src/gromacs/simd/tests/simd4_floatingpoint.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2014, 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 <math.h>
  38 #include "gromacs/math/utilities.h"
  39
  40 #include "simd4.h"
  41
  42 namespace gmx
  43 {
  44 namespace test
  45 {
  46 namespace
  47 {
  48
  49 /*! \cond internal */
  50 /*! \addtogroup module_simd */
  51 /*! \{ */
  52
  53 #ifdef GMX_SIMD4_HAVE_REAL
  54
  55 /*! \brief Test fixture for SIMD4 floating-point operations (identical to the SIMD4 \ref Simd4Test) */
  56 typedef Simd4Test Simd4FloatingpointTest;
  57
  58 TEST_F(Simd4FloatingpointTest, gmxSimd4SetZeroR)
  59 {
  60     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(0.0), gmx_simd4_setzero_r());
  61 }
  62
  63 TEST_F(Simd4FloatingpointTest, gmxSimd4Set1R)
  64 {
  65     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(1.0), gmx_simd4_set1_r(1.0));
  66 }
  67
  68 TEST_F(Simd4FloatingpointTest, gmxSimd4Load1R)
  69 {
  70     real r = 2.0;
  71     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(r), gmx_simd4_load1_r(&r));
  72 }
  73
  74 TEST_F(Simd4FloatingpointTest, gmxSimd4AddR)
  75 {
  76     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_5_7_9, gmx_simd4_add_r(rSimd4_1_2_3, rSimd4_4_5_6)); // 1+4=5, 2+5=7, 3+6=9
  77 }
  78
  79 TEST_F(Simd4FloatingpointTest, gmxSimd4SubR)
  80 {
  81     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_4_5_6, gmx_simd4_sub_r(rSimd4_5_7_9, rSimd4_1_2_3)); // 5-1=4, 7-2=5, 9-3=6
  82 }
  83
  84 TEST_F(Simd4FloatingpointTest, gmxSimd4MulR)
  85 {
  86     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(4, 10, 18), gmx_simd4_mul_r(rSimd4_1_2_3, rSimd4_4_5_6));
  87 }
  88
  89 TEST_F(Simd4FloatingpointTest, gmxSimd4FmaddR)
  90 {
  91     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(11, 18, 27), gmx_simd4_fmadd_r(rSimd4_1_2_3, rSimd4_4_5_6, rSimd4_7_8_9)); // 1*4+7, etc.
  92 }
  93
  94 TEST_F(Simd4FloatingpointTest, gmxSimd4FmsubR)
  95 {
  96     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-3, 2, 9), gmx_simd4_fmsub_r(rSimd4_1_2_3, rSimd4_4_5_6, rSimd4_7_8_9)); // 1*4-7, etc.
  97 }
  98
  99 TEST_F(Simd4FloatingpointTest, gmxSimd4FnmaddR)
 100 {
 101     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(3, -2, -9), gmx_simd4_fnmadd_r(rSimd4_1_2_3, rSimd4_4_5_6, rSimd4_7_8_9)); // -1*4+7, etc.
 102 }
 103
 104 TEST_F(Simd4FloatingpointTest, gmxSimd4FnmsubR)
 105 {
 106     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-11, -18, -27), gmx_simd4_fnmsub_r(rSimd4_1_2_3, rSimd4_4_5_6, rSimd4_7_8_9)); // -1*4-7, etc.
 107 }
 108
 109 TEST_F(Simd4FloatingpointTest, gmxSimd4FabsR)
 110 {
 111     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_1_2_3, gmx_simd4_fabs_r(rSimd4_1_2_3));    // fabs(x)=x
 112     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_1_2_3, gmx_simd4_fabs_r(rSimd4_m1_m2_m3)); // fabs(-x)=x
 113 }
 114
 115 TEST_F(Simd4FloatingpointTest, gmxSimd4FnegR)
 116 {
 117     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_m1_m2_m3, gmx_simd4_fneg_r(rSimd4_1_2_3));   // fneg(x)=-x
 118     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_1_2_3,   gmx_simd4_fneg_r(rSimd4_m1_m2_m3)); // fneg(-x)=x
 119 }
 120
 121 #ifdef GMX_SIMD4_HAVE_LOGICAL
 122 /* 1.3333282470703125 has mantissa 0101010101010101 (followed by zeros)
 123  * 1.79998779296875   has mantissa 1100110011001100 (followed by zeros)
 124  * 1.26666259765625   has mantissa 0100010001000100 (followed by zeros)
 125  * 1.8666534423828125 has mantissa 1101110111011101 (followed by zeros)
 126  *
 127  * Since all of them have the same exponent (2^0), the exponent will
 128  * not change with AND or OR operations.
