST_CTRL(size, count, stride) (((size) << 24) | ((count) << 16) | (stride))
67 
68 
69 // These constants are endian-agnostic
70 static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0); //{ 0.0, 0.0, 0.0, 0.0}
71 static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); //{ 0, 0, 0, 0,}
72 static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE,1); //{ 1, 1, 1, 1}
73 static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS16,-16); //{ -16, -16, -16, -16}
74 static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1); //{ -1, -1, -1, -1}
75 static Packet4f p4f_MZERO = (Packet4f) vec_sl((Packet4ui)p4i_MINUS1, (Packet4ui)p4i_MINUS1); //{ 0x80000000, 0x80000000, 0x80000000, 0x80000000}
76 #ifndef __VSX__
77 static Packet4f p4f_ONE = vec_ctf(p4i_ONE, 0); //{ 1.0, 1.0, 1.0, 1.0}
78 #endif
79 
80 static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 };
81 static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 };
82 
83 static Packet16uc p16uc_REVERSE32 = { 12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3 };
84 static Packet16uc p16uc_DUPLICATE32_HI = { 0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7 };
85 
86 // Mask alignment
87 #ifdef __PPC64__
88 #define _EIGEN_MASK_ALIGNMENT	0xfffffffffffffff0
89 #else
90 #define _EIGEN_MASK_ALIGNMENT	0xfffffff0
91 #endif
92 
93 #define _EIGEN_ALIGNED_PTR(x)	((std::ptrdiff_t)(x) & _EIGEN_MASK_ALIGNMENT)
94 
95 // Handle endianness properly while loading constants
96 // Define global static constants:
97 #ifdef _BIG_ENDIAN
98 static Packet16uc p16uc_FORWARD = vec_lvsl(0, (float*)0);
99 #ifdef __VSX__
100 static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
101 #endif
102 static Packet16uc p16uc_PSET32_WODD   = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 };
103 static Packet16uc p16uc_PSET32_WEVEN  = vec_sld(p16uc_DUPLICATE32_HI, (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 };
104 static Packet16uc p16uc_HALF64_0_16 = vec_sld((Packet16uc)p4i_ZERO, vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 3), 8);      //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16};
105 #else
106 static Packet16uc p16uc_FORWARD = p16uc_REVERSE32;
107 static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
108 static Packet16uc p16uc_PSET32_WODD = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 1), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 };
109 static Packet16uc p16uc_PSET32_WEVEN = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 };
110 static Packet16uc p16uc_HALF64_0_16 = vec_sld(vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 0), (Packet16uc)p4i_ZERO, 8);      //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16};
111 #endif // _BIG_ENDIAN
112 
113 static Packet16uc p16uc_PSET64_HI = (Packet16uc) vec_mergeh((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN);     //{ 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 };
114 static Packet16uc p16uc_PSET64_LO = (Packet16uc) vec_mergel((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN);     //{ 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 };
115 static Packet16uc p16uc_TRANSPOSE64_HI = p16uc_PSET64_HI + p16uc_HALF64_0_16;                                         //{ 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23};
116 static Packet16uc p16uc_TRANSPOSE64_LO = p16uc_PSET64_LO + p16uc_HALF64_0_16;                                         //{ 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31};
117 
118 static Packet16uc p16uc_COMPLEX32_REV = vec_sld(p16uc_REVERSE32, p16uc_REVERSE32, 8);                                         //{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 };
119 
120 #ifdef _BIG_ENDIAN
121 static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8);                                            //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
122 #else
123 static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_PSET64_HI, p16uc_PSET64_LO, 8);                                            //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
124 #endif // _BIG_ENDIAN
125 
126 #if EIGEN_HAS_BUILTIN(__builtin_prefetch) || EIGEN_COMP_GNUC
127   #define EIGEN_PPC_PREFETCH(ADDR) __builtin_prefetch(ADDR);
128 #else
129   #define EIGEN_PPC_PREFETCH(ADDR) asm( "   dcbt [%[addr]]\n" :: [addr] "r" (ADDR) : "cc" );
130 #endif
131 
132 template<> struct packet_traits<float>  : default_packet_traits
133 {
134   typedef Packet4f type;
135   typedef Packet4f half;
136   enum {
