src/gromacs/simd/tests/simd4_floatingpoint.cpp

   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 #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 "simd4.h"
  47
  48 #if GMX_SIMD
  49
  50 namespace gmx
  51 {
  52 namespace test
  53 {
  54 namespace
  55 {
  56
  57 /*! \cond internal */
  58 /*! \addtogroup module_simd */
  59 /*! \{ */
  60
  61 #if GMX_SIMD4_HAVE_REAL
  62
  63 /*! \brief Test fixture for SIMD4 floating-point operations (identical to the SIMD4 \ref Simd4Test) */
  64 typedef Simd4Test Simd4FloatingpointTest;
  65
  66 TEST_F(Simd4FloatingpointTest, setZero)
  67 {
  68     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(0.0), setZero());
  69 }
  70
  71 TEST_F(Simd4FloatingpointTest, set)
  72 {
  73     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(c1), Simd4Real(c1));
  74 }
  75
  76 TEST_F(Simd4FloatingpointTest, add)
  77 {
  78     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0 + c3, c1 + c4, c2 + c5 ),
  79                              rSimd4_c0c1c2 + rSimd4_c3c4c5);
  80
  81 }
  82
  83 TEST_F(Simd4FloatingpointTest, sub)
  84 {
  85     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0 - c3, c1 - c4, c2 - c5 ),
  86                              rSimd4_c0c1c2 - rSimd4_c3c4c5);
  87 }
  88
  89 TEST_F(Simd4FloatingpointTest, mul)
  90 {
  91     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0 * c3, c1 * c4, c2 * c5 ),
  92                              rSimd4_c0c1c2 * rSimd4_c3c4c5);
  93 }
  94
  95 TEST_F(Simd4FloatingpointTest, fma)
  96 {
  97     // The last bit of FMA operations depends on hardware, so we don't require exact match
  98     GMX_EXPECT_SIMD4_REAL_NEAR(setSimd4RealFrom3R(c0 * c3 + c6, c1 * c4 + c7, c2 * c5 + c8),
  99                                fma(rSimd4_c0c1c2, rSimd4_c3c4c5, rSimd4_c6c7c8));
 100 }
 101
 102 TEST_F(Simd4FloatingpointTest, fms)
 103 {
 104     // The last bit of FMA operations depends on hardware, so we don't require exact match
 105     GMX_EXPECT_SIMD4_REAL_NEAR(setSimd4RealFrom3R(c0 * c3 - c6, c1 * c4 - c7, c2 * c5 - c8),
 106                                fms(rSimd4_c0c1c2, rSimd4_c3c4c5, rSimd4_c6c7c8));
 107 }
 108
 109 TEST_F(Simd4FloatingpointTest, fnma)
 110 {
 111     // The last bit of FMA operations depends on hardware, so we don't require exact match
 112     GMX_EXPECT_SIMD4_REAL_NEAR(setSimd4RealFrom3R(c6 - c0 * c3, c7 - c1 * c4, c8 - c2 * c5),
 113                                fnma(rSimd4_c0c1c2, rSimd4_c3c4c5, rSimd4_c6c7c8));
 114 }
 115
 116 TEST_F(Simd4FloatingpointTest, fnms)
 117 {
 118     // The last bit of FMA operations depends on hardware, so we don't require exact match
 119     GMX_EXPECT_SIMD4_REAL_NEAR(setSimd4RealFrom3R(-c0 * c3 - c6, -c1 * c4 - c7, -c2 * c5 - c8),
 120                                fnms(rSimd4_c0c1c2, rSimd4_c3c4c5, rSimd4_c6c7c8));
 121 }
 122
 123 TEST_F(Simd4FloatingpointTest, abs)
 124 {
 125     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_c0c1c2, abs(rSimd4_c0c1c2)); // fabs(x)=x
 126     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_c0c1c2, abs(rSimd4_m0m1m2)); // fabs(-x)=x
 127 }
 128
 129 TEST_F(Simd4FloatingpointTest, neg)
 130 {
 131     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_m0m1m2, -(rSimd4_c0c1c2)); // fneg(x)=-x
 132     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_c0c1c2, -(rSimd4_m0m1m2)); // fneg(-x)=x
 133 }
 134
 135 #if GMX_SIMD_HAVE_LOGICAL
 136 TEST_F(Simd4FloatingpointTest, and)
 137 {
 138     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_logicalResultAnd,
 139                              (rSimd4_logicalA & rSimd4_logicalB));
 140 }
 141
 142 TEST_F(Simd4FloatingpointTest, or)
 143 {
 144     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_logicalResultOr,
 145                              (rSimd4_logicalA | rSimd4_logicalB));
 146 }
 147
 148 TEST_F(Simd4FloatingpointTest, xor)
 149 {
 150     /* Test xor by taking xor with a number and its negative. This should result
 151      * in only the sign bit being set. We then use this bit change the sign of
 152      * different numbers.
 153      */
 154     Simd4Real signbit = Simd4Real(c1) ^ Simd4Real(-c1);
 155     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-c2, c3, -c4), signbit ^ setSimd4RealFrom3R(c2, -c3, c4));
 156 }
 157
 158 TEST_F(Simd4FloatingpointTest, andNot)
 159 {
 160     /* Use xor (which we already tested, so fix that first if both tests fail)
 161      * to extract the sign bit, and then use andnot to take absolute values.
