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, 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 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 } // namespace
  89
  90
  91 template<int align, typename... Targs>
  92 static inline void gmx_simdcall
  93                    gatherLoadBySimdIntTranspose(const double* base, SimdDInt32 offset, SimdDouble* v, Targs... Fargs)
  94 {
  95     if (align > 1)
  96     {
  97         offset = fastMultiply<align>(offset);
  98     }
  99     constexpr size_t scale = sizeof(double);
 100     v->simdInternal_       = _mm512_i32gather_pd(offset.simdInternal_, base, scale);
 101     gatherLoadBySimdIntTranspose<1>(base + 1, offset, Fargs...);
 102 }
 103
 104 template<int align, typename... Targs>
 105 static inline void gmx_simdcall
 106                    gatherLoadUBySimdIntTranspose(const double* base, SimdDInt32 offset, SimdDouble* v, Targs... Fargs)
 107 {
 108     gatherLoadBySimdIntTranspose<align>(base, offset, v, Fargs...);
 109 }
 110
 111 template<int align, typename... Targs>
 112 static inline void gmx_simdcall
 113                    gatherLoadTranspose(const double* base, const std::int32_t offset[], SimdDouble* v, Targs... Fargs)
 114 {
 115     gatherLoadBySimdIntTranspose<align>(base, simdLoad(offset, SimdDInt32Tag()), v, Fargs...);
 116 }
 117
 118 template<int align, typename... Targs>
 119 static inline void gmx_simdcall
 120                    gatherLoadUTranspose(const double* base, const std::int32_t offset[], SimdDouble* v, Targs... Fargs)
 121 {
 122     gatherLoadBySimdIntTranspose<align>(base, simdLoad(offset, SimdDInt32Tag()), v, Fargs...);
 123 }
 124
 125 template<int align>
 126 static inline void gmx_simdcall transposeScatterStoreU(double*            base,
 127                                                        const std::int32_t offset[],
 128                                                        SimdDouble         v0,
 129                                                        SimdDouble         v1,
 130                                                        SimdDouble         v2)
 131 {
 132     SimdDInt32 simdoffset = simdLoad(offset, SimdDInt32Tag());
 133
 134     if (align > 1)
 135     {
 136         simdoffset = fastMultiply<align>(simdoffset);
 137         ;
 138     }
 139     constexpr size_t scale = sizeof(double);
 140     _mm512_i32scatter_pd(base, simdoffset.simdInternal_, v0.simdInternal_, scale);
 141     _mm512_i32scatter_pd(&(base[1]), simdoffset.simdInternal_, v1.simdInternal_, scale);
 142     _mm512_i32scatter_pd(&(base[2]), simdoffset.simdInternal_, v2.simdInternal_, scale);
 143 }
 144
 145 template<int align>
 146 static inline void gmx_simdcall
 147                    transposeScatterIncrU(double* base, const std::int32_t offset[], SimdDouble v0, SimdDouble v1, SimdDouble v2)
 148 {
 149     __m512d                                  t[4], t5, t6, t7, t8;
 150     alignas(GMX_SIMD_ALIGNMENT) std::int64_t o[8];
 151     // TODO: should use fastMultiply
 152     _mm512_store_epi64(o, _mm512_cvtepi32_epi64(_mm256_mullo_epi32(
 153                                   _mm256_load_si256((const __m256i*)(offset)), _mm256_set1_epi32(align))));
