src/gromacs/simd/tests/simd4_floatingpoint.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2014,2015,2016,2017,2018 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 #include "gmxpre.h"
  37
  38 #include <cmath>
  39
  40 #include "gromacs/math/utilities.h"
  41 #include "gromacs/simd/simd.h"
  42 #include "gromacs/utility/basedefinitions.h"
  43
  44 #include "testutils/testasserts.h"
  45
  46 #include "data.h"
  47 #include "simd4.h"
  48
  49 #if GMX_SIMD
  50
  51 namespace gmx
  52 {
  53 namespace test
  54 {
  55 namespace
  56 {
  57
  58 /*! \cond internal */
  59 /*! \addtogroup module_simd */
  60 /*! \{ */
  61
  62 #    if GMX_SIMD4_HAVE_REAL
  63
  64 /*! \brief Test fixture for SIMD4 floating-point operations (identical to the SIMD4 \ref Simd4Test) */
  65 typedef Simd4Test Simd4FloatingpointTest;
  66
  67 TEST_F(Simd4FloatingpointTest, setZero)
  68 {
  69     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(0.0), setZero());
  70 }
  71
  72 TEST_F(Simd4FloatingpointTest, set)
  73 {
  74     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(c1), Simd4Real(c1));
  75 }
  76
  77 TEST_F(Simd4FloatingpointTest, add)
  78 {
  79     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0 + c3, c1 + c4, c2 + c5), rSimd4_c0c1c2 + rSimd4_c3c4c5);
  80 }
  81
  82 TEST_F(Simd4FloatingpointTest, sub)
  83 {
  84     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0 - c3, c1 - c4, c2 - c5), rSimd4_c0c1c2 - rSimd4_c3c4c5);
  85 }
  86
  87 TEST_F(Simd4FloatingpointTest, mul)
  88 {
  89     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0 * c3, c1 * c4, c2 * c5), rSimd4_c0c1c2 * rSimd4_c3c4c5);
  90 }
  91
  92 TEST_F(Simd4FloatingpointTest, fma)
  93 {
  94     // The last bit of FMA operations depends on hardware, so we don't require exact match
  95     GMX_EXPECT_SIMD4_REAL_NEAR(setSimd4RealFrom3R(c0 * c3 + c6, c1 * c4 + c7, c2 * c5 + c8),
  96                                fma(rSimd4_c0c1c2, rSimd4_c3c4c5, rSimd4_c6c7c8));
  97 }
  98
  99 TEST_F(Simd4FloatingpointTest, fms)
 100 {
 101     // The last bit of FMA operations depends on hardware, so we don't require exact match
 102     GMX_EXPECT_SIMD4_REAL_NEAR(setSimd4RealFrom3R(c0 * c3 - c6, c1 * c4 - c7, c2 * c5 - c8),
 103                                fms(rSimd4_c0c1c2, rSimd4_c3c4c5, rSimd4_c6c7c8));
 104 }
 105
 106 TEST_F(Simd4FloatingpointTest, fnma)
 107 {
 108     // The last bit of FMA operations depends on hardware, so we don't require exact match
 109     GMX_EXPECT_SIMD4_REAL_NEAR(setSimd4RealFrom3R(c6 - c0 * c3, c7 - c1 * c4, c8 - c2 * c5),
 110                                fnma(rSimd4_c0c1c2, rSimd4_c3c4c5, rSimd4_c6c7c8));
 111 }
 112
 113 TEST_F(Simd4FloatingpointTest, fnms)
 114 {
 115     // The last bit of FMA operations depends on hardware, so we don't require exact match
 116     GMX_EXPECT_SIMD4_REAL_NEAR(setSimd4RealFrom3R(-c0 * c3 - c6, -c1 * c4 - c7, -c2 * c5 - c8),
 117                                fnms(rSimd4_c0c1c2, rSimd4_c3c4c5, rSimd4_c6c7c8));
 118 }
 119
 120 TEST_F(Simd4FloatingpointTest, abs)
 121 {
 122     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_c0c1c2, abs(rSimd4_c0c1c2)); // fabs(x)=x
 123     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_c0c1c2, abs(rSimd4_m0m1m2)); // fabs(-x)=x
 124 }
 125
 126 TEST_F(Simd4FloatingpointTest, neg)
 127 {
 128     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_m0m1m2, -(rSimd4_c0c1c2)); // fneg(x)=-x
