src/gromacs/simd/impl_x86_sse2/impl_x86_sse2_simd_float.h

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2014,2015,2016,2017, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 #ifndef GMX_SIMD_IMPL_X86_SSE2_SIMD_FLOAT_H
  36 #define GMX_SIMD_IMPL_X86_SSE2_SIMD_FLOAT_H
  37
  38 #include "config.h"
  39
  40 #include <cassert>
  41 #include <cstddef>
  42 #include <cstdint>
  43
  44 #include <emmintrin.h>
  45
  46 #include "gromacs/math/utilities.h"
  47
  48 namespace gmx
  49 {
  50
  51 class SimdFloat
  52 {
  53     public:
  54         SimdFloat() {}
  55
  56         SimdFloat(float f) : simdInternal_(_mm_set1_ps(f)) {}
  57
  58         // Internal utility constructor to simplify return statements
  59         SimdFloat(__m128 simd) : simdInternal_(simd) {}
  60
  61         __m128  simdInternal_;
  62 };
  63
  64 class SimdFInt32
  65 {
  66     public:
  67         SimdFInt32() {}
  68
  69         SimdFInt32(std::int32_t i) : simdInternal_(_mm_set1_epi32(i)) {}
  70
  71         // Internal utility constructor to simplify return statements
  72         SimdFInt32(__m128i simd) : simdInternal_(simd) {}
  73
  74         __m128i  simdInternal_;
  75 };
  76
  77 class SimdFBool
  78 {
  79     public:
  80         SimdFBool() {}
  81
  82         SimdFBool(bool b) : simdInternal_(_mm_castsi128_ps(_mm_set1_epi32( b ? 0xFFFFFFFF : 0))) {}
  83
  84         // Internal utility constructor to simplify return statements
  85         SimdFBool(__m128 simd) : simdInternal_(simd) {}
  86
  87         __m128  simdInternal_;
  88 };
  89
  90 class SimdFIBool
  91 {
  92     public:
  93         SimdFIBool() {}
  94
  95         SimdFIBool(bool b) : simdInternal_(_mm_set1_epi32( b ? 0xFFFFFFFF : 0)) {}
  96
  97         // Internal utility constructor to simplify return statements
  98         SimdFIBool(__m128i simd) : simdInternal_(simd) {}
  99
 100         __m128i  simdInternal_;
 101 };
 102
 103 static inline SimdFloat gmx_simdcall
 104 simdLoad(const float *m, SimdFloatTag = {})
 105 {
 106     assert(std::size_t(m) % 16 == 0);
 107     return {
 108                _mm_load_ps(m)
 109     };
 110 }
 111
 112 static inline void gmx_simdcall
 113 store(float *m, SimdFloat a)
 114 {
 115     assert(std::size_t(m) % 16 == 0);
 116     _mm_store_ps(m, a.simdInternal_);
 117 }
 118
 119 static inline SimdFloat gmx_simdcall
 120 simdLoadU(const float *m, SimdFloatTag = {})
 121 {
 122     return {
 123                _mm_loadu_ps(m)
 124     };
 125 }
 126
 127 static inline void gmx_simdcall
 128 storeU(float *m, SimdFloat a) { _mm_storeu_ps(m, a.simdInternal_); }
 129
 130 static inline SimdFloat gmx_simdcall
 131 setZeroF()
 132 {
 133     return {
 134                _mm_setzero_ps()
 135     };
 136 }
 137
 138 static inline SimdFInt32 gmx_simdcall
 139 simdLoad(const std::int32_t * m, SimdFInt32Tag)
 140 {
 141     assert(std::size_t(m) % 16 == 0);
 142     return {
 143                _mm_load_si128(reinterpret_cast<const __m128i *>(m))
 144     };
 145 }
 146
 147 static inline void gmx_simdcall
 148 store(std::int32_t * m, SimdFInt32 a)
 149 {
 150     assert(std::size_t(m) % 16 == 0);
 151     _mm_store_si128(reinterpret_cast<__m128i *>(m), a.simdInternal_);