137     Vectorizable = 1,
138     AlignedOnScalar = 1,
139     size=4,
140     HasHalfPacket = 1,
141 
142     HasAdd  = 1,
143     HasSub  = 1,
144     HasMul  = 1,
145     HasDiv  = 1,
146     HasMin  = 1,
147     HasMax  = 1,
148     HasAbs  = 1,
149     HasSin  = 0,
150     HasCos  = 0,
151     HasLog  = 0,
152     HasExp  = 1,
153 #ifdef __VSX__
154     HasSqrt = 1,
155 #if !EIGEN_COMP_CLANG
156     HasRsqrt = 1,
157 #else
158     HasRsqrt = 0,
159 #endif
160 #else
161     HasSqrt = 0,
162     HasRsqrt = 0,
163 #endif
164     HasRound = 1,
165     HasFloor = 1,
166     HasCeil = 1,
167     HasNegate = 1,
168     HasBlend = 1
169   };
170 };
171 template<> struct packet_traits<int>    : default_packet_traits
172 {
173   typedef Packet4i type;
174   typedef Packet4i half;
175   enum {
176     Vectorizable = 1,
177     AlignedOnScalar = 1,
178     size = 4,
179     HasHalfPacket = 0,
180 
181     HasAdd  = 1,
182     HasSub  = 1,
183     HasMul  = 1,
184     HasDiv  = 0,
185     HasBlend = 1
186   };
187 };
188 
189 
190 template<> struct unpacket_traits<Packet4f> { typedef float  type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; };
191 template<> struct unpacket_traits<Packet4i> { typedef int    type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; };
192 
193 inline std::ostream & operator <<(std::ostream & s, const Packet16uc & v)
194 {
195   union {
196     Packet16uc   v;
197     unsigned char n[16];
198   } vt;
199   vt.v = v;
200   for (int i=0; i< 16; i++)
201     s << (int)vt.n[i] << ", ";
202   return s;
203 }
204 
205 inline std::ostream & operator <<(std::ostream & s, const Packet4f & v)
206 {
207   union {
208     Packet4f   v;
209     float n[4];
210   } vt;
211   vt.v = v;
212   s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
213   return s;
214 }
215 
216 inline std::ostream & operator <<(std::ostream & s, const Packet4i & v)
217 {
218   union {
219     Packet4i   v;
220     int n[4];
221   } vt;
222   vt.v = v;
223   s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
224   return s;
225 }
226 
227 inline std::ostream & operator <<(std::ostream & s, const Packet4ui & v)
228 {
229   union {
230     Packet4ui   v;
231     unsigned int n[4];
232   } vt;
233   vt.v = v;
234   s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
235   return s;
236 }
237 
238 // Need to define them first or we get specialization after instantiation errors
239 template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from)
240 {
241   EIGEN_DEBUG_ALIGNED_LOAD
242 #ifdef __VSX__
243   return vec_vsx_ld(0, from);
244 #else
245   return vec_ld(0, from);
246 #endif
247 }
248 
249 template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*     from)
250 {
251   EIGEN_DEBUG_ALIGNED_LOAD
252 #ifdef __VSX__
253   return vec_vsx_ld(0, from);
254 #else
255   return vec_ld(0, from);
256 #endif
257 }
258 
259 template<> EIGEN_STRONG_INLINE void pstore<float>(float*   to, const Packet4f& from)
260 {
261   EIGEN_DEBUG_ALIGNED_STORE
262 #ifdef __VSX__
263   vec_vsx_st(from, 0, to);
264 #else
265   vec_st(from, 0, to);
266 #endif
267 }
268 
269 template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet4i& from)
270 {
271   EIGEN_DEBUG_ALIGNED_STORE
272 #ifdef __VSX__
273   vec_vsx_st(from, 0, to);
274 #else
275   vec_st(from, 0, to);
276 #endif
277 }
278 
279 template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&  from) {
280   Packet4f v = {from, from, from, from};
281   return v;
282 }
283 
284 template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int&    from)   {
285   Packet4i v = {from, from, from, from};
286   return v;
287 }
288 template<> EIGEN_STRONG_INLINE void
289 pbroadcast4<Packet4f>(const float *a,
290                       Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3)
291 {
292   a3 = pload<Packet4f>(a);
293   a0 = vec_splat(a3, 0);
294   a1 = vec_splat(a3, 1);
295   a2 = vec_splat(a3, 2);
296   a3 = vec_splat(a3, 3);
297 }
298 template<> EIGEN_STRONG_INLINE void
299 pbroadcast4<Packet4i>(const int *a,
300                       Packet4i& a0, Packet4i& a1, Packet4i& a2, Packet4i& a3)
301 {
302   a3 = pload<Packet4i>(a);
303   a0 = vec_splat(a3, 0);
304   a1 = vec_splat(a3, 1);
305   a2 = vec_splat(a3, 2);
306   a3 = vec_splat(a3, 3);
307 }
308 
309 template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
310 {
311   float EIGEN_ALIGN16 af[4];
312   af[0] = from[0*stride];
313   af[1] = from[1*stride];
314   af[2] = from[2*stride];
