2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015, 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>
51 #include "gromacs/gmxlib/gpu_utils/gpu_utils.h"
52 #include "gromacs/gmxlib/ocl_tools/oclutils.h"
53 #include "gromacs/legacyheaders/gmx_detect_hardware.h"
54 #include "gromacs/legacyheaders/typedefs.h"
55 #include "gromacs/legacyheaders/types/enums.h"
56 #include "gromacs/legacyheaders/types/force_flags.h"
57 #include "gromacs/legacyheaders/types/interaction_const.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/pbcutil/ishift.h"
64 #include "gromacs/timing/gpu_timing.h"
65 #include "gromacs/utility/cstringutil.h"
66 #include "gromacs/utility/fatalerror.h"
67 #include "gromacs/utility/smalloc.h"
69 #include "nbnxn_ocl_types.h"
72 /*! \brief This parameter should be determined heuristically from the
73 * kernel execution times
75 * This value is best for small systems on a single AMD Radeon R9 290X
76 * (and about 5% faster than 40, which is the default for CUDA
77 * devices). Larger simulation systems were quite insensitive to the
78 * value of this parameter.
80 static unsigned int gpu_min_ci_balanced_factor = 50;
82 /*! \brief Helper function for warning output
84 * We should actually be using md_print_warn in md_logging.c,
85 * but we can't include mpi.h in OpenCL code.
87 static void md_print_warn(FILE *fplog,
94 /* We should only print to stderr on the master node,
95 * in most cases fplog is only set on the master node, so this works.
98 fprintf(stderr, "\n");
99 vfprintf(stderr, fmt, ap);
100 fprintf(stderr, "\n");
104 fprintf(fplog, "\n");
105 vfprintf(fplog, fmt, ap);
106 fprintf(fplog, "\n");
111 /*! \brief Free device buffers
113 * If the pointers to the size variables are NULL no resetting happens.
115 void ocl_free_buffered(cl_mem d_ptr, int *n, int *nalloc)
117 cl_int gmx_unused cl_error;
121 cl_error = clReleaseMemObject(d_ptr);
122 assert(cl_error == CL_SUCCESS);
123 // TODO: handle errors
137 /*! \brief Reallocation device buffers
139 * Reallocation of the memory pointed by d_ptr and copying of the data from
140 * the location pointed by h_src host-side pointer is done. Allocation is
141 * buffered and therefore freeing is only needed if the previously allocated
142 * space is not enough.
143 * The H2D copy is launched in command queue s and can be done synchronously or
144 * asynchronously (the default is the latter).
145 * If copy_event is not NULL, on return it will contain an event object
146 * identifying the H2D copy. The event can further be used to queue a wait
147 * for this operation or to query profiling information.
148 * OpenCL equivalent of cu_realloc_buffered.
150 void ocl_realloc_buffered(cl_mem *d_dest, void *h_src,
152 int *curr_size, int *curr_alloc_size,
157 cl_event *copy_event = NULL)
161 if (d_dest == NULL || req_size < 0)
166 /* reallocate only if the data does not fit = allocation size is smaller
167 than the current requested size */
168 if (req_size > *curr_alloc_size)
170 /* only free if the array has already been initialized */
171 if (*curr_alloc_size >= 0)
173 ocl_free_buffered(*d_dest, curr_size, curr_alloc_size);
176 *curr_alloc_size = over_alloc_large(req_size);
178 *d_dest = clCreateBuffer(context, CL_MEM_READ_WRITE, *curr_alloc_size * type_size, NULL, &cl_error);
179 assert(cl_error == CL_SUCCESS);
180 // TODO: handle errors, check clCreateBuffer flags
183 /* size could have changed without actual reallocation */
184 *curr_size = req_size;
186 /* upload to device */
191 ocl_copy_H2D_async(*d_dest, h_src, 0, *curr_size * type_size, s, copy_event);
195 ocl_copy_H2D(*d_dest, h_src, 0, *curr_size * type_size, s);
200 /*! \brief Releases the input OpenCL buffer */
201 static void free_ocl_buffer(cl_mem *buffer)
203 cl_int gmx_unused cl_error;
205 assert(NULL != buffer);
209 cl_error = clReleaseMemObject(*buffer);
210 assert(CL_SUCCESS == cl_error);
215 /*! \brief Tabulates the Ewald Coulomb force and initializes the size/scale
216 * and the table GPU array.
