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.
36 * \brief Define OpenCL implementation of nbnxn_gpu_data_mgmt.h
38 * \author Anca Hamuraru <anca@streamcomputing.eu>
39 * \author Dimitrios Karkoulis <dimitris.karkoulis@gmail.com>
40 * \author Teemu Virolainen <teemu@streamcomputing.eu>
41 * \author Szilárd Páll <pall.szilard@gmail.com>
52 #include "gromacs/gpu_utils/gpu_utils.h"
53 #include "gromacs/gpu_utils/oclutils.h"
54 #include "gromacs/hardware/detecthardware.h"
55 #include "gromacs/hardware/gpu_hw_info.h"
56 #include "gromacs/math/vectypes.h"
57 #include "gromacs/mdlib/force_flags.h"
58 #include "gromacs/mdlib/nb_verlet.h"
59 #include "gromacs/mdlib/nbnxn_consts.h"
60 #include "gromacs/mdlib/nbnxn_gpu.h"
61 #include "gromacs/mdlib/nbnxn_gpu_data_mgmt.h"
62 #include "gromacs/mdlib/nbnxn_gpu_jit_support.h"
63 #include "gromacs/mdtypes/interaction_const.h"
64 #include "gromacs/mdtypes/md_enums.h"
65 #include "gromacs/pbcutil/ishift.h"
66 #include "gromacs/timing/gpu_timing.h"
67 #include "gromacs/utility/cstringutil.h"
68 #include "gromacs/utility/fatalerror.h"
69 #include "gromacs/utility/gmxassert.h"
70 #include "gromacs/utility/real.h"
71 #include "gromacs/utility/smalloc.h"
73 #include "nbnxn_ocl_internal.h"
74 #include "nbnxn_ocl_types.h"
76 /*! \brief This parameter should be determined heuristically from the
77 * kernel execution times
79 * This value is best for small systems on a single AMD Radeon R9 290X
80 * (and about 5% faster than 40, which is the default for CUDA
81 * devices). Larger simulation systems were quite insensitive to the
82 * value of this parameter.
84 static unsigned int gpu_min_ci_balanced_factor = 50;
87 /*! \brief Returns true if LJ combination rules are used in the non-bonded kernels.
89 * Full doc in nbnxn_ocl_internal.h */
90 bool useLjCombRule(int vdwType)
92 return (vdwType == evdwOclCUTCOMBGEOM ||
93 vdwType == evdwOclCUTCOMBLB);
96 /*! \brief Free device buffers
98 * If the pointers to the size variables are NULL no resetting happens.
100 void ocl_free_buffered(cl_mem d_ptr, int *n, int *nalloc)
102 cl_int gmx_unused cl_error;
106 cl_error = clReleaseMemObject(d_ptr);
107 assert(cl_error == CL_SUCCESS);
108 // TODO: handle errors
122 /*! \brief Reallocation device buffers
124 * Reallocation of the memory pointed by d_ptr and copying of the data from
125 * the location pointed by h_src host-side pointer is done. Allocation is
126 * buffered and therefore freeing is only needed if the previously allocated
127 * space is not enough.
128 * The H2D copy is launched in command queue s and can be done synchronously or
129 * asynchronously (the default is the latter).
130 * If copy_event is not NULL, on return it will contain an event object
131 * identifying the H2D copy. The event can further be used to queue a wait
132 * for this operation or to query profiling information.
133 * OpenCL equivalent of cu_realloc_buffered.
135 void ocl_realloc_buffered(cl_mem *d_dest, void *h_src,
137 int *curr_size, int *curr_alloc_size,
142 cl_event *copy_event = NULL)
144 if (d_dest == NULL || req_size < 0)
149 /* reallocate only if the data does not fit = allocation size is smaller
150 than the current requested size */
151 if (req_size > *curr_alloc_size)
153 cl_int gmx_unused cl_error;
155 /* only free if the array has already been initialized */
156 if (*curr_alloc_size >= 0)
158 ocl_free_buffered(*d_dest, curr_size, curr_alloc_size);
161 *curr_alloc_size = over_alloc_large(req_size);
163 *d_dest = clCreateBuffer(context, CL_MEM_READ_WRITE, *curr_alloc_size * type_size, NULL, &cl_error);
164 assert(cl_error == CL_SUCCESS);
165 // TODO: handle errors, check clCreateBuffer flags
168 /* size could have changed without actual reallocation */
169 *curr_size = req_size;
171 /* upload to device */
176 ocl_copy_H2D_async(*d_dest, h_src, 0, *curr_size * type_size, s, copy_event);
180 ocl_copy_H2D(*d_dest, h_src, 0, *curr_size * type_size, s);
185 /*! \brief Releases the input OpenCL buffer */
186 static void free_ocl_buffer(cl_mem *buffer)
188 cl_int gmx_unused cl_error;
190 assert(NULL != buffer);
194 cl_error = clReleaseMemObject(*buffer);
195 assert(CL_SUCCESS == cl_error);
200 /*! \brief Tabulates the Ewald Coulomb force and initializes the size/scale
201 * and the table GPU array.