 154     t5   = _mm512_unpacklo_pd(v0.simdInternal_, v1.simdInternal_);
 155     t6   = _mm512_unpackhi_pd(v0.simdInternal_, v1.simdInternal_);
 156     t7   = _mm512_unpacklo_pd(v2.simdInternal_, _mm512_setzero_pd());
 157     t8   = _mm512_unpackhi_pd(v2.simdInternal_, _mm512_setzero_pd());
 158     t[0] = _mm512_mask_permutex_pd(t5, avx512Int2Mask(0xCC), t7, 0x4E);
 159     t[1] = _mm512_mask_permutex_pd(t6, avx512Int2Mask(0xCC), t8, 0x4E);
 160     t[2] = _mm512_mask_permutex_pd(t7, avx512Int2Mask(0x33), t5, 0x4E);
 161     t[3] = _mm512_mask_permutex_pd(t8, avx512Int2Mask(0x33), t6, 0x4E);
 162     if (align < 4)
 163     {
 164         for (int i = 0; i < 4; i++)
 165         {
 166             _mm512_mask_storeu_pd(base + o[0 + i], avx512Int2Mask(7),
 167                                   _mm512_castpd256_pd512(_mm256_add_pd(_mm256_loadu_pd(base + o[0 + i]),
 168                                                                        _mm512_castpd512_pd256(t[i]))));
 169             _mm512_mask_storeu_pd(
 170                     base + o[4 + i], avx512Int2Mask(7),
 171                     _mm512_castpd256_pd512(_mm256_add_pd(_mm256_loadu_pd(base + o[4 + i]),
 172                                                          _mm512_extractf64x4_pd(t[i], 1))));
 173         }
 174     }
 175     else
 176     {
 177         if (align % 4 == 0)
 178         {
 179             for (int i = 0; i < 4; i++)
 180             {
 181                 _mm256_store_pd(base + o[0 + i], _mm256_add_pd(_mm256_load_pd(base + o[0 + i]),
 182                                                                _mm512_castpd512_pd256(t[i])));
 183                 _mm256_store_pd(base + o[4 + i], _mm256_add_pd(_mm256_load_pd(base + o[4 + i]),
 184                                                                _mm512_extractf64x4_pd(t[i], 1)));
 185             }
 186         }
 187         else
 188         {
 189             for (int i = 0; i < 4; i++)
 190             {
 191                 _mm256_storeu_pd(base + o[0 + i], _mm256_add_pd(_mm256_loadu_pd(base + o[0 + i]),
 192                                                                 _mm512_castpd512_pd256(t[i])));
 193                 _mm256_storeu_pd(base + o[4 + i], _mm256_add_pd(_mm256_loadu_pd(base + o[4 + i]),
 194                                                                 _mm512_extractf64x4_pd(t[i], 1)));
 195             }
 196         }
 197     }
 198 }
 199
 200 template<int align>
 201 static inline void gmx_simdcall
 202                    transposeScatterDecrU(double* base, const std::int32_t offset[], SimdDouble v0, SimdDouble v1, SimdDouble v2)
 203 {
 204     __m512d                                  t[4], t5, t6, t7, t8;
 205     alignas(GMX_SIMD_ALIGNMENT) std::int64_t o[8];
 206     // TODO: should use fastMultiply
 207     _mm512_store_epi64(o, _mm512_cvtepi32_epi64(_mm256_mullo_epi32(
 208                                   _mm256_load_si256((const __m256i*)(offset)), _mm256_set1_epi32(align))));
 209     t5   = _mm512_unpacklo_pd(v0.simdInternal_, v1.simdInternal_);
 210     t6   = _mm512_unpackhi_pd(v0.simdInternal_, v1.simdInternal_);
 211     t7   = _mm512_unpacklo_pd(v2.simdInternal_, _mm512_setzero_pd());
