/* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
*
- *
+ *
* This file is part of GROMACS.
- * Copyright (c) 2012-
+ * Copyright (c) 2012-
*
* Written by the Gromacs development team under coordination of
* David van der Spoel, Berk Hess, and Erik Lindahl.
*
* To help us fund GROMACS development, we humbly ask that you cite
* the research papers on the package. Check out http://www.gromacs.org
- *
+ *
* And Hey:
* Gnomes, ROck Monsters And Chili Sauce
*/
const __m128 half = _mm_set1_ps(0.5);
const __m128 one = _mm_set1_ps(1.0);
- __m128 lu = _mm_rsqrt_ps(x);
+ __m128 lu = _mm_rsqrt_ps(x);
- return _mm_macc_ps(_mm_nmacc_ps(x,_mm_mul_ps(lu,lu),one),_mm_mul_ps(lu,half),lu);
+ return _mm_macc_ps(_mm_nmacc_ps(x, _mm_mul_ps(lu, lu), one), _mm_mul_ps(lu, half), lu);
}
/* sqrt(x) - Do NOT use this (but rather invsqrt) if you actually need 1.0/sqrt(x) */
__m128 mask;
__m128 res;
- mask = _mm_cmp_ps(x,_mm_setzero_ps(),_CMP_EQ_OQ);
- res = _mm_andnot_ps(mask,gmx_mm_invsqrt_ps(x));
+ mask = _mm_cmp_ps(x, _mm_setzero_ps(), _CMP_EQ_OQ);
+ res = _mm_andnot_ps(mask, gmx_mm_invsqrt_ps(x));
- res = _mm_mul_ps(x,res);
+ res = _mm_mul_ps(x, res);
return res;
}
{
const __m128 two = _mm_set1_ps(2.0);
- __m128 lu = _mm_rcp_ps(x);
+ __m128 lu = _mm_rcp_ps(x);
- return _mm_mul_ps(lu,_mm_nmacc_ps(lu,x,two));
+ return _mm_mul_ps(lu, _mm_nmacc_ps(lu, x, two));
}
static gmx_inline __m128
{
const __m128 signmask = gmx_mm_castsi128_ps( _mm_set1_epi32(0x7FFFFFFF) );
- return _mm_and_ps(x,signmask);
+ return _mm_and_ps(x, signmask);
}
static __m128
const __m128 corr1 = _mm_set1_ps(-2.12194440e-4f);
const __m128 corr2 = _mm_set1_ps(0.693359375f);
- const __m128 CA_1 = _mm_set1_ps(0.070376836292f);
- const __m128 CB_0 = _mm_set1_ps(1.6714950086782716f);
- const __m128 CB_1 = _mm_set1_ps(-2.452088066061482f);
- const __m128 CC_0 = _mm_set1_ps(1.5220770854701728f);
- const __m128 CC_1 = _mm_set1_ps(-1.3422238433233642f);
- const __m128 CD_0 = _mm_set1_ps(1.386218787509749f);
- const __m128 CD_1 = _mm_set1_ps(0.35075468953796346f);
- const __m128 CE_0 = _mm_set1_ps(1.3429983063133937f);
- const __m128 CE_1 = _mm_set1_ps(1.807420826584643f);
-
- __m128 fexp,fexp1;
- __m128i iexp;
- __m128 mask;
- __m128 x1,x2;
- __m128 y;
- __m128 pA,pB,pC,pD,pE,tB,tC,tD,tE;
+ const __m128 CA_1 = _mm_set1_ps(0.070376836292f);
+ const __m128 CB_0 = _mm_set1_ps(1.6714950086782716f);
+ const __m128 CB_1 = _mm_set1_ps(-2.452088066061482f);
+ const __m128 CC_0 = _mm_set1_ps(1.5220770854701728f);
+ const __m128 CC_1 = _mm_set1_ps(-1.3422238433233642f);
+ const __m128 CD_0 = _mm_set1_ps(1.386218787509749f);
+ const __m128 CD_1 = _mm_set1_ps(0.35075468953796346f);
+ const __m128 CE_0 = _mm_set1_ps(1.3429983063133937f);
+ const __m128 CE_1 = _mm_set1_ps(1.807420826584643f);
+
+ __m128 fexp, fexp1;
+ __m128i iexp;
+ __m128 mask;
+ __m128 x1, x2;
+ __m128 y;
+ __m128 pA, pB, pC, pD, pE, tB, tC, tD, tE;
/* Separate x into exponent and mantissa, with a mantissa in the range [0.5..1[ (not IEEE754 standard!) */
- fexp = _mm_and_ps(x,expmask);
+ fexp = _mm_and_ps(x, expmask);
iexp = gmx_mm_castps_si128(fexp);
- iexp = _mm_srli_epi32(iexp,23);
- iexp = _mm_sub_epi32(iexp,expbase_m1);
+ iexp = _mm_srli_epi32(iexp, 23);
+ iexp = _mm_sub_epi32(iexp, expbase_m1);
- x = _mm_andnot_ps(expmask,x);
- x = _mm_or_ps(x,one);
- x = _mm_mul_ps(x,half);
+ x = _mm_andnot_ps(expmask, x);
+ x = _mm_or_ps(x, one);
+ x = _mm_mul_ps(x, half);
- mask = _mm_cmp_ps(x,invsq2,_CMP_LT_OQ);
+ mask = _mm_cmp_ps(x, invsq2, _CMP_LT_OQ);
- x = _mm_add_ps(x,_mm_and_ps(mask,x));
- x = _mm_sub_ps(x,one);
- iexp = _mm_add_epi32(iexp,gmx_mm_castps_si128(mask)); /* 0xFFFFFFFF = -1 as int */
+ x = _mm_add_ps(x, _mm_and_ps(mask, x));
+ x = _mm_sub_ps(x, one);
+ iexp = _mm_add_epi32(iexp, gmx_mm_castps_si128(mask)); /* 0xFFFFFFFF = -1 as int */
- x2 = _mm_mul_ps(x,x);
+ x2 = _mm_mul_ps(x, x);
- pA = _mm_mul_ps(CA_1,x);
+ pA = _mm_mul_ps(CA_1, x);
- pB = _mm_add_ps(x,CB_1);
- pC = _mm_add_ps(x,CC_1);
- pD = _mm_add_ps(x,CD_1);
- pE = _mm_add_ps(x,CE_1);
+ pB = _mm_add_ps(x, CB_1);
+ pC = _mm_add_ps(x, CC_1);
+ pD = _mm_add_ps(x, CD_1);
+ pE = _mm_add_ps(x, CE_1);
- pB = _mm_macc_ps(x,pB,CB_0);
- pC = _mm_macc_ps(x,pC,CC_0);
- pD = _mm_macc_ps(x,pD,CD_0);
- pE = _mm_macc_ps(x,pE,CE_0);
+ pB = _mm_macc_ps(x, pB, CB_0);
+ pC = _mm_macc_ps(x, pC, CC_0);
+ pD = _mm_macc_ps(x, pD, CD_0);
+ pE = _mm_macc_ps(x, pE, CE_0);
- pA = _mm_mul_ps(pA,pB);
- pC = _mm_mul_ps(pC,pD);
- pE = _mm_mul_ps(pE,x2);
- pA = _mm_mul_ps(pA,pC);
- y = _mm_mul_ps(pA,pE);
+ pA = _mm_mul_ps(pA, pB);
+ pC = _mm_mul_ps(pC, pD);
+ pE = _mm_mul_ps(pE, x2);
+ pA = _mm_mul_ps(pA, pC);
+ y = _mm_mul_ps(pA, pE);
fexp = _mm_cvtepi32_ps(iexp);
- y = _mm_macc_ps(fexp,corr1,y);
- y = _mm_nmacc_ps(half,x2,y);
