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 13 namespace Eigen { 14 15 namespace 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 31 typedef __m256 Packet8f; 32 typedef __m256i Packet8i; 33 typedef __m256d Packet4d; 34 35 template<> struct is_arithmetic<__m256> { enum { value = true }; }; 36 template<> struct is_arithmetic<__m256i> { enum { value = true }; }; 37 template<> 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 54 template<> 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 }; 78 template<> 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 99 100 template<> struct scalar_div_cost<float,true> { enum { value = 14 }; }; 101 template<> 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. 105 template<> 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 */ 115 116 template<> struct unpacket_traits<Packet8f> { typedef float type; typedef Packet4f half; enum {size=8, alignment=Aligned32}; }; 117 template<> struct unpacket_traits<Packet4d> { typedef double type; typedef Packet2d half; enum {size=4, alignment=Aligned32}; }; 118 template<> struct unpacket_traits<Packet8i> { typedef int type; typedef Packet4i half; enum {size=8, alignment=Aligned32}; }; 119 120 template<> EIGEN_STRONG_INLINE Packet8f pset1<Packet8f>(const float& from) { return _mm256_set1_ps(from); } 121 template<> EIGEN_STRONG_INLINE Packet4d pset1<Packet4d>(const double& from) { return _mm256_set1_pd(from); } 122 template<> EIGEN_STRONG_INLINE Packet8i pset1<Packet8i>(const int& from) { return _mm256_set1_epi32(from); } 123 124 template<> EIGEN_STRONG_INLINE Packet8f pload1<Packet8f>(const float* from) { return _mm256_broadcast_ss(from); } 125 template<> EIGEN_STRONG_INLINE Packet4d pload1<Packet4d>(const double* from) { return _mm256_broadcast_sd(from); } 126 127 template<> 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)); } 128 template<> 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)); } 129 130 template<> EIGEN_STRONG_INLINE Packet8f padd<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_add_ps(a,b); } 131 template<> EIGEN_STRONG_INLINE Packet4d padd<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_add_pd(a,b); } 132 133 template<> EIGEN_STRONG_INLINE Packet8f psub<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_sub_ps(a,b); } 134 template<> EIGEN_STRONG_INLINE Packet4d psub<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_sub_pd(a,b); } 135 136 template<> EIGEN_STRONG_INLINE Packet8f pnegate(const Packet8f& a) 137 { 138 return _mm256_sub_ps(_mm256_set1_ps(0.0),a); 139 } 140 template<> EIGEN_STRONG_INLINE Packet4d pnegate(const Packet4d& a) 141 { 142 return _mm256_sub_pd(_mm256_set1_pd(0.0),a); 143 } 144 145 template<> EIGEN_STRONG_INLINE Packet8f pconj(const Packet8f& a) { return a; } 146 template<> EIGEN_STRONG_INLINE Packet4d pconj(const Packet4d& a) { return a; } 147 template<> EIGEN_STRONG_INLINE Packet8i pconj(const Packet8i& a) { return a; } 148 149 template<> EIGEN_STRONG_INLINE Packet8f pmul<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_mul_ps(a,b); } 150 template<> EIGEN_STRONG_INLINE Packet4d pmul<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_mul_pd(a,b); } 151 152 153 template<> EIGEN_STRONG_INLINE Packet8f pdiv<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_div_ps(a,b); } 154 template<> EIGEN_STRONG_INLINE Packet4d pdiv<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_div_pd(a,b); } 155 template<> 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__ 161 template<> 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 } 174 template<> 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 185 186 template<> EIGEN_STRONG_INLINE Packet8f pmin<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_min_ps(a,b); } 187 template<> EIGEN_STRONG_INLINE Packet4d pmin<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_min_pd(a,b); } 188 189 template<> EIGEN_STRONG_INLINE Packet8f pmax<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_max_ps(a,b); } 190 template<> EIGEN_STRONG_INLINE Packet4d pmax<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_max_pd(a,b); } 191 192 template<> EIGEN_STRONG_INLINE Packet8f pround<Packet8f>(const Packet8f& a) { return _mm256_round_ps(a, _MM_FROUND_CUR_DIRECTION); } 193 template<> EIGEN_STRONG_INLINE Packet4d pround<Packet4d>(const Packet4d& a) { return _mm256_round_pd(a, _MM_FROUND_CUR_DIRECTION); } 194 195 template<> EIGEN_STRONG_INLINE Packet8f pceil<Packet8f>(const Packet8f& a) { return _mm256_ceil_ps(a); } 196 template<> EIGEN_STRONG_INLINE Packet4d pceil<Packet4d>(const Packet4d& a) { return _mm256_ceil_pd(a); } 197 198 template<> EIGEN_STRONG_INLINE Packet8f pfloor<Packet8f>(const Packet8f& a) { return _mm256_floor_ps(a); } 199 template<> EIGEN_STRONG_INLINE Packet4d pfloor<Packet4d>(const Packet4d& a) { return _mm256_floor_pd(a); } 200 201 template<> EIGEN_STRONG_INLINE Packet8f pand<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_and_ps(a,b); } 202 template<> EIGEN_STRONG_INLINE Packet4d pand<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_and_pd(a,b); } 203 204 template<> EIGEN_STRONG_INLINE Packet8f por<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_or_ps(a,b); } 205 template<> EIGEN_STRONG_INLINE Packet4d por<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_or_pd(a,b); } 206 207 template<> EIGEN_STRONG_INLINE Packet8f pxor<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_xor_ps(a,b); } 208 template<> EIGEN_STRONG_INLINE Packet4d pxor<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_xor_pd(a,b); } 209 210 template<> EIGEN_STRONG_INLINE Packet8f pandnot<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_andnot_ps(a,b); } 211 template<> EIGEN_STRONG_INLINE Packet4d pandnot<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_andnot_pd(a,b); } 212 213 template<> EIGEN_STRONG_INLINE Packet8f pload<Packet8f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm256_load_ps(from); } 214 template<> EIGEN_STRONG_INLINE Packet4d pload<Packet4d>(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm256_load_pd(from); } 215 template<> EIGEN_STRONG_INLINE Packet8i pload<Packet8i>(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm256_load_si256(reinterpret_cast<const __m256i*>(from)); } 216 217 template<> EIGEN_STRONG_INLINE Packet8f ploadu<Packet8f>(const float* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_ps(from); } 218 template<> EIGEN_STRONG_INLINE Packet4d ploadu<Packet4d>(const double* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_pd(from); } 219 template<> 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} 222 template<> 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} 237 template<> 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} 244 template<> 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 } 249 250 template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet8f& from) { EIGEN_DEBUG_ALIGNED_STORE _mm256_store_ps(to, from); } 251 template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet4d& from) { EIGEN_DEBUG_ALIGNED_STORE _mm256_store_pd(to, from); } 252 