Eigen  3.2.92
AltiVec/PacketMath.h
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2008-2014 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 = (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 #ifndef __VSX__
73 static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE,1); //{ 1, 1, 1, 1}
74 static Packet4f p4f_ONE = vec_ctf(p4i_ONE, 0); //{ 1.0, 1.0, 1.0, 1.0}
75 #endif
76 static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS16,-16); //{ -16, -16, -16, -16}
77 static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1); //{ -1, -1, -1, -1}
78 static Packet4f p4f_ZERO_ = (Packet4f) vec_sl((Packet4ui)p4i_MINUS1, (Packet4ui)p4i_MINUS1); //{ 0x80000000, 0x80000000, 0x80000000, 0x80000000}
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) ((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 static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
100 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 };
101 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 };
102 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};
103 #else
104 static Packet16uc p16uc_FORWARD = p16uc_REVERSE32;
105 static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
106 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 };
107 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 };
108 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};
109 #endif // _BIG_ENDIAN
110 
111 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 };
112 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 };
113 static Packet16uc p16uc_TRANSPOSE64_HI = vec_add(p16uc_PSET64_HI, p16uc_HALF64_0_16); //{ 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23};
114 static Packet16uc p16uc_TRANSPOSE64_LO = vec_add(p16uc_PSET64_LO, p16uc_HALF64_0_16); //{ 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31};
115 
116 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 };
117 
118 #ifdef _BIG_ENDIAN
119 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 };
120 #else
121 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 };
122 #endif // _BIG_ENDIAN
123 
124 template<> struct packet_traits<float> : default_packet_traits
125 {
126  typedef Packet4f type;
127  typedef Packet4f half;
128  enum {
129  Vectorizable = 1,
130  AlignedOnScalar = 1,
131  size=4,
132  HasHalfPacket=0,
133 
134  // FIXME check the Has*
135  HasDiv = 1,
136  HasSin = 0,
137  HasCos = 0,
138  HasLog = 1,
139  HasExp = 1,
140  HasSqrt = 0
141  };
142 };
143 template<> struct packet_traits<int> : default_packet_traits
144 {
145  typedef Packet4i type;
146  typedef Packet4i half;
147  enum {
148  // FIXME check the Has*
149  Vectorizable = 1,
150  AlignedOnScalar = 1,
151  size=4
152  };
153 };
154 
155 
156 template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; };
157 template<> struct unpacket_traits<Packet4i> { typedef int type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; };
158 
159 inline std::ostream & operator <<(std::ostream & s, const Packet16uc & v)
160 {
161  union {
162  Packet16uc v;
163  unsigned char n[16];
164  } vt;
165  vt.v = v;
166  for (int i=0; i< 16; i++)
167  s << (int)vt.n[i] << ", ";
168  return s;
169 }
170 
171 inline std::ostream & operator <<(std::ostream & s, const Packet4f & v)
172 {
173  union {
174  Packet4f v;
175  float n[4];
176  } vt;
177  vt.v = v;
178  s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
179  return s;
180 }
181 
182 inline std::ostream & operator <<(std::ostream & s, const Packet4i & v)
183 {
184  union {
185  Packet4i v;
186  int n[4];
187  } vt;
188  vt.v = v;
189  s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
190  return s;
191 }
192 
193 inline std::ostream & operator <<(std::ostream & s, const Packet4ui & v)
194 {
195  union {
196  Packet4ui v;
197  unsigned int n[4];
198  } vt;
199  vt.v = v;
200  s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
201  return s;
202 }
203 /*
204 inline std::ostream & operator <<(std::ostream & s, const Packetbi & v)
205 {
206  union {
207  Packet4bi v;
208  unsigned int n[4];
209  } vt;
210  vt.v = v;
211  s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
212  return s;
213 }*/
214 
215 
216 // Need to define them first or we get specialization after instantiation errors
217 template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); }
218 template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); }
219 
220 template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st(from, 0, to); }
