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.
42 #include "gromacs/math/units.h"
43 #include "gromacs/math/utilities.h"
44 #include "gromacs/simd/simd.h"
45 #include "gromacs/utility/basedefinitions.h"
47 #include "testutils/testasserts.h"
63 /*! \addtogroup module_simd */
66 # if GMX_SIMD_HAVE_REAL
68 /*! \brief Test fixture for floating-point tests (identical to the generic \ref SimdTest) */
69 typedef SimdTest SimdFloatingpointTest;
71 TEST_F(SimdFloatingpointTest, setZero)
73 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(0.0), setZero());
76 TEST_F(SimdFloatingpointTest, set)
79 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(c1), SimdReal(c1));
80 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(c0), SimdReal(*p));
83 TEST_F(SimdFloatingpointTest, add)
85 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 + c3, c1 + c4, c2 + c5), rSimd_c0c1c2 + rSimd_c3c4c5);
88 TEST_F(SimdFloatingpointTest, maskAdd)
90 SimdBool m = setSimdRealFrom3R(c6, 0, c7) != setZero();
91 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 + c3, c1 + 0.0, c2 + c5),
92 maskAdd(rSimd_c0c1c2, rSimd_c3c4c5, m));
95 TEST_F(SimdFloatingpointTest, sub)
97 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 - c3, c1 - c4, c2 - c5), rSimd_c0c1c2 - rSimd_c3c4c5);
100 TEST_F(SimdFloatingpointTest, mul)
102 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 * c3, c1 * c4, c2 * c5), rSimd_c0c1c2 * rSimd_c3c4c5);
105 TEST_F(SimdFloatingpointTest, maskzMul)
107 SimdBool m = setSimdRealFrom3R(c1, 0, c1) != setZero();
108 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 * c3, 0.0, c2 * c5),
109 maskzMul(rSimd_c0c1c2, rSimd_c3c4c5, m));
112 TEST_F(SimdFloatingpointTest, fma)
114 // The last bit of FMA operations depends on hardware, so we don't require exact match
115 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 * c3 + c6, c1 * c4 + c7, c2 * c5 + c8),
116 fma(rSimd_c0c1c2, rSimd_c3c4c5, rSimd_c6c7c8));
120 TEST_F(SimdFloatingpointTest, maskzFma)
122 SimdBool m = setSimdRealFrom3R(c2, 0, c3) != setZero();
123 // The last bit of FMA operations depends on hardware, so we don't require exact match
124 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 * c3 + c6, 0.0, c2 * c5 + c8),
125 maskzFma(rSimd_c0c1c2, rSimd_c3c4c5, rSimd_c6c7c8, m));
128 TEST_F(SimdFloatingpointTest, fms)
130 // The last bit of FMA operations depends on hardware, so we don't require exact match
131 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c0 * c3 - c6, c1 * c4 - c7, c2 * c5 - c8),
132 fms(rSimd_c0c1c2, rSimd_c3c4c5, rSimd_c6c7c8));
135 TEST_F(SimdFloatingpointTest, fnma)
137 // The last bit of FMA operations depends on hardware, so we don't require exact match
138 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(c6 - c0 * c3, c7 - c1 * c4, c8 - c2 * c5),
139 fnma(rSimd_c0c1c2, rSimd_c3c4c5, rSimd_c6c7c8));
142 TEST_F(SimdFloatingpointTest, fnms)
144 // The last bit of FMA operations depends on hardware, so we don't require exact match
145 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom3R(-c0 * c3 - c6, -c1 * c4 - c7, -c2 * c5 - c8),
146 fnms(rSimd_c0c1c2, rSimd_c3c4c5, rSimd_c6c7c8));
149 TEST_F(SimdFloatingpointTest, abs)
151 GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, abs(rSimd_c0c1c2)); // fabs(x)=x
152 GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, abs(rSimd_m0m1m2)); // fabs(-x)=x
155 TEST_F(SimdFloatingpointTest, neg)
157 GMX_EXPECT_SIMD_REAL_EQ(rSimd_m0m1m2, -(rSimd_c0c1c2)); // fneg(x)=-x
158 GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, -(rSimd_m0m1m2)); // fneg(-x)=x
161 # if GMX_SIMD_HAVE_LOGICAL
162 TEST_F(SimdFloatingpointTest, and)
164 GMX_EXPECT_SIMD_REAL_EQ(rSimd_logicalResultAnd, (rSimd_logicalA & rSimd_logicalB));
167 TEST_F(SimdFloatingpointTest, or)
169 GMX_EXPECT_SIMD_REAL_EQ(rSimd_logicalResultOr, (rSimd_logicalA | rSimd_logicalB));
172 TEST_F(SimdFloatingpointTest, xor)
174 /* Test xor by taking xor with a number and its negative. This should result
175 * in only the sign bit being set. We then use this bit change the sign of
178 SimdReal signbit = SimdReal(c1) ^ SimdReal(-c1);
179 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-c2, c3, -c4), (signbit ^ setSimdRealFrom3R(c2, -c3, c4)));
182 TEST_F(SimdFloatingpointTest, andNot)
184 /* Use xor (which we already tested, so fix that first if both tests fail)
185 * to extract the sign bit, and then use andnot to take absolute values.
