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 /* Note that floating-point constants in CUDA code should be suffixed
38 * with f (e.g. 0.5f), to stop the compiler producing intermediate
39 * code that is in double precision.
42 #if __CUDA_ARCH__ >= 300
43 #define REDUCE_SHUFFLE
44 /* On Kepler pre-loading i-atom types to shmem gives a few %,
45 but on Fermi it does not */
49 #if defined EL_EWALD_ANA || defined EL_EWALD_TAB
50 /* Note: convenience macro, needs to be undef-ed at the end of the file. */
54 #if defined EL_EWALD_ANY || defined EL_RF || defined LJ_EWALD
55 /* Macro to control the calculation of exclusion forces in the kernel
56 * We do that with Ewald (elec/vdw) and RF.
58 * Note: convenience macro, needs to be undef-ed at the end of the file.
60 #define EXCLUSION_FORCES
63 #if defined LJ_EWALD_COMB_GEOM || defined LJ_EWALD_COMB_LB
64 /* Note: convenience macro, needs to be undef-ed at the end of the file. */
69 Kernel launch parameters:
70 - #blocks = #pair lists, blockId = pair list Id
71 - #threads = CL_SIZE^2
72 - shmem = CL_SIZE^2 * sizeof(float)
74 Each thread calculates an i force-component taking one pair of i-j atoms.
76 #if __CUDA_ARCH__ >= 350
77 __launch_bounds__(64, 16)
81 __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _VF_prune_cuda)
83 __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _F_prune_cuda)
87 __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _VF_cuda)
89 __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _F_cuda)
92 (const cu_atomdata_t atdat,
93 const cu_nbparam_t nbparam,
94 const cu_plist_t plist,
97 /* convenience variables */
98 const nbnxn_sci_t *pl_sci = plist.sci;
102 nbnxn_cj4_t *pl_cj4 = plist.cj4;
103 const nbnxn_excl_t *excl = plist.excl;
104 const int *atom_types = atdat.atom_types;
105 int ntypes = atdat.ntypes;
106 const float4 *xq = atdat.xq;
108 const float3 *shift_vec = atdat.shift_vec;
109 float rcoulomb_sq = nbparam.rcoulomb_sq;
110 #ifdef VDW_CUTOFF_CHECK
111 float rvdw_sq = nbparam.rvdw_sq;
115 float lje_coeff2, lje_coeff6_6;
118 float two_k_rf = nbparam.two_k_rf;
121 float coulomb_tab_scale = nbparam.coulomb_tab_scale;
124 float beta2 = nbparam.ewald_beta*nbparam.ewald_beta;
125 float beta3 = nbparam.ewald_beta*nbparam.ewald_beta*nbparam.ewald_beta;
128 float rlist_sq = nbparam.rlist_sq;
133 float beta = nbparam.ewald_beta;
134 float ewald_shift = nbparam.sh_ewald;
136 float c_rf = nbparam.c_rf;
137 #endif /* EL_EWALD_ANY */
138 float *e_lj = atdat.e_lj;
139 float *e_el = atdat.e_el;
140 #endif /* CALC_ENERGIES */
142 /* thread/block/warp id-s */
143 unsigned int tidxi = threadIdx.x;
144 unsigned int tidxj = threadIdx.y;
145 unsigned int tidx = threadIdx.y * blockDim.x + threadIdx.x;
146 unsigned int bidx = blockIdx.x;
147 unsigned int widx = tidx / WARP_SIZE; /* warp index */
149 int sci, ci, cj, ci_offset,
151 cij4_start, cij4_end,
153 i, jm, j4, wexcl_idx;
155 r2, inv_r, inv_r2, inv_r6,
162 #if defined CALC_ENERGIES || defined LJ_POT_SWITCH
165 unsigned int wexcl, imask, mask_ji;
167 float3 xi, xj, rv, f_ij, fcj_buf, fshift_buf;
168 float3 fci_buf[NCL_PER_SUPERCL]; /* i force buffer */
171 /* shmem buffer for i x+q pre-loading */
172 extern __shared__ float4 xqib[];