 129  */
 130 TEST_F(Simd4FloatingpointTest, gmxSimd4AndR)
 131 {
 132     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(1.26666259765625),
 133                              gmx_simd4_and_r(gmx_simd4_set1_r(1.3333282470703125),
 134                                              gmx_simd4_set1_r(1.79998779296875)));
 135 }
 136
 137 TEST_F(Simd4FloatingpointTest, gmxSimd4OrR)
 138 {
 139     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(1.8666534423828125),
 140                              gmx_simd4_or_r(gmx_simd4_set1_r(1.3333282470703125),
 141                                             gmx_simd4_set1_r(1.79998779296875)));
 142 }
 143
 144 TEST_F(Simd4FloatingpointTest, gmxSimd4XorR)
 145 {
 146     /* Test xor by taking xor with a number and its negative. This should result
 147      * in only the sign bit being set. We then use this bit change the sign of
 148      * different numbers.
 149      */
 150     gmx_simd4_real_t signbit = gmx_simd4_xor_r(gmx_simd4_set1_r(1.5), gmx_simd4_set1_r(-1.5));
 151     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-1, 2, -3), gmx_simd4_xor_r(signbit, setSimd4RealFrom3R(1, -2, 3)));
 152 }
 153
 154 TEST_F(Simd4FloatingpointTest, gmxSimd4AndnotR)
 155 {
 156     /* Use xor (which we already tested, so fix that first if both tests fail)
 157      * to extract the sign bit, and then use andnot to take absolute values.
 158      */
 159     gmx_simd4_real_t signbit = gmx_simd4_xor_r(gmx_simd4_set1_r(1.5), gmx_simd4_set1_r(-1.5));
 160     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(1, 2, 3), gmx_simd4_andnot_r(signbit, setSimd4RealFrom3R(-1, 2, -3)));
 161 }
 162
 163 #endif
 164
 165 TEST_F(Simd4FloatingpointTest, gmxSimd4MaxR)
 166 {
 167     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(3, 2, 4), gmx_simd4_max_r(rSimd4_1_2_3, rSimd4_3_1_4));
 168     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(3, 2, 4), gmx_simd4_max_r(rSimd4_3_1_4, rSimd4_1_2_3));
 169     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-1, -1, -3), gmx_simd4_max_r(rSimd4_m1_m2_m3, rSimd4_m3_m1_m4));
 170     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-1, -1, -3), gmx_simd4_max_r(rSimd4_m3_m1_m4, rSimd4_m1_m2_m3));
 171 }
 172
 173 TEST_F(Simd4FloatingpointTest, gmxSimd4MinR)
 174 {
 175     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(1, 1, 3), gmx_simd4_min_r(rSimd4_1_2_3, rSimd4_3_1_4));
 176     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(1, 1, 3), gmx_simd4_min_r(rSimd4_3_1_4, rSimd4_1_2_3));
 177     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-3, -2, -4), gmx_simd4_min_r(rSimd4_m1_m2_m3, rSimd4_m3_m1_m4));
 178     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-3, -2, -4), gmx_simd4_min_r(rSimd4_m3_m1_m4, rSimd4_m1_m2_m3));
 179 }
 180
 181 TEST_F(Simd4FloatingpointTest, gmxSimd4RoundR)
 182 {
 183     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(2), gmx_simd4_round_r(gmx_simd4_set1_r(2.25)));
 184     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(4), gmx_simd4_round_r(gmx_simd4_set1_r(3.75)));
 185     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-2), gmx_simd4_round_r(gmx_simd4_set1_r(-2.25)));
 186     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-4), gmx_simd4_round_r(gmx_simd4_set1_r(-3.75)));
 187 }
 188
 189 TEST_F(Simd4FloatingpointTest, gmxSimd4TruncR)
 190 {
 191     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(2), gmx_simd4_trunc_r(rSimd4_2p25));
 192     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(3), gmx_simd4_trunc_r(rSimd4_3p75));
 193     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-2), gmx_simd4_trunc_r(rSimd4_m2p25));
 194     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-3), gmx_simd4_trunc_r(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     gmx_simd4_real_t x      = setSimd4RealFrom3R(4.0, M_PI, 1234567890.0);
 204     gmx_simd4_real_t ref    = setSimd4RealFrom3R(0.5, 1.0/sqrt(M_PI), 1.0/sqrt(1234567890.0));
 205
