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
43 #include "types/simple.h"
44 #include "types/nbnxn_pairlist.h"
45 #include "types/nb_verlet.h"
46 #include "types/ishift.h"
47 #include "types/force_flags.h"
48 #include "../nbnxn_consts.h"
51 #include "thread_mpi/atomic.h"
54 #include "nbnxn_cuda_types.h"
55 #include "../../gmxlib/cuda_tools/cudautils.cuh"
56 #include "nbnxn_cuda.h"
57 #include "nbnxn_cuda_data_mgmt.h"
60 /*! Texture reference for nonbonded parameters; bound to cu_nbparam_t.nbfp*/
61 texture<float, 1, cudaReadModeElementType> tex_nbfp;
63 /*! Texture reference for Ewald coulomb force table; bound to cu_nbparam_t.coulomb_tab */
64 texture<float, 1, cudaReadModeElementType> tex_coulomb_tab;
66 /* Convenience defines */
67 #define NCL_PER_SUPERCL (NBNXN_GPU_NCLUSTER_PER_SUPERCLUSTER)
68 #define CL_SIZE (NBNXN_GPU_CLUSTER_SIZE)
70 /***** The kernels come here *****/
71 #include "nbnxn_cuda_kernel_utils.cuh"
73 /* Generate all combinations of kernels through multiple inclusion:
74 F, F + E, F + prune, F + E + prune. */
76 #include "nbnxn_cuda_kernels.cuh"
77 /** Force & energy **/
79 #include "nbnxn_cuda_kernels.cuh"
82 /*** Pair-list pruning kernels ***/
85 #include "nbnxn_cuda_kernels.cuh"
86 /** Force & energy **/
88 #include "nbnxn_cuda_kernels.cuh"
92 /*! Nonbonded kernel function pointer type */
93 typedef void (*nbnxn_cu_kfunc_ptr_t)(const cu_atomdata_t,
98 /*********************************/
100 /* XXX always/never run the energy/pruning kernels -- only for benchmarking purposes */
101 static bool always_ener = (getenv("GMX_GPU_ALWAYS_ENER") != NULL);
102 static bool never_ener = (getenv("GMX_GPU_NEVER_ENER") != NULL);
103 static bool always_prune = (getenv("GMX_GPU_ALWAYS_PRUNE") != NULL);
106 /* Bit-pattern used for polling-based GPU synchronization. It is used as a float
107 * and corresponds to having the exponent set to the maximum (127 -- single
108 * precision) and the mantissa to 0.
110 static unsigned int poll_wait_pattern = (0x7FU << 23);
112 /*! Returns the number of blocks to be used for the nonbonded GPU kernel. */
113 static inline int calc_nb_kernel_nblock(int nwork_units, cuda_dev_info_t *dinfo)
119 max_grid_x_size = dinfo->prop.maxGridSize[0];
121 /* do we exceed the grid x dimension limit? */
122 if (nwork_units > max_grid_x_size)
124 gmx_fatal(FARGS, "Watch out system too large to simulate!\n"
125 "The number of nonbonded work units (=number of super-clusters) exceeds the"
126 "maximum grid size in x dimension (%d > %d)!", nwork_units, max_grid_x_size);
133 /* Constant arrays listing all kernel function pointers and enabling selection
134 of a kernel in an elegant manner. */
136 static const int nEnergyKernelTypes = 2; /* 0 - no energy, 1 - energy */
137 static const int nPruneKernelTypes = 2; /* 0 - no prune, 1 - prune */
139 /* Default kernels */
140 static const nbnxn_cu_kfunc_ptr_t
141 nb_default_kfunc_ptr[eelCuNR][nEnergyKernelTypes][nPruneKernelTypes] =
143 { { k_nbnxn_ewald, k_nbnxn_ewald_prune },
144 { k_nbnxn_ewald_ener, k_nbnxn_ewald_ener_prune } },
145 { { k_nbnxn_ewald_twin, k_nbnxn_ewald_twin_prune },
