src/gromacs/simd/tests/simd_floatingpoint.cpp

   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, 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 #include "gmxpre.h"
  36
  37 #include <cmath>
  38
  39 #include "gromacs/math/utilities.h"
  40 #include "gromacs/simd/simd.h"
  41 #include "gromacs/utility/basedefinitions.h"
  42
  43 #include "testutils/testasserts.h"
  44
  45 #include "data.h"
  46 #include "simd.h"
  47
  48 #if GMX_SIMD
  49
  50 namespace gmx
  51 {
  52 namespace test
  53 {
  54
  55 namespace
  56 {
  57
  58 /*! \cond internal */
  59 /*! \addtogroup module_simd */
  60 /*! \{ */
  61
  62 #if GMX_SIMD_HAVE_REAL
  63
  64 /*! \brief Test fixture for floating-point tests (identical to the generic \ref SimdTest) */
  65 typedef SimdTest SimdFloatingpointTest;
  66
  67 TEST_F(SimdFloatingpointTest, setZero)
  68 {
  69     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(0.0), setZero());
  70 }
  71
  72 TEST_F(SimdFloatingpointTest, set)
  73 {
  74     const real *p  = &c0;
  75     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(c1), SimdReal(c1));
  76     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(c0), SimdReal(*p));
  77 }
  78
  79 TEST_F(SimdFloatingpointTest, add)
  80 {
  81     GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 + c3, c1 + c4, c2 + c5 ),
  82                               rSimd_c0c1c2 + rSimd_c3c4c5);
  83 }
  84
  85 TEST_F(SimdFloatingpointTest, maskAdd)
  86 {
  87     SimdBool m = setSimdRealFrom3R(c6, 0, c7) != setZero();
  88     GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 + c3, c1 + 0.0, c2 + c5 ),
  89                               maskAdd(rSimd_c0c1c2, rSimd_c3c4c5, m));
  90 }
  91
  92 TEST_F(SimdFloatingpointTest, sub)
  93 {
  94     GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 - c3, c1 - c4, c2 - c5 ),
  95                               rSimd_c0c1c2 - rSimd_c3c4c5);
  96 }
  97
  98 TEST_F(SimdFloatingpointTest, mul)
  99 {
 100     GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 * c3, c1 * c4, c2 * c5 ),
 101                               rSimd_c0c1c2 * rSimd_c3c4c5);
 102 }
 103
 104 TEST_F(SimdFloatingpointTest, maskzMul)
 105 {
 106     SimdBool m = setSimdRealFrom3R(c1, 0, c1) != setZero();
 107     GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 * c3, 0.0, c2 * c5 ),
 108                               maskzMul(rSimd_c0c1c2, rSimd_c3c4c5, m));
 109 }
 110
 111 TEST_F(SimdFloatingpointTest, fma)
 112 {
 113     // The last bit of FMA operations depends on hardware, so we don't require exact match
 114     GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 * c3 + c6, c1 * c4 + c7, c2 * c5 + c8),
 115                               fma(rSimd_c0c1c2, rSimd_c3c4c5, rSimd_c6c7c8));
 116 }
 117
 118
 119 TEST_F(SimdFloatingpointTest, maskzFma)
 120 {
 121     SimdBool m = setSimdRealFrom3R(c2, 0, c3) != setZero();
 122     // The last bit of FMA operations depends on hardware, so we don't require exact match
 123     GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 * c3 + c6, 0.0, c2 * c5 + c8),
 124                               maskzFma(rSimd_c0c1c2, rSimd_c3c4c5, rSimd_c6c7c8, m));
 125 }
 126
 127 TEST_F(SimdFloatingpointTest, fms)
 128 {
 129     // The last bit of FMA operations depends on hardware, so we don't require exact match
 130     GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 * c3 - c6, c1 * c4 - c7, c2 * c5 - c8),
 131                               fms(rSimd_c0c1c2, rSimd_c3c4c5, rSimd_c6c7c8));
 132 }
 133
 134 TEST_F(SimdFloatingpointTest, fnma)
 135 {
 136     // The last bit of FMA operations depends on hardware, so we don't require exact match