315   af[3] = from[3*stride];
316  return pload<Packet4f>(af);
317 }
318 template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, Index stride)
319 {
320   int EIGEN_ALIGN16 ai[4];
321   ai[0] = from[0*stride];
322   ai[1] = from[1*stride];
323   ai[2] = from[2*stride];
324   ai[3] = from[3*stride];
325  return pload<Packet4i>(ai);
326 }
327 template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride)
328 {
329   float EIGEN_ALIGN16 af[4];
330   pstore<float>(af, from);
331   to[0*stride] = af[0];
332   to[1*stride] = af[1];
333   to[2*stride] = af[2];
334   to[3*stride] = af[3];
335 }
336 template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, Index stride)
337 {
338   int EIGEN_ALIGN16 ai[4];
339   pstore<int>((int *)ai, from);
340   to[0*stride] = ai[0];
341   to[1*stride] = ai[1];
342   to[2*stride] = ai[2];
343   to[3*stride] = ai[3];
344 }
345 
346 template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a) { return pset1<Packet4f>(a) + p4f_COUNTDOWN; }
347 template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a)   { return pset1<Packet4i>(a) + p4i_COUNTDOWN; }
348 
349 template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return a + b; }
350 template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return a + b; }
351 
352 template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return a - b; }
353 template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return a - b; }
354 
355 template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) { return p4f_ZERO - a; }
356 template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return p4i_ZERO - a; }
357 
358 template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
359 template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; }
360 
361 template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_madd(a,b, p4f_MZERO); }
362 template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) { return a * b; }
363 
364 template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b)
365 {
366 #ifndef __VSX__  // VSX actually provides a div instruction
367   Packet4f t, y_0, y_1;
368 
369   // Altivec does not offer a divide instruction, we have to do a reciprocal approximation
370   y_0 = vec_re(b);
371 
372   // Do one Newton-Raphson iteration to get the needed accuracy
373   t   = vec_nmsub(y_0, b, p4f_ONE);
374   y_1 = vec_madd(y_0, t, y_0);
375 
376   return vec_madd(a, y_1, p4f_MZERO);
377 #else
378   return vec_div(a, b);
379 #endif
380 }
381 
382 template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/)
383 { eigen_assert(false && "packet integer division are not supported by AltiVec");
384   return pset1<Packet4i>(0);
385 }
386 
387 // for some weird raisons, it has to be overloaded for packet of integers
388 template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vec_madd(a,b,c); }
389 template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return a*b + c; }
390 
391 template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_min(a, b); }
392 template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_min(a, b); }
393 
394 template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_max(a, b); }
395 template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); }
396 
397 template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, b); }
398 template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, b); }
399 
400 template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_or(a, b); }
401 template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_or(a, b); }
402 
403 template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_xor(a, b); }
404 template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_xor(a, b); }
405 
406 template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, vec_nor(b, b)); }
407 template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, vec_nor(b, b)); }
408 
409 template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) { return vec_round(a); }
410 template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const  Packet4f& a) { return vec_ceil(a); }
411 template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) { return vec_floor(a); }
412 
413 #ifdef _BIG_ENDIAN
414 template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
415 {
416   EIGEN_DEBUG_ALIGNED_LOAD
417   Packet16uc MSQ, LSQ;
418   Packet16uc mask;
419   MSQ = vec_ld(0, (unsigned char *)from);          // most significant quadword
