2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
36 #include "gromacs/math/utilities.h"
37 #include "gromacs/pbcutil/ishift.h"
38 /* Note that floating-point constants in CUDA code should be suffixed
39 * with f (e.g. 0.5f), to stop the compiler producing intermediate
40 * code that is in double precision.
43 #if __CUDA_ARCH__ >= 300
44 #define REDUCE_SHUFFLE
45 /* On Kepler pre-loading i-atom types to shmem gives a few %,
46 but on Fermi it does not */
50 #if defined EL_EWALD_ANA || defined EL_EWALD_TAB
51 /* Note: convenience macro, needs to be undef-ed at the end of the file. */
55 #if defined EL_EWALD_ANY || defined EL_RF || defined LJ_EWALD || (defined EL_CUTOFF && defined CALC_ENERGIES)
56 /* Macro to control the calculation of exclusion forces in the kernel
57 * We do that with Ewald (elec/vdw) and RF. Cut-off only has exclusion
60 * Note: convenience macro, needs to be undef-ed at the end of the file.
62 #define EXCLUSION_FORCES
65 #if defined LJ_EWALD_COMB_GEOM || defined LJ_EWALD_COMB_LB
66 /* Note: convenience macro, needs to be undef-ed at the end of the file. */
71 Kernel launch parameters:
72 - #blocks = #pair lists, blockId = pair list Id
73 - #threads = CL_SIZE^2
74 - shmem = CL_SIZE^2 * sizeof(float)
76 Each thread calculates an i force-component taking one pair of i-j atoms.
78 #if __CUDA_ARCH__ >= 350
79 __launch_bounds__(64, 16)
83 __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _VF_prune_cuda)
85 __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _F_prune_cuda)
89 __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _VF_cuda)
91 __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _F_cuda)
94 (const cu_atomdata_t atdat,
95 const cu_nbparam_t nbparam,
96 const cu_plist_t plist,
99 /* convenience variables */
100 const nbnxn_sci_t *pl_sci = plist.sci;
104 nbnxn_cj4_t *pl_cj4 = plist.cj4;
105 const nbnxn_excl_t *excl = plist.excl;
106 const int *atom_types = atdat.atom_types;
107 int ntypes = atdat.ntypes;
108 const float4 *xq = atdat.xq;
110 const float3 *shift_vec = atdat.shift_vec;
111 float rcoulomb_sq = nbparam.rcoulomb_sq;
112 #ifdef VDW_CUTOFF_CHECK
113 float rvdw_sq = nbparam.rvdw_sq;
117 float lje_coeff2, lje_coeff6_6;
120 float two_k_rf = nbparam.two_k_rf;
123 float coulomb_tab_scale = nbparam.coulomb_tab_scale;
126 float beta2 = nbparam.ewald_beta*nbparam.ewald_beta;
127 float beta3 = nbparam.ewald_beta*nbparam.ewald_beta*nbparam.ewald_beta;
130 float rlist_sq = nbparam.rlist_sq;
135 float beta = nbparam.ewald_beta;
136 float ewald_shift = nbparam.sh_ewald;
138 float c_rf = nbparam.c_rf;
139 #endif /* EL_EWALD_ANY */
140 float *e_lj = atdat.e_lj;
141 float *e_el = atdat.e_el;
142 #endif /* CALC_ENERGIES */
144 /* thread/block/warp id-s */
145 unsigned int tidxi = threadIdx.x;
146 unsigned int tidxj = threadIdx.y;
147 unsigned int tidx = threadIdx.y * blockDim.x + threadIdx.x;
148 unsigned int bidx = blockIdx.x;
149 unsigned int widx = tidx / WARP_SIZE; /* warp index */
151 int sci, ci, cj, ci_offset,
153 cij4_start, cij4_end,
155 i, jm, j4, wexcl_idx;
157 r2, inv_r, inv_r2, inv_r6,
164 #if defined CALC_ENERGIES || defined LJ_POT_SWITCH
167 unsigned int wexcl, imask, mask_ji;
169 float3 xi, xj, rv, f_ij, fcj_buf, fshift_buf;
170 float3 fci_buf[NCL_PER_SUPERCL]; /* i force buffer */
