1// This file is part of Eigen, a lightweight C++ template library 2// for linear algebra. 3// 4// Copyright (C) 2014 Benoit Steiner (benoit.steiner.goog@gmail.com) 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_AVX_H 11#define EIGEN_PACKET_MATH_AVX_H 12 13namespace Eigen { 14 15namespace internal { 16 17#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 18#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8 19#endif 20 21#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 22#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*)) 23#endif 24 25#ifdef __FMA__ 26#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD 27#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD 28#endif 29#endif 30 31typedef __m256 Packet8f; 32typedef __m256i Packet8i; 33typedef __m256d Packet4d; 34 35template<> struct is_arithmetic<__m256> { enum { value = true }; }; 36template<> struct is_arithmetic<__m256i> { enum { value = true }; }; 37template<> struct is_arithmetic<__m256d> { enum { value = true }; }; 38 39#define _EIGEN_DECLARE_CONST_Packet8f(NAME,X) \ 40 const Packet8f p8f_##NAME = pset1<Packet8f>(X) 41 42#define _EIGEN_DECLARE_CONST_Packet4d(NAME,X) \ 43 const Packet4d p4d_##NAME = pset1<Packet4d>(X) 44 45#define _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(NAME,X) \ 46 const Packet8f p8f_##NAME = _mm256_castsi256_ps(pset1<Packet8i>(X)) 47 48#define _EIGEN_DECLARE_CONST_Packet8i(NAME,X) \ 49 const Packet8i p8i_##NAME = pset1<Packet8i>(X) 50 51// Use the packet_traits defined in AVX512/PacketMath.h instead if we're going 52// to leverage AVX512 instructions. 53#ifndef EIGEN_VECTORIZE_AVX512 54template<> struct packet_traits<float> : default_packet_traits 55{ 56 typedef Packet8f type; 57 typedef Packet4f half; 58 enum { 59 Vectorizable = 1, 60 AlignedOnScalar = 1, 61 size=8, 62 HasHalfPacket = 1, 63 64 HasDiv = 1, 65 HasSin = EIGEN_FAST_MATH, 66 HasCos = 0, 67 HasLog = 1, 68 HasExp = 1, 69 HasSqrt = 1, 70 HasRsqrt = 1, 71 HasTanh = EIGEN_FAST_MATH, 72 HasBlend = 1, 73 HasRound = 1, 74 HasFloor = 1, 75 HasCeil = 1 76 }; 77}; 78template<> struct packet_traits<double> : default_packet_traits 79{ 80 typedef Packet4d type; 81 typedef Packet2d half; 82 enum { 83 Vectorizable = 1, 84 AlignedOnScalar = 1, 85 size=4, 86 HasHalfPacket = 1, 87 88 HasDiv = 1, 89 HasExp = 1, 90 HasSqrt = 1, 91 HasRsqrt = 1, 92 HasBlend = 1, 93 HasRound = 1, 94 HasFloor = 1, 95 HasCeil = 1 96 }; 97}; 98#endif 99 100template<> struct scalar_div_cost<float,true> { enum { value = 14 }; }; 101template<> struct scalar_div_cost<double,true> { enum { value = 16 }; }; 102 103/* Proper support for integers is only provided by AVX2. In the meantime, we'll 104 use SSE instructions and packets to deal with integers. 105template<> struct packet_traits<int> : default_packet_traits 106{ 107 typedef Packet8i type; 108 enum { 109 Vectorizable = 1, 110 AlignedOnScalar = 1, 111 size=8 112 }; 113}; 114*/ 115 116template<> struct unpacket_traits<Packet8f> { typedef float type; typedef Packet4f half; enum {size=8, alignment=Aligned32}; }; 117template<> struct unpacket_traits<Packet4d> { typedef double type; typedef Packet2d half; enum {size=4, alignment=Aligned32}; }; 118template<> struct unpacket_traits<Packet8i> { typedef int type; typedef Packet4i half; enum {size=8, alignment=Aligned32}; }; 119 120template<> EIGEN_STRONG_INLINE Packet8f pset1<Packet8f>(const float& from) { return _mm256_set1_ps(from); } 121template<> EIGEN_STRONG_INLINE Packet4d pset1<Packet4d>(const double& from) { return _mm256_set1_pd(from); } 122template<> EIGEN_STRONG_INLINE Packet8i pset1<Packet8i>(const int& from) { return _mm256_set1_epi32(from); } 123 124template<> EIGEN_STRONG_INLINE Packet8f pload1<Packet8f>(const float* from) { return _mm256_broadcast_ss(from); } 125template<> EIGEN_STRONG_INLINE Packet4d pload1<Packet4d>(const double* from) { return _mm256_broadcast_sd(from); } 126 127template<> EIGEN_STRONG_INLINE Packet8f plset<Packet8f>(const float& a) { return _mm256_add_ps(_mm256_set1_ps(a), _mm256_set_ps(7.0,6.0,5.0,4.0,3.0,2.0,1.0,0.0)); } 128template<> EIGEN_STRONG_INLINE Packet4d plset<Packet4d>(const double& a) { return _mm256_add_pd(_mm256_set1_pd(a), _mm256_set_pd(3.0,2.0,1.0,0.0)); } 129 130template<> EIGEN_STRONG_INLINE Packet8f padd<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_add_ps(a,b); } 131template<> EIGEN_STRONG_INLINE Packet4d padd<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_add_pd(a,b); } 132 133template<> EIGEN_STRONG_INLINE Packet8f psub<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_sub_ps(a,b); } 134template<> EIGEN_STRONG_INLINE Packet4d psub<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_sub_pd(a,b); } 135 136template<> EIGEN_STRONG_INLINE Packet8f pnegate(const Packet8f& a) 137{ 138 return _mm256_sub_ps(_mm256_set1_ps(0.0),a); 139} 140template<> EIGEN_STRONG_INLINE Packet4d pnegate(const Packet4d& a) 141{ 142 return _mm256_sub_pd(_mm256_set1_pd(0.0),a); 143} 144 145template<> EIGEN_STRONG_INLINE Packet8f pconj(const Packet8f& a) { return a; } 146template<> EIGEN_STRONG_INLINE Packet4d pconj(const Packet4d& a) { return a; } 147template<> EIGEN_STRONG_INLINE Packet8i pconj(const Packet8i& a) { return a; } 148 149template<> EIGEN_STRONG_INLINE Packet8f pmul<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_mul_ps(a,b); } 150template<> EIGEN_STRONG_INLINE Packet4d pmul<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_mul_pd(a,b); } 151 152 153template<> EIGEN_STRONG_INLINE Packet8f pdiv<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_div_ps(a,b); } 154template<> EIGEN_STRONG_INLINE Packet4d pdiv<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_div_pd(a,b); } 155template<> EIGEN_STRONG_INLINE Packet8i pdiv<Packet8i>(const Packet8i& /*a*/, const Packet8i& /*b*/) 156{ eigen_assert(false && "packet integer division are not supported by AVX"); 157 return pset1<Packet8i>(0); 158} 159 160#ifdef __FMA__ 161template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f& b, const Packet8f& c) { 162#if ( EIGEN_COMP_GNUC_STRICT || (EIGEN_COMP_CLANG && (EIGEN_COMP_CLANG<308)) ) 163 // clang stupidly generates a vfmadd213ps instruction plus some vmovaps on registers, 164 // and gcc stupidly generates a vfmadd132ps instruction, 165 // so let's enforce it to generate a vfmadd231ps instruction since the most common use case is to accumulate 166 // the result of the product. 167 Packet8f res = c; 168 __asm__("vfmadd231ps %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b)); 169 return res; 170#else 171 return _mm256_fmadd_ps(a,b,c); 172#endif 173} 174template<> EIGEN_STRONG_INLINE Packet4d pmadd(const Packet4d& a, const Packet4d& b, const Packet4d& c) { 175#if ( EIGEN_COMP_GNUC_STRICT || (EIGEN_COMP_CLANG && (EIGEN_COMP_CLANG<308)) ) 176 // see above 177 Packet4d res = c; 178 __asm__("vfmadd231pd %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b)); 179 return res; 180#else 181 return _mm256_fmadd_pd(a,b,c); 182#endif 183} 184#endif 185 186template<> EIGEN_STRONG_INLINE Packet8f