146 { k_nbnxn_ewald_twin_ener, k_nbnxn_ewald_twin_ener_prune } },
147 { { k_nbnxn_rf, k_nbnxn_rf_prune },
148 { k_nbnxn_rf_ener, k_nbnxn_rf_ener_prune } },
149 { { k_nbnxn_ewald, k_nbnxn_ewald_prune },
150 { k_nbnxn_cutoff_ener, k_nbnxn_cutoff_ener_prune } },
154 static const nbnxn_cu_kfunc_ptr_t
155 nb_legacy_kfunc_ptr[eelCuNR][nEnergyKernelTypes][nPruneKernelTypes] =
157 { { k_nbnxn_ewald_legacy, k_nbnxn_ewald_prune_legacy },
158 { k_nbnxn_ewald_ener_legacy, k_nbnxn_ewald_ener_prune_legacy } },
159 { { k_nbnxn_ewald_twin_legacy, k_nbnxn_ewald_twin_prune_legacy },
160 { k_nbnxn_ewald_twin_ener_legacy, k_nbnxn_ewald_twin_ener_prune_legacy } },
161 { { k_nbnxn_rf_legacy, k_nbnxn_rf_prune_legacy },
162 { k_nbnxn_rf_ener_legacy, k_nbnxn_rf_ener_prune_legacy } },
163 { { k_nbnxn_ewald_legacy, k_nbnxn_ewald_prune_legacy },
164 { k_nbnxn_cutoff_ener_legacy, k_nbnxn_cutoff_ener_prune_legacy } },
167 /*! Return a pointer to the kernel version to be executed at the current step. */
168 static inline nbnxn_cu_kfunc_ptr_t select_nbnxn_kernel(int kver, int eeltype,
169 bool bDoEne, bool bDoPrune)
171 assert(kver < eNbnxnCuKNR);
172 assert(eeltype < eelCuNR);
174 if (NBNXN_KVER_LEGACY(kver))
176 return nb_legacy_kfunc_ptr[eeltype][bDoEne][bDoPrune];
180 return nb_default_kfunc_ptr[eeltype][bDoEne][bDoPrune];
184 /*! Calculates the amount of shared memory required for kernel version in use. */
185 static inline int calc_shmem_required(int kver)
189 /* size of shmem (force-buffers/xq/atom type preloading) */
190 if (NBNXN_KVER_LEGACY(kver))
192 /* i-atom x+q in shared memory */
193 shmem = NCL_PER_SUPERCL * CL_SIZE * sizeof(float4);
194 /* force reduction buffers in shared memory */
195 shmem += CL_SIZE * CL_SIZE * 3 * sizeof(float);
199 /* NOTE: with the default kernel on sm3.0 we need shmem only for pre-loading */
200 /* i-atom x+q in shared memory */
201 shmem = NCL_PER_SUPERCL * CL_SIZE * sizeof(float4);
203 /* i-atom types in shared memory */
204 shmem += NCL_PER_SUPERCL * CL_SIZE * sizeof(int);
206 #if __CUDA_ARCH__ < 300
207 /* force reduction buffers in shared memory */
208 shmem += CL_SIZE * CL_SIZE * 3 * sizeof(float);
215 /*! As we execute nonbonded workload in separate streams, before launching
216 the kernel we need to make sure that he following operations have completed:
217 - atomdata allocation and related H2D transfers (every nstlist step);
218 - pair list H2D transfer (every nstlist step);
219 - shift vector H2D transfer (every nstlist step);
220 - force (+shift force and energy) output clearing (every step).
222 These operations are issued in the local stream at the beginning of the step
223 and therefore always complete before the local kernel launch. The non-local
224 kernel is launched after the local on the same device/context, so this is
225 inherently scheduled after the operations in the local stream (including the
227 However, for the sake of having a future-proof implementation, we use the
228 misc_ops_done event to record the point in time when the above operations
229 are finished and synchronize with this event in the non-local stream.