203 * If called with an already allocated table, it just re-uploads the
206 static void init_ewald_coulomb_force_table(const interaction_const_t *ic,
208 const gmx_device_runtime_data_t *runData)
214 if (nbp->coulomb_tab_climg2d != NULL)
216 free_ocl_buffer(&(nbp->coulomb_tab_climg2d));
219 /* Switched from using textures to using buffers */
220 // TODO: decide which alternative is most efficient - textures or buffers.
222 cl_image_format array_format;
224 array_format.image_channel_data_type = CL_FLOAT;
225 array_format.image_channel_order = CL_R;
227 coul_tab = clCreateImage2D(runData->context, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
228 &array_format, tabsize, 1, 0, ftmp, &cl_error);
231 coul_tab = clCreateBuffer(runData->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, ic->tabq_size*sizeof(cl_float), ic->tabq_coul_F, &cl_error);
232 assert(cl_error == CL_SUCCESS);
233 // TODO: handle errors, check clCreateBuffer flags
235 nbp->coulomb_tab_climg2d = coul_tab;
236 nbp->coulomb_tab_size = ic->tabq_size;
237 nbp->coulomb_tab_scale = ic->tabq_scale;
241 /*! \brief Initializes the atomdata structure first time, it only gets filled at
244 static void init_atomdata_first(cl_atomdata_t *ad, int ntypes, gmx_device_runtime_data_t *runData)
250 /* An element of the shift_vec device buffer has the same size as one element
251 of the host side shift_vec buffer. */
252 ad->shift_vec_elem_size = sizeof(*(((nbnxn_atomdata_t*)0)->shift_vec));
254 // TODO: handle errors, check clCreateBuffer flags
255 ad->shift_vec = clCreateBuffer(runData->context, CL_MEM_READ_WRITE, SHIFTS * ad->shift_vec_elem_size, NULL, &cl_error);
256 assert(cl_error == CL_SUCCESS);
257 ad->bShiftVecUploaded = false;
259 /* An element of the fshift device buffer has the same size as one element
260 of the host side fshift buffer. */
261 ad->fshift_elem_size = sizeof(*(((cl_nb_staging_t*)0)->fshift));
263 ad->fshift = clCreateBuffer(runData->context, CL_MEM_READ_WRITE, SHIFTS * ad->fshift_elem_size, NULL, &cl_error);
264 assert(cl_error == CL_SUCCESS);
265 // TODO: handle errors, check clCreateBuffer flags
267 ad->e_lj = clCreateBuffer(runData->context, CL_MEM_READ_WRITE, sizeof(float), NULL, &cl_error);
268 assert(cl_error == CL_SUCCESS);
269 // TODO: handle errors, check clCreateBuffer flags
271 ad->e_el = clCreateBuffer(runData->context, CL_MEM_READ_WRITE, sizeof(float), NULL, &cl_error);
272 assert(cl_error == CL_SUCCESS);
273 // TODO: handle errors, check clCreateBuffer flags
275 /* initialize to NULL pointers to data that is not allocated here and will
276 need reallocation in nbnxn_gpu_init_atomdata */
280 /* size -1 indicates that the respective array hasn't been initialized yet */
285 /*! \brief Copies all parameters related to the cut-off from ic to nbp
287 static void set_cutoff_parameters(cl_nbparam_t *nbp,
288 const interaction_const_t *ic)
290 nbp->ewald_beta = ic->ewaldcoeff_q;
291 nbp->sh_ewald = ic->sh_ewald;
292 nbp->epsfac = ic->epsfac;
293 nbp->two_k_rf = 2.0 * ic->k_rf;
294 nbp->c_rf = ic->c_rf;
295 nbp->rvdw_sq = ic->rvdw * ic->rvdw;
296 nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
297 nbp->rlist_sq = ic->rlist * ic->rlist;
299 nbp->sh_lj_ewald = ic->sh_lj_ewald;
300 nbp->ewaldcoeff_lj = ic->ewaldcoeff_lj;
302 nbp->rvdw_switch = ic->rvdw_switch;
303 nbp->dispersion_shift = ic->dispersion_shift;
304 nbp->repulsion_shift = ic->repulsion_shift;
305 nbp->vdw_switch = ic->vdw_switch;
308 /*! \brief Returns the kinds of electrostatics and Vdw OpenCL
309 * kernels that will be used.
