2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2014,2015,2016,2017,2018 by the GROMACS development team.
5 * Copyright (c) 2019,2020,2021, 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.
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.
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.
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.
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.
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.
40 #include "gromacs/math/units.h"
41 #include "gromacs/math/utilities.h"
42 #include "gromacs/simd/simd.h"
43 #include "gromacs/utility/basedefinitions.h"
45 #include "testutils/testasserts.h"
60 /*! \addtogroup module_simd */
63 # if GMX_SIMD4_HAVE_REAL
65 /*! \brief Test fixture for SIMD4 floating-point operations (identical to the SIMD4 \ref Simd4Test) */
66 typedef Simd4Test Simd4FloatingpointTest;
68 TEST_F(Simd4FloatingpointTest, setZero)
70 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(0.0), setZero());
73 TEST_F(Simd4FloatingpointTest, set)
75 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(c1), Simd4Real(c1));
78 TEST_F(Simd4FloatingpointTest, add)
80 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0 + c3, c1 + c4, c2 + c5), rSimd4_c0c1c2 + rSimd4_c3c4c5);
83 TEST_F(Simd4FloatingpointTest, sub)
85 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0 - c3, c1 - c4, c2 - c5), rSimd4_c0c1c2 - rSimd4_c3c4c5);
88 TEST_F(Simd4FloatingpointTest, mul)
90 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0 * c3, c1 * c4, c2 * c5), rSimd4_c0c1c2 * rSimd4_c3c4c5);
93 TEST_F(Simd4FloatingpointTest, fma)
95 // The last bit of FMA operations depends on hardware, so we don't require exact match
96 GMX_EXPECT_SIMD4_REAL_NEAR(setSimd4RealFrom3R(c0 * c3 + c6, c1 * c4 + c7, c2 * c5 + c8),
97 fma(rSimd4_c0c1c2, rSimd4_c3c4c5, rSimd4_c6c7c8));
100 TEST_F(Simd4FloatingpointTest, fms)
102 // The last bit of FMA operations depends on hardware, so we don't require exact match
103 GMX_EXPECT_SIMD4_REAL_NEAR(setSimd4RealFrom3R(c0 * c3 - c6, c1 * c4 - c7, c2 * c5 - c8),
104 fms(rSimd4_c0c1c2, rSimd4_c3c4c5, rSimd4_c6c7c8));
107 TEST_F(Simd4FloatingpointTest, fnma)
109 // The last bit of FMA operations depends on hardware, so we don't require exact match
110 GMX_EXPECT_SIMD4_REAL_NEAR(setSimd4RealFrom3R(c6 - c0 * c3, c7 - c1 * c4, c8 - c2 * c5),
111 fnma(rSimd4_c0c1c2, rSimd4_c3c4c5, rSimd4_c6c7c8));
114 TEST_F(Simd4FloatingpointTest, fnms)
116 // The last bit of FMA operations depends on hardware, so we don't require exact match
117 GMX_EXPECT_SIMD4_REAL_NEAR(setSimd4RealFrom3R(-c0 * c3 - c6, -c1 * c4 - c7, -c2 * c5 - c8),
118 fnms(rSimd4_c0c1c2, rSimd4_c3c4c5, rSimd4_c6c7c8));
121 TEST_F(Simd4FloatingpointTest, abs)
123 GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_c0c1c2, abs(rSimd4_c0c1c2)); // fabs(x)=x
124 GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_c0c1c2, abs(rSimd4_m0m1m2)); // fabs(-x)=x
127 TEST_F(Simd4FloatingpointTest, neg)
129 GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_m0m1m2, -(rSimd4_c0c1c2)); // fneg(x)=-x
130 GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_c0c1c2, -(rSimd4_m0m1m2)); // fneg(-x)=x
133 # if GMX_SIMD_HAVE_LOGICAL
134 TEST_F(Simd4FloatingpointTest, and)
136 GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_logicalResultAnd, (rSimd4_logicalA & rSimd4_logicalB));
139 TEST_F(Simd4FloatingpointTest, or)
141 GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_logicalResultOr, (rSimd4_logicalA | rSimd4_logicalB));
144 TEST_F(Simd4FloatingpointTest, xor)
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
150 Simd4Real signbit = Simd4Real(c1) ^ Simd4Real(-c1);
151 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-c2, c3, -c4), signbit ^ setSimd4RealFrom3R(c2, -c3, c4));
154 TEST_F(Simd4FloatingpointTest, andNot)
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.
