2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016, 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.
38 * CUDA non-bonded kernel used through preprocessor-based code generation
39 * of multiple kernel flavors, see nbnxn_cuda_kernels.cuh.
41 * NOTE: No include fence as it is meant to be included multiple times.
43 * \author Szilárd Páll <pall.szilard@gmail.com>
44 * \author Berk Hess <hess@kth.se>
45 * \ingroup module_mdlib
48 #include "gromacs/gpu_utils/cuda_arch_utils.cuh"
49 #include "gromacs/math/utilities.h"
50 #include "gromacs/pbcutil/ishift.h"
51 /* Note that floating-point constants in CUDA code should be suffixed
52 * with f (e.g. 0.5f), to stop the compiler producing intermediate
53 * code that is in double precision.
56 #if GMX_PTX_ARCH < 300
57 #error "nbnxn_cuda_kernel.cuh included with GMX_PTX_ARCH < 300"
60 #if defined EL_EWALD_ANA || defined EL_EWALD_TAB
61 /* Note: convenience macro, needs to be undef-ed at the end of the file. */
65 #if defined EL_EWALD_ANY || defined EL_RF || defined LJ_EWALD || (defined EL_CUTOFF && defined CALC_ENERGIES)
66 /* Macro to control the calculation of exclusion forces in the kernel
67 * We do that with Ewald (elec/vdw) and RF. Cut-off only has exclusion
70 * Note: convenience macro, needs to be undef-ed at the end of the file.
72 #define EXCLUSION_FORCES
75 #if defined LJ_EWALD_COMB_GEOM || defined LJ_EWALD_COMB_LB
76 /* Note: convenience macro, needs to be undef-ed at the end of the file. */
80 #if defined LJ_COMB_GEOM || defined LJ_COMB_LB
85 Kernel launch parameters:
86 - #blocks = #pair lists, blockId = pair list Id
87 - #threads = NTHREAD_Z * c_clSize^2
88 - shmem = see nbnxn_cuda.cu:calc_shmem_required()
90 Each thread calculates an i force-component taking one pair of i-j atoms.
94 /*! \brief Compute capability dependent definition of kernel launch configuration parameters.
96 * NTHREAD_Z controls the number of j-clusters processed concurrently on NTHREAD_Z
97 * warp-pairs per block.
99 * - On CC 2.0-3.5, 5.0, and 5.2, NTHREAD_Z == 1, translating to 64 th/block with 16
100 * blocks/multiproc, is the fastest even though this setup gives low occupancy.
101 * NTHREAD_Z > 1 results in excessive register spilling unless the minimum blocks
102 * per multiprocessor is reduced proportionally to get the original number of max
103 * threads in flight (and slightly lower performance).
104 * - On CC 3.7 and 6.0 there are enough registers to double the number of threads; using
105 * NTHREADS_Z == 2 is fastest with 16 blocks (TODO: test with RF and other kernels
106 * with low-register use).
108 * Note that the current kernel implementation only supports NTHREAD_Z > 1 with
109 * shuffle-based reduction, hence CC >= 3.0.
112 /* Kernel launch bounds for different compute capabilities. The value of NTHREAD_Z
113 * determines the number of threads per block and it is chosen such that
114 * 16 blocks/multiprocessor can be kept in flight.
115 * - CC 3.0/3.5/5.x, >=6.1: NTHREAD_Z=1, (64, 16) bounds
116 * - CC 3.7, 6.0: NTHREAD_Z=2, (128, 16) bounds
118 * Note: convenience macros, need to be undef-ed at the end of the file.