311 * Respectively, these values are from enum eelOcl and enum
314 map_interaction_types_to_gpu_kernel_flavors(const interaction_const_t *ic,
319 if (ic->vdwtype == evdwCUT)
321 switch (ic->vdw_modifier)
324 case eintmodPOTSHIFT:
328 *gpu_vdwtype = evdwOclCUT;
331 *gpu_vdwtype = evdwOclCUTCOMBGEOM;
334 *gpu_vdwtype = evdwOclCUTCOMBLB;
337 gmx_incons("The requested LJ combination rule is not implemented in the OpenCL GPU accelerated kernels!");
341 case eintmodFORCESWITCH:
342 *gpu_vdwtype = evdwOclFSWITCH;
344 case eintmodPOTSWITCH:
345 *gpu_vdwtype = evdwOclPSWITCH;
348 gmx_incons("The requested VdW interaction modifier is not implemented in the GPU accelerated kernels!");
352 else if (ic->vdwtype == evdwPME)
354 if (ic->ljpme_comb_rule == ljcrGEOM)
356 *gpu_vdwtype = evdwOclEWALDGEOM;
360 *gpu_vdwtype = evdwOclEWALDLB;
365 gmx_incons("The requested VdW type is not implemented in the GPU accelerated kernels!");
368 if (ic->eeltype == eelCUT)
370 *gpu_eeltype = eelOclCUT;
372 else if (EEL_RF(ic->eeltype))
374 *gpu_eeltype = eelOclRF;
376 else if ((EEL_PME(ic->eeltype) || ic->eeltype == eelEWALD))
378 /* Initially rcoulomb == rvdw, so it's surely not twin cut-off. */
379 *gpu_eeltype = nbnxn_gpu_pick_ewald_kernel_type(false);
383 /* Shouldn't happen, as this is checked when choosing Verlet-scheme */
384 gmx_incons("The requested electrostatics type is not implemented in the GPU accelerated kernels!");
388 /*! \brief Initializes the nonbonded parameter data structure.
390 static void init_nbparam(cl_nbparam_t *nbp,
391 const interaction_const_t *ic,
392 const nbnxn_atomdata_t *nbat,
393 const gmx_device_runtime_data_t *runData)
395 int ntypes, nnbfp, nnbfp_comb;
399 ntypes = nbat->ntype;
401 set_cutoff_parameters(nbp, ic);
403 map_interaction_types_to_gpu_kernel_flavors(ic,
408 if (ic->vdwtype == evdwPME)
410 if (ic->ljpme_comb_rule == ljcrGEOM)
412 assert(nbat->comb_rule == ljcrGEOM);
416 assert(nbat->comb_rule == ljcrLB);
419 /* generate table for PME */
420 nbp->coulomb_tab_climg2d = NULL;
421 if (nbp->eeltype == eelOclEWALD_TAB || nbp->eeltype == eelOclEWALD_TAB_TWIN)
423 init_ewald_coulomb_force_table(ic, nbp, runData);
426 // TODO: improvement needed.
427 // The image2d is created here even if eeltype is not eelCuEWALD_TAB or eelCuEWALD_TAB_TWIN because the OpenCL kernels
428 // don't accept NULL values for image2D parameters.
430 /* Switched from using textures to using buffers */
431 // TODO: decide which alternative is most efficient - textures or buffers.
433 cl_image_format array_format;
435 array_format.image_channel_data_type = CL_FLOAT;
436 array_format.image_channel_order = CL_R;
438 nbp->coulomb_tab_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE,
439 &array_format, 1, 1, 0, NULL, &cl_error);
442 nbp->coulomb_tab_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY, sizeof(cl_float), NULL, &cl_error);
443 // TODO: handle errors
446 nnbfp = 2*ntypes*ntypes;
447 nnbfp_comb = 2*ntypes;
450 /* Switched from using textures to using buffers */
451 // TODO: decide which alternative is most efficient - textures or buffers.
453 cl_image_format array_format;
455 array_format.image_channel_data_type = CL_FLOAT;
456 array_format.image_channel_order = CL_R;
458 nbp->nbfp_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
459 &array_format, nnbfp, 1, 0, nbat->nbfp, &cl_error);
462 nbp->nbfp_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, nnbfp*sizeof(cl_float), nbat->nbfp, &cl_error);
463 assert(cl_error == CL_SUCCESS);
464 // TODO: handle errors
466 if (ic->vdwtype == evdwPME)
468 /* Switched from using textures to using buffers */
469 // TODO: decide which alternative is most efficient - textures or buffers.
470 /* nbp->nbfp_comb_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
471 &array_format, nnbfp_comb, 1, 0, nbat->nbfp_comb, &cl_error);*/
472 nbp->nbfp_comb_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, nnbfp_comb*sizeof(cl_float), nbat->nbfp_comb, &cl_error);
475 assert(cl_error == CL_SUCCESS);
476 // TODO: handle errors
480 // TODO: improvement needed.
481 // The image2d is created here even if vdwtype is not evdwPME because the OpenCL kernels
482 // don't accept NULL values for image2D parameters.
483 /* Switched from using textures to using buffers */
484 // TODO: decide which alternative is most efficient - textures or buffers.