231 void nbnxn_cuda_launch_kernel(nbnxn_cuda_ptr_t cu_nb,
232 const nbnxn_atomdata_t *nbatom,
237 int adat_begin, adat_len; /* local/nonlocal offset and length used for xq and f */
238 /* CUDA kernel launch-related stuff */
240 dim3 dim_block, dim_grid;
241 nbnxn_cu_kfunc_ptr_t nb_kernel = NULL; /* fn pointer to the nonbonded kernel */
243 cu_atomdata_t *adat = cu_nb->atdat;
244 cu_nbparam_t *nbp = cu_nb->nbparam;
245 cu_plist_t *plist = cu_nb->plist[iloc];
246 cu_timers_t *t = cu_nb->timers;
247 cudaStream_t stream = cu_nb->stream[iloc];
249 bool bCalcEner = flags & GMX_FORCE_VIRIAL;
250 bool bCalcFshift = flags & GMX_FORCE_VIRIAL;
251 bool bDoTime = cu_nb->bDoTime;
253 /* turn energy calculation always on/off (for debugging/testing only) */
254 bCalcEner = (bCalcEner || always_ener) && !never_ener;
256 /* don't launch the kernel if there is no work to do */
257 if (plist->nsci == 0)
262 /* calculate the atom data index range based on locality */
266 adat_len = adat->natoms_local;
270 adat_begin = adat->natoms_local;
271 adat_len = adat->natoms - adat->natoms_local;
274 /* When we get here all misc operations issues in the local stream are done,
275 so we record that in the local stream and wait for it in the nonlocal one. */
276 if (cu_nb->bUseTwoStreams)
278 if (iloc == eintLocal)
280 stat = cudaEventRecord(cu_nb->misc_ops_done, stream);
281 CU_RET_ERR(stat, "cudaEventRecord on misc_ops_done failed");
285 stat = cudaStreamWaitEvent(stream, cu_nb->misc_ops_done, 0);
286 CU_RET_ERR(stat, "cudaStreamWaitEvent on misc_ops_done failed");
290 /* beginning of timed HtoD section */
293 stat = cudaEventRecord(t->start_nb_h2d[iloc], stream);
294 CU_RET_ERR(stat, "cudaEventRecord failed");
298 cu_copy_H2D_async(adat->xq + adat_begin, nbatom->x + adat_begin * 4,
299 adat_len * sizeof(*adat->xq), stream);
303 stat = cudaEventRecord(t->stop_nb_h2d[iloc], stream);
304 CU_RET_ERR(stat, "cudaEventRecord failed");
307 /* beginning of timed nonbonded calculation section */
310 stat = cudaEventRecord(t->start_nb_k[iloc], stream);
311 CU_RET_ERR(stat, "cudaEventRecord failed");
314 /* get the pointer to the kernel flavor we need to use */
315 nb_kernel = select_nbnxn_kernel(cu_nb->kernel_ver, nbp->eeltype, bCalcEner,
316 plist->bDoPrune || always_prune);
318 /* kernel launch config */
319 nblock = calc_nb_kernel_nblock(plist->nsci, cu_nb->dev_info);
320 dim_block = dim3(CL_SIZE, CL_SIZE, 1);
321 dim_grid = dim3(nblock, 1, 1);
322 shmem = calc_shmem_required(cu_nb->kernel_ver);
326 fprintf(debug, "GPU launch configuration:\n\tThread block: %dx%dx%d\n\t"
327 "Grid: %dx%d\n\t#Super-clusters/clusters: %d/%d (%d)\n",
328 dim_block.x, dim_block.y, dim_block.z,
329 dim_grid.x, dim_grid.y, plist->nsci*NCL_PER_SUPERCL,
330 NCL_PER_SUPERCL, plist->na_c);
333 nb_kernel<<<dim_grid, dim_block, shmem, stream>>>(*adat, *nbp, *plist, bCalcFshift);
334 CU_LAUNCH_ERR("k_calc_nb");
338 stat = cudaEventRecord(t->stop_nb_k[iloc], stream);
339 CU_RET_ERR(stat, "cudaEventRecord failed");
343 void nbnxn_cuda_launch_cpyback(nbnxn_cuda_ptr_t cu_nb,
344 const nbnxn_atomdata_t *nbatom,
349 int adat_begin, adat_len, adat_end; /* local/nonlocal offset and length used for xq and f */
352 /* determine interaction locality from atom locality */
357 else if (NONLOCAL_A(aloc))
364 sprintf(stmp, "Invalid atom locality passed (%d); valid here is only "
365 "local (%d) or nonlocal (%d)", aloc, eatLocal, eatNonlocal);
369 cu_atomdata_t *adat = cu_nb->atdat;
370 cu_timers_t *t = cu_nb->timers;
371 bool bDoTime = cu_nb->bDoTime;
372 cudaStream_t stream = cu_nb->stream[iloc];
374 bool bCalcEner = flags & GMX_FORCE_VIRIAL;
375 bool bCalcFshift = flags & GMX_FORCE_VIRIAL;
377 /* don't launch copy-back if there was no work to do */
378 if (cu_nb->plist[iloc]->nsci == 0)
383 /* calculate the atom data index range based on locality */
387 adat_len = adat->natoms_local;
388 adat_end = cu_nb->atdat->natoms_local;
392 adat_begin = adat->natoms_local;
393 adat_len = adat->natoms - adat->natoms_local;
394 adat_end = cu_nb->atdat->natoms;