218 * If called with an already allocated table, it just re-uploads the
221 static void init_ewald_coulomb_force_table(const interaction_const_t *ic,
223 const gmx_device_info_t *dev_info)
229 if (nbp->coulomb_tab_climg2d != NULL)
231 free_ocl_buffer(&(nbp->coulomb_tab_climg2d));
234 /* Switched from using textures to using buffers */
235 // TODO: decide which alternative is most efficient - textures or buffers.
237 cl_image_format array_format;
239 array_format.image_channel_data_type = CL_FLOAT;
240 array_format.image_channel_order = CL_R;
242 coul_tab = clCreateImage2D(dev_info->context, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
243 &array_format, tabsize, 1, 0, ftmp, &cl_error);
246 coul_tab = clCreateBuffer(dev_info->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, ic->tabq_size*sizeof(cl_float), ic->tabq_coul_F, &cl_error);
247 assert(cl_error == CL_SUCCESS);
248 // TODO: handle errors, check clCreateBuffer flags
250 nbp->coulomb_tab_climg2d = coul_tab;
251 nbp->coulomb_tab_size = ic->tabq_size;
252 nbp->coulomb_tab_scale = ic->tabq_scale;
256 /*! \brief Initializes the atomdata structure first time, it only gets filled at
259 static void init_atomdata_first(cl_atomdata_t *ad, int ntypes, gmx_device_info_t *dev_info)
265 /* An element of the shift_vec device buffer has the same size as one element
266 of the host side shift_vec buffer. */
267 ad->shift_vec_elem_size = sizeof(*(((nbnxn_atomdata_t*)0)->shift_vec));
269 // TODO: handle errors, check clCreateBuffer flags
270 ad->shift_vec = clCreateBuffer(dev_info->context, CL_MEM_READ_WRITE, SHIFTS * ad->shift_vec_elem_size, NULL, &cl_error);
271 assert(cl_error == CL_SUCCESS);
272 ad->bShiftVecUploaded = false;
274 /* An element of the fshift device buffer has the same size as one element
275 of the host side fshift buffer. */
276 ad->fshift_elem_size = sizeof(*(((cl_nb_staging_t*)0)->fshift));
278 ad->fshift = clCreateBuffer(dev_info->context, CL_MEM_READ_WRITE, SHIFTS * ad->fshift_elem_size, NULL, &cl_error);
279 assert(cl_error == CL_SUCCESS);
280 // TODO: handle errors, check clCreateBuffer flags
282 ad->e_lj = clCreateBuffer(dev_info->context, CL_MEM_READ_WRITE, sizeof(float), NULL, &cl_error);
283 assert(cl_error == CL_SUCCESS);
284 // TODO: handle errors, check clCreateBuffer flags
286 ad->e_el = clCreateBuffer(dev_info->context, CL_MEM_READ_WRITE, sizeof(float), NULL, &cl_error);
287 assert(cl_error == CL_SUCCESS);
288 // TODO: handle errors, check clCreateBuffer flags
290 /* initialize to NULL pointers to data that is not allocated here and will
291 need reallocation in nbnxn_gpu_init_atomdata */
295 /* size -1 indicates that the respective array hasn't been initialized yet */
300 /*! \brief Copies all parameters related to the cut-off from ic to nbp
302 static void set_cutoff_parameters(cl_nbparam_t *nbp,
303 const interaction_const_t *ic)
305 nbp->ewald_beta = ic->ewaldcoeff_q;
306 nbp->sh_ewald = ic->sh_ewald;
307 nbp->epsfac = ic->epsfac;
308 nbp->two_k_rf = 2.0 * ic->k_rf;
309 nbp->c_rf = ic->c_rf;
310 nbp->rvdw_sq = ic->rvdw * ic->rvdw;
311 nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
312 nbp->rlist_sq = ic->rlist * ic->rlist;
314 nbp->sh_lj_ewald = ic->sh_lj_ewald;
315 nbp->ewaldcoeff_lj = ic->ewaldcoeff_lj;
317 nbp->rvdw_switch = ic->rvdw_switch;
318 nbp->dispersion_shift = ic->dispersion_shift;
319 nbp->repulsion_shift = ic->repulsion_shift;
320 nbp->vdw_switch = ic->vdw_switch;
323 /*! \brief Returns the kinds of electrostatics and Vdw OpenCL
324 * kernels that will be used.