+ y = _mm_macc_ps(fexp, corr1, y);
+ y = _mm_nmacc_ps(half, x2, y);
- x2 = _mm_add_ps(x,y);
- x2 = _mm_macc_ps(fexp,corr2,x2);
+ x2 = _mm_add_ps(x, y);
+ x2 = _mm_macc_ps(fexp, corr2, x2);
return x2;
}
const __m128 CA1 = _mm_set1_ps(6.931472028550421E-001);
const __m128 CA0 = _mm_set1_ps(1.0f);
- __m128 valuemask;
- __m128i iexppart;
- __m128 fexppart;
- __m128 intpart;
- __m128 x2;
- __m128 p0,p1;
+ __m128 valuemask;
+ __m128i iexppart;
+ __m128 fexppart;
+ __m128 intpart;
+ __m128 x2;
+ __m128 p0, p1;
iexppart = _mm_cvtps_epi32(x);
- intpart = _mm_round_ps(x,_MM_FROUND_TO_NEAREST_INT);
- iexppart = _mm_slli_epi32(_mm_add_epi32(iexppart,expbase),23);
- valuemask = _mm_cmp_ps(arglimit,gmx_mm_abs_ps(x),_CMP_GE_OQ);
- fexppart = _mm_and_ps(valuemask,gmx_mm_castsi128_ps(iexppart));
-
- x = _mm_sub_ps(x,intpart);
- x2 = _mm_mul_ps(x,x);
-
- p0 = _mm_macc_ps(CA6,x2,CA4);
- p1 = _mm_macc_ps(CA5,x2,CA3);
- p0 = _mm_macc_ps(p0,x2,CA2);
- p1 = _mm_macc_ps(p1,x2,CA1);
- p0 = _mm_macc_ps(p0,x2,CA0);
- p0 = _mm_macc_ps(p1,x,p0);
- x = _mm_mul_ps(p0,fexppart);
-
- return x;
+ intpart = _mm_round_ps(x, _MM_FROUND_TO_NEAREST_INT);
+ iexppart = _mm_slli_epi32(_mm_add_epi32(iexppart, expbase), 23);
+ valuemask = _mm_cmp_ps(arglimit, gmx_mm_abs_ps(x), _CMP_GE_OQ);
+ fexppart = _mm_and_ps(valuemask, gmx_mm_castsi128_ps(iexppart));
+
+ x = _mm_sub_ps(x, intpart);
+ x2 = _mm_mul_ps(x, x);
+
+ p0 = _mm_macc_ps(CA6, x2, CA4);
+ p1 = _mm_macc_ps(CA5, x2, CA3);
+ p0 = _mm_macc_ps(p0, x2, CA2);
+ p1 = _mm_macc_ps(p1, x2, CA1);
+ p0 = _mm_macc_ps(p0, x2, CA0);
+ p0 = _mm_macc_ps(p1, x, p0);
+ x = _mm_mul_ps(p0, fexppart);
+
+ return x;
}
const __m128 CC0 = _mm_set1_ps(5.0000001201e-1f);
const __m128 one = _mm_set1_ps(1.0f);
- __m128 y,x2;
- __m128 p0,p1;
- __m128 valuemask;
- __m128i iexppart;
- __m128 fexppart;
- __m128 intpart;
+ __m128 y, x2;
+ __m128 p0, p1;
+ __m128 valuemask;
+ __m128i iexppart;
+ __m128 fexppart;
+ __m128 intpart;
- y = _mm_mul_ps(x,argscale);
+ y = _mm_mul_ps(x, argscale);
iexppart = _mm_cvtps_epi32(y);
- intpart = _mm_round_ps(y,_MM_FROUND_TO_NEAREST_INT);
+ intpart = _mm_round_ps(y, _MM_FROUND_TO_NEAREST_INT);
- iexppart = _mm_slli_epi32(_mm_add_epi32(iexppart,expbase),23);
- valuemask = _mm_cmp_ps(arglimit,gmx_mm_abs_ps(y),_CMP_GE_OQ);
- fexppart = _mm_and_ps(valuemask,gmx_mm_castsi128_ps(iexppart));
+ iexppart = _mm_slli_epi32(_mm_add_epi32(iexppart, expbase), 23);
+ valuemask = _mm_cmp_ps(arglimit, gmx_mm_abs_ps(y), _CMP_GE_OQ);
+ fexppart = _mm_and_ps(valuemask, gmx_mm_castsi128_ps(iexppart));
/* Extended precision arithmetics */
- x = _mm_nmacc_ps(invargscale0,intpart,x);
- x = _mm_nmacc_ps(invargscale1,intpart,x);
+ x = _mm_nmacc_ps(invargscale0, intpart, x);
+ x = _mm_nmacc_ps(invargscale1, intpart, x);
- x2 = _mm_mul_ps(x,x);
+ x2 = _mm_mul_ps(x, x);
- p1 = _mm_macc_ps(CC5,x2,CC3);
- p0 = _mm_macc_ps(CC4,x2,CC2);
- p1 = _mm_macc_ps(p1,x2,CC1);
- p0 = _mm_macc_ps(p0,x2,CC0);
- p0 = _mm_macc_ps(p1,x,p0);
- p0 = _mm_macc_ps(p0,x2,one);
+ p1 = _mm_macc_ps(CC5, x2, CC3);
+ p0 = _mm_macc_ps(CC4, x2, CC2);
+ p1 = _mm_macc_ps(p1, x2, CC1);
+ p0 = _mm_macc_ps(p0, x2, CC0);
+ p0 = _mm_macc_ps(p1, x, p0);
+ p0 = _mm_macc_ps(p0, x2, one);
- x = _mm_add_ps(x,p0);
+ x = _mm_add_ps(x, p0);
- x = _mm_mul_ps(x,fexppart);
+ x = _mm_mul_ps(x, fexppart);
return x;
}
const __m128 one = _mm_set1_ps(1.0f);
const __m128 two = _mm_set1_ps(2.0f);
- __m128 x2,x4,y;
- __m128 z,q,t,t2,w,w2;
- __m128 pA0,pA1,pB0,pB1,pC0,pC1;
- __m128 expmx2,corr;
- __m128 res_erf,res_erfc,res;
- __m128 mask;
+ __m128 x2, x4, y;
+ __m128 z, q, t, t2, w, w2;
+ __m128 pA0, pA1, pB0, pB1, pC0, pC1;
+ __m128 expmx2, corr;
+ __m128 res_erf, res_erfc, res;
+ __m128 mask;
/* Calculate erf() */
- x2 = _mm_mul_ps(x,x);
- x4 = _mm_mul_ps(x2,x2);
-
- pA0 = _mm_macc_ps(CA6,x4,CA4);
- pA1 = _mm_macc_ps(CA5,x4,CA3);
- pA0 = _mm_macc_ps(pA0,x4,CA2);
- pA1 = _mm_macc_ps(pA1,x4,CA1);
- pA0 = _mm_mul_ps(pA0,x4);
- pA0 = _mm_macc_ps(pA1,x2,pA0);
+ x2 = _mm_mul_ps(x, x);
+ x4 = _mm_mul_ps(x2, x2);
+
+ pA0 = _mm_macc_ps(CA6, x4, CA4);
+ pA1 = _mm_macc_ps(CA5, x4, CA3);
+ pA0 = _mm_macc_ps(pA0, x4, CA2);
+ pA1 = _mm_macc_ps(pA1, x4, CA1);
+ pA0 = _mm_mul_ps(pA0, x4);
+ pA0 = _mm_macc_ps(pA1, x2, pA0);
/* Constant term must come last for precision reasons */
- pA0 = _mm_add_ps(pA0,CA0);
+ pA0 = _mm_add_ps(pA0, CA0);
- res_erf = _mm_mul_ps(x,pA0);
+ res_erf = _mm_mul_ps(x, pA0);
/* Calculate erfc */
y = gmx_mm_abs_ps(x);
t = gmx_mm_inv_ps(y);
- w = _mm_sub_ps(t,one);
- t2 = _mm_mul_ps(t,t);
- w2 = _mm_mul_ps(w,w);
+ w = _mm_sub_ps(t, one);
+ t2 = _mm_mul_ps(t, t);
+ w2 = _mm_mul_ps(w, w);
/*
* We cannot simply calculate exp(-x2) directly in single precision, since
* that will lose a couple of bits of precision due to the multiplication.