 212     t8   = _mm512_unpackhi_pd(v2.simdInternal_, _mm512_setzero_pd());
 213     t[0] = _mm512_mask_permutex_pd(t5, avx512Int2Mask(0xCC), t7, 0x4E);
 214     t[2] = _mm512_mask_permutex_pd(t7, avx512Int2Mask(0x33), t5, 0x4E);
 215     t[1] = _mm512_mask_permutex_pd(t6, avx512Int2Mask(0xCC), t8, 0x4E);
 216     t[3] = _mm512_mask_permutex_pd(t8, avx512Int2Mask(0x33), t6, 0x4E);
 217     if (align < 4)
 218     {
 219         for (int i = 0; i < 4; i++)
 220         {
 221             _mm512_mask_storeu_pd(base + o[0 + i], avx512Int2Mask(7),
 222                                   _mm512_castpd256_pd512(_mm256_sub_pd(_mm256_loadu_pd(base + o[0 + i]),
 223                                                                        _mm512_castpd512_pd256(t[i]))));
 224             _mm512_mask_storeu_pd(
 225                     base + o[4 + i], avx512Int2Mask(7),
 226                     _mm512_castpd256_pd512(_mm256_sub_pd(_mm256_loadu_pd(base + o[4 + i]),
 227                                                          _mm512_extractf64x4_pd(t[i], 1))));
 228         }
 229     }
 230     else
 231     {
 232         if (align % 4 == 0)
 233         {
 234             for (int i = 0; i < 4; i++)
 235             {
 236                 _mm256_store_pd(base + o[0 + i], _mm256_sub_pd(_mm256_load_pd(base + o[0 + i]),
 237                                                                _mm512_castpd512_pd256(t[i])));
 238                 _mm256_store_pd(base + o[4 + i], _mm256_sub_pd(_mm256_load_pd(base + o[4 + i]),
 239                                                                _mm512_extractf64x4_pd(t[i], 1)));
 240             }
 241         }
 242         else
 243         {
 244             for (int i = 0; i < 4; i++)
 245             {
 246                 _mm256_storeu_pd(base + o[0 + i], _mm256_sub_pd(_mm256_loadu_pd(base + o[0 + i]),
 247                                                                 _mm512_castpd512_pd256(t[i])));
 248                 _mm256_storeu_pd(base + o[4 + i], _mm256_sub_pd(_mm256_loadu_pd(base + o[4 + i]),
 249                                                                 _mm512_extractf64x4_pd(t[i], 1)));
 250             }
 251         }
 252     }
 253 }
 254
 255 static inline void gmx_simdcall expandScalarsToTriplets(SimdDouble  scalar,
 256                                                         SimdDouble* triplets0,
 257                                                         SimdDouble* triplets1,
 258                                                         SimdDouble* triplets2)
 259 {
 260     triplets0->simdInternal_ = _mm512_castsi512_pd(_mm512_permutexvar_epi32(
 261             _mm512_set_epi32(5, 4, 5, 4, 3, 2, 3, 2, 3, 2, 1, 0, 1, 0, 1, 0),
 262             _mm512_castpd_si512(scalar.simdInternal_)));
 263     triplets1->simdInternal_ = _mm512_castsi512_pd(_mm512_permutexvar_epi32(
 264             _mm512_set_epi32(11, 10, 9, 8, 9, 8, 9, 8, 7, 6, 7, 6, 7, 6, 5, 4),
 265             _mm512_castpd_si512(scalar.simdInternal_)));
 266     triplets2->simdInternal_ = _mm512_castsi512_pd(_mm512_permutexvar_epi32(
 267             _mm512_set_epi32(15, 14, 15, 14, 15, 14, 13, 12, 13, 12, 13, 12, 11, 10, 11, 10),