159 Simd4Real signbit = Simd4Real(c1) ^ Simd4Real(-c1);
160 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c2, c3, c4),
161 andNot(signbit, setSimd4RealFrom3R(-c2, c3, -c4)));
166 TEST_F(Simd4FloatingpointTest, max)
168 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c3, c1, c4), max(rSimd4_c0c1c2, rSimd4_c3c0c4));
169 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c3, c1, c4), max(rSimd4_c3c0c4, rSimd4_c0c1c2));
170 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-c0, -c0, -c2), max(rSimd4_m0m1m2, rSimd4_m3m0m4));
171 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-c0, -c0, -c2), max(rSimd4_m3m0m4, rSimd4_m0m1m2));
174 TEST_F(Simd4FloatingpointTest, min)
176 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0, c0, c2), min(rSimd4_c0c1c2, rSimd4_c3c0c4));
177 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0, c0, c2), min(rSimd4_c3c0c4, rSimd4_c0c1c2));
178 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-c3, -c1, -c4), min(rSimd4_m0m1m2, rSimd4_m3m0m4));
179 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(-c3, -c1, -c4), min(rSimd4_m3m0m4, rSimd4_m0m1m2));
182 TEST_F(Simd4FloatingpointTest, round)
184 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(2), round(Simd4Real(2.25)));
185 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(4), round(Simd4Real(3.75)));
186 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-2), round(Simd4Real(-2.25)));
187 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-4), round(Simd4Real(-3.75)));
190 TEST_F(Simd4FloatingpointTest, trunc)
192 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(2), trunc(rSimd4_2p25));
193 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(3), trunc(rSimd4_3p75));
194 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-2), trunc(rSimd4_m2p25));
195 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom1R(-3), trunc(rSimd4_m3p75));
198 /* We do extensive 1/sqrt(x) and 1/x accuracy testing in the tests for
199 * the SIMD math functions, so we just make sure the lookup instructions
200 * appear to work for a few values here.
202 TEST_F(Simd4FloatingpointTest, gmxSimd4RsqrtR)
204 Simd4Real x = setSimd4RealFrom3R(4.0, M_PI, 1234567890.0);
205 Simd4Real ref = setSimd4RealFrom3R(0.5, 1.0 / std::sqrt(M_PI), 1.0 / std::sqrt(1234567890.0));
206 int shiftbits = std::numeric_limits<real>::digits - GMX_SIMD_RSQRT_BITS;
213 // The allowed Ulp deviation is 2 to the power of the number of mantissa
214 // digits, minus the number of bits provided by the table lookup
215 setUlpTol(1LL << shiftbits);
216 GMX_EXPECT_SIMD4_REAL_NEAR(ref, rsqrt(x));
219 TEST_F(Simd4FloatingpointTest, cmpEqAndSelectByMask)
221 Simd4Bool eq = rSimd4_c4c6c8 == rSimd4_c6c7c8;
222 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(0, 0, c2), selectByMask(rSimd4_c0c1c2, eq));
225 TEST_F(Simd4FloatingpointTest, selectByNotMask)
227 Simd4Bool eq = rSimd4_c4c6c8 == rSimd4_c6c7c8;
228 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0, c1, 0), selectByNotMask(rSimd4_c0c1c2, eq));
231 TEST_F(Simd4FloatingpointTest, cmpNe)
233 Simd4Bool eq = rSimd4_c4c6c8 != rSimd4_c6c7c8;
234 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0, c1, 0), selectByMask(rSimd4_c0c1c2, eq));
237 TEST_F(Simd4FloatingpointTest, cmpLe)
239 Simd4Bool le = rSimd4_c4c6c8 <= rSimd4_c6c7c8;
240 GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_c0c1c2, selectByMask(rSimd4_c0c1c2, le));
243 TEST_F(Simd4FloatingpointTest, cmpLt)
245 Simd4Bool lt = rSimd4_c4c6c8 < rSimd4_c6c7c8;
246 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c0, c1, 0), selectByMask(rSimd4_c0c1c2, lt));
249 TEST_F(Simd4FloatingpointTest, andB)
251 Simd4Bool eq = rSimd4_c4c6c8 == rSimd4_c6c7c8;
252 Simd4Bool le = rSimd4_c4c6c8 <= rSimd4_c6c7c8;
253 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(0, 0, c2), selectByMask(rSimd4_c0c1c2, (eq && le)));
256 TEST_F(Simd4FloatingpointTest, orB)
258 Simd4Bool eq = rSimd4_c4c6c8 == rSimd4_c6c7c8;
259 Simd4Bool lt = rSimd4_c4c6c8 < rSimd4_c6c7c8;
260 GMX_EXPECT_SIMD4_REAL_EQ(rSimd4_c0c1c2, selectByMask(rSimd4_c0c1c2, (eq || lt)));
263 TEST_F(Simd4FloatingpointTest, anyTrue)
267 /* this test is a bit tricky since we don't know the simd width.