* low-order minimax expansion on [0,0.1].
*/
- z = _mm_and_ps(y,sieve);
- q = _mm_mul_ps( _mm_sub_ps(z,y) , _mm_add_ps(z,y) );
-
- corr = _mm_macc_ps(CD4,q,CD3);
- corr = _mm_macc_ps(corr,q,CD2);
- corr = _mm_macc_ps(corr,q,one);
- corr = _mm_macc_ps(corr,q,one);
-
- expmx2 = gmx_mm_exp_ps( _mm_or_ps( signbit , _mm_mul_ps(z,z) ) );
- expmx2 = _mm_mul_ps(expmx2,corr);
-
- pB1 = _mm_macc_ps(CB9,w2,CB7);
- pB0 = _mm_macc_ps(CB8,w2,CB6);
- pB1 = _mm_macc_ps(pB1,w2,CB5);
- pB0 = _mm_macc_ps(pB0,w2,CB4);
- pB1 = _mm_macc_ps(pB1,w2,CB3);
- pB0 = _mm_macc_ps(pB0,w2,CB2);
- pB1 = _mm_macc_ps(pB1,w2,CB1);
- pB0 = _mm_macc_ps(pB0,w2,CB0);
- pB0 = _mm_macc_ps(pB1,w,pB0);
-
- pC0 = _mm_macc_ps(CC10,t2,CC8);
- pC1 = _mm_macc_ps(CC9,t2,CC7);
- pC0 = _mm_macc_ps(pC0,t2,CC6);
- pC1 = _mm_macc_ps(pC1,t2,CC5);
- pC0 = _mm_macc_ps(pC0,t2,CC4);
- pC1 = _mm_macc_ps(pC1,t2,CC3);
- pC0 = _mm_macc_ps(pC0,t2,CC2);
- pC1 = _mm_macc_ps(pC1,t2,CC1);
-
- pC0 = _mm_macc_ps(pC0,t2,CC0);
- pC0 = _mm_macc_ps(pC1,t,pC0);
- pC0 = _mm_mul_ps(pC0,t);
+ z = _mm_and_ps(y, sieve);
+ q = _mm_mul_ps( _mm_sub_ps(z, y), _mm_add_ps(z, y) );
+
+ corr = _mm_macc_ps(CD4, q, CD3);
+ corr = _mm_macc_ps(corr, q, CD2);
+ corr = _mm_macc_ps(corr, q, one);
+ corr = _mm_macc_ps(corr, q, one);
+
+ expmx2 = gmx_mm_exp_ps( _mm_or_ps( signbit, _mm_mul_ps(z, z) ) );
+ expmx2 = _mm_mul_ps(expmx2, corr);
+
+ pB1 = _mm_macc_ps(CB9, w2, CB7);
+ pB0 = _mm_macc_ps(CB8, w2, CB6);
+ pB1 = _mm_macc_ps(pB1, w2, CB5);
+ pB0 = _mm_macc_ps(pB0, w2, CB4);
+ pB1 = _mm_macc_ps(pB1, w2, CB3);
+ pB0 = _mm_macc_ps(pB0, w2, CB2);
+ pB1 = _mm_macc_ps(pB1, w2, CB1);
+ pB0 = _mm_macc_ps(pB0, w2, CB0);
+ pB0 = _mm_macc_ps(pB1, w, pB0);
+
+ pC0 = _mm_macc_ps(CC10, t2, CC8);
+ pC1 = _mm_macc_ps(CC9, t2, CC7);
+ pC0 = _mm_macc_ps(pC0, t2, CC6);
+ pC1 = _mm_macc_ps(pC1, t2, CC5);
+ pC0 = _mm_macc_ps(pC0, t2, CC4);
+ pC1 = _mm_macc_ps(pC1, t2, CC3);
+ pC0 = _mm_macc_ps(pC0, t2, CC2);
+ pC1 = _mm_macc_ps(pC1, t2, CC1);
+
+ pC0 = _mm_macc_ps(pC0, t2, CC0);
+ pC0 = _mm_macc_ps(pC1, t, pC0);
+ pC0 = _mm_mul_ps(pC0, t);
/* SELECT pB0 or pC0 for erfc() */
- mask = _mm_cmp_ps(two,y,_CMP_LT_OQ);
- res_erfc = _mm_blendv_ps(pB0,pC0,mask);
- res_erfc = _mm_mul_ps(res_erfc,expmx2);
+ mask = _mm_cmp_ps(two, y, _CMP_LT_OQ);
+ res_erfc = _mm_blendv_ps(pB0, pC0, mask);
+ res_erfc = _mm_mul_ps(res_erfc, expmx2);
/* erfc(x<0) = 2-erfc(|x|) */
- mask = _mm_cmp_ps(x,_mm_setzero_ps(),_CMP_LT_OQ);
- res_erfc = _mm_blendv_ps(res_erfc,_mm_sub_ps(two,res_erfc),mask);
+ mask = _mm_cmp_ps(x, _mm_setzero_ps(), _CMP_LT_OQ);
+ res_erfc = _mm_blendv_ps(res_erfc, _mm_sub_ps(two, res_erfc), mask);
/* Select erf() or erfc() */
- mask = _mm_cmp_ps(y,_mm_set1_ps(0.75f),_CMP_LT_OQ);
- res = _mm_blendv_ps(_mm_sub_ps(one,res_erfc),res_erf,mask);
+ mask = _mm_cmp_ps(y, _mm_set1_ps(0.75f), _CMP_LT_OQ);
+ res = _mm_blendv_ps(_mm_sub_ps(one, res_erfc), res_erf, mask);
return res;
}
const __m128 one = _mm_set1_ps(1.0f);
const __m128 two = _mm_set1_ps(2.0f);
- __m128 x2,x4,y;
- __m128 z,q,t,t2,w,w2;
- __m128 pA0,pA1,pB0,pB1,pC0,pC1;
- __m128 expmx2,corr;
- __m128 res_erf,res_erfc,res;
- __m128 mask;
+ __m128 x2, x4, y;
+ __m128 z, q, t, t2, w, w2;
+ __m128 pA0, pA1, pB0, pB1, pC0, pC1;
+ __m128 expmx2, corr;
+ __m128 res_erf, res_erfc, res;
+ __m128 mask;
/* Calculate erf() */
- x2 = _mm_mul_ps(x,x);
- x4 = _mm_mul_ps(x2,x2);
-
- pA0 = _mm_macc_ps(CA6,x4,CA4);
- pA1 = _mm_macc_ps(CA5,x4,CA3);
- pA0 = _mm_macc_ps(pA0,x4,CA2);
- pA1 = _mm_macc_ps(pA1,x4,CA1);
- pA1 = _mm_mul_ps(pA1,x2);
- pA0 = _mm_macc_ps(pA0,x4,pA1);
+ x2 = _mm_mul_ps(x, x);
+ x4 = _mm_mul_ps(x2, x2);
+
+ pA0 = _mm_macc_ps(CA6, x4, CA4);
+ pA1 = _mm_macc_ps(CA5, x4, CA3);
+ pA0 = _mm_macc_ps(pA0, x4, CA2);
+ pA1 = _mm_macc_ps(pA1, x4, CA1);
+ pA1 = _mm_mul_ps(pA1, x2);
+ pA0 = _mm_macc_ps(pA0, x4, pA1);
/* Constant term must come last for precision reasons */
- pA0 = _mm_add_ps(pA0,CA0);
+ pA0 = _mm_add_ps(pA0, CA0);
- res_erf = _mm_mul_ps(x,pA0);
+ res_erf = _mm_mul_ps(x, pA0);
/* Calculate erfc */
y = gmx_mm_abs_ps(x);
t = gmx_mm_inv_ps(y);
- w = _mm_sub_ps(t,one);
- t2 = _mm_mul_ps(t,t);
- w2 = _mm_mul_ps(w,w);
+ w = _mm_sub_ps(t, one);
+ t2 = _mm_mul_ps(t, t);
+ w2 = _mm_mul_ps(w, w);
/*
* We cannot simply calculate exp(-x2) directly in single precision, since
* that will lose a couple of bits of precision due to the multiplication.