template<> EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet8i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm256_storeu_si256(reinterpret_cast<__m256i*>(to), from); } 253 254 template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet8f& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_ps(to, from); } 255 template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet4d& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_pd(to, from); } 256 template<> 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); 260 template<> 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 } 265 template<> 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 } 269 270 template<> 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 } 284 template<> 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 } 293 294 template<> EIGEN_STRONG_INLINE void pstore1<Packet8f>(float* to, const float& a) 295 { 296 Packet8f pa = pset1<Packet8f>(a); 297 pstore(to, pa); 298 } 299 template<> EIGEN_STRONG_INLINE void pstore1<Packet4d>(double* to, const double& a) 300 { 301 Packet4d pa = pset1<Packet4d>(a); 302 pstore(to, pa); 303 } 304 template<> 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 311 template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } 312 template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } 313 template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } 314 #endif 315 316 template<> EIGEN_STRONG_INLINE float pfirst<Packet8f>(const Packet8f& a) { 317 return _mm_cvtss_f32(_mm256_castps256_ps128(a)); 318 } 319 template<> EIGEN_STRONG_INLINE double pfirst<Packet4d>(const Packet4d& a) { 320 return _mm_cvtsd_f64(_mm256_castpd256_pd128(a)); 321 } 322 template<> EIGEN_STRONG_INLINE int pfirst<Packet8i>(const Packet8i& a) { 323 return _mm_cvtsi128_si32(_mm256_castsi256_si128(a)); 324 } 325 326 327 template<> 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 } 332 template<> 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 // G+Smo 338 //__m256d swap_halves = _mm256_permute2f128_pd(a,a,1); 339 // return _mm256_permute_pd(swap_halves,5); 340 } 341 342 // pabs should be ok 343 template<> EIGEN_STRONG_INLINE Packet8f pabs(const Packet8f& a) 344 { 345 const Packet8f mask = _mm256_castsi256_ps(_mm256_setr_epi32(0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF)); 346 return _mm256_and_ps(a,mask); 347 } 348 template<> EIGEN_STRONG_INLINE Packet4d pabs(const Packet4d& a) 349 { 350 const Packet4d mask = _mm256_castsi256_pd(_mm256_setr_epi32(0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF)); 351 return _mm256_and_pd(a,mask); 352 } 353 354 // preduxp should be ok 355 // FIXME: why is this ok? why isn't the simply implementation working as expected? 356 template<> EIGEN_STRONG_INLINE Packet8f preduxp<Packet8f>(const Packet8f* vecs) 357 { 358 __m256 hsum1 = _mm256_hadd_ps(vecs[0], vecs[1]); 359 __m256 hsum2 = _mm256_hadd_ps(vecs[2], vecs[3]); 360 __m256 hsum3 = _mm256_hadd_ps(vecs[4], vecs[5]); 361 __m256 hsum4 = _mm256_hadd_ps(vecs[6], vecs[7]); 362 363 __m256 hsum5 = _mm256_hadd_ps(hsum1, hsum1); 364 __m256 hsum6 = _mm256_hadd_ps(hsum2, hsum2); 365 __m256 hsum7 = _mm256_hadd_ps(hsum3, hsum3); 366 __m256 hsum8 = _mm256_hadd_ps(hsum4, hsum4); 367 368 __m256 perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23); 369 __m256 perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23); 370 __m256 perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23); 371 __m256 perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23); 372 373 __m256 sum1 = _mm256_add_ps(perm1, hsum5); 374 __m256 sum2 = _mm256_add_ps(perm2, hsum6); 375 __m256 sum3 = _mm256_add_ps(perm3, hsum7); 376 __m256 sum4 = _mm256_add_ps(perm4, hsum8); 377 378 __m256 blend1 = _mm256_blend_ps(sum1, sum2, 0xcc); 379 __m256 blend2 = _mm256_blend_ps(sum3, sum4, 0xcc); 380 381 __m256 final = _mm256_blend_ps(blend1, blend2, 0xf0); 382 return final; 383 } 384 template<> EIGEN_STRONG_INLINE Packet4d preduxp<Packet4d>(const Packet4d* vecs) 385 { 386 Packet4d tmp0, tmp1; 387 388 tmp0 = _mm256_hadd_pd(vecs[0], vecs[1]); 389 tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1)); 390 391 tmp1 = _mm256_hadd_pd(vecs[2], vecs[3]); 392 tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1)); 393 394 return _mm256_blend_pd(tmp0, tmp1, 0xC); 395 } 396 397 template<> EIGEN_STRONG_INLINE float predux<Packet8f>(const Packet8f& a) 398 { 399 return predux(Packet4f(_mm_add_ps(_mm256_castps256_ps128(a),_mm256_extractf128_ps(a,1)))); 400 } 401 template<> EIGEN_STRONG_INLINE double predux<Packet4d>(const Packet4d& a) 402 { 403 return predux(Packet2d(_mm_add_pd(_mm256_castpd256_pd128(a),_mm256_extractf128_pd(a,1)))); 404 } 405 406 template<> EIGEN_STRONG_INLINE Packet4f predux_downto4<Packet8f>(const Packet8f& a) 407 { 408 return _mm_add_ps(_mm256_castps256_ps128(a),_mm256_extractf128_ps(a,1)); 409 } 410 411 template<> EIGEN_STRONG_INLINE float predux_mul<Packet8f>(const Packet8f& a) 412 { 413 Packet8f tmp; 414 tmp = _mm256_mul_ps(a, _mm256_permute2f128_ps(a,a,1)); 415 tmp = _mm256_mul_ps(tmp, _mm256_shuffle_ps(tmp,tmp,_MM_SHUFFLE(1,0,3,2))); 416 return pfirst(_mm256_mul_ps(tmp, _mm256_shuffle_ps(tmp,tmp,1))); 417 } 418 template<> EIGEN_STRONG_INLINE double predux_mul<Packet4d>(const Packet4d& a) 419 { 420 Packet4d tmp; 421 tmp = _mm256_mul_pd(a, _mm256_permute2f128_pd(a,a,1)); 422 return pfirst(_mm256_mul_pd(tmp, _mm256_shuffle_pd(tmp,tmp,1))); 423 } 424 425 template<> EIGEN_STRONG_INLINE float predux_min<Packet8f>(const Packet8f& a) 426 { 427 Packet8f tmp = _mm256_min_ps(a, _mm256_permute2f128_ps(a,a,1)); 428 tmp = _mm256_min_ps(tmp, _mm256_shuffle_ps(tmp,tmp,_MM_SHUFFLE(1,0,3,2))); 429 return pfirst(_mm256_min_ps(tmp, _mm256_shuffle_ps(tmp,tmp,1))); 430 } 431 template<> EIGEN_STRONG_INLINE double predux_min<Packet4d>(const Packet4d& a) 432 { 433 Packet4d tmp = _mm256_min_pd(a, _mm256_permute2f128_pd(a,a,1)); 434 return pfirst(_mm256_min_pd(tmp, _mm256_shuffle_pd(tmp, tmp, 1))); 435 } 436 437 template<> EIGEN_STRONG_INLINE float predux_max<Packet8f>(const Packet8f& a) 438 { 439 Packet8f tmp = _mm256_max_ps(a, _mm256_permute2f128_ps(a,a,1)); 440 tmp = _mm256_max_ps(tmp, _mm256_shuffle_ps(tmp,tmp,_MM_SHUFFLE(1,0,3,2))); 441 return pfirst(_mm256_max_ps(tmp, _mm256_shuffle_ps(tmp,tmp,1))); 442 } 443 444 template<> EIGEN_STRONG_INLINE double predux_max<Packet4d>(const Packet4d& a) 445 { 446 Packet4d tmp = _mm256_max_pd(a, _mm256_permute2f128_pd(a,a,1)); 447 return pfirst(_mm256_max_pd(tmp, _mm256_shuffle_pd(tmp, tmp, 1))); 448 } 449 450 451 template<int Offset> 452 struct palign_impl<Offset,Packet8f> 453 { 454 static EIGEN_STRONG_INLINE void run(Packet8f& first, const Packet8f& second) 455 { 456 if (Offset==1) 457 { 458 first = _mm256_blend_ps(first, second, 1); 459 Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(0,3,2,1)); 460 Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1); 461 first = _mm256_blend_ps(tmp1, tmp2, 0x88); 462 } 463 else if (Offset==2) 464 { 465 first = _mm256_blend_ps(first, second, 3); 466 Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(1,0,3,2)); 467 Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1); 