221 template<> EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st(from, 0, to); }
222 
223 template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) {
224  // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
225  float EIGEN_ALIGN16 af[4];
226  af[0] = from;
227  Packet4f vc = pload<Packet4f>(af);
228  vc = vec_splat(vc, 0);
229  return vc;
230 }
231 
232 template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) {
233  int EIGEN_ALIGN16 ai[4];
234  ai[0] = from;
235  Packet4i vc = pload<Packet4i>(ai);
236  vc = vec_splat(vc, 0);
237  return vc;
238 }
239 template<> EIGEN_STRONG_INLINE void
240 pbroadcast4<Packet4f>(const float *a,
241  Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3)
242 {
243  a3 = pload<Packet4f>(a);
244  a0 = vec_splat(a3, 0);
245  a1 = vec_splat(a3, 1);
246  a2 = vec_splat(a3, 2);
247  a3 = vec_splat(a3, 3);
248 }
249 template<> EIGEN_STRONG_INLINE void
250 pbroadcast4<Packet4i>(const int *a,
251  Packet4i& a0, Packet4i& a1, Packet4i& a2, Packet4i& a3)
252 {
253  a3 = pload<Packet4i>(a);
254  a0 = vec_splat(a3, 0);
255  a1 = vec_splat(a3, 1);
256  a2 = vec_splat(a3, 2);
257  a3 = vec_splat(a3, 3);
258 }
259 
260 template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
261 {
262  float EIGEN_ALIGN16 af[4];
263  af[0] = from[0*stride];
264  af[1] = from[1*stride];
265  af[2] = from[2*stride];
266  af[3] = from[3*stride];
267  return pload<Packet4f>(af);
268 }
269 template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, Index stride)
270 {
271  int EIGEN_ALIGN16 ai[4];
272  ai[0] = from[0*stride];
273  ai[1] = from[1*stride];
274  ai[2] = from[2*stride];
275  ai[3] = from[3*stride];
276  return pload<Packet4i>(ai);
277 }
278 template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride)
279 {
280  float EIGEN_ALIGN16 af[4];
281  pstore<float>(af, from);
282  to[0*stride] = af[0];
283  to[1*stride] = af[1];
284  to[2*stride] = af[2];
285  to[3*stride] = af[3];
286 }
287 template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, Index stride)
288 {
289  int EIGEN_ALIGN16 ai[4];
290  pstore<int>((int *)ai, from);
291  to[0*stride] = ai[0];
292  to[1*stride] = ai[1];
293  to[2*stride] = ai[2];
294  to[3*stride] = ai[3];
295 }
296 
297 template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a) { return vec_add(pset1<Packet4f>(a), p4f_COUNTDOWN); }
298 template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a) { return vec_add(pset1<Packet4i>(a), p4i_COUNTDOWN); }
299 
300 template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_add(a,b); }
301 template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_add(a,b); }
302 
303 template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_sub(a,b); }
304 template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_sub(a,b); }
305 
306 template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) { return psub<Packet4f>(p4f_ZERO, a); }
307 template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return psub<Packet4i>(p4i_ZERO, a); }
308 
309 template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
310 template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; }
311 
312 template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_madd(a,b,p4f_ZERO); }
313 /* Commented out: it's actually slower than processing it scalar
314  *
315 template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b)
316 {
317  // Detailed in: http://freevec.org/content/32bit_signed_integer_multiplication_altivec
318  //Set up constants, variables
319  Packet4i a1, b1, bswap, low_prod, high_prod, prod, prod_, v1sel;
320 
321  // Get the absolute values
322  a1 = vec_abs(a);
323  b1 = vec_abs(b);
324 
325  // Get the signs using xor
326  Packet4bi sgn = (Packet4bi) vec_cmplt(vec_xor(a, b), p4i_ZERO);
327 
328  // Do the multiplication for the asbolute values.
329  bswap = (Packet4i) vec_rl((Packet4ui) b1, (Packet4ui) p4i_MINUS16 );
330  low_prod = vec_mulo((Packet8i) a1, (Packet8i)b1);
331  high_prod = vec_msum((Packet8i) a1, (Packet8i) bswap, p4i_ZERO);
332  high_prod = (Packet4i) vec_sl((Packet4ui) high_prod, (Packet4ui) p4i_MINUS16);
333  prod = vec_add( low_prod, high_prod );
334 
335  // NOR the product and select only the negative elements according to the sign mask
336  prod_ = vec_nor(prod, prod);
337  prod_ = vec_sel(p4i_ZERO, prod_, sgn);
338 
339  // Add 1 to the result to get the negative numbers
340  v1sel = vec_sel(p4i_ZERO, p4i_ONE, sgn);
341  prod_ = vec_add(prod_, v1sel);
342 
343  // Merge the results back to the final vector.
