2 * Note: this file was generated by the Gromacs sse4_1_single kernel generator.
4 * This source code is part of
8 * Copyright (c) 2001-2012, The GROMACS Development Team
10 * Gromacs is a library for molecular simulation and trajectory analysis,
11 * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
12 * a full list of developers and information, check out http://www.gromacs.org
14 * This program is free software; you can redistribute it and/or modify it under
15 * the terms of the GNU Lesser General Public License as published by the Free
16 * Software Foundation; either version 2 of the License, or (at your option) any
19 * To help fund GROMACS development, we humbly ask that you cite
20 * the papers people have written on it - you can find them on the website.
28 #include "../nb_kernel.h"
29 #include "types/simple.h"
33 #include "gmx_math_x86_sse4_1_single.h"
34 #include "kernelutil_x86_sse4_1_single.h"
37 * Gromacs nonbonded kernel: nb_kernel_ElecEwSw_VdwLJSw_GeomW4P1_VF_sse4_1_single
38 * Electrostatics interaction: Ewald
39 * VdW interaction: LennardJones
40 * Geometry: Water4-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecEwSw_VdwLJSw_GeomW4P1_VF_sse4_1_single
45 (t_nblist * gmx_restrict nlist,
46 rvec * gmx_restrict xx,
47 rvec * gmx_restrict ff,
48 t_forcerec * gmx_restrict fr,
49 t_mdatoms * gmx_restrict mdatoms,
50 nb_kernel_data_t * gmx_restrict kernel_data,
51 t_nrnb * gmx_restrict nrnb)
53 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
54 * just 0 for non-waters.
55 * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
56 * jnr indices corresponding to data put in the four positions in the SIMD register.
58 int i_shift_offset,i_coord_offset,outeriter,inneriter;
59 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
60 int jnrA,jnrB,jnrC,jnrD;
61 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
62 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
63 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
65 real *shiftvec,*fshift,*x,*f;
66 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
68 __m128 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
70 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
72 __m128 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
74 __m128 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
76 __m128 ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
77 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
78 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
79 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
80 __m128 dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
81 __m128 dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
82 __m128 dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
83 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
86 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
89 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
90 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
92 __m128 ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
94 __m128 rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
95 real rswitch_scalar,d_scalar;
96 __m128 dummy_mask,cutoff_mask;
97 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
98 __m128 one = _mm_set1_ps(1.0);
99 __m128 two = _mm_set1_ps(2.0);
105 jindex = nlist->jindex;
107 shiftidx = nlist->shift;
109 shiftvec = fr->shift_vec[0];
110 fshift = fr->fshift[0];
111 facel = _mm_set1_ps(fr->epsfac);
112 charge = mdatoms->chargeA;
113 nvdwtype = fr->ntype;
115 vdwtype = mdatoms->typeA;
117 sh_ewald = _mm_set1_ps(fr->ic->sh_ewald);
118 ewtab = fr->ic->tabq_coul_FDV0;
119 ewtabscale = _mm_set1_ps(fr->ic->tabq_scale);
120 ewtabhalfspace = _mm_set1_ps(0.5/fr->ic->tabq_scale);
122 /* Setup water-specific parameters */
123 inr = nlist->iinr[0];
124 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
125 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
126 iq3 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
127 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
129 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
130 rcutoff_scalar = fr->rcoulomb;
131 rcutoff = _mm_set1_ps(rcutoff_scalar);
132 rcutoff2 = _mm_mul_ps(rcutoff,rcutoff);
134 rswitch_scalar = fr->rcoulomb_switch;
135 rswitch = _mm_set1_ps(rswitch_scalar);
136 /* Setup switch parameters */
137 d_scalar = rcutoff_scalar-rswitch_scalar;
138 d = _mm_set1_ps(d_scalar);
139 swV3 = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
140 swV4 = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
141 swV5 = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
142 swF2 = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
143 swF3 = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
144 swF4 = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
146 /* Avoid stupid compiler warnings */
147 jnrA = jnrB = jnrC = jnrD = 0;
156 for(iidx=0;iidx<4*DIM;iidx++)
161 /* Start outer loop over neighborlists */
162 for(iidx=0; iidx<nri; iidx++)
164 /* Load shift vector for this list */
165 i_shift_offset = DIM*shiftidx[iidx];
167 /* Load limits for loop over neighbors */
168 j_index_start = jindex[iidx];
169 j_index_end = jindex[iidx+1];
171 /* Get outer coordinate index */
173 i_coord_offset = DIM*inr;
175 /* Load i particle coords and add shift vector */
176 gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
177 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
179 fix0 = _mm_setzero_ps();
180 fiy0 = _mm_setzero_ps();
181 fiz0 = _mm_setzero_ps();
182 fix1 = _mm_setzero_ps();
183 fiy1 = _mm_setzero_ps();
184 fiz1 = _mm_setzero_ps();
185 fix2 = _mm_setzero_ps();
186 fiy2 = _mm_setzero_ps();
187 fiz2 = _mm_setzero_ps();
188 fix3 = _mm_setzero_ps();
189 fiy3 = _mm_setzero_ps();
190 fiz3 = _mm_setzero_ps();
192 /* Reset potential sums */
193 velecsum = _mm_setzero_ps();
194 vvdwsum = _mm_setzero_ps();
196 /* Start inner kernel loop */
197 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
200 /* Get j neighbor index, and coordinate index */
205 j_coord_offsetA = DIM*jnrA;
206 j_coord_offsetB = DIM*jnrB;
207 j_coord_offsetC = DIM*jnrC;
208 j_coord_offsetD = DIM*jnrD;
210 /* load j atom coordinates */
211 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
212 x+j_coord_offsetC,x+j_coord_offsetD,
215 /* Calculate displacement vector */
216 dx00 = _mm_sub_ps(ix0,jx0);
217 dy00 = _mm_sub_ps(iy0,jy0);
218 dz00 = _mm_sub_ps(iz0,jz0);
219 dx10 = _mm_sub_ps(ix1,jx0);
220 dy10 = _mm_sub_ps(iy1,jy0);
221 dz10 = _mm_sub_ps(iz1,jz0);
222 dx20 = _mm_sub_ps(ix2,jx0);
223 dy20 = _mm_sub_ps(iy2,jy0);
224 dz20 = _mm_sub_ps(iz2,jz0);
225 dx30 = _mm_sub_ps(ix3,jx0);
226 dy30 = _mm_sub_ps(iy3,jy0);
227 dz30 = _mm_sub_ps(iz3,jz0);
229 /* Calculate squared distance and things based on it */
230 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
231 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
232 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
233 rsq30 = gmx_mm_calc_rsq_ps(dx30,dy30,dz30);
235 rinv00 = gmx_mm_invsqrt_ps(rsq00);
236 rinv10 = gmx_mm_invsqrt_ps(rsq10);
237 rinv20 = gmx_mm_invsqrt_ps(rsq20);
238 rinv30 = gmx_mm_invsqrt_ps(rsq30);
240 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
241 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
242 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
243 rinvsq30 = _mm_mul_ps(rinv30,rinv30);
245 /* Load parameters for j particles */