485 /* nbp->nbfp_comb_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE,
486 &array_format, 1, 1, 0, NULL, &cl_error);*/
487 nbp->nbfp_comb_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY, sizeof(cl_float), NULL, &cl_error);
490 assert(cl_error == CL_SUCCESS);
491 // TODO: handle errors
496 //! This function is documented in the header file
497 void nbnxn_gpu_pme_loadbal_update_param(const nonbonded_verlet_t *nbv,
498 const interaction_const_t *ic)
500 if (!nbv || nbv->grp[0].kernel_type != nbnxnk8x8x8_GPU)
504 gmx_nbnxn_ocl_t *nb = nbv->gpu_nbv;
505 cl_nbparam_t *nbp = nb->nbparam;
507 set_cutoff_parameters(nbp, ic);
509 nbp->eeltype = nbnxn_gpu_pick_ewald_kernel_type(ic->rcoulomb != ic->rvdw);
511 init_ewald_coulomb_force_table(ic, nb->nbparam, nb->dev_rundata);
514 /*! \brief Initializes the pair list data structure.
516 static void init_plist(cl_plist_t *pl)
518 /* initialize to NULL pointers to data that is not allocated here and will
519 need reallocation in nbnxn_gpu_init_pairlist */
524 /* size -1 indicates that the respective array hasn't been initialized yet */
531 pl->excl_nalloc = -1;
532 pl->bDoPrune = false;
535 /*! \brief Initializes the timer data structure.
537 static void init_timers(cl_timers_t gmx_unused *t, bool gmx_unused bUseTwoStreams)
539 /* Nothing to initialize for OpenCL */
542 /*! \brief Initializes the timings data structure.
544 static void init_timings(gmx_wallclock_gpu_t *t)
553 for (i = 0; i < 2; i++)
555 for (j = 0; j < 2; j++)
557 t->ktime[i][j].t = 0.0;
558 t->ktime[i][j].c = 0;
563 /*! \brief Creates context for OpenCL GPU given by \p mygpu
565 * A fatal error results if creation fails.
567 * \param[inout] runtimeData runtime data including program and context
568 * \param[in] devInfo device info struct
569 * \param[in] rank MPI rank (for error reporting)
572 nbnxn_gpu_create_context(gmx_device_runtime_data_t *runtimeData,
573 const gmx_device_info_t *devInfo,
576 cl_context_properties context_properties[3];
577 cl_platform_id platform_id;
578 cl_device_id device_id;
582 assert(runtimeData != NULL);
583 assert(devInfo != NULL);
585 platform_id = devInfo->ocl_gpu_id.ocl_platform_id;
586 device_id = devInfo->ocl_gpu_id.ocl_device_id;
588 context_properties[0] = CL_CONTEXT_PLATFORM;
589 context_properties[1] = (cl_context_properties) platform_id;
590 context_properties[2] = 0; /* Terminates the list of properties */
592 context = clCreateContext(context_properties, 1, &device_id, NULL, NULL, &cl_error);
593 if (CL_SUCCESS != cl_error)
595 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s:\n OpenCL error %d: %s",
597 devInfo->device_name,
598 cl_error, ocl_get_error_string(cl_error).c_str());
602 runtimeData->context = context;
605 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
606 static cl_kernel nbnxn_gpu_create_kernel(gmx_nbnxn_ocl_t *nb,
607 const char *kernel_name)
612 kernel = clCreateKernel(nb->dev_rundata->program, kernel_name, &cl_error);
613 if (CL_SUCCESS != cl_error)
615 gmx_fatal(FARGS, "Failed to create kernel '%s' for GPU #%s: OpenCL error %d",
617 nb->dev_info->device_name,
624 /*! \brief Clears nonbonded shift force output array and energy outputs on the GPU.
627 nbnxn_ocl_clear_e_fshift(gmx_nbnxn_ocl_t *nb)
631 cl_atomdata_t * adat = nb->atdat;
632 cl_command_queue ls = nb->stream[eintLocal];
634 size_t local_work_size[3] = {1, 1, 1};
635 size_t global_work_size[3] = {1, 1, 1};
637 cl_int shifts = SHIFTS*3;
641 cl_kernel zero_e_fshift = nb->kernel_zero_e_fshift;
643 local_work_size[0] = 64;
644 // Round the total number of threads up from the array size
645 global_work_size[0] = ((shifts + local_work_size[0] - 1)/local_work_size[0])*local_work_size[0];
648 cl_error = clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->fshift));
649 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_lj));
650 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_el));
651 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_uint), &shifts);
652 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
654 cl_error = clEnqueueNDRangeKernel(ls, zero_e_fshift, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL);
655 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
658 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
659 static void nbnxn_gpu_init_kernels(gmx_nbnxn_ocl_t *nb)
661 /* Init to 0 main kernel arrays */
662 /* They will be later on initialized in select_nbnxn_kernel */
663 memset(nb->kernel_ener_noprune_ptr, 0, sizeof(nb->kernel_ener_noprune_ptr));
664 memset(nb->kernel_ener_prune_ptr, 0, sizeof(nb->kernel_ener_prune_ptr));
665 memset(nb->kernel_noener_noprune_ptr, 0, sizeof(nb->kernel_noener_noprune_ptr));
666 memset(nb->kernel_noener_prune_ptr, 0, sizeof(nb->kernel_noener_prune_ptr));
668 /* Init auxiliary kernels */
669 nb->kernel_memset_f = nbnxn_gpu_create_kernel(nb, "memset_f");
670 nb->kernel_memset_f2 = nbnxn_gpu_create_kernel(nb, "memset_f2");
671 nb->kernel_memset_f3 = nbnxn_gpu_create_kernel(nb, "memset_f3");
672 nb->kernel_zero_e_fshift = nbnxn_gpu_create_kernel(nb, "zero_e_fshift");
675 /*! \brief Initializes simulation constant data.
