src/gromacs/simd/impl_x86_avx_512/impl_x86_avx_512_util_double.h

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2014-2018, The GROMACS development team.
   5  * Copyright (c) 2019,2020, by the GROMACS development team, led by
   6  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   7  * and including many others, as listed in the AUTHORS file in the
   8  * top-level source directory and at http://www.gromacs.org.
   9  *
  10  * GROMACS is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public License
  12  * as published by the Free Software Foundation; either version 2.1
  13  * of the License, or (at your option) any later version.
  14  *
  15  * GROMACS is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with GROMACS; if not, see
  22  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  23  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  24  *
  25  * If you want to redistribute modifications to GROMACS, please
  26  * consider that scientific software is very special. Version
  27  * control is crucial - bugs must be traceable. We will be happy to
  28  * consider code for inclusion in the official distribution, but
  29  * derived work must not be called official GROMACS. Details are found
  30  * in the README & COPYING files - if they are missing, get the
  31  * official version at http://www.gromacs.org.
  32  *
  33  * To help us fund GROMACS development, we humbly ask that you cite
  34  * the research papers on the package. Check out http://www.gromacs.org.
  35  */
  36
  37 #ifndef GMX_SIMD_IMPL_X86_AVX_512_UTIL_DOUBLE_H
  38 #define GMX_SIMD_IMPL_X86_AVX_512_UTIL_DOUBLE_H
  39
  40 #include "config.h"
  41
  42 #include <cassert>
  43 #include <cstdint>
  44
  45 #include <immintrin.h>
  46
  47 #include "gromacs/utility/basedefinitions.h"
  48
  49 #include "impl_x86_avx_512_general.h"
  50 #include "impl_x86_avx_512_simd_double.h"
  51
  52 namespace gmx
  53 {
  54
  55 static const int c_simdBestPairAlignmentDouble = 2;
  56
  57 namespace
  58 {
  59 // Multiply function optimized for powers of 2, for which it is done by
  60 // shifting. Currently up to 8 is accelerated. Could be accelerated for any
  61 // number with a constexpr log2 function.
  62 template<int n>
  63 static inline SimdDInt32 fastMultiply(SimdDInt32 x)
  64 {
  65     if (n == 2)
  66     {
  67         return _mm256_slli_epi32(x.simdInternal_, 1);
  68     }
  69     else if (n == 4)
  70     {
  71         return _mm256_slli_epi32(x.simdInternal_, 2);
  72     }
  73     else if (n == 8)
  74     {
  75         return _mm256_slli_epi32(x.simdInternal_, 3);
  76     }
  77     else
  78     {
  79         return x * n;
  80     }
  81 }
  82
  83 template<int align>
  84 static inline void gmx_simdcall gatherLoadBySimdIntTranspose(const double*, SimdDInt32)
  85 {
  86     // Nothing to do. Termination of recursion.
  87 }
  88
  89 /* This is an internal helper function used by decr3Hsimd(...).
  90  */
  91 inline void gmx_simdcall decrHsimd(double* m, SimdDouble a)
  92 {
  93     __m256d t;
  94
  95     assert(std::size_t(m) % 32 == 0);
  96
  97     a.simdInternal_ = _mm512_add_pd(a.simdInternal_,
  98                                     _mm512_shuffle_f64x2(a.simdInternal_, a.simdInternal_, 0xEE));
  99     t               = _mm256_load_pd(m);
 100     t               = _mm256_sub_pd(t, _mm512_castpd512_pd256(a.simdInternal_));
 101     _mm256_store_pd(m, t);
 102 }
 103 } // namespace
 104
 105
 106 template<int align, typename... Targs>
 107 static inline void gmx_simdcall
 108                    gatherLoadBySimdIntTranspose(const double* base, SimdDInt32 offset, SimdDouble* v, Targs... Fargs)
 109 {
 110     if (align > 1)
 111     {
 112         offset = fastMultiply<align>(offset);
 113     }
 114     constexpr size_t scale = sizeof(double);
