src/gromacs/simd/impl_x86_mic/impl_x86_mic_util_double.h

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2014,2015,2016,2017,2018 by 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_MIC_UTIL_DOUBLE_H
  38 #define GMX_SIMD_IMPL_X86_MIC_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_mic_simd_double.h"
  50
  51 namespace gmx
  52 {
  53
  54 namespace
  55 {
  56 /* This is an internal helper function used by decr3Hsimd(...).
  57  */
  58 inline void gmx_simdcall decrHsimd(double* m, SimdDouble a)
  59 {
  60     __m512d t;
  61
  62     assert(std::size_t(m) % 32 == 0);
  63
  64     t               = _mm512_extload_pd(m, _MM_UPCONV_PD_NONE, _MM_BROADCAST_4X8, _MM_HINT_NONE);
  65     a.simdInternal_ = _mm512_add_pd(
  66             a.simdInternal_,
  67             _mm512_castps_pd(_mm512_permute4f128_ps(_mm512_castpd_ps(a.simdInternal_), _MM_PERM_BADC)));
  68     t = _mm512_sub_pd(t, a.simdInternal_);
  69     _mm512_mask_packstorelo_pd(m, _mm512_int2mask(0x0F), t);
  70 }
  71 } // namespace
  72
  73 // On MIC it is better to use scatter operations, so we define the load routines
  74 // that use a SIMD offset variable first.
  75
  76 template<int align>
  77 static inline void gmx_simdcall gatherLoadBySimdIntTranspose(const double* base,
  78                                                              SimdDInt32    simdoffset,
  79                                                              SimdDouble*   v0,
  80                                                              SimdDouble*   v1,
  81                                                              SimdDouble*   v2,
  82                                                              SimdDouble*   v3)
  83 {
  84     assert((size_t)base % 32 == 0);
  85     assert(align % 4 == 0);
  86
  87     // All instructions might be latency ~4 on MIC, so we use shifts where we
  88     // only need a single instruction (since the shift parameter is an immediate),
  89     // but multiplication otherwise.
  90     if (align == 4)
  91     {
  92         simdoffset.simdInternal_ = _mm512_slli_epi32(simdoffset.simdInternal_, 2);
  93     }
  94     else if (align == 8)
  95     {
  96         simdoffset.simdInternal_ = _mm512_slli_epi32(simdoffset.simdInternal_, 3);
  97     }
  98     else
  99     {
 100         simdoffset = simdoffset * SimdDInt32(align);
 101     }
 102
 103     v0->simdInternal_ = _mm512_i32logather_pd(simdoffset.simdInternal_, base, sizeof(double));
 104     v1->simdInternal_ = _mm512_i32logather_pd(simdoffset.simdInternal_, base + 1, sizeof(double));
 105     v2->simdInternal_ = _mm512_i32logather_pd(simdoffset.simdInternal_, base + 2, sizeof(double));
 106     v3->simdInternal_ = _mm512_i32logather_pd(simdoffset.simdInternal_, base + 3, sizeof(double));
 107 }
 108
 109 template<int align>
 110 static inline void gmx_simdcall gatherLoadUBySimdIntTranspose(const double* base,
 111                                                               SimdDInt32    simdoffset,
 112                                                               SimdDouble*   v0,
 113                                                               SimdDouble*   v1)
 114 {
 115     // All instructions might be latency ~4 on MIC, so we use shifts where we
 116     // only need a single instruction (since the shift parameter is an immediate),
 117     // but multiplication otherwise.