397 /* beginning of timed D2H section */
400 stat = cudaEventRecord(t->start_nb_d2h[iloc], stream);
401 CU_RET_ERR(stat, "cudaEventRecord failed");
404 if (!cu_nb->bUseStreamSync)
406 /* For safety reasons set a few (5%) forces to NaN. This way even if the
407 polling "hack" fails with some future NVIDIA driver we'll get a crash. */
408 for (int i = adat_begin; i < 3*adat_end + 2; i += adat_len/20)
411 nbatom->out[0].f[i] = NAN;
414 if (numeric_limits<float>::has_quiet_NaN)
416 nbatom->out[0].f[i] = numeric_limits<float>::quiet_NaN();
421 nbatom->out[0].f[i] = GMX_REAL_MAX;
426 /* Set the last four bytes of the force array to a bit pattern
427 which can't be the result of the force calculation:
428 max exponent (127) and zero mantissa. */
429 *(unsigned int*)&nbatom->out[0].f[adat_end*3 - 1] = poll_wait_pattern;
432 /* With DD the local D2H transfer can only start after the non-local
433 has been launched. */
434 if (iloc == eintLocal && cu_nb->bUseTwoStreams)
436 stat = cudaStreamWaitEvent(stream, cu_nb->nonlocal_done, 0);
437 CU_RET_ERR(stat, "cudaStreamWaitEvent on nonlocal_done failed");
441 cu_copy_D2H_async(nbatom->out[0].f + adat_begin * 3, adat->f + adat_begin,
442 (adat_len)*sizeof(*adat->f), stream);
444 /* After the non-local D2H is launched the nonlocal_done event can be
445 recorded which signals that the local D2H can proceed. This event is not
446 placed after the non-local kernel because we first need the non-local
448 if (iloc == eintNonlocal)
450 stat = cudaEventRecord(cu_nb->nonlocal_done, stream);
451 CU_RET_ERR(stat, "cudaEventRecord on nonlocal_done failed");
454 /* only transfer energies in the local stream */
460 cu_copy_D2H_async(cu_nb->nbst.fshift, adat->fshift,
461 SHIFTS * sizeof(*cu_nb->nbst.fshift), stream);
467 cu_copy_D2H_async(cu_nb->nbst.e_lj, adat->e_lj,
468 sizeof(*cu_nb->nbst.e_lj), stream);
469 cu_copy_D2H_async(cu_nb->nbst.e_el, adat->e_el,
470 sizeof(*cu_nb->nbst.e_el), stream);
476 stat = cudaEventRecord(t->stop_nb_d2h[iloc], stream);
477 CU_RET_ERR(stat, "cudaEventRecord failed");
481 /* Atomic compare-exchange operation on unsigned values. It is used in
482 * polling wait for the GPU.
484 static inline bool atomic_cas(volatile unsigned int *ptr,
491 return tMPI_Atomic_cas((tMPI_Atomic_t *)ptr, oldval, newval);
493 gmx_incons("Atomic operations not available, atomic_cas() should not have been called!");
498 void nbnxn_cuda_wait_gpu(nbnxn_cuda_ptr_t cu_nb,
499 const nbnxn_atomdata_t *nbatom,
501 float *e_lj, float *e_el, rvec *fshift)
504 int i, adat_end, iloc = -1;
505 volatile unsigned int *poll_word;
507 /* determine interaction locality from atom locality */
512 else if (NONLOCAL_A(aloc))
519 sprintf(stmp, "Invalid atom locality passed (%d); valid here is only "
520 "local (%d) or nonlocal (%d)", aloc, eatLocal, eatNonlocal);
524 cu_plist_t *plist = cu_nb->plist[iloc];
525 cu_timers_t *timers = cu_nb->timers;
526 wallclock_gpu_t *timings = cu_nb->timings;
527 nb_staging nbst = cu_nb->nbst;
529 bool bCalcEner = flags & GMX_FORCE_VIRIAL;
530 bool bCalcFshift = flags & GMX_FORCE_VIRIAL;
532 /* turn energy calculation always on/off (for debugging/testing only) */
533 bCalcEner = (bCalcEner || always_ener) && !never_ener;
535 /* don't launch wait/update timers & counters if there was no work to do
537 NOTE: if timing with multiple GPUs (streams) becomes possible, the
538 counters could end up being inconsistent due to not being incremented
539 on some of the nodes! */
540 if (cu_nb->plist[iloc]->nsci == 0)
545 /* calculate the atom data index range based on locality */
548 adat_end = cu_nb->atdat->natoms_local;
552 adat_end = cu_nb->atdat->natoms;
555 if (cu_nb->bUseStreamSync)
557 stat = cudaStreamSynchronize(cu_nb->stream[iloc]);
558 CU_RET_ERR(stat, "cudaStreamSynchronize failed in cu_blockwait_nb");
562 /* Busy-wait until we get the signal pattern set in last byte
563 * of the l/nl float vector. This pattern corresponds to a floating
564 * point number which can't be the result of the force calculation
565 * (maximum, 127 exponent and 0 mantissa).