246 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
247 charge+jnrC+0,charge+jnrD+0);
248 vdwjidx0A = 2*vdwtype[jnrA+0];
249 vdwjidx0B = 2*vdwtype[jnrB+0];
250 vdwjidx0C = 2*vdwtype[jnrC+0];
251 vdwjidx0D = 2*vdwtype[jnrD+0];
253 /**************************
254 * CALCULATE INTERACTIONS *
255 **************************/
257 if (gmx_mm_any_lt(rsq00,rcutoff2))
260 r00 = _mm_mul_ps(rsq00,rinv00);
262 /* Compute parameters for interactions between i and j atoms */
263 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
264 vdwparam+vdwioffset0+vdwjidx0B,
265 vdwparam+vdwioffset0+vdwjidx0C,
266 vdwparam+vdwioffset0+vdwjidx0D,
269 /* LENNARD-JONES DISPERSION/REPULSION */
271 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
272 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
273 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
274 vvdw = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
275 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
277 d = _mm_sub_ps(r00,rswitch);
278 d = _mm_max_ps(d,_mm_setzero_ps());
279 d2 = _mm_mul_ps(d,d);
280 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
282 dsw = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
284 /* Evaluate switch function */
285 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
286 fvdw = _mm_sub_ps( _mm_mul_ps(fvdw,sw) , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
287 vvdw = _mm_mul_ps(vvdw,sw);
288 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
290 /* Update potential sum for this i atom from the interaction with this j atom. */
291 vvdw = _mm_and_ps(vvdw,cutoff_mask);
292 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
296 fscal = _mm_and_ps(fscal,cutoff_mask);
298 /* Calculate temporary vectorial force */
299 tx = _mm_mul_ps(fscal,dx00);
300 ty = _mm_mul_ps(fscal,dy00);
301 tz = _mm_mul_ps(fscal,dz00);
303 /* Update vectorial force */
304 fix0 = _mm_add_ps(fix0,tx);
305 fiy0 = _mm_add_ps(fiy0,ty);
306 fiz0 = _mm_add_ps(fiz0,tz);
308 fjptrA = f+j_coord_offsetA;
309 fjptrB = f+j_coord_offsetB;
310 fjptrC = f+j_coord_offsetC;
311 fjptrD = f+j_coord_offsetD;
312 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
316 /**************************
317 * CALCULATE INTERACTIONS *
318 **************************/
320 if (gmx_mm_any_lt(rsq10,rcutoff2))
323 r10 = _mm_mul_ps(rsq10,rinv10);
325 /* Compute parameters for interactions between i and j atoms */
326 qq10 = _mm_mul_ps(iq1,jq0);
328 /* EWALD ELECTROSTATICS */
330 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
331 ewrt = _mm_mul_ps(r10,ewtabscale);
332 ewitab = _mm_cvttps_epi32(ewrt);
333 eweps = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
334 ewitab = _mm_slli_epi32(ewitab,2);
335 ewtabF = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
336 ewtabD = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
337 ewtabV = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
338 ewtabFn = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
339 _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
340 felec = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
341 velec = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
342 velec = _mm_mul_ps(qq10,_mm_sub_ps(rinv10,velec));
343 felec = _mm_mul_ps(_mm_mul_ps(qq10,rinv10),_mm_sub_ps(rinvsq10,felec));
345 d = _mm_sub_ps(r10,rswitch);
346 d = _mm_max_ps(d,_mm_setzero_ps());
347 d2 = _mm_mul_ps(d,d);
348 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
350 dsw = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
352 /* Evaluate switch function */
353 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
354 felec = _mm_sub_ps( _mm_mul_ps(felec,sw) , _mm_mul_ps(rinv10,_mm_mul_ps(velec,dsw)) );
355 velec = _mm_mul_ps(velec,sw);
356 cutoff_mask = _mm_cmplt_ps(rsq10,rcutoff2);
358 /* Update potential sum for this i atom from the interaction with this j atom. */
359 velec = _mm_and_ps(velec,cutoff_mask);
360 velecsum = _mm_add_ps(velecsum,velec);
364 fscal = _mm_and_ps(fscal,cutoff_mask);
366 /* Calculate temporary vectorial force */
367 tx = _mm_mul_ps(fscal,dx10);
368 ty = _mm_mul_ps(fscal,dy10);
369 tz = _mm_mul_ps(fscal,dz10);
371 /* Update vectorial force */
372 fix1 = _mm_add_ps(fix1,tx);
373 fiy1 = _mm_add_ps(fiy1,ty);
374 fiz1 = _mm_add_ps(fiz1,tz);
376 fjptrA = f+j_coord_offsetA;
377 fjptrB = f+j_coord_offsetB;
378 fjptrC = f+j_coord_offsetC;
379 fjptrD = f+j_coord_offsetD;
380 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
384 /**************************
385 * CALCULATE INTERACTIONS *
386 **************************/
388 if (gmx_mm_any_lt(rsq20,rcutoff2))
391 r20 = _mm_mul_ps(rsq20,rinv20);
393 /* Compute parameters for interactions between i and j atoms */
394 qq20 = _mm_mul_ps(iq2,jq0);
396 /* EWALD ELECTROSTATICS */
398 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
399 ewrt = _mm_mul_ps(r20,ewtabscale);
400 ewitab = _mm_cvttps_epi32(ewrt);
401 eweps = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
402 ewitab = _mm_slli_epi32(ewitab,2);
403 ewtabF = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
404 ewtabD = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
405 ewtabV = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
406 ewtabFn = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
407 _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
408 felec = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
409 velec = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
410 velec = _mm_mul_ps(qq20,_mm_sub_ps(rinv20,velec));
411 felec = _mm_mul_ps(_mm_mul_ps(qq20,rinv20),_mm_sub_ps(rinvsq20,felec));
413 d = _mm_sub_ps(r20,rswitch);
414 d = _mm_max_ps(d,_mm_setzero_ps());
415 d2 = _mm_mul_ps(d,d);
416 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
418 dsw = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
420 /* Evaluate switch function */
421 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
422 felec = _mm_sub_ps( _mm_mul_ps(felec,sw) , _mm_mul_ps(rinv20,_mm_mul_ps(velec,dsw)) );
423 velec = _mm_mul_ps(velec,sw);
424 cutoff_mask = _mm_cmplt_ps(rsq20,rcutoff2);
426 /* Update potential sum for this i atom from the interaction with this j atom. */
427 velec = _mm_and_ps(velec,cutoff_mask);
428 velecsum = _mm_add_ps(velecsum,velec);
432 fscal = _mm_and_ps(fscal,cutoff_mask);
434 /* Calculate temporary vectorial force */
435 tx = _mm_mul_ps(fscal,dx20);
436 ty = _mm_mul_ps(fscal,dy20);
437 tz = _mm_mul_ps(fscal,dz20);
439 /* Update vectorial force */
440 fix2 = _mm_add_ps(fix2,tx);
441 fiy2 = _mm_add_ps(fiy2,ty);
442 fiz2 = _mm_add_ps(fiz2,tz);
444 fjptrA = f+j_coord_offsetA;
445 fjptrB = f+j_coord_offsetB;
446 fjptrC = f+j_coord_offsetC;
447 fjptrD = f+j_coord_offsetD;
448 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
452 /**************************
453 * CALCULATE INTERACTIONS *
454 **************************/
456 if (gmx_mm_any_lt(rsq30,rcutoff2))
459 r30 = _mm_mul_ps(rsq30,rinv30);
461 /* Compute parameters for interactions between i and j atoms */
462 qq30 = _mm_mul_ps(iq3,jq0);
464 /* EWALD ELECTROSTATICS */
466 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
467 ewrt = _mm_mul_ps(r30,ewtabscale);
468 ewitab = _mm_cvttps_epi32(ewrt);
469 eweps = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
470 ewitab = _mm_slli_epi32(ewitab,2);
471 ewtabF = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
472 ewtabD = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
473 ewtabV = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