 162      */
 163     Simd4Real signbit = Simd4Real(c1) ^ Simd4Real(-c1);
 164     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c2, c3, c4), andNot(signbit, setSimd4RealFrom3R(-c2, c3, -c4)));
 165 }
 166
 167 #endif
 168
 169 TEST_F(Simd4FloatingpointTest, max)
 170 {
 171     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R( c3,  c1,  c4), max(rSimd4_c0c1c2, rSimd4_c3c0c4));
 172     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R( c3,  c1,  c4), max(rSimd4_c3c0c4, rSimd4_c0c1c2));
 173     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-c0, -c0, -c2), max(rSimd4_m0m1m2, rSimd4_m3m0m4));
 174     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-c0, -c0, -c2), max(rSimd4_m3m0m4, rSimd4_m0m1m2));
 175 }
 176
 177 TEST_F(Simd4FloatingpointTest, min)
 178 {
 179     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R( c0,  c0,  c2), min(rSimd4_c0c1c2, rSimd4_c3c0c4));
 180     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R( c0,  c0,  c2), min(rSimd4_c3c0c4, rSimd4_c0c1c2));
 181     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-c3, -c1, -c4), min(rSimd4_m0m1m2, rSimd4_m3m0m4));
 182     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-c3, -c1, -c4), min(rSimd4_m3m0m4, rSimd4_m0m1m2));
 183 }
 184
 185 TEST_F(Simd4FloatingpointTest, round)
 186 {
 187     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(2), round(Simd4Real(2.25)));
 188     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(4), round(Simd4Real(3.75)));
 189     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-2), round(Simd4Real(-2.25)));
 190     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-4), round(Simd4Real(-3.75)));
 191 }
 192
 193 TEST_F(Simd4FloatingpointTest, trunc)
 194 {
 195     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(2), trunc(rSimd4_2p25));
 196     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(3), trunc(rSimd4_3p75));
 197     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-2), trunc(rSimd4_m2p25));
 198     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-3), trunc(rSimd4_m3p75));
 199 }
 200
 201 /* We do extensive 1/sqrt(x) and 1/x accuracy testing in the tests for
 202  * the SIMD math functions, so we just make sure the lookup instructions
 203  * appear to work for a few values here.
 204  */
 205 TEST_F(Simd4FloatingpointTest, gmxSimd4RsqrtR)
 206 {
 207     Simd4Real        x                   = setSimd4RealFrom3R(4.0, M_PI, 1234567890.0);
 208     Simd4Real        ref                 = setSimd4RealFrom3R(0.5, 1.0/std::sqrt(M_PI), 1.0/std::sqrt(1234567890.0));
 209     int              shiftbits           = std::numeric_limits<real>::digits-GMX_SIMD_RSQRT_BITS;
 210
 211     if (shiftbits < 0)
 212     {
 213         shiftbits = 0;
 214     }
 215
 216     // The allowed Ulp deviation is 2 to the power of the number of mantissa
 217     // digits, minus the number of bits provided by the table lookup
 218     setUlpTol(1LL << shiftbits);
 219     GMX_EXPECT_SIMD4_REAL_NEAR(ref, rsqrt(x));
 220 }
 221
 222 TEST_F(Simd4FloatingpointTest, cmpEqAndSelectByMask)
 223 {
 224     Simd4Bool eq   = rSimd4_c4c6c8 == rSimd4_c6c7c8;
 225     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(0, 0, c2), selectByMask(rSimd4_c0c1c2, eq));
 226 }
 227
 228 TEST_F(Simd4FloatingpointTest, selectByNotMask)
 229 {
 230     Simd4Bool eq   = rSimd4_c4c6c8 == rSimd4_c6c7c8;
 231     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0, c1, 0), selectByNotMask(rSimd4_c0c1c2, eq));
 232 }
 233
 234 TEST_F(Simd4FloatingpointTest, cmpNe)
 235 {
 236     Simd4Bool eq   = rSimd4_c4c6c8 != rSimd4_c6c7c8;
 237     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0, c1, 0), selectByMask(rSimd4_c0c1c2, eq));
 238 }
 239
 240 TEST_F(Simd4FloatingpointTest, cmpLe)
 241 {
 242     Simd4Bool le   = rSimd4_c4c6c8 <= rSimd4_c6c7c8;
 243     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_c0c1c2, selectByMask(rSimd4_c0c1c2, le));
 244 }
 245
 246 TEST_F(Simd4FloatingpointTest, cmpLt)
 247 {
 248     Simd4Bool lt   = rSimd4_c4c6c8 < rSimd4_c6c7c8;
 249     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0, c1, 0), selectByMask(rSimd4_c0c1c2, lt));
 250 }
 251
 252 TEST_F(Simd4FloatingpointTest, andB)
 253 {
 254     Simd4Bool eq   = rSimd4_c4c6c8 == rSimd4_c6c7c8;
 255     Simd4Bool le   = rSimd4_c4c6c8 <= rSimd4_c6c7c8;
 256     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(0, 0, c2), selectByMask(rSimd4_c0c1c2, (eq && le)));
 257 }
 258
 259 TEST_F(Simd4FloatingpointTest, orB)
 260 {
 261     Simd4Bool eq   = rSimd4_c4c6c8 == rSimd4_c6c7c8;
 262     Simd4Bool lt   = rSimd4_c4c6c8  < rSimd4_c6c7c8;
 263     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_c0c1c2, selectByMask(rSimd4_c0c1c2, (eq || lt)));
 264 }
 265
 266 TEST_F(Simd4FloatingpointTest, anyTrue)
 267 {
 268     Simd4Bool eq;
 269
 270     /* this test is a bit tricky since we don't know the simd width.