 129     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_c0c1c2, -(rSimd4_m0m1m2)); // fneg(-x)=x
 130 }
 131
 132 #        if GMX_SIMD_HAVE_LOGICAL
 133 TEST_F(Simd4FloatingpointTest, and)
 134 {
 135     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_logicalResultAnd, (rSimd4_logicalA & rSimd4_logicalB));
 136 }
 137
 138 TEST_F(Simd4FloatingpointTest, or)
 139 {
 140     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_logicalResultOr, (rSimd4_logicalA | rSimd4_logicalB));
 141 }
 142
 143 TEST_F(Simd4FloatingpointTest, xor)
 144 {
 145     /* Test xor by taking xor with a number and its negative. This should result
 146      * in only the sign bit being set. We then use this bit change the sign of
 147      * different numbers.
 148      */
 149     Simd4Real signbit = Simd4Real(c1) ^ Simd4Real(-c1);
 150     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-c2, c3, -c4), signbit ^ setSimd4RealFrom3R(c2, -c3, c4));
 151 }
 152
 153 TEST_F(Simd4FloatingpointTest, andNot)
 154 {
 155     /* Use xor (which we already tested, so fix that first if both tests fail)
 156      * to extract the sign bit, and then use andnot to take absolute values.
 157      */
 158     Simd4Real signbit = Simd4Real(c1) ^ Simd4Real(-c1);
 159     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c2, c3, c4),
 160                              andNot(signbit, setSimd4RealFrom3R(-c2, c3, -c4)));
 161 }
 162
 163 #        endif
 164
 165 TEST_F(Simd4FloatingpointTest, max)
 166 {
 167     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c3, c1, c4), max(rSimd4_c0c1c2, rSimd4_c3c0c4));
 168     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c3, c1, c4), max(rSimd4_c3c0c4, rSimd4_c0c1c2));
 169     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-c0, -c0, -c2), max(rSimd4_m0m1m2, rSimd4_m3m0m4));
 170     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-c0, -c0, -c2), max(rSimd4_m3m0m4, rSimd4_m0m1m2));
 171 }
 172
 173 TEST_F(Simd4FloatingpointTest, min)
 174 {
 175     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0, c0, c2), min(rSimd4_c0c1c2, rSimd4_c3c0c4));
 176     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0, c0, c2), min(rSimd4_c3c0c4, rSimd4_c0c1c2));
 177     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-c3, -c1, -c4), min(rSimd4_m0m1m2, rSimd4_m3m0m4));
 178     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-c3, -c1, -c4), min(rSimd4_m3m0m4, rSimd4_m0m1m2));
 179 }
 180
 181 TEST_F(Simd4FloatingpointTest, round)
 182 {
 183     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(2), round(Simd4Real(2.25)));
 184     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(4), round(Simd4Real(3.75)));
 185     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-2), round(Simd4Real(-2.25)));
 186     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-4), round(Simd4Real(-3.75)));
 187 }
 188
 189 TEST_F(Simd4FloatingpointTest, trunc)
 190 {
 191     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(2), trunc(rSimd4_2p25));
 192     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(3), trunc(rSimd4_3p75));
 193     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-2), trunc(rSimd4_m2p25));
 194     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-3), trunc(rSimd4_m3p75));
 195 }
 196
 197 /* We do extensive 1/sqrt(x) and 1/x accuracy testing in the tests for
 198  * the SIMD math functions, so we just make sure the lookup instructions
 199  * appear to work for a few values here.