344  prod = vec_sel(prod, prod_, sgn);
345 
346  return prod;
347 }
348 */
349 template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b)
350 {
351 #ifndef __VSX__ // VSX actually provides a div instruction
352  Packet4f t, y_0, y_1;
353 
354  // Altivec does not offer a divide instruction, we have to do a reciprocal approximation
355  y_0 = vec_re(b);
356 
357  // Do one Newton-Raphson iteration to get the needed accuracy
358  t = vec_nmsub(y_0, b, p4f_ONE);
359  y_1 = vec_madd(y_0, t, y_0);
360 
361  return vec_madd(a, y_1, p4f_ZERO);
362 #else
363  return vec_div(a, b);
364 #endif
365 }
366 
367 template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/)
368 { eigen_assert(false && "packet integer division are not supported by AltiVec");
369  return pset1<Packet4i>(0);
370 }
371 
372 // for some weird raisons, it has to be overloaded for packet of integers
373 template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vec_madd(a, b, c); }
374 template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd(pmul(a,b), c); }
375 
376 template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_min(a, b); }
377 template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_min(a, b); }
378 
379 template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_max(a, b); }
380 template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); }
381 
382 template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, b); }
383 template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, b); }
384 
385 template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_or(a, b); }
386 template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_or(a, b); }
387 
388 template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_xor(a, b); }
389 template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_xor(a, b); }
390 
391 template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, vec_nor(b, b)); }
392 template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, vec_nor(b, b)); }
393 
394 #ifdef _BIG_ENDIAN
395 template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
396 {
397  EIGEN_DEBUG_ALIGNED_LOAD
398  Packet16uc MSQ, LSQ;
399  Packet16uc mask;
400  MSQ = vec_ld(0, (unsigned char *)from); // most significant quadword
401  LSQ = vec_ld(15, (unsigned char *)from); // least significant quadword
402  mask = vec_lvsl(0, from); // create the permute mask
403  return (Packet4f) vec_perm(MSQ, LSQ, mask); // align the data
404 
405 }
406 template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
407 {
408  EIGEN_DEBUG_ALIGNED_LOAD
409  // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
410  Packet16uc MSQ, LSQ;
411  Packet16uc mask;
412  MSQ = vec_ld(0, (unsigned char *)from); // most significant quadword
413  LSQ = vec_ld(15, (unsigned char *)from); // least significant quadword
414  mask = vec_lvsl(0, from); // create the permute mask
415  return (Packet4i) vec_perm(MSQ, LSQ, mask); // align the data
416 }
417 #else
418 // We also need ot redefine little endian loading of Packet4i/Packet4f using VSX
419 template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
420 {
421  EIGEN_DEBUG_ALIGNED_LOAD
422  return (Packet4i) vec_vsx_ld((long)from & 15, (const int*) _EIGEN_ALIGNED_PTR(from));
423 }
424 template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
425 {
426  EIGEN_DEBUG_ALIGNED_LOAD
427  return (Packet4f) vec_vsx_ld((long)from & 15, (const float*) _EIGEN_ALIGNED_PTR(from));
428 }
429 #endif
430 
431 template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from)
432 {
433  Packet4f p;
434  if((ptrdiff_t(from) % 16) == 0) p = pload<Packet4f>(from);
435  else p = ploadu<Packet4f>(from);
436  return vec_perm(p, p, p16uc_DUPLICATE32_HI);
437 }
438 template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int* from)
439 {
440  Packet4i p;
441  if((ptrdiff_t(from) % 16) == 0) p = pload<Packet4i>(from);
442  else p = ploadu<Packet4i>(from);
443  return vec_perm(p, p, p16uc_DUPLICATE32_HI);
444 }
445 
446 #ifdef _BIG_ENDIAN
447 template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from)
448 {
449  EIGEN_DEBUG_UNALIGNED_STORE
450  // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
451  // Warning: not thread safe!
