2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016,2017,2018,2019, 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_nbnxm
48 #include "gromacs/gpu_utils/cuda_arch_utils.cuh"
49 #include "gromacs/gpu_utils/cuda_kernel_utils.cuh"
50 #include "gromacs/math/utilities.h"
51 #include "gromacs/pbcutil/ishift.h"
52 /* Note that floating-point constants in CUDA code should be suffixed
53 * with f (e.g. 0.5f), to stop the compiler producing intermediate
54 * code that is in double precision.
57 #if defined EL_EWALD_ANA || defined EL_EWALD_TAB
58 /* Note: convenience macro, needs to be undef-ed at the end of the file. */
62 #if defined EL_EWALD_ANY || defined EL_RF || defined LJ_EWALD || (defined EL_CUTOFF && defined CALC_ENERGIES)
63 /* Macro to control the calculation of exclusion forces in the kernel
64 * We do that with Ewald (elec/vdw) and RF. Cut-off only has exclusion
67 * Note: convenience macro, needs to be undef-ed at the end of the file.
69 #define EXCLUSION_FORCES
72 #if defined LJ_EWALD_COMB_GEOM || defined LJ_EWALD_COMB_LB
73 /* Note: convenience macro, needs to be undef-ed at the end of the file. */
77 #if defined LJ_COMB_GEOM || defined LJ_COMB_LB
82 Kernel launch parameters:
83 - #blocks = #pair lists, blockId = pair list Id
84 - #threads = NTHREAD_Z * c_clSize^2
85 - shmem = see nbnxn_cuda.cu:calc_shmem_required_nonbonded()
87 Each thread calculates an i force-component taking one pair of i-j atoms.
91 /*! \brief Compute capability dependent definition of kernel launch configuration parameters.
93 * NTHREAD_Z controls the number of j-clusters processed concurrently on NTHREAD_Z
94 * warp-pairs per block.
96 * - On CC 3.0-3.5, and >=5.0 NTHREAD_Z == 1, translating to 64 th/block with 16
97 * blocks/multiproc, is the fastest even though this setup gives low occupancy
99 * NTHREAD_Z > 1 results in excessive register spilling unless the minimum blocks
100 * per multiprocessor is reduced proportionally to get the original number of max
101 * threads in flight (and slightly lower performance).
102 * - On CC 3.7 there are enough registers to double the number of threads; using
103 * NTHREADS_Z == 2 is fastest with 16 blocks (TODO: test with RF and other kernels
104 * with low-register use).
106 * Note that the current kernel implementation only supports NTHREAD_Z > 1 with
107 * shuffle-based reduction, hence CC >= 3.0.
110 * NOTEs on Volta / CUDA 9 extensions:
112 * - While active thread masks are required for the warp collectives
113 * (we use any and shfl), the kernel is designed such that all conditions
114 * (other than the inner-most distance check) including loop trip counts
115 * are warp-synchronous. Therefore, we don't need ballot to compute the
116 * active masks as these are all full-warp masks.
118 * - TODO: reconsider the use of __syncwarp(): its only role is currently to prevent
119 * WAR hazard due to the cj preload; we should try to replace it with direct
120 * loads (which may be faster given the improved L1 on Volta).
123 /* Kernel launch bounds for different compute capabilities. The value of NTHREAD_Z
124 * determines the number of threads per block and it is chosen such that
125 * 16 blocks/multiprocessor can be kept in flight.
126 * - CC 3.0,3.5, and >=5.0: NTHREAD_Z=1, (64, 16) bounds
127 * - CC 3.7: NTHREAD_Z=2, (128, 16) bounds
129 * Note: convenience macros, need to be undef-ed at the end of the file.