* low-order minimax expansion on [0,0.1].
*/
- z = _mm_and_ps(y,sieve);
- q = _mm_mul_ps( _mm_sub_ps(z,y) , _mm_add_ps(z,y) );
-
- corr = _mm_macc_ps(CD4,q,CD3);
- corr = _mm_macc_ps(corr,q,CD2);
- corr = _mm_macc_ps(corr,q,one);
- corr = _mm_macc_ps(corr,q,one);
-
- expmx2 = gmx_mm_exp_ps( _mm_or_ps( signbit , _mm_mul_ps(z,z) ) );
- expmx2 = _mm_mul_ps(expmx2,corr);
-
- pB1 = _mm_macc_ps(CB9,w2,CB7);
- pB0 = _mm_macc_ps(CB8,w2,CB6);
- pB1 = _mm_macc_ps(pB1,w2,CB5);
- pB0 = _mm_macc_ps(pB0,w2,CB4);
- pB1 = _mm_macc_ps(pB1,w2,CB3);
- pB0 = _mm_macc_ps(pB0,w2,CB2);
- pB1 = _mm_macc_ps(pB1,w2,CB1);
- pB0 = _mm_macc_ps(pB0,w2,CB0);
- pB0 = _mm_macc_ps(pB1,w,pB0);
-
- pC0 = _mm_macc_ps(CC10,t2,CC8);
- pC1 = _mm_macc_ps(CC9,t2,CC7);
- pC0 = _mm_macc_ps(pC0,t2,CC6);
- pC1 = _mm_macc_ps(pC1,t2,CC5);
- pC0 = _mm_macc_ps(pC0,t2,CC4);
- pC1 = _mm_macc_ps(pC1,t2,CC3);
- pC0 = _mm_macc_ps(pC0,t2,CC2);
- pC1 = _mm_macc_ps(pC1,t2,CC1);
-
- pC0 = _mm_macc_ps(pC0,t2,CC0);
- pC0 = _mm_macc_ps(pC1,t,pC0);
- pC0 = _mm_mul_ps(pC0,t);
+ z = _mm_and_ps(y, sieve);
+ q = _mm_mul_ps( _mm_sub_ps(z, y), _mm_add_ps(z, y) );
+
+ corr = _mm_macc_ps(CD4, q, CD3);
+ corr = _mm_macc_ps(corr, q, CD2);
+ corr = _mm_macc_ps(corr, q, one);
+ corr = _mm_macc_ps(corr, q, one);
+
+ expmx2 = gmx_mm_exp_ps( _mm_or_ps( signbit, _mm_mul_ps(z, z) ) );
+ expmx2 = _mm_mul_ps(expmx2, corr);
+
+ pB1 = _mm_macc_ps(CB9, w2, CB7);
+ pB0 = _mm_macc_ps(CB8, w2, CB6);
+ pB1 = _mm_macc_ps(pB1, w2, CB5);
+ pB0 = _mm_macc_ps(pB0, w2, CB4);
+ pB1 = _mm_macc_ps(pB1, w2, CB3);
+ pB0 = _mm_macc_ps(pB0, w2, CB2);
+ pB1 = _mm_macc_ps(pB1, w2, CB1);
+ pB0 = _mm_macc_ps(pB0, w2, CB0);
+ pB0 = _mm_macc_ps(pB1, w, pB0);
+
+ pC0 = _mm_macc_ps(CC10, t2, CC8);
+ pC1 = _mm_macc_ps(CC9, t2, CC7);
+ pC0 = _mm_macc_ps(pC0, t2, CC6);
+ pC1 = _mm_macc_ps(pC1, t2, CC5);
+ pC0 = _mm_macc_ps(pC0, t2, CC4);
+ pC1 = _mm_macc_ps(pC1, t2, CC3);
+ pC0 = _mm_macc_ps(pC0, t2, CC2);
+ pC1 = _mm_macc_ps(pC1, t2, CC1);
+
+ pC0 = _mm_macc_ps(pC0, t2, CC0);
+ pC0 = _mm_macc_ps(pC1, t, pC0);
+ pC0 = _mm_mul_ps(pC0, t);
/* SELECT pB0 or pC0 for erfc() */
- mask = _mm_cmp_ps(two,y,_CMP_LT_OQ);
- res_erfc = _mm_blendv_ps(pB0,pC0,mask);
- res_erfc = _mm_mul_ps(res_erfc,expmx2);
+ mask = _mm_cmp_ps(two, y, _CMP_LT_OQ);
+ res_erfc = _mm_blendv_ps(pB0, pC0, mask);
+ res_erfc = _mm_mul_ps(res_erfc, expmx2);
/* erfc(x<0) = 2-erfc(|x|) */
- mask = _mm_cmp_ps(x,_mm_setzero_ps(),_CMP_LT_OQ);
- res_erfc = _mm_blendv_ps(res_erfc,_mm_sub_ps(two,res_erfc),mask);
+ mask = _mm_cmp_ps(x, _mm_setzero_ps(), _CMP_LT_OQ);
+ res_erfc = _mm_blendv_ps(res_erfc, _mm_sub_ps(two, res_erfc), mask);
/* Select erf() or erfc() */
- mask = _mm_cmp_ps(y,_mm_set1_ps(0.75f),_CMP_LT_OQ);
- res = _mm_blendv_ps(res_erfc,_mm_sub_ps(one,res_erf),mask);
+ mask = _mm_cmp_ps(y, _mm_set1_ps(0.75f), _CMP_LT_OQ);
+ res = _mm_blendv_ps(res_erfc, _mm_sub_ps(one, res_erf), mask);
return res;
}
/* Calculate the force correction due to PME analytically.