187 SimdReal signbit = SimdReal(c1) ^ SimdReal(-c1);
188 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c2, c3, c4),
189 andNot(signbit, setSimdRealFrom3R(-c2, c3, -c4)));
194 TEST_F(SimdFloatingpointTest, max)
196 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c3, c1, c4), max(rSimd_c0c1c2, rSimd_c3c0c4));
197 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c3, c1, c4), max(rSimd_c3c0c4, rSimd_c0c1c2));
198 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-c0, -c0, -c2), max(rSimd_m0m1m2, rSimd_m3m0m4));
199 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-c0, -c0, -c2), max(rSimd_m3m0m4, rSimd_m0m1m2));
202 TEST_F(SimdFloatingpointTest, min)
204 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, c0, c2), min(rSimd_c0c1c2, rSimd_c3c0c4));
205 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, c0, c2), min(rSimd_c3c0c4, rSimd_c0c1c2));
206 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-c3, -c1, -c4), min(rSimd_m0m1m2, rSimd_m3m0m4));
207 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(-c3, -c1, -c4), min(rSimd_m3m0m4, rSimd_m0m1m2));
210 TEST_F(SimdFloatingpointTest, round)
212 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(2), round(rSimd_2p25));
213 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(4), round(rSimd_3p75));
214 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-2), round(rSimd_m2p25));
215 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-4), round(rSimd_m3p75));
218 TEST_F(SimdFloatingpointTest, roundMode)
220 /* Rounding mode needs to be consistent between round and cvtR2I */
221 SimdReal x0 = setSimdRealFrom3R(0.5, 11.5, 99.5);
222 SimdReal x1 = setSimdRealFrom3R(-0.5, -11.5, -99.5);
224 GMX_EXPECT_SIMD_REAL_EQ(round(x0), cvtI2R(cvtR2I(x0)));
225 GMX_EXPECT_SIMD_REAL_EQ(round(x1), cvtI2R(cvtR2I(x1)));
228 TEST_F(SimdFloatingpointTest, trunc)
230 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(2), trunc(rSimd_2p25));
231 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(3), trunc(rSimd_3p75));
232 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-2), trunc(rSimd_m2p25));
233 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom1R(-3), trunc(rSimd_m3p75));
236 // We explicitly test the exponent/mantissa routines with double precision data,
237 // since these usually rely on direct manipulation and shift of the SIMD registers,
238 // where it is easy to make mistakes with single vs double precision.
240 TEST_F(SimdFloatingpointTest, frexp)
246 fraction = frexp(rSimd_Exp, &exponent);
247 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(0.609548660288905419513128,
248 0.5833690139241746175358116,
249 -0.584452007502232362412542),
251 GMX_EXPECT_SIMD_INT_EQ(setSimdIntFrom3I(61, -40, 55), exponent);
253 // Test the unsafe flavor too, in case they use different branches
254 fraction = frexp<MathOptimization::Unsafe>(rSimd_Exp, &exponent);
255 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(0.609548660288905419513128,
256 0.5833690139241746175358116,
257 -0.584452007502232362412542),
259 GMX_EXPECT_SIMD_INT_EQ(setSimdIntFrom3I(61, -40, 55), exponent);
261 // Use ulp testing with 0 bit ulp tolerance for testing to separate 0.0 and -0.0
263 fraction = frexp(setSimdRealFrom1R(0.0), &exponent);
264 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom1R(0.0), fraction);
265 GMX_EXPECT_SIMD_INT_EQ(setSimdIntFrom1I(0), exponent);
268 fraction = frexp(setSimdRealFrom1R(-0.0), &exponent);
269 GMX_EXPECT_SIMD_REAL_NEAR(setSimdRealFrom1R(-0.0), fraction);
270 GMX_EXPECT_SIMD_INT_EQ(setSimdIntFrom1I(0), exponent);
272 // Reset to default ulp tolerance
273 setUlpTol(defaultRealUlpTol());
275 # if GMX_SIMD_HAVE_DOUBLE && GMX_DOUBLE
276 // Test exponents larger than what fit in single precision, as well as mixtures of 0 and non-zero values, to
277 // make sure the shuffling operations in the double-precision implementations don't do anything bad.