468 first = _mm256_blend_ps(tmp1, tmp2, 0xcc); 469 } 470 else if (Offset==3) 471 { 472 first = _mm256_blend_ps(first, second, 7); 473 Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(2,1,0,3)); 474 Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1); 475 first = _mm256_blend_ps(tmp1, tmp2, 0xee); 476 } 477 else if (Offset==4) 478 { 479 first = _mm256_blend_ps(first, second, 15); 480 Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(3,2,1,0)); 481 Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1); 482 first = _mm256_permute_ps(tmp2, _MM_SHUFFLE(3,2,1,0)); 483 } 484 else if (Offset==5) 485 { 486 first = _mm256_blend_ps(first, second, 31); 487 first = _mm256_permute2f128_ps(first, first, 1); 488 Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(0,3,2,1)); 489 first = _mm256_permute2f128_ps(tmp, tmp, 1); 490 first = _mm256_blend_ps(tmp, first, 0x88); 491 } 492 else if (Offset==6) 493 { 494 first = _mm256_blend_ps(first, second, 63); 495 first = _mm256_permute2f128_ps(first, first, 1); 496 Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(1,0,3,2)); 497 first = _mm256_permute2f128_ps(tmp, tmp, 1); 498 first = _mm256_blend_ps(tmp, first, 0xcc); 499 } 500 else if (Offset==7) 501 { 502 first = _mm256_blend_ps(first, second, 127); 503 first = _mm256_permute2f128_ps(first, first, 1); 504 Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(2,1,0,3)); 505 first = _mm256_permute2f128_ps(tmp, tmp, 1); 506 first = _mm256_blend_ps(tmp, first, 0xee); 507 } 508 } 509 }; 510 511 template<int Offset> 512 struct palign_impl<Offset,Packet4d> 513 { 514 static EIGEN_STRONG_INLINE void run(Packet4d& first, const Packet4d& second) 515 { 516 if (Offset==1) 517 { 518 first = _mm256_blend_pd(first, second, 1); 519 __m256d tmp = _mm256_permute_pd(first, 5); 520 first = _mm256_permute2f128_pd(tmp, tmp, 1); 521 first = _mm256_blend_pd(tmp, first, 0xA); 522 } 523 else if (Offset==2) 524 { 525 first = _mm256_blend_pd(first, second, 3); 526 first = _mm256_permute2f128_pd(first, first, 1); 527 } 528 else if (Offset==3) 529 { 530 first = _mm256_blend_pd(first, second, 7); 531 __m256d tmp = _mm256_permute_pd(first, 5); 532 first = _mm256_permute2f128_pd(tmp, tmp, 1); 533 first = _mm256_blend_pd(tmp, first, 5); 534 } 535 } 536 }; 537 538 EIGEN_DEVICE_FUNC inline void 539 ptranspose(PacketBlock<Packet8f,8>& kernel) { 540 __m256 T0 = _mm256_unpacklo_ps(kernel.packet[0], kernel.packet[1]); 541 __m256 T1 = _mm256_unpackhi_ps(kernel.packet[0], kernel.packet[1]); 542 __m256 T2 = _mm256_unpacklo_ps(kernel.packet[2], kernel.packet[3]); 543 __m256 T3 = _mm256_unpackhi_ps(kernel.packet[2], kernel.packet[3]); 544 __m256 T4 = _mm256_unpacklo_ps(kernel.packet[4], kernel.packet[5]); 545 __m256 T5 = _mm256_unpackhi_ps(kernel.packet[4], kernel.packet[5]); 546 __m256 T6 = _mm256_unpacklo_ps(kernel.packet[6], kernel.packet[7]); 547 __m256 T7 = _mm256_unpackhi_ps(kernel.packet[6], kernel.packet[7]); 548 __m256 S0 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(1,0,1,0)); 549 __m256 S1 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(3,2,3,2)); 550 __m256 S2 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(1,0,1,0)); 551 __m256 S3 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(3,2,3,2)); 552 __m256 S4 = _mm256_shuffle_ps(T4,T6,_MM_SHUFFLE(1,0,1,0)); 553 __m256 S5 = _mm256_shuffle_ps(T4,T6,_MM_SHUFFLE(3,2,3,2)); 554 __m256 S6 = _mm256_shuffle_ps(T5,T7,_MM_SHUFFLE(1,0,1,0)); 555 __m256 S7 = _mm256_shuffle_ps(T5,T7,_MM_SHUFFLE(3,2,3,2)); 556 kernel.packet[0] = _mm256_permute2f128_ps(S0, S4, 0x20); 