 137     GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c6 - c0 * c3, c7 - c1 * c4, c8 - c2 * c5),
 138                               fnma(rSimd_c0c1c2, rSimd_c3c4c5, rSimd_c6c7c8));
 139 }
 140
 141 TEST_F(SimdFloatingpointTest, fnms)
 142 {
 143     // The last bit of FMA operations depends on hardware, so we don't require exact match
 144     GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(-c0 * c3 - c6, -c1 * c4 - c7, -c2 * c5 - c8),
 145                               fnms(rSimd_c0c1c2, rSimd_c3c4c5, rSimd_c6c7c8));
 146 }
 147
 148 TEST_F(SimdFloatingpointTest, abs)
 149 {
 150     GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, abs(rSimd_c0c1c2)); // fabs(x)=x
 151     GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, abs(rSimd_m0m1m2)); // fabs(-x)=x
 152 }
 153
 154 TEST_F(SimdFloatingpointTest, neg)
 155 {
 156     GMX_EXPECT_SIMD_REAL_EQ(rSimd_m0m1m2, -(rSimd_c0c1c2)); // fneg(x)=-x
 157     GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, -(rSimd_m0m1m2)); // fneg(-x)=x
 158 }
 159
 160 #if GMX_SIMD_HAVE_LOGICAL
 161 TEST_F(SimdFloatingpointTest, and)
 162 {
 163     GMX_EXPECT_SIMD_REAL_EQ(rSimd_logicalResultAnd,
 164                             (rSimd_logicalA & rSimd_logicalB));
 165 }
 166
 167 TEST_F(SimdFloatingpointTest, or)
 168 {
 169     GMX_EXPECT_SIMD_REAL_EQ(rSimd_logicalResultOr,
 170                             (rSimd_logicalA | rSimd_logicalB));
 171 }
 172
 173 TEST_F(SimdFloatingpointTest, xor)
 174 {
 175     /* Test xor by taking xor with a number and its negative. This should result
 176      * in only the sign bit being set. We then use this bit change the sign of
 177      * different numbers.
 178      */
 179     SimdReal signbit = SimdReal(c1) ^ SimdReal(-c1);
 180     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-c2, c3, -c4), (signbit ^ setSimdRealFrom3R(c2, -c3, c4)));
 181 }
 182
 183 TEST_F(SimdFloatingpointTest, andNot)
 184 {
 185     /* Use xor (which we already tested, so fix that first if both tests fail)
 186      * to extract the sign bit, and then use andnot to take absolute values.
 187      */
 188     SimdReal signbit = SimdReal(c1) ^ SimdReal(-c1);
 189     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c2, c3, c4), andNot(signbit, setSimdRealFrom3R(-c2, c3, -c4)));
 190 }
 191
 192 #endif
 193
 194 TEST_F(SimdFloatingpointTest, max)
 195 {
 196     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R( c3,  c1,  c4), max(rSimd_c0c1c2, rSimd_c3c0c4));
 197     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R( c3,  c1,  c4), max(rSimd_c3c0c4, rSimd_c0c1c2));
 198     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-c0, -c0, -c2), max(rSimd_m0m1m2, rSimd_m3m0m4));
 199     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-c0, -c0, -c2), max(rSimd_m3m0m4, rSimd_m0m1m2));
 200 }
 201
 202 TEST_F(SimdFloatingpointTest, min)
 203 {
 204     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R( c0,  c0,  c2), min(rSimd_c0c1c2, rSimd_c3c0c4));
 205     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R( c0,  c0,  c2), min(rSimd_c3c0c4, rSimd_c0c1c2));
 206     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-c3, -c1, -c4), min(rSimd_m0m1m2, rSimd_m3m0m4));
 207     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-c3, -c1, -c4), min(rSimd_m3m0m4, rSimd_m0m1m2));
 208 }
 209
 210 TEST_F(SimdFloatingpointTest, round)
 211 {
 212     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(2), round(rSimd_2p25));
 213     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(4), round(rSimd_3p75));
 214     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-2), round(rSimd_m2p25));
 215     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-4), round(rSimd_m3p75));