131 #if GMX_PTX_ARCH == 370
132 #define NTHREAD_Z (2)
133 #define MIN_BLOCKS_PER_MP (16)
135 #define NTHREAD_Z (1)
136 #define MIN_BLOCKS_PER_MP (16)
137 #endif /* GMX_PTX_ARCH == 370 */
138 #define THREADS_PER_BLOCK (c_clSize*c_clSize*NTHREAD_Z)
140 #if GMX_PTX_ARCH >= 350
142 __launch_bounds__(THREADS_PER_BLOCK, MIN_BLOCKS_PER_MP)
144 __launch_bounds__(THREADS_PER_BLOCK)
145 #endif /* GMX_PTX_ARCH >= 350 */
148 __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _VF_prune_cuda)
150 __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _F_prune_cuda)
151 #endif /* CALC_ENERGIES */
154 __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _VF_cuda)
156 __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _F_cuda)
157 #endif /* CALC_ENERGIES */
158 #endif /* PRUNE_NBL */
159 (const cu_atomdata_t atdat,
160 const cu_nbparam_t nbparam,
161 const cu_plist_t plist,
163 #ifdef FUNCTION_DECLARATION_ONLY
164 ; /* Only do function declaration, omit the function body. */
167 /* convenience variables */
168 const nbnxn_sci_t *pl_sci = plist.sci;
172 nbnxn_cj4_t *pl_cj4 = plist.cj4;
173 const nbnxn_excl_t *excl = plist.excl;
175 const int *atom_types = atdat.atom_types;
176 int ntypes = atdat.ntypes;
178 const float2 *lj_comb = atdat.lj_comb;
179 float2 ljcp_i, ljcp_j;
181 const float4 *xq = atdat.xq;
183 const float3 *shift_vec = atdat.shift_vec;
184 float rcoulomb_sq = nbparam.rcoulomb_sq;
185 #ifdef VDW_CUTOFF_CHECK
186 float rvdw_sq = nbparam.rvdw_sq;
190 float lje_coeff2, lje_coeff6_6;
193 float two_k_rf = nbparam.two_k_rf;
196 float beta2 = nbparam.ewald_beta*nbparam.ewald_beta;
197 float beta3 = nbparam.ewald_beta*nbparam.ewald_beta*nbparam.ewald_beta;
200 float rlist_sq = nbparam.rlistOuter_sq;
205 float beta = nbparam.ewald_beta;
206 float ewald_shift = nbparam.sh_ewald;
208 float c_rf = nbparam.c_rf;
209 #endif /* EL_EWALD_ANY */
210 float *e_lj = atdat.e_lj;
211 float *e_el = atdat.e_el;
212 #endif /* CALC_ENERGIES */
214 /* thread/block/warp id-s */
215 unsigned int tidxi = threadIdx.x;
216 unsigned int tidxj = threadIdx.y;
217 unsigned int tidx = threadIdx.y * blockDim.x + threadIdx.x;
219 unsigned int tidxz = 0;
221 unsigned int tidxz = threadIdx.z;
223 unsigned int bidx = blockIdx.x;
224 unsigned int widx = tidx / warp_size; /* warp index */
228 cij4_start, cij4_end;
232 int i, jm, j4, wexcl_idx;
235 #if !defined LJ_COMB_LB || defined CALC_ENERGIES
236 float inv_r6, c6, c12;
239 float sigma, epsilon;
246 #if defined CALC_ENERGIES || defined LJ_POT_SWITCH
249 unsigned int wexcl, imask, mask_ji;
251 float3 xi, xj, rv, f_ij, fcj_buf;
252 float3 fci_buf[c_numClPerSupercl]; /* i force buffer */
255 /*! i-cluster interaction mask for a super-cluster with all c_numClPerSupercl=8 bits set */
256 const unsigned superClInteractionMask = ((1U << c_numClPerSupercl) - 1U);
258 /*********************************************************************
259 * Set up shared memory pointers.
260 * sm_nextSlotPtr should always be updated to point to the "next slot",
261 * that is past the last point where data has been stored.