 268             _mm512_castpd_si512(scalar.simdInternal_)));
 269 }
 270
 271
 272 static inline double gmx_simdcall
 273                      reduceIncr4ReturnSum(double* m, SimdDouble v0, SimdDouble v1, SimdDouble v2, SimdDouble v3)
 274 {
 275     __m512d t0, t2;
 276     __m256d t3, t4;
 277
 278     assert(std::size_t(m) % 32 == 0);
 279
 280     t0 = _mm512_add_pd(v0.simdInternal_, _mm512_permute_pd(v0.simdInternal_, 0x55));
 281     t2 = _mm512_add_pd(v2.simdInternal_, _mm512_permute_pd(v2.simdInternal_, 0x55));
 282     t0 = _mm512_mask_add_pd(t0, avx512Int2Mask(0xAA), v1.simdInternal_,
 283                             _mm512_permute_pd(v1.simdInternal_, 0x55));
 284     t2 = _mm512_mask_add_pd(t2, avx512Int2Mask(0xAA), v3.simdInternal_,
 285                             _mm512_permute_pd(v3.simdInternal_, 0x55));
 286     t0 = _mm512_add_pd(t0, _mm512_shuffle_f64x2(t0, t0, 0x4E));
 287     t0 = _mm512_mask_add_pd(t0, avx512Int2Mask(0xF0), t2, _mm512_shuffle_f64x2(t2, t2, 0x4E));
 288     t0 = _mm512_add_pd(t0, _mm512_shuffle_f64x2(t0, t0, 0xB1));
 289     t0 = _mm512_mask_shuffle_f64x2(t0, avx512Int2Mask(0x0C), t0, t0, 0xEE);
 290
 291     t3 = _mm512_castpd512_pd256(t0);
 292     t4 = _mm256_load_pd(m);
 293     t4 = _mm256_add_pd(t4, t3);
 294     _mm256_store_pd(m, t4);
 295
 296     t0 = _mm512_add_pd(t0, _mm512_permutex_pd(t0, 0x4E));
 297     t0 = _mm512_add_pd(t0, _mm512_permutex_pd(t0, 0xB1));
 298
 299     return _mm_cvtsd_f64(_mm512_castpd512_pd128(t0));
 300 }
 301
 302 static inline SimdDouble gmx_simdcall loadDualHsimd(const double* m0, const double* m1)
 303 {
 304     assert(std::size_t(m0) % 32 == 0);
 305     assert(std::size_t(m1) % 32 == 0);
 306
 307     return { _mm512_insertf64x4(_mm512_castpd256_pd512(_mm256_load_pd(m0)), _mm256_load_pd(m1), 1) };
 308 }
 309
 310 static inline SimdDouble gmx_simdcall loadDuplicateHsimd(const double* m)
 311 {
 312     assert(std::size_t(m) % 32 == 0);
 313
 314     return { _mm512_broadcast_f64x4(_mm256_load_pd(m)) };
 315 }
 316
 317 static inline SimdDouble gmx_simdcall loadU1DualHsimd(const double* m)
 318 {
 319     return { _mm512_insertf64x4(_mm512_broadcastsd_pd(_mm_load_sd(m)),
 320                                 _mm256_broadcastsd_pd(_mm_load_sd(m + 1)), 1) };
 321 }
 322
 323
 324 static inline void gmx_simdcall storeDualHsimd(double* m0, double* m1, SimdDouble a)
 325 {
 326     assert(std::size_t(m0) % 32 == 0);
 327     assert(std::size_t(m1) % 32 == 0);
 328
 329     _mm256_store_pd(m0, _mm512_castpd512_pd256(a.simdInternal_));
 330     _mm256_store_pd(m1, _mm512_extractf64x4_pd(a.simdInternal_, 1));
 331 }
 332
 333 static inline void gmx_simdcall incrDualHsimd(double* m0, double* m1, SimdDouble a)
 334 {
 335     assert(std::size_t(m0) % 32 == 0);
 336     assert(std::size_t(m1) % 32 == 0);
 337
 338     __m256d x;
 339
 340     // Lower half
 341     x = _mm256_load_pd(m0);
 342     x = _mm256_add_pd(x, _mm512_castpd512_pd256(a.simdInternal_));
 343     _mm256_store_pd(m0, x);
 344
 345     // Upper half
 346     x = _mm256_load_pd(m1);
 347     x = _mm256_add_pd(x, _mm512_extractf64x4_pd(a.simdInternal_, 1));
 348     _mm256_store_pd(m1, x);
 349 }
 350