452  Packet16uc MSQ, LSQ, edges;
453  Packet16uc edgeAlign, align;
454 
455  MSQ = vec_ld(0, (unsigned char *)to); // most significant quadword
456  LSQ = vec_ld(15, (unsigned char *)to); // least significant quadword
457  edgeAlign = vec_lvsl(0, to); // permute map to extract edges
458  edges=vec_perm(LSQ,MSQ,edgeAlign); // extract the edges
459  align = vec_lvsr( 0, to ); // permute map to misalign data
460  MSQ = vec_perm(edges,(Packet16uc)from,align); // misalign the data (MSQ)
461  LSQ = vec_perm((Packet16uc)from,edges,align); // misalign the data (LSQ)
462  vec_st( LSQ, 15, (unsigned char *)to ); // Store the LSQ part first
463  vec_st( MSQ, 0, (unsigned char *)to ); // Store the MSQ part
464 }
465 template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from)
466 {
467  EIGEN_DEBUG_UNALIGNED_STORE
468  // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
469  // Warning: not thread safe!
470  Packet16uc MSQ, LSQ, edges;
471  Packet16uc edgeAlign, align;
472 
473  MSQ = vec_ld(0, (unsigned char *)to); // most significant quadword
474  LSQ = vec_ld(15, (unsigned char *)to); // least significant quadword
475  edgeAlign = vec_lvsl(0, to); // permute map to extract edges
476  edges=vec_perm(LSQ, MSQ, edgeAlign); // extract the edges
477  align = vec_lvsr( 0, to ); // permute map to misalign data
478  MSQ = vec_perm(edges, (Packet16uc) from, align); // misalign the data (MSQ)
479  LSQ = vec_perm((Packet16uc) from, edges, align); // misalign the data (LSQ)
480  vec_st( LSQ, 15, (unsigned char *)to ); // Store the LSQ part first
481  vec_st( MSQ, 0, (unsigned char *)to ); // Store the MSQ part
482 }
483 #else
484 // We also need ot redefine little endian loading of Packet4i/Packet4f using VSX
485 template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from)
486 {
487  EIGEN_DEBUG_ALIGNED_STORE
488  vec_vsx_st(from, (long)to & 15, (int*) _EIGEN_ALIGNED_PTR(to));
489 }
490 template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from)
491 {
492  EIGEN_DEBUG_ALIGNED_STORE
493  vec_vsx_st(from, (long)to & 15, (float*) _EIGEN_ALIGNED_PTR(to));
494 }
495 #endif
496 
497 #ifndef __VSX__
498 template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); }
499 template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); }
500 #endif
501 
502 template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; }
503 template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; }
504 
505 template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) { return (Packet4f)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE32); }
506 template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) { return (Packet4i)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE32); }
507 
508 template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) { return vec_abs(a); }
509 template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vec_abs(a); }
510 
511 template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
512 {
513  Packet4f b, sum;
514  b = (Packet4f) vec_sld(a, a, 8);
515  sum = vec_add(a, b);
516  b = (Packet4f) vec_sld(sum, sum, 4);
517  sum = vec_add(sum, b);
518  return pfirst(sum);
519 }
520 
521 template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
522 {
523  Packet4f v[4], sum[4];
524 
525  // It's easier and faster to transpose then add as columns
526  // Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation
527  // Do the transpose, first set of moves
528  v[0] = vec_mergeh(vecs[0], vecs[2]);
529  v[1] = vec_mergel(vecs[0], vecs[2]);
530  v[2] = vec_mergeh(vecs[1], vecs[3]);
531  v[3] = vec_mergel(vecs[1], vecs[3]);
532  // Get the resulting vectors
533  sum[0] = vec_mergeh(v[0], v[2]);
534  sum[1] = vec_mergel(v[0], v[2]);
535  sum[2] = vec_mergeh(v[1], v[3]);
536  sum[3] = vec_mergel(v[1], v[3]);
537 
538  // Now do the summation:
539  // Lines 0+1
540  sum[0] = vec_add(sum[0], sum[1]);
541  // Lines 2+3
542  sum[1] = vec_add(sum[2], sum[3]);
543  // Add the results
544  sum[0] = vec_add(sum[0], sum[1]);
545 
546  return sum[0];
547 }
548 
549 template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