677 * Initializes members of the atomdata and nbparam structs and
678 * clears e/fshift output buffers.
680 static void nbnxn_ocl_init_const(gmx_nbnxn_ocl_t *nb,
681 const interaction_const_t *ic,
682 const nonbonded_verlet_group_t *nbv_group)
684 init_atomdata_first(nb->atdat, nbv_group[0].nbat->ntype, nb->dev_rundata);
685 init_nbparam(nb->nbparam, ic, nbv_group[0].nbat, nb->dev_rundata);
689 //! This function is documented in the header file
690 void nbnxn_gpu_init(gmx_nbnxn_ocl_t **p_nb,
691 const gmx_gpu_info_t *gpu_info,
692 const gmx_gpu_opt_t *gpu_opt,
693 const interaction_const_t *ic,
694 nonbonded_verlet_group_t *nbv_grp,
697 gmx_bool bLocalAndNonlocal)
701 cl_command_queue_properties queue_properties;
714 snew(nb->nbparam, 1);
715 snew(nb->plist[eintLocal], 1);
716 if (bLocalAndNonlocal)
718 snew(nb->plist[eintNonlocal], 1);
721 nb->bUseTwoStreams = bLocalAndNonlocal;
724 snew(nb->timings, 1);
726 /* set device info, just point it to the right GPU among the detected ones */
727 nb->dev_info = gpu_info->gpu_dev + gpu_opt->dev_use[my_gpu_index];
728 snew(nb->dev_rundata, 1);
730 /* init to NULL the debug buffer */
731 nb->debug_buffer = NULL;
734 ocl_pmalloc((void**)&nb->nbst.e_lj, sizeof(*nb->nbst.e_lj));
735 ocl_pmalloc((void**)&nb->nbst.e_el, sizeof(*nb->nbst.e_el));
736 ocl_pmalloc((void**)&nb->nbst.fshift, SHIFTS * sizeof(*nb->nbst.fshift));
738 init_plist(nb->plist[eintLocal]);
740 /* OpenCL timing disabled if GMX_DISABLE_OCL_TIMING is defined. */
741 nb->bDoTime = (getenv("GMX_DISABLE_OCL_TIMING") == NULL);
743 /* Create queues only after bDoTime has been initialized */
746 queue_properties = CL_QUEUE_PROFILING_ENABLE;
750 queue_properties = 0;
753 nbnxn_gpu_create_context(nb->dev_rundata, nb->dev_info, rank);
755 /* local/non-local GPU streams */
756 nb->stream[eintLocal] = clCreateCommandQueue(nb->dev_rundata->context, nb->dev_info->ocl_gpu_id.ocl_device_id, queue_properties, &cl_error);
757 if (CL_SUCCESS != cl_error)
759 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s: OpenCL error %d",
761 nb->dev_info->device_name,
766 if (nb->bUseTwoStreams)
768 init_plist(nb->plist[eintNonlocal]);
770 nb->stream[eintNonlocal] = clCreateCommandQueue(nb->dev_rundata->context, nb->dev_info->ocl_gpu_id.ocl_device_id, queue_properties, &cl_error);
771 if (CL_SUCCESS != cl_error)
773 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s: OpenCL error %d",
775 nb->dev_info->device_name,
783 init_timers(nb->timers, nb->bUseTwoStreams);
784 init_timings(nb->timings);
787 nbnxn_ocl_init_const(nb, ic, nbv_grp);
789 /* Enable LJ param manual prefetch for AMD or if we request through env. var.
790 * TODO: decide about NVIDIA
792 nb->bPrefetchLjParam =
793 (getenv("GMX_OCL_DISABLE_I_PREFETCH") == NULL) &&
794 ((nb->dev_info->vendor_e == OCL_VENDOR_AMD) || (getenv("GMX_OCL_ENABLE_I_PREFETCH") != NULL));
796 /* NOTE: in CUDA we pick L1 cache configuration for the nbnxn kernels here,
797 * but sadly this is not supported in OpenCL (yet?). Consider adding it if
798 * it becomes supported.