474 ewtabFn = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
475 _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
476 felec = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
477 velec = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
478 velec = _mm_mul_ps(qq30,_mm_sub_ps(rinv30,velec));
479 felec = _mm_mul_ps(_mm_mul_ps(qq30,rinv30),_mm_sub_ps(rinvsq30,felec));
481 d = _mm_sub_ps(r30,rswitch);
482 d = _mm_max_ps(d,_mm_setzero_ps());
483 d2 = _mm_mul_ps(d,d);
484 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
486 dsw = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
488 /* Evaluate switch function */
489 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
490 felec = _mm_sub_ps( _mm_mul_ps(felec,sw) , _mm_mul_ps(rinv30,_mm_mul_ps(velec,dsw)) );
491 velec = _mm_mul_ps(velec,sw);
492 cutoff_mask = _mm_cmplt_ps(rsq30,rcutoff2);
494 /* Update potential sum for this i atom from the interaction with this j atom. */
495 velec = _mm_and_ps(velec,cutoff_mask);
496 velecsum = _mm_add_ps(velecsum,velec);
500 fscal = _mm_and_ps(fscal,cutoff_mask);
502 /* Calculate temporary vectorial force */
503 tx = _mm_mul_ps(fscal,dx30);
504 ty = _mm_mul_ps(fscal,dy30);
505 tz = _mm_mul_ps(fscal,dz30);
507 /* Update vectorial force */
508 fix3 = _mm_add_ps(fix3,tx);
509 fiy3 = _mm_add_ps(fiy3,ty);
510 fiz3 = _mm_add_ps(fiz3,tz);
512 fjptrA = f+j_coord_offsetA;
513 fjptrB = f+j_coord_offsetB;
514 fjptrC = f+j_coord_offsetC;
515 fjptrD = f+j_coord_offsetD;
516 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
520 /* Inner loop uses 254 flops */
526 /* Get j neighbor index, and coordinate index */
527 jnrlistA = jjnr[jidx];
528 jnrlistB = jjnr[jidx+1];
529 jnrlistC = jjnr[jidx+2];
530 jnrlistD = jjnr[jidx+3];
531 /* Sign of each element will be negative for non-real atoms.
532 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
533 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
535 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
536 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
537 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
538 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
539 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
540 j_coord_offsetA = DIM*jnrA;
541 j_coord_offsetB = DIM*jnrB;
542 j_coord_offsetC = DIM*jnrC;
543 j_coord_offsetD = DIM*jnrD;
545 /* load j atom coordinates */
546 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
547 x+j_coord_offsetC,x+j_coord_offsetD,
550 /* Calculate displacement vector */
551 dx00 = _mm_sub_ps(ix0,jx0);
552 dy00 = _mm_sub_ps(iy0,jy0);
553 dz00 = _mm_sub_ps(iz0,jz0);
554 dx10 = _mm_sub_ps(ix1,jx0);
555 dy10 = _mm_sub_ps(iy1,jy0);
556 dz10 = _mm_sub_ps(iz1,jz0);
557 dx20 = _mm_sub_ps(ix2,jx0);
558 dy20 = _mm_sub_ps(iy2,jy0);
559 dz20 = _mm_sub_ps(iz2,jz0);
560 dx30 = _mm_sub_ps(ix3,jx0);
561 dy30 = _mm_sub_ps(iy3,jy0);
562 dz30 = _mm_sub_ps(iz3,jz0);
564 /* Calculate squared distance and things based on it */
565 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
566 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
567 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
568 rsq30 = gmx_mm_calc_rsq_ps(dx30,dy30,dz30);
570 rinv00 = gmx_mm_invsqrt_ps(rsq00);
571 rinv10 = gmx_mm_invsqrt_ps(rsq10);
572 rinv20 = gmx_mm_invsqrt_ps(rsq20);
573 rinv30 = gmx_mm_invsqrt_ps(rsq30);
575 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
576 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
577 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
578 rinvsq30 = _mm_mul_ps(rinv30,rinv30);
580 /* Load parameters for j particles */
581 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
582 charge+jnrC+0,charge+jnrD+0);
583 vdwjidx0A = 2*vdwtype[jnrA+0];
584 vdwjidx0B = 2*vdwtype[jnrB+0];
585 vdwjidx0C = 2*vdwtype[jnrC+0];
586 vdwjidx0D = 2*vdwtype[jnrD+0];
588 /**************************
589 * CALCULATE INTERACTIONS *
590 **************************/
592 if (gmx_mm_any_lt(rsq00,rcutoff2))
595 r00 = _mm_mul_ps(rsq00,rinv00);
596 r00 = _mm_andnot_ps(dummy_mask,r00);
598 /* Compute parameters for interactions between i and j atoms */
599 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
600 vdwparam+vdwioffset0+vdwjidx0B,
601 vdwparam+vdwioffset0+vdwjidx0C,
602 vdwparam+vdwioffset0+vdwjidx0D,
605 /* LENNARD-JONES DISPERSION/REPULSION */
607 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
608 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
609 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
610 vvdw = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
611 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
613 d = _mm_sub_ps(r00,rswitch);
614 d = _mm_max_ps(d,_mm_setzero_ps());
615 d2 = _mm_mul_ps(d,d);
616 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
618 dsw = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
620 /* Evaluate switch function */
621 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
622 fvdw = _mm_sub_ps( _mm_mul_ps(fvdw,sw) , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
623 vvdw = _mm_mul_ps(vvdw,sw);
624 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
626 /* Update potential sum for this i atom from the interaction with this j atom. */
627 vvdw = _mm_and_ps(vvdw,cutoff_mask);
628 vvdw = _mm_andnot_ps(dummy_mask,vvdw);
629 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
633 fscal = _mm_and_ps(fscal,cutoff_mask);
635 fscal = _mm_andnot_ps(dummy_mask,fscal);
637 /* Calculate temporary vectorial force */
638 tx = _mm_mul_ps(fscal,dx00);
639 ty = _mm_mul_ps(fscal,dy00);
640 tz = _mm_mul_ps(fscal,dz00);
642 /* Update vectorial force */
643 fix0 = _mm_add_ps(fix0,tx);
644 fiy0 = _mm_add_ps(fiy0,ty);
645 fiz0 = _mm_add_ps(fiz0,tz);
647 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
648 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
649 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
650 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
651 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
655 /**************************
656 * CALCULATE INTERACTIONS *
657 **************************/
659 if (gmx_mm_any_lt(rsq10,rcutoff2))
662 r10 = _mm_mul_ps(rsq10,rinv10);
663 r10 = _mm_andnot_ps(dummy_mask,r10);
665 /* Compute parameters for interactions between i and j atoms */
666 qq10 = _mm_mul_ps(iq1,jq0);
668 /* EWALD ELECTROSTATICS */
670 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
671 ewrt = _mm_mul_ps(r10,ewtabscale);
672 ewitab = _mm_cvttps_epi32(ewrt);
673 eweps = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
674 ewitab = _mm_slli_epi32(ewitab,2);
675 ewtabF = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
676 ewtabD = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
677 ewtabV = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
678 ewtabFn = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
679 _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
680 felec = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
681 velec = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
682 velec = _mm_mul_ps(qq10,_mm_sub_ps(rinv10,velec));
683 felec = _mm_mul_ps(_mm_mul_ps(qq10,rinv10),_mm_sub_ps(rinvsq10,felec));
685 d = _mm_sub_ps(r10,rswitch);
686 d = _mm_max_ps(d,_mm_setzero_ps());
687 d2 = _mm_mul_ps(d,d);
688 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
690 dsw = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
