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, 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.
42 #include "gromacs/math/utilities.h"
43 #include "gromacs/simd/simd.h"
44 #include "gromacs/utility/basedefinitions.h"
46 #include "testutils/testasserts.h"
62 /*! \addtogroup module_simd */
65 # if GMX_SIMD_HAVE_REAL
67 /*! \brief Test fixture for floating-point tests (identical to the generic \ref SimdTest) */
68 typedef SimdTest SimdFloatingpointTest;
70 TEST_F(SimdFloatingpointTest, setZero)
72 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(0.0), setZero());
75 TEST_F(SimdFloatingpointTest, set)
78 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(c1), SimdReal(c1));
79 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(c0), SimdReal(*p));
82 TEST_F(SimdFloatingpointTest, add)
84 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 + c3, c1 + c4, c2 + c5), rSimd_c0c1c2 + rSimd_c3c4c5);
87 TEST_F(SimdFloatingpointTest, maskAdd)
89 SimdBool m = setSimdRealFrom3R(c6, 0, c7) != setZero();
90 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 + c3, c1 + 0.0, c2 + c5),
91 maskAdd(rSimd_c0c1c2, rSimd_c3c4c5, m));
94 TEST_F(SimdFloatingpointTest, sub)
96 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 - c3, c1 - c4, c2 - c5), rSimd_c0c1c2 - rSimd_c3c4c5);
99 TEST_F(SimdFloatingpointTest, mul)
101 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 * c3, c1 * c4, c2 * c5), rSimd_c0c1c2 * rSimd_c3c4c5);
104 TEST_F(SimdFloatingpointTest, maskzMul)
106 SimdBool m = setSimdRealFrom3R(c1, 0, c1) != setZero();
107 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 * c3, 0.0, c2 * c5),
108 maskzMul(rSimd_c0c1c2, rSimd_c3c4c5, m));
111 TEST_F(SimdFloatingpointTest, fma)
113 // The last bit of FMA operations depends on hardware, so we don't require exact match
114 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 * c3 + c6, c1 * c4 + c7, c2 * c5 + c8),
115 fma(rSimd_c0c1c2, rSimd_c3c4c5, rSimd_c6c7c8));
119 TEST_F(SimdFloatingpointTest, maskzFma)
121 SimdBool m = setSimdRealFrom3R(c2, 0, c3) != setZero();
122 // The last bit of FMA operations depends on hardware, so we don't require exact match
123 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 * c3 + c6, 0.0, c2 * c5 + c8),
124 maskzFma(rSimd_c0c1c2, rSimd_c3c4c5, rSimd_c6c7c8, m));
127 TEST_F(SimdFloatingpointTest, fms)
129 // The last bit of FMA operations depends on hardware, so we don't require exact match
130 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 * c3 - c6, c1 * c4 - c7, c2 * c5 - c8),
131 fms(rSimd_c0c1c2, rSimd_c3c4c5, rSimd_c6c7c8));
134 TEST_F(SimdFloatingpointTest, fnma)
136 // The last bit of FMA operations depends on hardware, so we don't require exact match
137 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c6 - c0 * c3, c7 - c1 * c4, c8 - c2 * c5),
138 fnma(rSimd_c0c1c2, rSimd_c3c4c5, rSimd_c6c7c8));
141 TEST_F(SimdFloatingpointTest, fnms)
143 // The last bit of FMA operations depends on hardware, so we don't require exact match
144 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(-c0 * c3 - c6, -c1 * c4 - c7, -c2 * c5 - c8),
145 fnms(rSimd_c0c1c2, rSimd_c3c4c5, rSimd_c6c7c8));
148 TEST_F(SimdFloatingpointTest, abs)
150 GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, abs(rSimd_c0c1c2)); // fabs(x)=x
151 GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, abs(rSimd_m0m1m2)); // fabs(-x)=x
154 TEST_F(SimdFloatingpointTest, neg)