 271      * We cannot check for truth values for "any" element beyond the first,
 272      * since that part of the data will not be used if simd width is 1.
 273      */
 274     eq = (rSimd4_c4c6c8 == setSimd4RealFrom3R(c4, 0, 0));
 275     EXPECT_TRUE(anyTrue(eq));
 276
 277     eq = (rSimd4_c0c1c2 == rSimd4_c3c4c5);
 278     EXPECT_FALSE(anyTrue(eq));
 279 }
 280
 281 TEST_F(Simd4FloatingpointTest, blend)
 282 {
 283     Simd4Bool lt   = rSimd4_c4c6c8 < rSimd4_c6c7c8;
 284     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c3, c4, c2), blend(rSimd4_c0c1c2, rSimd4_c3c4c5, lt));
 285 }
 286
 287 TEST_F(Simd4FloatingpointTest, reduce)
 288 {
 289     // The horizontal sum of the SIMD variable depends on the width, so
 290     // simply store it an extra time and calculate what the sum should be
 291     std::vector<real> v   = simd4Real2Vector(rSimd4_c3c4c5);
 292     real              sum = 0.0;
 293
 294     for (int i = 0; i < GMX_SIMD4_WIDTH; i++)
 295     {
 296         sum += v[i];
 297     }
 298
 299     EXPECT_REAL_EQ_TOL(sum, reduce(rSimd4_c3c4c5), defaultRealTolerance() );
 300 }
 301
 302
 303 TEST_F(Simd4FloatingpointTest, dotProduct)
 304 {
 305     real res = c0*c3 + c1*c4 + c2*c5;
 306
 307     EXPECT_REAL_EQ_TOL(res, dotProduct(rSimd4_c0c1c2, rSimd4_c3c4c5), defaultRealTolerance());
 308 }
 309
 310 TEST_F(Simd4FloatingpointTest, transpose)
 311 {
 312     Simd4Real        v0, v1, v2, v3;
 313     int              i;
 314     // aligned pointers
 315     GMX_ALIGNED(real, GMX_SIMD4_WIDTH) p0[4*GMX_SIMD4_WIDTH];
 316     real          *  p1 = p0 + GMX_SIMD4_WIDTH;
 317     real          *  p2 = p0 + 2*GMX_SIMD4_WIDTH;
 318     real          *  p3 = p0 + 3*GMX_SIMD4_WIDTH;
 319
 320     // Assign data with tens as row, single-digit as column
 321     for (i = 0; i < 4; i++)
 322     {
 323         // Scale by 1+100*eps to use low bits tii
 324         p0[i] = (0*10 + i*1) * (1.0 + 100*GMX_REAL_EPS);
 325         p1[i] = (1*10 + i*1) * (1.0 + 100*GMX_REAL_EPS);
 326         p2[i] = (2*10 + i*1) * (1.0 + 100*GMX_REAL_EPS);
 327         p3[i] = (3*10 + i*1) * (1.0 + 100*GMX_REAL_EPS);
 328     }
 329
 330     v0 = load4(p0);
 331     v1 = load4(p1);
 332     v2 = load4(p2);
 333     v3 = load4(p3);
 334
 335     transpose(&v0, &v1, &v2, &v3);
 336
 337     store4(p0, v0);
 338     store4(p1, v1);
 339     store4(p2, v2);
 340     store4(p3, v3);
 341
 342     for (i = 0; i < 4; i++)
 343     {
 344         EXPECT_REAL_EQ_TOL( (i*10+0) * (1.0 + 100*GMX_REAL_EPS), p0[i], defaultRealTolerance());
 345         EXPECT_REAL_EQ_TOL( (i*10+1) * (1.0 + 100*GMX_REAL_EPS), p1[i], defaultRealTolerance());
 346         EXPECT_REAL_EQ_TOL( (i*10+2) * (1.0 + 100*GMX_REAL_EPS), p2[i], defaultRealTolerance());
 347         EXPECT_REAL_EQ_TOL( (i*10+3) * (1.0 + 100*GMX_REAL_EPS), p3[i], defaultRealTolerance());
 348     }
 349 }
 350
 351 #endif      // GMX_SIMD4_HAVE_REAL
 352
 353 /*! \} */
 354 /*! \endcond */
 355
 356 }      // namespace
 357 }      // namespace
 358 }      // namespace
 359
 360 #endif // GMX_SIMD