800 nbnxn_gpu_compile_kernels(nb);
801 nbnxn_gpu_init_kernels(nb);
803 /* clear energy and shift force outputs */
804 nbnxn_ocl_clear_e_fshift(nb);
810 fprintf(debug, "Initialized OpenCL data structures.\n");
814 /*! \brief Clears the first natoms_clear elements of the GPU nonbonded force output array.
816 static void nbnxn_ocl_clear_f(gmx_nbnxn_ocl_t *nb, int natoms_clear)
818 if (natoms_clear == 0)
824 cl_atomdata_t * adat = nb->atdat;
825 cl_command_queue ls = nb->stream[eintLocal];
826 cl_float value = 0.0f;
828 size_t local_work_size[3] = {1, 1, 1};
829 size_t global_work_size[3] = {1, 1, 1};
833 cl_kernel memset_f = nb->kernel_memset_f;
835 cl_uint natoms_flat = natoms_clear * (sizeof(rvec)/sizeof(real));
837 local_work_size[0] = 64;
838 // Round the total number of threads up from the array size
839 global_work_size[0] = ((natoms_flat + local_work_size[0] - 1)/local_work_size[0])*local_work_size[0];
843 cl_error = clSetKernelArg(memset_f, arg_no++, sizeof(cl_mem), &(adat->f));
844 cl_error |= clSetKernelArg(memset_f, arg_no++, sizeof(cl_float), &value);
845 cl_error |= clSetKernelArg(memset_f, arg_no++, sizeof(cl_uint), &natoms_flat);
846 assert(cl_error == CL_SUCCESS);
848 cl_error = clEnqueueNDRangeKernel(ls, memset_f, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL);
849 assert(cl_error == CL_SUCCESS);
852 //! This function is documented in the header file
854 nbnxn_gpu_clear_outputs(gmx_nbnxn_ocl_t *nb,
857 nbnxn_ocl_clear_f(nb, nb->atdat->natoms);
858 /* clear shift force array and energies if the outputs were
859 used in the current step */
860 if (flags & GMX_FORCE_VIRIAL)
862 nbnxn_ocl_clear_e_fshift(nb);
865 /* kick off buffer clearing kernel to ensure concurrency with constraints/update */
866 cl_int gmx_unused cl_error;
867 cl_error = clFlush(nb->stream[eintLocal]);
868 assert(CL_SUCCESS == cl_error);
871 //! This function is documented in the header file
872 void nbnxn_gpu_init_pairlist(gmx_nbnxn_ocl_t *nb,
873 const nbnxn_pairlist_t *h_plist,
877 cl_command_queue stream = nb->stream[iloc];
878 cl_plist_t *d_plist = nb->plist[iloc];
880 if (d_plist->na_c < 0)
882 d_plist->na_c = h_plist->na_ci;
886 if (d_plist->na_c != h_plist->na_ci)
888 sprintf(sbuf, "In cu_init_plist: the #atoms per cell has changed (from %d to %d)",
889 d_plist->na_c, h_plist->na_ci);
894 ocl_realloc_buffered(&d_plist->sci, h_plist->sci, sizeof(nbnxn_sci_t),
895 &d_plist->nsci, &d_plist->sci_nalloc,
897 nb->dev_rundata->context,
898 stream, true, &(nb->timers->pl_h2d_sci[iloc]));
900 ocl_realloc_buffered(&d_plist->cj4, h_plist->cj4, sizeof(nbnxn_cj4_t),
901 &d_plist->ncj4, &d_plist->cj4_nalloc,
903 nb->dev_rundata->context,
904 stream, true, &(nb->timers->pl_h2d_cj4[iloc]));
906 ocl_realloc_buffered(&d_plist->excl, h_plist->excl, sizeof(nbnxn_excl_t),
907 &d_plist->nexcl, &d_plist->excl_nalloc,
909 nb->dev_rundata->context,
910 stream, true, &(nb->timers->pl_h2d_excl[iloc]));
912 /* need to prune the pair list during the next step */
913 d_plist->bDoPrune = true;
916 //! This function is documented in the header file
917 void nbnxn_gpu_upload_shiftvec(gmx_nbnxn_ocl_t *nb,
918 const nbnxn_atomdata_t *nbatom)
920 cl_atomdata_t *adat = nb->atdat;
921 cl_command_queue ls = nb->stream[eintLocal];
923 /* only if we have a dynamic box */
924 if (nbatom->bDynamicBox || !adat->bShiftVecUploaded)
926 ocl_copy_H2D_async(adat->shift_vec, nbatom->shift_vec, 0,