pmin<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_min_ps(a,b); } 187template<> EIGEN_STRONG_INLINE Packet4d pmin<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_min_pd(a,b); } 188 189template<> EIGEN_STRONG_INLINE Packet8f pmax<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_max_ps(a,b); } 190template<> EIGEN_STRONG_INLINE Packet4d pmax<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_max_pd(a,b); } 191 192template<> EIGEN_STRONG_INLINE Packet8f pround<Packet8f>(const Packet8f& a) { return _mm256_round_ps(a, _MM_FROUND_CUR_DIRECTION); } 193template<> EIGEN_STRONG_INLINE Packet4d pround<Packet4d>(const Packet4d& a) { return _mm256_round_pd(a, _MM_FROUND_CUR_DIRECTION); } 194 195template<> EIGEN_STRONG_INLINE Packet8f pceil<Packet8f>(const Packet8f& a) { return _mm256_ceil_ps(a); } 196template<> EIGEN_STRONG_INLINE Packet4d pceil<Packet4d>(const Packet4d& a) { return _mm256_ceil_pd(a); } 197 198template<> EIGEN_STRONG_INLINE Packet8f pfloor<Packet8f>(const Packet8f& a) { return _mm256_floor_ps(a); } 199template<> EIGEN_STRONG_INLINE Packet4d pfloor<Packet4d>(const Packet4d& a) { return _mm256_floor_pd(a); } 200 201template<> EIGEN_STRONG_INLINE Packet8f pand<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_and_ps(a,b); } 202template<> EIGEN_STRONG_INLINE Packet4d pand<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_and_pd(a,b); } 203 204template<> EIGEN_STRONG_INLINE Packet8f por<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_or_ps(a,b); } 205template<> EIGEN_STRONG_INLINE Packet4d por<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_or_pd(a,b); } 206 207template<> EIGEN_STRONG_INLINE Packet8f pxor<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_xor_ps(a,b); } 208template<> EIGEN_STRONG_INLINE Packet4d pxor<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_xor_pd(a,b); } 209 210template<> EIGEN_STRONG_INLINE Packet8f pandnot<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_andnot_ps(a,b); } 211template<> EIGEN_STRONG_INLINE Packet4d pandnot<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_andnot_pd(a,b); } 212 213template<> EIGEN_STRONG_INLINE Packet8f pload<Packet8f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm256_load_ps(from); } 214template<> EIGEN_STRONG_INLINE Packet4d pload<Packet4d>(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm256_load_pd(from); } 215template<> EIGEN_STRONG_INLINE Packet8i pload<Packet8i>(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm256_load_si256(reinterpret_cast<const __m256i*>(from)); } 216 217template<> EIGEN_STRONG_INLINE Packet8f ploadu<Packet8f>(const float* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_ps(from); } 218template<> EIGEN_STRONG_INLINE Packet4d ploadu<Packet4d>(const double* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_pd(from); } 219template<> EIGEN_STRONG_INLINE Packet8i ploadu<Packet8i>(const int* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_si256(reinterpret_cast<const __m256i*>(from)); } 220 221// Loads 4 floats from memory a returns the packet {a0, a0 a1, a1, a2, a2, a3, a3} 222template<> EIGEN_STRONG_INLINE Packet8f ploaddup<Packet8f>(const float* from) 223{ 224 // TODO try to find a way to avoid the need of a temporary register 225// Packet8f tmp = _mm256_castps128_ps256(_mm_loadu_ps(from)); 226// tmp = _mm256_insertf128_ps(tmp, _mm_movehl_ps(_mm256_castps256_ps128(tmp),_mm256_castps256_ps128(tmp)), 1); 227// return _mm256_unpacklo_ps(tmp,tmp); 228 229 // _mm256_insertf128_ps is very slow on Haswell, thus: 