PacketMath.h
Go to the documentation of this file.
00001 // This file is part of Eigen, a lightweight C++ template library
00002 // for linear algebra.
00003 //
00004 // Copyright (C) 2008 Konstantinos Margaritis <markos@codex.gr>
00005 //
00006 // This Source Code Form is subject to the terms of the Mozilla
00007 // Public License v. 2.0. If a copy of the MPL was not distributed
00008 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
00009 
00010 #ifndef EIGEN_PACKET_MATH_ALTIVEC_H
00011 #define EIGEN_PACKET_MATH_ALTIVEC_H
00012 
00013 namespace Eigen {
00014 
00015 namespace internal {
00016 
00017 #ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
00018 #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 4
00019 #endif
00020 
00021 #ifndef EIGEN_HAS_FUSE_CJMADD
00022 #define EIGEN_HAS_FUSE_CJMADD 1
00023 #endif
00024 
00025 // NOTE Altivec has 32 registers, but Eigen only accepts a value of 8 or 16
00026 #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
00027 #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 16
00028 #endif
00029 
00030 typedef __vector float          Packet4f;
00031 typedef __vector int            Packet4i;
00032 typedef __vector unsigned int   Packet4ui;
00033 typedef __vector __bool int     Packet4bi;
00034 typedef __vector short int      Packet8i;
00035 typedef __vector unsigned char  Packet16uc;
00036 
00037 // We don't want to write the same code all the time, but we need to reuse the constants
00038 // and it doesn't really work to declare them global, so we define macros instead
00039 
00040 #define _EIGEN_DECLARE_CONST_FAST_Packet4f(NAME,X) \
00041   Packet4f p4f_##NAME = (Packet4f) vec_splat_s32(X)
00042 
00043 #define _EIGEN_DECLARE_CONST_FAST_Packet4i(NAME,X) \
00044   Packet4i p4i_##NAME = vec_splat_s32(X)
00045 
00046 #define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
00047   Packet4f p4f_##NAME = pset1<Packet4f>(X)
00048 
00049 #define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
00050   Packet4f p4f_##NAME = vreinterpretq_f32_u32(pset1<int>(X))
00051 
00052 #define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
00053   Packet4i p4i_##NAME = pset1<Packet4i>(X)
00054 
00055 #define DST_CHAN 1
00056 #define DST_CTRL(size, count, stride) (((size) << 24) | ((count) << 16) | (stride))
00057 
00058 // Define global static constants:
00059 static Packet4f p4f_COUNTDOWN = { 3.0, 2.0, 1.0, 0.0 };
00060 static Packet4i p4i_COUNTDOWN = { 3, 2, 1, 0 };
00061 static Packet16uc p16uc_REVERSE = {12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3};
00062 static Packet16uc p16uc_FORWARD = vec_lvsl(0, (float*)0);
00063 static Packet16uc p16uc_DUPLICATE = {0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7};
00064 
00065 static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0);
00066 static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0);
00067 static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE,1);
00068 static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS16,-16);
00069 static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1);
00070 static Packet4f p4f_ONE = vec_ctf(p4i_ONE, 0);
00071 static Packet4f p4f_ZERO_ = (Packet4f) vec_sl((Packet4ui)p4i_MINUS1, (Packet4ui)p4i_MINUS1);
00072 
00073 template<> struct packet_traits<float>  : default_packet_traits
00074 {
00075   typedef Packet4f type;
00076   enum {
00077     Vectorizable = 1,
00078     AlignedOnScalar = 1,
00079     size=4,
00080 
00081     // FIXME check the Has*
00082     HasSin  = 0,
00083     HasCos  = 0,
00084     HasLog  = 0,
00085     HasExp  = 0,
00086     HasSqrt = 0
00087   };
00088 };
00089 template<> struct packet_traits<int>    : default_packet_traits
00090 {
00091   typedef Packet4i type;
00092   enum {
00093     // FIXME check the Has*
00094     Vectorizable = 1,
00095     AlignedOnScalar = 1,
00096     size=4
00097   };
00098 };
00099 
00100 template<> struct unpacket_traits<Packet4f> { typedef float  type; enum {size=4}; };
00101 template<> struct unpacket_traits<Packet4i> { typedef int    type; enum {size=4}; };
00102 /*