420   LSQ = vec_ld(15, (unsigned char *)from);         // least significant quadword
421   mask = vec_lvsl(0, from);                        // create the permute mask
422   return (Packet4f) vec_perm(MSQ, LSQ, mask);           // align the data
423 
424 }
425 template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
426 {
427   EIGEN_DEBUG_ALIGNED_LOAD
428   // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
429   Packet16uc MSQ, LSQ;
430   Packet16uc mask;
431   MSQ = vec_ld(0, (unsigned char *)from);          // most significant quadword
432   LSQ = vec_ld(15, (unsigned char *)from);         // least significant quadword
433   mask = vec_lvsl(0, from);                        // create the permute mask
434   return (Packet4i) vec_perm(MSQ, LSQ, mask);    // align the data
435 }
436 #else
437 // We also need ot redefine little endian loading of Packet4i/Packet4f using VSX
438 template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
439 {
440   EIGEN_DEBUG_UNALIGNED_LOAD
441   return (Packet4i) vec_vsx_ld((long)from & 15, (const int*) _EIGEN_ALIGNED_PTR(from));
442 }
443 template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
444 {
445   EIGEN_DEBUG_UNALIGNED_LOAD
446   return (Packet4f) vec_vsx_ld((long)from & 15, (const float*) _EIGEN_ALIGNED_PTR(from));
447 }
448 #endif
449 
450 template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float*   from)
451 {
452   Packet4f p;
453   if((std::ptrdiff_t(from) % 16) == 0)  p = pload<Packet4f>(from);
454   else                                  p = ploadu<Packet4f>(from);
455   return vec_perm(p, p, p16uc_DUPLICATE32_HI);
456 }
457 template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int*     from)
458 {
459   Packet4i p;
460   if((std::ptrdiff_t(from) % 16) == 0)  p = pload<Packet4i>(from);
461   else                                  p = ploadu<Packet4i>(from);
462   return vec_perm(p, p, p16uc_DUPLICATE32_HI);
463 }
464 
465 #ifdef _BIG_ENDIAN
466 template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*  to, const Packet4f& from)
467 {
468   EIGEN_DEBUG_UNALIGNED_STORE
469   // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
470   // Warning: not thread safe!
471   Packet16uc MSQ, LSQ, edges;
472   Packet16uc edgeAlign, align;
473 
474   MSQ = vec_ld(0, (unsigned char *)to);                     // most significant quadword
475   LSQ = vec_ld(15, (unsigned char *)to);                    // least significant quadword
476   edgeAlign = vec_lvsl(0, to);                              // permute map to extract edges
477   edges=vec_perm(LSQ,MSQ,edgeAlign);                        // extract the edges
478   align = vec_lvsr( 0, to );                                // permute map to misalign data
479   MSQ = vec_perm(edges,(Packet16uc)from,align);             // misalign the data (MSQ)
480   LSQ = vec_perm((Packet16uc)from,edges,align);             // misalign the data (LSQ)
481   vec_st( LSQ, 15, (unsigned char *)to );                   // Store the LSQ part first
482   vec_st( MSQ, 0, (unsigned char *)to );                    // Store the MSQ part
483 }
484 template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*      to, const Packet4i& from)
485 {
486   EIGEN_DEBUG_UNALIGNED_STORE
487   // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
488   // Warning: not thread safe!
489   Packet16uc MSQ, LSQ, edges;
490   Packet16uc edgeAlign, align;
491 
492   MSQ = vec_ld(0, (unsigned char *)to);                     // most significant quadword
493   LSQ = vec_ld(15, (unsigned char *)to);                    // least significant quadword
494   edgeAlign = vec_lvsl(0, to);                              // permute map to extract edges
495   edges=vec_perm(LSQ, MSQ, edgeAlign);                      // extract the edges
496   align = vec_lvsr( 0, to );                                // permute map to misalign data
497   MSQ = vec_perm(edges, (Packet16uc) from, align);          // misalign the data (MSQ)
498   LSQ = vec_perm((Packet16uc) from, edges, align);          // misalign the data (LSQ)
499   vec_st( LSQ, 15, (unsigned char *)to );                   // Store the LSQ part first
500   vec_st( MSQ, 0, (unsigned char *)to );                    // Store the MSQ part
501 }
502 #else
503 // We also need ot redefine little endian loading of Packet4i/Packet4f using VSX
504 template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*       to, const Packet4i& from)
505 {
506   EIGEN_DEBUG_ALIGNED_STORE
