1 /* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
4 * This source code is part of
8 * GROningen MAchine for Chemical Simulations
10 * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
11 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
12 * Copyright (c) 2001-2012, The GROMACS development team,
13 * check out http://www.gromacs.org for more information.
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License
17 * as published by the Free Software Foundation; either version 2
18 * of the License, or (at your option) any later version.
20 * If you want to redistribute modifications, please consider that
21 * scientific software is very special. Version control is crucial -
22 * bugs must be traceable. We will be happy to consider code for
23 * inclusion in the official distribution, but derived work must not
24 * be called official GROMACS. Details are found in the README & COPYING
25 * files - if they are missing, get the official version at www.gromacs.org.
27 * To help us fund GROMACS development, we humbly ask that you cite
28 * the papers on the package - you can find them in the top README file.
30 * For more info, check our website at http://www.gromacs.org
33 * Gallium Rubidium Oxygen Manganese Argon Carbon Silicon
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 */
50 Kernel launch parameters:
51 - #blocks = #pair lists, blockId = pair list Id
52 - #threads = CL_SIZE^2
53 - shmem = CL_SIZE^2 * sizeof(float)
55 Each thread calculates an i force-component taking one pair of i-j atoms.
59 __global__ void NB_KERNEL_FUNC_NAME(k_nbnxn, _ener_prune)
61 __global__ void NB_KERNEL_FUNC_NAME(k_nbnxn, _prune)
65 __global__ void NB_KERNEL_FUNC_NAME(k_nbnxn, _ener)
67 __global__ void NB_KERNEL_FUNC_NAME(k_nbnxn)
70 (const cu_atomdata_t atdat,
71 const cu_nbparam_t nbparam,
72 const cu_plist_t plist,
75 /* convenience variables */
76 const nbnxn_sci_t *pl_sci = plist.sci;
80 nbnxn_cj4_t *pl_cj4 = plist.cj4;
81 const nbnxn_excl_t *excl = plist.excl;
82 const int *atom_types = atdat.atom_types;
83 int ntypes = atdat.ntypes;
84 const float4 *xq = atdat.xq;
86 const float3 *shift_vec = atdat.shift_vec;
87 float rcoulomb_sq = nbparam.rcoulomb_sq;
88 #ifdef VDW_CUTOFF_CHECK
89 float rvdw_sq = nbparam.rvdw_sq;
93 float two_k_rf = nbparam.two_k_rf;
96 float coulomb_tab_scale = nbparam.coulomb_tab_scale;
99 float rlist_sq = nbparam.rlist_sq;
103 float lj_shift = nbparam.sh_invrc6;
105 float beta = nbparam.ewald_beta;
106 float ewald_shift = nbparam.sh_ewald;
108 float c_rf = nbparam.c_rf;
110 float *e_lj = atdat.e_lj;
111 float *e_el = atdat.e_el;
114 /* thread/block/warp id-s */
115 unsigned int tidxi = threadIdx.x;
116 unsigned int tidxj = threadIdx.y;
117 unsigned int tidx = threadIdx.y * blockDim.x + threadIdx.x;
118 unsigned int bidx = blockIdx.x;
119 unsigned int widx = tidx / WARP_SIZE; /* warp index */
121 int sci, ci, cj, ci_offset,
123 cij4_start, cij4_end,
125 i, jm, j4, wexcl_idx;
127 r2, inv_r, inv_r2, inv_r6,
134 unsigned int wexcl, imask, mask_ji;
136 float3 xi, xj, rv, f_ij, fcj_buf, fshift_buf;
137 float3 fci_buf[NCL_PER_SUPERCL]; /* i force buffer */
140 /* shmem buffer for i x+q pre-loading */
141 extern __shared__ float4 xqib[];
142 /* shmem buffer for cj, for both warps separately */
143 int *cjs = (int *)(xqib + NCL_PER_SUPERCL * CL_SIZE);
145 /* shmem buffer for i atom-type pre-loading */
146 int *atib = (int *)(cjs + 2 * NBNXN_GPU_JGROUP_SIZE);