 115     v->simdInternal_       = _mm512_i32gather_pd(offset.simdInternal_, base, scale);
 116     gatherLoadBySimdIntTranspose<1>(base + 1, offset, Fargs...);
 117 }
 118
 119 template<int align, typename... Targs>
 120 static inline void gmx_simdcall
 121                    gatherLoadUBySimdIntTranspose(const double* base, SimdDInt32 offset, SimdDouble* v, Targs... Fargs)
 122 {
 123     gatherLoadBySimdIntTranspose<align>(base, offset, v, Fargs...);
 124 }
 125
 126 template<int align, typename... Targs>
 127 static inline void gmx_simdcall
 128                    gatherLoadTranspose(const double* base, const std::int32_t offset[], SimdDouble* v, Targs... Fargs)
 129 {
 130     gatherLoadBySimdIntTranspose<align>(base, simdLoad(offset, SimdDInt32Tag()), v, Fargs...);
 131 }
 132
 133 template<int align, typename... Targs>
 134 static inline void gmx_simdcall
 135                    gatherLoadUTranspose(const double* base, const std::int32_t offset[], SimdDouble* v, Targs... Fargs)
 136 {
 137     gatherLoadBySimdIntTranspose<align>(base, simdLoad(offset, SimdDInt32Tag()), v, Fargs...);
 138 }
 139
 140 template<int align>
 141 static inline void gmx_simdcall transposeScatterStoreU(double*            base,
 142                                                        const std::int32_t offset[],
 143                                                        SimdDouble         v0,
 144                                                        SimdDouble         v1,
 145                                                        SimdDouble         v2)
 146 {
 147     SimdDInt32 simdoffset = simdLoad(offset, SimdDInt32Tag());
 148
 149     if (align > 1)
 150     {
 151         simdoffset = fastMultiply<align>(simdoffset);
 152         ;
 153     }
 154     constexpr size_t scale = sizeof(double);
 155     _mm512_i32scatter_pd(base, simdoffset.simdInternal_, v0.simdInternal_, scale);
 156     _mm512_i32scatter_pd(&(base[1]), simdoffset.simdInternal_, v1.simdInternal_, scale);
 157     _mm512_i32scatter_pd(&(base[2]), simdoffset.simdInternal_, v2.simdInternal_, scale);
 158 }
 159
 160 template<int align>
 161 static inline void gmx_simdcall
 162                    transposeScatterIncrU(double* base, const std::int32_t offset[], SimdDouble v0, SimdDouble v1, SimdDouble v2)
 163 {
 164     __m512d                                  t[4], t5, t6, t7, t8;
 165     alignas(GMX_SIMD_ALIGNMENT) std::int64_t o[8];
 166     // TODO: should use fastMultiply
 167     _mm512_store_epi64(o, _mm512_cvtepi32_epi64(_mm256_mullo_epi32(
 168                                   _mm256_load_si256((const __m256i*)(offset)), _mm256_set1_epi32(align))));
 169     t5   = _mm512_unpacklo_pd(v0.simdInternal_, v1.simdInternal_);
 170     t6   = _mm512_unpackhi_pd(v0.simdInternal_, v1.simdInternal_);
 171     t7   = _mm512_unpacklo_pd(v2.simdInternal_, _mm512_setzero_pd());
 172     t8   = _mm512_unpackhi_pd(v2.simdInternal_, _mm512_setzero_pd());
 173     t[0] = _mm512_mask_permutex_pd(t5, avx512Int2Mask(0xCC), t7, 0x4E);
 174     t[1] = _mm512_mask_permutex_pd(t6, avx512Int2Mask(0xCC), t8, 0x4E);
 175     t[2] = _mm512_mask_permutex_pd(t7, avx512Int2Mask(0x33), t5, 0x4E);
 176     t[3] = _mm512_mask_permutex_pd(t8, avx512Int2Mask(0x33), t6, 0x4E);
 177     if (align < 4)
 178     {
 179         for (int i = 0; i < 4; i++)
 180         {
 181             _mm512_mask_storeu_pd(base + o[0 + i], avx512Int2Mask(7),
 182                                   _mm512_castpd256_pd512(_mm256_add_pd(_mm256_loadu_pd(base + o[0 + i]),
 183                                                                        _mm512_castpd512_pd256(t[i]))));
 184             _mm512_mask_storeu_pd(
 185                     base + o[4 + i], avx512Int2Mask(7),
 186                     _mm512_castpd256_pd512(_mm256_add_pd(_mm256_loadu_pd(base + o[4 + i]),