326 * Respectively, these values are from enum eelOcl and enum
329 map_interaction_types_to_gpu_kernel_flavors(const interaction_const_t *ic,
333 if (ic->vdwtype == evdwCUT)
335 switch (ic->vdw_modifier)
338 case eintmodPOTSHIFT:
339 *gpu_vdwtype = evdwOclCUT;
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_info_t *dev_info)
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,
407 if (ic->vdwtype == evdwPME)
409 if (ic->ljpme_comb_rule == ljcrGEOM)
411 assert(nbat->comb_rule == ljcrGEOM);
415 assert(nbat->comb_rule == ljcrLB);
418 /* generate table for PME */
419 nbp->coulomb_tab_climg2d = NULL;
420 if (nbp->eeltype == eelOclEWALD_TAB || nbp->eeltype == eelOclEWALD_TAB_TWIN)
422 init_ewald_coulomb_force_table(ic, nbp, dev_info);
425 // TODO: improvement needed.
426 // The image2d is created here even if eeltype is not eelCuEWALD_TAB or eelCuEWALD_TAB_TWIN because the OpenCL kernels
427 // don't accept NULL values for image2D parameters.
429 /* Switched from using textures to using buffers */
430 // TODO: decide which alternative is most efficient - textures or buffers.
432 cl_image_format array_format;
434 array_format.image_channel_data_type = CL_FLOAT;
435 array_format.image_channel_order = CL_R;
437 nbp->coulomb_tab_climg2d = clCreateImage2D(dev_info->context, CL_MEM_READ_WRITE,
438 &array_format, 1, 1, 0, NULL, &cl_error);
441 nbp->coulomb_tab_climg2d = clCreateBuffer(dev_info->context, CL_MEM_READ_ONLY, sizeof(cl_float), NULL, &cl_error);
442 // TODO: handle errors
445 nnbfp = 2*ntypes*ntypes;
446 nnbfp_comb = 2*ntypes;
449 /* Switched from using textures to using buffers */
450 // TODO: decide which alternative is most efficient - textures or buffers.
452 cl_image_format array_format;
454 array_format.image_channel_data_type = CL_FLOAT;
455 array_format.image_channel_order = CL_R;
457 nbp->nbfp_climg2d = clCreateImage2D(dev_info->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
458 &array_format, nnbfp, 1, 0, nbat->nbfp, &cl_error);
461 nbp->nbfp_climg2d = clCreateBuffer(dev_info->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, nnbfp*sizeof(cl_float), nbat->nbfp, &cl_error);
462 assert(cl_error == CL_SUCCESS);
463 // TODO: handle errors
465 if (ic->vdwtype == evdwPME)
467 /* Switched from using textures to using buffers */
468 // TODO: decide which alternative is most efficient - textures or buffers.
469 /* nbp->nbfp_comb_climg2d = clCreateImage2D(dev_info->context, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
470 &array_format, nnbfp_comb, 1, 0, nbat->nbfp_comb, &cl_error);*/
471 nbp->nbfp_comb_climg2d = clCreateBuffer(dev_info->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, nnbfp_comb*sizeof(cl_float), nbat->nbfp_comb, &cl_error);
474 assert(cl_error == CL_SUCCESS);
475 // TODO: handle errors
479 // TODO: improvement needed.
480 // The image2d is created here even if vdwtype is not evdwPME because the OpenCL kernels
481 // don't accept NULL values for image2D parameters.
482 /* Switched from using textures to using buffers */
483 // TODO: decide which alternative is most efficient - textures or buffers.