149 #ifndef REDUCE_SHUFFLE
150 /* shmem j force buffer */
152 float *f_buf = (float *)(atib + NCL_PER_SUPERCL * CL_SIZE);
154 float *f_buf = (float *)(cjs + 2 * NBNXN_GPU_JGROUP_SIZE);
158 nb_sci = pl_sci[bidx]; /* my i super-cluster's index = current bidx */
159 sci = nb_sci.sci; /* super-cluster */
160 cij4_start = nb_sci.cj4_ind_start; /* first ...*/
161 cij4_end = nb_sci.cj4_ind_end; /* and last index of j clusters */
163 /* Store the i-atom x and q in shared memory */
164 /* Note: the thread indexing here is inverted with respect to the
165 inner-loop as this results in slightly higher performance */
166 ci = sci * NCL_PER_SUPERCL + tidxi;
167 ai = ci * CL_SIZE + tidxj;
168 xqib[tidxi * CL_SIZE + tidxj] = xq[ai] + shift_vec[nb_sci.shift];
170 ci = sci * NCL_PER_SUPERCL + tidxj;
171 ai = ci * CL_SIZE + tidxi;
172 atib[tidxj * CL_SIZE + tidxi] = atom_types[ai];
176 for(ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
178 fci_buf[ci_offset] = make_float3(0.0f);
185 #if defined EL_EWALD || defined EL_RF
186 if (nb_sci.shift == CENTRAL && pl_cj4[cij4_start].cj[0] == sci*NCL_PER_SUPERCL)
188 /* we have the diagonal: add the charge self interaction energy term */
189 for (i = 0; i < NCL_PER_SUPERCL; i++)
191 qi = xqib[i * CL_SIZE + tidxi].w;
194 /* divide the self term equally over the j-threads */
197 E_el *= -nbparam.epsfac*0.5f*c_rf;
199 E_el *= -nbparam.epsfac*beta*M_FLOAT_1_SQRTPI; /* last factor 1/sqrt(pi) */
205 /* skip central shifts when summing shift forces */
206 if (nb_sci.shift == CENTRAL)
211 fshift_buf = make_float3(0.0f);
213 /* loop over the j clusters = seen by any of the atoms in the current super-cluster */
214 for (j4 = cij4_start; j4 < cij4_end; j4++)
216 wexcl_idx = pl_cj4[j4].imei[widx].excl_ind;
217 imask = pl_cj4[j4].imei[widx].imask;
218 wexcl = excl[wexcl_idx].pair[(tidx) & (WARP_SIZE - 1)];
224 /* Pre-load cj into shared memory on both warps separately */
225 if ((tidxj == 0 || tidxj == 4) && tidxi < NBNXN_GPU_JGROUP_SIZE)
227 cjs[tidxi + tidxj * NBNXN_GPU_JGROUP_SIZE / 4] = pl_cj4[j4].cj[tidxi];
230 /* Unrolling this loop
231 - with pruning leads to register spilling;
232 - on Kepler is much slower;
233 - doesn't work on CUDA <v4.1
234 Tested with nvcc 3.2 - 5.0.7 */
235 #if !defined PRUNE_NBL && __CUDA_ARCH__ < 300 && CUDA_VERSION >= 4010
238 for (jm = 0; jm < NBNXN_GPU_JGROUP_SIZE; jm++)
240 if (imask & (supercl_interaction_mask << (jm * NCL_PER_SUPERCL)))
242 mask_ji = (1U << (jm * NCL_PER_SUPERCL));
244 cj = cjs[jm + (tidxj & 4) * NBNXN_GPU_JGROUP_SIZE / 4];
245 aj = cj * CL_SIZE + tidxj;
247 /* load j atom data */
249 xj = make_float3(xqbuf.x, xqbuf.y, xqbuf.z);
250 qj_f = nbparam.epsfac * xqbuf.w;
251 typej = atom_types[aj];
253 fcj_buf = make_float3(0.0f);
255 /* The PME and RF kernels don't unroll with CUDA <v4.1. */
256 #if !defined PRUNE_NBL && !(CUDA_VERSION < 4010 && (defined EL_EWALD || defined EL_RF))
259 for(i = 0; i < NCL_PER_SUPERCL; i++)
263 ci_offset = i; /* i force buffer offset */
265 ci = sci * NCL_PER_SUPERCL + i; /* i cluster index */
266 ai = ci * CL_SIZE + tidxi; /* i atom index */
268 /* all threads load an atom from i cluster ci into shmem! */
269 xqbuf = xqib[i * CL_SIZE + tidxi];
270 xi = make_float3(xqbuf.x, xqbuf.y, xqbuf.z);
272 /* distance between i and j atoms */
277 /* If _none_ of the atoms pairs are in cutoff range,
278 the bit corresponding to the current
279 cluster-pair in imask gets set to 0. */
280 if (!__any(r2 < rlist_sq))
