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) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
00005 // Copyright (C) 2010 Konstantinos Margaritis <markos@codex.gr>
00006 // Heavily based on Gael's SSE version.
00007 //
00008 // This Source Code Form is subject to the terms of the Mozilla
00009 // Public License v. 2.0. If a copy of the MPL was not distributed
00010 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
00011 
00012 #ifndef EIGEN_PACKET_MATH_NEON_H
00013 #define EIGEN_PACKET_MATH_NEON_H
00014 
00015 namespace Eigen {
00016 
00017 namespace internal {
00018 
00019 #ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
00020 #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8
00021 #endif
00022 
00023 // FIXME NEON has 16 quad registers, but since the current register allocator
00024 // is so bad, it is much better to reduce it to 8
00025 #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
00026 #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 8
00027 #endif
00028 
00029 typedef float32x4_t Packet4f;
00030 typedef int32x4_t   Packet4i;
00031 typedef uint32x4_t  Packet4ui;
00032 
00033 #define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
00034   const Packet4f p4f_##NAME = pset1<Packet4f>(X)
00035 
00036 #define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
00037   const Packet4f p4f_##NAME = vreinterpretq_f32_u32(pset1<int>(X))
00038 
00039 #define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
00040   const Packet4i p4i_##NAME = pset1<Packet4i>(X)
00041 
00042 #if defined(__llvm__) && !defined(__clang__)
00043   //Special treatment for Apple's llvm-gcc, its NEON packet types are unions
00044   #define EIGEN_INIT_NEON_PACKET2(X, Y)       {{X, Y}}
00045   #define EIGEN_INIT_NEON_PACKET4(X, Y, Z, W) {{X, Y, Z, W}}
00046 #else
00047   //Default initializer for packets
00048   #define EIGEN_INIT_NEON_PACKET2(X, Y)       {X, Y}
00049   #define EIGEN_INIT_NEON_PACKET4(X, Y, Z, W) {X, Y, Z, W}
00050 #endif
00051     
00052 #ifndef __pld
00053 #define __pld(x) asm volatile ( "   pld [%[addr]]\n" :: [addr] "r" (x) : "cc" );
00054 #endif
00055 
00056 template<> struct packet_traits<float>  : default_packet_traits
00057 {
00058   typedef Packet4f type;
00059   enum {
00060     Vectorizable = 1,
00061     AlignedOnScalar = 1,
00062     size = 4,
00063    
00064     HasDiv  = 1,
00065     // FIXME check the Has*
00066     HasSin  = 0,
00067     HasCos  = 0,
00068     HasLog  = 0,
00069     HasExp  = 0,
00070     HasSqrt = 0
00071   };
00072 };
00073 template<> struct packet_traits<int>    : default_packet_traits
00074 {
00075   typedef Packet4i type;
00076   enum {
00077     Vectorizable = 1,
00078     AlignedOnScalar = 1,
00079     size=4
00080     // FIXME check the Has*
00081   };
00082 };
00083 
00084 #if EIGEN_GNUC_AT_MOST(4,4) && !defined(__llvm__)
00085 // workaround gcc 4.2, 4.3 and 4.4 compilatin issue
00086 EIGEN_STRONG_INLINE float32x4_t vld1q_f32(const float* x) { return ::vld1q_f32((const float32_t*)x); }
00087 EIGEN_STRONG_INLINE float32x2_t vld1_f32 (const float* x) { return ::vld1_f32 ((const float32_t*)x); }
00088 EIGEN_STRONG_INLINE void        vst1q_f32(float* to, float32x4_t from) { ::vst1q_f32((float32_t*)to,from); }
00089 EIGEN_STRONG_INLINE void        vst1_f32 (float* to, float32x2_t from) { ::vst1_f32 ((float32_t*)to,from); }
00090 #endif
00091 
00092 template<> struct unpacket_traits<Packet4f> { typedef float  type; enum {size=4}; };
00093 template<> struct unpacket_traits<Packet4i> { typedef int    type; enum {size=4}; };
00094 
00095 template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&  from) { return vdupq_n_f32(from); }
00096 template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int&    from)   { return vdupq_n_s32(from); }
00097 
00098 template<> EIGEN_STRONG_INLINE Packet4f plset<float>(const float& a)
00099 {
00100   Packet4f countdown = EIGEN_INIT_NEON_PACKET4(0, 1, 2, 3);
00101   return vaddq_f32(pset1<Packet4f>(a), countdown);
00102 }
00103 template<> EIGEN_STRONG_INLINE Packet4i plset<int>(const int& a)
00104 {
00105   Packet4i countdown = EIGEN_INIT_NEON_PACKET4(0, 1, 2, 3);
00106   return vaddq_s32(pset1<Packet4i>(a), countdown);
00107 }
00108 
00109 template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return vaddq_f32(a,b); }
00110 template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return vaddq_s32(a,b); }
00111 
00112 template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return vsubq_f32(a,b); }
00113 template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return vsubq_s32(a,b); }
00114 
00115 template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) { return vnegq_f32(a); }
00116 template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return vnegq_s32(a); }
00117 
00118 template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