156 GMX_EXPECT_SIMD_REAL_EQ(rSimd_m0m1m2, -(rSimd_c0c1c2)); // fneg(x)=-x
157 GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, -(rSimd_m0m1m2)); // fneg(-x)=x
160 # if GMX_SIMD_HAVE_LOGICAL
161 TEST_F(SimdFloatingpointTest, and)
163 GMX_EXPECT_SIMD_REAL_EQ(rSimd_logicalResultAnd, (rSimd_logicalA & rSimd_logicalB));
166 TEST_F(SimdFloatingpointTest, or)
168 GMX_EXPECT_SIMD_REAL_EQ(rSimd_logicalResultOr, (rSimd_logicalA | rSimd_logicalB));
171 TEST_F(SimdFloatingpointTest, xor)
173 /* Test xor by taking xor with a number and its negative. This should result
174 * in only the sign bit being set. We then use this bit change the sign of
177 SimdReal signbit = SimdReal(c1) ^ SimdReal(-c1);
178 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-c2, c3, -c4), (signbit ^ setSimdRealFrom3R(c2, -c3, c4)));
181 TEST_F(SimdFloatingpointTest, andNot)
183 /* Use xor (which we already tested, so fix that first if both tests fail)
184 * to extract the sign bit, and then use andnot to take absolute values.
186 SimdReal signbit = SimdReal(c1) ^ SimdReal(-c1);
187 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c2, c3, c4),
188 andNot(signbit, setSimdRealFrom3R(-c2, c3, -c4)));
193 TEST_F(SimdFloatingpointTest, max)
195 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c3, c1, c4), max(rSimd_c0c1c2, rSimd_c3c0c4));
196 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c3, c1, c4), max(rSimd_c3c0c4, rSimd_c0c1c2));
197 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-c0, -c0, -c2), max(rSimd_m0m1m2, rSimd_m3m0m4));
198 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-c0, -c0, -c2), max(rSimd_m3m0m4, rSimd_m0m1m2));
201 TEST_F(SimdFloatingpointTest, min)
203 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, c0, c2), min(rSimd_c0c1c2, rSimd_c3c0c4));
204 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, c0, c2), min(rSimd_c3c0c4, rSimd_c0c1c2));
205 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-c3, -c1, -c4), min(rSimd_m0m1m2, rSimd_m3m0m4));
206 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-c3, -c1, -c4), min(rSimd_m3m0m4, rSimd_m0m1m2));
209 TEST_F(SimdFloatingpointTest, round)
211 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(2), round(rSimd_2p25));
212 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(4), round(rSimd_3p75));
213 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-2), round(rSimd_m2p25));
214 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-4), round(rSimd_m3p75));
217 TEST_F(SimdFloatingpointTest, roundMode)
219 /* Rounding mode needs to be consistent between round and cvtR2I */
220 SimdReal x0 = setSimdRealFrom3R(0.5, 11.5, 99.5);
221 SimdReal x1 = setSimdRealFrom3R(-0.5, -11.5, -99.5);
223 GMX_EXPECT_SIMD_REAL_EQ(round(x0), cvtI2R(cvtR2I(x0)));
224 GMX_EXPECT_SIMD_REAL_EQ(round(x1), cvtI2R(cvtR2I(x1)));
227 TEST_F(SimdFloatingpointTest, trunc)
229 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(2), trunc(rSimd_2p25));
230 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(3), trunc(rSimd_3p75));
231 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-2), trunc(rSimd_m2p25));
232 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-3), trunc(rSimd_m3p75));
235 // We explicitly test the exponent/mantissa routines with double precision data,
236 // since these usually rely on direct manipulation and shift of the SIMD registers,
237 // where it is easy to make mistakes with single vs double precision.