550 {
551  Packet4i sum;
552  sum = vec_sums(a, p4i_ZERO);
553 #ifdef _BIG_ENDIAN
554  sum = vec_sld(sum, p4i_ZERO, 12);
555 #else
556  sum = vec_sld(p4i_ZERO, sum, 4);
557 #endif
558  return pfirst(sum);
559 }
560 
561 template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
562 {
563  Packet4i v[4], sum[4];
564 
565  // It's easier and faster to transpose then add as columns
566  // Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation
567  // Do the transpose, first set of moves
568  v[0] = vec_mergeh(vecs[0], vecs[2]);
569  v[1] = vec_mergel(vecs[0], vecs[2]);
570  v[2] = vec_mergeh(vecs[1], vecs[3]);
571  v[3] = vec_mergel(vecs[1], vecs[3]);
572  // Get the resulting vectors
573  sum[0] = vec_mergeh(v[0], v[2]);
574  sum[1] = vec_mergel(v[0], v[2]);
575  sum[2] = vec_mergeh(v[1], v[3]);
576  sum[3] = vec_mergel(v[1], v[3]);
577 
578  // Now do the summation:
579  // Lines 0+1
580  sum[0] = vec_add(sum[0], sum[1]);
581  // Lines 2+3
582  sum[1] = vec_add(sum[2], sum[3]);
583  // Add the results
584  sum[0] = vec_add(sum[0], sum[1]);
585 
586  return sum[0];
587 }
588 
589 // Other reduction functions:
590 // mul
591 template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
592 {
593  Packet4f prod;
594  prod = pmul(a, (Packet4f)vec_sld(a, a, 8));
595  return pfirst(pmul(prod, (Packet4f)vec_sld(prod, prod, 4)));
596 }
597 
598 template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
599 {
600  EIGEN_ALIGN16 int aux[4];
601  pstore(aux, a);
602  return aux[0] * aux[1] * aux[2] * aux[3];
603 }
604 
605 // min
606 template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
607 {
608  Packet4f b, res;
609  b = vec_min(a, vec_sld(a, a, 8));
610  res = vec_min(b, vec_sld(b, b, 4));
611  return pfirst(res);
612 }
613 
614 template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a)
615 {
616  Packet4i b, res;
617  b = vec_min(a, vec_sld(a, a, 8));
618  res = vec_min(b, vec_sld(b, b, 4));
619  return pfirst(res);
620 }
621 
622 // max
623 template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
624 {
625  Packet4f b, res;
626  b = vec_max(a, vec_sld(a, a, 8));
627  res = vec_max(b, vec_sld(b, b, 4));
628  return pfirst(res);
629 }
630 
631 template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
632 {
633  Packet4i b, res;
634  b = vec_max(a, vec_sld(a, a, 8));
635  res = vec_max(b, vec_sld(b, b, 4));
636  return pfirst(res);
637 }
638 
639 template<int Offset>
640 struct palign_impl<Offset,Packet4f>
641 {
642  static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
643  {
644 #ifdef _BIG_ENDIAN
645  switch (Offset % 4) {
646  case 1:
647  first = vec_sld(first, second, 4); break;
648  case 2:
649  first = vec_sld(first, second, 8); break;
650  case 3:
651  first = vec_sld(first, second, 12); break;
652  }
653 #else
654  switch (Offset % 4) {
655  case 1:
656  first = vec_sld(second, first, 12); break;
657  case 2:
658  first = vec_sld(second, first, 8); break;
659  case 3:
660  first = vec_sld(second, first, 4); break;
661  }
662 #endif
663  }
664 };
665 
666 template<int Offset>
667 struct palign_impl<Offset,Packet4i>
668 {
669  static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second)
670  {
671 #ifdef _BIG_ENDIAN
672  switch (Offset % 4) {
673  case 1:
674  first = vec_sld(first, second, 4); break;
675  case 2:
676  first = vec_sld(first, second, 8); break;
677  case 3:
678  first = vec_sld(first, second, 12); break;
679  }
680 #else
681  switch (Offset % 4) {
682  case 1:
683  first = vec_sld(second, first, 12); break;
684  case 2:
685  first = vec_sld(second, first, 8); break;
686  case 3:
687  first = vec_sld(second, first, 4); break;
688  }
689 #endif
690  }
691 };
692 
693 EIGEN_DEVICE_FUNC inline void
694 ptranspose(PacketBlock<Packet4f,4>& kernel) {
695  Packet4f t0, t1, t2, t3;
696  t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]);
697  t1 = vec_mergel(kernel.packet[0], kernel.packet[2]);
698  t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]);
699  t3 = vec_mergel(kernel.packet[1], kernel.packet[3]);
700  kernel.packet[0] = vec_mergeh(t0, t2);