557 kernel.packet[1] = _mm256_permute2f128_ps(S1, S5, 0x20); 558 kernel.packet[2] = _mm256_permute2f128_ps(S2, S6, 0x20); 559 kernel.packet[3] = _mm256_permute2f128_ps(S3, S7, 0x20); 560 kernel.packet[4] = _mm256_permute2f128_ps(S0, S4, 0x31); 561 kernel.packet[5] = _mm256_permute2f128_ps(S1, S5, 0x31); 562 kernel.packet[6] = _mm256_permute2f128_ps(S2, S6, 0x31); 563 kernel.packet[7] = _mm256_permute2f128_ps(S3, S7, 0x31); 564 } 565 566 EIGEN_DEVICE_FUNC inline void 567 ptranspose(PacketBlock<Packet8f,4>& kernel) { 568 __m256 T0 = _mm256_unpacklo_ps(kernel.packet[0], kernel.packet[1]); 569 __m256 T1 = _mm256_unpackhi_ps(kernel.packet[0], kernel.packet[1]); 570 __m256 T2 = _mm256_unpacklo_ps(kernel.packet[2], kernel.packet[3]); 571 __m256 T3 = _mm256_unpackhi_ps(kernel.packet[2], kernel.packet[3]); 572 573 __m256 S0 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(1,0,1,0)); 574 __m256 S1 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(3,2,3,2)); 575 __m256 S2 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(1,0,1,0)); 576 __m256 S3 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(3,2,3,2)); 577 578 kernel.packet[0] = _mm256_permute2f128_ps(S0, S1, 0x20); 579 kernel.packet[1] = _mm256_permute2f128_ps(S2, S3, 0x20); 580 kernel.packet[2] = _mm256_permute2f128_ps(S0, S1, 0x31); 581 kernel.packet[3] = _mm256_permute2f128_ps(S2, S3, 0x31); 582 } 583 584 EIGEN_DEVICE_FUNC inline void 585 ptranspose(PacketBlock<Packet4d,4>& kernel) { 586 __m256d T0 = _mm256_shuffle_pd(kernel.packet[0], kernel.packet[1], 15); 587 __m256d T1 = _mm256_shuffle_pd(kernel.packet[0], kernel.packet[1], 0); 588 __m256d T2 = _mm256_shuffle_pd(kernel.packet[2], kernel.packet[3], 15); 589 __m256d T3 = _mm256_shuffle_pd(kernel.packet[2], kernel.packet[3], 0); 590 591 kernel.packet[1] = _mm256_permute2f128_pd(T0, T2, 32); 592 kernel.packet[3] = _mm256_permute2f128_pd(T0, T2, 49); 593 kernel.packet[0] = _mm256_permute2f128_pd(T1, T3, 32); 594 kernel.packet[2] = _mm256_permute2f128_pd(T1, T3, 49); 595 } 596 597 template<> EIGEN_STRONG_INLINE Packet8f pblend(const Selector<8>& ifPacket, const Packet8f& thenPacket, const Packet8f& elsePacket) { 598 const __m256 zero = _mm256_setzero_ps(); 599 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]); 600 __m256 false_mask = _mm256_cmp_ps(select, zero, _CMP_EQ_UQ); 601 return _mm256_blendv_ps(thenPacket, elsePacket, false_mask); 602 } 603 template<> EIGEN_STRONG_INLINE Packet4d pblend(const Selector<4>& ifPacket, const Packet4d& thenPacket, const Packet4d& elsePacket) { 604 const __m256d zero = _mm256_setzero_pd(); 605 const __m256d select = _mm256_set_pd(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]); 606 __m256d false_mask = _mm256_cmp_pd(select, zero, _CMP_EQ_UQ); 607 return _mm256_blendv_pd(thenPacket, elsePacket, false_mask); 608 } 609 610 template<> EIGEN_STRONG_INLINE Packet8f pinsertfirst(const Packet8f& a, float b) 611 { 612 return _mm256_blend_ps(a,pset1<Packet8f>(b),1); 613 } 614 615 template<> EIGEN_STRONG_INLINE Packet4d pinsertfirst(const Packet4d& a, double b) 616 { 617 return _mm256_blend_pd(a,pset1<Packet4d>(b),1); 618 } 619 620 template<> EIGEN_STRONG_INLINE Packet8f pinsertlast(const Packet8f& a, float b) 621 { 622 return _mm256_blend_ps(a,pset1<Packet8f>(b),(1<<7)); 623 } 624 625 template<> EIGEN_STRONG_INLINE Packet4d pinsertlast(const Packet4d& a, double b) 626 { 627 return _mm256_blend_pd(a,pset1<Packet4d>(b),(1<<3)); 628 } 629 630 } // end namespace internal 631 632 } // end namespace Eigen 633 634 #endif // EIGEN_PACKET_MATH_AVX_H 635