230 Packet8f tmp = _mm256_broadcast_ps((const __m128*)(const void*)from); 231 // mimic an "inplace" permutation of the lower 128bits using a blend 232 tmp = _mm256_blend_ps(tmp,_mm256_castps128_ps256(_mm_permute_ps( _mm256_castps256_ps128(tmp), _MM_SHUFFLE(1,0,1,0))), 15); 233 // then we can perform a consistent permutation on the global register to get everything in shape: 234 return _mm256_permute_ps(tmp, _MM_SHUFFLE(3,3,2,2)); 235} 236// Loads 2 doubles from memory a returns the packet {a0, a0 a1, a1} 237template<> EIGEN_STRONG_INLINE Packet4d ploaddup<Packet4d>(const double* from) 238{ 239 Packet4d tmp = _mm256_broadcast_pd((const __m128d*)(const void*)from); 240 return _mm256_permute_pd(tmp, 3<<2); 241} 242 243// Loads 2 floats from memory a returns the packet {a0, a0 a0, a0, a1, a1, a1, a1} 244template<> EIGEN_STRONG_INLINE Packet8f ploadquad<Packet8f>(const float* from) 245{ 246 Packet8f tmp = _mm256_castps128_ps256(_mm_broadcast_ss(from)); 247 return _mm256_insertf128_ps(tmp, _mm_broadcast_ss(from+1), 1); 248} 249 250template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet8f& from) { EIGEN_DEBUG_ALIGNED_STORE _mm256_store_ps(to, from); } 251template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet4d& from) { EIGEN_DEBUG_ALIGNED_STORE _mm256_store_pd(to, from); } 252template<> EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet8i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm256_storeu_si256(reinterpret_cast<__m256i*>(to), from); } 253 254template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet8f& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_ps(to, from); } 255template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet4d& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_pd(to, from); } 256template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet8i& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_si256(reinterpret_cast<__m256i*>(to), from); } 257 258// NOTE: leverage _mm256_i32gather_ps and _mm256_i32gather_pd if AVX2 instructions are available 259// NOTE: for the record the following seems to be slower: return _mm256_i32gather_ps(from, _mm256_set1_epi32(stride), 4); 260template<> EIGEN_DEVICE_FUNC inline Packet8f pgather<float, Packet8f>(const float* from, Index stride) 261{ 262 return _mm256_set_ps(from[7*stride], from[6*stride], from[5*stride], from[4*stride], 263 from[3*stride], from[2*stride], from[1*stride], from[0*stride]); 264} 265template<> EIGEN_DEVICE_FUNC inline Packet4d pgather<double, Packet4d>(const double* from, Index stride) 266{ 267 return _mm256_set_pd(from[3*stride], from[2*stride], from[1*stride], from[0*stride]); 268} 269 270template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet8f>(float* to, const Packet8f& from, Index stride) 271{ 272 __m128 low = _mm256_extractf128_ps(from, 0); 273 to[stride*0] = _mm_cvtss_f32(low); 274 to[stride*1] = _mm_cvtss_f32(_mm_shuffle_ps(low, low, 1)); 275 to[stride*2] = _mm_cvtss_f32(_mm_shuffle_ps(low, low, 2)); 276 to[stride*3] = _mm_cvtss_f32(_mm_shuffle_ps(low, low, 3)); 277 278 __m128 high = _mm256_extractf128_ps(from, 1); 279 to[stride*4] = _mm_cvtss_f32(high); 280 to[stride*5] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 1)); 281 to[stride*6] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 2)); 282 to[stride*7] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 3)); 283} 284template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet4d>(double* to, const Packet4d& from, Index stride) 285{ 286 __m128d