 118     if (align == 2)
 119     {
 120         simdoffset.simdInternal_ = _mm512_slli_epi32(simdoffset.simdInternal_, 1);
 121     }
 122     else if (align == 4)
 123     {
 124         simdoffset.simdInternal_ = _mm512_slli_epi32(simdoffset.simdInternal_, 2);
 125     }
 126     else if (align == 8)
 127     {
 128         simdoffset.simdInternal_ = _mm512_slli_epi32(simdoffset.simdInternal_, 3);
 129     }
 130     else
 131     {
 132         simdoffset = simdoffset * SimdDInt32(align);
 133     }
 134
 135     v0->simdInternal_ = _mm512_i32logather_pd(simdoffset.simdInternal_, base, sizeof(double));
 136     v1->simdInternal_ = _mm512_i32logather_pd(simdoffset.simdInternal_, base + 1, sizeof(double));
 137 }
 138
 139 template<int align>
 140 static inline void gmx_simdcall gatherLoadBySimdIntTranspose(const double* base,
 141                                                              SimdDInt32    simdoffset,
 142                                                              SimdDouble*   v0,
 143                                                              SimdDouble*   v1)
 144 {
 145     assert(std::size_t(base) % 16 == 0);
 146     assert(align % 2 == 0);
 147     gatherLoadUBySimdIntTranspose<align>(base, simdoffset, v0, v1);
 148 }
 149
 150
 151 template<int align>
 152 static inline void gmx_simdcall gatherLoadTranspose(const double*      base,
 153                                                     const std::int32_t offset[],
 154                                                     SimdDouble*        v0,
 155                                                     SimdDouble*        v1,
 156                                                     SimdDouble*        v2,
 157                                                     SimdDouble*        v3)
 158 {
 159     gatherLoadBySimdIntTranspose<align>(base, simdLoad(offset, SimdDInt32Tag()), v0, v1, v2, v3);
 160 }
 161
 162 template<int align>
 163 static inline void gmx_simdcall
 164                    gatherLoadTranspose(const double* base, const std::int32_t offset[], SimdDouble* v0, SimdDouble* v1)
 165 {
 166     gatherLoadBySimdIntTranspose<align>(base, simdLoad(offset, SimdDInt32Tag()), v0, v1);
 167 }
 168
 169 static const int c_simdBestPairAlignmentDouble = 2;
 170
 171 template<int align>
 172 static inline void gmx_simdcall gatherLoadUTranspose(const double*      base,
 173                                                      const std::int32_t offset[],
 174                                                      SimdDouble*        v0,
 175                                                      SimdDouble*        v1,
 176                                                      SimdDouble*        v2)
 177 {
 178     SimdDInt32 simdoffset;
 179
 180     assert(std::size_t(offset) % 32 == 0);
 181
 182     simdoffset = simdLoad(offset, SimdDInt32Tag());
 183
 184     // All instructions might be latency ~4 on MIC, so we use shifts where we
 185     // only need a single instruction (since the shift parameter is an immediate),
 186     // but multiplication otherwise.
 187     if (align == 4)
 188     {
 189         simdoffset.simdInternal_ = _mm512_slli_epi32(simdoffset.simdInternal_, 2);
 190     }
 191     else if (align == 8)
 192     {
 193         simdoffset.simdInternal_ = _mm512_slli_epi32(simdoffset.simdInternal_, 3);
 194     }
 195     else
 196     {
 197         simdoffset = simdoffset * SimdDInt32(align);
 198     }
 199
 200     v0->simdInternal_ = _mm512_i32logather_pd(simdoffset.simdInternal_, base, sizeof(double));
 201     v1->simdInternal_ = _mm512_i32logather_pd(simdoffset.simdInternal_, base + 1, sizeof(double));
 202     v2->simdInternal_ = _mm512_i32logather_pd(simdoffset.simdInternal_, base + 2, sizeof(double));
 203 }
 204
 205 template<int align>
 206 static inline void gmx_simdcall transposeScatterStoreU(double*            base,
 207                                                        const std::int32_t offset[],
 208                                                        SimdDouble         v0,
 209                                                        SimdDouble         v1,
 210                                                        SimdDouble         v2)
 211 {
 212     SimdDInt32 simdoffset;
 213
 214     assert(std::size_t(offset) % 32 == 0);
 215
 216     simdoffset = simdLoad(offset, SimdDInt32Tag());
 217
 218     // All instructions might be latency ~4 on MIC, so we use shifts where we
 219     // only need a single instruction (since the shift parameter is an immediate),
 220     // but multiplication otherwise.