701  kernel.packet[1] = vec_mergel(t0, t2);
702  kernel.packet[2] = vec_mergeh(t1, t3);
703  kernel.packet[3] = vec_mergel(t1, t3);
704 }
705 
706 EIGEN_DEVICE_FUNC inline void
707 ptranspose(PacketBlock<Packet4i,4>& kernel) {
708  Packet4i t0, t1, t2, t3;
709  t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]);
710  t1 = vec_mergel(kernel.packet[0], kernel.packet[2]);
711  t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]);
712  t3 = vec_mergel(kernel.packet[1], kernel.packet[3]);
713  kernel.packet[0] = vec_mergeh(t0, t2);
714  kernel.packet[1] = vec_mergel(t0, t2);
715  kernel.packet[2] = vec_mergeh(t1, t3);
716  kernel.packet[3] = vec_mergel(t1, t3);
717 }
718 
719 
720 //---------- double ----------
721 #ifdef __VSX__
722 typedef __vector double Packet2d;
723 typedef __vector unsigned long long Packet2ul;
724 typedef __vector long long Packet2l;
725 
726 static Packet2l p2l_ZERO = (Packet2l) p4i_ZERO;
727 static Packet2d p2d_ONE = { 1.0, 1.0 };
728 static Packet2d p2d_ZERO = (Packet2d) p4f_ZERO;
729 static Packet2d p2d_ZERO_ = { -0.0, -0.0 };
730 
731 #ifdef _BIG_ENDIAN
732 static Packet2d p2d_COUNTDOWN = (Packet2d) vec_sld((Packet16uc) p2d_ZERO, (Packet16uc) p2d_ONE, 8);
733 #else
734 static Packet2d p2d_COUNTDOWN = (Packet2d) vec_sld((Packet16uc) p2d_ONE, (Packet16uc) p2d_ZERO, 8);
735 #endif
736 
737 static EIGEN_STRONG_INLINE Packet2d vec_splat_dbl(Packet2d& a, int index)
738 {
739  switch (index) {
740  case 0:
741  return (Packet2d) vec_perm(a, a, p16uc_PSET64_HI);
742  case 1:
743  return (Packet2d) vec_perm(a, a, p16uc_PSET64_LO);
744  }
745  return a;
746 }
747 
748 template<> struct packet_traits<double> : default_packet_traits
749 {
750  typedef Packet2d type;
751  typedef Packet2d half;
752  enum {
753  Vectorizable = 1,
754  AlignedOnScalar = 1,
755  size=2,
756  HasHalfPacket = 0,
757 
758  HasDiv = 1,
759  HasExp = 1,
760  HasSqrt = 0
761  };
762 };
763 
764 template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; };
765 
766 
767 inline std::ostream & operator <<(std::ostream & s, const Packet2d & v)
768 {
769  union {
770  Packet2d v;
771  double n[2];
772  } vt;
773  vt.v = v;
774  s << vt.n[0] << ", " << vt.n[1];
775  return s;
776 }
777 
778 // Need to define them first or we get specialization after instantiation errors
779 template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return (Packet2d) vec_ld(0, (const float *) from); } //FIXME
780 
781 template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st((Packet4f)from, 0, (float *)to); }
782 
783 template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) {
784  double EIGEN_ALIGN16 af[2];
785  af[0] = from;
786  Packet2d vc = pload<Packet2d>(af);
787  vc = vec_splat_dbl(vc, 0);
788  return vc;
789 }
790 template<> EIGEN_STRONG_INLINE void
791 pbroadcast4<Packet2d>(const double *a,
792  Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3)
793 {
794  a1 = pload<Packet2d>(a);
795  a0 = vec_splat_dbl(a1, 0);
796  a1 = vec_splat_dbl(a1, 1);
797  a3 = pload<Packet2d>(a+2);
798  a2 = vec_splat_dbl(a3, 0);
799  a3 = vec_splat_dbl(a3, 1);
800 }
801 template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride)
802 {
803  double EIGEN_ALIGN16 af[2];
804  af[0] = from[0*stride];
805  af[1] = from[1*stride];
806  return pload<Packet2d>(af);
807 }
808 template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride)
809 {
810  double EIGEN_ALIGN16 af[2];
811  pstore<double>(af, from);
812  to[0*stride] = af[0];
813  to[1*stride] = af[1];
814 }
815 template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) { return vec_add(pset1<Packet2d>(a), p2d_COUNTDOWN); }
816 
817 template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_add(a,b); }
818 
819 template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_sub(a,b); }
820 
821 template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return psub<Packet2d>(p2d_ZERO, a); }
822 
823 template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; }
824 
825 template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_madd(a,b,p2d_ZERO); }