120 #if GMX_PTX_ARCH == 370 || GMX_PTX_ARCH == 600
121 #define NTHREAD_Z (2)
122 #define MIN_BLOCKS_PER_MP (16)
124 #define NTHREAD_Z (1)
125 #define MIN_BLOCKS_PER_MP (16)
126 #endif /* GMX_PTX_ARCH == 370 || GMX_PTX_ARCH == 600 */
127 #define THREADS_PER_BLOCK (c_clSize*c_clSize*NTHREAD_Z)
129 #if GMX_PTX_ARCH >= 350
130 #if (GMX_PTX_ARCH <= 210) && (NTHREAD_Z > 1)
131 #error NTHREAD_Z > 1 will give incorrect results on CC 2.x
134 __launch_bounds__(THREADS_PER_BLOCK, MIN_BLOCKS_PER_MP)
136 __launch_bounds__(THREADS_PER_BLOCK)
137 #endif /* GMX_PTX_ARCH >= 350 */
140 __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _VF_prune_cuda)
142 __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _F_prune_cuda)
143 #endif /* CALC_ENERGIES */
146 __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _VF_cuda)
148 __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _F_cuda)
149 #endif /* CALC_ENERGIES */
150 #endif /* PRUNE_NBL */
151 (const cu_atomdata_t atdat,
152 const cu_nbparam_t nbparam,
153 const cu_plist_t plist,
155 #ifdef FUNCTION_DECLARATION_ONLY
156 ; /* Only do function declaration, omit the function body. */
159 /* convenience variables */
160 const nbnxn_sci_t *pl_sci = plist.sci;
164 nbnxn_cj4_t *pl_cj4 = plist.cj4;
165 const nbnxn_excl_t *excl = plist.excl;
167 const int *atom_types = atdat.atom_types;
168 int ntypes = atdat.ntypes;
170 const float2 *lj_comb = atdat.lj_comb;
171 float2 ljcp_i, ljcp_j;
173 const float4 *xq = atdat.xq;
175 const float3 *shift_vec = atdat.shift_vec;
176 float rcoulomb_sq = nbparam.rcoulomb_sq;
177 #ifdef VDW_CUTOFF_CHECK
178 float rvdw_sq = nbparam.rvdw_sq;
182 float lje_coeff2, lje_coeff6_6;
185 float two_k_rf = nbparam.two_k_rf;
188 float coulomb_tab_scale = nbparam.coulomb_tab_scale;
191 float beta2 = nbparam.ewald_beta*nbparam.ewald_beta;
192 float beta3 = nbparam.ewald_beta*nbparam.ewald_beta*nbparam.ewald_beta;
195 float rlist_sq = nbparam.rlist_sq;
200 float beta = nbparam.ewald_beta;
201 float ewald_shift = nbparam.sh_ewald;
203 float c_rf = nbparam.c_rf;
204 #endif /* EL_EWALD_ANY */
205 float *e_lj = atdat.e_lj;
206 float *e_el = atdat.e_el;
207 #endif /* CALC_ENERGIES */
209 /* thread/block/warp id-s */
210 unsigned int tidxi = threadIdx.x;
211 unsigned int tidxj = threadIdx.y;
212 unsigned int tidx = threadIdx.y * blockDim.x + threadIdx.x;
214 unsigned int tidxz = 0;
216 unsigned int tidxz = threadIdx.z;
218 unsigned int bidx = blockIdx.x;
219 unsigned int widx = tidx / warp_size; /* warp index */
223 cij4_start, cij4_end;
227 int i, jm, j4, wexcl_idx;
230 #if !defined LJ_COMB_LB || defined CALC_ENERGIES
231 float inv_r6, c6, c12;
234 float sigma, epsilon;
241 #if defined CALC_ENERGIES || defined LJ_POT_SWITCH
244 unsigned int wexcl, imask, mask_ji;
246 float3 xi, xj, rv, f_ij, fcj_buf;
247 float3 fci_buf[c_numClPerSupercl]; /* i force buffer */
250 /*! i-cluster interaction mask for a super-cluster with all c_numClPerSupercl=8 bits set */
251 const unsigned superClInteractionMask = ((1U << c_numClPerSupercl) - 1U);
253 /* shmem buffer for i x+q pre-loading */