low = _mm256_extractf128_pd(from, 0); 287 to[stride*0] = _mm_cvtsd_f64(low); 288 to[stride*1] = _mm_cvtsd_f64(_mm_shuffle_pd(low, low, 1)); 289 __m128d high = _mm256_extractf128_pd(from, 1); 290 to[stride*2] = _mm_cvtsd_f64(high); 291 to[stride*3] = _mm_cvtsd_f64(_mm_shuffle_pd(high, high, 1)); 292} 293 294template<> EIGEN_STRONG_INLINE void pstore1<Packet8f>(float* to, const float& a) 295{ 296 Packet8f pa = pset1<Packet8f>(a); 297 pstore(to, pa); 298} 299template<> EIGEN_STRONG_INLINE void pstore1<Packet4d>(double* to, const double& a) 300{ 301 Packet4d pa = pset1<Packet4d>(a); 302 pstore(to, pa); 303} 304template<> EIGEN_STRONG_INLINE void pstore1<Packet8i>(int* to, const int& a) 305{ 306 Packet8i pa = pset1<Packet8i>(a); 307 pstore(to, pa); 308} 309 310#ifndef EIGEN_VECTORIZE_AVX512 311template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } 312template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } 313template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } 314#endif 315 316template<> EIGEN_STRONG_INLINE float pfirst<Packet8f>(const Packet8f& a) { 317 return _mm_cvtss_f32(_mm256_castps256_ps128(a)); 318} 319template<> EIGEN_STRONG_INLINE double pfirst<Packet4d>(const Packet4d& a) { 320 return _mm_cvtsd_f64(_mm256_castpd256_pd128(a)); 321} 322template<> EIGEN_STRONG_INLINE int pfirst<Packet8i>(const Packet8i& a) { 323 return _mm_cvtsi128_si32(_mm256_castsi256_si128(a)); 324} 325 326 327template<> EIGEN_STRONG_INLINE Packet8f preverse(const Packet8f& a) 328{ 329 __m256 tmp = _mm256_shuffle_ps(a,a,0x1b); 330 return _mm256_permute2f128_ps(tmp, tmp, 1); 331} 332template<> EIGEN_STRONG_INLINE Packet4d preverse(const Packet4d& a) 333{ 334 __m256d tmp = _mm256_shuffle_pd(a,a,5); 335 return _mm256_permute2f128_pd(tmp, tmp, 1); 336 337 __m256d swap_halves = _mm256_permute2f128_pd(a,a,1); 338 return _mm256_permute_pd(swap_halves,5); 339} 340 341// pabs should be ok 342template<> EIGEN_STRONG_INLINE Packet8f pabs(const Packet8f& a) 343{ 344 const Packet8f mask = _mm256_castsi256_ps(_mm256_setr_epi32(0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF)); 345 return _mm256_and_ps(a,mask); 346} 347template<> EIGEN_STRONG_INLINE Packet4d pabs(const Packet4d& a) 348{ 349 const Packet4d mask = _mm256_castsi256_pd(_mm256_setr_epi32(0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF)); 350 return _mm256_and_pd(a,mask); 351} 352 353// preduxp should be ok 354// FIXME: why is this ok? why isn't the simply implementation working as expected? 355template<> EIGEN_STRONG_INLINE Packet8f preduxp<Packet8f>(const Packet8f* vecs) 356{ 357 __m256 hsum1 = _mm256_hadd_ps(vecs[0], vecs[1]); 358 __m256 hsum2 = _mm256_hadd_ps(vecs[2], vecs[3]); 359 __m256 hsum3 = _mm256_hadd_ps(vecs[4], vecs[5]); 360 __m256 hsum4 = _mm256_hadd_ps(vecs[6], vecs[7]); 361 362 __m256 hsum5 = _mm256_hadd_ps(hsum1, hsum1); 363 __m256 hsum6 = _mm256_hadd_ps(hsum2, hsum2); 364 __m256 hsum7 = _mm256_hadd_ps(hsum3, hsum3); 365 __m256 hsum8 = _mm256_hadd_ps(hsum4, hsum4); 366 367 __m256 perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23); 368 __m256 perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23); 369 __m256 perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23); 370 __m256 perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23); 371 372 __m256 sum1 = _mm256_add_ps(perm1, hsum5); 373 __m256 sum2 = _mm256_add_ps(perm2, hsum6); 374 __m256 sum3 = _mm256_add_ps(perm3, hsum7); 375 __m256 sum4 = _mm256_add_ps(perm4, hsum8); 376 377 __m256 blend1 = _mm256_blend_ps(sum1, sum2, 