*
* This routine is meant to enable analytical evaluation of the
- * direct-space PME electrostatic force to avoid tables.
+ * direct-space PME electrostatic force to avoid tables.
*
* The direct-space potential should be Erfc(beta*r)/r, but there
- * are some problems evaluating that:
+ * are some problems evaluating that:
*
- * First, the error function is difficult (read: expensive) to
+ * First, the error function is difficult (read: expensive) to
* approxmiate accurately for intermediate to large arguments, and
* this happens already in ranges of beta*r that occur in simulations.
* Second, we now try to avoid calculating potentials in Gromacs but
* V= 1/r - Erf(beta*r)/r
*
* The first term we already have from the inverse square root, so
- * that we can leave out of this routine.
+ * that we can leave out of this routine.
*
* For pme tolerances of 1e-3 to 1e-8 and cutoffs of 0.5nm to 1.8nm,
- * the argument beta*r will be in the range 0.15 to ~4. Use your
+ * the argument beta*r will be in the range 0.15 to ~4. Use your
* favorite plotting program to realize how well-behaved Erf(z)/z is
- * in this range!
+ * in this range!
*
- * We approximate f(z)=erf(z)/z with a rational minimax polynomial.
+ * We approximate f(z)=erf(z)/z with a rational minimax polynomial.
* However, it turns out it is more efficient to approximate f(z)/z and
* then only use even powers. This is another minor optimization, since
* we actually WANT f(z)/z, because it is going to be multiplied by
- * the vector between the two atoms to get the vectorial force. The
+ * the vector between the two atoms to get the vectorial force. The
* fastest flops are the ones we can avoid calculating!
- *
+ *
* So, here's how it should be used:
*
* 1. Calculate r^2.
* 3. Evaluate this routine with z^2 as the argument.
* 4. The return value is the expression:
*
- *
+ *
* 2*exp(-z^2) erf(z)
* ------------ - --------
* sqrt(Pi)*z^2 z^3
- *
+ *
* 5. Multiply the entire expression by beta^3. This will get you
*
* beta^3*2*exp(-z^2) beta^3*erf(z)
* ------------------ - ---------------
* sqrt(Pi)*z^2 z^3
- *
+ *
* or, switching back to r (z=r*beta):
*
* 2*beta*exp(-r^2*beta^2) erf(r*beta)
const __m128 FN2 = _mm_set1_ps(-0.019278317264888380590f);
const __m128 FN1 = _mm_set1_ps(0.069670166153766424023f);
const __m128 FN0 = _mm_set1_ps(-0.75225204789749321333f);
-
+
const __m128 FD4 = _mm_set1_ps(0.0011193462567257629232f);
const __m128 FD3 = _mm_set1_ps(0.014866955030185295499f);
const __m128 FD2 = _mm_set1_ps(0.11583842382862377919f);
const __m128 FD1 = _mm_set1_ps(0.50736591960530292870f);
const __m128 FD0 = _mm_set1_ps(1.0f);
-
- __m128 z4;
- __m128 polyFN0,polyFN1,polyFD0,polyFD1;
-
- z4 = _mm_mul_ps(z2,z2);
-
- polyFD0 = _mm_macc_ps(FD4,z4,FD2);
- polyFD1 = _mm_macc_ps(FD3,z4,FD1);
- polyFD0 = _mm_macc_ps(polyFD0,z4,FD0);
- polyFD0 = _mm_macc_ps(polyFD1,z2,polyFD0);
-
+
+ __m128 z4;
+ __m128 polyFN0, polyFN1, polyFD0, polyFD1;
+
+ z4 = _mm_mul_ps(z2, z2);
+
+ polyFD0 = _mm_macc_ps(FD4, z4, FD2);
+ polyFD1 = _mm_macc_ps(FD3, z4, FD1);
+ polyFD0 = _mm_macc_ps(polyFD0, z4, FD0);
+ polyFD0 = _mm_macc_ps(polyFD1, z2, polyFD0);
+
polyFD0 = gmx_mm_inv_ps(polyFD0);
-
- polyFN0 = _mm_macc_ps(FN6,z4,FN4);
- polyFN1 = _mm_macc_ps(FN5,z4,FN3);
- polyFN0 = _mm_macc_ps(polyFN0,z4,FN2);
- polyFN1 = _mm_macc_ps(polyFN1,z4,FN1);
- polyFN0 = _mm_macc_ps(polyFN0,z4,FN0);
- polyFN0 = _mm_macc_ps(polyFN1,z2,polyFN0);
-
- return _mm_mul_ps(polyFN0,polyFD0);
+
+ polyFN0 = _mm_macc_ps(FN6, z4, FN4);
+ polyFN1 = _mm_macc_ps(FN5, z4, FN3);
+ polyFN0 = _mm_macc_ps(polyFN0, z4, FN2);
+ polyFN1 = _mm_macc_ps(polyFN1, z4, FN1);
+ polyFN0 = _mm_macc_ps(polyFN0, z4, FN0);
+ polyFN0 = _mm_macc_ps(polyFN1, z2, polyFN0);
+
+ return _mm_mul_ps(polyFN0, polyFD0);
}
* 3. Evaluate this routine with z^2 as the argument.
* 4. The return value is the expression:
*
- *
+ *
* erf(z)
* --------
* z
- *
+ *
* 5. Multiply the entire expression by beta and switching back to r (z=r*beta):
*
* erf(r*beta)
* -----------
- * r
+ * r
*
* 6. Add the result to 1/r, multiply by the product of the charges,
* and you have your potential.