484 /* nbp->nbfp_comb_climg2d = clCreateImage2D(dev_info->context, CL_MEM_READ_WRITE,
485 &array_format, 1, 1, 0, NULL, &cl_error);*/
486 nbp->nbfp_comb_climg2d = clCreateBuffer(dev_info->context, CL_MEM_READ_ONLY, sizeof(cl_float), NULL, &cl_error);
489 assert(cl_error == CL_SUCCESS);
490 // TODO: handle errors
495 //! This function is documented in the header file
496 void nbnxn_gpu_pme_loadbal_update_param(const nonbonded_verlet_t *nbv,
497 const interaction_const_t *ic)
499 if (!nbv || nbv->grp[0].kernel_type != nbnxnk8x8x8_GPU)
503 gmx_nbnxn_ocl_t *nb = nbv->gpu_nbv;
504 cl_nbparam_t *nbp = nb->nbparam;
506 set_cutoff_parameters(nbp, ic);
508 nbp->eeltype = nbnxn_gpu_pick_ewald_kernel_type(ic->rcoulomb != ic->rvdw);
510 init_ewald_coulomb_force_table(ic, nb->nbparam, nb->dev_info);
513 /*! \brief Initializes the pair list data structure.
515 static void init_plist(cl_plist_t *pl)
517 /* initialize to NULL pointers to data that is not allocated here and will
518 need reallocation in nbnxn_gpu_init_pairlist */
523 /* size -1 indicates that the respective array hasn't been initialized yet */
530 pl->excl_nalloc = -1;
531 pl->bDoPrune = false;
534 /*! \brief Initializes the timer data structure.
536 static void init_timers(cl_timers_t gmx_unused *t, bool gmx_unused bUseTwoStreams)
538 /* Nothing to initialize for OpenCL */
541 /*! \brief Initializes the timings data structure.
543 static void init_timings(gmx_wallclock_gpu_t *t)
552 for (i = 0; i < 2; i++)
554 for (j = 0; j < 2; j++)
556 t->ktime[i][j].t = 0.0;
557 t->ktime[i][j].c = 0;
562 /*! \brief Creates context for OpenCL GPU given by \p mygpu
564 * A fatal error results if creation fails.
566 * \param[inout] nb Manages OpenCL non-bonded calculations;
567 * contexts returned in dev_info members
568 * \param[in] rank MPI rank (for error reporting)
571 nbnxn_gpu_create_context(gmx_nbnxn_ocl_t *nb,
574 cl_context_properties context_properties[3];
575 cl_platform_id platform_id;
576 cl_device_id device_id;
580 platform_id = nb->dev_info->ocl_gpu_id.ocl_platform_id;
581 device_id = nb->dev_info->ocl_gpu_id.ocl_device_id;
583 context_properties[0] = CL_CONTEXT_PLATFORM;
584 context_properties[1] = (cl_context_properties) platform_id;
585 context_properties[2] = 0; /* Terminates the list of properties */
587 context = clCreateContext(context_properties, 1, &device_id, NULL, NULL, &cl_error);
588 if (CL_SUCCESS != cl_error)
590 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s: OpenCL error %d",
592 nb->dev_info->device_name,
597 nb->dev_info->context = context;
600 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
601 static cl_kernel nbnxn_gpu_create_kernel(gmx_nbnxn_ocl_t *nb,
602 const char *kernel_name)
607 kernel = clCreateKernel(nb->dev_info->program, kernel_name, &cl_error);
608 if (CL_SUCCESS != cl_error)
610 gmx_fatal(FARGS, "Failed to create kernel '%s' for GPU #%s: OpenCL error %d",
612 nb->dev_info->device_name,
619 /*! \brief Clears nonbonded shift force output array and energy outputs on the GPU.