263 extern __shared__ char sm_dynamicShmem[];
264 char *sm_nextSlotPtr = sm_dynamicShmem;
265 static_assert(sizeof(char) == 1, "The shared memory offset calculation assumes that char is 1 byte");
267 /* shmem buffer for i x+q pre-loading */
268 float4 *xqib = (float4 *)sm_nextSlotPtr;
269 sm_nextSlotPtr += (c_numClPerSupercl * c_clSize * sizeof(*xqib));
271 /* shmem buffer for cj, for each warp separately */
272 int *cjs = (int *)(sm_nextSlotPtr);
273 /* the cjs buffer's use expects a base pointer offset for pairs of warps in the j-concurrent execution */
274 cjs += tidxz * c_nbnxnGpuClusterpairSplit * c_nbnxnGpuJgroupSize;
275 sm_nextSlotPtr += (NTHREAD_Z * c_nbnxnGpuClusterpairSplit * c_nbnxnGpuJgroupSize * sizeof(*cjs));
278 /* shmem buffer for i atom-type pre-loading */
279 int *atib = (int *)sm_nextSlotPtr;
280 sm_nextSlotPtr += (c_numClPerSupercl * c_clSize * sizeof(*atib));
282 /* shmem buffer for i-atom LJ combination rule parameters */
283 float2 *ljcpib = (float2 *)sm_nextSlotPtr;
284 sm_nextSlotPtr += (c_numClPerSupercl * c_clSize * sizeof(*ljcpib));
286 /*********************************************************************/
288 nb_sci = pl_sci[bidx]; /* my i super-cluster's index = current bidx */
289 sci = nb_sci.sci; /* super-cluster */
290 cij4_start = nb_sci.cj4_ind_start; /* first ...*/
291 cij4_end = nb_sci.cj4_ind_end; /* and last index of j clusters */
295 /* Pre-load i-atom x and q into shared memory */
296 ci = sci * c_numClPerSupercl + tidxj;
297 ai = ci * c_clSize + tidxi;
299 float *shiftptr = (float *)&shift_vec[nb_sci.shift];
300 xqbuf = xq[ai] + make_float4(LDG(shiftptr), LDG(shiftptr + 1), LDG(shiftptr + 2), 0.0f);
301 xqbuf.w *= nbparam.epsfac;
302 xqib[tidxj * c_clSize + tidxi] = xqbuf;
305 /* Pre-load the i-atom types into shared memory */
306 atib[tidxj * c_clSize + tidxi] = atom_types[ai];
308 /* Pre-load the LJ combination parameters into shared memory */
309 ljcpib[tidxj * c_clSize + tidxi] = lj_comb[ai];
314 for (i = 0; i < c_numClPerSupercl; i++)
316 fci_buf[i] = make_float3(0.0f);
320 /* TODO: we are trading registers with flops by keeping lje_coeff-s, try re-calculating it later */
321 lje_coeff2 = nbparam.ewaldcoeff_lj*nbparam.ewaldcoeff_lj;
322 lje_coeff6_6 = lje_coeff2*lje_coeff2*lje_coeff2*c_oneSixth;
330 #ifdef EXCLUSION_FORCES /* Ewald or RF */
331 if (nb_sci.shift == CENTRAL && pl_cj4[cij4_start].cj[0] == sci*c_numClPerSupercl)
333 /* we have the diagonal: add the charge and LJ self interaction energy term */
334 for (i = 0; i < c_numClPerSupercl; i++)
336 #if defined EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF
337 qi = xqib[i * c_clSize + tidxi].w;
342 #if DISABLE_CUDA_TEXTURES
343 E_lj += LDG(&nbparam.nbfp[atom_types[(sci*c_numClPerSupercl + i)*c_clSize + tidxi]*(ntypes + 1)*2]);
345 E_lj += tex1Dfetch<float>(nbparam.nbfp_texobj, atom_types[(sci*c_numClPerSupercl + i)*c_clSize + tidxi]*(ntypes + 1)*2);
350 /* divide the self term(s) equally over the j-threads, then multiply with the coefficients. */
352 E_lj /= c_clSize*NTHREAD_Z;
353 E_lj *= 0.5f*c_oneSixth*lje_coeff6_6;
356 #if defined EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF
357 /* Correct for epsfac^2 due to adding qi^2 */
358 E_el /= nbparam.epsfac*c_clSize*NTHREAD_Z;
359 #if defined EL_RF || defined EL_CUTOFF
362 E_el *= -beta*M_FLOAT_1_SQRTPI; /* last factor 1/sqrt(pi) */
364 #endif /* EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF */
366 #endif /* EXCLUSION_FORCES */
368 #endif /* CALC_ENERGIES */
370 #ifdef EXCLUSION_FORCES
371 const int nonSelfInteraction = !(nb_sci.shift == CENTRAL & tidxj <= tidxi);
374 /* loop over the j clusters = seen by any of the atoms in the current super-cluster;
375 * The loop stride NTHREAD_Z ensures that consecutive warps-pairs are assigned
376 * consecutive j4's entries.