254 extern __shared__ float4 xqib[];
256 /* shmem buffer for cj, for each warp separately */
257 int *cjs = ((int *)(xqib + c_numClPerSupercl * c_clSize)) + tidxz * c_nbnxnGpuClusterpairSplit * c_nbnxnGpuJgroupSize;
258 int *cjs_end = ((int *)(xqib + c_numClPerSupercl * c_clSize)) + NTHREAD_Z * c_nbnxnGpuClusterpairSplit * c_nbnxnGpuJgroupSize;
260 /* shmem buffer for i atom-type pre-loading */
263 /* shmem buffer for i-atom LJ combination rule parameters */
264 float2 *ljcpib = (float2 *)cjs_end;
267 nb_sci = pl_sci[bidx]; /* my i super-cluster's index = current bidx */
268 sci = nb_sci.sci; /* super-cluster */
269 cij4_start = nb_sci.cj4_ind_start; /* first ...*/
270 cij4_end = nb_sci.cj4_ind_end; /* and last index of j clusters */
274 /* Pre-load i-atom x and q into shared memory */
275 ci = sci * c_numClPerSupercl + tidxj;
276 ai = ci * c_clSize + tidxi;
278 xqbuf = xq[ai] + shift_vec[nb_sci.shift];
279 xqbuf.w *= nbparam.epsfac;
280 xqib[tidxj * c_clSize + tidxi] = xqbuf;
283 /* Pre-load the i-atom types into shared memory */
284 atib[tidxj * c_clSize + tidxi] = atom_types[ai];
286 /* Pre-load the LJ combination parameters into shared memory */
287 ljcpib[tidxj * c_clSize + tidxi] = lj_comb[ai];
292 for (i = 0; i < c_numClPerSupercl; i++)
294 fci_buf[i] = make_float3(0.0f);
298 /* TODO: we are trading registers with flops by keeping lje_coeff-s, try re-calculating it later */
299 lje_coeff2 = nbparam.ewaldcoeff_lj*nbparam.ewaldcoeff_lj;
300 lje_coeff6_6 = lje_coeff2*lje_coeff2*lje_coeff2*c_oneSixth;
308 #ifdef EXCLUSION_FORCES /* Ewald or RF */
309 if (nb_sci.shift == CENTRAL && pl_cj4[cij4_start].cj[0] == sci*c_numClPerSupercl)
311 /* we have the diagonal: add the charge and LJ self interaction energy term */
312 for (i = 0; i < c_numClPerSupercl; i++)
314 #if defined EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF
315 qi = xqib[i * c_clSize + tidxi].w;
320 E_lj += tex1Dfetch<float>(nbparam.nbfp_texobj, atom_types[(sci*c_numClPerSupercl + i)*c_clSize + tidxi]*(ntypes + 1)*2);
324 /* divide the self term(s) equally over the j-threads, then multiply with the coefficients. */
326 E_lj /= c_clSize*NTHREAD_Z;
327 E_lj *= 0.5f*c_oneSixth*lje_coeff6_6;
330 #if defined EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF
331 /* Correct for epsfac^2 due to adding qi^2 */
332 E_el /= nbparam.epsfac*c_clSize*NTHREAD_Z;
333 #if defined EL_RF || defined EL_CUTOFF
336 E_el *= -beta*M_FLOAT_1_SQRTPI; /* last factor 1/sqrt(pi) */
338 #endif /* EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF */
340 #endif /* EXCLUSION_FORCES */
342 #endif /* CALC_ENERGIES */
344 /* loop over the j clusters = seen by any of the atoms in the current super-cluster */
345 for (j4 = cij4_start + tidxz; j4 < cij4_end; j4 += NTHREAD_Z)
347 wexcl_idx = pl_cj4[j4].imei[widx].excl_ind;
348 imask = pl_cj4[j4].imei[widx].imask;
349 wexcl = excl[wexcl_idx].pair[(tidx) & (warp_size - 1)];
355 /* Pre-load cj into shared memory on both warps separately */
356 if ((tidxj == 0 || tidxj == 4) && tidxi < c_nbnxnGpuJgroupSize)