278 fraction = frexp(rSimd_ExpDouble1, &exponent);
279 GMX_EXPECT_SIMD_REAL_EQ(
280 setSimdRealFrom3R(0.0, 0.5236473618795619566768096, -0.9280331023751380303821179), fraction);
281 GMX_EXPECT_SIMD_INT_EQ(setSimdIntFrom3I(0, -461, 673), exponent);
283 fraction = frexp(rSimd_ExpDouble2, &exponent);
284 GMX_EXPECT_SIMD_REAL_EQ(
285 setSimdRealFrom3R(0.6206306194761728178832527, 0.0, -0.9280331023751380303821179), fraction);
286 GMX_EXPECT_SIMD_INT_EQ(setSimdIntFrom3I(588, 0, 673), exponent);
290 TEST_F(SimdFloatingpointTest, ldexp)
292 SimdReal one = setSimdRealFrom1R(1.0);
294 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(pow(2.0, 60.0), pow(2.0, -41.0), pow(2.0, 54.0)),
295 ldexp<MathOptimization::Unsafe>(one, setSimdIntFrom3I(60, -41, 54)));
296 # if GMX_SIMD_HAVE_DOUBLE && GMX_DOUBLE
297 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(pow(2.0, 587.0), pow(2.0, -462.0), pow(2.0, 672.0)),
298 ldexp<MathOptimization::Unsafe>(one, setSimdIntFrom3I(587, -462, 672)));
300 // The default safe version must be able to handle very negative arguments too
301 GMX_EXPECT_SIMD_REAL_EQ(setZero(), ldexp(one, setSimdIntFrom3I(-2000, -1000000, -1000000000)));
305 * We do extensive 1/sqrt(x) and 1/x accuracy testing in the math module, so
306 * we just make sure the lookup instructions appear to work here
309 TEST_F(SimdFloatingpointTest, rsqrt)
311 SimdReal x = setSimdRealFrom3R(4.0, M_PI, 1234567890.0);
312 SimdReal ref = setSimdRealFrom3R(0.5, 1.0 / std::sqrt(M_PI), 1.0 / std::sqrt(1234567890.0));
313 int shiftbits = std::numeric_limits<real>::digits - GMX_SIMD_RSQRT_BITS;
320 /* Set the allowed ulp error as 2 to the power of the number of bits in
321 * the mantissa that do not have to be correct after the table lookup.
323 setUlpTol(1LL << shiftbits);
324 GMX_EXPECT_SIMD_REAL_NEAR(ref, rsqrt(x));
327 TEST_F(SimdFloatingpointTest, maskzRsqrt)
329 SimdReal x = setSimdRealFrom3R(M_PI, -4.0, 0.0);
330 // simdCmpLe is tested separately further down
331 SimdBool m = setZero() < x;
332 SimdReal ref = setSimdRealFrom3R(1.0 / std::sqrt(M_PI), 0.0, 0.0);
333 int shiftbits = std::numeric_limits<real>::digits - GMX_SIMD_RSQRT_BITS;
340 /* Set the allowed ulp error as 2 to the power of the number of bits in
341 * the mantissa that do not have to be correct after the table lookup.