const __m128 VN2 = _mm_set1_ps(0.038657983986041781264f);
const __m128 VN1 = _mm_set1_ps(0.11285044772717598220f);
const __m128 VN0 = _mm_set1_ps(1.1283802385263030286f);
-
+
const __m128 VD3 = _mm_set1_ps(0.0066752224023576045451f);
const __m128 VD2 = _mm_set1_ps(0.078647795836373922256f);
const __m128 VD1 = _mm_set1_ps(0.43336185284710920150f);
const __m128 VD0 = _mm_set1_ps(1.0f);
-
- __m128 z4;
- __m128 polyVN0,polyVN1,polyVD0,polyVD1;
-
- z4 = _mm_mul_ps(z2,z2);
-
- polyVD1 = _mm_macc_ps(VD3,z4,VD1);
- polyVD0 = _mm_macc_ps(VD2,z4,VD0);
- polyVD0 = _mm_macc_ps(polyVD1,z2,polyVD0);
-
+
+ __m128 z4;
+ __m128 polyVN0, polyVN1, polyVD0, polyVD1;
+
+ z4 = _mm_mul_ps(z2, z2);
+
+ polyVD1 = _mm_macc_ps(VD3, z4, VD1);
+ polyVD0 = _mm_macc_ps(VD2, z4, VD0);
+ polyVD0 = _mm_macc_ps(polyVD1, z2, polyVD0);
+
polyVD0 = gmx_mm_inv_ps(polyVD0);
-
- polyVN0 = _mm_macc_ps(VN6,z4,VN4);
- polyVN1 = _mm_macc_ps(VN5,z4,VN3);
- polyVN0 = _mm_macc_ps(polyVN0,z4,VN2);
- polyVN1 = _mm_macc_ps(polyVN1,z4,VN1);
- polyVN0 = _mm_macc_ps(polyVN0,z4,VN0);
- polyVN0 = _mm_macc_ps(polyVN1,z2,polyVN0);
-
- return _mm_mul_ps(polyVN0,polyVD0);
+
+ polyVN0 = _mm_macc_ps(VN6, z4, VN4);
+ polyVN1 = _mm_macc_ps(VN5, z4, VN3);
+ polyVN0 = _mm_macc_ps(polyVN0, z4, VN2);
+ polyVN1 = _mm_macc_ps(polyVN1, z4, VN1);
+ polyVN0 = _mm_macc_ps(polyVN0, z4, VN0);
+ polyVN0 = _mm_macc_ps(polyVN1, z2, polyVN0);
+
+ return _mm_mul_ps(polyVN0, polyVD0);
}
static int
-gmx_mm_sincos_ps(__m128 x,
+gmx_mm_sincos_ps(__m128 x,
__m128 *sinval,
__m128 *cosval)
{
- const __m128 two_over_pi = _mm_set1_ps(2.0/M_PI);
- const __m128 half = _mm_set1_ps(0.5);
- const __m128 one = _mm_set1_ps(1.0);
+ const __m128 two_over_pi = _mm_set1_ps(2.0/M_PI);
+ const __m128 half = _mm_set1_ps(0.5);
+ const __m128 one = _mm_set1_ps(1.0);
const __m128i izero = _mm_set1_epi32(0);
const __m128i ione = _mm_set1_epi32(1);
const __m128i ithree = _mm_set1_epi32(3);
const __m128 signbit = gmx_mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
- const __m128 CA1 = _mm_set1_ps(1.5703125f);
- const __m128 CA2 = _mm_set1_ps(4.837512969970703125e-4f);
- const __m128 CA3 = _mm_set1_ps(7.54978995489188216e-8f);
+ const __m128 CA1 = _mm_set1_ps(1.5703125f);
+ const __m128 CA2 = _mm_set1_ps(4.837512969970703125e-4f);
+ const __m128 CA3 = _mm_set1_ps(7.54978995489188216e-8f);
- const __m128 CC0 = _mm_set1_ps(-0.0013602249f);
- const __m128 CC1 = _mm_set1_ps(0.0416566950f);
- const __m128 CC2 = _mm_set1_ps(-0.4999990225f);
- const __m128 CS0 = _mm_set1_ps(-0.0001950727f);
- const __m128 CS1 = _mm_set1_ps(0.0083320758f);
- const __m128 CS2 = _mm_set1_ps(-0.1666665247f);
+ const __m128 CC0 = _mm_set1_ps(-0.0013602249f);
+ const __m128 CC1 = _mm_set1_ps(0.0416566950f);
+ const __m128 CC2 = _mm_set1_ps(-0.4999990225f);
+ const __m128 CS0 = _mm_set1_ps(-0.0001950727f);
+ const __m128 CS1 = _mm_set1_ps(0.0083320758f);
+ const __m128 CS2 = _mm_set1_ps(-0.1666665247f);
- __m128 y,y2;
- __m128 z;
- __m128i iz;
- __m128i offset_sin,offset_cos;
- __m128 tmp1,tmp2;
- __m128 mask_sin,mask_cos;
- __m128 tmp_sin,tmp_cos;
+ __m128 y, y2;
+ __m128 z;
+ __m128i iz;
+ __m128i offset_sin, offset_cos;
+ __m128 tmp1, tmp2;
+ __m128 mask_sin, mask_cos;
+ __m128 tmp_sin, tmp_cos;
- y = _mm_mul_ps(x,two_over_pi);
- y = _mm_add_ps(y,_mm_or_ps(_mm_and_ps(y,signbit),half));
+ y = _mm_mul_ps(x, two_over_pi);
+ y = _mm_add_ps(y, _mm_or_ps(_mm_and_ps(y, signbit), half));
iz = _mm_cvttps_epi32(y);
- z = _mm_round_ps(y,_MM_FROUND_TO_ZERO);
+ z = _mm_round_ps(y, _MM_FROUND_TO_ZERO);
- offset_sin = _mm_and_si128(iz,ithree);
- offset_cos = _mm_add_epi32(iz,ione);
+ offset_sin = _mm_and_si128(iz, ithree);
+ offset_cos = _mm_add_epi32(iz, ione);
/* Extended precision arithmethic to achieve full precision */
- y = _mm_nmacc_ps(z,CA1,x);
- y = _mm_nmacc_ps(z,CA2,y);
- y = _mm_nmacc_ps(z,CA3,y);
+ y = _mm_nmacc_ps(z, CA1, x);
+ y = _mm_nmacc_ps(z, CA2, y);
+ y = _mm_nmacc_ps(z, CA3, y);
- y2 = _mm_mul_ps(y,y);
+ y2 = _mm_mul_ps(y, y);
- tmp1 = _mm_macc_ps(CC0,y2,CC1);
- tmp2 = _mm_macc_ps(CS0,y2,CS1);
- tmp1 = _mm_macc_ps(tmp1,y2,CC2);
- tmp2 = _mm_macc_ps(tmp2,y2,CS2);
+ tmp1 = _mm_macc_ps(CC0, y2, CC1);
+ tmp2 = _mm_macc_ps(CS0, y2, CS1);
+ tmp1 = _mm_macc_ps(tmp1, y2, CC2);
+ tmp2 = _mm_macc_ps(tmp2, y2, CS2);
- tmp1 = _mm_macc_ps(tmp1,y2,one);
+ tmp1 = _mm_macc_ps(tmp1, y2, one);
- tmp2 = _mm_macc_ps(tmp2,_mm_mul_ps(y,y2),y);
+ tmp2 = _mm_macc_ps(tmp2, _mm_mul_ps(y, y2), y);
- mask_sin = gmx_mm_castsi128_ps(_mm_cmpeq_epi32( _mm_and_si128(offset_sin,ione), izero));