507   vec_vsx_st(from, (long)to & 15, (int*) _EIGEN_ALIGNED_PTR(to));
508 }
509 template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*   to, const Packet4f& from)
510 {
511   EIGEN_DEBUG_ALIGNED_STORE
512   vec_vsx_st(from, (long)to & 15, (float*) _EIGEN_ALIGNED_PTR(to));
513 }
514 #endif
515 
516 template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr)    { EIGEN_PPC_PREFETCH(addr); }
517 template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*     addr)    { EIGEN_PPC_PREFETCH(addr); }
518 
519 template<> EIGEN_STRONG_INLINE float  pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x; vec_ste(a, 0, &x); return x; }
520 template<> EIGEN_STRONG_INLINE int    pfirst<Packet4i>(const Packet4i& a) { int   EIGEN_ALIGN16 x; vec_ste(a, 0, &x); return x; }
521 
522 template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a)
523 {
524   return reinterpret_cast<Packet4f>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE32));
525 }
526 template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a)
527 {
528   return reinterpret_cast<Packet4i>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE32)); }
529 
530 template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) { return vec_abs(a); }
531 template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vec_abs(a); }
532 
533 template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
534 {
535   Packet4f b, sum;
536   b   = vec_sld(a, a, 8);
537   sum = a + b;
538   b   = vec_sld(sum, sum, 4);
539   sum += b;
540   return pfirst(sum);
541 }
542 
543 template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
544 {
545   Packet4f v[4], sum[4];
546 
547   // It's easier and faster to transpose then add as columns
548   // Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation
549   // Do the transpose, first set of moves
550   v[0] = vec_mergeh(vecs[0], vecs[2]);
551   v[1] = vec_mergel(vecs[0], vecs[2]);
552   v[2] = vec_mergeh(vecs[1], vecs[3]);
553   v[3] = vec_mergel(vecs[1], vecs[3]);
554   // Get the resulting vectors
555   sum[0] = vec_mergeh(v[0], v[2]);
556   sum[1] = vec_mergel(v[0], v[2]);
557   sum[2] = vec_mergeh(v[1], v[3]);
558   sum[3] = vec_mergel(v[1], v[3]);
559 
560   // Now do the summation:
561   // Lines 0+1
562   sum[0] = sum[0] + sum[1];
563   // Lines 2+3
564   sum[1] = sum[2] + sum[3];
565   // Add the results
566   sum[0] = sum[0] + sum[1];
567 
568   return sum[0];
569 }
570 
571 template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
572 {
573   Packet4i sum;
574   sum = vec_sums(a, p4i_ZERO);
575 #ifdef _BIG_ENDIAN
576   sum = vec_sld(sum, p4i_ZERO, 12);
577 #else
578   sum = vec_sld(p4i_ZERO, sum, 4);
579 #endif
580   return pfirst(sum);
581 }
582 
583 template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
584 {
585   Packet4i v[4], sum[4];
586 
587   // It's easier and faster to transpose then add as columns
588   // Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation
589   // Do the transpose, first set of moves
590   v[0] = vec_mergeh(vecs[0], vecs[2]);
591   v[1] = vec_mergel(vecs[0], vecs[2]);
592   v[2] = vec_mergeh(vecs[1], vecs[3]);
593   v[3] = vec_mergel(vecs[1], vecs[3]);
594   // Get the resulting vectors
595   sum[0] = vec_mergeh(v[0], v[2]);
596   sum[1] = vec_mergel(v[0], v[2]);
597   sum[2] = vec_mergeh(v[1], v[3]);
598   sum[3] = vec_mergel(v[1], v[3]);
599 
600   // Now do the summation:
601   // Lines 0+1
602   sum[0] = sum[0] + sum[1];
603   // Lines 2+3
604   sum[1] = sum[2] + sum[3];
605   // Add the results
606   sum[0] = sum[0] + sum[1];
607 
608   return sum[0];
609 }
610 
611 // Other reduction functions:
612 // mul
613 template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
614 {
615   Packet4f prod;
616   prod = pmul(a, vec_sld(a, a, 8));
617   return pfirst(pmul(prod, vec_sld(prod, prod, 4)));
618 }
619 
620 template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
621 {
622   EIGEN_ALIGN16 int aux[4];
623   pstore(aux, a);
624   return aux[0] * aux[1] * aux[2] * aux[3];
625 }
626 
627 // min
628 template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
629 {
630   Packet4f b, res;
631   b = vec_min(a, vec_sld(a, a, 8));
632   res = vec_min(b, vec_sld(b, b, 4));
633   return pfirst(res);
634 }
635 
636 template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a)
637 {
638   Packet4i b, res;
639   b = vec_min(a, vec_sld(a, a, 8));
640   res = vec_min(b, vec_sld(b, b, 4));
641   return pfirst(res);
642 }
643 
644 // max
645 template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