 206     // The allowed Ulp deviation is 2 to the power of the number of mantissa
 207     // digits, minus the number of bits provided by the table lookup
 208     setUlpTol(1LL << (std::numeric_limits<real>::digits-GMX_SIMD_RSQRT_BITS));
 209     GMX_EXPECT_SIMD4_REAL_NEAR(ref, gmx_simd4_rsqrt_r(x));
 210 }
 211
 212 TEST_F(Simd4FloatingpointTest, gmxSimd4BoolCmpEqAndBlendZeroR)
 213 {
 214     gmx_simd4_bool_t eq   = gmx_simd4_cmpeq_r(rSimd4_5_7_9, rSimd4_7_8_9);
 215     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(0, 0, 3), gmx_simd4_blendzero_r(rSimd4_1_2_3, eq));
 216 }
 217
 218 TEST_F(Simd4FloatingpointTest, gmxSimd4BlendNotZeroR)
 219 {
 220     gmx_simd4_bool_t eq   = gmx_simd4_cmpeq_r(rSimd4_5_7_9, rSimd4_7_8_9);
 221     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(1, 2, 0), gmx_simd4_blendnotzero_r(rSimd4_1_2_3, eq));
 222 }
 223
 224 TEST_F(Simd4FloatingpointTest, gmxSimd4BoolCmpLER)
 225 {
 226     gmx_simd4_bool_t le   = gmx_simd4_cmple_r(rSimd4_5_7_9, rSimd4_7_8_9);
 227     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_1_2_3, gmx_simd4_blendzero_r(rSimd4_1_2_3, le));
 228 }
 229
 230 TEST_F(Simd4FloatingpointTest, gmxSimd4BoolCmpLTR)
 231 {
 232     gmx_simd4_bool_t lt   = gmx_simd4_cmplt_r(rSimd4_5_7_9, rSimd4_7_8_9);
 233     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(1, 2, 0), gmx_simd4_blendzero_r(rSimd4_1_2_3, lt));
 234 }
 235
 236 TEST_F(Simd4FloatingpointTest, gmxSimd4BoolAndB)
 237 {
 238     gmx_simd4_bool_t eq   = gmx_simd4_cmpeq_r(rSimd4_5_7_9, rSimd4_7_8_9);
 239     gmx_simd4_bool_t le   = gmx_simd4_cmple_r(rSimd4_5_7_9, rSimd4_7_8_9);
 240     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(0, 0, 3), gmx_simd4_blendzero_r(rSimd4_1_2_3, gmx_simd4_and_b(eq, le)));
 241 }
 242
 243 TEST_F(Simd4FloatingpointTest, gmxSimd4BoolOrB)
 244 {
 245     gmx_simd4_bool_t eq   = gmx_simd4_cmpeq_r(rSimd4_5_7_9, rSimd4_7_8_9);
 246     gmx_simd4_bool_t lt   = gmx_simd4_cmplt_r(rSimd4_5_7_9, rSimd4_7_8_9);
 247     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(1, 2, 3), gmx_simd4_blendzero_r(rSimd4_1_2_3, gmx_simd4_or_b(eq, lt)));
 248 }
 249
 250 TEST_F(Simd4FloatingpointTest, gmxSimd4AnytrueB)
 251 {
 252     gmx_simd4_bool_t eq;
 253
 254     /* this test is a bit tricky since we don't know the simd width.
 255      * We cannot check for truth values for "any" element beyond the first,
 256      * since that part of the data will not be used if simd width is 1.
 257      */
 258     eq = gmx_simd4_cmpeq_r(rSimd4_5_7_9, setSimd4RealFrom3R(5, 0, 0));
 259     EXPECT_NE(0, gmx_simd4_anytrue_b(eq));
 260
 261     eq = gmx_simd4_cmpeq_r(rSimd4_1_2_3, rSimd4_4_5_6);
 262     EXPECT_EQ(0, gmx_simd4_anytrue_b(eq));
 263 }
 264
 265 TEST_F(Simd4FloatingpointTest, gmxSimd4BlendvR)
 266 {
 267     gmx_simd4_bool_t lt   = gmx_simd4_cmplt_r(rSimd4_5_7_9, rSimd4_7_8_9);
 268     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(4, 5, 3), gmx_simd4_blendv_r(rSimd4_1_2_3, rSimd4_4_5_6, lt));
 269 }
 270
 271 TEST_F(Simd4FloatingpointTest, gmxSimd4ReduceR)
 272 {
 273     // The horizontal sum of the SIMD variable depends on the width, so
 274     // simply store it an extra time and calculate what the sum should be
 275     std::vector<real> v   = simd4Real2Vector(rSimd4_1_2_3);
 276     real              sum = 0.0;
 277
 278     for (int i = 0; i < GMX_SIMD4_WIDTH; i++)
 279     {
 280         sum += v[i];
 281     }
 282
 283     EXPECT_EQ(sum, gmx_simd4_reduce_r(rSimd4_1_2_3));
 284 }
 285
 286
 287 TEST_F(Simd4FloatingpointTest, gmxSimd4Dotproduct3R)
 288 {
 289     gmx_simd4_real_t v1 = setSimd4RealFrom3R(1, 4, 5);
 290     gmx_simd4_real_t v2 = setSimd4RealFrom3R(3, 8, 2);
 291 #    ifdef GMX_DOUBLE
 292     EXPECT_DOUBLE_EQ(45.0, gmx_simd4_dotproduct3_r(v1, v2));
 293 #    else
 294     EXPECT_FLOAT_EQ(45.0, gmx_simd4_dotproduct3_r(v1, v2));
 295 #    endif
 296 }
 297
 298 #endif      // GMX_SIMD4_HAVE_REAL
 299
 300 /*! \} */
 301 /*! \endcond */
 302
 303 }      // namespace
 304 }      // namespace
 305 }      // namespace