00119 template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; }
00120 
00121 template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return vmulq_f32(a,b); }
00122 template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) { return vmulq_s32(a,b); }
00123 
00124 template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b)
00125 {
00126   Packet4f inv, restep, div;
00127 
00128   // NEON does not offer a divide instruction, we have to do a reciprocal approximation
00129   // However NEON in contrast to other SIMD engines (AltiVec/SSE), offers
00130   // a reciprocal estimate AND a reciprocal step -which saves a few instructions
00131   // vrecpeq_f32() returns an estimate to 1/b, which we will finetune with
00132   // Newton-Raphson and vrecpsq_f32()
00133   inv = vrecpeq_f32(b);
00134 
00135   // This returns a differential, by which we will have to multiply inv to get a better
00136   // approximation of 1/b.
00137   restep = vrecpsq_f32(b, inv);
00138   inv = vmulq_f32(restep, inv);
00139 
00140   // Finally, multiply a by 1/b and get the wanted result of the division.
00141   div = vmulq_f32(a, inv);
00142 
00143   return div;
00144 }
00145 template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/)
00146 { eigen_assert(false && "packet integer division are not supported by NEON");
00147   return pset1<Packet4i>(0);
00148 }
00149 
00150 // for some weird raisons, it has to be overloaded for packet of integers
00151 template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vmlaq_f32(c,a,b); }
00152 template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return vmlaq_s32(c,a,b); }
00153 
00154 template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return vminq_f32(a,b); }
00155 template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vminq_s32(a,b); }
00156 
00157 template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return vmaxq_f32(a,b); }
00158 template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vmaxq_s32(a,b); }
00159 
00160 // Logical Operations are not supported for float, so we have to reinterpret casts using NEON intrinsics
00161 template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b)
00162 {
00163   return vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b)));
00164 }
00165 template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vandq_s32(a,b); }
00166 
00167 template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b)
00168 {
00169   return vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b)));
00170 }
00171 template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return vorrq_s32(a,b); }
00172 
00173 template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b)
00174 {
00175   return vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b)));
00176 }
00177 template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return veorq_s32(a,b); }
00178 
00179 template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b)
00180 {
00181   return vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b)));
00182 }
00183 template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return vbicq_s32(a,b); }
00184 
00185 template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_f32(from); }
00186 template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*   from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_s32(from); }
00187 
00188 template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_f32(from); }
00189 template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)   { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_s32(from); }
00190 
00191 template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float*   from)
00192 {
00193   float32x2_t lo, hi;
00194   lo = vld1_dup_f32(from);
00195   hi = vld1_dup_f32(from+1);
00196   return vcombine_f32(lo, hi);
00197 }
00198 template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int*     from)
00199 {
00200   int32x2_t lo, hi;
00201   lo = vld1_dup_s32(from);
00202   hi = vld1_dup_s32(from+1);
00203   return vcombine_s32(lo, hi);
00204 }
00205 
00206 template<> EIGEN_STRONG_INLINE void pstore<float>(float*   to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_f32(to, from); }
00207 template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_s32(to, from); }
00208 
00209 template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*  to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f32(to, from); }
00210 template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*      to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_s32(to, from); }
00211 
00212 template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { __pld(addr); }
00213 template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*     addr) { __pld(addr); }
00214 
00215 // FIXME only store the 2 first elements ?