00103 inline std::ostream & operator <<(std::ostream & s, const Packet4f & v)
00104 {
00105   union {
00106     Packet4f   v;
00107     float n[4];
00108   } vt;
00109   vt.v = v;
00110   s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
00111   return s;
00112 }
00113 
00114 inline std::ostream & operator <<(std::ostream & s, const Packet4i & v)
00115 {
00116   union {
00117     Packet4i   v;
00118     int n[4];
00119   } vt;
00120   vt.v = v;
00121   s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
00122   return s;
00123 }
00124 
00125 inline std::ostream & operator <<(std::ostream & s, const Packet4ui & v)
00126 {
00127   union {
00128     Packet4ui   v;
00129     unsigned int n[4];
00130   } vt;
00131   vt.v = v;
00132   s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
00133   return s;
00134 }
00135 
00136 inline std::ostream & operator <<(std::ostream & s, const Packetbi & v)
00137 {
00138   union {
00139     Packet4bi v;
00140     unsigned int n[4];
00141   } vt;
00142   vt.v = v;
00143   s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
00144   return s;
00145 }
00146 */
00147 template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&  from) {
00148   // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
00149   float EIGEN_ALIGN16 af[4];
00150   af[0] = from;
00151   Packet4f vc = vec_ld(0, af);
00152   vc = vec_splat(vc, 0);
00153   return vc;
00154 }
00155 
00156 template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int&    from)   {
00157   int EIGEN_ALIGN16 ai[4];
00158   ai[0] = from;
00159   Packet4i vc = vec_ld(0, ai);
00160   vc = vec_splat(vc, 0);
00161   return vc;
00162 }
00163 
00164 template<> EIGEN_STRONG_INLINE Packet4f plset<float>(const float& a) { return vec_add(pset1<Packet4f>(a), p4f_COUNTDOWN); }
00165 template<> EIGEN_STRONG_INLINE Packet4i plset<int>(const int& a)     { return vec_add(pset1<Packet4i>(a), p4i_COUNTDOWN); }
00166 
00167 template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_add(a,b); }
00168 template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_add(a,b); }
00169 
00170 template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_sub(a,b); }
00171 template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_sub(a,b); }
00172 
00173 template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) { return psub<Packet4f>(p4f_ZERO, a); }
00174 template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return psub<Packet4i>(p4i_ZERO, a); }
00175 
00176 template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
00177 template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; }
00178 
00179 template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_madd(a,b,p4f_ZERO); }
00180 /* Commented out: it's actually slower than processing it scalar
00181  *
00182 template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b)
00183 {
00184   // Detailed in: http://freevec.org/content/32bit_signed_integer_multiplication_altivec
00185   //Set up constants, variables
00186   Packet4i a1, b1, bswap, low_prod, high_prod, prod, prod_, v1sel;
00187 
00188   // Get the absolute values
00189   a1  = vec_abs(a);
00190   b1  = vec_abs(b);
00191 
00192   // Get the signs using xor
00193   Packet4bi sgn = (Packet4bi) vec_cmplt(vec_xor(a, b), p4i_ZERO);
00194 
00195   // Do the multiplication for the asbolute values.
00196   bswap = (Packet4i) vec_rl((Packet4ui) b1, (Packet4ui) p4i_MINUS16 );
00197   low_prod = vec_mulo((Packet8i) a1, (Packet8i)b1);
00198   high_prod = vec_msum((Packet8i) a1, (Packet8i) bswap, p4i_ZERO);
00199   high_prod = (Packet4i) vec_sl((Packet4ui) high_prod, (Packet4ui) p4i_MINUS16);
00200   prod = vec_add( low_prod, high_prod );
00201 
00202   // NOR the product and select only the negative elements according to the sign mask
00203   prod_ = vec_nor(prod, prod);
00204   prod_ = vec_sel(p4i_ZERO, prod_, sgn);
00205 
00206   // Add 1 to the result to get the negative numbers
00207   v1sel = vec_sel(p4i_ZERO, p4i_ONE, sgn);
00208   prod_ = vec_add(prod_, v1sel);
00209 
00210   // Merge the results back to the final vector.