286 int_bit = (wexcl & mask_ji) ? 1.0f : 0.0f;
288 /* cutoff & exclusion check */
289 #if defined EL_EWALD || defined EL_RF
290 if (r2 < rcoulomb_sq *
291 (nb_sci.shift != CENTRAL || ci != cj || tidxj > tidxi))
293 if (r2 < rcoulomb_sq * int_bit)
296 /* load the rest of the i-atom parameters */
299 typei = atib[i * CL_SIZE + tidxi];
301 typei = atom_types[ai];
304 /* LJ 6*C6 and 12*C12 */
305 c6 = tex1Dfetch(tex_nbfp, 2 * (ntypes * typei + typej));
306 c12 = tex1Dfetch(tex_nbfp, 2 * (ntypes * typei + typej) + 1);
308 /* avoid NaN for excluded pairs at r=0 */
309 r2 += (1.0f - int_bit) * NBNXN_AVOID_SING_R2_INC;
312 inv_r2 = inv_r * inv_r;
313 inv_r6 = inv_r2 * inv_r2 * inv_r2;
314 #if defined EL_EWALD || defined EL_RF
315 /* We could mask inv_r2, but with Ewald
316 * masking both inv_r6 and F_invr is faster */
320 F_invr = inv_r6 * (c12 * inv_r6 - c6) * inv_r2;
323 E_lj_p = int_bit * (c12 * (inv_r6 * inv_r6 - lj_shift * lj_shift) * 0.08333333f - c6 * (inv_r6 - lj_shift) * 0.16666667f);
326 #ifdef VDW_CUTOFF_CHECK
327 /* this enables twin-range cut-offs (rvdw < rcoulomb <= rlist) */
328 vdw_in_range = (r2 < rvdw_sq) ? 1.0f : 0.0f;
329 F_invr *= vdw_in_range;
331 E_lj_p *= vdw_in_range;
340 F_invr += qi * qj_f * inv_r2 * inv_r;
343 F_invr += qi * qj_f * (int_bit*inv_r2 * inv_r - two_k_rf);
346 F_invr += qi * qj_f * (int_bit*inv_r2 - interpolate_coulomb_force_r(r2 * inv_r, coulomb_tab_scale)) * inv_r;
351 E_el += qi * qj_f * (inv_r - c_rf);
354 E_el += qi * qj_f * (int_bit*inv_r + 0.5f * two_k_rf * r2 - c_rf);
357 /* 1.0f - erff is faster than erfcf */
358 E_el += qi * qj_f * (inv_r * (int_bit - erff(r2 * inv_r * beta)) - int_bit * ewald_shift);
363 /* accumulate j forces in registers */
366 /* accumulate i forces in registers */
367 fci_buf[ci_offset] += f_ij;
371 /* shift the mask bit by 1 */
375 /* reduce j forces */
376 #ifdef REDUCE_SHUFFLE
377 reduce_force_j_warp_shfl(fcj_buf, f, tidxi, aj);
379 /* store j forces in shmem */
380 f_buf[ tidx] = fcj_buf.x;
381 f_buf[ FBUF_STRIDE + tidx] = fcj_buf.y;
382 f_buf[2 * FBUF_STRIDE + tidx] = fcj_buf.z;
384 reduce_force_j_generic(f_buf, f, tidxi, tidxj, aj);
389 /* Update the imask with the new one which does not contain the
390 out of range clusters anymore. */
391 pl_cj4[j4].imei[widx].imask = imask;
396 /* reduce i forces */
397 for(ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
399 ai = (sci * NCL_PER_SUPERCL + ci_offset) * CL_SIZE + tidxi;
400 #ifdef REDUCE_SHUFFLE
401 reduce_force_i_warp_shfl(fci_buf[ci_offset], f,
402 &fshift_buf, bCalcFshift,
405 f_buf[ tidx] = fci_buf[ci_offset].x;
406 f_buf[ FBUF_STRIDE + tidx] = fci_buf[ci_offset].y;
407 f_buf[2 * FBUF_STRIDE + tidx] = fci_buf[ci_offset].z;
409 reduce_force_i(f_buf, f,
410 &fshift_buf, bCalcFshift,
416 /* add up local shift forces into global mem */
417 #ifdef REDUCE_SHUFFLE
418 if (bCalcFshift && (tidxj == 0 || tidxj == 4))
420 if (bCalcFshift && tidxj == 0)
423 atomicAdd(&atdat.fshift[nb_sci.shift].x, fshift_buf.x);
424 atomicAdd(&atdat.fshift[nb_sci.shift].y, fshift_buf.y);
425 atomicAdd(&atdat.fshift[nb_sci.shift].z, fshift_buf.z);
429 #ifdef REDUCE_SHUFFLE
430 /* reduce the energies over warps and store into global memory */
431 reduce_energy_warp_shfl(E_lj, E_el, e_lj, e_el, tidx);
433 /* flush the energies to shmem and reduce them */
435 f_buf[FBUF_STRIDE + tidx] = E_el;
436 reduce_energy_pow2(f_buf + (tidx & WARP_SIZE), e_lj, e_el, tidx & ~WARP_SIZE);