268 * We cannot check for truth values for "any" element beyond the first,
269 * since that part of the data will not be used if simd width is 1.
271 eq = (rSimd4_c4c6c8 == setSimd4RealFrom3R(c4, 0, 0));
272 EXPECT_TRUE(anyTrue(eq));
274 eq = (rSimd4_c0c1c2 == rSimd4_c3c4c5);
275 EXPECT_FALSE(anyTrue(eq));
278 TEST_F(Simd4FloatingpointTest, blend)
280 Simd4Bool lt = rSimd4_c4c6c8 < rSimd4_c6c7c8;
281 GMX_EXPECT_SIMD4_REAL_EQ(setSimd4RealFrom3R(c3, c4, c2), blend(rSimd4_c0c1c2, rSimd4_c3c4c5, lt));
284 TEST_F(Simd4FloatingpointTest, reduce)
286 // The horizontal sum of the SIMD variable depends on the width, so
287 // simply store it an extra time and calculate what the sum should be
288 std::vector<real> v = simd4Real2Vector(rSimd4_c3c4c5);
291 for (int i = 0; i < GMX_SIMD4_WIDTH; i++)
296 EXPECT_REAL_EQ_TOL(sum, reduce(rSimd4_c3c4c5), defaultRealTolerance());
300 TEST_F(Simd4FloatingpointTest, dotProduct)
302 real res = c0 * c3 + c1 * c4 + c2 * c5;
304 EXPECT_REAL_EQ_TOL(res, dotProduct(rSimd4_c0c1c2, rSimd4_c3c4c5), defaultRealTolerance());
307 TEST_F(Simd4FloatingpointTest, transpose)
309 Simd4Real v0, v1, v2, v3;
312 alignas(GMX_SIMD_ALIGNMENT) real p0[4 * GMX_SIMD4_WIDTH];
313 real* p1 = p0 + GMX_SIMD4_WIDTH;
314 real* p2 = p0 + 2 * GMX_SIMD4_WIDTH;
315 real* p3 = p0 + 3 * GMX_SIMD4_WIDTH;
317 // Assign data with tens as row, single-digit as column
318 for (i = 0; i < 4; i++)
320 // Scale by 1+100*eps to use low bits tii
321 p0[i] = (0 * 10 + i * 1) * (1.0 + 100 * GMX_REAL_EPS);
322 p1[i] = (1 * 10 + i * 1) * (1.0 + 100 * GMX_REAL_EPS);
323 p2[i] = (2 * 10 + i * 1) * (1.0 + 100 * GMX_REAL_EPS);
324 p3[i] = (3 * 10 + i * 1) * (1.0 + 100 * GMX_REAL_EPS);
332 transpose(&v0, &v1, &v2, &v3);
339 for (i = 0; i < 4; i++)
341 EXPECT_REAL_EQ_TOL((i * 10 + 0) * (1.0 + 100 * GMX_REAL_EPS), p0[i], defaultRealTolerance());
342 EXPECT_REAL_EQ_TOL((i * 10 + 1) * (1.0 + 100 * GMX_REAL_EPS), p1[i], defaultRealTolerance());
343 EXPECT_REAL_EQ_TOL((i * 10 + 2) * (1.0 + 100 * GMX_REAL_EPS), p2[i], defaultRealTolerance());
344 EXPECT_REAL_EQ_TOL((i * 10 + 3) * (1.0 + 100 * GMX_REAL_EPS), p3[i], defaultRealTolerance());
348 # endif // GMX_SIMD4_HAVE_REAL