00216 template<> EIGEN_STRONG_INLINE float  pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; vst1q_f32(x, a); return x[0]; }
00217 template<> EIGEN_STRONG_INLINE int    pfirst<Packet4i>(const Packet4i& a) { int   EIGEN_ALIGN16 x[4]; vst1q_s32(x, a); return x[0]; }
00218 
00219 template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) {
00220   float32x2_t a_lo, a_hi;
00221   Packet4f a_r64;
00222 
00223   a_r64 = vrev64q_f32(a);
00224   a_lo = vget_low_f32(a_r64);
00225   a_hi = vget_high_f32(a_r64);
00226   return vcombine_f32(a_hi, a_lo);
00227 }
00228 template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) {
00229   int32x2_t a_lo, a_hi;
00230   Packet4i a_r64;
00231 
00232   a_r64 = vrev64q_s32(a);
00233   a_lo = vget_low_s32(a_r64);
00234   a_hi = vget_high_s32(a_r64);
00235   return vcombine_s32(a_hi, a_lo);
00236 }
00237 template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) { return vabsq_f32(a); }
00238 template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vabsq_s32(a); }
00239 
00240 template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
00241 {
00242   float32x2_t a_lo, a_hi, sum;
00243 
00244   a_lo = vget_low_f32(a);
00245   a_hi = vget_high_f32(a);
00246   sum = vpadd_f32(a_lo, a_hi);
00247   sum = vpadd_f32(sum, sum);
00248   return vget_lane_f32(sum, 0);
00249 }
00250 
00251 template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
00252 {
00253   float32x4x2_t vtrn1, vtrn2, res1, res2;
00254   Packet4f sum1, sum2, sum;
00255 
00256   // NEON zip performs interleaving of the supplied vectors.
00257   // We perform two interleaves in a row to acquire the transposed vector
00258   vtrn1 = vzipq_f32(vecs[0], vecs[2]);
00259   vtrn2 = vzipq_f32(vecs[1], vecs[3]);
00260   res1 = vzipq_f32(vtrn1.val[0], vtrn2.val[0]);
00261   res2 = vzipq_f32(vtrn1.val[1], vtrn2.val[1]);
00262 
00263   // Do the addition of the resulting vectors
00264   sum1 = vaddq_f32(res1.val[0], res1.val[1]);
00265   sum2 = vaddq_f32(res2.val[0], res2.val[1]);
00266   sum = vaddq_f32(sum1, sum2);
00267 
00268   return sum;
00269 }
00270 
00271 template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
00272 {
00273   int32x2_t a_lo, a_hi, sum;
00274 
00275   a_lo = vget_low_s32(a);
00276   a_hi = vget_high_s32(a);
00277   sum = vpadd_s32(a_lo, a_hi);
00278   sum = vpadd_s32(sum, sum);
00279   return vget_lane_s32(sum, 0);
00280 }
00281 
00282 template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
00283 {
00284   int32x4x2_t vtrn1, vtrn2, res1, res2;
00285   Packet4i sum1, sum2, sum;
00286 
00287   // NEON zip performs interleaving of the supplied vectors.