692 /* Evaluate switch function */
693 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
694 felec = _mm_sub_ps( _mm_mul_ps(felec,sw) , _mm_mul_ps(rinv10,_mm_mul_ps(velec,dsw)) );
695 velec = _mm_mul_ps(velec,sw);
696 cutoff_mask = _mm_cmplt_ps(rsq10,rcutoff2);
698 /* Update potential sum for this i atom from the interaction with this j atom. */
699 velec = _mm_and_ps(velec,cutoff_mask);
700 velec = _mm_andnot_ps(dummy_mask,velec);
701 velecsum = _mm_add_ps(velecsum,velec);
705 fscal = _mm_and_ps(fscal,cutoff_mask);
707 fscal = _mm_andnot_ps(dummy_mask,fscal);
709 /* Calculate temporary vectorial force */
710 tx = _mm_mul_ps(fscal,dx10);
711 ty = _mm_mul_ps(fscal,dy10);
712 tz = _mm_mul_ps(fscal,dz10);
714 /* Update vectorial force */
715 fix1 = _mm_add_ps(fix1,tx);
716 fiy1 = _mm_add_ps(fiy1,ty);
717 fiz1 = _mm_add_ps(fiz1,tz);
719 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
720 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
721 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
722 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
723 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
727 /**************************
728 * CALCULATE INTERACTIONS *
729 **************************/
731 if (gmx_mm_any_lt(rsq20,rcutoff2))
734 r20 = _mm_mul_ps(rsq20,rinv20);
735 r20 = _mm_andnot_ps(dummy_mask,r20);
737 /* Compute parameters for interactions between i and j atoms */
738 qq20 = _mm_mul_ps(iq2,jq0);
740 /* EWALD ELECTROSTATICS */
742 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
743 ewrt = _mm_mul_ps(r20,ewtabscale);
744 ewitab = _mm_cvttps_epi32(ewrt);
745 eweps = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
746 ewitab = _mm_slli_epi32(ewitab,2);
747 ewtabF = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
748 ewtabD = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
749 ewtabV = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
750 ewtabFn = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
751 _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
752 felec = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
753 velec = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
754 velec = _mm_mul_ps(qq20,_mm_sub_ps(rinv20,velec));
755 felec = _mm_mul_ps(_mm_mul_ps(qq20,rinv20),_mm_sub_ps(rinvsq20,felec));
757 d = _mm_sub_ps(r20,rswitch);
758 d = _mm_max_ps(d,_mm_setzero_ps());
759 d2 = _mm_mul_ps(d,d);
760 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
762 dsw = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
764 /* Evaluate switch function */
765 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
766 felec = _mm_sub_ps( _mm_mul_ps(felec,sw) , _mm_mul_ps(rinv20,_mm_mul_ps(velec,dsw)) );
767 velec = _mm_mul_ps(velec,sw);
768 cutoff_mask = _mm_cmplt_ps(rsq20,rcutoff2);
770 /* Update potential sum for this i atom from the interaction with this j atom. */
771 velec = _mm_and_ps(velec,cutoff_mask);
772 velec = _mm_andnot_ps(dummy_mask,velec);
773 velecsum = _mm_add_ps(velecsum,velec);
777 fscal = _mm_and_ps(fscal,cutoff_mask);
779 fscal = _mm_andnot_ps(dummy_mask,fscal);
781 /* Calculate temporary vectorial force */
782 tx = _mm_mul_ps(fscal,dx20);
783 ty = _mm_mul_ps(fscal,dy20);
784 tz = _mm_mul_ps(fscal,dz20);
786 /* Update vectorial force */
787 fix2 = _mm_add_ps(fix2,tx);
788 fiy2 = _mm_add_ps(fiy2,ty);
789 fiz2 = _mm_add_ps(fiz2,tz);
791 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
792 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
793 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
794 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
795 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
799 /**************************
800 * CALCULATE INTERACTIONS *
801 **************************/
803 if (gmx_mm_any_lt(rsq30,rcutoff2))
806 r30 = _mm_mul_ps(rsq30,rinv30);
807 r30 = _mm_andnot_ps(dummy_mask,r30);
809 /* Compute parameters for interactions between i and j atoms */
810 qq30 = _mm_mul_ps(iq3,jq0);
812 /* EWALD ELECTROSTATICS */
814 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
815 ewrt = _mm_mul_ps(r30,ewtabscale);
816 ewitab = _mm_cvttps_epi32(ewrt);
817 eweps = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
818 ewitab = _mm_slli_epi32(ewitab,2);
819 ewtabF = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
820 ewtabD = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
821 ewtabV = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
822 ewtabFn = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
823 _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
824 felec = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
825 velec = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
826 velec = _mm_mul_ps(qq30,_mm_sub_ps(rinv30,velec));
827 felec = _mm_mul_ps(_mm_mul_ps(qq30,rinv30),_mm_sub_ps(rinvsq30,felec));
829 d = _mm_sub_ps(r30,rswitch);
830 d = _mm_max_ps(d,_mm_setzero_ps());
831 d2 = _mm_mul_ps(d,d);
832 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
834 dsw = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
836 /* Evaluate switch function */
837 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
838 felec = _mm_sub_ps( _mm_mul_ps(felec,sw) , _mm_mul_ps(rinv30,_mm_mul_ps(velec,dsw)) );
839 velec = _mm_mul_ps(velec,sw);
840 cutoff_mask = _mm_cmplt_ps(rsq30,rcutoff2);
842 /* Update potential sum for this i atom from the interaction with this j atom. */
843 velec = _mm_and_ps(velec,cutoff_mask);
844 velec = _mm_andnot_ps(dummy_mask,velec);
845 velecsum = _mm_add_ps(velecsum,velec);
849 fscal = _mm_and_ps(fscal,cutoff_mask);
851 fscal = _mm_andnot_ps(dummy_mask,fscal);
853 /* Calculate temporary vectorial force */
854 tx = _mm_mul_ps(fscal,dx30);
855 ty = _mm_mul_ps(fscal,dy30);
856 tz = _mm_mul_ps(fscal,dz30);
858 /* Update vectorial force */
859 fix3 = _mm_add_ps(fix3,tx);
860 fiy3 = _mm_add_ps(fiy3,ty);
861 fiz3 = _mm_add_ps(fiz3,tz);
863 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
864 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
865 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
866 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
867 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
871 /* Inner loop uses 258 flops */
874 /* End of innermost loop */
876 gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
877 f+i_coord_offset,fshift+i_shift_offset);
880 /* Update potential energies */
881 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
882 gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
884 /* Increment number of inner iterations */
885 inneriter += j_index_end - j_index_start;
887 /* Outer loop uses 26 flops */
890 /* Increment number of outer iterations */
893 /* Update outer/inner flops */
895 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_VF,outeriter*26 + inneriter*258);
898 * Gromacs nonbonded kernel: nb_kernel_ElecEwSw_VdwLJSw_GeomW4P1_F_sse4_1_single
899 * Electrostatics interaction: Ewald
900 * VdW interaction: LennardJones
901 * Geometry: Water4-Particle
902 * Calculate force/pot: Force
905 nb_kernel_ElecEwSw_VdwLJSw_GeomW4P1_F_sse4_1_single
906 (t_nblist * gmx_restrict nlist,
907 rvec * gmx_restrict xx,
908 rvec * gmx_restrict ff,
909 t_forcerec * gmx_restrict fr,
910 t_mdatoms * gmx_restrict mdatoms,
911 nb_kernel_data_t * gmx_restrict kernel_data,
912 t_nrnb * gmx_restrict nrnb)
914 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
915 * just 0 for non-waters.