622 nbnxn_ocl_clear_e_fshift(gmx_nbnxn_ocl_t *nb)
626 cl_atomdata_t * adat = nb->atdat;
627 cl_command_queue ls = nb->stream[eintLocal];
629 size_t local_work_size[3] = {1, 1, 1};
630 size_t global_work_size[3] = {1, 1, 1};
632 cl_int shifts = SHIFTS*3;
636 cl_kernel zero_e_fshift = nb->kernel_zero_e_fshift;
638 local_work_size[0] = 64;
639 global_work_size[0] = ((shifts/64)*64) + ((shifts%64) ? 64 : 0);
642 cl_error = clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->fshift));
643 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_lj));
644 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_el));
645 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_uint), &shifts);
646 assert(cl_error == CL_SUCCESS);
648 cl_error = clEnqueueNDRangeKernel(ls, zero_e_fshift, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL);
649 assert(cl_error == CL_SUCCESS);
653 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
654 static void nbnxn_gpu_init_kernels(gmx_nbnxn_ocl_t *nb)
656 /* Init to 0 main kernel arrays */
657 /* They will be later on initialized in select_nbnxn_kernel */
658 memset(nb->kernel_ener_noprune_ptr, 0, sizeof(nb->kernel_ener_noprune_ptr));
659 memset(nb->kernel_ener_prune_ptr, 0, sizeof(nb->kernel_ener_prune_ptr));
660 memset(nb->kernel_noener_noprune_ptr, 0, sizeof(nb->kernel_noener_noprune_ptr));
661 memset(nb->kernel_noener_prune_ptr, 0, sizeof(nb->kernel_noener_prune_ptr));
663 /* Init auxiliary kernels */
664 nb->kernel_memset_f = nbnxn_gpu_create_kernel(nb, "memset_f");
665 nb->kernel_memset_f2 = nbnxn_gpu_create_kernel(nb, "memset_f2");
666 nb->kernel_memset_f3 = nbnxn_gpu_create_kernel(nb, "memset_f3");
667 nb->kernel_zero_e_fshift = nbnxn_gpu_create_kernel(nb, "zero_e_fshift");
670 //! This function is documented in the header file
671 void nbnxn_gpu_init(FILE gmx_unused *fplog,
672 gmx_nbnxn_ocl_t **p_nb,
673 const gmx_gpu_info_t *gpu_info,
674 const gmx_gpu_opt_t *gpu_opt,
675 const interaction_const_t *ic,
676 nonbonded_verlet_group_t *nbv_grp,
679 gmx_bool bLocalAndNonlocal)
684 bool gmx_unused bStreamSync;
685 bool gmx_unused bNoStreamSync;
686 bool gmx_unused bTMPIAtomics;
687 bool gmx_unused bX86;
688 bool gmx_unused bOldDriver;
690 cl_command_queue_properties queue_properties;
703 snew(nb->nbparam, 1);
704 snew(nb->plist[eintLocal], 1);
705 if (bLocalAndNonlocal)
707 snew(nb->plist[eintNonlocal], 1);
710 nb->bUseTwoStreams = bLocalAndNonlocal;
713 snew(nb->timings, 1);
715 /* set device info, just point it to the right GPU among the detected ones */
716 nb->dev_info = gpu_info->gpu_dev + gpu_opt->dev_use[my_gpu_index];
718 /* init to NULL the debug buffer */
719 nb->debug_buffer = NULL;
722 ocl_pmalloc((void**)&nb->nbst.e_lj, sizeof(*nb->nbst.e_lj));
723 ocl_pmalloc((void**)&nb->nbst.e_el, sizeof(*nb->nbst.e_el));
724 ocl_pmalloc((void**)&nb->nbst.fshift, SHIFTS * sizeof(*nb->nbst.fshift));
726 init_plist(nb->plist[eintLocal]);
728 /* OpenCL timing disabled if GMX_DISABLE_OCL_TIMING is defined. */
729 nb->bDoTime = (getenv("GMX_DISABLE_OCL_TIMING") == NULL);
731 /* Create queues only after bDoTime has been initialized */
734 queue_properties = CL_QUEUE_PROFILING_ENABLE;
738 queue_properties = 0;
741 nbnxn_gpu_create_context(nb, rank);
743 /* local/non-local GPU streams */
744 nb->stream[eintLocal] = clCreateCommandQueue(nb->dev_info->context, nb->dev_info->ocl_gpu_id.ocl_device_id, queue_properties, &cl_error);
745 if (CL_SUCCESS != cl_error)
747 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s: OpenCL error %d",
749 nb->dev_info->device_name,
754 if (nb->bUseTwoStreams)
756 init_plist(nb->plist[eintNonlocal]);
758 nb->stream[eintNonlocal] = clCreateCommandQueue(nb->dev_info->context, nb->dev_info->ocl_gpu_id.ocl_device_id, queue_properties, &cl_error);
759 if (CL_SUCCESS != cl_error)
761 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s: OpenCL error %d",
763 nb->dev_info->device_name,
771 init_timers(nb->timers, nb->bUseTwoStreams);
772 init_timings(nb->timings);
775 // TODO: check if it's worth implementing for NVIDIA GPUs
776 ///////////* set the kernel type for the current GPU */
777 ///////////* pick L1 cache configuration */
778 //////////nbnxn_gpu_set_cacheconfig(nb->dev_info);
780 init_atomdata_first(nb->atdat, nbv_grp[0].nbat->ntype, nb->dev_info);
781 init_nbparam(nb->nbparam, ic, nbv_grp[0].nbat, nb->dev_info);
782 nbnxn_gpu_compile_kernels(nb);
783 nbnxn_gpu_init_kernels(nb);
784 // TODO put this elsewhere? also mirror it in cuda
785 nbnxn_ocl_clear_e_fshift(nb);
791 fprintf(debug, "Initialized OpenCL data structures.\n");
795 /*! \brief Clears the first natoms_clear elements of the GPU nonbonded force output array.