 216 }
 217
 218 TEST_F(SimdFloatingpointTest, roundMode)
 219 {
 220     /* Rounding mode needs to be consistent between round and cvtR2I */
 221     SimdReal x0  = setSimdRealFrom3R(0.5, 11.5, 99.5);
 222     SimdReal x1  = setSimdRealFrom3R(-0.5, -11.5, -99.5);
 223
 224     GMX_EXPECT_SIMD_REAL_EQ(round(x0), cvtI2R(cvtR2I(x0)));
 225     GMX_EXPECT_SIMD_REAL_EQ(round(x1), cvtI2R(cvtR2I(x1)));
 226 }
 227
 228 TEST_F(SimdFloatingpointTest, trunc)
 229 {
 230     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(2), trunc(rSimd_2p25));
 231     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(3), trunc(rSimd_3p75));
 232     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-2), trunc(rSimd_m2p25));
 233     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-3), trunc(rSimd_m3p75));
 234 }
 235
 236 // We explicitly test the exponent/mantissa routines with double precision data,
 237 // since these usually rely on direct manipulation and shift of the SIMD registers,
 238 // where it is easy to make mistakes with single vs double precision.
 239
 240 TEST_F(SimdFloatingpointTest, frexp)
 241 {
 242     SimdReal  fraction;
 243     SimdInt32 exponent;
 244
 245     fraction = frexp(rSimd_Exp, &exponent);
 246
 247     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(0.609548660288905419513128,
 248                                               0.5833690139241746175358116,
 249                                               -0.584452007502232362412542),
 250                             fraction);
 251     GMX_EXPECT_SIMD_INT_EQ(setSimdIntFrom3I(61, -40, 55), exponent);
 252
 253
 254 #if GMX_SIMD_HAVE_DOUBLE && GMX_DOUBLE
 255     fraction = frexp(rSimd_ExpDouble, &exponent);
 256
 257     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(0.6206306194761728178832527,
 258                                               0.5236473618795619566768096,
 259                                               -0.9280331023751380303821179),
 260                             fraction);
 261     GMX_EXPECT_SIMD_INT_EQ(setSimdIntFrom3I(588, -461, 673), exponent);
 262 #endif
 263 }
 264
 265 TEST_F(SimdFloatingpointTest, ldexp)
 266 {
 267     SimdReal one = setSimdRealFrom1R(1.0);
 268
 269     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(pow(2.0, 60.0), pow(2.0, -41.0), pow(2.0, 54.0)),
 270                             ldexp<MathOptimization::Unsafe>(one, setSimdIntFrom3I(60, -41, 54)));
 271 #if GMX_SIMD_HAVE_DOUBLE && GMX_DOUBLE
 272     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(pow(2.0, 587.0), pow(2.0, -462.0), pow(2.0, 672.0)),
 273                             ldexp<MathOptimization::Unsafe>(one, setSimdIntFrom3I(587, -462, 672)));
 274 #endif
 275     // The default safe version must be able to handle very negative arguments too
 276     GMX_EXPECT_SIMD_REAL_EQ(setZero(), ldexp(one, setSimdIntFrom3I(-2000, -1000000, -1000000000)));
 277 }
 278
 279 /*
 280  * We do extensive 1/sqrt(x) and 1/x accuracy testing in the math module, so
 281  * we just make sure the lookup instructions appear to work here
 282  */
 283
 284 TEST_F(SimdFloatingpointTest, rsqrt)
 285 {
 286     SimdReal        x                  = setSimdRealFrom3R(4.0, M_PI, 1234567890.0);
 287     SimdReal        ref                = setSimdRealFrom3R(0.5, 1.0/std::sqrt(M_PI), 1.0/std::sqrt(1234567890.0));
 288     int             shiftbits          = std::numeric_limits<real>::digits-GMX_SIMD_RSQRT_BITS;
 289
 290     if (shiftbits < 0)
 291     {
 292         shiftbits = 0;
 293     }
 294
 295     /* Set the allowed ulp error as 2 to the power of the number of bits in
 296      * the mantissa that do not have to be correct after the table lookup.