916 * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
917 * jnr indices corresponding to data put in the four positions in the SIMD register.
919 int i_shift_offset,i_coord_offset,outeriter,inneriter;
920 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
921 int jnrA,jnrB,jnrC,jnrD;
922 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
923 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
924 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
926 real *shiftvec,*fshift,*x,*f;
927 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
929 __m128 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
931 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
933 __m128 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
935 __m128 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
937 __m128 ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
938 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
939 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
940 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
941 __m128 dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
942 __m128 dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
943 __m128 dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
944 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
947 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
950 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
951 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
953 __m128 ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
955 __m128 rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
956 real rswitch_scalar,d_scalar;
957 __m128 dummy_mask,cutoff_mask;
958 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
959 __m128 one = _mm_set1_ps(1.0);
960 __m128 two = _mm_set1_ps(2.0);
966 jindex = nlist->jindex;
968 shiftidx = nlist->shift;
970 shiftvec = fr->shift_vec[0];
971 fshift = fr->fshift[0];
972 facel = _mm_set1_ps(fr->epsfac);
973 charge = mdatoms->chargeA;
974 nvdwtype = fr->ntype;
976 vdwtype = mdatoms->typeA;
978 sh_ewald = _mm_set1_ps(fr->ic->sh_ewald);
979 ewtab = fr->ic->tabq_coul_FDV0;
980 ewtabscale = _mm_set1_ps(fr->ic->tabq_scale);
981 ewtabhalfspace = _mm_set1_ps(0.5/fr->ic->tabq_scale);
983 /* Setup water-specific parameters */
984 inr = nlist->iinr[0];
985 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
986 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
987 iq3 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
988 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
990 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
991 rcutoff_scalar = fr->rcoulomb;
992 rcutoff = _mm_set1_ps(rcutoff_scalar);
993 rcutoff2 = _mm_mul_ps(rcutoff,rcutoff);
995 rswitch_scalar = fr->rcoulomb_switch;
996 rswitch = _mm_set1_ps(rswitch_scalar);
997 /* Setup switch parameters */
998 d_scalar = rcutoff_scalar-rswitch_scalar;
999 d = _mm_set1_ps(d_scalar);
1000 swV3 = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
1001 swV4 = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1002 swV5 = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1003 swF2 = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
1004 swF3 = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1005 swF4 = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1007 /* Avoid stupid compiler warnings */
1008 jnrA = jnrB = jnrC = jnrD = 0;
1009 j_coord_offsetA = 0;
1010 j_coord_offsetB = 0;
1011 j_coord_offsetC = 0;
1012 j_coord_offsetD = 0;
1017 for(iidx=0;iidx<4*DIM;iidx++)
1019 scratch[iidx] = 0.0;
1022 /* Start outer loop over neighborlists */
1023 for(iidx=0; iidx<nri; iidx++)
1025 /* Load shift vector for this list */
1026 i_shift_offset = DIM*shiftidx[iidx];
1028 /* Load limits for loop over neighbors */
1029 j_index_start = jindex[iidx];
1030 j_index_end = jindex[iidx+1];
1032 /* Get outer coordinate index */
1034 i_coord_offset = DIM*inr;
1036 /* Load i particle coords and add shift vector */
1037 gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1038 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1040 fix0 = _mm_setzero_ps();
1041 fiy0 = _mm_setzero_ps();
1042 fiz0 = _mm_setzero_ps();
1043 fix1 = _mm_setzero_ps();
1044 fiy1 = _mm_setzero_ps();
1045 fiz1 = _mm_setzero_ps();
1046 fix2 = _mm_setzero_ps();
1047 fiy2 = _mm_setzero_ps();
1048 fiz2 = _mm_setzero_ps();
1049 fix3 = _mm_setzero_ps();
1050 fiy3 = _mm_setzero_ps();
1051 fiz3 = _mm_setzero_ps();
1053 /* Start inner kernel loop */
1054 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1057 /* Get j neighbor index, and coordinate index */
1059 jnrB = jjnr[jidx+1];
1060 jnrC = jjnr[jidx+2];
1061 jnrD = jjnr[jidx+3];
1062 j_coord_offsetA = DIM*jnrA;
1063 j_coord_offsetB = DIM*jnrB;
1064 j_coord_offsetC = DIM*jnrC;
1065 j_coord_offsetD = DIM*jnrD;
1067 /* load j atom coordinates */
1068 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1069 x+j_coord_offsetC,x+j_coord_offsetD,
1072 /* Calculate displacement vector */
1073 dx00 = _mm_sub_ps(ix0,jx0);
1074 dy00 = _mm_sub_ps(iy0,jy0);
1075 dz00 = _mm_sub_ps(iz0,jz0);
1076 dx10 = _mm_sub_ps(ix1,jx0);
1077 dy10 = _mm_sub_ps(iy1,jy0);
1078 dz10 = _mm_sub_ps(iz1,jz0);
1079 dx20 = _mm_sub_ps(ix2,jx0);
1080 dy20 = _mm_sub_ps(iy2,jy0);
1081 dz20 = _mm_sub_ps(iz2,jz0);
1082 dx30 = _mm_sub_ps(ix3,jx0);
1083 dy30 = _mm_sub_ps(iy3,jy0);
1084 dz30 = _mm_sub_ps(iz3,jz0);
1086 /* Calculate squared distance and things based on it */
1087 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1088 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1089 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1090 rsq30 = gmx_mm_calc_rsq_ps(dx30,dy30,dz30);
1092 rinv00 = gmx_mm_invsqrt_ps(rsq00);
1093 rinv10 = gmx_mm_invsqrt_ps(rsq10);
1094 rinv20 = gmx_mm_invsqrt_ps(rsq20);
1095 rinv30 = gmx_mm_invsqrt_ps(rsq30);
1097 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
1098 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
1099 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
1100 rinvsq30 = _mm_mul_ps(rinv30,rinv30);
1102 /* Load parameters for j particles */
1103 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
1104 charge+jnrC+0,charge+jnrD+0);
1105 vdwjidx0A = 2*vdwtype[jnrA+0];
1106 vdwjidx0B = 2*vdwtype[jnrB+0];
1107 vdwjidx0C = 2*vdwtype[jnrC+0];
1108 vdwjidx0D = 2*vdwtype[jnrD+0];
1110 /**************************
1111 * CALCULATE INTERACTIONS *
1112 **************************/
1114 if (gmx_mm_any_lt(rsq00,rcutoff2))
1117 r00 = _mm_mul_ps(rsq00,rinv00);
1119 /* Compute parameters for interactions between i and j atoms */
