00001
00002
00003
00004
00005
00006
00007
00008
00009
00010
00011
00012
00013
00014
00015
00016
00017
00018
00019
00020
00021
00022
00023
00024
00025
00026
00027 #ifndef EIGEN_PACKET_MATH_NEON_H
00028 #define EIGEN_PACKET_MATH_NEON_H
00029
00030 namespace Eigen {
00031
00032 namespace internal {
00033
00034 #ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
00035 #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8
00036 #endif
00037
00038
00039
00040 #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
00041 #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 8
00042 #endif
00043
00044 typedef float32x4_t Packet4f;
00045 typedef int32x4_t Packet4i;
00046 typedef uint32x4_t Packet4ui;
00047
00048 #define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
00049 const Packet4f p4f_##NAME = pset1<Packet4f>(X)
00050
00051 #define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
00052 const Packet4f p4f_##NAME = vreinterpretq_f32_u32(pset1<int>(X))
00053
00054 #define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
00055 const Packet4i p4i_##NAME = pset1<Packet4i>(X)
00056
00057 #if defined(__llvm__) && !defined(__clang__)
00058
00059 #define EIGEN_INIT_NEON_PACKET2(X, Y) {{X, Y}}
00060 #define EIGEN_INIT_NEON_PACKET4(X, Y, Z, W) {{X, Y, Z, W}}
00061 #else
00062
00063 #define EIGEN_INIT_NEON_PACKET2(X, Y) {X, Y}
00064 #define EIGEN_INIT_NEON_PACKET4(X, Y, Z, W) {X, Y, Z, W}
00065 #endif
00066
00067 #ifndef __pld
00068 #define __pld(x) asm volatile ( " pld [%[addr]]\n" :: [addr] "r" (x) : "cc" );
00069 #endif
00070
00071 template<> struct packet_traits<float> : default_packet_traits
00072 {
00073 typedef Packet4f type;
00074 enum {
00075 Vectorizable = 1,
00076 AlignedOnScalar = 1,
00077 size = 4,
00078
00079 HasDiv = 1,
00080
00081 HasSin = 0,
00082 HasCos = 0,
00083 HasLog = 0,
00084 HasExp = 0,
00085 HasSqrt = 0
00086 };
00087 };
00088 template<> struct packet_traits<int> : default_packet_traits
00089 {
00090 typedef Packet4i type;
00091 enum {
00092 Vectorizable = 1,
00093 AlignedOnScalar = 1,
00094 size=4
00095
00096 };
00097 };
00098
00099 #if EIGEN_GNUC_AT_MOST(4,4) && !defined(__llvm__)
00100
00101 EIGEN_STRONG_INLINE float32x4_t vld1q_f32(const float* x) { return ::vld1q_f32((const float32_t*)x); }
00102 EIGEN_STRONG_INLINE float32x2_t vld1_f32 (const float* x) { return ::vld1_f32 ((const float32_t*)x); }
00103 EIGEN_STRONG_INLINE void vst1q_f32(float* to, float32x4_t from) { ::vst1q_f32((float32_t*)to,from); }
00104 EIGEN_STRONG_INLINE void vst1_f32 (float* to, float32x2_t from) { ::vst1_f32 ((float32_t*)to,from); }
00105 #endif
00106
00107 template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4}; };
00108 template<> struct unpacket_traits<Packet4i> { typedef int type; enum {size=4}; };
00109
00110 template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { return vdupq_n_f32(from); }
00111 template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { return vdupq_n_s32(from); }
00112
00113 template<> EIGEN_STRONG_INLINE Packet4f plset<float>(const float& a)
00114 {
00115 Packet4f countdown = EIGEN_INIT_NEON_PACKET4(0, 1, 2, 3);
00116 return vaddq_f32(pset1<Packet4f>(a), countdown);
00117 }
00118 template<> EIGEN_STRONG_INLINE Packet4i plset<int>(const int& a)
00119 {
00120 Packet4i countdown = EIGEN_INIT_NEON_PACKET4(0, 1, 2, 3);
00121 return vaddq_s32(pset1<Packet4i>(a), countdown);
00122 }
00123
00124 template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return vaddq_f32(a,b); }
00125 template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return vaddq_s32(a,b); }
00126
00127 template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return vsubq_f32(a,b); }
00128 template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return vsubq_s32(a,b); }
00129
00130 template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) { return vnegq_f32(a); }
00131 template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return vnegq_s32(a); }
00132
00133 template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return vmulq_f32(a,b); }
00134 template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) { return vmulq_s32(a,b); }
00135
00136 template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b)
00137 {
00138 Packet4f inv, restep, div;
00139
00140
00141
00142
00143
00144
00145 inv = vrecpeq_f32(b);
00146
00147
00148