00211   prod = vec_sel(prod, prod_, sgn);
00212 
00213   return prod;
00214 }
00215 */
00216 template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b)
00217 {
00218   Packet4f t, y_0, y_1, res;
00219 
00220   // Altivec does not offer a divide instruction, we have to do a reciprocal approximation
00221   y_0 = vec_re(b);
00222 
00223   // Do one Newton-Raphson iteration to get the needed accuracy
00224   t   = vec_nmsub(y_0, b, p4f_ONE);
00225   y_1 = vec_madd(y_0, t, y_0);
00226 
00227   res = vec_madd(a, y_1, p4f_ZERO);
00228   return res;
00229 }
00230 
00231 template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/)
00232 { eigen_assert(false && "packet integer division are not supported by AltiVec");
00233   return pset1<Packet4i>(0);
00234 }
00235 
00236 // for some weird raisons, it has to be overloaded for packet of integers
00237 template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vec_madd(a, b, c); }
00238 template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd(pmul(a,b), c); }
00239 
00240 template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_min(a, b); }
00241 template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_min(a, b); }
00242 
00243 template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_max(a, b); }
00244 template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); }
00245 
00246 // Logical Operations are not supported for float, so we have to reinterpret casts using NEON intrinsics
00247 template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, b); }
00248 template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, b); }
00249 
00250 template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_or(a, b); }
00251 template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_or(a, b); }
00252 
00253 template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_xor(a, b); }
00254 template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_xor(a, b); }
00255 
00256 template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, vec_nor(b, b)); }
00257 template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, vec_nor(b, b)); }
00258 
00259 template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); }
00260 template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*     from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); }
00261 
00262 template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
00263 {
00264   EIGEN_DEBUG_ALIGNED_LOAD
00265   // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
00266   Packet16uc MSQ, LSQ;
00267   Packet16uc mask;
00268   MSQ = vec_ld(0, (unsigned char *)from);          // most significant quadword
00269   LSQ = vec_ld(15, (unsigned char *)from);         // least significant quadword
00270   mask = vec_lvsl(0, from);                        // create the permute mask
00271   return (Packet4f) vec_perm(MSQ, LSQ, mask);           // align the data
00272 
00273 }
00274 template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
00275 {
00276   EIGEN_DEBUG_ALIGNED_LOAD
00277   // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
00278   Packet16uc MSQ, LSQ;
00279   Packet16uc mask;
00280   MSQ = vec_ld(0, (unsigned char *)from);          // most significant quadword
00281   LSQ = vec_ld(15, (unsigned char *)from);         // least significant quadword
00282   mask = vec_lvsl(0, from);                        // create the permute mask
00283   return (Packet4i) vec_perm(MSQ, LSQ, mask);    // align the data
00284 }
00285 
00286 template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float*   from)
00287 {
00288   Packet4f p;
00289   if((ptrdiff_t(&from) % 16) == 0)  p = pload<Packet4f>(from);
00290   else                              p = ploadu<Packet4f>(from);
00291   return vec_perm(p, p, p16uc_DUPLICATE);
00292 }
00293 template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int*     from)
00294 {
00295   Packet4i p;
00296   if((ptrdiff_t(&from) % 16) == 0)  p = pload<Packet4i>(from);
00297   else                              p = ploadu<Packet4i>(from);
00298   return vec_perm(p, p, p16uc_DUPLICATE);
00299 }
00300 
00301 template<> EIGEN_STRONG_INLINE void pstore<float>(float*   to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st(from, 0, to); }
00302 template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st(from, 0, to); }
00303 
00304 template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*  to, const Packet4f& from)
00305 {
00306   EIGEN_DEBUG_UNALIGNED_STORE
00307   // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
00308   // Warning: not thread safe!