797 static void nbnxn_ocl_clear_f(gmx_nbnxn_ocl_t *nb, int natoms_clear)
801 cl_atomdata_t * adat = nb->atdat;
802 cl_command_queue ls = nb->stream[eintLocal];
803 cl_float value = 0.0f;
805 size_t local_work_size[3] = {1, 1, 1};
806 size_t global_work_size[3] = {1, 1, 1};
810 cl_kernel memset_f = nb->kernel_memset_f;
812 cl_uint natoms_flat = natoms_clear * (sizeof(rvec)/sizeof(real));
814 local_work_size[0] = 64;
815 global_work_size[0] = ((natoms_flat/local_work_size[0])*local_work_size[0]) + ((natoms_flat%local_work_size[0]) ? local_work_size[0] : 0);
818 cl_error = clSetKernelArg(memset_f, arg_no++, sizeof(cl_mem), &(adat->f));
819 cl_error |= clSetKernelArg(memset_f, arg_no++, sizeof(cl_float), &value);
820 cl_error |= clSetKernelArg(memset_f, arg_no++, sizeof(cl_uint), &natoms_flat);
821 assert(cl_error == CL_SUCCESS);
823 cl_error = clEnqueueNDRangeKernel(ls, memset_f, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL);
824 assert(cl_error == CL_SUCCESS);
827 //! This function is documented in the header file
829 nbnxn_gpu_clear_outputs(gmx_nbnxn_ocl_t *nb,
832 nbnxn_ocl_clear_f(nb, nb->atdat->natoms);
833 /* clear shift force array and energies if the outputs were
834 used in the current step */
835 if (flags & GMX_FORCE_VIRIAL)
837 nbnxn_ocl_clear_e_fshift(nb);
841 //! This function is documented in the header file
842 void nbnxn_gpu_init_pairlist(gmx_nbnxn_ocl_t *nb,
843 const nbnxn_pairlist_t *h_plist,
847 cl_command_queue stream = nb->stream[iloc];
848 cl_plist_t *d_plist = nb->plist[iloc];
850 if (d_plist->na_c < 0)
852 d_plist->na_c = h_plist->na_ci;
856 if (d_plist->na_c != h_plist->na_ci)
858 sprintf(sbuf, "In cu_init_plist: the #atoms per cell has changed (from %d to %d)",
859 d_plist->na_c, h_plist->na_ci);
864 ocl_realloc_buffered(&d_plist->sci, h_plist->sci, sizeof(nbnxn_sci_t),
865 &d_plist->nsci, &d_plist->sci_nalloc,
867 nb->dev_info->context,
868 stream, true, &(nb->timers->pl_h2d_sci[iloc]));
870 ocl_realloc_buffered(&d_plist->cj4, h_plist->cj4, sizeof(nbnxn_cj4_t),
871 &d_plist->ncj4, &d_plist->cj4_nalloc,
873 nb->dev_info->context,
874 stream, true, &(nb->timers->pl_h2d_cj4[iloc]));
876 ocl_realloc_buffered(&d_plist->excl, h_plist->excl, sizeof(nbnxn_excl_t),
877 &d_plist->nexcl, &d_plist->excl_nalloc,
879 nb->dev_info->context,
880 stream, true, &(nb->timers->pl_h2d_excl[iloc]));
882 /* need to prune the pair list during the next step */
883 d_plist->bDoPrune = true;
886 //! This function is documented in the header file
887 void nbnxn_gpu_upload_shiftvec(gmx_nbnxn_ocl_t *nb,
888 const nbnxn_atomdata_t *nbatom)
890 cl_atomdata_t *adat = nb->atdat;
891 cl_command_queue ls = nb->stream[eintLocal];
893 /* only if we have a dynamic box */