1120 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
1121 vdwparam+vdwioffset0+vdwjidx0B,
1122 vdwparam+vdwioffset0+vdwjidx0C,
1123 vdwparam+vdwioffset0+vdwjidx0D,
1126 /* LENNARD-JONES DISPERSION/REPULSION */
1128 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1129 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
1130 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
1131 vvdw = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
1132 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
1134 d = _mm_sub_ps(r00,rswitch);
1135 d = _mm_max_ps(d,_mm_setzero_ps());
1136 d2 = _mm_mul_ps(d,d);
1137 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
1139 dsw = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
1141 /* Evaluate switch function */
1142 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1143 fvdw = _mm_sub_ps( _mm_mul_ps(fvdw,sw) , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
1144 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
1148 fscal = _mm_and_ps(fscal,cutoff_mask);
1150 /* Calculate temporary vectorial force */
1151 tx = _mm_mul_ps(fscal,dx00);
1152 ty = _mm_mul_ps(fscal,dy00);
1153 tz = _mm_mul_ps(fscal,dz00);
1155 /* Update vectorial force */
1156 fix0 = _mm_add_ps(fix0,tx);
1157 fiy0 = _mm_add_ps(fiy0,ty);
1158 fiz0 = _mm_add_ps(fiz0,tz);
1160 fjptrA = f+j_coord_offsetA;
1161 fjptrB = f+j_coord_offsetB;
1162 fjptrC = f+j_coord_offsetC;
1163 fjptrD = f+j_coord_offsetD;
1164 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
1168 /**************************
1169 * CALCULATE INTERACTIONS *
1170 **************************/
1172 if (gmx_mm_any_lt(rsq10,rcutoff2))
1175 r10 = _mm_mul_ps(rsq10,rinv10);
1177 /* Compute parameters for interactions between i and j atoms */
1178 qq10 = _mm_mul_ps(iq1,jq0);
1180 /* EWALD ELECTROSTATICS */
1182 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1183 ewrt = _mm_mul_ps(r10,ewtabscale);
1184 ewitab = _mm_cvttps_epi32(ewrt);
1185 eweps = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
1186 ewitab = _mm_slli_epi32(ewitab,2);
1187 ewtabF = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1188 ewtabD = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1189 ewtabV = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1190 ewtabFn = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1191 _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1192 felec = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1193 velec = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1194 velec = _mm_mul_ps(qq10,_mm_sub_ps(rinv10,velec));
1195 felec = _mm_mul_ps(_mm_mul_ps(qq10,rinv10),_mm_sub_ps(rinvsq10,felec));
1197 d = _mm_sub_ps(r10,rswitch);
1198 d = _mm_max_ps(d,_mm_setzero_ps());
1199 d2 = _mm_mul_ps(d,d);
1200 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
1202 dsw = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
1204 /* Evaluate switch function */
1205 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1206 felec = _mm_sub_ps( _mm_mul_ps(felec,sw) , _mm_mul_ps(rinv10,_mm_mul_ps(velec,dsw)) );
1207 cutoff_mask = _mm_cmplt_ps(rsq10,rcutoff2);
1211 fscal = _mm_and_ps(fscal,cutoff_mask);
1213 /* Calculate temporary vectorial force */
1214 tx = _mm_mul_ps(fscal,dx10);
1215 ty = _mm_mul_ps(fscal,dy10);
1216 tz = _mm_mul_ps(fscal,dz10);
1218 /* Update vectorial force */
1219 fix1 = _mm_add_ps(fix1,tx);
1220 fiy1 = _mm_add_ps(fiy1,ty);
1221 fiz1 = _mm_add_ps(fiz1,tz);
1223 fjptrA = f+j_coord_offsetA;
1224 fjptrB = f+j_coord_offsetB;
1225 fjptrC = f+j_coord_offsetC;
1226 fjptrD = f+j_coord_offsetD;
1227 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
1231 /**************************
1232 * CALCULATE INTERACTIONS *
1233 **************************/
1235 if (gmx_mm_any_lt(rsq20,rcutoff2))
1238 r20 = _mm_mul_ps(rsq20,rinv20);
1240 /* Compute parameters for interactions between i and j atoms */
1241 qq20 = _mm_mul_ps(iq2,jq0);
1243 /* EWALD ELECTROSTATICS */
1245 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1246 ewrt = _mm_mul_ps(r20,ewtabscale);
1247 ewitab = _mm_cvttps_epi32(ewrt);
1248 eweps = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
1249 ewitab = _mm_slli_epi32(ewitab,2);
1250 ewtabF = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1251 ewtabD = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1252 ewtabV = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1253 ewtabFn = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1254 _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1255 felec = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1256 velec = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1257 velec = _mm_mul_ps(qq20,_mm_sub_ps(rinv20,velec));
1258 felec = _mm_mul_ps(_mm_mul_ps(qq20,rinv20),_mm_sub_ps(rinvsq20,felec));
1260 d = _mm_sub_ps(r20,rswitch);
1261 d = _mm_max_ps(d,_mm_setzero_ps());
1262 d2 = _mm_mul_ps(d,d);
1263 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
1265 dsw = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
1267 /* Evaluate switch function */
1268 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1269 felec = _mm_sub_ps( _mm_mul_ps(felec,sw) , _mm_mul_ps(rinv20,_mm_mul_ps(velec,dsw)) );
1270 cutoff_mask = _mm_cmplt_ps(rsq20,rcutoff2);
1274 fscal = _mm_and_ps(fscal,cutoff_mask);
1276 /* Calculate temporary vectorial force */
1277 tx = _mm_mul_ps(fscal,dx20);
1278 ty = _mm_mul_ps(fscal,dy20);
1279 tz = _mm_mul_ps(fscal,dz20);
1281 /* Update vectorial force */
1282 fix2 = _mm_add_ps(fix2,tx);
1283 fiy2 = _mm_add_ps(fiy2,ty);
1284 fiz2 = _mm_add_ps(fiz2,tz);
1286 fjptrA = f+j_coord_offsetA;
1287 fjptrB = f+j_coord_offsetB;
1288 fjptrC = f+j_coord_offsetC;
1289 fjptrD = f+j_coord_offsetD;
1290 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
1294 /**************************
1295 * CALCULATE INTERACTIONS *
1296 **************************/
1298 if (gmx_mm_any_lt(rsq30,rcutoff2))
1301 r30 = _mm_mul_ps(rsq30,rinv30);
1303 /* Compute parameters for interactions between i and j atoms */
1304 qq30 = _mm_mul_ps(iq3,jq0);
1306 /* EWALD ELECTROSTATICS */
1308 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1309 ewrt = _mm_mul_ps(r30,ewtabscale);
1310 ewitab = _mm_cvttps_epi32(ewrt);
1311 eweps = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
1312 ewitab = _mm_slli_epi32(ewitab,2);