00309   Packet16uc MSQ, LSQ, edges;
00310   Packet16uc edgeAlign, align;
00311 
00312   MSQ = vec_ld(0, (unsigned char *)to);                     // most significant quadword
00313   LSQ = vec_ld(15, (unsigned char *)to);                    // least significant quadword
00314   edgeAlign = vec_lvsl(0, to);                              // permute map to extract edges
00315   edges=vec_perm(LSQ,MSQ,edgeAlign);                        // extract the edges
00316   align = vec_lvsr( 0, to );                                // permute map to misalign data
00317   MSQ = vec_perm(edges,(Packet16uc)from,align);             // misalign the data (MSQ)
00318   LSQ = vec_perm((Packet16uc)from,edges,align);             // misalign the data (LSQ)
00319   vec_st( LSQ, 15, (unsigned char *)to );                   // Store the LSQ part first
00320   vec_st( MSQ, 0, (unsigned char *)to );                    // Store the MSQ part
00321 }
00322 template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*      to, const Packet4i& from)
00323 {
00324   EIGEN_DEBUG_UNALIGNED_STORE
00325   // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
00326   // Warning: not thread safe!
00327   Packet16uc MSQ, LSQ, edges;
00328   Packet16uc edgeAlign, align;
00329 
00330   MSQ = vec_ld(0, (unsigned char *)to);                     // most significant quadword
00331   LSQ = vec_ld(15, (unsigned char *)to);                    // least significant quadword
00332   edgeAlign = vec_lvsl(0, to);                              // permute map to extract edges
00333   edges=vec_perm(LSQ, MSQ, edgeAlign);                      // extract the edges
00334   align = vec_lvsr( 0, to );                                // permute map to misalign data
00335   MSQ = vec_perm(edges, (Packet16uc) from, align);          // misalign the data (MSQ)
00336   LSQ = vec_perm((Packet16uc) from, edges, align);          // misalign the data (LSQ)
00337   vec_st( LSQ, 15, (unsigned char *)to );                   // Store the LSQ part first
00338   vec_st( MSQ, 0, (unsigned char *)to );                    // Store the MSQ part
00339 }
00340 
00341 template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); }
00342 template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*     addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); }
00343 
00344 template<> EIGEN_STRONG_INLINE float  pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; }
00345 template<> EIGEN_STRONG_INLINE int    pfirst<Packet4i>(const Packet4i& a) { int   EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; }
00346 
00347 template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) { return (Packet4f)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE); }
00348 template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) { return (Packet4i)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE); }
00349 
00350 template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) { return vec_abs(a); }
00351 template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vec_abs(a); }
00352 
00353 template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
00354 {
00355   Packet4f b, sum;
00356   b   = (Packet4f) vec_sld(a, a, 8);
00357   sum = vec_add(a, b);
00358   b   = (Packet4f) vec_sld(sum, sum, 4);
00359   sum = vec_add(sum, b);
00360   return pfirst(sum);
00361 }
00362 
00363 template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
00364 {
00365   Packet4f v[4], sum[4];
00366 
00367   // It's easier and faster to transpose then add as columns
00368   // Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation
00369   // Do the transpose, first set of moves
00370   v[0] = vec_mergeh(vecs[0], vecs[2]);
00371   v[1] = vec_mergel(vecs[0], vecs[2]);
00372   v[2] = vec_mergeh(vecs[1], vecs[3]);
00373   v[3] = vec_mergel(vecs[1], vecs[3]);