646 {
647   Packet4f b, res;
648   b = vec_max(a, vec_sld(a, a, 8));
649   res = vec_max(b, vec_sld(b, b, 4));
650   return pfirst(res);
651 }
652 
653 template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
654 {
655   Packet4i b, res;
656   b = vec_max(a, vec_sld(a, a, 8));
657   res = vec_max(b, vec_sld(b, b, 4));
658   return pfirst(res);
659 }
660 
661 template<int Offset>
662 struct palign_impl<Offset,Packet4f>
663 {
664   static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
665   {
666 #ifdef _BIG_ENDIAN
667     switch (Offset % 4) {
668     case 1:
669       first = vec_sld(first, second, 4); break;
670     case 2:
671       first = vec_sld(first, second, 8); break;
672     case 3:
673       first = vec_sld(first, second, 12); break;
674     }
675 #else
676     switch (Offset % 4) {
677     case 1:
678       first = vec_sld(second, first, 12); break;
679     case 2:
680       first = vec_sld(second, first, 8); break;
681     case 3:
682       first = vec_sld(second, first, 4); break;
683     }
684 #endif
685   }
686 };
687 
688 template<int Offset>
689 struct palign_impl<Offset,Packet4i>
690 {
691   static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second)
692   {
693 #ifdef _BIG_ENDIAN
694     switch (Offset % 4) {
695     case 1:
696       first = vec_sld(first, second, 4); break;
697     case 2:
698       first = vec_sld(first, second, 8); break;
699     case 3:
700       first = vec_sld(first, second, 12); break;
701     }
702 #else
703     switch (Offset % 4) {
704     case 1:
705       first = vec_sld(second, first, 12); break;
706     case 2:
707       first = vec_sld(second, first, 8); break;
708     case 3:
709       first = vec_sld(second, first, 4); break;
710     }
711 #endif
712   }
713 };
714 
715 EIGEN_DEVICE_FUNC inline void
716 ptranspose(PacketBlock<Packet4f,4>& kernel) {
717   Packet4f t0, t1, t2, t3;
718   t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]);
719   t1 = vec_mergel(kernel.packet[0], kernel.packet[2]);
720   t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]);
721   t3 = vec_mergel(kernel.packet[1], kernel.packet[3]);
722   kernel.packet[0] = vec_mergeh(t0, t2);
723   kernel.packet[1] = vec_mergel(t0, t2);
724   kernel.packet[2] = vec_mergeh(t1, t3);
725   kernel.packet[3] = vec_mergel(t1, t3);
726 }
727 
728 EIGEN_DEVICE_FUNC inline void
729 ptranspose(PacketBlock<Packet4i,4>& kernel) {
730   Packet4i t0, t1, t2, t3;
731   t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]);
732   t1 = vec_mergel(kernel.packet[0], kernel.packet[2]);
733   t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]);
734   t3 = vec_mergel(kernel.packet[1], kernel.packet[3]);
735   kernel.packet[0] = vec_mergeh(t0, t2);
736   kernel.packet[1] = vec_mergel(t0, t2);
737   kernel.packet[2] = vec_mergeh(t1, t3);
738   kernel.packet[3] = vec_mergel(t1, t3);
739 }
740 
741 template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) {
742   Packet4ui select = { ifPacket.select[0], ifPacket.select[1], ifPacket.select[2], ifPacket.select[3] };
743   Packet4ui mask = reinterpret_cast<Packet4ui>(vec_cmpeq(reinterpret_cast<Packet4ui>(select), reinterpret_cast<Packet4ui>(p4i_ONE)));
744   return vec_sel(elsePacket, thenPacket, mask);
745 }
746 
747 template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) {
748   Packet4ui select = { ifPacket.select[0], ifPacket.select[1], ifPacket.select[2], ifPacket.select[3] };
749   Packet4ui mask = reinterpret_cast<Packet4ui>(vec_cmpeq(reinterpret_cast<Packet4ui>(select), reinterpret_cast<Packet4ui>(p4i_ONE)));
750   return vec_sel(elsePacket, thenPacket, mask);
751 }
752 
753 
754 //---------- double ----------
755 #ifdef __VSX__
756 typedef __vector double              Packet2d;
757 typedef __vector unsigned long long  Packet2ul;
758 typedef __vector long long           Packet2l;
759 #if EIGEN_COMP_CLANG
760 typedef Packet2ul                    Packet2bl;
761 #else
762 typedef __vector __bool long         Packet2bl;
763 #endif
764 
765 static Packet2l  p2l_ONE  = { 1, 1 };
766 static Packet2l  p2l_ZERO = reinterpret_cast<Packet2l>(p4i_ZERO);
767 static Packet2d  p2d_ONE  = { 1.0, 1.0 };
768 static Packet2d  p2d_ZERO = reinterpret_cast<Packet2d>(p4f_ZERO);
769 static Packet2d  p2d_MZERO = { -0.0, -0.0 };
770 
771 #ifdef _BIG_ENDIAN
772 static Packet2d p2d_COUNTDOWN = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(p2d_ZERO), reinterpret_cast<Packet4f>(p2d_ONE), 8));
773 #else