1313 ewtabF = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1314 ewtabD = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1315 ewtabV = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1316 ewtabFn = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1317 _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1318 felec = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1319 velec = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1320 velec = _mm_mul_ps(qq30,_mm_sub_ps(rinv30,velec));
1321 felec = _mm_mul_ps(_mm_mul_ps(qq30,rinv30),_mm_sub_ps(rinvsq30,felec));
1323 d = _mm_sub_ps(r30,rswitch);
1324 d = _mm_max_ps(d,_mm_setzero_ps());
1325 d2 = _mm_mul_ps(d,d);
1326 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
1328 dsw = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
1330 /* Evaluate switch function */
1331 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1332 felec = _mm_sub_ps( _mm_mul_ps(felec,sw) , _mm_mul_ps(rinv30,_mm_mul_ps(velec,dsw)) );
1333 cutoff_mask = _mm_cmplt_ps(rsq30,rcutoff2);
1337 fscal = _mm_and_ps(fscal,cutoff_mask);
1339 /* Calculate temporary vectorial force */
1340 tx = _mm_mul_ps(fscal,dx30);
1341 ty = _mm_mul_ps(fscal,dy30);
1342 tz = _mm_mul_ps(fscal,dz30);
1344 /* Update vectorial force */
1345 fix3 = _mm_add_ps(fix3,tx);
1346 fiy3 = _mm_add_ps(fiy3,ty);
1347 fiz3 = _mm_add_ps(fiz3,tz);
1349 fjptrA = f+j_coord_offsetA;
1350 fjptrB = f+j_coord_offsetB;
1351 fjptrC = f+j_coord_offsetC;
1352 fjptrD = f+j_coord_offsetD;
1353 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
1357 /* Inner loop uses 242 flops */
1360 if(jidx<j_index_end)
1363 /* Get j neighbor index, and coordinate index */
1364 jnrlistA = jjnr[jidx];
1365 jnrlistB = jjnr[jidx+1];
1366 jnrlistC = jjnr[jidx+2];
1367 jnrlistD = jjnr[jidx+3];
1368 /* Sign of each element will be negative for non-real atoms.
1369 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1370 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1372 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1373 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
1374 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
1375 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
1376 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
1377 j_coord_offsetA = DIM*jnrA;
1378 j_coord_offsetB = DIM*jnrB;
1379 j_coord_offsetC = DIM*jnrC;
1380 j_coord_offsetD = DIM*jnrD;
1382 /* load j atom coordinates */
1383 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1384 x+j_coord_offsetC,x+j_coord_offsetD,
1387 /* Calculate displacement vector */
1388 dx00 = _mm_sub_ps(ix0,jx0);
1389 dy00 = _mm_sub_ps(iy0,jy0);
1390 dz00 = _mm_sub_ps(iz0,jz0);
1391 dx10 = _mm_sub_ps(ix1,jx0);
1392 dy10 = _mm_sub_ps(iy1,jy0);
1393 dz10 = _mm_sub_ps(iz1,jz0);
1394 dx20 = _mm_sub_ps(ix2,jx0);
1395 dy20 = _mm_sub_ps(iy2,jy0);
1396 dz20 = _mm_sub_ps(iz2,jz0);
1397 dx30 = _mm_sub_ps(ix3,jx0);
1398 dy30 = _mm_sub_ps(iy3,jy0);
1399 dz30 = _mm_sub_ps(iz3,jz0);
1401 /* Calculate squared distance and things based on it */
1402 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1403 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1404 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1405 rsq30 = gmx_mm_calc_rsq_ps(dx30,dy30,dz30);
1407 rinv00 = gmx_mm_invsqrt_ps(rsq00);
1408 rinv10 = gmx_mm_invsqrt_ps(rsq10);
1409 rinv20 = gmx_mm_invsqrt_ps(rsq20);
1410 rinv30 = gmx_mm_invsqrt_ps(rsq30);
1412 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
1413 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
1414 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
1415 rinvsq30 = _mm_mul_ps(rinv30,rinv30);
1417 /* Load parameters for j particles */
1418 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
1419 charge+jnrC+0,charge+jnrD+0);
1420 vdwjidx0A = 2*vdwtype[jnrA+0];
1421 vdwjidx0B = 2*vdwtype[jnrB+0];
1422 vdwjidx0C = 2*vdwtype[jnrC+0];
1423 vdwjidx0D = 2*vdwtype[jnrD+0];
1425 /**************************
1426 * CALCULATE INTERACTIONS *
1427 **************************/
1429 if (gmx_mm_any_lt(rsq00,rcutoff2))
1432 r00 = _mm_mul_ps(rsq00,rinv00);
1433 r00 = _mm_andnot_ps(dummy_mask,r00);
1435 /* Compute parameters for interactions between i and j atoms */
1436 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
1437 vdwparam+vdwioffset0+vdwjidx0B,
1438 vdwparam+vdwioffset0+vdwjidx0C,
1439 vdwparam+vdwioffset0+vdwjidx0D,
1442 /* LENNARD-JONES DISPERSION/REPULSION */
1444 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1445 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
1446 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
1447 vvdw = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
1448 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
1450 d = _mm_sub_ps(r00,rswitch);
1451 d = _mm_max_ps(d,_mm_setzero_ps());
1452 d2 = _mm_mul_ps(d,d);
1453 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
1455 dsw = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
1457 /* Evaluate switch function */
1458 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1459 fvdw = _mm_sub_ps( _mm_mul_ps(fvdw,sw) , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
1460 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
1464 fscal = _mm_and_ps(fscal,cutoff_mask);
1466 fscal = _mm_andnot_ps(dummy_mask,fscal);
1468 /* Calculate temporary vectorial force */
1469 tx = _mm_mul_ps(fscal,dx00);
1470 ty = _mm_mul_ps(fscal,dy00);
1471 tz = _mm_mul_ps(fscal,dz00);
1473 /* Update vectorial force */
1474 fix0 = _mm_add_ps(fix0,tx);
1475 fiy0 = _mm_add_ps(fiy0,ty);
1476 fiz0 = _mm_add_ps(fiz0,tz);
1478 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1479 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1480 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1481 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1482 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
1486 /**************************
1487 * CALCULATE INTERACTIONS *
1488 **************************/
1490 if (gmx_mm_any_lt(rsq10,rcutoff2))
1493 r10 = _mm_mul_ps(rsq10,rinv10);
1494 r10 = _mm_andnot_ps(dummy_mask,r10);
1496 /* Compute parameters for interactions between i and j atoms */
1497 qq10 = _mm_mul_ps(iq1,jq0);
1499 /* EWALD ELECTROSTATICS */
1501 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1502 ewrt = _mm_mul_ps(r10,ewtabscale);
1503 ewitab = _mm_cvttps_epi32(ewrt);
1504 eweps = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
1505 ewitab = _mm_slli_epi32(ewitab,2);
1506 ewtabF = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1507 ewtabD = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1508 ewtabV = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1509 ewtabFn = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1510 _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1511 felec = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1512 velec = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1513 velec = _mm_mul_ps(qq10,_mm_sub_ps(rinv10,velec));