 152 }
 153
 154 static inline SimdFInt32 gmx_simdcall
 155 simdLoadU(const std::int32_t *m, SimdFInt32Tag)
 156 {
 157     return {
 158                _mm_loadu_si128(reinterpret_cast<const __m128i *>(m))
 159     };
 160 }
 161
 162 static inline void gmx_simdcall
 163 storeU(std::int32_t * m, SimdFInt32 a)
 164 {
 165     _mm_storeu_si128(reinterpret_cast<__m128i *>(m), a.simdInternal_);
 166 }
 167
 168 static inline SimdFInt32 gmx_simdcall
 169 setZeroFI()
 170 {
 171     return {
 172                _mm_setzero_si128()
 173     };
 174 }
 175
 176
 177 // Override for SSE4.1 and higher
 178 #if GMX_SIMD_X86_SSE2
 179 template<int index>
 180 static inline std::int32_t gmx_simdcall
 181 extract(SimdFInt32 a)
 182 {
 183     return _mm_cvtsi128_si32( _mm_srli_si128(a.simdInternal_, 4 * index) );
 184 }
 185 #endif
 186
 187 static inline SimdFloat gmx_simdcall
 188 operator&(SimdFloat a, SimdFloat b)
 189 {
 190     return {
 191                _mm_and_ps(a.simdInternal_, b.simdInternal_)
 192     };
 193 }
 194
 195 static inline SimdFloat gmx_simdcall
 196 andNot(SimdFloat a, SimdFloat b)
 197 {
 198     return {
 199                _mm_andnot_ps(a.simdInternal_, b.simdInternal_)
 200     };
 201 }
 202
 203 static inline SimdFloat gmx_simdcall
 204 operator|(SimdFloat a, SimdFloat b)
 205 {
 206     return {
 207                _mm_or_ps(a.simdInternal_, b.simdInternal_)
 208     };
 209 }
 210
 211 static inline SimdFloat gmx_simdcall
 212 operator^(SimdFloat a, SimdFloat b)
 213 {
 214     return {
 215                _mm_xor_ps(a.simdInternal_, b.simdInternal_)
 216     };
 217 }
 218
 219 static inline SimdFloat gmx_simdcall
 220 operator+(SimdFloat a, SimdFloat b)
 221 {
 222     return {
 223                _mm_add_ps(a.simdInternal_, b.simdInternal_)
 224     };
 225 }
 226
 227 static inline SimdFloat gmx_simdcall
 228 operator-(SimdFloat a, SimdFloat b)
 229 {
 230     return {
 231                _mm_sub_ps(a.simdInternal_, b.simdInternal_)
 232     };
 233 }
 234
 235 static inline SimdFloat gmx_simdcall
 236 operator-(SimdFloat x)
 237 {
 238     return {
 239                _mm_xor_ps(x.simdInternal_, _mm_set1_ps(GMX_FLOAT_NEGZERO))
 240     };
 241 }
 242
 243 static inline SimdFloat gmx_simdcall
 244 operator*(SimdFloat a, SimdFloat b)
 245 {
 246     return {
 247                _mm_mul_ps(a.simdInternal_, b.simdInternal_)
 248     };
 249 }
 250
 251 // Override for AVX-128-FMA and higher
 252 #if GMX_SIMD_X86_SSE2 || GMX_SIMD_X86_SSE4_1
 253 static inline SimdFloat gmx_simdcall
 254 fma(SimdFloat a, SimdFloat b, SimdFloat c)
 255 {
 256     return {
 257                _mm_add_ps(_mm_mul_ps(a.simdInternal_, b.simdInternal_), c.simdInternal_)
 258     };
 259 }
 260
 261 static inline SimdFloat gmx_simdcall
 262 fms(SimdFloat a, SimdFloat b, SimdFloat c)
 263 {
 264     return {
 265                _mm_sub_ps(_mm_mul_ps(a.simdInternal_, b.simdInternal_), c.simdInternal_)
 266     };
 267 }
 268
 269 static inline SimdFloat gmx_simdcall
 270 fnma(SimdFloat a, SimdFloat b, SimdFloat c)
 271 {
 272     return {
 273                _mm_sub_ps(c.simdInternal_, _mm_mul_ps(a.simdInternal_, b.simdInternal_))
 274     };
 275 }
 276