 297      */
 298     setUlpTol(1LL << shiftbits);
 299     GMX_EXPECT_SIMD_REAL_NEAR(ref, rsqrt(x));
 300 }
 301
 302 TEST_F(SimdFloatingpointTest, maskzRsqrt)
 303 {
 304     SimdReal        x                  = setSimdRealFrom3R(M_PI, -4.0, 0.0);
 305     // simdCmpLe is tested separately further down
 306     SimdBool        m                  = setZero() < x;
 307     SimdReal        ref                = setSimdRealFrom3R(1.0/std::sqrt(M_PI), 0.0, 0.0);
 308     int             shiftbits          = std::numeric_limits<real>::digits-GMX_SIMD_RSQRT_BITS;
 309
 310     if (shiftbits < 0)
 311     {
 312         shiftbits = 0;
 313     }
 314
 315     /* Set the allowed ulp error as 2 to the power of the number of bits in
 316      * the mantissa that do not have to be correct after the table lookup.
 317      */
 318     setUlpTol(1LL << shiftbits);
 319     GMX_EXPECT_SIMD_REAL_NEAR(ref, maskzRsqrt(x, m));
 320 }
 321
 322 TEST_F(SimdFloatingpointTest, rcp)
 323 {
 324     SimdReal        x                  = setSimdRealFrom3R(4.0, M_PI, 1234567890.0);
 325     SimdReal        ref                = setSimdRealFrom3R(0.25, 1.0/M_PI, 1.0/1234567890.0);
 326     int             shiftbits          = std::numeric_limits<real>::digits-GMX_SIMD_RCP_BITS;
 327
 328     if (shiftbits < 0)
 329     {
 330         shiftbits = 0;
 331     }
 332
 333     /* Set the allowed ulp error as 2 to the power of the number of bits in
 334      * the mantissa that do not have to be correct after the table lookup.
 335      */
 336     setUlpTol(1LL << shiftbits);
 337     GMX_EXPECT_SIMD_REAL_NEAR(ref, rcp(x));
 338 }
 339
 340 TEST_F(SimdFloatingpointTest, maskzRcp)
 341 {
 342     SimdReal        x                  = setSimdRealFrom3R(M_PI, 0.0, -1234567890.0);
 343     SimdBool        m                  = (x != setZero());
 344     SimdReal        ref                = setSimdRealFrom3R(1.0/M_PI, 0.0, -1.0/1234567890.0);
 345     int             shiftbits          = std::numeric_limits<real>::digits-GMX_SIMD_RCP_BITS;
 346
 347     if (shiftbits < 0)
 348     {
 349         shiftbits = 0;
 350     }
 351
 352     /* Set the allowed ulp error as 2 to the power of the number of bits in
 353      * the mantissa that do not have to be correct after the table lookup.