1514 felec = _mm_mul_ps(_mm_mul_ps(qq10,rinv10),_mm_sub_ps(rinvsq10,felec));
1516 d = _mm_sub_ps(r10,rswitch);
1517 d = _mm_max_ps(d,_mm_setzero_ps());
1518 d2 = _mm_mul_ps(d,d);
1519 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
1521 dsw = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
1523 /* Evaluate switch function */
1524 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1525 felec = _mm_sub_ps( _mm_mul_ps(felec,sw) , _mm_mul_ps(rinv10,_mm_mul_ps(velec,dsw)) );
1526 cutoff_mask = _mm_cmplt_ps(rsq10,rcutoff2);
1530 fscal = _mm_and_ps(fscal,cutoff_mask);
1532 fscal = _mm_andnot_ps(dummy_mask,fscal);
1534 /* Calculate temporary vectorial force */
1535 tx = _mm_mul_ps(fscal,dx10);
1536 ty = _mm_mul_ps(fscal,dy10);
1537 tz = _mm_mul_ps(fscal,dz10);
1539 /* Update vectorial force */
1540 fix1 = _mm_add_ps(fix1,tx);
1541 fiy1 = _mm_add_ps(fiy1,ty);
1542 fiz1 = _mm_add_ps(fiz1,tz);
1544 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1545 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1546 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1547 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1548 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
1552 /**************************
1553 * CALCULATE INTERACTIONS *
1554 **************************/
1556 if (gmx_mm_any_lt(rsq20,rcutoff2))
1559 r20 = _mm_mul_ps(rsq20,rinv20);
1560 r20 = _mm_andnot_ps(dummy_mask,r20);
1562 /* Compute parameters for interactions between i and j atoms */
1563 qq20 = _mm_mul_ps(iq2,jq0);
1565 /* EWALD ELECTROSTATICS */
1567 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1568 ewrt = _mm_mul_ps(r20,ewtabscale);
1569 ewitab = _mm_cvttps_epi32(ewrt);
1570 eweps = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
1571 ewitab = _mm_slli_epi32(ewitab,2);
1572 ewtabF = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1573 ewtabD = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1574 ewtabV = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1575 ewtabFn = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1576 _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1577 felec = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1578 velec = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1579 velec = _mm_mul_ps(qq20,_mm_sub_ps(rinv20,velec));
1580 felec = _mm_mul_ps(_mm_mul_ps(qq20,rinv20),_mm_sub_ps(rinvsq20,felec));
1582 d = _mm_sub_ps(r20,rswitch);
1583 d = _mm_max_ps(d,_mm_setzero_ps());
1584 d2 = _mm_mul_ps(d,d);
1585 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
1587 dsw = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
1589 /* Evaluate switch function */
1590 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1591 felec = _mm_sub_ps( _mm_mul_ps(felec,sw) , _mm_mul_ps(rinv20,_mm_mul_ps(velec,dsw)) );
1592 cutoff_mask = _mm_cmplt_ps(rsq20,rcutoff2);
1596 fscal = _mm_and_ps(fscal,cutoff_mask);
1598 fscal = _mm_andnot_ps(dummy_mask,fscal);
1600 /* Calculate temporary vectorial force */
1601 tx = _mm_mul_ps(fscal,dx20);
1602 ty = _mm_mul_ps(fscal,dy20);
1603 tz = _mm_mul_ps(fscal,dz20);
1605 /* Update vectorial force */
1606 fix2 = _mm_add_ps(fix2,tx);
1607 fiy2 = _mm_add_ps(fiy2,ty);
1608 fiz2 = _mm_add_ps(fiz2,tz);
1610 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1611 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1612 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1613 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1614 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
1618 /**************************
1619 * CALCULATE INTERACTIONS *
1620 **************************/
1622 if (gmx_mm_any_lt(rsq30,rcutoff2))
1625 r30 = _mm_mul_ps(rsq30,rinv30);
1626 r30 = _mm_andnot_ps(dummy_mask,r30);
1628 /* Compute parameters for interactions between i and j atoms */
1629 qq30 = _mm_mul_ps(iq3,jq0);
1631 /* EWALD ELECTROSTATICS */
1633 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1634 ewrt = _mm_mul_ps(r30,ewtabscale);
1635 ewitab = _mm_cvttps_epi32(ewrt);
1636 eweps = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
1637 ewitab = _mm_slli_epi32(ewitab,2);
1638 ewtabF = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1639 ewtabD = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
1640 ewtabV = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
1641 ewtabFn = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
1642 _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
1643 felec = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
1644 velec = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
1645 velec = _mm_mul_ps(qq30,_mm_sub_ps(rinv30,velec));
1646 felec = _mm_mul_ps(_mm_mul_ps(qq30,rinv30),_mm_sub_ps(rinvsq30,felec));
1648 d = _mm_sub_ps(r30,rswitch);
1649 d = _mm_max_ps(d,_mm_setzero_ps());
1650 d2 = _mm_mul_ps(d,d);
1651 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
1653 dsw = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
1655 /* Evaluate switch function */
1656 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1657 felec = _mm_sub_ps( _mm_mul_ps(felec,sw) , _mm_mul_ps(rinv30,_mm_mul_ps(velec,dsw)) );
1658 cutoff_mask = _mm_cmplt_ps(rsq30,rcutoff2);
1662 fscal = _mm_and_ps(fscal,cutoff_mask);
1664 fscal = _mm_andnot_ps(dummy_mask,fscal);
1666 /* Calculate temporary vectorial force */
1667 tx = _mm_mul_ps(fscal,dx30);
1668 ty = _mm_mul_ps(fscal,dy30);
1669 tz = _mm_mul_ps(fscal,dz30);
1671 /* Update vectorial force */
1672 fix3 = _mm_add_ps(fix3,tx);
1673 fiy3 = _mm_add_ps(fiy3,ty);
1674 fiz3 = _mm_add_ps(fiz3,tz);
1676 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1677 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1678 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1679 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1680 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
1684 /* Inner loop uses 246 flops */
1687 /* End of innermost loop */
1689 gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1690 f+i_coord_offset,fshift+i_shift_offset);
1692 /* Increment number of inner iterations */
1693 inneriter += j_index_end - j_index_start;
1695 /* Outer loop uses 24 flops */
1698 /* Increment number of outer iterations */
1701 /* Update outer/inner flops */
1703 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4_F,outeriter*24 + inneriter*246);