 277 static inline SimdFloat gmx_simdcall
 278 fnms(SimdFloat a, SimdFloat b, SimdFloat c)
 279 {
 280     return {
 281                _mm_sub_ps(_mm_setzero_ps(), _mm_add_ps(_mm_mul_ps(a.simdInternal_, b.simdInternal_), c.simdInternal_))
 282     };
 283 }
 284 #endif
 285
 286 static inline SimdFloat gmx_simdcall
 287 rsqrt(SimdFloat x)
 288 {
 289     return {
 290                _mm_rsqrt_ps(x.simdInternal_)
 291     };
 292 }
 293
 294 static inline SimdFloat gmx_simdcall
 295 rcp(SimdFloat x)
 296 {
 297     return {
 298                _mm_rcp_ps(x.simdInternal_)
 299     };
 300 }
 301
 302 static inline SimdFloat gmx_simdcall
 303 maskAdd(SimdFloat a, SimdFloat b, SimdFBool m)
 304 {
 305     return {
 306                _mm_add_ps(a.simdInternal_, _mm_and_ps(b.simdInternal_, m.simdInternal_))
 307     };
 308 }
 309
 310 static inline SimdFloat gmx_simdcall
 311 maskzMul(SimdFloat a, SimdFloat b, SimdFBool m)
 312 {
 313     return {
 314                _mm_and_ps(_mm_mul_ps(a.simdInternal_, b.simdInternal_), m.simdInternal_)
 315     };
 316 }
 317
 318 static inline SimdFloat gmx_simdcall
 319 maskzFma(SimdFloat a, SimdFloat b, SimdFloat c, SimdFBool m)
 320 {
 321     return {
 322                _mm_and_ps(_mm_add_ps(_mm_mul_ps(a.simdInternal_, b.simdInternal_), c.simdInternal_), m.simdInternal_)
 323     };
 324 }
 325
 326 // Override for SSE4.1 and higher
 327 #if GMX_SIMD_X86_SSE2
 328 static inline SimdFloat gmx_simdcall
 329 maskzRsqrt(SimdFloat x, SimdFBool m)
 330 {
 331 #ifndef NDEBUG
 332     x.simdInternal_ = _mm_or_ps(_mm_andnot_ps(m.simdInternal_, _mm_set1_ps(1.0f)), _mm_and_ps(m.simdInternal_, x.simdInternal_));
 333 #endif
 334     return {
 335                _mm_and_ps(_mm_rsqrt_ps(x.simdInternal_), m.simdInternal_)
 336     };
 337 }
 338
 339 static inline SimdFloat gmx_simdcall
 340 maskzRcp(SimdFloat x, SimdFBool m)
 341 {
 342 #ifndef NDEBUG
 343     x.simdInternal_ = _mm_or_ps(_mm_andnot_ps(m.simdInternal_, _mm_set1_ps(1.0f)), _mm_and_ps(m.simdInternal_, x.simdInternal_));
 344 #endif
 345     return {
 346                _mm_and_ps(_mm_rcp_ps(x.simdInternal_), m.simdInternal_)
 347     };
 348 }
 349 #endif
 350
 351 static inline SimdFloat gmx_simdcall
 352 abs(SimdFloat x)
 353 {
 354     return {
 355                _mm_andnot_ps( _mm_set1_ps(GMX_FLOAT_NEGZERO), x.simdInternal_ )
 356     };
 357 }
 358
 359 static inline SimdFloat gmx_simdcall
 360 max(SimdFloat a, SimdFloat b)
 361 {
 362     return {
 363                _mm_max_ps(a.simdInternal_, b.simdInternal_)
 364     };
 365 }
 366
 367 static inline SimdFloat gmx_simdcall
 368 min(SimdFloat a, SimdFloat b)
 369 {
 370     return {
 371                _mm_min_ps(a.simdInternal_, b.simdInternal_)
 372     };
 373 }
 374
 375 // Override for SSE4.1 and higher
 376 #if GMX_SIMD_X86_SSE2
 377 static inline SimdFloat gmx_simdcall
 378 round(SimdFloat x)
 379 {
 380     return {
 381                _mm_cvtepi32_ps( _mm_cvtps_epi32(x.simdInternal_) )
 382     };
 383 }
 384
 385 static inline SimdFloat gmx_simdcall
 386 trunc(SimdFloat x)
 387 {
 388     return {
 389                _mm_cvtepi32_ps( _mm_cvttps_epi32(x.simdInternal_) )
 390     };
 391 }
 392
 393 #endif