826 template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_div(a,b); }
827 
828 // for some weird raisons, it has to be overloaded for packet of integers
829 template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vec_madd(a, b, c); }
830 
831 template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_min(a, b); }
832 
833 template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_max(a, b); }
834 
835 template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, b); }
836 
837 template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_or(a, b); }
838 
839 template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_xor(a, b); }
840 
841 template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, vec_nor(b, b)); }
842 
843 template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from)
844 {
845  EIGEN_DEBUG_ALIGNED_LOAD
846  return (Packet2d) vec_vsx_ld((long)from & 15, (const float*) _EIGEN_ALIGNED_PTR(from));
847 }
848 template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double* from)
849 {
850  Packet2d p;
851  if((ptrdiff_t(from) % 16) == 0) p = pload<Packet2d>(from);
852  else p = ploadu<Packet2d>(from);
853  return vec_perm(p, p, p16uc_PSET64_HI);
854 }
855 
856 template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from)
857 {
858  EIGEN_DEBUG_ALIGNED_STORE
859  vec_vsx_st((Packet4f)from, (long)to & 15, (float*) _EIGEN_ALIGNED_PTR(to));
860 }
861 
862 template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { vec_dstt((const float *) addr, DST_CTRL(2,2,32), DST_CHAN); }
863 
864 template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { double EIGEN_ALIGN16 x[2]; pstore(x, a); return x[0]; }
865 
866 template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) { return (Packet2d)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE64); }
867 
868 template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) { return vec_abs(a); }
869 
870 template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
871 {
872  Packet2d b, sum;
873  b = (Packet2d) vec_sld((Packet4ui) a, (Packet4ui)a, 8);
874  sum = vec_add(a, b);
875  return pfirst(sum);
876 }
877 
878 template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
879 {
880  Packet2d v[2], sum;
881  v[0] = vec_add(vecs[0], (Packet2d) vec_sld((Packet4ui) vecs[0], (Packet4ui) vecs[0], 8));
882  v[1] = vec_add(vecs[1], (Packet2d) vec_sld((Packet4ui) vecs[1], (Packet4ui) vecs[1], 8));
883 
884 #ifdef _BIG_ENDIAN
885  sum = (Packet2d) vec_sld((Packet4ui) v[0], (Packet4ui) v[1], 8);
886 #else
887  sum = (Packet2d) vec_sld((Packet4ui) v[1], (Packet4ui) v[0], 8);
888 #endif
889 
890  return sum;
891 }
892 // Other reduction functions:
893 // mul
894 template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a)
895 {
896  return pfirst(pmul(a, (Packet2d)vec_sld((Packet4ui) a, (Packet4ui) a, 8)));
897 }
898 
899 // min
900 template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a)
901 {
902  return pfirst(vec_min(a, (Packet2d) vec_sld((Packet4ui) a, (Packet4ui) a, 8)));
903 }
904 
905 // max
906 template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a)
907 {
908  return pfirst(vec_max(a, (Packet2d) vec_sld((Packet4ui) a, (Packet4ui) a, 8)));
909 }
910 
911 template<int Offset>
912 struct palign_impl<Offset,Packet2d>
913 {
914  static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second)
915  {
916  if (Offset == 1)
917 #ifdef _BIG_ENDIAN
918  first = (Packet2d) vec_sld((Packet4ui) first, (Packet4ui) second, 8);
919 #else
920  first = (Packet2d) vec_sld((Packet4ui) second, (Packet4ui) first, 8);
921 #endif
922  }
923 };
924 
925 EIGEN_DEVICE_FUNC inline void
926 ptranspose(PacketBlock<Packet2d,2>& kernel) {
927  Packet2d t0, t1;
928  t0 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_HI);
929  t1 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_LO);
930  kernel.packet[0] = t0;
931  kernel.packet[1] = t1;
932 }
933 
934 #endif // __VSX__
935 } // end namespace internal
936 
937 } // end namespace Eigen
938 
939 #endif // EIGEN_PACKET_MATH_ALTIVEC_H
Definition: LDLT.h:16
Definition: Constants.h:230
Definition: Eigen_Colamd.h:54