 187                                                          _mm512_extractf64x4_pd(t[i], 1))));
 188         }
 189     }
 190     else
 191     {
 192         if (align % 4 == 0)
 193         {
 194             for (int i = 0; i < 4; i++)
 195             {
 196                 _mm256_store_pd(base + o[0 + i], _mm256_add_pd(_mm256_load_pd(base + o[0 + i]),
 197                                                                _mm512_castpd512_pd256(t[i])));
 198                 _mm256_store_pd(base + o[4 + i], _mm256_add_pd(_mm256_load_pd(base + o[4 + i]),
 199                                                                _mm512_extractf64x4_pd(t[i], 1)));
 200             }
 201         }
 202         else
 203         {
 204             for (int i = 0; i < 4; i++)
 205             {
 206                 _mm256_storeu_pd(base + o[0 + i], _mm256_add_pd(_mm256_loadu_pd(base + o[0 + i]),
 207                                                                 _mm512_castpd512_pd256(t[i])));
 208                 _mm256_storeu_pd(base + o[4 + i], _mm256_add_pd(_mm256_loadu_pd(base + o[4 + i]),
 209                                                                 _mm512_extractf64x4_pd(t[i], 1)));
 210             }
 211         }
 212     }
 213 }
 214
 215 template<int align>
 216 static inline void gmx_simdcall
 217                    transposeScatterDecrU(double* base, const std::int32_t offset[], SimdDouble v0, SimdDouble v1, SimdDouble v2)
 218 {
 219     __m512d                                  t[4], t5, t6, t7, t8;
 220     alignas(GMX_SIMD_ALIGNMENT) std::int64_t o[8];
 221     // TODO: should use fastMultiply
 222     _mm512_store_epi64(o, _mm512_cvtepi32_epi64(_mm256_mullo_epi32(
 223                                   _mm256_load_si256((const __m256i*)(offset)), _mm256_set1_epi32(align))));
 224     t5   = _mm512_unpacklo_pd(v0.simdInternal_, v1.simdInternal_);
 225     t6   = _mm512_unpackhi_pd(v0.simdInternal_, v1.simdInternal_);
 226     t7   = _mm512_unpacklo_pd(v2.simdInternal_, _mm512_setzero_pd());
 227     t8   = _mm512_unpackhi_pd(v2.simdInternal_, _mm512_setzero_pd());
 228     t[0] = _mm512_mask_permutex_pd(t5, avx512Int2Mask(0xCC), t7, 0x4E);
 229     t[2] = _mm512_mask_permutex_pd(t7, avx512Int2Mask(0x33), t5, 0x4E);
 230     t[1] = _mm512_mask_permutex_pd(t6, avx512Int2Mask(0xCC), t8, 0x4E);
 231     t[3] = _mm512_mask_permutex_pd(t8, avx512Int2Mask(0x33), t6, 0x4E);
 232     if (align < 4)
 233     {
 234         for (int i = 0; i < 4; i++)
 235         {
 236             _mm512_mask_storeu_pd(base + o[0 + i], avx512Int2Mask(7),
 237                                   _mm512_castpd256_pd512(_mm256_sub_pd(_mm256_loadu_pd(base + o[0 + i]),
 238                                                                        _mm512_castpd512_pd256(t[i]))));
 239             _mm512_mask_storeu_pd(
 240                     base + o[4 + i], avx512Int2Mask(7),
 241                     _mm512_castpd256_pd512(_mm256_sub_pd(_mm256_loadu_pd(base + o[4 + i]),
 242                                                          _mm512_extractf64x4_pd(t[i], 1))));
 243         }
 244     }
 245     else
 246     {
 247         if (align % 4 == 0)
 248         {
 249             for (int i = 0; i < 4; i++)
 250             {
 251                 _mm256_store_pd(base + o[0 + i], _mm256_sub_pd(_mm256_load_pd(base + o[0 + i]),
 252                                                                _mm512_castpd512_pd256(t[i])));
 253                 _mm256_store_pd(base + o[4 + i], _mm256_sub_pd(_mm256_load_pd(base + o[4 + i]),
 254                                                                _mm512_extractf64x4_pd(t[i], 1)));
 255             }
 256         }
 257         else
 258         {
 259             for (int i = 0; i < 4; i++)