378 for (j4 = cij4_start + tidxz; j4 < cij4_end; j4 += NTHREAD_Z)
380 wexcl_idx = pl_cj4[j4].imei[widx].excl_ind;
381 imask = pl_cj4[j4].imei[widx].imask;
382 wexcl = excl[wexcl_idx].pair[(tidx) & (warp_size - 1)];
388 /* Pre-load cj into shared memory on both warps separately */
389 if ((tidxj == 0 | tidxj == 4) & (tidxi < c_nbnxnGpuJgroupSize))
391 cjs[tidxi + tidxj * c_nbnxnGpuJgroupSize/c_splitClSize] = pl_cj4[j4].cj[tidxi];
393 gmx_syncwarp(c_fullWarpMask);
395 /* Unrolling this loop
396 - with pruning leads to register spilling;
397 - on Kepler and later it is much slower;
398 Tested with up to nvcc 7.5 */
399 for (jm = 0; jm < c_nbnxnGpuJgroupSize; jm++)
401 if (imask & (superClInteractionMask << (jm * c_numClPerSupercl)))
403 mask_ji = (1U << (jm * c_numClPerSupercl));
405 cj = cjs[jm + (tidxj & 4) * c_nbnxnGpuJgroupSize/c_splitClSize];
406 aj = cj * c_clSize + tidxj;
408 /* load j atom data */
410 xj = make_float3(xqbuf.x, xqbuf.y, xqbuf.z);
413 typej = atom_types[aj];
415 ljcp_j = lj_comb[aj];
418 fcj_buf = make_float3(0.0f);
420 #if !defined PRUNE_NBL
423 for (i = 0; i < c_numClPerSupercl; i++)
427 ci = sci * c_numClPerSupercl + i; /* i cluster index */
429 /* all threads load an atom from i cluster ci into shmem! */
430 xqbuf = xqib[i * c_clSize + tidxi];
431 xi = make_float3(xqbuf.x, xqbuf.y, xqbuf.z);
433 /* distance between i and j atoms */
438 /* If _none_ of the atoms pairs are in cutoff range,
439 the bit corresponding to the current
440 cluster-pair in imask gets set to 0. */
441 if (!gmx_any_sync(c_fullWarpMask, r2 < rlist_sq))
447 int_bit = (wexcl & mask_ji) ? 1.0f : 0.0f;
449 /* cutoff & exclusion check */
450 #ifdef EXCLUSION_FORCES
451 if ((r2 < rcoulomb_sq) * (nonSelfInteraction | (ci != cj)))
453 if ((r2 < rcoulomb_sq) * int_bit)
456 /* load the rest of the i-atom parameters */
460 /* LJ 6*C6 and 12*C12 */
461 typei = atib[i * c_clSize + tidxi];
462 fetch_nbfp_c6_c12(c6, c12, nbparam, ntypes * typei + typej);
464 ljcp_i = ljcpib[i * c_clSize + tidxi];
466 c6 = ljcp_i.x * ljcp_j.x;
467 c12 = ljcp_i.y * ljcp_j.y;
469 /* LJ 2^(1/6)*sigma and 12*epsilon */
470 sigma = ljcp_i.x + ljcp_j.x;
471 epsilon = ljcp_i.y * ljcp_j.y;
472 #if defined CALC_ENERGIES || defined LJ_FORCE_SWITCH || defined LJ_POT_SWITCH
473 convert_sigma_epsilon_to_c6_c12(sigma, epsilon, &c6, &c12);
475 #endif /* LJ_COMB_GEOM */
478 // Ensure distance do not become so small that r^-12 overflows
479 r2 = max(r2, NBNXN_MIN_RSQ);
482 inv_r2 = inv_r * inv_r;
483 #if !defined LJ_COMB_LB || defined CALC_ENERGIES
484 inv_r6 = inv_r2 * inv_r2 * inv_r2;
485 #ifdef EXCLUSION_FORCES
486 /* We could mask inv_r2, but with Ewald
487 * masking both inv_r6 and F_invr is faster */
489 #endif /* EXCLUSION_FORCES */
491 F_invr = inv_r6 * (c12 * inv_r6 - c6) * inv_r2;
492 #if defined CALC_ENERGIES || defined LJ_POT_SWITCH
493 E_lj_p = int_bit * (c12 * (inv_r6 * inv_r6 + nbparam.repulsion_shift.cpot)*c_oneTwelveth -
494 c6 * (inv_r6 + nbparam.dispersion_shift.cpot)*c_oneSixth);
496 #else /* !LJ_COMB_LB || CALC_ENERGIES */
497 float sig_r = sigma*inv_r;
498 float sig_r2 = sig_r*sig_r;
499 float sig_r6 = sig_r2*sig_r2*sig_r2;
500 #ifdef EXCLUSION_FORCES
502 #endif /* EXCLUSION_FORCES */
504 F_invr = epsilon * sig_r6 * (sig_r6 - 1.0f) * inv_r2;
