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 "config.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 TEST_F(Simd4FloatingpointTest, gmxSimd4AndR)
 123 {
 124     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_Bits3, gmx_simd4_and_r(rSimd4_Bits1, rSimd4_Bits2)); // Bits1 & Bits2 = Bits3
 125 }
 126
 127 TEST_F(Simd4FloatingpointTest, gmxSimd4AndnotR)
 128 {
 129     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_Bits4, gmx_simd4_andnot_r(rSimd4_Bits1, rSimd4_Bits2)); // (~Bits1) & Bits2 = Bits3
 130 }
 131
 132 TEST_F(Simd4FloatingpointTest, gmxSimd4OrR)
 133 {
 134     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_Bits5, gmx_simd4_or_r(rSimd4_Bits1, rSimd4_Bits2)); // Bits1 | Bits2 = Bits3
 135 }
 136
 137 TEST_F(Simd4FloatingpointTest, gmxSimd4XorR)
 138 {
 139     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_Bits6, gmx_simd4_xor_r(rSimd4_Bits1, rSimd4_Bits2)); // Bits1 ^ Bits2 = Bits3
 140 }
 141 #endif
 142
 143 TEST_F(Simd4FloatingpointTest, gmxSimd4MaxR)
 144 {
 145     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(3, 2, 4), gmx_simd4_max_r(rSimd4_1_2_3, rSimd4_3_1_4));
 146     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(3, 2, 4), gmx_simd4_max_r(rSimd4_3_1_4, rSimd4_1_2_3));
 147     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-1, -1, -3), gmx_simd4_max_r(rSimd4_m1_m2_m3, rSimd4_m3_m1_m4));
 148     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-1, -1, -3), gmx_simd4_max_r(rSimd4_m3_m1_m4, rSimd4_m1_m2_m3));
 149 }
 150
 151 TEST_F(Simd4FloatingpointTest, gmxSimd4MinR)
 152 {
 153     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(1, 1, 3), gmx_simd4_min_r(rSimd4_1_2_3, rSimd4_3_1_4));
 154     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(1, 1, 3), gmx_simd4_min_r(rSimd4_3_1_4, rSimd4_1_2_3));
 155     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-3, -2, -4), gmx_simd4_min_r(rSimd4_m1_m2_m3, rSimd4_m3_m1_m4));
 156     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-3, -2, -4), gmx_simd4_min_r(rSimd4_m3_m1_m4, rSimd4_m1_m2_m3));
 157 }
 158
 159 TEST_F(Simd4FloatingpointTest, gmxSimd4RoundR)
 160 {
 161     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(2), gmx_simd4_round_r(gmx_simd4_set1_r(2.25)));
 162     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(4), gmx_simd4_round_r(gmx_simd4_set1_r(3.75)));
 163     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-2), gmx_simd4_round_r(gmx_simd4_set1_r(-2.25)));
 164     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-4), gmx_simd4_round_r(gmx_simd4_set1_r(-3.75)));
 165 }
 166
 167 TEST_F(Simd4FloatingpointTest, gmxSimd4TruncR)
 168 {
 169     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(2), gmx_simd4_trunc_r(rSimd4_2p25));
 170     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(3), gmx_simd4_trunc_r(rSimd4_3p75));
 171     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-2), gmx_simd4_trunc_r(rSimd4_m2p25));
 172     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-3), gmx_simd4_trunc_r(rSimd4_m3p75));
 173 }
 174
 175 /* We do extensive 1/sqrt(x) and 1/x accuracy testing in the tests for
 176  * the SIMD math functions, so we just make sure the lookup instructions
 177  * appear to work for a few values here.