566 * The polling uses atomic compare-exchange.
568 poll_word = (volatile unsigned int*)&nbatom->out[0].f[adat_end*3 - 1];
569 while (atomic_cas(poll_word, poll_wait_pattern, poll_wait_pattern)) {}
572 /* timing data accumulation */
575 /* only increase counter once (at local F wait) */
579 timings->ktime[plist->bDoPrune ? 1 : 0][bCalcEner ? 1 : 0].c += 1;
583 timings->ktime[plist->bDoPrune ? 1 : 0][bCalcEner ? 1 : 0].t +=
584 cu_event_elapsed(timers->start_nb_k[iloc], timers->stop_nb_k[iloc]);
586 /* X/q H2D and F D2H timings */
587 timings->nb_h2d_t += cu_event_elapsed(timers->start_nb_h2d[iloc],
588 timers->stop_nb_h2d[iloc]);
589 timings->nb_d2h_t += cu_event_elapsed(timers->start_nb_d2h[iloc],
590 timers->stop_nb_d2h[iloc]);
592 /* only count atdat and pair-list H2D at pair-search step */
595 /* atdat transfer timing (add only once, at local F wait) */
599 timings->pl_h2d_t += cu_event_elapsed(timers->start_atdat,
603 timings->pl_h2d_t += cu_event_elapsed(timers->start_pl_h2d[iloc],
604 timers->stop_pl_h2d[iloc]);
608 /* add up energies and shift forces (only once at local F wait) */
619 for (i = 0; i < SHIFTS; i++)
621 fshift[i][0] += nbst.fshift[i].x;
622 fshift[i][1] += nbst.fshift[i].y;
623 fshift[i][2] += nbst.fshift[i].z;
628 /* turn off pruning (doesn't matter if this is pair-search step or not) */
629 plist->bDoPrune = false;
632 /*! Return the reference to the nbfp texture. */
633 const struct texture<float, 1, cudaReadModeElementType>& nbnxn_cuda_get_nbfp_texref()
638 /*! Return the reference to the coulomb_tab. */
639 const struct texture<float, 1, cudaReadModeElementType>& nbnxn_cuda_get_coulomb_tab_texref()
641 return tex_coulomb_tab;
644 /*! Set up the cache configuration for the non-bonded kernels,
646 void nbnxn_cuda_set_cacheconfig(cuda_dev_info_t *devinfo)
650 for (int i = 0; i < eelCuNR; i++)
651 for (int j = 0; j < nEnergyKernelTypes; j++)
652 for (int k = 0; k < nPruneKernelTypes; k++)
654 /* Legacy kernel 16/48 kB Shared/L1 */
655 stat = cudaFuncSetCacheConfig(nb_legacy_kfunc_ptr[i][j][k], cudaFuncCachePreferL1);
656 CU_RET_ERR(stat, "cudaFuncSetCacheConfig failed");
658 if (devinfo->prop.major >= 3)
660 /* Default kernel on sm 3.x 48/16 kB Shared/L1 */
661 stat = cudaFuncSetCacheConfig(nb_default_kfunc_ptr[i][j][k], cudaFuncCachePreferShared);
665 /* On Fermi prefer L1 gives 2% higher performance */
666 /* Default kernel on sm_2.x 16/48 kB Shared/L1 */
667 stat = cudaFuncSetCacheConfig(nb_default_kfunc_ptr[i][j][k], cudaFuncCachePreferL1);
669 CU_RET_ERR(stat, "cudaFuncSetCacheConfig failed");