 221     if (align == 4)
 222     {
 223         simdoffset.simdInternal_ = _mm512_slli_epi32(simdoffset.simdInternal_, 2);
 224     }
 225     else if (align == 8)
 226     {
 227         simdoffset.simdInternal_ = _mm512_slli_epi32(simdoffset.simdInternal_, 3);
 228     }
 229     else
 230     {
 231         simdoffset = simdoffset * SimdDInt32(align);
 232     }
 233
 234     _mm512_i32loscatter_pd(base, simdoffset.simdInternal_, v0.simdInternal_, sizeof(double));
 235     _mm512_i32loscatter_pd(base + 1, simdoffset.simdInternal_, v1.simdInternal_, sizeof(double));
 236     _mm512_i32loscatter_pd(base + 2, simdoffset.simdInternal_, v2.simdInternal_, sizeof(double));
 237 }
 238
 239 template<int align>
 240 static inline void gmx_simdcall
 241                    transposeScatterIncrU(double* base, const std::int32_t offset[], SimdDouble v0, SimdDouble v1, SimdDouble v2)
 242 {
 243     alignas(GMX_SIMD_ALIGNMENT) double rdata0[GMX_SIMD_DOUBLE_WIDTH];
 244     alignas(GMX_SIMD_ALIGNMENT) double rdata1[GMX_SIMD_DOUBLE_WIDTH];
 245     alignas(GMX_SIMD_ALIGNMENT) double rdata2[GMX_SIMD_DOUBLE_WIDTH];
 246
 247     store(rdata0, v0);
 248     store(rdata1, v1);
 249     store(rdata2, v2);
 250
 251     for (int i = 0; i < GMX_SIMD_DOUBLE_WIDTH; i++)
 252     {
 253         base[align * offset[i] + 0] += rdata0[i];
 254         base[align * offset[i] + 1] += rdata1[i];
 255         base[align * offset[i] + 2] += rdata2[i];
 256     }
 257 }
 258
 259 template<int align>
 260 static inline void gmx_simdcall
 261                    transposeScatterDecrU(double* base, const std::int32_t offset[], SimdDouble v0, SimdDouble v1, SimdDouble v2)
 262 {
 263     alignas(GMX_SIMD_ALIGNMENT) double rdata0[GMX_SIMD_DOUBLE_WIDTH];
 264     alignas(GMX_SIMD_ALIGNMENT) double rdata1[GMX_SIMD_DOUBLE_WIDTH];
 265     alignas(GMX_SIMD_ALIGNMENT) double rdata2[GMX_SIMD_DOUBLE_WIDTH];
 266
 267     store(rdata0, v0);
 268     store(rdata1, v1);
 269     store(rdata2, v2);
 270
 271     for (int i = 0; i < GMX_SIMD_DOUBLE_WIDTH; i++)
 272     {
 273         base[align * offset[i] + 0] -= rdata0[i];
 274         base[align * offset[i] + 1] -= rdata1[i];
 275         base[align * offset[i] + 2] -= rdata2[i];
 276     }
 277 }
 278
 279 static inline void gmx_simdcall expandScalarsToTriplets(SimdDouble  scalar,
 280                                                         SimdDouble* triplets0,
 281                                                         SimdDouble* triplets1,
 282                                                         SimdDouble* triplets2)
 283 {
 284     triplets0->simdInternal_ = _mm512_castsi512_pd(
 285             _mm512_permutevar_epi32(_mm512_set_epi32(5, 4, 5, 4, 3, 2, 3, 2, 3, 2, 1, 0, 1, 0, 1, 0),
 286                                     _mm512_castpd_si512(scalar.simdInternal_)));
 287     triplets1->simdInternal_ = _mm512_castsi512_pd(_mm512_permutevar_epi32(
 288             _mm512_set_epi32(11, 10, 9, 8, 9, 8, 9, 8, 7, 6, 7, 6, 7, 6, 5, 4),
 289             _mm512_castpd_si512(scalar.simdInternal_)));
 290     triplets2->simdInternal_ = _mm512_castsi512_pd(_mm512_permutevar_epi32(
 291             _mm512_set_epi32(15, 14, 15, 14, 15, 14, 13, 12, 13, 12, 13, 12, 11, 10, 11, 10),
 292             _mm512_castpd_si512(scalar.simdInternal_)));
 293 }
 294
 295
 296 static inline double gmx_simdcall
 297                      reduceIncr4ReturnSum(double* m, SimdDouble v0, SimdDouble v1, SimdDouble v2, SimdDouble v3)
 298 {
 299     double  d;
 300     __m512d t0, t1, t2, t3;
 301
 302     assert(std::size_t(m) % 32 == 0);