343 setUlpTol(1LL << shiftbits);
344 GMX_EXPECT_SIMD_REAL_NEAR(ref, maskzRsqrt(x, m));
347 TEST_F(SimdFloatingpointTest, rcp)
349 SimdReal x = setSimdRealFrom3R(4.0, M_PI, 1234567890.0);
350 SimdReal ref = setSimdRealFrom3R(0.25, 1.0 / M_PI, 1.0 / 1234567890.0);
351 int shiftbits = std::numeric_limits<real>::digits - GMX_SIMD_RCP_BITS;
358 /* Set the allowed ulp error as 2 to the power of the number of bits in
359 * the mantissa that do not have to be correct after the table lookup.
361 setUlpTol(1LL << shiftbits);
362 GMX_EXPECT_SIMD_REAL_NEAR(ref, rcp(x));
365 TEST_F(SimdFloatingpointTest, maskzRcp)
367 SimdReal x = setSimdRealFrom3R(M_PI, 0.0, -1234567890.0);
368 SimdBool m = (x != setZero());
369 SimdReal ref = setSimdRealFrom3R(1.0 / M_PI, 0.0, -1.0 / 1234567890.0);
370 int shiftbits = std::numeric_limits<real>::digits - GMX_SIMD_RCP_BITS;
377 /* Set the allowed ulp error as 2 to the power of the number of bits in
378 * the mantissa that do not have to be correct after the table lookup.
380 setUlpTol(1LL << shiftbits);
381 GMX_EXPECT_SIMD_REAL_NEAR(ref, maskzRcp(x, m));
384 TEST_F(SimdFloatingpointTest, cmpEqAndSelectByMask)
386 SimdBool eq = rSimd_c4c6c8 == rSimd_c6c7c8;
387 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(0, 0, c2), selectByMask(rSimd_c0c1c2, eq));
390 TEST_F(SimdFloatingpointTest, selectByNotMask)
392 SimdBool eq = rSimd_c4c6c8 == rSimd_c6c7c8;
393 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, c1, 0), selectByNotMask(rSimd_c0c1c2, eq));
396 TEST_F(SimdFloatingpointTest, cmpNe)
398 SimdBool eq = rSimd_c4c6c8 != rSimd_c6c7c8;
399 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, c1, 0), selectByMask(rSimd_c0c1c2, eq));
402 TEST_F(SimdFloatingpointTest, cmpLe)
404 SimdBool le = rSimd_c4c6c8 <= rSimd_c6c7c8;
405 GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, selectByMask(rSimd_c0c1c2, le));
408 TEST_F(SimdFloatingpointTest, cmpLt)
410 SimdBool lt = rSimd_c4c6c8 < rSimd_c6c7c8;
411 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, c1, 0), selectByMask(rSimd_c0c1c2, lt));
414 # if GMX_SIMD_HAVE_INT32_LOGICAL || GMX_SIMD_HAVE_LOGICAL
415 TEST_F(SimdFloatingpointTest, testBits)
417 SimdBool eq = testBits(setSimdRealFrom3R(c1, 0, c1));
418 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c0, 0, c2), selectByMask(rSimd_c0c1c2, eq));
420 // Test if we detect only the sign bit being set
421 eq = testBits(setSimdRealFrom1R(GMX_REAL_NEGZERO));
422 GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, selectByMask(rSimd_c0c1c2, eq));
426 TEST_F(SimdFloatingpointTest, andB)
428 SimdBool eq = rSimd_c4c6c8 == rSimd_c6c7c8;
429 SimdBool le = rSimd_c4c6c8 <= rSimd_c6c7c8;
430 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(0, 0, c2), selectByMask(rSimd_c0c1c2, (eq && le)));
433 TEST_F(SimdFloatingpointTest, orB)
435 SimdBool eq = rSimd_c4c6c8 == rSimd_c6c7c8;
436 SimdBool lt = rSimd_c4c6c8 < rSimd_c6c7c8;
437 GMX_EXPECT_SIMD_REAL_EQ(rSimd_c0c1c2, selectByMask(rSimd_c0c1c2, (eq || lt)));
440 TEST_F(SimdFloatingpointTest, anyTrueB)
442 alignas(GMX_SIMD_ALIGNMENT) std::array<real, GMX_SIMD_REAL_WIDTH> mem{};
444 // Test the false case