927 SHIFTS * adat->shift_vec_elem_size, ls, NULL);
928 adat->bShiftVecUploaded = true;
932 //! This function is documented in the header file
933 void nbnxn_gpu_init_atomdata(gmx_nbnxn_ocl_t *nb,
934 const struct nbnxn_atomdata_t *nbat)
939 bool bDoTime = nb->bDoTime;
940 cl_timers_t *timers = nb->timers;
941 cl_atomdata_t *d_atdat = nb->atdat;
942 cl_command_queue ls = nb->stream[eintLocal];
944 natoms = nbat->natoms;
947 /* need to reallocate if we have to copy more atoms than the amount of space
948 available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
949 if (natoms > d_atdat->nalloc)
951 nalloc = over_alloc_small(natoms);
953 /* free up first if the arrays have already been initialized */
954 if (d_atdat->nalloc != -1)
956 ocl_free_buffered(d_atdat->f, &d_atdat->natoms, &d_atdat->nalloc);
957 ocl_free_buffered(d_atdat->xq, NULL, NULL);
958 ocl_free_buffered(d_atdat->lj_comb, NULL, NULL);
959 ocl_free_buffered(d_atdat->atom_types, NULL, NULL);
962 d_atdat->f_elem_size = sizeof(rvec);
964 // TODO: handle errors, check clCreateBuffer flags
965 d_atdat->f = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_WRITE, nalloc * d_atdat->f_elem_size, NULL, &cl_error);
966 assert(CL_SUCCESS == cl_error);
968 // TODO: change the flag to read-only
969 d_atdat->xq = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_WRITE, nalloc * sizeof(cl_float4), NULL, &cl_error);
970 assert(CL_SUCCESS == cl_error);
971 // TODO: handle errors, check clCreateBuffer flags
973 if (useLjCombRule(nb->nbparam->vdwtype))
975 // TODO: change the flag to read-only
976 d_atdat->lj_comb = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_WRITE, nalloc * sizeof(cl_float2), NULL, &cl_error);
977 assert(CL_SUCCESS == cl_error);
978 // TODO: handle errors, check clCreateBuffer flags
982 // TODO: change the flag to read-only
983 d_atdat->atom_types = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_WRITE, nalloc * sizeof(int), NULL, &cl_error);
984 assert(CL_SUCCESS == cl_error);
985 // TODO: handle errors, check clCreateBuffer flags
988 d_atdat->nalloc = nalloc;
992 d_atdat->natoms = natoms;
993 d_atdat->natoms_local = nbat->natoms_local;
995 /* need to clear GPU f output if realloc happened */
998 nbnxn_ocl_clear_f(nb, nalloc);
1001 if (useLjCombRule(nb->nbparam->vdwtype))
1003 ocl_copy_H2D_async(d_atdat->lj_comb, nbat->lj_comb, 0,
1004 natoms*sizeof(cl_float2), ls, bDoTime ? &(timers->atdat) : NULL);
1008 ocl_copy_H2D_async(d_atdat->atom_types, nbat->type, 0,
1009 natoms*sizeof(int), ls, bDoTime ? &(timers->atdat) : NULL);
1013 /* kick off the tasks enqueued above to ensure concurrency with the search */
1014 cl_error = clFlush(ls);
1015 assert(CL_SUCCESS == cl_error);
1018 /*! \brief Releases an OpenCL kernel pointer */
1019 void free_kernel(cl_kernel *kernel_ptr)
1021 cl_int gmx_unused cl_error;
1023 assert(NULL != kernel_ptr);
1027 cl_error = clReleaseKernel(*kernel_ptr);
1028 assert(cl_error == CL_SUCCESS);
1034 /*! \brief Releases a list of OpenCL kernel pointers */
1035 void free_kernels(cl_kernel *kernels, int count)
1039 for (i = 0; i < count; i++)
1041 free_kernel(kernels + i);
1045 /*! \brief Free the OpenCL runtime data (context and program).
1047 * The function releases the OpenCL context and program assuciated with the
1048 * device that the calling PP rank is running on.
1050 * \param runData [in] porinter to the structure with runtime data.