00374   // Get the resulting vectors
00375   sum[0] = vec_mergeh(v[0], v[2]);
00376   sum[1] = vec_mergel(v[0], v[2]);
00377   sum[2] = vec_mergeh(v[1], v[3]);
00378   sum[3] = vec_mergel(v[1], v[3]);
00379 
00380   // Now do the summation:
00381   // Lines 0+1
00382   sum[0] = vec_add(sum[0], sum[1]);
00383   // Lines 2+3
00384   sum[1] = vec_add(sum[2], sum[3]);
00385   // Add the results
00386   sum[0] = vec_add(sum[0], sum[1]);
00387 
00388   return sum[0];
00389 }
00390 
00391 template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
00392 {
00393   Packet4i sum;
00394   sum = vec_sums(a, p4i_ZERO);
00395   sum = vec_sld(sum, p4i_ZERO, 12);
00396   return pfirst(sum);
00397 }
00398 
00399 template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
00400 {
00401   Packet4i v[4], sum[4];
00402 
00403   // It's easier and faster to transpose then add as columns
00404   // Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation
00405   // Do the transpose, first set of moves
00406   v[0] = vec_mergeh(vecs[0], vecs[2]);
00407   v[1] = vec_mergel(vecs[0], vecs[2]);
00408   v[2] = vec_mergeh(vecs[1], vecs[3]);
00409   v[3] = vec_mergel(vecs[1], vecs[3]);
00410   // Get the resulting vectors
00411   sum[0] = vec_mergeh(v[0], v[2]);
00412   sum[1] = vec_mergel(v[0], v[2]);
00413   sum[2] = vec_mergeh(v[1], v[3]);
00414   sum[3] = vec_mergel(v[1], v[3]);
00415 
00416   // Now do the summation:
00417   // Lines 0+1
00418   sum[0] = vec_add(sum[0], sum[1]);
00419   // Lines 2+3
00420   sum[1] = vec_add(sum[2], sum[3]);
00421   // Add the results
00422   sum[0] = vec_add(sum[0], sum[1]);
00423 
00424   return sum[0];
00425 }
00426 
00427 // Other reduction functions:
00428 // mul
00429 template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
00430 {
00431   Packet4f prod;
00432   prod = pmul(a, (Packet4f)vec_sld(a, a, 8));
00433   return pfirst(pmul(prod, (Packet4f)vec_sld(prod, prod, 4)));
00434 }
00435 
00436 template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
00437 {
00438   EIGEN_ALIGN16 int aux[4];
00439   pstore(aux, a);
00440   return aux[0] * aux[1] * aux[2] * aux[3];
00441 }
00442 
00443 // min
00444 template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
00445 {
00446   Packet4f b, res;
00447   b = vec_min(a, vec_sld(a, a, 8));
00448   res = vec_min(b, vec_sld(b, b, 4));
00449   return pfirst(res);
00450 }
00451 
00452 template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a)
00453 {
00454   Packet4i b, res;
00455   b = vec_min(a, vec_sld(a, a, 8));
00456   res = vec_min(b, vec_sld(b, b, 4));
00457   return pfirst(res);
00458 }
00459 
00460 // max
00461 template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
00462 {
00463   Packet4f b, res;
00464   b = vec_max(a, vec_sld(a, a, 8));
00465   res = vec_max(b, vec_sld(b, b, 4));
00466   return pfirst(res);
00467 }
00468 
00469 template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
00470 {
00471   Packet4i b, res;
00472   b = vec_max(a, vec_sld(a, a, 8));
00473   res = vec_max(b, vec_sld(b, b, 4));
00474   return pfirst(res);
00475 }
00476 
00477 template<int Offset>
00478 struct palign_impl<Offset,Packet4f>
00479 {
00480   static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
00481   {
00482     if (Offset!=0)
00483       first = vec_sld(first, second, Offset*4);
00484   }
00485 };
00486 
00487 template<int Offset>
00488 struct palign_impl<Offset,Packet4i>
00489 {
00490   static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second)
00491   {
00492     if (Offset!=0)
00493       first = vec_sld(first, second, Offset*4);
00494   }
00495 };
00496 
00497 } // end namespace internal
00498 
00499 } // end namespace Eigen
00500 
00501 #endif // EIGEN_PACKET_MATH_ALTIVEC_H


acado
Author(s): Milan Vukov, Rien Quirynen
autogenerated on Sat Jun 8 2019 19:38:15