0xcc); 378 __m256 blend2 = _mm256_blend_ps(sum3, sum4, 0xcc); 379 380 __m256 final = _mm256_blend_ps(blend1, blend2, 0xf0); 381 return final; 382} 383template<> EIGEN_STRONG_INLINE Packet4d preduxp<Packet4d>(const Packet4d* vecs) 384{ 385 Packet4d tmp0, tmp1; 386 387 tmp0 = _mm256_hadd_pd(vecs[0], vecs[1]); 388 tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1)); 389 390 tmp1 = _mm256_hadd_pd(vecs[2], vecs[3]); 391 tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1)); 392 393 return _mm256_blend_pd(tmp0, tmp1, 0xC); 394} 395 396template<> EIGEN_STRONG_INLINE float predux<Packet8f>(const Packet8f& a) 397{ 398 return predux(Packet4f(_mm_add_ps(_mm256_castps256_ps128(a),_mm256_extractf128_ps(a,1)))); 399} 400template<> EIGEN_STRONG_INLINE double predux<Packet4d>(const Packet4d& a) 401{ 402 return predux(Packet2d(_mm_add_pd(_mm256_castpd256_pd128(a),_mm256_extractf128_pd(a,1)))); 403} 404 405template<> EIGEN_STRONG_INLINE Packet4f predux_downto4<Packet8f>(const Packet8f& a) 406{ 407 return _mm_add_ps(_mm256_castps256_ps128(a),_mm256_extractf128_ps(a,1)); 408} 409 410template<> EIGEN_STRONG_INLINE float predux_mul<Packet8f>(const Packet8f& a) 411{ 412 Packet8f tmp; 413 tmp = _mm256_mul_ps(a, _mm256_permute2f128_ps(a,a,1)); 414 tmp = _mm256_mul_ps(tmp, _mm256_shuffle_ps(tmp,tmp,_MM_SHUFFLE(1,0,3,2))); 415 return pfirst(_mm256_mul_ps(tmp, _mm256_shuffle_ps(tmp,tmp,1))); 416} 417template<> EIGEN_STRONG_INLINE double predux_mul<Packet4d>(const Packet4d& a) 418{ 419 Packet4d tmp; 420 tmp = _mm256_mul_pd(a, _mm256_permute2f128_pd(a,a,1)); 421 return pfirst(_mm256_mul_pd(tmp, _mm256_shuffle_pd(tmp,tmp,1))); 422} 423 424template<> EIGEN_STRONG_INLINE float predux_min<Packet8f>(const Packet8f& a) 425{ 426 Packet8f tmp = _mm256_min_ps(a, _mm256_permute2f128_ps(a,a,1)); 427 tmp = _mm256_min_ps(tmp, _mm256_shuffle_ps(tmp,tmp,_MM_SHUFFLE(1,0,3,2))); 428 return pfirst(_mm256_min_ps(tmp, _mm256_shuffle_ps(tmp,tmp,1))); 429} 430template<> EIGEN_STRONG_INLINE double predux_min<Packet4d>(const Packet4d& a) 431{ 432 Packet4d tmp = _mm256_min_pd(a, _mm256_permute2f128_pd(a,a,1)); 433 return pfirst(_mm256_min_pd(tmp, _mm256_shuffle_pd(tmp, tmp, 1))); 434} 435 436template<> EIGEN_STRONG_INLINE float predux_max<Packet8f>(const Packet8f& a) 437{ 438 Packet8f tmp = _mm256_max_ps(a, _mm256_permute2f128_ps(a,a,1)); 439 tmp = _mm256_max_ps(tmp, _mm256_shuffle_ps(tmp,tmp,_MM_SHUFFLE(1,0,3,2))); 440 return pfirst(_mm256_max_ps(tmp, _mm256_shuffle_ps(tmp,tmp,1))); 441} 442 443template<> EIGEN_STRONG_INLINE double predux_max<Packet4d>(const Packet4d& a) 444{ 445 Packet4d tmp = _mm256_max_pd(a, _mm256_permute2f128_pd(a,a,1)); 446 return pfirst(_mm256_max_pd(tmp, _mm256_shuffle_pd(tmp, tmp, 1))); 447} 448 449 450template<int Offset> 451struct palign_impl<Offset,Packet8f> 452{ 453 static EIGEN_STRONG_INLINE void run(Packet8f& first, const Packet8f& second) 454 { 455 if (Offset==1) 456 { 457 first = _mm256_blend_ps(first, second, 1); 458 Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(0,3,2,1)); 459 Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1); 460 first = _mm256_blend_ps(tmp1, tmp2, 0x88); 461 } 462 else if (Offset==2) 463 { 464 first = _mm256_blend_ps(first, second, 3); 465 Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(1,0,3,2)); 466 Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1); 467 first = _mm256_blend_ps(tmp1, tmp2, 0xcc); 468 } 469 else if (Offset==3) 470 { 471 first = _mm256_blend_ps(first, second, 7); 472 Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(2,1,0,3)); 473 Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1); 474 first = _mm256_blend_ps(tmp1, tmp2, 0xee); 475 } 476 else if (Offset==4) 477 { 478 first = _mm256_blend_ps(first, second, 15); 479 Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(3,2,1,0)); 480 Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1); 481 first = _mm256_permute_ps(tmp2, _MM_SHUFFLE(3,2,1,0)); 482 } 483 else if (Offset==5) 484 { 485 first = _mm256_blend_ps(first, second, 31); 486 first = _mm256_permute2f128_ps(first, first, 1); 487 Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(0,3,2,1)); 488 first = _mm256_permute2f128_ps(tmp, tmp, 1); 489 first = _mm256_blend_ps(tmp, first, 0x88); 490 } 491 else if (Offset==6) 492 { 493 first = _mm256_blend_ps(first, second, 63); 494 first = _mm256_permute2f128_ps(first, first, 1); 495 Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(1,0,3,2)); 496 first = _mm256_permute2f128_ps(tmp, tmp, 1); 497 first = _mm256_blend_ps(tmp, first, 0xcc); 498 } 499 else if (Offset==7) 500 { 501 first = _mm256_blend_ps(first, second, 127); 502 first = _mm256_permute2f128_ps(first, first, 1); 503 Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(2,1,0,3)); 504 first = _mm256_permute2f128_ps(tmp, tmp, 1); 505 first = _mm256_blend_ps(tmp, first, 0xee); 506 } 507 } 508}; 509 510template<int Offset> 511struct palign_impl<Offset,Packet4d> 512{ 513 static EIGEN_STRONG_INLINE void run(Packet4d& first, const Packet4d& second) 514 { 515 if (Offset==1) 516 { 517 first = _mm256_blend_pd(first, second, 1); 518 __m256d tmp = _mm256_permute_pd(first, 5); 519 first = _mm256_permute2f128_pd(tmp, tmp, 1); 520 first = _mm256_blend_pd(tmp, first, 0xA); 521 } 522 else if (Offset==2) 523 { 524 first = _mm256_blend_pd(first, second, 3); 525 first = _mm256_permute2f128_pd(first, first, 1); 526 } 527 else if (Offset==3) 528 { 529 first = _mm256_blend_pd(first, second, 7); 530 __m256d tmp = _mm256_permute_pd(first, 5); 531 first = _mm256_permute2f128_pd(tmp, tmp, 1); 532 first = _mm256_blend_pd(tmp, first, 5); 533 } 534 } 535}; 536 537EIGEN_DEVICE_FUNC inline void 538ptranspose(PacketBlock<Packet8f,8>& kernel) { 539 __m256 T0 = _mm256_unpacklo_ps(kernel.packet[0], kernel.packet[1]); 540 __m256 T1 = _mm256_unpackhi_ps(kernel.packet[0], kernel.packet[1]); 541 __m256 T2 = _mm256_unpacklo_ps(kernel.packet[2], kernel.packet[3]); 542 __m256 T3 = _mm256_unpackhi_ps(kernel.packet[2], kernel.packet[3]); 543 __m256 T4 = _mm256_unpacklo_ps(kernel.packet[4], kernel.packet[5]); 544 __m256 T5 = _mm256_unpackhi_ps(kernel.packet[4], kernel.packet[5]); 545 __m256 T6 = _mm256_unpacklo_ps(kernel.packet[6], kernel.packet[7]); 546 __m256 T7 = _mm256_unpackhi_ps(kernel.packet[6], kernel.packet[7]); 547 __m256 S0 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(1,0,1,0)); 548 __m256 S1 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(3,2,3,2)); 549 __m256 S2 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(1,0,1,0)); 550 __m256 S3 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(3,2,3,2)); 551 __m256 S4 = _mm256_shuffle_ps(T4,T6,_MM_SHUFFLE(1,0,1,0)); 552 __m256 S5 = _mm256_shuffle_ps(T4,T6,_MM_SHUFFLE(3,2,3,2)); 553 __m256 S6 = _mm256_shuffle_ps(T5,T7,_MM_SHUFFLE(1,0,1,0)); 554 __m256 S7 = _mm256_shuffle_ps(T5,T7,_MM_SHUFFLE(3,2,3,2)); 555 kernel.packet[0] = _mm256_permute2f128_ps(S0, S4, 0x20); 556 kernel.packet[1] = _mm256_permute2f128_ps(S1, S5, 0x20); 557 kernel.packet[2] = _mm256_permute2f128_ps(S2, S6, 0x20); 558 