1052 static void free_gpu_device_runtime_data(gmx_device_runtime_data_t *runData)
1054 if (runData == NULL)
1059 cl_int gmx_unused cl_error;
1061 if (runData->context)
1063 cl_error = clReleaseContext(runData->context);
1064 runData->context = NULL;
1065 assert(CL_SUCCESS == cl_error);
1068 if (runData->program)
1070 cl_error = clReleaseProgram(runData->program);
1071 runData->program = NULL;
1072 assert(CL_SUCCESS == cl_error);
1077 //! This function is documented in the header file
1078 void nbnxn_gpu_free(gmx_nbnxn_ocl_t *nb)
1083 kernel_count = sizeof(nb->kernel_ener_noprune_ptr) / sizeof(nb->kernel_ener_noprune_ptr[0][0]);
1084 free_kernels((cl_kernel*)nb->kernel_ener_noprune_ptr, kernel_count);
1086 kernel_count = sizeof(nb->kernel_ener_prune_ptr) / sizeof(nb->kernel_ener_prune_ptr[0][0]);
1087 free_kernels((cl_kernel*)nb->kernel_ener_prune_ptr, kernel_count);
1089 kernel_count = sizeof(nb->kernel_noener_noprune_ptr) / sizeof(nb->kernel_noener_noprune_ptr[0][0]);
1090 free_kernels((cl_kernel*)nb->kernel_noener_noprune_ptr, kernel_count);
1092 kernel_count = sizeof(nb->kernel_noener_prune_ptr) / sizeof(nb->kernel_noener_prune_ptr[0][0]);
1093 free_kernels((cl_kernel*)nb->kernel_noener_prune_ptr, kernel_count);
1095 free_kernel(&(nb->kernel_memset_f));
1096 free_kernel(&(nb->kernel_memset_f2));
1097 free_kernel(&(nb->kernel_memset_f3));
1098 free_kernel(&(nb->kernel_zero_e_fshift));
1101 free_ocl_buffer(&(nb->atdat->xq));
1102 free_ocl_buffer(&(nb->atdat->f));
1103 free_ocl_buffer(&(nb->atdat->e_lj));
1104 free_ocl_buffer(&(nb->atdat->e_el));
1105 free_ocl_buffer(&(nb->atdat->fshift));
1106 free_ocl_buffer(&(nb->atdat->lj_comb));
1107 free_ocl_buffer(&(nb->atdat->atom_types));
1108 free_ocl_buffer(&(nb->atdat->shift_vec));
1112 free_ocl_buffer(&(nb->nbparam->nbfp_climg2d));
1113 free_ocl_buffer(&(nb->nbparam->nbfp_comb_climg2d));
1114 free_ocl_buffer(&(nb->nbparam->coulomb_tab_climg2d));
1118 free_ocl_buffer(&(nb->plist[eintLocal]->sci));
1119 free_ocl_buffer(&(nb->plist[eintLocal]->cj4));
1120 free_ocl_buffer(&(nb->plist[eintLocal]->excl));
1121 sfree(nb->plist[eintLocal]);
1122 if (nb->bUseTwoStreams)
1124 free_ocl_buffer(&(nb->plist[eintNonlocal]->sci));
1125 free_ocl_buffer(&(nb->plist[eintNonlocal]->cj4));
1126 free_ocl_buffer(&(nb->plist[eintNonlocal]->excl));
1127 sfree(nb->plist[eintNonlocal]);
1131 ocl_pfree(nb->nbst.e_lj);
1132 nb->nbst.e_lj = NULL;
1134 ocl_pfree(nb->nbst.e_el);
1135 nb->nbst.e_el = NULL;
1137 ocl_pfree(nb->nbst.fshift);
1138 nb->nbst.fshift = NULL;
1140 /* Free debug buffer */
1141 free_ocl_buffer(&nb->debug_buffer);
1143 /* Free command queues */
1144 clReleaseCommandQueue(nb->stream[eintLocal]);
1145 nb->stream[eintLocal] = NULL;
1146 if (nb->bUseTwoStreams)
1148 clReleaseCommandQueue(nb->stream[eintNonlocal]);
1149 nb->stream[eintNonlocal] = NULL;
1151 /* Free other events */
1152 if (nb->nonlocal_done)
1154 clReleaseEvent(nb->nonlocal_done);
1155 nb->nonlocal_done = NULL;
1157 if (nb->misc_ops_and_local_H2D_done)
1159 clReleaseEvent(nb->misc_ops_and_local_H2D_done);
1160 nb->misc_ops_and_local_H2D_done = NULL;
1163 free_gpu_device_runtime_data(nb->dev_rundata);
1164 sfree(nb->dev_rundata);
1166 /* Free timers and timings */
1173 fprintf(debug, "Cleaned up OpenCL data structures.\n");
1178 //! This function is documented in the header file
1179 gmx_wallclock_gpu_t * nbnxn_gpu_get_timings(gmx_nbnxn_ocl_t *nb)
1181 return (nb != NULL && nb->bDoTime) ? nb->timings : NULL;
1184 //! This function is documented in the header file
1185 void nbnxn_gpu_reset_timings(nonbonded_verlet_t* nbv)
1187 if (nbv->gpu_nbv && nbv->gpu_nbv->bDoTime)
1189 init_timings(nbv->gpu_nbv->timings);
1193 //! This function is documented in the header file
1194 int nbnxn_gpu_min_ci_balanced(gmx_nbnxn_ocl_t *nb)
1197 gpu_min_ci_balanced_factor * nb->dev_info->compute_units : 0;
1200 //! This function is documented in the header file
1201 gmx_bool nbnxn_gpu_is_kernel_ewald_analytical(const gmx_nbnxn_ocl_t *nb)
1203 return ((nb->nbparam->eeltype == eelOclEWALD_ANA) ||
1204 (nb->nbparam->eeltype == eelOclEWALD_ANA_TWIN));