239 TEST_F(SimdFloatingpointTest, frexp)
244 fraction = frexp(rSimd_Exp, &exponent);
246 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(0.609548660288905419513128, 0.5833690139241746175358116,
247 -0.584452007502232362412542),
249 GMX_EXPECT_SIMD_INT_EQ(setSimdIntFrom3I(61, -40, 55), exponent);
252 # if GMX_SIMD_HAVE_DOUBLE && GMX_DOUBLE
253 fraction = frexp(rSimd_ExpDouble, &exponent);
255 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(0.6206306194761728178832527, 0.5236473618795619566768096,
256 -0.9280331023751380303821179),
258 GMX_EXPECT_SIMD_INT_EQ(setSimdIntFrom3I(588, -461, 673), exponent);
262 TEST_F(SimdFloatingpointTest, ldexp)
264 SimdReal one = setSimdRealFrom1R(1.0);
266 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(pow(2.0, 60.0), pow(2.0, -41.0), pow(2.0, 54.0)),
267 ldexp<MathOptimization::Unsafe>(one, setSimdIntFrom3I(60, -41, 54)));
268 # if GMX_SIMD_HAVE_DOUBLE && GMX_DOUBLE
269 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(pow(2.0, 587.0), pow(2.0, -462.0), pow(2.0, 672.0)),
270 ldexp<MathOptimization::Unsafe>(one, setSimdIntFrom3I(587, -462, 672)));
272 // The default safe version must be able to handle very negative arguments too
273 GMX_EXPECT_SIMD_REAL_EQ(setZero(), ldexp(one, setSimdIntFrom3I(-2000, -1000000, -1000000000)));
277 * We do extensive 1/sqrt(x) and 1/x accuracy testing in the math module, so
278 * we just make sure the lookup instructions appear to work here
281 TEST_F(SimdFloatingpointTest, rsqrt)
283 SimdReal x = setSimdRealFrom3R(4.0, M_PI, 1234567890.0);
284 SimdReal ref = setSimdRealFrom3R(0.5, 1.0 / std::sqrt(M_PI), 1.0 / std::sqrt(1234567890.0));
285 int shiftbits = std::numeric_limits<real>::digits - GMX_SIMD_RSQRT_BITS;
292 /* Set the allowed ulp error as 2 to the power of the number of bits in
293 * the mantissa that do not have to be correct after the table lookup.
295 setUlpTol(1LL << shiftbits);
296 GMX_EXPECT_SIMD_REAL_NEAR(ref, rsqrt(x));
299 TEST_F(SimdFloatingpointTest, maskzRsqrt)
301 SimdReal x = setSimdRealFrom3R(M_PI, -4.0, 0.0);
302 // simdCmpLe is tested separately further down
303 SimdBool m = setZero() < x;
304 SimdReal ref = setSimdRealFrom3R(1.0 / std::sqrt(M_PI), 0.0, 0.0);
305 int shiftbits = std::numeric_limits<real>::digits - GMX_SIMD_RSQRT_BITS;
312 /* Set the allowed ulp error as 2 to the power of the number of bits in
313 * the mantissa that do not have to be correct after the table lookup.
315 setUlpTol(1LL << shiftbits);
316 GMX_EXPECT_SIMD_REAL_NEAR(ref, maskzRsqrt(x, m));
319 TEST_F(SimdFloatingpointTest, rcp)
321 SimdReal x = setSimdRealFrom3R(4.0, M_PI, 1234567890.0);
322 SimdReal ref = setSimdRealFrom3R(0.25, 1.0 / M_PI, 1.0 / 1234567890.0);
323 int shiftbits = std::numeric_limits<real>::digits - GMX_SIMD_RCP_BITS;
330 /* Set the allowed ulp error as 2 to the power of the number of bits in
331 * the mantissa that do not have to be correct after the table lookup.