505 #endif /* !LJ_COMB_LB || CALC_ENERGIES */
507 #ifdef LJ_FORCE_SWITCH
509 calculate_force_switch_F_E(nbparam, c6, c12, inv_r, r2, &F_invr, &E_lj_p);
511 calculate_force_switch_F(nbparam, c6, c12, inv_r, r2, &F_invr);
512 #endif /* CALC_ENERGIES */
513 #endif /* LJ_FORCE_SWITCH */
517 #ifdef LJ_EWALD_COMB_GEOM
519 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);
521 calculate_lj_ewald_comb_geom_F(nbparam, typei, typej, r2, inv_r2, lje_coeff2, lje_coeff6_6, &F_invr);
522 #endif /* CALC_ENERGIES */
523 #elif defined LJ_EWALD_COMB_LB
524 calculate_lj_ewald_comb_LB_F_E(nbparam, typei, typej, r2, inv_r2, lje_coeff2, lje_coeff6_6,
526 int_bit, &F_invr, &E_lj_p
529 #endif /* CALC_ENERGIES */
531 #endif /* LJ_EWALD_COMB_GEOM */
532 #endif /* LJ_EWALD */
536 calculate_potential_switch_F_E(nbparam, inv_r, r2, &F_invr, &E_lj_p);
538 calculate_potential_switch_F(nbparam, inv_r, r2, &F_invr, &E_lj_p);
539 #endif /* CALC_ENERGIES */
540 #endif /* LJ_POT_SWITCH */
542 #ifdef VDW_CUTOFF_CHECK
543 /* Separate VDW cut-off check to enable twin-range cut-offs
544 * (rvdw < rcoulomb <= rlist)
546 vdw_in_range = (r2 < rvdw_sq) ? 1.0f : 0.0f;
547 F_invr *= vdw_in_range;
549 E_lj_p *= vdw_in_range;
551 #endif /* VDW_CUTOFF_CHECK */
559 #ifdef EXCLUSION_FORCES
560 F_invr += qi * qj_f * int_bit * inv_r2 * inv_r;
562 F_invr += qi * qj_f * inv_r2 * inv_r;
566 F_invr += qi * qj_f * (int_bit*inv_r2 * inv_r - two_k_rf);
568 #if defined EL_EWALD_ANA
569 F_invr += qi * qj_f * (int_bit*inv_r2*inv_r + pmecorrF(beta2*r2)*beta3);
570 #elif defined EL_EWALD_TAB
571 F_invr += qi * qj_f * (int_bit*inv_r2 -
572 interpolate_coulomb_force_r(nbparam, r2 * inv_r)) * inv_r;
573 #endif /* EL_EWALD_ANA/TAB */
577 E_el += qi * qj_f * (int_bit*inv_r - c_rf);
580 E_el += qi * qj_f * (int_bit*inv_r + 0.5f * two_k_rf * r2 - c_rf);
583 /* 1.0f - erff is faster than erfcf */
584 E_el += qi * qj_f * (inv_r * (int_bit - erff(r2 * inv_r * beta)) - int_bit * ewald_shift);
585 #endif /* EL_EWALD_ANY */
589 /* accumulate j forces in registers */
592 /* accumulate i forces in registers */
597 /* shift the mask bit by 1 */
601 /* reduce j forces */
602 reduce_force_j_warp_shfl(fcj_buf, f, tidxi, aj, c_fullWarpMask);
606 /* Update the imask with the new one which does not contain the
607 out of range clusters anymore. */
608 pl_cj4[j4].imei[widx].imask = imask;
611 // avoid shared memory WAR hazards between loop iterations
612 gmx_syncwarp(c_fullWarpMask);
615 /* skip central shifts when summing shift forces */
616 if (nb_sci.shift == CENTRAL)
621 float fshift_buf = 0.0f;
623 /* reduce i forces */
624 for (i = 0; i < c_numClPerSupercl; i++)
626 ai = (sci * c_numClPerSupercl + i) * c_clSize + tidxi;
627 reduce_force_i_warp_shfl(fci_buf[i], f,
628 &fshift_buf, bCalcFshift,
629 tidxj, ai, c_fullWarpMask);
632 /* add up local shift forces into global mem, tidxj indexes x,y,z */
633 if (bCalcFshift && (tidxj & 3) < 3)
635 atomicAdd(&(atdat.fshift[nb_sci.shift].x) + (tidxj & 3), fshift_buf);
639 /* reduce the energies over warps and store into global memory */
640 reduce_energy_warp_shfl(E_lj, E_el, e_lj, e_el, tidx, c_fullWarpMask);
643 #endif /* FUNCTION_DECLARATION_ONLY */
646 #undef MIN_BLOCKS_PER_MP
647 #undef THREADS_PER_BLOCK
650 #undef EXCLUSION_FORCES