774 static Packet2d p2d_COUNTDOWN = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(p2d_ONE), reinterpret_cast<Packet4f>(p2d_ZERO), 8));
775 #endif
776 
777 template<int index> Packet2d vec_splat_dbl(Packet2d& a);
778 
779 template<> EIGEN_STRONG_INLINE Packet2d vec_splat_dbl<0>(Packet2d& a)
780 {
781   return reinterpret_cast<Packet2d>(vec_perm(a, a, p16uc_PSET64_HI));
782 }
783 
784 template<> EIGEN_STRONG_INLINE Packet2d vec_splat_dbl<1>(Packet2d& a)
785 {
786   return reinterpret_cast<Packet2d>(vec_perm(a, a, p16uc_PSET64_LO));
787 }
788 
789 template<> struct packet_traits<double> : default_packet_traits
790 {
791   typedef Packet2d type;
792   typedef Packet2d half;
793   enum {
794     Vectorizable = 1,
795     AlignedOnScalar = 1,
796     size=2,
797     HasHalfPacket = 1,
798 
799     HasAdd  = 1,
800     HasSub  = 1,
801     HasMul  = 1,
802     HasDiv  = 1,
803     HasMin  = 1,
804     HasMax  = 1,
805     HasAbs  = 1,
806     HasSin  = 0,
807     HasCos  = 0,
808     HasLog  = 0,
809     HasExp  = 1,
810     HasSqrt = 1,
811     HasRsqrt = 1,
812     HasRound = 1,
813     HasFloor = 1,
814     HasCeil = 1,
815     HasNegate = 1,
816     HasBlend = 1
817   };
818 };
819 
820 template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; };
821 
822 inline std::ostream & operator <<(std::ostream & s, const Packet2l & v)
823 {
824   union {
825     Packet2l   v;
826     int64_t n[2];
827   } vt;
828   vt.v = v;
829   s << vt.n[0] << ", " << vt.n[1];
830   return s;
831 }
832 
833 inline std::ostream & operator <<(std::ostream & s, const Packet2d & v)
834 {
835   union {
836     Packet2d   v;
837     double n[2];
838   } vt;
839   vt.v = v;
840   s << vt.n[0] << ", " << vt.n[1];
841   return s;
842 }
843 
844 // Need to define them first or we get specialization after instantiation errors
845 template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from)
846 {
847   EIGEN_DEBUG_ALIGNED_LOAD
848 #ifdef __VSX__
849   return vec_vsx_ld(0, from);
850 #else
851   return vec_ld(0, from);
852 #endif
853 }
854 
855 template<> EIGEN_STRONG_INLINE void pstore<double>(double*   to, const Packet2d& from)
856 {
857   EIGEN_DEBUG_ALIGNED_STORE
858 #ifdef __VSX__
859   vec_vsx_st(from, 0, to);
860 #else
861   vec_st(from, 0, to);
862 #endif
863 }
864 
865 template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double&  from) {
866   Packet2d v = {from, from};
867   return v;
868 }
869 
870 template<> EIGEN_STRONG_INLINE void
871 pbroadcast4<Packet2d>(const double *a,
872                       Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3)
873 {
874   a1 = pload<Packet2d>(a);
875   a0 = vec_splat_dbl<0>(a1);
876   a1 = vec_splat_dbl<1>(a1);
877   a3 = pload<Packet2d>(a+2);
878   a2 = vec_splat_dbl<0>(a3);
879   a3 = vec_splat_dbl<1>(a3);
880 }
881 
882 template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride)
883 {
884   double EIGEN_ALIGN16 af[2];
885   af[0] = from[0*stride];
886   af[1] = from[1*stride];
887  return pload<Packet2d>(af);
888 }
889 template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride)
890 {
891   double EIGEN_ALIGN16 af[2];
892   pstore<double>(af, from);
893   to[0*stride] = af[0];
894   to[1*stride] = af[1];
895 }
896 
897 template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) { return pset1<Packet2d>(a) + p2d_COUNTDOWN; }
898 
899 template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return a + b; }
900 
901 template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return a - b; }
902 
903 template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return p2d_ZERO - a; }
904 
905 template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; }
906 
907 template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_madd(a,b,p2d_MZERO); }
908 template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_div(a,b); }
909 
910 // for some weird raisons, it has to be overloaded for packet of integers
911 template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vec_madd(a, b, c); }
912 
913 template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_min(a, b); }
914 
915 template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_max(a, b); }
916 
917 template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, b); }
918 
919 template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_or(a, b); }
920 
921 template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_xor(a, b); }
922 