 178  */
 179 TEST_F(Simd4FloatingpointTest, gmxSimd4RsqrtR)
 180 {
 181     gmx_simd4_real_t x      = setSimd4RealFrom3R(4.0, M_PI, 1234567890.0);
 182     gmx_simd4_real_t ref    = setSimd4RealFrom3R(0.5, 1.0/sqrt(M_PI), 1.0/sqrt(1234567890.0));
 183
 184     // The allowed Ulp deviation is 2 to the power of the number of mantissa
 185     // digits, minus the number of bits provided by the table lookup
 186     setUlpTol(1LL << (std::numeric_limits<real>::digits-GMX_SIMD_RSQRT_BITS));
 187     GMX_EXPECT_SIMD4_REAL_NEAR(ref, gmx_simd4_rsqrt_r(x));
 188 }
 189
 190 TEST_F(Simd4FloatingpointTest, gmxSimd4BoolCmpEqAndBlendZeroR)
 191 {
 192     gmx_simd4_bool_t eq   = gmx_simd4_cmpeq_r(rSimd4_5_7_9, rSimd4_7_8_9);
 193     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(0, 0, 3), gmx_simd4_blendzero_r(rSimd4_1_2_3, eq));
 194 }
 195
 196 TEST_F(Simd4FloatingpointTest, gmxSimd4BlendNotZeroR)
 197 {
 198     gmx_simd4_bool_t eq   = gmx_simd4_cmpeq_r(rSimd4_5_7_9, rSimd4_7_8_9);
 199     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(1, 2, 0), gmx_simd4_blendnotzero_r(rSimd4_1_2_3, eq));
 200 }
 201
 202 TEST_F(Simd4FloatingpointTest, gmxSimd4BoolCmpLER)
 203 {
 204     gmx_simd4_bool_t le   = gmx_simd4_cmple_r(rSimd4_5_7_9, rSimd4_7_8_9);
 205     GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_1_2_3, gmx_simd4_blendzero_r(rSimd4_1_2_3, le));
 206 }
 207
 208 TEST_F(Simd4FloatingpointTest, gmxSimd4BoolCmpLTR)
 209 {
 210     gmx_simd4_bool_t lt   = gmx_simd4_cmplt_r(rSimd4_5_7_9, rSimd4_7_8_9);
 211     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(1, 2, 0), gmx_simd4_blendzero_r(rSimd4_1_2_3, lt));
 212 }
 213
 214 TEST_F(Simd4FloatingpointTest, gmxSimd4BoolAndB)
 215 {
 216     gmx_simd4_bool_t eq   = gmx_simd4_cmpeq_r(rSimd4_5_7_9, rSimd4_7_8_9);
 217     gmx_simd4_bool_t le   = gmx_simd4_cmple_r(rSimd4_5_7_9, rSimd4_7_8_9);
 218     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(0, 0, 3), gmx_simd4_blendzero_r(rSimd4_1_2_3, gmx_simd4_and_b(eq, le)));
 219 }
 220
 221 TEST_F(Simd4FloatingpointTest, gmxSimd4BoolOrB)
 222 {
 223     gmx_simd4_bool_t eq   = gmx_simd4_cmpeq_r(rSimd4_5_7_9, rSimd4_7_8_9);
 224     gmx_simd4_bool_t lt   = gmx_simd4_cmplt_r(rSimd4_5_7_9, rSimd4_7_8_9);
 225     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(1, 2, 3), gmx_simd4_blendzero_r(rSimd4_1_2_3, gmx_simd4_or_b(eq, lt)));
 226 }
 227
 228 TEST_F(Simd4FloatingpointTest, gmxSimd4AnytrueB)
 229 {
 230     gmx_simd4_bool_t eq;
 231
 232     /* this test is a bit tricky since we don't know the simd width.
 233      * We cannot check for truth values for "any" element beyond the first,
 234      * since that part of the data will not be used if simd width is 1.
 235      */
 236     eq = gmx_simd4_cmpeq_r(rSimd4_5_7_9, setSimd4RealFrom3R(5, 0, 0));
 237     EXPECT_NE(0, gmx_simd4_anytrue_b(eq));
 238
 239     eq = gmx_simd4_cmpeq_r(rSimd4_1_2_3, rSimd4_4_5_6);
 240     EXPECT_EQ(0, gmx_simd4_anytrue_b(eq));
 241 }
 242
 243 TEST_F(Simd4FloatingpointTest, gmxSimd4BlendvR)
 244 {
 245     gmx_simd4_bool_t lt   = gmx_simd4_cmplt_r(rSimd4_5_7_9, rSimd4_7_8_9);
 246     GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(4, 5, 3), gmx_simd4_blendv_r(rSimd4_1_2_3, rSimd4_4_5_6, lt));
 247 }
 248
 249 TEST_F(Simd4FloatingpointTest, gmxSimd4ReduceR)
 250 {
 251     // The horizontal sum of the SIMD variable depends on the width, so
 252     // simply store it an extra time and calculate what the sum should be
 253     std::vector<real> v   = simd4Real2Vector(rSimd4_1_2_3);
 254     real              sum = 0.0;
 255
 256     for (int i = 0; i < GMX_SIMD4_WIDTH; i++)
 257     {
 258         sum += v[i];
 259     }
 260
 261     EXPECT_EQ(sum, gmx_simd4_reduce_r(rSimd4_1_2_3));
 262 }
 263
 264
 265 TEST_F(Simd4FloatingpointTest, gmxSimd4Dotproduct3R)
 266 {
 267     gmx_simd4_real_t v1 = setSimd4RealFrom3R(1, 4, 5);
 268     gmx_simd4_real_t v2 = setSimd4RealFrom3R(3, 8, 2);
 269 #    ifdef GMX_DOUBLE
 270     EXPECT_DOUBLE_EQ(45.0, gmx_simd4_dotproduct3_r(v1, v2));
 271 #    else
 272     EXPECT_FLOAT_EQ(45.0, gmx_simd4_dotproduct3_r(v1, v2));
 273 #    endif
 274 }
 275
 276 #endif      // GMX_SIMD4_HAVE_REAL
 277
 278 /*! \} */
 279 /*! \endcond */
 280
 281 }      // namespace
 282 }      // namespace
 283 }      // namespace