333 setUlpTol(1LL << shiftbits);
334 GMX_EXPECT_SIMD_REAL_NEAR(ref, rcp(x));
337 TEST_F(SimdFloatingpointTest, maskzRcp)
339 SimdReal x = setSimdRealFrom3R(M_PI, 0.0, -1234567890.0);
340 SimdBool m = (x != setZero());
341 SimdReal ref = setSimdRealFrom3R(1.0 / M_PI, 0.0, -1.0 / 1234567890.0);
342 int shiftbits = std::numeric_limits<real>::digits - GMX_SIMD_RCP_BITS;
349 /* Set the allowed ulp error as 2 to the power of the number of bits in
350 * the mantissa that do not have to be correct after the table lookup.
352 setUlpTol(1LL << shiftbits);
353 GMX_EXPECT_SIMD_REAL_NEAR(ref, maskzRcp(x, m));
356 TEST_F(SimdFloatingpointTest, cmpEqAndSelectByMask)
358 SimdBool eq = rSimd_c4c6c8 == rSimd_c6c7c8;
359 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(0, 0, c2), selectByMask(rSimd_c0c1c2, eq));
362 TEST_F(SimdFloatingpointTest, selectByNotMask)
364 SimdBool eq = rSimd_c4c6c8 == rSimd_c6c7c8;
365 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, c1, 0), selectByNotMask(rSimd_c0c1c2, eq));
368 TEST_F(SimdFloatingpointTest, cmpNe)
370 SimdBool eq = rSimd_c4c6c8 != rSimd_c6c7c8;
371 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, c1, 0), selectByMask(rSimd_c0c1c2, eq));
374 TEST_F(SimdFloatingpointTest, cmpLe)
376 SimdBool le = rSimd_c4c6c8 <= rSimd_c6c7c8;
377 GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, selectByMask(rSimd_c0c1c2, le));
380 TEST_F(SimdFloatingpointTest, cmpLt)
382 SimdBool lt = rSimd_c4c6c8 < rSimd_c6c7c8;
383 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, c1, 0), selectByMask(rSimd_c0c1c2, lt));
386 # if GMX_SIMD_HAVE_INT32_LOGICAL || GMX_SIMD_HAVE_LOGICAL
387 TEST_F(SimdFloatingpointTest, testBits)
389 SimdBool eq = testBits(setSimdRealFrom3R(c1, 0, c1));
390 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, 0, c2), selectByMask(rSimd_c0c1c2, eq));
392 // Test if we detect only the sign bit being set
393 eq = testBits(setSimdRealFrom1R(GMX_REAL_NEGZERO));
394 GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, selectByMask(rSimd_c0c1c2, eq));
398 TEST_F(SimdFloatingpointTest, andB)
400 SimdBool eq = rSimd_c4c6c8 == rSimd_c6c7c8;
401 SimdBool le = rSimd_c4c6c8 <= rSimd_c6c7c8;
402 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(0, 0, c2), selectByMask(rSimd_c0c1c2, (eq && le)));
405 TEST_F(SimdFloatingpointTest, orB)
407 SimdBool eq = rSimd_c4c6c8 == rSimd_c6c7c8;
408 SimdBool lt = rSimd_c4c6c8 < rSimd_c6c7c8;
409 GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, selectByMask(rSimd_c0c1c2, (eq || lt)));
412 TEST_F(SimdFloatingpointTest, anyTrueB)
414 alignas(GMX_SIMD_ALIGNMENT) std::array<real, GMX_SIMD_REAL_WIDTH> mem{};
416 // Test the false case
417 EXPECT_FALSE(anyTrue(setZero() < load<SimdReal>(mem.data())));
419 // Test each bit (these should all be true)
420 for (int i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
424 EXPECT_TRUE(anyTrue(setZero() < load<SimdReal>(mem.data())))