 351 static inline void gmx_simdcall decrHsimd(double* m, SimdDouble a)
 352 {
 353     __m256d t;
 354
 355     assert(std::size_t(m) % 32 == 0);
 356
 357     a.simdInternal_ = _mm512_add_pd(a.simdInternal_,
 358                                     _mm512_shuffle_f64x2(a.simdInternal_, a.simdInternal_, 0xEE));
 359     t               = _mm256_load_pd(m);
 360     t               = _mm256_sub_pd(t, _mm512_castpd512_pd256(a.simdInternal_));
 361     _mm256_store_pd(m, t);
 362 }
 363
 364
 365 template<int align>
 366 static inline void gmx_simdcall gatherLoadTransposeHsimd(const double*      base0,
 367                                                          const double*      base1,
 368                                                          const std::int32_t offset[],
 369                                                          SimdDouble*        v0,
 370                                                          SimdDouble*        v1)
 371 {
 372     __m128i idx0, idx1;
 373     __m256i idx;
 374     __m512d tmp1, tmp2;
 375
 376     assert(std::size_t(offset) % 16 == 0);
 377     assert(std::size_t(base0) % 16 == 0);
 378     assert(std::size_t(base1) % 16 == 0);
 379
 380     idx0 = _mm_load_si128(reinterpret_cast<const __m128i*>(offset));
 381
 382     static_assert(align == 2 || align == 4, "If more are needed use fastMultiply");
 383     idx0 = _mm_slli_epi32(idx0, align == 2 ? 1 : 2);
 384
 385     idx1 = _mm_add_epi32(idx0, _mm_set1_epi32(1));
 386
 387     idx = _mm256_inserti128_si256(_mm256_castsi128_si256(idx0), idx1, 1);
 388
 389     constexpr size_t scale = sizeof(double);
 390     tmp1 = _mm512_i32gather_pd(idx, base0, scale); // TODO: Might be faster to use invidual loads
 391     tmp2 = _mm512_i32gather_pd(idx, base1, scale);
 392
 393     v0->simdInternal_ = _mm512_shuffle_f64x2(tmp1, tmp2, 0x44);
 394     v1->simdInternal_ = _mm512_shuffle_f64x2(tmp1, tmp2, 0xEE);
 395 }
 396
 397 static inline double gmx_simdcall reduceIncr4ReturnSumHsimd(double* m, SimdDouble v0, SimdDouble v1)
 398 {
 399     __m512d t0;
 400     __m256d t2, t3;
 401
 402     assert(std::size_t(m) % 32 == 0);
 403
 404     t0 = _mm512_add_pd(v0.simdInternal_, _mm512_permutex_pd(v0.simdInternal_, 0x4E));
 405     t0 = _mm512_mask_add_pd(t0, avx512Int2Mask(0xCC), v1.simdInternal_,
 406                             _mm512_permutex_pd(v1.simdInternal_, 0x4E));
 407     t0 = _mm512_add_pd(t0, _mm512_permutex_pd(t0, 0xB1));
 408     t0 = _mm512_mask_shuffle_f64x2(t0, avx512Int2Mask(0xAA), t0, t0, 0xEE);
 409
 410     t2 = _mm512_castpd512_pd256(t0);
 411     t3 = _mm256_load_pd(m);
 412     t3 = _mm256_add_pd(t3, t2);
 413     _mm256_store_pd(m, t3);
 414
 415     t0 = _mm512_add_pd(t0, _mm512_permutex_pd(t0, 0x4E));
 416     t0 = _mm512_add_pd(t0, _mm512_permutex_pd(t0, 0xB1));
 417
 418     return _mm_cvtsd_f64(_mm512_castpd512_pd128(t0));
 419 }
 420
 421 static inline SimdDouble gmx_simdcall loadU4NOffset(const double* m, int offset)
 422 {
 423     return { _mm512_insertf64x4(_mm512_castpd256_pd512(_mm256_loadu_pd(m)),
 424                                 _mm256_loadu_pd(m + offset), 1) };
 425 }
 426
 427 } // namespace gmx
 428
 429 #endif // GMX_SIMD_IMPL_X86_AVX_512_UTIL_DOUBLE_H