kernel.packet[3] = _mm256_permute2f128_ps(S3, S7, 0x20); 559 kernel.packet[4] = _mm256_permute2f128_ps(S0, S4, 0x31); 560 kernel.packet[5] = _mm256_permute2f128_ps(S1, S5, 0x31); 561 kernel.packet[6] = _mm256_permute2f128_ps(S2, S6, 0x31); 562 kernel.packet[7] = _mm256_permute2f128_ps(S3, S7, 0x31); 563} 564 565EIGEN_DEVICE_FUNC inline void 566ptranspose(PacketBlock<Packet8f,4>& kernel) { 567 __m256 T0 = _mm256_unpacklo_ps(kernel.packet[0], kernel.packet[1]); 568 __m256 T1 = _mm256_unpackhi_ps(kernel.packet[0], kernel.packet[1]); 569 __m256 T2 = _mm256_unpacklo_ps(kernel.packet[2], kernel.packet[3]); 570 __m256 T3 = _mm256_unpackhi_ps(kernel.packet[2], kernel.packet[3]); 571 572 __m256 S0 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(1,0,1,0)); 573 __m256 S1 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(3,2,3,2)); 574 __m256 S2 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(1,0,1,0)); 575 __m256 S3 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(3,2,3,2)); 576 577 kernel.packet[0] = _mm256_permute2f128_ps(S0, S1, 0x20); 578 kernel.packet[1] = _mm256_permute2f128_ps(S2, S3, 0x20); 579 kernel.packet[2] = _mm256_permute2f128_ps(S0, S1, 0x31); 580 kernel.packet[3] = _mm256_permute2f128_ps(S2, S3, 0x31); 581} 582 583EIGEN_DEVICE_FUNC inline void 584ptranspose(PacketBlock<Packet4d,4>& kernel) { 585 __m256d T0 = _mm256_shuffle_pd(kernel.packet[0], kernel.packet[1], 15); 586 __m256d T1 = _mm256_shuffle_pd(kernel.packet[0], kernel.packet[1], 0); 587 __m256d T2 = _mm256_shuffle_pd(kernel.packet[2], kernel.packet[3], 15); 588 __m256d T3 = _mm256_shuffle_pd(kernel.packet[2], kernel.packet[3], 0); 589 590 kernel.packet[1] = _mm256_permute2f128_pd(T0, T2, 32); 591 kernel.packet[3] = _mm256_permute2f128_pd(T0, T2, 49); 592 kernel.packet[0] = _mm256_permute2f128_pd(T1, T3, 32); 593 kernel.packet[2] = _mm256_permute2f128_pd(T1, T3, 49); 594} 595 596template<> EIGEN_STRONG_INLINE Packet8f pblend(const Selector<8>& ifPacket, const Packet8f& thenPacket, const Packet8f& elsePacket) { 597 const __m256 zero = _mm256_setzero_ps(); 598 const __m256 select = _mm256_set_ps(ifPacket.select[7], ifPacket.select[6], ifPacket.select[5], ifPacket.select[4], ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]); 599 __m256 false_mask = _mm256_cmp_ps(select, zero, _CMP_EQ_UQ); 600 return _mm256_blendv_ps(thenPacket, elsePacket, false_mask); 601} 602template<> EIGEN_STRONG_INLINE Packet4d pblend(const Selector<4>& ifPacket, const Packet4d& thenPacket, const Packet4d& elsePacket) { 603 const __m256d zero = _mm256_setzero_pd(); 604 const __m256d select = _mm256_set_pd(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]); 605 __m256d false_mask = _mm256_cmp_pd(select, zero, _CMP_EQ_UQ); 606 return _mm256_blendv_pd(thenPacket, elsePacket, false_mask); 607} 608 609template<> EIGEN_STRONG_INLINE Packet8f pinsertfirst(const Packet8f& a, float b) 610{ 611 return _mm256_blend_ps(a,pset1<Packet8f>(b),1); 612} 613 614template<> EIGEN_STRONG_INLINE Packet4d pinsertfirst(const Packet4d& a, double b) 615{ 616 return _mm256_blend_pd(a,pset1<Packet4d>(b),1); 617} 618 619template<> EIGEN_STRONG_INLINE Packet8f pinsertlast(const Packet8f& a, float b) 620{ 621 return _mm256_blend_ps(a,pset1<Packet8f>(b),(1<<7)); 622} 623 624template<> EIGEN_STRONG_INLINE Packet4d pinsertlast(const Packet4d& a, double b) 625{ 626 return _mm256_blend_pd(a,pset1<Packet4d>(b),(1<<3)); 627} 628 629} // end namespace internal 630 631} // end namespace Eigen 632 633#endif // EIGEN_PACKET_MATH_AVX_H 634