 303
 304     t0 = _mm512_swizzle_pd(_mm512_mask_blend_pd(_mm512_int2mask(0x33), v0.simdInternal_, v2.simdInternal_),
 305                            _MM_SWIZ_REG_BADC);
 306     t2 = _mm512_mask_blend_pd(_mm512_int2mask(0x33), v2.simdInternal_, v0.simdInternal_);
 307     t1 = _mm512_swizzle_pd(_mm512_mask_blend_pd(_mm512_int2mask(0x33), v1.simdInternal_, v3.simdInternal_),
 308                            _MM_SWIZ_REG_BADC);
 309     t3 = _mm512_mask_blend_pd(_mm512_int2mask(0x33), v3.simdInternal_, v1.simdInternal_);
 310     t0 = _mm512_add_pd(t0, t2);
 311     t1 = _mm512_add_pd(t1, t3);
 312
 313     t2 = _mm512_swizzle_pd(_mm512_mask_blend_pd(_mm512_int2mask(0b01010101), t0, t1), _MM_SWIZ_REG_CDAB);
 314     t3 = _mm512_mask_blend_pd(_mm512_int2mask(0b01010101), t1, t0);
 315     t2 = _mm512_add_pd(t2, t3);
 316
 317     t2 = _mm512_add_pd(t2, _mm512_castps_pd(_mm512_permute4f128_ps(_mm512_castpd_ps(t2), _MM_PERM_BADC)));
 318
 319     t0 = _mm512_mask_extload_pd(
 320             _mm512_undefined_pd(), _mm512_int2mask(0xF), m, _MM_UPCONV_PD_NONE, _MM_BROADCAST_4X8, _MM_HINT_NONE);
 321     t0 = _mm512_add_pd(t0, t2);
 322     _mm512_mask_packstorelo_pd(m, _mm512_int2mask(0xF), t0);
 323
 324     t2 = _mm512_add_pd(t2, _mm512_swizzle_pd(t2, _MM_SWIZ_REG_BADC));
 325     t2 = _mm512_add_pd(t2, _mm512_swizzle_pd(t2, _MM_SWIZ_REG_CDAB));
 326
 327     _mm512_mask_packstorelo_pd(&d, _mm512_mask2int(0x01), t2);
 328     return d;
 329 }
 330
 331 static inline SimdDouble gmx_simdcall loadDualHsimd(const double* m0, const double* m1)
 332 {
 333     assert(std::size_t(m0) % 32 == 0);
 334     assert(std::size_t(m1) % 32 == 0);
 335
 336     return _mm512_mask_extload_pd(_mm512_extload_pd(m0, _MM_UPCONV_PD_NONE, _MM_BROADCAST_4X8, _MM_HINT_NONE),
 337                                   _mm512_int2mask(0xF0),
 338                                   m1,
 339                                   _MM_UPCONV_PD_NONE,
 340                                   _MM_BROADCAST_4X8,
 341                                   _MM_HINT_NONE);
 342 }
 343
 344 static inline SimdDouble gmx_simdcall loadDuplicateHsimd(const double* m)
 345 {
 346     assert(std::size_t(m) % 32 == 0);
 347
 348     return _mm512_extload_pd(m, _MM_UPCONV_PD_NONE, _MM_BROADCAST_4X8, _MM_HINT_NONE);
 349 }
 350
 351 static inline SimdDouble gmx_simdcall loadU1DualHsimd(const double* m)
 352 {
 353     return _mm512_mask_extload_pd(_mm512_extload_pd(m, _MM_UPCONV_PD_NONE, _MM_BROADCAST_1X8, _MM_HINT_NONE),
 354                                   _mm512_int2mask(0xF0),
 355                                   m + 1,
 356                                   _MM_UPCONV_PD_NONE,
 357                                   _MM_BROADCAST_1X8,
 358                                   _MM_HINT_NONE);
 359 }
 360
 361
 362 static inline void gmx_simdcall storeDualHsimd(double* m0, double* m1, SimdDouble a)
 363 {
 364     assert(std::size_t(m0) % 32 == 0);
 365     assert(std::size_t(m1) % 32 == 0);
 366
 367     _mm512_mask_packstorelo_pd(m0, _mm512_int2mask(0x0F), a.simdInternal_);
 368     _mm512_mask_packstorelo_pd(m1, _mm512_int2mask(0xF0), a.simdInternal_);
 369 }
 370
 371 static inline void gmx_simdcall incrDualHsimd(double* m0, double* m1, SimdDouble a)
 372 {
 373     assert(std::size_t(m0) % 32 == 0);
 374     assert(std::size_t(m1) % 32 == 0);
 375
 376     __m512d x;
 377
 378     // Update lower half
 379     x = _mm512_extload_pd(m0, _MM_UPCONV_PD_NONE, _MM_BROADCAST_4X8, _MM_HINT_NONE);
 380     x = _mm512_add_pd(x, a.simdInternal_);