 354      */
 355     setUlpTol(1LL << shiftbits);
 356     GMX_EXPECT_SIMD_REAL_NEAR(ref, maskzRcp(x, m));
 357 }
 358
 359 TEST_F(SimdFloatingpointTest, cmpEqAndSelectByMask)
 360 {
 361     SimdBool eq   = rSimd_c4c6c8 == rSimd_c6c7c8;
 362     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(0, 0, c2), selectByMask(rSimd_c0c1c2, eq));
 363 }
 364
 365 TEST_F(SimdFloatingpointTest, selectByNotMask)
 366 {
 367     SimdBool eq   = rSimd_c4c6c8 == rSimd_c6c7c8;
 368     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, c1, 0), selectByNotMask(rSimd_c0c1c2, eq));
 369 }
 370
 371 TEST_F(SimdFloatingpointTest, cmpNe)
 372 {
 373     SimdBool eq   = rSimd_c4c6c8 != rSimd_c6c7c8;
 374     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, c1, 0), selectByMask(rSimd_c0c1c2, eq));
 375 }
 376
 377 TEST_F(SimdFloatingpointTest, cmpLe)
 378 {
 379     SimdBool le   = rSimd_c4c6c8 <= rSimd_c6c7c8;
 380     GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, selectByMask(rSimd_c0c1c2, le));
 381 }
 382
 383 TEST_F(SimdFloatingpointTest, cmpLt)
 384 {
 385     SimdBool lt   = rSimd_c4c6c8 < rSimd_c6c7c8;
 386     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, c1, 0), selectByMask(rSimd_c0c1c2, lt));
 387 }
 388
 389 #if GMX_SIMD_HAVE_INT32_LOGICAL || GMX_SIMD_HAVE_LOGICAL
 390 TEST_F(SimdFloatingpointTest, testBits)
 391 {
 392     SimdBool eq   = testBits(setSimdRealFrom3R(c1, 0, c1));
 393     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, 0, c2), selectByMask(rSimd_c0c1c2, eq));
 394
 395     // Test if we detect only the sign bit being set
 396     eq            = testBits(setSimdRealFrom1R(GMX_REAL_NEGZERO));
 397     GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, selectByMask(rSimd_c0c1c2, eq));
 398 }
 399 #endif
 400
 401 TEST_F(SimdFloatingpointTest, andB)
 402 {
 403     SimdBool eq   = rSimd_c4c6c8 == rSimd_c6c7c8;
 404     SimdBool le   = rSimd_c4c6c8 <= rSimd_c6c7c8;
 405     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(0, 0, c2), selectByMask(rSimd_c0c1c2, (eq && le)));
 406 }
 407
 408 TEST_F(SimdFloatingpointTest, orB)
 409 {
 410     SimdBool eq   = rSimd_c4c6c8 == rSimd_c6c7c8;
 411     SimdBool lt   = rSimd_c4c6c8  < rSimd_c6c7c8;
 412     GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, selectByMask(rSimd_c0c1c2, (eq || lt)));
 413 }
 414
 415 TEST_F(SimdFloatingpointTest, anyTrueB)
 416 {
 417     SimdBool eq;
 418
 419     /* this test is a bit tricky since we don't know the simd width.
 420      * We cannot check for truth values for "any" element beyond the first,
 421      * since that part of the data will not be used if simd width is 1.
 422      */
 423     eq = rSimd_c4c6c8 == setSimdRealFrom3R(c4, 0, 0);
 424     EXPECT_TRUE(anyTrue(eq));
 425
 426     eq = rSimd_c0c1c2 == rSimd_c3c4c5;
 427     EXPECT_FALSE(anyTrue(eq));
 428 }
 429
 430 TEST_F(SimdFloatingpointTest, blend)
 431 {
 432     SimdBool lt   = rSimd_c4c6c8 < rSimd_c6c7c8;
 433     GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c3, c4, c2), blend(rSimd_c0c1c2, rSimd_c3c4c5, lt));
 434 }
 435
 436 TEST_F(SimdFloatingpointTest, reduce)
 437 {
 438     // The horizontal sum of the SIMD variable depends on the width, so
 439     // simply store it an extra time and calculate what the sum should be
 440     std::vector<real> v   = simdReal2Vector(rSimd_c3c4c5);
 441     real              sum = 0.0;
 442
 443     for (int i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
 444     {
 445         sum += v[i];
 446     }
 447
 448     EXPECT_REAL_EQ_TOL(sum, reduce(rSimd_c3c4c5), defaultRealTolerance() );
 449 }
 450
 451 #endif      // GMX_SIMD_HAVE_REAL
 452
 453 #if GMX_SIMD_HAVE_FLOAT && GMX_SIMD_HAVE_DOUBLE
 454 TEST_F(SimdFloatingpointTest, cvtFloat2Double)
 455 {
 456     alignas(GMX_SIMD_ALIGNMENT) float   f[GMX_SIMD_FLOAT_WIDTH];
 457     alignas(GMX_SIMD_ALIGNMENT) double  d[GMX_SIMD_FLOAT_WIDTH];  // Yes, double array length should be same as float
 458
 459     int                               i;
 460     SimdFloat                         vf;
 461     SimdDouble                        vd0;
 462     FloatingPointTolerance            tolerance(defaultRealTolerance());
 463
 464     for (i = 0; i < GMX_SIMD_FLOAT_WIDTH; i++)
 465     {
 466         // Scale by 1+100*eps to use low bits too.