 200  */
 201 TEST_F(Simd4FloatingpointTest, gmxSimd4RsqrtR)
 202 {
 203     Simd4Real x   = setSimd4RealFrom3R(4.0, M_PI, 1234567890.0);
 204     Simd4Real ref = setSimd4RealFrom3R(0.5, 1.0 / std::sqrt(M_PI), 1.0 / std::sqrt(1234567890.0));
 205     int       shiftbits = std::numeric_limits<real>::digits - GMX_SIMD_RSQRT_BITS;
 206
 207     if (shiftbits < 0)
 208     {
 209         shiftbits = 0;
 210     }
 211
 212     // The allowed Ulp deviation is 2 to the power of the number of mantissa
 213     // digits, minus the number of bits provided by the table lookup
 214     setUlpTol(1LL << shiftbits);
 215     GMX_EXPECT_SIMD4_REAL_NEAR(ref, rsqrt(x));
 216 }
 217
 218 TEST_F(Simd4FloatingpointTest, cmpEqAndSelectByMask)
 219 {
 220     Simd4Bool eq = rSimd4_c4c6c8 == rSimd4_c6c7c8;
 221     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(0, 0, c2), selectByMask(rSimd4_c0c1c2, eq));
 222 }
 223
 224 TEST_F(Simd4FloatingpointTest, selectByNotMask)
 225 {
 226     Simd4Bool eq = rSimd4_c4c6c8 == rSimd4_c6c7c8;
 227     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0, c1, 0), selectByNotMask(rSimd4_c0c1c2, eq));
 228 }
 229
 230 TEST_F(Simd4FloatingpointTest, cmpNe)
 231 {
 232     Simd4Bool eq = rSimd4_c4c6c8 != rSimd4_c6c7c8;
 233     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0, c1, 0), selectByMask(rSimd4_c0c1c2, eq));
 234 }
 235
 236 TEST_F(Simd4FloatingpointTest, cmpLe)
 237 {
 238     Simd4Bool le = rSimd4_c4c6c8 <= rSimd4_c6c7c8;
 239     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_c0c1c2, selectByMask(rSimd4_c0c1c2, le));
 240 }
 241
 242 TEST_F(Simd4FloatingpointTest, cmpLt)
 243 {
 244     Simd4Bool lt = rSimd4_c4c6c8 < rSimd4_c6c7c8;
 245     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0, c1, 0), selectByMask(rSimd4_c0c1c2, lt));
 246 }
 247
 248 TEST_F(Simd4FloatingpointTest, andB)
 249 {
 250     Simd4Bool eq = rSimd4_c4c6c8 == rSimd4_c6c7c8;
 251     Simd4Bool le = rSimd4_c4c6c8 <= rSimd4_c6c7c8;
 252     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(0, 0, c2), selectByMask(rSimd4_c0c1c2, (eq && le)));
 253 }
 254
 255 TEST_F(Simd4FloatingpointTest, orB)
 256 {
 257     Simd4Bool eq = rSimd4_c4c6c8 == rSimd4_c6c7c8;
 258     Simd4Bool lt = rSimd4_c4c6c8 < rSimd4_c6c7c8;
 259     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_c0c1c2, selectByMask(rSimd4_c0c1c2, (eq || lt)));
 260 }
 261
 262 TEST_F(Simd4FloatingpointTest, anyTrue)
 263 {
 264     Simd4Bool eq;
 265
 266     /* this test is a bit tricky since we don't know the simd width.
 267      * We cannot check for truth values for "any" element beyond the first,
 268      * since that part of the data will not be used if simd width is 1.