894 if (nbatom->bDynamicBox || !adat->bShiftVecUploaded)
896 ocl_copy_H2D_async(adat->shift_vec, nbatom->shift_vec, 0,
897 SHIFTS * adat->shift_vec_elem_size, ls, NULL);
898 adat->bShiftVecUploaded = true;
902 //! This function is documented in the header file
903 void nbnxn_gpu_init_atomdata(gmx_nbnxn_ocl_t *nb,
904 const struct nbnxn_atomdata_t *nbat)
909 bool bDoTime = nb->bDoTime;
910 cl_timers_t *timers = nb->timers;
911 cl_atomdata_t *d_atdat = nb->atdat;
912 cl_command_queue ls = nb->stream[eintLocal];
914 natoms = nbat->natoms;
917 /* need to reallocate if we have to copy more atoms than the amount of space
918 available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
919 if (natoms > d_atdat->nalloc)
921 nalloc = over_alloc_small(natoms);
923 /* free up first if the arrays have already been initialized */
924 if (d_atdat->nalloc != -1)
926 ocl_free_buffered(d_atdat->f, &d_atdat->natoms, &d_atdat->nalloc);
927 ocl_free_buffered(d_atdat->xq, NULL, NULL);
928 ocl_free_buffered(d_atdat->atom_types, NULL, NULL);
931 d_atdat->f_elem_size = sizeof(rvec);
933 // TODO: handle errors, check clCreateBuffer flags
934 d_atdat->f = clCreateBuffer(nb->dev_info->context, CL_MEM_READ_WRITE, nalloc * d_atdat->f_elem_size, NULL, &cl_error);
935 assert(CL_SUCCESS == cl_error);
937 d_atdat->xq = clCreateBuffer(nb->dev_info->context, CL_MEM_READ_WRITE, nalloc * sizeof(cl_float4), NULL, &cl_error);
938 assert(CL_SUCCESS == cl_error);
939 // TODO: handle errors, check clCreateBuffer flags
941 d_atdat->atom_types = clCreateBuffer(nb->dev_info->context, CL_MEM_READ_WRITE, nalloc * sizeof(int), NULL, &cl_error);
942 assert(CL_SUCCESS == cl_error);
943 // TODO: handle errors, check clCreateBuffer flags
945 d_atdat->nalloc = nalloc;
949 d_atdat->natoms = natoms;
950 d_atdat->natoms_local = nbat->natoms_local;
952 /* need to clear GPU f output if realloc happened */
955 nbnxn_ocl_clear_f(nb, nalloc);
958 ocl_copy_H2D_async(d_atdat->atom_types, nbat->type, 0,
959 natoms*sizeof(int), ls, bDoTime ? &(timers->atdat) : NULL);
962 /*! \brief Releases an OpenCL kernel pointer */
963 void free_kernel(cl_kernel *kernel_ptr)
965 cl_int gmx_unused cl_error;
967 assert(NULL != kernel_ptr);
971 cl_error = clReleaseKernel(*kernel_ptr);
972 assert(cl_error == CL_SUCCESS);
978 /*! \brief Releases a list of OpenCL kernel pointers */
979 void free_kernels(cl_kernel *kernels, int count)
983 for (i = 0; i < count; i++)
985 free_kernel(kernels + i);
989 //! This function is documented in the header file
990 void nbnxn_gpu_free(gmx_nbnxn_ocl_t *nb)
995 kernel_count = sizeof(nb->kernel_ener_noprune_ptr) / sizeof(nb->kernel_ener_noprune_ptr[0][0]);
996 free_kernels((cl_kernel*)nb->kernel_ener_noprune_ptr, kernel_count);