173 /* shmem buffer for i x+q pre-loading */
174 extern __shared__ float4 xqib[];
175 /* shmem buffer for cj, for both warps separately */
176 int *cjs = (int *)(xqib + NCL_PER_SUPERCL * CL_SIZE);
178 /* shmem buffer for i atom-type pre-loading */
179 int *atib = (int *)(cjs + 2 * NBNXN_GPU_JGROUP_SIZE);
182 #ifndef REDUCE_SHUFFLE
183 /* shmem j force buffer */
185 float *f_buf = (float *)(atib + NCL_PER_SUPERCL * CL_SIZE);
187 float *f_buf = (float *)(cjs + 2 * NBNXN_GPU_JGROUP_SIZE);
191 nb_sci = pl_sci[bidx]; /* my i super-cluster's index = current bidx */
192 sci = nb_sci.sci; /* super-cluster */
193 cij4_start = nb_sci.cj4_ind_start; /* first ...*/
194 cij4_end = nb_sci.cj4_ind_end; /* and last index of j clusters */
196 /* Pre-load i-atom x and q into shared memory */
197 ci = sci * NCL_PER_SUPERCL + tidxj;
198 ai = ci * CL_SIZE + tidxi;
199 xqib[tidxj * CL_SIZE + tidxi] = xq[ai] + shift_vec[nb_sci.shift];
201 /* Pre-load the i-atom types into shared memory */
202 atib[tidxj * CL_SIZE + tidxi] = atom_types[ai];
206 for (ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
208 fci_buf[ci_offset] = make_float3(0.0f);
212 /* TODO: we are trading registers with flops by keeping lje_coeff-s, try re-calculating it later */
213 lje_coeff2 = nbparam.ewaldcoeff_lj*nbparam.ewaldcoeff_lj;
214 lje_coeff6_6 = lje_coeff2*lje_coeff2*lje_coeff2*ONE_SIXTH_F;
215 #endif /* LJ_EWALD */
222 #if defined EXCLUSION_FORCES /* Ewald or RF */
223 if (nb_sci.shift == CENTRAL && pl_cj4[cij4_start].cj[0] == sci*NCL_PER_SUPERCL)
225 /* we have the diagonal: add the charge and LJ self interaction energy term */
226 for (i = 0; i < NCL_PER_SUPERCL; i++)
228 #if defined EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF
229 qi = xqib[i * CL_SIZE + tidxi].w;
235 E_lj += tex1Dfetch<float>(nbparam.nbfp_texobj, atom_types[(sci*NCL_PER_SUPERCL + i)*CL_SIZE + tidxi]*(ntypes + 1)*2);
237 E_lj += tex1Dfetch(nbfp_texref, atom_types[(sci*NCL_PER_SUPERCL + i)*CL_SIZE + tidxi]*(ntypes + 1)*2);
238 #endif /* USE_TEXOBJ */
239 #endif /* LJ_EWALD */
243 /* divide the self term(s) equally over the j-threads, then multiply with the coefficients. */
246 E_lj *= 0.5f*ONE_SIXTH_F*lje_coeff6_6;
247 #endif /* LJ_EWALD */
249 #if defined EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF
251 #if defined EL_RF || defined EL_CUTOFF
252 E_el *= -nbparam.epsfac*0.5f*c_rf;
254 E_el *= -nbparam.epsfac*beta*M_FLOAT_1_SQRTPI; /* last factor 1/sqrt(pi) */
256 #endif /* EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF */
258 #endif /* EXCLUSION_FORCES */
260 #endif /* CALC_ENERGIES */
262 /* skip central shifts when summing shift forces */
263 if (nb_sci.shift == CENTRAL)
268 fshift_buf = make_float3(0.0f);
270 /* loop over the j clusters = seen by any of the atoms in the current super-cluster */
271 for (j4 = cij4_start; j4 < cij4_end; j4++)
273 wexcl_idx = pl_cj4[j4].imei[widx].excl_ind;
274 imask = pl_cj4[j4].imei[widx].imask;
275 wexcl = excl[wexcl_idx].pair[(tidx) & (WARP_SIZE - 1)];
281 /* Pre-load cj into shared memory on both warps separately */
282 if ((tidxj == 0 || tidxj == 4) && tidxi < NBNXN_GPU_JGROUP_SIZE)
284 cjs[tidxi + tidxj * NBNXN_GPU_JGROUP_SIZE / 4] = pl_cj4[j4].cj[tidxi];
287 /* Unrolling this loop
288 - with pruning leads to register spilling;