425 << "Not detecting true in element " << i;
429 TEST_F(SimdFloatingpointTest, blend)
431 SimdBool lt = rSimd_c4c6c8 < rSimd_c6c7c8;
432 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c3, c4, c2), blend(rSimd_c0c1c2, rSimd_c3c4c5, lt));
435 TEST_F(SimdFloatingpointTest, reduce)
437 // The horizontal sum of the SIMD variable depends on the width, so
438 // simply store it an extra time and calculate what the sum should be
439 std::vector<real> v = simdReal2Vector(rSimd_c3c4c5);
442 for (int i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
447 EXPECT_REAL_EQ_TOL(sum, reduce(rSimd_c3c4c5), defaultRealTolerance());
450 # endif // GMX_SIMD_HAVE_REAL
452 # if GMX_SIMD_HAVE_FLOAT && GMX_SIMD_HAVE_DOUBLE
453 TEST_F(SimdFloatingpointTest, cvtFloat2Double)
455 alignas(GMX_SIMD_ALIGNMENT) float f[GMX_SIMD_FLOAT_WIDTH];
456 alignas(GMX_SIMD_ALIGNMENT) double d[GMX_SIMD_FLOAT_WIDTH]; // Yes, double array length should be same as float
461 FloatingPointTolerance tolerance(defaultRealTolerance());
463 for (i = 0; i < GMX_SIMD_FLOAT_WIDTH; i++)
465 // Scale by 1+100*eps to use low bits too.
466 // Due to the conversions we want to avoid being too sensitive to fluctuations in last bit
467 f[i] = i * (1.0 + 100 * GMX_FLOAT_EPS);
470 vf = load<SimdFloat>(f);
471 # if (GMX_SIMD_FLOAT_WIDTH == 2 * GMX_SIMD_DOUBLE_WIDTH)
473 cvtF2DD(vf, &vd0, &vd1);
474 store(d + GMX_SIMD_DOUBLE_WIDTH, vd1); // Store upper part halfway through array
475 # elif (GMX_SIMD_FLOAT_WIDTH == GMX_SIMD_DOUBLE_WIDTH)
478 # error Width of float SIMD must either be identical to double, or twice the width.
480 store(d, vd0); // store lower (or whole) part from start of vector
482 for (i = 0; i < GMX_SIMD_FLOAT_WIDTH; i++)
484 EXPECT_REAL_EQ_TOL(f[i], d[i], tolerance);
488 TEST_F(SimdFloatingpointTest, cvtDouble2Float)
490 alignas(GMX_SIMD_ALIGNMENT) float f[GMX_SIMD_FLOAT_WIDTH];
491 alignas(GMX_SIMD_ALIGNMENT) double d[GMX_SIMD_FLOAT_WIDTH]; // Yes, double array length should be same as float
495 FloatingPointTolerance tolerance(defaultRealTolerance());
497 // This fills elements for pd1 too when double width is 2*single width
498 for (i = 0; i < GMX_SIMD_FLOAT_WIDTH; i++)
500 // Scale by 1+eps to use low bits too.
501 // Due to the conversions we want to avoid being too sensitive to fluctuations in last bit
502 d[i] = i * (1.0 + 100 * GMX_FLOAT_EPS);
505 vd0 = load<SimdDouble>(d);
506 # if (GMX_SIMD_FLOAT_WIDTH == 2 * GMX_SIMD_DOUBLE_WIDTH)
507 SimdDouble vd1 = load<SimdDouble>(d + GMX_SIMD_DOUBLE_WIDTH); // load upper half of data
508 vf = cvtDD2F(vd0, vd1);
509 # elif (GMX_SIMD_FLOAT_WIDTH == GMX_SIMD_DOUBLE_WIDTH)
512 # error Width of float SIMD must either be identical to double, or twice the width.
516 // This will check elements in pd1 too when double width is 2*single width
517 for (i = 0; i < GMX_SIMD_FLOAT_WIDTH; i++)
519 EXPECT_FLOAT_EQ_TOL(d[i], f[i], tolerance);
522 # endif // GMX_SIMD_HAVE_FLOAT && GMX_SIMD_HAVE_DOUBLE