- mask_cos = gmx_mm_castsi128_ps(_mm_cmpeq_epi32( _mm_and_si128(offset_cos,ione), izero));
+ mask_sin = gmx_mm_castsi128_ps(_mm_cmpeq_epi32( _mm_and_si128(offset_sin, ione), izero));
+ mask_cos = gmx_mm_castsi128_ps(_mm_cmpeq_epi32( _mm_and_si128(offset_cos, ione), izero));
- tmp_sin = _mm_blendv_ps(tmp1,tmp2,mask_sin);
- tmp_cos = _mm_blendv_ps(tmp1,tmp2,mask_cos);
+ tmp_sin = _mm_blendv_ps(tmp1, tmp2, mask_sin);
+ tmp_cos = _mm_blendv_ps(tmp1, tmp2, mask_cos);
- mask_sin = gmx_mm_castsi128_ps(_mm_cmpeq_epi32( _mm_and_si128(offset_sin,itwo), izero));
- mask_cos = gmx_mm_castsi128_ps(_mm_cmpeq_epi32( _mm_and_si128(offset_cos,itwo), izero));
+ mask_sin = gmx_mm_castsi128_ps(_mm_cmpeq_epi32( _mm_and_si128(offset_sin, itwo), izero));
+ mask_cos = gmx_mm_castsi128_ps(_mm_cmpeq_epi32( _mm_and_si128(offset_cos, itwo), izero));
- tmp1 = _mm_xor_ps(signbit,tmp_sin);
- tmp2 = _mm_xor_ps(signbit,tmp_cos);
+ tmp1 = _mm_xor_ps(signbit, tmp_sin);
+ tmp2 = _mm_xor_ps(signbit, tmp_cos);
- *sinval = _mm_blendv_ps(tmp1,tmp_sin,mask_sin);
- *cosval = _mm_blendv_ps(tmp2,tmp_cos,mask_cos);
+ *sinval = _mm_blendv_ps(tmp1, tmp_sin, mask_sin);
+ *cosval = _mm_blendv_ps(tmp2, tmp_cos, mask_cos);
return 0;
}
static __m128
gmx_mm_sin_ps(__m128 x)
{
- __m128 s,c;
- gmx_mm_sincos_ps(x,&s,&c);
+ __m128 s, c;
+ gmx_mm_sincos_ps(x, &s, &c);
return s;
}
static __m128
gmx_mm_cos_ps(__m128 x)
{
- __m128 s,c;
- gmx_mm_sincos_ps(x,&s,&c);
+ __m128 s, c;
+ gmx_mm_sincos_ps(x, &s, &c);
return c;
}
static __m128
gmx_mm_tan_ps(__m128 x)
{
- __m128 sinval,cosval;
+ __m128 sinval, cosval;
__m128 tanval;
- gmx_mm_sincos_ps(x,&sinval,&cosval);
+ gmx_mm_sincos_ps(x, &sinval, &cosval);
- tanval = _mm_mul_ps(sinval,gmx_mm_inv_ps(cosval));
+ tanval = _mm_mul_ps(sinval, gmx_mm_inv_ps(cosval));
return tanval;
}
const __m128 CC2 = _mm_set1_ps(7.4953002686E-2f);
const __m128 CC1 = _mm_set1_ps(1.6666752422E-1f);
- __m128 sign;
- __m128 mask;
- __m128 xabs;
- __m128 z,z1,z2,q,q1,q2;
- __m128 pA,pB;
+ __m128 sign;
+ __m128 mask;
+ __m128 xabs;
+ __m128 z, z1, z2, q, q1, q2;
+ __m128 pA, pB;
- sign = _mm_andnot_ps(signmask,x);
- xabs = _mm_and_ps(x,signmask);
+ sign = _mm_andnot_ps(signmask, x);
+ xabs = _mm_and_ps(x, signmask);
- mask = _mm_cmp_ps(xabs,half,_CMP_GT_OQ);
+ mask = _mm_cmp_ps(xabs, half, _CMP_GT_OQ);
- z1 = _mm_mul_ps(half, _mm_sub_ps(one,xabs));
- q1 = _mm_mul_ps(z1,gmx_mm_invsqrt_ps(z1));
- q1 = _mm_andnot_ps(_mm_cmp_ps(xabs,one,_CMP_EQ_OQ),q1);
+ z1 = _mm_mul_ps(half, _mm_sub_ps(one, xabs));
+ q1 = _mm_mul_ps(z1, gmx_mm_invsqrt_ps(z1));
+ q1 = _mm_andnot_ps(_mm_cmp_ps(xabs, one, _CMP_EQ_OQ), q1);
q2 = xabs;
- z2 = _mm_mul_ps(q2,q2);
+ z2 = _mm_mul_ps(q2, q2);
- z = _mm_or_ps( _mm_and_ps(mask,z1) , _mm_andnot_ps(mask,z2) );
- q = _mm_or_ps( _mm_and_ps(mask,q1) , _mm_andnot_ps(mask,q2) );
+ z = _mm_or_ps( _mm_and_ps(mask, z1), _mm_andnot_ps(mask, z2) );
+ q = _mm_or_ps( _mm_and_ps(mask, q1), _mm_andnot_ps(mask, q2) );
- z2 = _mm_mul_ps(z,z);
+ z2 = _mm_mul_ps(z, z);
- pA = _mm_macc_ps(CC5,z2,CC3);
- pB = _mm_macc_ps(CC4,z2,CC2);
+ pA = _mm_macc_ps(CC5, z2, CC3);
+ pB = _mm_macc_ps(CC4, z2, CC2);
- pA = _mm_macc_ps(pA,z2,CC1);
- pA = _mm_mul_ps(pA,z);
+ pA = _mm_macc_ps(pA, z2, CC1);
+ pA = _mm_mul_ps(pA, z);
- z = _mm_macc_ps(pB,z2,pA);
+ z = _mm_macc_ps(pB, z2, pA);
- z = _mm_macc_ps(z,q,q);
+ z = _mm_macc_ps(z, q, q);
- q2 = _mm_sub_ps(halfpi,z);
- q2 = _mm_sub_ps(q2,z);
+ q2 = _mm_sub_ps(halfpi, z);
+ q2 = _mm_sub_ps(q2, z);
- z = _mm_or_ps( _mm_and_ps(mask,q2) , _mm_andnot_ps(mask,z) );
+ z = _mm_or_ps( _mm_and_ps(mask, q2), _mm_andnot_ps(mask, z) );
- mask = _mm_cmp_ps(xabs,limitlow,_CMP_GT_OQ);
- z = _mm_or_ps( _mm_and_ps(mask,z) , _mm_andnot_ps(mask,xabs) );
+ mask = _mm_cmp_ps(xabs, limitlow, _CMP_GT_OQ);
+ z = _mm_or_ps( _mm_and_ps(mask, z), _mm_andnot_ps(mask, xabs) );
- z = _mm_xor_ps(z,sign);
+ z = _mm_xor_ps(z, sign);
return z;
}
const __m128 pi_ps = _mm_set1_ps(M_PI);
const __m128 halfpi_ps = _mm_set1_ps(M_PI/2.0f);
- __m128 mask1;
- __m128 mask2;
- __m128 xabs;
- __m128 z,z1,z2,z3;
+ __m128 mask1;
+ __m128 mask2;
+ __m128 xabs;
+ __m128 z, z1, z2, z3;
- xabs = _mm_and_ps(x,signmask);
- mask1 = _mm_cmp_ps(xabs,half_ps,_CMP_GT_OQ);
- mask2 = _mm_cmp_ps(x,_mm_setzero_ps(),_CMP_GT_OQ);
+ xabs = _mm_and_ps(x, signmask);
+ mask1 = _mm_cmp_ps(xabs, half_ps, _CMP_GT_OQ);
+ mask2 = _mm_cmp_ps(x, _mm_setzero_ps(), _CMP_GT_OQ);
- z = _mm_mul_ps(half_ps,_mm_sub_ps(one_ps,xabs));
- z = _mm_mul_ps(z,gmx_mm_invsqrt_ps(z));
- z = _mm_andnot_ps(_mm_cmp_ps(xabs,one_ps,_CMP_EQ_OQ),z);
+ z = _mm_mul_ps(half_ps, _mm_sub_ps(one_ps, xabs));