289 - on Kepler is much slower;
290 - doesn't work on CUDA <v4.1
291 Tested with nvcc 3.2 - 5.0.7 */
292 #if !defined PRUNE_NBL && __CUDA_ARCH__ < 300 && CUDA_VERSION >= 4010
295 for (jm = 0; jm < NBNXN_GPU_JGROUP_SIZE; jm++)
297 if (imask & (supercl_interaction_mask << (jm * NCL_PER_SUPERCL)))
299 mask_ji = (1U << (jm * NCL_PER_SUPERCL));
301 cj = cjs[jm + (tidxj & 4) * NBNXN_GPU_JGROUP_SIZE / 4];
302 aj = cj * CL_SIZE + tidxj;
304 /* load j atom data */
306 xj = make_float3(xqbuf.x, xqbuf.y, xqbuf.z);
307 qj_f = nbparam.epsfac * xqbuf.w;
308 typej = atom_types[aj];
310 fcj_buf = make_float3(0.0f);
312 /* The PME and RF kernels don't unroll with CUDA <v4.1. */
313 #if !defined PRUNE_NBL && !(CUDA_VERSION < 4010 && defined EXCLUSION_FORCES)
316 for (i = 0; i < NCL_PER_SUPERCL; i++)
320 ci_offset = i; /* i force buffer offset */
322 ci = sci * NCL_PER_SUPERCL + i; /* i cluster index */
323 ai = ci * CL_SIZE + tidxi; /* i atom index */
325 /* all threads load an atom from i cluster ci into shmem! */
326 xqbuf = xqib[i * CL_SIZE + tidxi];
327 xi = make_float3(xqbuf.x, xqbuf.y, xqbuf.z);
329 /* distance between i and j atoms */
334 /* If _none_ of the atoms pairs are in cutoff range,
335 the bit corresponding to the current
336 cluster-pair in imask gets set to 0. */
337 if (!__any(r2 < rlist_sq))
343 int_bit = (wexcl & mask_ji) ? 1.0f : 0.0f;
345 /* cutoff & exclusion check */
346 #ifdef EXCLUSION_FORCES
347 if (r2 < rcoulomb_sq *
348 (nb_sci.shift != CENTRAL || ci != cj || tidxj > tidxi))
350 if (r2 < rcoulomb_sq * int_bit)
353 /* load the rest of the i-atom parameters */
356 typei = atib[i * CL_SIZE + tidxi];
358 typei = atom_types[ai];
361 /* LJ 6*C6 and 12*C12 */
363 c6 = tex1Dfetch<float>(nbparam.nbfp_texobj, 2 * (ntypes * typei + typej));
364 c12 = tex1Dfetch<float>(nbparam.nbfp_texobj, 2 * (ntypes * typei + typej) + 1);
366 c6 = tex1Dfetch(nbfp_texref, 2 * (ntypes * typei + typej));
367 c12 = tex1Dfetch(nbfp_texref, 2 * (ntypes * typei + typej) + 1);
368 #endif /* USE_TEXOBJ */
371 /* avoid NaN for excluded pairs at r=0 */
372 r2 += (1.0f - int_bit) * NBNXN_AVOID_SING_R2_INC;
375 inv_r2 = inv_r * inv_r;
376 inv_r6 = inv_r2 * inv_r2 * inv_r2;
377 #if defined EXCLUSION_FORCES
378 /* We could mask inv_r2, but with Ewald
379 * masking both inv_r6 and F_invr is faster */
381 #endif /* EXCLUSION_FORCES */
383 F_invr = inv_r6 * (c12 * inv_r6 - c6) * inv_r2;
384 #if defined CALC_ENERGIES || defined LJ_POT_SWITCH
385 E_lj_p = int_bit * (c12 * (inv_r6 * inv_r6 + nbparam.repulsion_shift.cpot)*ONE_TWELVETH_F -
386 c6 * (inv_r6 + nbparam.dispersion_shift.cpot)*ONE_SIXTH_F);
389 #ifdef LJ_FORCE_SWITCH
391 calculate_force_switch_F_E(nbparam, c6, c12, inv_r, r2, &F_invr, &E_lj_p);
393 calculate_force_switch_F(nbparam, c6, c12, inv_r, r2, &F_invr);
394 #endif /* CALC_ENERGIES */
395 #endif /* LJ_FORCE_SWITCH */
399 #ifdef LJ_EWALD_COMB_GEOM
401 calculate_lj_ewald_comb_geom_F_E(nbparam, typei, typej, r2, inv_r2, lje_coeff2, lje_coeff6_6, int_bit, &F_invr, &E_lj_p);
403 calculate_lj_ewald_comb_geom_F(nbparam, typei, typej, r2, inv_r2, lje_coeff2, lje_coeff6_6, &F_invr);
404 #endif /* CALC_ENERGIES */
405 #elif defined LJ_EWALD_COMB_LB
406 calculate_lj_ewald_comb_LB_F_E(nbparam, typei, typej, r2, inv_r2, lje_coeff2, lje_coeff6_6,