 394
 395 static inline SimdFloat gmx_simdcall
 396 frexp(SimdFloat value, SimdFInt32 * exponent)
 397 {
 398     const __m128  exponentMask   = _mm_castsi128_ps(_mm_set1_epi32(0x7F800000));
 399     const __m128  mantissaMask   = _mm_castsi128_ps(_mm_set1_epi32(0x807FFFFF));
 400     const __m128i exponentBias   = _mm_set1_epi32(126); // add 1 to make our definition identical to frexp()
 401     const __m128  half           = _mm_set1_ps(0.5f);
 402     __m128i       iExponent;
 403
 404     iExponent               = _mm_castps_si128(_mm_and_ps(value.simdInternal_, exponentMask));
 405     iExponent               = _mm_sub_epi32(_mm_srli_epi32(iExponent, 23), exponentBias);
 406     exponent->simdInternal_ = iExponent;
 407
 408     return {
 409                _mm_or_ps( _mm_and_ps(value.simdInternal_, mantissaMask), half)
 410     };
 411 }
 412
 413 // Override for SSE4.1
 414 #if GMX_SIMD_X86_SSE2
 415 template <MathOptimization opt = MathOptimization::Safe>
 416 static inline SimdFloat gmx_simdcall
 417 ldexp(SimdFloat value, SimdFInt32 exponent)
 418 {
 419     const __m128i exponentBias = _mm_set1_epi32(127);
 420     __m128i       iExponent;
 421
 422     iExponent = _mm_add_epi32(exponent.simdInternal_, exponentBias);
 423
 424     if (opt == MathOptimization::Safe)
 425     {
 426         // Make sure biased argument is not negative
 427         iExponent = _mm_and_si128(iExponent, _mm_cmpgt_epi32(iExponent, _mm_setzero_si128()));
 428     }
 429
 430     iExponent = _mm_slli_epi32( iExponent, 23);
 431
 432     return {
 433                _mm_mul_ps(value.simdInternal_, _mm_castsi128_ps(iExponent))
 434     };
 435 }
 436 #endif
 437
 438 // Override for AVX-128-FMA and higher
 439 #if GMX_SIMD_X86_SSE2 || GMX_SIMD_X86_SSE4_1
 440 static inline float gmx_simdcall
 441 reduce(SimdFloat a)
 442 {
 443     // Shuffle has latency 1/throughput 1, followed by add with latency 3, t-put 1.
 444     // This is likely faster than using _mm_hadd_ps, which has latency 5, t-put 2.
 445     a.simdInternal_ = _mm_add_ps(a.simdInternal_, _mm_shuffle_ps(a.simdInternal_, a.simdInternal_, _MM_SHUFFLE(1, 0, 3, 2)));
 446     a.simdInternal_ = _mm_add_ss(a.simdInternal_, _mm_shuffle_ps(a.simdInternal_, a.simdInternal_, _MM_SHUFFLE(0, 3, 2, 1)));
 447     return *reinterpret_cast<float *>(&a);
 448 }
 449 #endif
 450
 451 static inline SimdFBool gmx_simdcall
 452 operator==(SimdFloat a, SimdFloat b)
 453 {
 454     return {
 455                _mm_cmpeq_ps(a.simdInternal_, b.simdInternal_)
 456     };
 457 }
 458
 459 static inline SimdFBool gmx_simdcall
 460 operator!=(SimdFloat a, SimdFloat b)
 461 {
 462     return {
 463                _mm_cmpneq_ps(a.simdInternal_, b.simdInternal_)
 464     };
 465 }
 466
 467 static inline SimdFBool gmx_simdcall
 468 operator<(SimdFloat a, SimdFloat b)
 469 {
 470     return {
 471                _mm_cmplt_ps(a.simdInternal_, b.simdInternal_)
 472     };
 473 }
 474
 475 static inline SimdFBool gmx_simdcall
 476 operator<=(SimdFloat a, SimdFloat b)
 477 {
 478     return {
 479                _mm_cmple_ps(a.simdInternal_, b.simdInternal_)
 480     };
 481 }
 482
 483 static inline SimdFBool gmx_simdcall
 484 testBits(SimdFloat a)