358 cjs[tidxi + tidxj * c_nbnxnGpuJgroupSize/c_splitClSize] = pl_cj4[j4].cj[tidxi];
361 /* Unrolling this loop
362 - with pruning leads to register spilling;
363 - on Kepler and later it is much slower;
364 Tested with up to nvcc 7.5 */
365 for (jm = 0; jm < c_nbnxnGpuJgroupSize; jm++)
367 if (imask & (superClInteractionMask << (jm * c_numClPerSupercl)))
369 mask_ji = (1U << (jm * c_numClPerSupercl));
371 cj = cjs[jm + (tidxj & 4) * c_nbnxnGpuJgroupSize/c_splitClSize];
372 aj = cj * c_clSize + tidxj;
374 /* load j atom data */
376 xj = make_float3(xqbuf.x, xqbuf.y, xqbuf.z);
379 typej = atom_types[aj];
381 ljcp_j = lj_comb[aj];
384 fcj_buf = make_float3(0.0f);
386 #if !defined PRUNE_NBL
389 for (i = 0; i < c_numClPerSupercl; i++)
393 ci = sci * c_numClPerSupercl + i; /* i cluster index */
395 /* all threads load an atom from i cluster ci into shmem! */
396 xqbuf = xqib[i * c_clSize + tidxi];
397 xi = make_float3(xqbuf.x, xqbuf.y, xqbuf.z);
399 /* distance between i and j atoms */
404 /* If _none_ of the atoms pairs are in cutoff range,
405 the bit corresponding to the current
406 cluster-pair in imask gets set to 0. */
407 if (!__any(r2 < rlist_sq))
413 int_bit = (wexcl & mask_ji) ? 1.0f : 0.0f;
415 /* cutoff & exclusion check */
416 #ifdef EXCLUSION_FORCES
417 if (r2 < rcoulomb_sq *
418 (nb_sci.shift != CENTRAL || ci != cj || tidxj > tidxi))
420 if (r2 < rcoulomb_sq * int_bit)
423 /* load the rest of the i-atom parameters */
427 /* LJ 6*C6 and 12*C12 */
428 typei = atib[i * c_clSize + tidxi];
429 c6 = tex1Dfetch<float>(nbparam.nbfp_texobj, 2 * (ntypes * typei + typej));
430 c12 = tex1Dfetch<float>(nbparam.nbfp_texobj, 2 * (ntypes * typei + typej) + 1);
432 ljcp_i = ljcpib[i * c_clSize + tidxi];
434 c6 = ljcp_i.x * ljcp_j.x;
435 c12 = ljcp_i.y * ljcp_j.y;
437 /* LJ 2^(1/6)*sigma and 12*epsilon */
438 sigma = ljcp_i.x + ljcp_j.x;
439 epsilon = ljcp_i.y * ljcp_j.y;
440 #if defined CALC_ENERGIES || defined LJ_FORCE_SWITCH || defined LJ_POT_SWITCH
441 convert_sigma_epsilon_to_c6_c12(sigma, epsilon, &c6, &c12);
443 #endif /* LJ_COMB_GEOM */
446 // Ensure distance do not become so small that r^-12 overflows
447 r2 = max(r2, NBNXN_MIN_RSQ);
450 inv_r2 = inv_r * inv_r;
451 #if !defined LJ_COMB_LB || defined CALC_ENERGIES
452 inv_r6 = inv_r2 * inv_r2 * inv_r2;
453 #ifdef EXCLUSION_FORCES
454 /* We could mask inv_r2, but with Ewald
455 * masking both inv_r6 and F_invr is faster */
457 #endif /* EXCLUSION_FORCES */
459 F_invr = inv_r6 * (c12 * inv_r6 - c6) * inv_r2;
460 #if defined CALC_ENERGIES || defined LJ_POT_SWITCH
461 E_lj_p = int_bit * (c12 * (inv_r6 * inv_r6 + nbparam.repulsion_shift.cpot)*c_oneTwelveth -
462 c6 * (inv_r6 + nbparam.dispersion_shift.cpot)*c_oneSixth);
464 #else /* !LJ_COMB_LB || CALC_ENERGIES */
465 float sig_r = sigma*inv_r;
466 float sig_r2 = sig_r*sig_r;
467 float sig_r6 = sig_r2*sig_r2*sig_r2;
468 #ifdef EXCLUSION_FORCES
470 #endif /* EXCLUSION_FORCES */
472 F_invr = epsilon * sig_r6 * (sig_r6 - 1.0f) * inv_r2;
473 #endif /* !LJ_COMB_LB || CALC_ENERGIES */
475 #ifdef LJ_FORCE_SWITCH