00288   // We perform two interleaves in a row to acquire the transposed vector
00289   vtrn1 = vzipq_s32(vecs[0], vecs[2]);
00290   vtrn2 = vzipq_s32(vecs[1], vecs[3]);
00291   res1 = vzipq_s32(vtrn1.val[0], vtrn2.val[0]);
00292   res2 = vzipq_s32(vtrn1.val[1], vtrn2.val[1]);
00293 
00294   // Do the addition of the resulting vectors
00295   sum1 = vaddq_s32(res1.val[0], res1.val[1]);
00296   sum2 = vaddq_s32(res2.val[0], res2.val[1]);
00297   sum = vaddq_s32(sum1, sum2);
00298 
00299   return sum;
00300 }
00301 
00302 // Other reduction functions:
00303 // mul
00304 template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
00305 {
00306   float32x2_t a_lo, a_hi, prod;
00307 
00308   // Get a_lo = |a1|a2| and a_hi = |a3|a4|
00309   a_lo = vget_low_f32(a);
00310   a_hi = vget_high_f32(a);
00311   // Get the product of a_lo * a_hi -> |a1*a3|a2*a4|
00312   prod = vmul_f32(a_lo, a_hi);
00313   // Multiply prod with its swapped value |a2*a4|a1*a3|
00314   prod = vmul_f32(prod, vrev64_f32(prod));
00315 
00316   return vget_lane_f32(prod, 0);
00317 }
00318 template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
00319 {
00320   int32x2_t a_lo, a_hi, prod;
00321 
00322   // Get a_lo = |a1|a2| and a_hi = |a3|a4|
00323   a_lo = vget_low_s32(a);
00324   a_hi = vget_high_s32(a);
00325   // Get the product of a_lo * a_hi -> |a1*a3|a2*a4|
00326   prod = vmul_s32(a_lo, a_hi);
00327   // Multiply prod with its swapped value |a2*a4|a1*a3|
00328   prod = vmul_s32(prod, vrev64_s32(prod));
00329 
00330   return vget_lane_s32(prod, 0);
00331 }
00332 
00333 // min
00334 template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
00335 {
00336   float32x2_t a_lo, a_hi, min;
00337 
00338   a_lo = vget_low_f32(a);
00339   a_hi = vget_high_f32(a);
00340   min = vpmin_f32(a_lo, a_hi);
00341   min = vpmin_f32(min, min);
00342 
00343   return vget_lane_f32(min, 0);
00344 }
00345 
00346 template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a)
00347 {
00348   int32x2_t a_lo, a_hi, min;
00349 
00350   a_lo = vget_low_s32(a);
00351   a_hi = vget_high_s32(a);
00352   min = vpmin_s32(a_lo, a_hi);
00353   min = vpmin_s32(min, min);
00354   
00355   return vget_lane_s32(min, 0);
00356 }
00357 
00358 // max
00359 template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
00360 {
00361   float32x2_t a_lo, a_hi, max;
00362 
00363   a_lo = vget_low_f32(a);
00364   a_hi = vget_high_f32(a);
00365   max = vpmax_f32(a_lo, a_hi);
00366   max = vpmax_f32(max, max);
00367 
00368   return vget_lane_f32(max, 0);
00369 }
00370 
00371 template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
00372 {
00373   int32x2_t a_lo, a_hi, max;
00374 
00375   a_lo = vget_low_s32(a);
00376   a_hi = vget_high_s32(a);
00377   max = vpmax_s32(a_lo, a_hi);
00378 
00379   return vget_lane_s32(max, 0);
00380 }
00381 
00382 // this PALIGN_NEON business is to work around a bug in LLVM Clang 3.0 causing incorrect compilation errors,
00383 // see bug 347 and this LLVM bug: http://llvm.org/bugs/show_bug.cgi?id=11074
00384 #define PALIGN_NEON(Offset,Type,Command) \
00385 template<>\
00386 struct palign_impl<Offset,Type>\
00387 {\
00388     EIGEN_STRONG_INLINE static void run(Type& first, const Type& second)\
00389     {\
00390         if (Offset!=0)\
00391             first = Command(first, second, Offset);\
00392     }\
00393 };\
00394 
00395 PALIGN_NEON(0,Packet4f,vextq_f32)
00396 PALIGN_NEON(1,Packet4f,vextq_f32)
00397 PALIGN_NEON(2,Packet4f,vextq_f32)
00398 PALIGN_NEON(3,Packet4f,vextq_f32)
00399 PALIGN_NEON(0,Packet4i,vextq_s32)
00400 PALIGN_NEON(1,Packet4i,vextq_s32)
00401 PALIGN_NEON(2,Packet4i,vextq_s32)
00402 PALIGN_NEON(3,Packet4i,vextq_s32)
00403     
00404 #undef PALIGN_NEON
00405 
00406 } // end namespace internal
00407 
00408 } // end namespace Eigen
00409 
00410 #endif // EIGEN_PACKET_MATH_NEON_H


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