408 int_bit, &F_invr, &E_lj_p
411 #endif /* CALC_ENERGIES */
413 #endif /* LJ_EWALD_COMB_GEOM */
414 #endif /* LJ_EWALD */
416 #ifdef VDW_CUTOFF_CHECK
417 /* Separate VDW cut-off check to enable twin-range cut-offs
418 * (rvdw < rcoulomb <= rlist)
420 vdw_in_range = (r2 < rvdw_sq) ? 1.0f : 0.0f;
421 F_invr *= vdw_in_range;
423 E_lj_p *= vdw_in_range;
425 #endif /* VDW_CUTOFF_CHECK */
429 calculate_potential_switch_F_E(nbparam, c6, c12, inv_r, r2, &F_invr, &E_lj_p);
431 calculate_potential_switch_F(nbparam, c6, c12, inv_r, r2, &F_invr, &E_lj_p);
432 #endif /* CALC_ENERGIES */
433 #endif /* LJ_POT_SWITCH */
441 #ifdef EXCLUSION_FORCES
442 F_invr += qi * qj_f * int_bit * inv_r2 * inv_r;
444 F_invr += qi * qj_f * inv_r2 * inv_r;
448 F_invr += qi * qj_f * (int_bit*inv_r2 * inv_r - two_k_rf);
450 #if defined EL_EWALD_ANA
451 F_invr += qi * qj_f * (int_bit*inv_r2*inv_r + pmecorrF(beta2*r2)*beta3);
452 #elif defined EL_EWALD_TAB
453 F_invr += qi * qj_f * (int_bit*inv_r2 -
455 interpolate_coulomb_force_r(nbparam.coulomb_tab_texobj, r2 * inv_r, coulomb_tab_scale)
457 interpolate_coulomb_force_r(r2 * inv_r, coulomb_tab_scale)
458 #endif /* USE_TEXOBJ */
460 #endif /* EL_EWALD_ANA/TAB */
464 E_el += qi * qj_f * (int_bit*inv_r - c_rf);
467 E_el += qi * qj_f * (int_bit*inv_r + 0.5f * two_k_rf * r2 - c_rf);
470 /* 1.0f - erff is faster than erfcf */
471 E_el += qi * qj_f * (inv_r * (int_bit - erff(r2 * inv_r * beta)) - int_bit * ewald_shift);
472 #endif /* EL_EWALD_ANY */
476 /* accumulate j forces in registers */
479 /* accumulate i forces in registers */
480 fci_buf[ci_offset] += f_ij;
484 /* shift the mask bit by 1 */
488 /* reduce j forces */
489 #ifdef REDUCE_SHUFFLE
490 reduce_force_j_warp_shfl(fcj_buf, f, tidxi, aj);
492 /* store j forces in shmem */
493 f_buf[ tidx] = fcj_buf.x;
494 f_buf[ FBUF_STRIDE + tidx] = fcj_buf.y;
495 f_buf[2 * FBUF_STRIDE + tidx] = fcj_buf.z;
497 reduce_force_j_generic(f_buf, f, tidxi, tidxj, aj);
502 /* Update the imask with the new one which does not contain the
503 out of range clusters anymore. */
504 pl_cj4[j4].imei[widx].imask = imask;
509 /* reduce i forces */
510 for (ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
512 ai = (sci * NCL_PER_SUPERCL + ci_offset) * CL_SIZE + tidxi;
513 #ifdef REDUCE_SHUFFLE
514 reduce_force_i_warp_shfl(fci_buf[ci_offset], f,
515 &fshift_buf, bCalcFshift,
518 f_buf[ tidx] = fci_buf[ci_offset].x;
519 f_buf[ FBUF_STRIDE + tidx] = fci_buf[ci_offset].y;
520 f_buf[2 * FBUF_STRIDE + tidx] = fci_buf[ci_offset].z;
522 reduce_force_i(f_buf, f,
523 &fshift_buf, bCalcFshift,
529 /* add up local shift forces into global mem */
530 #ifdef REDUCE_SHUFFLE
531 if (bCalcFshift && (tidxj == 0 || tidxj == 4))
533 if (bCalcFshift && tidxj == 0)
536 atomicAdd(&atdat.fshift[nb_sci.shift].x, fshift_buf.x);
537 atomicAdd(&atdat.fshift[nb_sci.shift].y, fshift_buf.y);
538 atomicAdd(&atdat.fshift[nb_sci.shift].z, fshift_buf.z);
542 #ifdef REDUCE_SHUFFLE
543 /* reduce the energies over warps and store into global memory */
544 reduce_energy_warp_shfl(E_lj, E_el, e_lj, e_el, tidx);
546 /* flush the energies to shmem and reduce them */
548 f_buf[FBUF_STRIDE + tidx] = E_el;
549 reduce_energy_pow2(f_buf + (tidx & WARP_SIZE), e_lj, e_el, tidx & ~WARP_SIZE);
555 #undef EXCLUSION_FORCES