 269      */
 270     eq = (rSimd4_c4c6c8 == setSimd4RealFrom3R(c4, 0, 0));
 271     EXPECT_TRUE(anyTrue(eq));
 272
 273     eq = (rSimd4_c0c1c2 == rSimd4_c3c4c5);
 274     EXPECT_FALSE(anyTrue(eq));
 275 }
 276
 277 TEST_F(Simd4FloatingpointTest, blend)
 278 {
 279     Simd4Bool lt = rSimd4_c4c6c8 < rSimd4_c6c7c8;
 280     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c3, c4, c2), blend(rSimd4_c0c1c2, rSimd4_c3c4c5, lt));
 281 }
 282
 283 TEST_F(Simd4FloatingpointTest, reduce)
 284 {
 285     // The horizontal sum of the SIMD variable depends on the width, so
 286     // simply store it an extra time and calculate what the sum should be
 287     std::vector<real> v   = simd4Real2Vector(rSimd4_c3c4c5);
 288     real              sum = 0.0;
 289
 290     for (int i = 0; i < GMX_SIMD4_WIDTH; i++)
 291     {
 292         sum += v[i];
 293     }
 294
 295     EXPECT_REAL_EQ_TOL(sum, reduce(rSimd4_c3c4c5), defaultRealTolerance());
 296 }
 297
 298
 299 TEST_F(Simd4FloatingpointTest, dotProduct)
 300 {
 301     real res = c0 * c3 + c1 * c4 + c2 * c5;
 302
 303     EXPECT_REAL_EQ_TOL(res, dotProduct(rSimd4_c0c1c2, rSimd4_c3c4c5), defaultRealTolerance());
 304 }
 305
 306 TEST_F(Simd4FloatingpointTest, transpose)
 307 {
 308     Simd4Real v0, v1, v2, v3;
 309     int       i;
 310     // aligned pointers
 311     alignas(GMX_SIMD_ALIGNMENT) real p0[4 * GMX_SIMD4_WIDTH];
 312     real*                            p1 = p0 + GMX_SIMD4_WIDTH;
 313     real*                            p2 = p0 + 2 * GMX_SIMD4_WIDTH;
 314     real*                            p3 = p0 + 3 * GMX_SIMD4_WIDTH;
 315
 316     // Assign data with tens as row, single-digit as column
 317     for (i = 0; i < 4; i++)
 318     {
 319         // Scale by 1+100*eps to use low bits tii
 320         p0[i] = (0 * 10 + i * 1) * (1.0 + 100 * GMX_REAL_EPS);
 321         p1[i] = (1 * 10 + i * 1) * (1.0 + 100 * GMX_REAL_EPS);
 322         p2[i] = (2 * 10 + i * 1) * (1.0 + 100 * GMX_REAL_EPS);
 323         p3[i] = (3 * 10 + i * 1) * (1.0 + 100 * GMX_REAL_EPS);
 324     }
 325
 326     v0 = load4(p0);
 327     v1 = load4(p1);
 328     v2 = load4(p2);
 329     v3 = load4(p3);
 330
 331     transpose(&v0, &v1, &v2, &v3);
 332
 333     store4(p0, v0);
 334     store4(p1, v1);
 335     store4(p2, v2);
 336     store4(p3, v3);
 337
 338     for (i = 0; i < 4; i++)
 339     {
 340         EXPECT_REAL_EQ_TOL((i * 10 + 0) * (1.0 + 100 * GMX_REAL_EPS), p0[i], defaultRealTolerance());
 341         EXPECT_REAL_EQ_TOL((i * 10 + 1) * (1.0 + 100 * GMX_REAL_EPS), p1[i], defaultRealTolerance());
 342         EXPECT_REAL_EQ_TOL((i * 10 + 2) * (1.0 + 100 * GMX_REAL_EPS), p2[i], defaultRealTolerance());
 343         EXPECT_REAL_EQ_TOL((i * 10 + 3) * (1.0 + 100 * GMX_REAL_EPS), p3[i], defaultRealTolerance());
 344     }
 345 }
 346
 347 #    endif // GMX_SIMD4_HAVE_REAL
 348
 349 /*! \} */
 350 /*! \endcond */
 351
 352 } // namespace
 353 } // namespace test
 354 } // namespace gmx
 355
 356 #endif // GMX_SIMD