 260             {
 261                 _mm256_storeu_pd(base + o[0 + i], _mm256_sub_pd(_mm256_loadu_pd(base + o[0 + i]),
 262                                                                 _mm512_castpd512_pd256(t[i])));
 263                 _mm256_storeu_pd(base + o[4 + i], _mm256_sub_pd(_mm256_loadu_pd(base + o[4 + i]),
 264                                                                 _mm512_extractf64x4_pd(t[i], 1)));
 265             }
 266         }
 267     }
 268 }
 269
 270 static inline void gmx_simdcall expandScalarsToTriplets(SimdDouble  scalar,
 271                                                         SimdDouble* triplets0,
 272                                                         SimdDouble* triplets1,
 273                                                         SimdDouble* triplets2)
 274 {
 275     triplets0->simdInternal_ = _mm512_castsi512_pd(_mm512_permutexvar_epi32(
 276             _mm512_set_epi32(5, 4, 5, 4, 3, 2, 3, 2, 3, 2, 1, 0, 1, 0, 1, 0),
 277             _mm512_castpd_si512(scalar.simdInternal_)));
 278     triplets1->simdInternal_ = _mm512_castsi512_pd(_mm512_permutexvar_epi32(
 279             _mm512_set_epi32(11, 10, 9, 8, 9, 8, 9, 8, 7, 6, 7, 6, 7, 6, 5, 4),
 280             _mm512_castpd_si512(scalar.simdInternal_)));
 281     triplets2->simdInternal_ = _mm512_castsi512_pd(_mm512_permutexvar_epi32(
 282             _mm512_set_epi32(15, 14, 15, 14, 15, 14, 13, 12, 13, 12, 13, 12, 11, 10, 11, 10),
 283             _mm512_castpd_si512(scalar.simdInternal_)));
 284 }
 285
 286
 287 static inline double gmx_simdcall
 288                      reduceIncr4ReturnSum(double* m, SimdDouble v0, SimdDouble v1, SimdDouble v2, SimdDouble v3)
 289 {
 290     __m512d t0, t2;
 291     __m256d t3, t4;
 292
 293     assert(std::size_t(m) % 32 == 0);
 294
 295     t0 = _mm512_add_pd(v0.simdInternal_, _mm512_permute_pd(v0.simdInternal_, 0x55));
 296     t2 = _mm512_add_pd(v2.simdInternal_, _mm512_permute_pd(v2.simdInternal_, 0x55));
 297     t0 = _mm512_mask_add_pd(t0, avx512Int2Mask(0xAA), v1.simdInternal_,
 298                             _mm512_permute_pd(v1.simdInternal_, 0x55));
 299     t2 = _mm512_mask_add_pd(t2, avx512Int2Mask(0xAA), v3.simdInternal_,
 300                             _mm512_permute_pd(v3.simdInternal_, 0x55));
 301     t0 = _mm512_add_pd(t0, _mm512_shuffle_f64x2(t0, t0, 0x4E));
 302     t0 = _mm512_mask_add_pd(t0, avx512Int2Mask(0xF0), t2, _mm512_shuffle_f64x2(t2, t2, 0x4E));
 303     t0 = _mm512_add_pd(t0, _mm512_shuffle_f64x2(t0, t0, 0xB1));
 304     t0 = _mm512_mask_shuffle_f64x2(t0, avx512Int2Mask(0x0C), t0, t0, 0xEE);
 305
 306     t3 = _mm512_castpd512_pd256(t0);
 307     t4 = _mm256_load_pd(m);
 308     t4 = _mm256_add_pd(t4, t3);
 309     _mm256_store_pd(m, t4);
 310
 311     t0 = _mm512_add_pd(t0, _mm512_permutex_pd(t0, 0x4E));
 312     t0 = _mm512_add_pd(t0, _mm512_permutex_pd(t0, 0xB1));
 313
 314     return _mm_cvtsd_f64(_mm512_castpd512_pd128(t0));
 315 }
 316
 317 static inline SimdDouble gmx_simdcall loadDualHsimd(const double* m0, const double* m1)
 318 {
 319     assert(std::size_t(m0) % 32 == 0);
 320     assert(std::size_t(m1) % 32 == 0);
 321
 322     return { _mm512_insertf64x4(_mm512_castpd256_pd512(_mm256_load_pd(m0)), _mm256_load_pd(m1), 1) };
 323 }
 324
 325 static inline SimdDouble gmx_simdcall loadDuplicateHsimd(const double* m)
 326 {
 327     assert(std::size_t(m) % 32 == 0);
 328
 329     return { _mm512_broadcast_f64x4(_mm256_load_pd(m)) };
 330 }
 331
 332 static inline SimdDouble gmx_simdcall loadU1DualHsimd(const double* m)
 333 {
 334     return { _mm512_insertf64x4(_mm512_broadcastsd_pd(_mm_load_sd(m)),
 335                                 _mm256_broadcastsd_pd(_mm_load_sd(m + 1)), 1) };
 336 }
 337
 338
 339 static inline void gmx_simdcall storeDualHsimd(double* m0, double* m1, SimdDouble a)