923 template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, vec_nor(b, b)); }
924 
925 template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return vec_round(a); }
926 template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const  Packet2d& a) { return vec_ceil(a); }
927 template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return vec_floor(a); }
928 
929 template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from)
930 {
931   EIGEN_DEBUG_ALIGNED_LOAD
932   return (Packet2d) vec_vsx_ld((long)from & 15, (const double*) _EIGEN_ALIGNED_PTR(from));
933 }
934 
935 template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double*   from)
936 {
937   Packet2d p;
938   if((std::ptrdiff_t(from) % 16) == 0)  p = pload<Packet2d>(from);
939   else                                  p = ploadu<Packet2d>(from);
940   return vec_splat_dbl<0>(p);
941 }
942 
943 template<> EIGEN_STRONG_INLINE void pstoreu<double>(double*  to, const Packet2d& from)
944 {
945   EIGEN_DEBUG_ALIGNED_STORE
946   vec_vsx_st((Packet4f)from, (long)to & 15, (float*) _EIGEN_ALIGNED_PTR(to));
947 }
948 
949 template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { EIGEN_PPC_PREFETCH(addr); }
950 
951 template<> EIGEN_STRONG_INLINE double  pfirst<Packet2d>(const Packet2d& a) { double EIGEN_ALIGN16 x[2]; pstore<double>(x, a); return x[0]; }
952 
953 template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a)
954 {
955   return reinterpret_cast<Packet2d>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE64));
956 }
957 template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) { return vec_abs(a); }
958 
959 template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
960 {
961   Packet2d b, sum;
962   b   = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(a), reinterpret_cast<Packet4f>(a), 8));
963   sum = a + b;
964   return pfirst<Packet2d>(sum);
965 }
966 
967 template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
968 {
969   Packet2d v[2], sum;
970   v[0] = vecs[0] + reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(vecs[0]), reinterpret_cast<Packet4f>(vecs[0]), 8));
971   v[1] = vecs[1] + reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(vecs[1]), reinterpret_cast<Packet4f>(vecs[1]), 8));
972 
973 #ifdef _BIG_ENDIAN
974   sum = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(v[0]), reinterpret_cast<Packet4f>(v[1]), 8));
975 #else
976   sum = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(v[1]), reinterpret_cast<Packet4f>(v[0]), 8));
977 #endif
978 
979   return sum;
980 }
981 // Other reduction functions:
982 // mul
983 template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a)
984 {
985   return pfirst(pmul(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(a), reinterpret_cast<Packet4ui>(a), 8))));
986 }
987 
988 // min
989 template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a)
990 {
991   return pfirst(pmin(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(a), reinterpret_cast<Packet4ui>(a), 8))));
992 }
993 
994 // max
995 template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a)
996 {
997   return pfirst(pmax(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(a), reinterpret_cast<Packet4ui>(a), 8))));
998 }
999 
1000 template<int Offset>
1001 struct palign_impl<Offset,Packet2d>
1002 {
1003   static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second)
1004   {
1005     if (Offset == 1)
1006 #ifdef _BIG_ENDIAN
1007       first = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(first), reinterpret_cast<Packet4ui>(second), 8));
1008 #else
1009       first = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(second), reinterpret_cast<Packet4ui>(first), 8));
1010 #endif
1011   }
1012 };
1013 
1014 EIGEN_DEVICE_FUNC inline void
1015 ptranspose(PacketBlock<Packet2d,2>& kernel) {
1016   Packet2d t0, t1;
1017   t0 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_HI);
1018   t1 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_LO);
1019   kernel.packet[0] = t0;
1020   kernel.packet[1] = t1;
1021 }
1022 
1023 template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) {
1024   Packet2l select = { ifPacket.select[0], ifPacket.select[1] };
1025   Packet2bl mask = vec_cmpeq(reinterpret_cast<Packet2d>(select), reinterpret_cast<Packet2d>(p2l_ONE));
1026   return vec_sel(elsePacket, thenPacket, mask);
1027 }
1028 #endif // __VSX__
1029 } // end namespace internal
1030 
1031 } // end namespace Eigen
1032 
1033 #endif // EIGEN_PACKET_MATH_ALTIVEC_H
1034