 467         // Due to the conversions we want to avoid being too sensitive to fluctuations in last bit
 468         f[i] = i * (1.0 + 100*GMX_FLOAT_EPS);
 469     }
 470
 471     vf = load<SimdFloat>(f);
 472 #if (GMX_SIMD_FLOAT_WIDTH == 2*GMX_SIMD_DOUBLE_WIDTH)
 473     SimdDouble vd1;
 474     cvtF2DD(vf, &vd0, &vd1);
 475     store(d + GMX_SIMD_DOUBLE_WIDTH, vd1); // Store upper part halfway through array
 476 #elif (GMX_SIMD_FLOAT_WIDTH == GMX_SIMD_DOUBLE_WIDTH)
 477     vd0 = cvtF2D(vf);
 478 #else
 479 #    error Width of float SIMD must either be identical to double, or twice the width.
 480 #endif
 481     store(d, vd0); // store lower (or whole) part from start of vector
 482
 483     for (i = 0; i < GMX_SIMD_FLOAT_WIDTH; i++)
 484     {
 485         EXPECT_REAL_EQ_TOL(f[i], d[i], tolerance);
 486     }
 487 }
 488
 489 TEST_F(SimdFloatingpointTest, cvtDouble2Float)
 490 {
 491     alignas(GMX_SIMD_ALIGNMENT) float     f[GMX_SIMD_FLOAT_WIDTH];
 492     alignas(GMX_SIMD_ALIGNMENT) double    d[GMX_SIMD_FLOAT_WIDTH];  // Yes, double array length should be same as float
 493     int                               i;
 494     SimdFloat                         vf;
 495     SimdDouble                        vd0;
 496     FloatingPointTolerance            tolerance(defaultRealTolerance());
 497
 498     // This fills elements for pd1 too when double width is 2*single width
 499     for (i = 0; i < GMX_SIMD_FLOAT_WIDTH; i++)
 500     {
 501         // Scale by 1+eps to use low bits too.
 502         // Due to the conversions we want to avoid being too sensitive to fluctuations in last bit
 503         d[i] = i * (1.0 + 100*GMX_FLOAT_EPS);
 504     }
 505
 506     vd0 = load<SimdDouble>(d);
 507 #if (GMX_SIMD_FLOAT_WIDTH == 2*GMX_SIMD_DOUBLE_WIDTH)
 508     SimdDouble vd1 = load<SimdDouble>(d + GMX_SIMD_DOUBLE_WIDTH); // load upper half of data
 509     vf = cvtDD2F(vd0, vd1);
 510 #elif (GMX_SIMD_FLOAT_WIDTH == GMX_SIMD_DOUBLE_WIDTH)
 511     vf = cvtD2F(vd0);
 512 #else
 513 #    error Width of float SIMD must either be identical to double, or twice the width.
 514 #endif
 515     store(f, vf);
 516
 517     // This will check elements in pd1 too when double width is 2*single width
 518     for (i = 0; i < GMX_SIMD_FLOAT_WIDTH; i++)
 519     {
 520         EXPECT_FLOAT_EQ_TOL(d[i], f[i], tolerance);
 521     }
 522 }
 523 #endif      // GMX_SIMD_HAVE_FLOAT && GMX_SIMD_HAVE_DOUBLE
 524
 525 /*! \} */
 526 /*! \endcond */
 527
 528 }      // namespace
 529 }      // namespace
 530 }      // namespace
 531
 532 #endif // GMX_SIMD