 485 {
 486     __m128i ia  = _mm_castps_si128(a.simdInternal_);
 487     __m128i res = _mm_andnot_si128( _mm_cmpeq_epi32(ia, _mm_setzero_si128()), _mm_cmpeq_epi32(ia, ia));
 488
 489     return {
 490                _mm_castsi128_ps(res)
 491     };
 492 }
 493
 494 static inline SimdFBool gmx_simdcall
 495 operator&&(SimdFBool a, SimdFBool b)
 496 {
 497     return {
 498                _mm_and_ps(a.simdInternal_, b.simdInternal_)
 499     };
 500 }
 501
 502 static inline SimdFBool gmx_simdcall
 503 operator||(SimdFBool a, SimdFBool b)
 504 {
 505     return {
 506                _mm_or_ps(a.simdInternal_, b.simdInternal_)
 507     };
 508 }
 509
 510 static inline bool gmx_simdcall
 511 anyTrue(SimdFBool a) { return _mm_movemask_ps(a.simdInternal_) != 0; }
 512
 513 static inline SimdFloat gmx_simdcall
 514 selectByMask(SimdFloat a, SimdFBool mask)
 515 {
 516     return {
 517                _mm_and_ps(a.simdInternal_, mask.simdInternal_)
 518     };
 519 }
 520
 521 static inline SimdFloat gmx_simdcall
 522 selectByNotMask(SimdFloat a, SimdFBool mask)
 523 {
 524     return {
 525                _mm_andnot_ps(mask.simdInternal_, a.simdInternal_)
 526     };
 527 }
 528
 529 // Override for SSE4.1 and higher
 530 #if GMX_SIMD_X86_SSE2
 531 static inline SimdFloat gmx_simdcall
 532 blend(SimdFloat a, SimdFloat b, SimdFBool sel)
 533 {
 534     return {
 535                _mm_or_ps(_mm_andnot_ps(sel.simdInternal_, a.simdInternal_), _mm_and_ps(sel.simdInternal_, b.simdInternal_))
 536     };
 537 }
 538 #endif
 539
 540 static inline SimdFInt32 gmx_simdcall
 541 operator&(SimdFInt32 a, SimdFInt32 b)
 542 {
 543     return {
 544                _mm_and_si128(a.simdInternal_, b.simdInternal_)
 545     };
 546 }
 547
 548 static inline SimdFInt32 gmx_simdcall
 549 andNot(SimdFInt32 a, SimdFInt32 b)
 550 {
 551     return {
 552                _mm_andnot_si128(a.simdInternal_, b.simdInternal_)
 553     };
 554 }
 555
 556 static inline SimdFInt32 gmx_simdcall
 557 operator|(SimdFInt32 a, SimdFInt32 b)
 558 {
 559     return {
 560                _mm_or_si128(a.simdInternal_, b.simdInternal_)
 561     };
 562 }
 563
 564 static inline SimdFInt32 gmx_simdcall
 565 operator^(SimdFInt32 a, SimdFInt32 b)
 566 {
 567     return {
 568                _mm_xor_si128(a.simdInternal_, b.simdInternal_)
 569     };
 570 }
 571
 572 static inline SimdFInt32 gmx_simdcall
 573 operator+(SimdFInt32 a, SimdFInt32 b)
 574 {
 575     return {
 576                _mm_add_epi32(a.simdInternal_, b.simdInternal_)
 577     };
 578 }
 579
 580 static inline SimdFInt32 gmx_simdcall
 581 operator-(SimdFInt32 a, SimdFInt32 b)
 582 {
 583     return {
 584                _mm_sub_epi32(a.simdInternal_, b.simdInternal_)
 585     };
 586 }
 587
 588 // Override for SSE4.1 and higher
 589 #if GMX_SIMD_X86_SSE2
 590 static inline SimdFInt32 gmx_simdcall
 591 operator*(SimdFInt32 a, SimdFInt32 b)
 592 {
 593     __m128i a1 = _mm_srli_si128(a.simdInternal_, 4); // - a[3] a[2] a[1]
 594     __m128i b1 = _mm_srli_si128(b.simdInternal_, 4); // - b[3] b[2] b[1]
 595     __m128i c  = _mm_mul_epu32(a.simdInternal_, b.simdInternal_);
 596     __m128i c1 = _mm_mul_epu32(a1, b1);
 597