173 /* shmem buffer for cj, for both warps separately */
174 int *cjs = (int *)(xqib + NCL_PER_SUPERCL * CL_SIZE);
176 /* shmem buffer for i atom-type pre-loading */
177 int *atib = (int *)(cjs + 2 * NBNXN_GPU_JGROUP_SIZE);
180 #ifndef REDUCE_SHUFFLE
181 /* shmem j force buffer */
183 float *f_buf = (float *)(atib + NCL_PER_SUPERCL * CL_SIZE);
185 float *f_buf = (float *)(cjs + 2 * NBNXN_GPU_JGROUP_SIZE);
189 nb_sci = pl_sci[bidx]; /* my i super-cluster's index = current bidx */
190 sci = nb_sci.sci; /* super-cluster */
191 cij4_start = nb_sci.cj4_ind_start; /* first ...*/
192 cij4_end = nb_sci.cj4_ind_end; /* and last index of j clusters */
194 /* Store the i-atom x and q in shared memory */
195 /* Note: the thread indexing here is inverted with respect to the
196 inner-loop as this results in slightly higher performance */
197 ci = sci * NCL_PER_SUPERCL + tidxi;
198 ai = ci * CL_SIZE + tidxj;
199 xqib[tidxi * CL_SIZE + tidxj] = xq[ai] + shift_vec[nb_sci.shift];
201 ci = sci * NCL_PER_SUPERCL + tidxj;
202 ai = ci * CL_SIZE + tidxi;
203 atib[tidxj * CL_SIZE + tidxi] = atom_types[ai];
207 for (ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
209 fci_buf[ci_offset] = make_float3(0.0f);
213 /* TODO: we are trading registers with flops by keeping lje_coeff-s, try re-calculating it later */
214 lje_coeff2 = nbparam.ewaldcoeff_lj*nbparam.ewaldcoeff_lj;
215 lje_coeff6_6 = lje_coeff2*lje_coeff2*lje_coeff2*ONE_SIXTH_F;
216 #endif /* LJ_EWALD */
223 #if defined EXCLUSION_FORCES /* Ewald or RF */
224 if (nb_sci.shift == CENTRAL && pl_cj4[cij4_start].cj[0] == sci*NCL_PER_SUPERCL)
226 /* we have the diagonal: add the charge and LJ self interaction energy term */
227 for (i = 0; i < NCL_PER_SUPERCL; i++)
229 #if defined EL_EWALD_ANY || defined EL_RF
230 qi = xqib[i * CL_SIZE + tidxi].w;
236 E_lj += tex1Dfetch<float>(nbparam.nbfp_texobj, atom_types[(sci*NCL_PER_SUPERCL + i)*CL_SIZE + tidxi]*(ntypes + 1)*2);
238 E_lj += tex1Dfetch(nbfp_texref, atom_types[(sci*NCL_PER_SUPERCL + i)*CL_SIZE + tidxi]*(ntypes + 1)*2);
239 #endif /* USE_TEXOBJ */
240 #endif /* LJ_EWALD */
244 /* divide the self term(s) equally over the j-threads, then multiply with the coefficients. */
247 E_lj *= 0.5f*ONE_SIXTH_F*lje_coeff6_6;
248 #endif /* LJ_EWALD */
250 #if defined EL_EWALD_ANY || defined EL_RF
253 E_el *= -nbparam.epsfac*0.5f*c_rf;
255 E_el *= -nbparam.epsfac*beta*M_FLOAT_1_SQRTPI; /* last factor 1/sqrt(pi) */
257 #endif /* EL_EWALD_ANY || defined EL_RF */
259 #endif /* EXCLUSION_FORCES */
261 #endif /* CALC_ENERGIES */
263 /* skip central shifts when summing shift forces */
264 if (nb_sci.shift == CENTRAL)
269 fshift_buf = make_float3(0.0f);
271 /* loop over the j clusters = seen by any of the atoms in the current super-cluster */
272 for (j4 = cij4_start; j4 < cij4_end; j4++)
274 wexcl_idx = pl_cj4[j4].imei[widx].excl_ind;
275 imask = pl_cj4[j4].imei[widx].imask;
276 wexcl = excl[wexcl_idx].pair[(tidx) & (WARP_SIZE - 1)];
282 /* Pre-load cj into shared memory on both warps separately */
283 if ((tidxj == 0 || tidxj == 4) && tidxi < NBNXN_GPU_JGROUP_SIZE)
285 cjs[tidxi + tidxj * NBNXN_GPU_JGROUP_SIZE / 4] = pl_cj4[j4].cj[tidxi];
288 /* Unrolling this loop
289 - with pruning leads to register spilling;