 381     _mm512_mask_packstorelo_pd(m0, _mm512_int2mask(0x0F), x);
 382
 383     // Update upper half
 384     x = _mm512_extload_pd(m1, _MM_UPCONV_PD_NONE, _MM_BROADCAST_4X8, _MM_HINT_NONE);
 385     x = _mm512_add_pd(x, a.simdInternal_);
 386     _mm512_mask_packstorelo_pd(m1, _mm512_int2mask(0xF0), x);
 387 }
 388
 389 static inline void gmx_simdcall decr3Hsimd(double* m, SimdDouble a0, SimdDouble a1, SimdDouble a2)
 390 {
 391     assert(std::size_t(m) % 32 == 0);
 392     decrHsimd(m, a0);
 393     decrHsimd(m + GMX_SIMD_DOUBLE_WIDTH / 2, a1);
 394     decrHsimd(m + GMX_SIMD_DOUBLE_WIDTH, a2);
 395 }
 396
 397
 398 template<int align>
 399 static inline void gmx_simdcall gatherLoadTransposeHsimd(const double*      base0,
 400                                                          const double*      base1,
 401                                                          const std::int32_t offset[],
 402                                                          SimdDouble*        v0,
 403                                                          SimdDouble*        v1)
 404 {
 405     __m512i idx0, idx1, idx;
 406     __m512d tmp1, tmp2;
 407
 408     assert(std::size_t(offset) % 16 == 0);
 409     assert(std::size_t(base0) % 16 == 0);
 410     assert(std::size_t(base1) % 16 == 0);
 411     assert(std::size_t(align) % 2 == 0);
 412
 413     idx0 = _mm512_extload_epi32(offset, _MM_UPCONV_EPI32_NONE, _MM_BROADCAST_4X16, _MM_HINT_NONE);
 414
 415     idx0 = _mm512_mullo_epi32(idx0, _mm512_set1_epi32(align));
 416     idx1 = _mm512_add_epi32(idx0, _mm512_set1_epi32(1));
 417
 418     idx = _mm512_mask_permute4f128_epi32(idx0, _mm512_int2mask(0x00F0), idx1, _MM_PERM_AAAA);
 419
 420     tmp1 = _mm512_i32logather_pd(idx, base0, sizeof(double));
 421     tmp2 = _mm512_i32logather_pd(idx, base1, sizeof(double));
 422
 423     v0->simdInternal_ = _mm512_castps_pd(_mm512_mask_permute4f128_ps(
 424             _mm512_castpd_ps(tmp1), _mm512_int2mask(0xFF00), _mm512_castpd_ps(tmp2), _MM_PERM_BABA));
 425     v1->simdInternal_ = _mm512_castps_pd(_mm512_mask_permute4f128_ps(
 426             _mm512_castpd_ps(tmp2), _mm512_int2mask(0x00FF), _mm512_castpd_ps(tmp1), _MM_PERM_DCDC));
 427 }
 428
 429 static inline double gmx_simdcall reduceIncr4ReturnSumHsimd(double* m, SimdDouble v0, SimdDouble v1)
 430 {
 431     double  d;
 432     __m512d t0, t1;
 433
 434     assert(std::size_t(m) % 32 == 0);
 435
 436     t0 = _mm512_add_pd(v0.simdInternal_, _mm512_swizzle_pd(v0.simdInternal_, _MM_SWIZ_REG_BADC));
 437     t0 = _mm512_mask_add_pd(t0,
 438                             _mm512_int2mask(0xCC),
 439                             v1.simdInternal_,
 440                             _mm512_swizzle_pd(v1.simdInternal_, _MM_SWIZ_REG_BADC));
 441     t0 = _mm512_add_pd(t0, _mm512_swizzle_pd(t0, _MM_SWIZ_REG_CDAB));
 442     t0 = _mm512_castps_pd(_mm512_mask_permute4f128_ps(
 443             _mm512_castpd_ps(t0), _mm512_int2mask(0xCCCC), _mm512_castpd_ps(t0), _MM_PERM_DCDC));
 444
 445     t1 = _mm512_mask_extload_pd(
 446             _mm512_undefined_pd(), _mm512_int2mask(0xF), m, _MM_UPCONV_PD_NONE, _MM_BROADCAST_4X8, _MM_HINT_NONE);
 447     t1 = _mm512_add_pd(t1, t0);
 448     _mm512_mask_packstorelo_pd(m, _mm512_int2mask(0xF), t1);
 449
 450     t0 = _mm512_add_pd(t0, _mm512_swizzle_pd(t0, _MM_SWIZ_REG_BADC));
 451     t0 = _mm512_add_pd(t0, _mm512_swizzle_pd(t0, _MM_SWIZ_REG_CDAB));
 452
 453     _mm512_mask_packstorelo_pd(&d, _mm512_mask2int(0x03), t0);
 454     return d;
 455 }
 456
 457 } // namespace gmx
 458
 459 #endif // GMX_SIMD_IMPL_X86_MIC_UTIL_DOUBLE_H