00149 restep = vrecpsq_f32(b, inv);
00150 inv = vmulq_f32(restep, inv);
00151
00152
00153 div = vmulq_f32(a, inv);
00154
00155 return div;
00156 }
00157 template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& , const Packet4i& )
00158 { eigen_assert(false && "packet integer division are not supported by NEON");
00159 return pset1<Packet4i>(0);
00160 }
00161
00162
00163 template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vmlaq_f32(c,a,b); }
00164 template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return vmlaq_s32(c,a,b); }
00165
00166 template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return vminq_f32(a,b); }
00167 template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vminq_s32(a,b); }
00168
00169 template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return vmaxq_f32(a,b); }
00170 template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vmaxq_s32(a,b); }
00171
00172
00173 template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b)
00174 {
00175 return vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b)));
00176 }
00177 template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vandq_s32(a,b); }
00178
00179 template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b)
00180 {
00181 return vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b)));
00182 }
00183 template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return vorrq_s32(a,b); }
00184
00185 template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b)
00186 {
00187 return vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b)));
00188 }
00189 template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return veorq_s32(a,b); }
00190
00191 template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b)
00192 {
00193 return vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b)));
00194 }
00195 template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return vbicq_s32(a,b); }
00196
00197 template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_f32(from); }
00198 template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_s32(from); }
00199
00200 template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_f32(from); }
00201 template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_s32(from); }
00202
00203 template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from)
00204 {
00205 float32x2_t lo, hi;
00206 lo = vdup_n_f32(*from);
00207 hi = vdup_n_f32(*(from+1));
00208 return vcombine_f32(lo, hi);
00209 }
00210 template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int* from)
00211 {
00212 int32x2_t lo, hi;
00213 lo = vdup_n_s32(*from);
00214 hi = vdup_n_s32(*(from+1));
00215 return vcombine_s32(lo, hi);
00216 }
00217
00218 template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_f32(to, from); }
00219 template<> EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_s32(to, from); }
00220
00221 template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f32(to, from); }
00222 template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_s32(to, from); }
00223
00224 template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { __pld(addr); }
00225 template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { __pld(addr); }
00226
00227
00228 template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; vst1q_f32(x, a); return x[0]; }
00229 template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int EIGEN_ALIGN16 x[4]; vst1q_s32(x, a); return x[0]; }
00230
00231 template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) {
00232 float32x2_t a_lo, a_hi;
00233 Packet4f a_r64;
00234
00235 a_r64 = vrev64q_f32(a);
00236 a_lo = vget_low_f32(a_r64);
00237 a_hi = vget_high_f32(a_r64);
00238 return vcombine_f32(a_hi, a_lo);
00239 }
00240 template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) {
00241 int32x2_t a_lo, a_hi;
00242 Packet4i a_r64;
00243
00244 a_r64 = vrev64q_s32(a);
00245 a_lo = vget_low_s32(a_r64);
00246 a_hi = vget_high_s32(a_r64);
00247 return vcombine_s32(a_hi, a_lo);
00248 }
00249 template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) { return vabsq_f32(a); }
00250 template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vabsq_s32(a); }
00251
00252 template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
00253 {
00254 float32x2_t a_lo, a_hi, sum;
00255 float s[2];
00256
00257 a_lo = vget_low_f32(a);