477 calculate_force_switch_F_E(nbparam, c6, c12, inv_r, r2, &F_invr, &E_lj_p);
479 calculate_force_switch_F(nbparam, c6, c12, inv_r, r2, &F_invr);
480 #endif /* CALC_ENERGIES */
481 #endif /* LJ_FORCE_SWITCH */
485 #ifdef LJ_EWALD_COMB_GEOM
487 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);
489 calculate_lj_ewald_comb_geom_F(nbparam, typei, typej, r2, inv_r2, lje_coeff2, lje_coeff6_6, &F_invr);
490 #endif /* CALC_ENERGIES */
491 #elif defined LJ_EWALD_COMB_LB
492 calculate_lj_ewald_comb_LB_F_E(nbparam, typei, typej, r2, inv_r2, lje_coeff2, lje_coeff6_6,
494 int_bit, &F_invr, &E_lj_p
497 #endif /* CALC_ENERGIES */
499 #endif /* LJ_EWALD_COMB_GEOM */
500 #endif /* LJ_EWALD */
504 calculate_potential_switch_F_E(nbparam, c6, c12, inv_r, r2, &F_invr, &E_lj_p);
506 calculate_potential_switch_F(nbparam, c6, c12, inv_r, r2, &F_invr, &E_lj_p);
507 #endif /* CALC_ENERGIES */
508 #endif /* LJ_POT_SWITCH */
510 #ifdef VDW_CUTOFF_CHECK
511 /* Separate VDW cut-off check to enable twin-range cut-offs
512 * (rvdw < rcoulomb <= rlist)
514 vdw_in_range = (r2 < rvdw_sq) ? 1.0f : 0.0f;
515 F_invr *= vdw_in_range;
517 E_lj_p *= vdw_in_range;
519 #endif /* VDW_CUTOFF_CHECK */
527 #ifdef EXCLUSION_FORCES
528 F_invr += qi * qj_f * int_bit * inv_r2 * inv_r;
530 F_invr += qi * qj_f * inv_r2 * inv_r;
534 F_invr += qi * qj_f * (int_bit*inv_r2 * inv_r - two_k_rf);
536 #if defined EL_EWALD_ANA
537 F_invr += qi * qj_f * (int_bit*inv_r2*inv_r + pmecorrF(beta2*r2)*beta3);
538 #elif defined EL_EWALD_TAB
539 F_invr += qi * qj_f * (int_bit*inv_r2 -
540 interpolate_coulomb_force_r(nbparam.coulomb_tab_texobj, r2 * inv_r, coulomb_tab_scale)) * inv_r;
541 #endif /* EL_EWALD_ANA/TAB */
545 E_el += qi * qj_f * (int_bit*inv_r - c_rf);
548 E_el += qi * qj_f * (int_bit*inv_r + 0.5f * two_k_rf * r2 - c_rf);
551 /* 1.0f - erff is faster than erfcf */
552 E_el += qi * qj_f * (inv_r * (int_bit - erff(r2 * inv_r * beta)) - int_bit * ewald_shift);
553 #endif /* EL_EWALD_ANY */
557 /* accumulate j forces in registers */
560 /* accumulate i forces in registers */
565 /* shift the mask bit by 1 */
569 /* reduce j forces */
570 reduce_force_j_warp_shfl(fcj_buf, f, tidxi, aj);
574 /* Update the imask with the new one which does not contain the
575 out of range clusters anymore. */
576 pl_cj4[j4].imei[widx].imask = imask;
581 /* skip central shifts when summing shift forces */
582 if (nb_sci.shift == CENTRAL)
587 float fshift_buf = 0.0f;
589 /* reduce i forces */
590 for (i = 0; i < c_numClPerSupercl; i++)
592 ai = (sci * c_numClPerSupercl + i) * c_clSize + tidxi;
593 reduce_force_i_warp_shfl(fci_buf[i], f,
594 &fshift_buf, bCalcFshift,
598 /* add up local shift forces into global mem, tidxj indexes x,y,z */
599 if (bCalcFshift && (tidxj & 3) < 3)
601 atomicAdd(&(atdat.fshift[nb_sci.shift].x) + (tidxj & 3), fshift_buf);
605 /* reduce the energies over warps and store into global memory */
606 reduce_energy_warp_shfl(E_lj, E_el, e_lj, e_el, tidx);
609 #endif /* FUNCTION_DECLARATION_ONLY */
612 #undef MIN_BLOCKS_PER_MP
613 #undef THREADS_PER_BLOCK
616 #undef EXCLUSION_FORCES