 340 {
 341     assert(std::size_t(m0) % 32 == 0);
 342     assert(std::size_t(m1) % 32 == 0);
 343
 344     _mm256_store_pd(m0, _mm512_castpd512_pd256(a.simdInternal_));
 345     _mm256_store_pd(m1, _mm512_extractf64x4_pd(a.simdInternal_, 1));
 346 }
 347
 348 static inline void gmx_simdcall incrDualHsimd(double* m0, double* m1, SimdDouble a)
 349 {
 350     assert(std::size_t(m0) % 32 == 0);
 351     assert(std::size_t(m1) % 32 == 0);
 352
 353     __m256d x;
 354
 355     // Lower half
 356     x = _mm256_load_pd(m0);
 357     x = _mm256_add_pd(x, _mm512_castpd512_pd256(a.simdInternal_));
 358     _mm256_store_pd(m0, x);
 359
 360     // Upper half
 361     x = _mm256_load_pd(m1);
 362     x = _mm256_add_pd(x, _mm512_extractf64x4_pd(a.simdInternal_, 1));
 363     _mm256_store_pd(m1, x);
 364 }
 365
 366 static inline void gmx_simdcall decr3Hsimd(double* m, SimdDouble a0, SimdDouble a1, SimdDouble a2)
 367 {
 368     decrHsimd(m, a0);
 369     decrHsimd(m + GMX_SIMD_DOUBLE_WIDTH / 2, a1);
 370     decrHsimd(m + GMX_SIMD_DOUBLE_WIDTH, a2);
 371 }
 372
 373 template<int align>
 374 static inline void gmx_simdcall gatherLoadTransposeHsimd(const double*      base0,
 375                                                          const double*      base1,
 376                                                          const std::int32_t offset[],
 377                                                          SimdDouble*        v0,
 378                                                          SimdDouble*        v1)
 379 {
 380     __m128i idx0, idx1;
 381     __m256i idx;
 382     __m512d tmp1, tmp2;
 383
 384     assert(std::size_t(offset) % 16 == 0);
 385     assert(std::size_t(base0) % 16 == 0);
 386     assert(std::size_t(base1) % 16 == 0);
 387
 388     idx0 = _mm_load_si128(reinterpret_cast<const __m128i*>(offset));
 389
 390     static_assert(align == 2 || align == 4, "If more are needed use fastMultiply");
 391     idx0 = _mm_slli_epi32(idx0, align == 2 ? 1 : 2);
 392
 393     idx1 = _mm_add_epi32(idx0, _mm_set1_epi32(1));
 394
 395     idx = _mm256_inserti128_si256(_mm256_castsi128_si256(idx0), idx1, 1);
 396
 397     constexpr size_t scale = sizeof(double);
 398     tmp1 = _mm512_i32gather_pd(idx, base0, scale); // TODO: Might be faster to use invidual loads
 399     tmp2 = _mm512_i32gather_pd(idx, base1, scale);
 400
 401     v0->simdInternal_ = _mm512_shuffle_f64x2(tmp1, tmp2, 0x44);
 402     v1->simdInternal_ = _mm512_shuffle_f64x2(tmp1, tmp2, 0xEE);
 403 }
 404
 405 static inline double gmx_simdcall reduceIncr4ReturnSumHsimd(double* m, SimdDouble v0, SimdDouble v1)
 406 {
 407     __m512d t0;
 408     __m256d t2, t3;
 409
 410     assert(std::size_t(m) % 32 == 0);
 411
 412     t0 = _mm512_add_pd(v0.simdInternal_, _mm512_permutex_pd(v0.simdInternal_, 0x4E));
 413     t0 = _mm512_mask_add_pd(t0, avx512Int2Mask(0xCC), v1.simdInternal_,
 414                             _mm512_permutex_pd(v1.simdInternal_, 0x4E));
 415     t0 = _mm512_add_pd(t0, _mm512_permutex_pd(t0, 0xB1));
 416     t0 = _mm512_mask_shuffle_f64x2(t0, avx512Int2Mask(0xAA), t0, t0, 0xEE);
 417
 418     t2 = _mm512_castpd512_pd256(t0);
 419     t3 = _mm256_load_pd(m);
 420     t3 = _mm256_add_pd(t3, t2);
 421     _mm256_store_pd(m, t3);
 422
 423     t0 = _mm512_add_pd(t0, _mm512_permutex_pd(t0, 0x4E));
 424     t0 = _mm512_add_pd(t0, _mm512_permutex_pd(t0, 0xB1));
 425
 426     return _mm_cvtsd_f64(_mm512_castpd512_pd128(t0));
 427 }
 428
 429 static inline SimdDouble gmx_simdcall loadU4NOffset(const double* m, int offset)
 430 {
 431     return { _mm512_insertf64x4(_mm512_castpd256_pd512(_mm256_loadu_pd(m)),
 432                                 _mm256_loadu_pd(m + offset), 1) };
 433 }
 434
 435 } // namespace gmx
 436
 437 #endif // GMX_SIMD_IMPL_X86_AVX_512_UTIL_DOUBLE_H