00258 a_hi = vget_high_f32(a);
00259 sum = vpadd_f32(a_lo, a_hi);
00260 sum = vpadd_f32(sum, sum);
00261 vst1_f32(s, sum);
00262
00263 return s[0];
00264 }
00265
00266 template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
00267 {
00268 float32x4x2_t vtrn1, vtrn2, res1, res2;
00269 Packet4f sum1, sum2, sum;
00270
00271
00272
00273 vtrn1 = vzipq_f32(vecs[0], vecs[2]);
00274 vtrn2 = vzipq_f32(vecs[1], vecs[3]);
00275 res1 = vzipq_f32(vtrn1.val[0], vtrn2.val[0]);
00276 res2 = vzipq_f32(vtrn1.val[1], vtrn2.val[1]);
00277
00278
00279 sum1 = vaddq_f32(res1.val[0], res1.val[1]);
00280 sum2 = vaddq_f32(res2.val[0], res2.val[1]);
00281 sum = vaddq_f32(sum1, sum2);
00282
00283 return sum;
00284 }
00285
00286 template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
00287 {
00288 int32x2_t a_lo, a_hi, sum;
00289 int32_t s[2];
00290
00291 a_lo = vget_low_s32(a);
00292 a_hi = vget_high_s32(a);
00293 sum = vpadd_s32(a_lo, a_hi);
00294 sum = vpadd_s32(sum, sum);
00295 vst1_s32(s, sum);
00296
00297 return s[0];
00298 }
00299
00300 template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
00301 {
00302 int32x4x2_t vtrn1, vtrn2, res1, res2;
00303 Packet4i sum1, sum2, sum;
00304
00305
00306
00307 vtrn1 = vzipq_s32(vecs[0], vecs[2]);
00308 vtrn2 = vzipq_s32(vecs[1], vecs[3]);
00309 res1 = vzipq_s32(vtrn1.val[0], vtrn2.val[0]);
00310 res2 = vzipq_s32(vtrn1.val[1], vtrn2.val[1]);
00311
00312
00313 sum1 = vaddq_s32(res1.val[0], res1.val[1]);
00314 sum2 = vaddq_s32(res2.val[0], res2.val[1]);
00315 sum = vaddq_s32(sum1, sum2);
00316
00317 return sum;
00318 }
00319
00320
00321
00322 template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
00323 {
00324 float32x2_t a_lo, a_hi, prod;
00325 float s[2];
00326
00327
00328 a_lo = vget_low_f32(a);
00329 a_hi = vget_high_f32(a);
00330
00331 prod = vmul_f32(a_lo, a_hi);
00332
00333 prod = vmul_f32(prod, vrev64_f32(prod));
00334 vst1_f32(s, prod);
00335
00336 return s[0];
00337 }
00338 template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
00339 {
00340 int32x2_t a_lo, a_hi, prod;
00341 int32_t s[2];
00342
00343
00344 a_lo = vget_low_s32(a);
00345 a_hi = vget_high_s32(a);
00346
00347 prod = vmul_s32(a_lo, a_hi);
00348
00349 prod = vmul_s32(prod, vrev64_s32(prod));
00350 vst1_s32(s, prod);
00351
00352 return s[0];
00353 }
00354
00355
00356 template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
00357 {
00358 float32x2_t a_lo, a_hi, min;
00359 float s[2];
00360
00361 a_lo = vget_low_f32(a);
00362 a_hi = vget_high_f32(a);
00363 min = vpmin_f32(a_lo, a_hi);
00364 min = vpmin_f32(min, min);
00365 vst1_f32(s, min);
00366
00367 return s[0];
00368 }
00369 template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a)
00370 {
00371 int32x2_t a_lo, a_hi, min;
00372 int32_t s[2];
00373
00374 a_lo = vget_low_s32(a);
00375 a_hi = vget_high_s32(a);
00376 min = vpmin_s32(a_lo, a_hi);
00377 min = vpmin_s32(min, min);
00378 vst1_s32(s, min);
00379
00380 return s[0];
00381 }
00382
00383
00384 template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
00385 {
00386 float32x2_t a_lo, a_hi, max;
00387 float s[2];
00388
00389 a_lo = vget_low_f32(a);
00390 a_hi = vget_high_f32(a);
00391 max = vpmax_f32(a_lo, a_hi);
00392 max = vpmax_f32(max, max);
00393 vst1_f32(s, max);
00394
00395 return s[0];
00396 }
00397 template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
00398 {
00399 int32x2_t a_lo, a_hi, max;
00400 int32_t s[2];
00401
00402 a_lo = vget_low_s32(a);
00403 a_hi = vget_high_s32(a);
00404 max = vpmax_s32(a_lo, a_hi);
00405 max = vpmax_s32(max, max);
00406 vst1_s32(s, max);
00407
00408 return s[0];
00409 }
00410
00411
00412
00413 #define PALIGN_NEON(Offset,Type,Command) \
00414 template<>\
00415 struct palign_impl<Offset,Type>\
00416 {\
00417 EIGEN_STRONG_INLINE static void run(Type& first, const Type& second)\
00418 {\
00419 if (Offset!=0)\
00420 first = Command(first, second, Offset);\
00421 }\
00422 };\
00423
00424 PALIGN_NEON(0,Packet4f,vextq_f32)
00425 PALIGN_NEON(1,Packet4f,vextq_f32)
00426 PALIGN_NEON(2,Packet4f,vextq_f32)
00427 PALIGN_NEON(3,Packet4f,vextq_f32)
00428 PALIGN_NEON(0,Packet4i,vextq_s32)
00429 PALIGN_NEON(1,Packet4i,vextq_s32)
00430 PALIGN_NEON(2,Packet4i,vextq_s32)
00431 PALIGN_NEON(3,Packet4i,vextq_s32)
00432
00433 #undef PALIGN_NEON
00434
00435 }
00436
00437 }
00438
00439 #endif // EIGEN_PACKET_MATH_NEON_H