445 EXPECT_FALSE(anyTrue(setZero() < load<SimdReal>(mem.data())));
447 // Test each bit (these should all be true)
448 for (int i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
452 EXPECT_TRUE(anyTrue(setZero() < load<SimdReal>(mem.data())))
453 << "Not detecting true in element " << i;
457 TEST_F(SimdFloatingpointTest, blend)
459 SimdBool lt = rSimd_c4c6c8 < rSimd_c6c7c8;
460 GMX_EXPECT_SIMD_REAL_EQ(setSimdRealFrom3R(c3, c4, c2), blend(rSimd_c0c1c2, rSimd_c3c4c5, lt));
463 TEST_F(SimdFloatingpointTest, reduce)
465 // The horizontal sum of the SIMD variable depends on the width, so
466 // simply store it an extra time and calculate what the sum should be
467 std::vector<real> v = simdReal2Vector(rSimd_c3c4c5);
470 for (int i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
475 EXPECT_REAL_EQ_TOL(sum, reduce(rSimd_c3c4c5), defaultRealTolerance());
478 # endif // GMX_SIMD_HAVE_REAL
480 # if GMX_SIMD_HAVE_FLOAT && GMX_SIMD_HAVE_DOUBLE
481 TEST_F(SimdFloatingpointTest, cvtFloat2Double)
483 alignas(GMX_SIMD_ALIGNMENT) float f[GMX_SIMD_FLOAT_WIDTH];
484 alignas(GMX_SIMD_ALIGNMENT) double d[GMX_SIMD_FLOAT_WIDTH]; // Yes, double array length should be same as float
489 FloatingPointTolerance tolerance(defaultRealTolerance());
491 for (i = 0; i < GMX_SIMD_FLOAT_WIDTH; i++)
493 // Scale by 1+100*eps to use low bits too.
494 // Due to the conversions we want to avoid being too sensitive to fluctuations in last bit
495 f[i] = i * (1.0 + 100 * GMX_FLOAT_EPS);
498 vf = load<SimdFloat>(f);
499 # if (GMX_SIMD_FLOAT_WIDTH == 2 * GMX_SIMD_DOUBLE_WIDTH)
501 cvtF2DD(vf, &vd0, &vd1);
502 store(d + GMX_SIMD_DOUBLE_WIDTH, vd1); // Store upper part halfway through array
503 # elif (GMX_SIMD_FLOAT_WIDTH == GMX_SIMD_DOUBLE_WIDTH)
506 # error Width of float SIMD must either be identical to double, or twice the width.
508 store(d, vd0); // store lower (or whole) part from start of vector
510 for (i = 0; i < GMX_SIMD_FLOAT_WIDTH; i++)
512 EXPECT_REAL_EQ_TOL(f[i], d[i], tolerance);
516 TEST_F(SimdFloatingpointTest, cvtDouble2Float)
518 alignas(GMX_SIMD_ALIGNMENT) float f[GMX_SIMD_FLOAT_WIDTH];
519 alignas(GMX_SIMD_ALIGNMENT) double d[GMX_SIMD_FLOAT_WIDTH]; // Yes, double array length should be same as float
523 FloatingPointTolerance tolerance(defaultRealTolerance());
525 // This fills elements for pd1 too when double width is 2*single width
526 for (i = 0; i < GMX_SIMD_FLOAT_WIDTH; i++)
528 // Scale by 1+eps to use low bits too.
529 // Due to the conversions we want to avoid being too sensitive to fluctuations in last bit
530 d[i] = i * (1.0 + 100 * GMX_FLOAT_EPS);
533 vd0 = load<SimdDouble>(d);
534 # if (GMX_SIMD_FLOAT_WIDTH == 2 * GMX_SIMD_DOUBLE_WIDTH)
535 SimdDouble vd1 = load<SimdDouble>(d + GMX_SIMD_DOUBLE_WIDTH); // load upper half of data
536 vf = cvtDD2F(vd0, vd1);
537 # elif (GMX_SIMD_FLOAT_WIDTH == GMX_SIMD_DOUBLE_WIDTH)
540 # error Width of float SIMD must either be identical to double, or twice the width.
544 // This will check elements in pd1 too when double width is 2*single width
545 for (i = 0; i < GMX_SIMD_FLOAT_WIDTH; i++)
547 EXPECT_FLOAT_EQ_TOL(d[i], f[i], tolerance);
550 # endif // GMX_SIMD_HAVE_FLOAT && GMX_SIMD_HAVE_DOUBLE