 598     c  = _mm_shuffle_epi32(c, _MM_SHUFFLE(3, 1, 2, 0));  // - - a[2]*b[2] a[0]*b[0]
 599     c1 = _mm_shuffle_epi32(c1, _MM_SHUFFLE(3, 1, 2, 0)); // - - a[3]*b[3] a[1]*b[1]
 600
 601     return {
 602                _mm_unpacklo_epi32(c, c1)
 603     };
 604 }
 605 #endif
 606
 607 static inline SimdFIBool gmx_simdcall
 608 operator==(SimdFInt32 a, SimdFInt32 b)
 609 {
 610     return {
 611                _mm_cmpeq_epi32(a.simdInternal_, b.simdInternal_)
 612     };
 613 }
 614
 615 static inline SimdFIBool gmx_simdcall
 616 testBits(SimdFInt32 a)
 617 {
 618     __m128i x   = a.simdInternal_;
 619     __m128i res = _mm_andnot_si128( _mm_cmpeq_epi32(x, _mm_setzero_si128()), _mm_cmpeq_epi32(x, x));
 620
 621     return {
 622                res
 623     };
 624 }
 625
 626 static inline SimdFIBool gmx_simdcall
 627 operator<(SimdFInt32 a, SimdFInt32 b)
 628 {
 629     return {
 630                _mm_cmplt_epi32(a.simdInternal_, b.simdInternal_)
 631     };
 632 }
 633
 634 static inline SimdFIBool gmx_simdcall
 635 operator&&(SimdFIBool a, SimdFIBool b)
 636 {
 637     return {
 638                _mm_and_si128(a.simdInternal_, b.simdInternal_)
 639     };
 640 }
 641
 642 static inline SimdFIBool gmx_simdcall
 643 operator||(SimdFIBool a, SimdFIBool b)
 644 {
 645     return {
 646                _mm_or_si128(a.simdInternal_, b.simdInternal_)
 647     };
 648 }
 649
 650 static inline bool gmx_simdcall
 651 anyTrue(SimdFIBool a) { return _mm_movemask_epi8(a.simdInternal_) != 0; }
 652
 653 static inline SimdFInt32 gmx_simdcall
 654 selectByMask(SimdFInt32 a, SimdFIBool mask)
 655 {
 656     return {
 657                _mm_and_si128(a.simdInternal_, mask.simdInternal_)
 658     };
 659 }
 660
 661 static inline SimdFInt32 gmx_simdcall
 662 selectByNotMask(SimdFInt32 a, SimdFIBool mask)
 663 {
 664     return {
 665                _mm_andnot_si128(mask.simdInternal_, a.simdInternal_)
 666     };
 667 }
 668
 669 // Override for SSE4.1 and higher
 670 #if GMX_SIMD_X86_SSE2
 671 static inline SimdFInt32 gmx_simdcall
 672 blend(SimdFInt32 a, SimdFInt32 b, SimdFIBool sel)
 673 {
 674     return {
 675                _mm_or_si128(_mm_andnot_si128(sel.simdInternal_, a.simdInternal_), _mm_and_si128(sel.simdInternal_, b.simdInternal_))
 676     };
 677 }
 678 #endif
 679
 680 static inline SimdFInt32 gmx_simdcall
 681 cvtR2I(SimdFloat a)
 682 {
 683     return {
 684                _mm_cvtps_epi32(a.simdInternal_)
 685     };
 686 }
 687
 688 static inline SimdFInt32 gmx_simdcall
 689 cvttR2I(SimdFloat a)
 690 {
 691     return {
 692                _mm_cvttps_epi32(a.simdInternal_)
 693     };
 694 }
 695
 696 static inline SimdFloat gmx_simdcall
 697 cvtI2R(SimdFInt32 a)
 698 {
 699     return {
 700                _mm_cvtepi32_ps(a.simdInternal_)
 701     };
 702 }
 703
 704 static inline SimdFIBool gmx_simdcall
 705 cvtB2IB(SimdFBool a)
 706 {
 707     return {
 708                _mm_castps_si128(a.simdInternal_)
 709     };
 710 }
 711
 712 static inline SimdFBool gmx_simdcall
 713 cvtIB2B(SimdFIBool a)
 714 {
 715     return {
 716                _mm_castsi128_ps(a.simdInternal_)
 717     };
 718 }
 719
 720 }      // namespace gmx
 721
 722 #endif // GMX_SIMD_IMPL_X86_SSE2_SIMD_FLOAT_H