+ z = _mm_mul_ps(z, gmx_mm_invsqrt_ps(z));
+ z = _mm_andnot_ps(_mm_cmp_ps(xabs, one_ps, _CMP_EQ_OQ), z);
- z = _mm_blendv_ps(x,z,mask1);
+ z = _mm_blendv_ps(x, z, mask1);
z = gmx_mm_asin_ps(z);
- z2 = _mm_add_ps(z,z);
- z1 = _mm_sub_ps(pi_ps,z2);
- z3 = _mm_sub_ps(halfpi_ps,z);
+ z2 = _mm_add_ps(z, z);
+ z1 = _mm_sub_ps(pi_ps, z2);
+ z3 = _mm_sub_ps(halfpi_ps, z);
- z = _mm_blendv_ps(z1,z2,mask2);
- z = _mm_blendv_ps(z3,z,mask1);
+ z = _mm_blendv_ps(z1, z2, mask2);
+ z = _mm_blendv_ps(z3, z, mask1);
return z;
}
const __m128 CC7 = _mm_set1_ps(-1.38776856032E-1);
const __m128 CC9 = _mm_set1_ps(8.05374449538e-2f);
- __m128 sign;
- __m128 mask1,mask2;
- __m128 y,z1,z2;
- __m128 x2,x4;
- __m128 sum1,sum2;
+ __m128 sign;
+ __m128 mask1, mask2;
+ __m128 y, z1, z2;
+ __m128 x2, x4;
+ __m128 sum1, sum2;
- sign = _mm_andnot_ps(signmask,x);
- x = _mm_and_ps(x,signmask);
+ sign = _mm_andnot_ps(signmask, x);
+ x = _mm_and_ps(x, signmask);
- mask1 = _mm_cmp_ps(x,limit1,_CMP_GT_OQ);
- mask2 = _mm_cmp_ps(x,limit2,_CMP_GT_OQ);
+ mask1 = _mm_cmp_ps(x, limit1, _CMP_GT_OQ);
+ mask2 = _mm_cmp_ps(x, limit2, _CMP_GT_OQ);
- z1 = _mm_mul_ps(_mm_add_ps(x,mone),gmx_mm_inv_ps(_mm_sub_ps(x,mone)));
- z2 = _mm_mul_ps(mone,gmx_mm_inv_ps(x));
+ z1 = _mm_mul_ps(_mm_add_ps(x, mone), gmx_mm_inv_ps(_mm_sub_ps(x, mone)));
+ z2 = _mm_mul_ps(mone, gmx_mm_inv_ps(x));
- y = _mm_and_ps(mask1,quarterpi);
- y = _mm_blendv_ps(y,halfpi,mask2);
+ y = _mm_and_ps(mask1, quarterpi);
+ y = _mm_blendv_ps(y, halfpi, mask2);
- x = _mm_blendv_ps(x,z1,mask1);
- x = _mm_blendv_ps(x,z2,mask2);
+ x = _mm_blendv_ps(x, z1, mask1);
+ x = _mm_blendv_ps(x, z2, mask2);
- x2 = _mm_mul_ps(x,x);
- x4 = _mm_mul_ps(x2,x2);
+ x2 = _mm_mul_ps(x, x);
+ x4 = _mm_mul_ps(x2, x2);
- sum1 = _mm_macc_ps(CC9,x4,CC5);
- sum2 = _mm_macc_ps(CC7,x4,CC3);
- sum1 = _mm_mul_ps(sum1,x4);
- sum1 = _mm_macc_ps(sum2,x2,sum1);
+ sum1 = _mm_macc_ps(CC9, x4, CC5);
+ sum2 = _mm_macc_ps(CC7, x4, CC3);
+ sum1 = _mm_mul_ps(sum1, x4);
+ sum1 = _mm_macc_ps(sum2, x2, sum1);
- sum1 = _mm_sub_ps(sum1,mone);
- y = _mm_macc_ps(sum1,x,y);
+ sum1 = _mm_sub_ps(sum1, mone);
+ y = _mm_macc_ps(sum1, x, y);
- y = _mm_xor_ps(y,sign);
+ y = _mm_xor_ps(y, sign);
return y;
}
const __m128 halfpi = _mm_set1_ps(M_PI/2.0);
const __m128 minushalfpi = _mm_set1_ps(-M_PI/2.0);
- __m128 z,z1,z3,z4;
- __m128 w;
- __m128 maskx_lt,maskx_eq;
- __m128 masky_lt,masky_eq;
- __m128 mask1,mask2,mask3,mask4,maskall;
+ __m128 z, z1, z3, z4;
+ __m128 w;
+ __m128 maskx_lt, maskx_eq;
+ __m128 masky_lt, masky_eq;
+ __m128 mask1, mask2, mask3, mask4, maskall;
- maskx_lt = _mm_cmp_ps(x,_mm_setzero_ps(),_CMP_LT_OQ);
- masky_lt = _mm_cmp_ps(y,_mm_setzero_ps(),_CMP_LT_OQ);
- maskx_eq = _mm_cmp_ps(x,_mm_setzero_ps(),_CMP_EQ_OQ);
- masky_eq = _mm_cmp_ps(y,_mm_setzero_ps(),_CMP_EQ_OQ);
+ maskx_lt = _mm_cmp_ps(x, _mm_setzero_ps(), _CMP_LT_OQ);
+ masky_lt = _mm_cmp_ps(y, _mm_setzero_ps(), _CMP_LT_OQ);
+ maskx_eq = _mm_cmp_ps(x, _mm_setzero_ps(), _CMP_EQ_OQ);
+ masky_eq = _mm_cmp_ps(y, _mm_setzero_ps(), _CMP_EQ_OQ);
- z = _mm_mul_ps(y,gmx_mm_inv_ps(x));
+ z = _mm_mul_ps(y, gmx_mm_inv_ps(x));
z = gmx_mm_atan_ps(z);
- mask1 = _mm_and_ps(maskx_eq,masky_lt);
- mask2 = _mm_andnot_ps(maskx_lt,masky_eq);
- mask3 = _mm_andnot_ps( _mm_or_ps(masky_lt,masky_eq) , maskx_eq);
- mask4 = _mm_and_ps(masky_eq,maskx_lt);
+ mask1 = _mm_and_ps(maskx_eq, masky_lt);
+ mask2 = _mm_andnot_ps(maskx_lt, masky_eq);
+ mask3 = _mm_andnot_ps( _mm_or_ps(masky_lt, masky_eq), maskx_eq);
+ mask4 = _mm_and_ps(masky_eq, maskx_lt);
- maskall = _mm_or_ps( _mm_or_ps(mask1,mask2), _mm_or_ps(mask3,mask4) );
+ maskall = _mm_or_ps( _mm_or_ps(mask1, mask2), _mm_or_ps(mask3, mask4) );
- z = _mm_andnot_ps(maskall,z);
- z1 = _mm_and_ps(mask1,minushalfpi);
- z3 = _mm_and_ps(mask3,halfpi);
- z4 = _mm_and_ps(mask4,pi);
+ z = _mm_andnot_ps(maskall, z);
+ z1 = _mm_and_ps(mask1, minushalfpi);
+ z3 = _mm_and_ps(mask3, halfpi);
+ z4 = _mm_and_ps(mask4, pi);
- z = _mm_or_ps( _mm_or_ps(z,z1), _mm_or_ps(z3,z4) );
+ z = _mm_or_ps( _mm_or_ps(z, z1), _mm_or_ps(z3, z4) );
- mask1 = _mm_andnot_ps(masky_lt,maskx_lt);
- mask2 = _mm_and_ps(maskx_lt,masky_lt);
+ mask1 = _mm_andnot_ps(masky_lt, maskx_lt);
+ mask2 = _mm_and_ps(maskx_lt, masky_lt);
- w = _mm_or_ps( _mm_and_ps(mask1,pi), _mm_and_ps(mask2,minuspi) );
- w = _mm_andnot_ps(maskall,w);
+ w = _mm_or_ps( _mm_and_ps(mask1, pi), _mm_and_ps(mask2, minuspi) );
+ w = _mm_andnot_ps(maskall, w);
- z = _mm_add_ps(z,w);
+ z = _mm_add_ps(z, w);
return z;
}
biod.pnpi.spb.ru / alexxy / gromacs.git / blobdiff