290 - on Kepler is much slower;
291 - doesn't work on CUDA <v4.1
292 Tested with nvcc 3.2 - 5.0.7 */
293 #if !defined PRUNE_NBL && __CUDA_ARCH__ < 300 && CUDA_VERSION >= 4010
296 for (jm = 0; jm < NBNXN_GPU_JGROUP_SIZE; jm++)
298 if (imask & (supercl_interaction_mask << (jm * NCL_PER_SUPERCL)))
300 mask_ji = (1U << (jm * NCL_PER_SUPERCL));
302 cj = cjs[jm + (tidxj & 4) * NBNXN_GPU_JGROUP_SIZE / 4];
303 aj = cj * CL_SIZE + tidxj;
305 /* load j atom data */
307 xj = make_float3(xqbuf.x, xqbuf.y, xqbuf.z);
308 qj_f = nbparam.epsfac * xqbuf.w;
309 typej = atom_types[aj];
311 fcj_buf = make_float3(0.0f);
313 /* The PME and RF kernels don't unroll with CUDA <v4.1. */
314 #if !defined PRUNE_NBL && !(CUDA_VERSION < 4010 && (defined EL_EWALD_ANY || defined EL_RF))
317 for (i = 0; i < NCL_PER_SUPERCL; i++)
321 ci_offset = i; /* i force buffer offset */
323 ci = sci * NCL_PER_SUPERCL + i; /* i cluster index */
324 ai = ci * CL_SIZE + tidxi; /* i atom index */
326 /* all threads load an atom from i cluster ci into shmem! */
327 xqbuf = xqib[i * CL_SIZE + tidxi];
328 xi = make_float3(xqbuf.x, xqbuf.y, xqbuf.z);
330 /* distance between i and j atoms */
335 /* If _none_ of the atoms pairs are in cutoff range,
336 the bit corresponding to the current
337 cluster-pair in imask gets set to 0. */
338 if (!__any(r2 < rlist_sq))
344 int_bit = (wexcl & mask_ji) ? 1.0f : 0.0f;
346 /* cutoff & exclusion check */
347 #ifdef EXCLUSION_FORCES
348 if (r2 < rcoulomb_sq *
349 (nb_sci.shift != CENTRAL || ci != cj || tidxj > tidxi))
351 if (r2 < rcoulomb_sq * int_bit)
354 /* load the rest of the i-atom parameters */
357 typei = atib[i * CL_SIZE + tidxi];
359 typei = atom_types[ai];
362 /* LJ 6*C6 and 12*C12 */
364 c6 = tex1Dfetch<float>(nbparam.nbfp_texobj, 2 * (ntypes * typei + typej));
365 c12 = tex1Dfetch<float>(nbparam.nbfp_texobj, 2 * (ntypes * typei + typej) + 1);
367 c6 = tex1Dfetch(nbfp_texref, 2 * (ntypes * typei + typej));
368 c12 = tex1Dfetch(nbfp_texref, 2 * (ntypes * typei + typej) + 1);
369 #endif /* USE_TEXOBJ */
372 /* avoid NaN for excluded pairs at r=0 */
373 r2 += (1.0f - int_bit) * NBNXN_AVOID_SING_R2_INC;
376 inv_r2 = inv_r * inv_r;
377 inv_r6 = inv_r2 * inv_r2 * inv_r2;
378 #if defined EXCLUSION_FORCES
379 /* We could mask inv_r2, but with Ewald
380 * masking both inv_r6 and F_invr is faster */
382 #endif /* EXCLUSION_FORCES */
384 F_invr = inv_r6 * (c12 * inv_r6 - c6) * inv_r2;
385 #if defined CALC_ENERGIES || defined LJ_POT_SWITCH
386 E_lj_p = int_bit * (c12 * (inv_r6 * inv_r6 + nbparam.repulsion_shift.cpot)*ONE_TWELVETH_F -
387 c6 * (inv_r6 + nbparam.dispersion_shift.cpot)*ONE_SIXTH_F);
390 #ifdef LJ_FORCE_SWITCH
392 calculate_force_switch_F_E(nbparam, c6, c12, inv_r, r2, &F_invr, &E_lj_p);
394 calculate_force_switch_F(nbparam, c6, c12, inv_r, r2, &F_invr);
395 #endif /* CALC_ENERGIES */
396 #endif /* LJ_FORCE_SWITCH */
400 #ifdef LJ_EWALD_COMB_GEOM
402 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);
404 calculate_lj_ewald_comb_geom_F(nbparam, typei, typej, r2, inv_r2, lje_coeff2, lje_coeff6_6, &F_invr);
405 #endif /* CALC_ENERGIES */
406 #elif defined LJ_EWALD_COMB_LB
407 calculate_lj_ewald_comb_LB_F_E(nbparam, typei, typej, r2, inv_r2, lje_coeff2, lje_coeff6_6,
409 int_bit, &F_invr, &E_lj_p
412 #endif /* CALC_ENERGIES */
414 #endif /* LJ_EWALD_COMB_GEOM */
415 #endif /* LJ_EWALD */
417 #ifdef VDW_CUTOFF_CHECK
418 /* Separate VDW cut-off check to enable twin-range cut-offs
419 * (rvdw < rcoulomb <= rlist)
421 vdw_in_range = (r2 < rvdw_sq) ? 1.0f : 0.0f;
422 F_invr *= vdw_in_range;
424 E_lj_p *= vdw_in_range;
426 #endif /* VDW_CUTOFF_CHECK */
430 calculate_potential_switch_F_E(nbparam, c6, c12, inv_r, r2, &F_invr, &E_lj_p);
432 calculate_potential_switch_F(nbparam, c6, c12, inv_r, r2, &F_invr, &E_lj_p);
433 #endif /* CALC_ENERGIES */
434 #endif /* LJ_POT_SWITCH */
442 F_invr += qi * qj_f * inv_r2 * inv_r;
445 F_invr += qi * qj_f * (int_bit*inv_r2 * inv_r - two_k_rf);
447 #if defined EL_EWALD_ANA
448 F_invr += qi * qj_f * (int_bit*inv_r2*inv_r + pmecorrF(beta2*r2)*beta3);
449 #elif defined EL_EWALD_TAB
450 F_invr += qi * qj_f * (int_bit*inv_r2 -
452 interpolate_coulomb_force_r(nbparam.coulomb_tab_texobj, r2 * inv_r, coulomb_tab_scale)
454 interpolate_coulomb_force_r(r2 * inv_r, coulomb_tab_scale)
455 #endif /* USE_TEXOBJ */
457 #endif /* EL_EWALD_ANA/TAB */
461 E_el += qi * qj_f * (inv_r - c_rf);
464 E_el += qi * qj_f * (int_bit*inv_r + 0.5f * two_k_rf * r2 - c_rf);
467 /* 1.0f - erff is faster than erfcf */
468 E_el += qi * qj_f * (inv_r * (int_bit - erff(r2 * inv_r * beta)) - int_bit * ewald_shift);
469 #endif /* EL_EWALD_ANY */
473 /* accumulate j forces in registers */
476 /* accumulate i forces in registers */
477 fci_buf[ci_offset] += f_ij;
481 /* shift the mask bit by 1 */
485 /* reduce j forces */
486 #ifdef REDUCE_SHUFFLE
487 reduce_force_j_warp_shfl(fcj_buf, f, tidxi, aj);
489 /* store j forces in shmem */
490 f_buf[ tidx] = fcj_buf.x;
491 f_buf[ FBUF_STRIDE + tidx] = fcj_buf.y;
492 f_buf[2 * FBUF_STRIDE + tidx] = fcj_buf.z;
494 reduce_force_j_generic(f_buf, f, tidxi, tidxj, aj);
499 /* Update the imask with the new one which does not contain the
500 out of range clusters anymore. */
501 pl_cj4[j4].imei[widx].imask = imask;
506 /* reduce i forces */
507 for (ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
509 ai = (sci * NCL_PER_SUPERCL + ci_offset) * CL_SIZE + tidxi;
510 #ifdef REDUCE_SHUFFLE
511 reduce_force_i_warp_shfl(fci_buf[ci_offset], f,
512 &fshift_buf, bCalcFshift,
515 f_buf[ tidx] = fci_buf[ci_offset].x;
516 f_buf[ FBUF_STRIDE + tidx] = fci_buf[ci_offset].y;
517 f_buf[2 * FBUF_STRIDE + tidx] = fci_buf[ci_offset].z;
519 reduce_force_i(f_buf, f,
520 &fshift_buf, bCalcFshift,
526 /* add up local shift forces into global mem */
527 #ifdef REDUCE_SHUFFLE
528 if (bCalcFshift && (tidxj == 0 || tidxj == 4))
530 if (bCalcFshift && tidxj == 0)
533 atomicAdd(&atdat.fshift[nb_sci.shift].x, fshift_buf.x);
534 atomicAdd(&atdat.fshift[nb_sci.shift].y, fshift_buf.y);
535 atomicAdd(&atdat.fshift[nb_sci.shift].z, fshift_buf.z);
539 #ifdef REDUCE_SHUFFLE
540 /* reduce the energies over warps and store into global memory */
541 reduce_energy_warp_shfl(E_lj, E_el, e_lj, e_el, tidx);
543 /* flush the energies to shmem and reduce them */
545 f_buf[FBUF_STRIDE + tidx] = E_el;
546 reduce_energy_pow2(f_buf + (tidx & WARP_SIZE), e_lj, e_el, tidx & ~WARP_SIZE);
552 #undef EXCLUSION_FORCES