998 kernel_count = sizeof(nb->kernel_ener_prune_ptr) / sizeof(nb->kernel_ener_prune_ptr[0][0]);
999 free_kernels((cl_kernel*)nb->kernel_ener_prune_ptr, kernel_count);
1001 kernel_count = sizeof(nb->kernel_noener_noprune_ptr) / sizeof(nb->kernel_noener_noprune_ptr[0][0]);
1002 free_kernels((cl_kernel*)nb->kernel_noener_noprune_ptr, kernel_count);
1004 kernel_count = sizeof(nb->kernel_noener_prune_ptr) / sizeof(nb->kernel_noener_prune_ptr[0][0]);
1005 free_kernels((cl_kernel*)nb->kernel_noener_prune_ptr, kernel_count);
1007 free_kernel(&(nb->kernel_memset_f));
1008 free_kernel(&(nb->kernel_memset_f2));
1009 free_kernel(&(nb->kernel_memset_f3));
1010 free_kernel(&(nb->kernel_zero_e_fshift));
1013 free_ocl_buffer(&(nb->atdat->xq));
1014 free_ocl_buffer(&(nb->atdat->f));
1015 free_ocl_buffer(&(nb->atdat->e_lj));
1016 free_ocl_buffer(&(nb->atdat->e_el));
1017 free_ocl_buffer(&(nb->atdat->fshift));
1018 free_ocl_buffer(&(nb->atdat->atom_types));
1019 free_ocl_buffer(&(nb->atdat->shift_vec));
1023 free_ocl_buffer(&(nb->nbparam->nbfp_climg2d));
1024 free_ocl_buffer(&(nb->nbparam->nbfp_comb_climg2d));
1025 free_ocl_buffer(&(nb->nbparam->coulomb_tab_climg2d));
1029 free_ocl_buffer(&(nb->plist[eintLocal]->sci));
1030 free_ocl_buffer(&(nb->plist[eintLocal]->cj4));
1031 free_ocl_buffer(&(nb->plist[eintLocal]->excl));
1032 sfree(nb->plist[eintLocal]);
1033 if (nb->bUseTwoStreams)
1035 free_ocl_buffer(&(nb->plist[eintNonlocal]->sci));
1036 free_ocl_buffer(&(nb->plist[eintNonlocal]->cj4));
1037 free_ocl_buffer(&(nb->plist[eintNonlocal]->excl));
1038 sfree(nb->plist[eintNonlocal]);
1042 ocl_pfree(nb->nbst.e_lj);
1043 nb->nbst.e_lj = NULL;
1045 ocl_pfree(nb->nbst.e_el);
1046 nb->nbst.e_el = NULL;
1048 ocl_pfree(nb->nbst.fshift);
1049 nb->nbst.fshift = NULL;
1051 /* Free debug buffer */
1052 free_ocl_buffer(&nb->debug_buffer);
1054 /* Free command queues */
1055 clReleaseCommandQueue(nb->stream[eintLocal]);
1056 nb->stream[eintLocal] = NULL;
1057 if (nb->bUseTwoStreams)
1059 clReleaseCommandQueue(nb->stream[eintNonlocal]);
1060 nb->stream[eintNonlocal] = NULL;
1062 /* Free other events */
1063 if (nb->nonlocal_done)
1065 clReleaseEvent(nb->nonlocal_done);
1066 nb->nonlocal_done = NULL;
1068 if (nb->misc_ops_done)
1070 clReleaseEvent(nb->misc_ops_done);
1071 nb->misc_ops_done = NULL;
1074 /* Free timers and timings */
1081 fprintf(debug, "Cleaned up OpenCL data structures.\n");
1085 //! This function is documented in the header file
1086 gmx_wallclock_gpu_t * nbnxn_gpu_get_timings(gmx_nbnxn_ocl_t *nb)
1088 return (nb != NULL && nb->bDoTime) ? nb->timings : NULL;
1091 //! This function is documented in the header file
1092 void nbnxn_gpu_reset_timings(nonbonded_verlet_t* nbv)
1094 if (nbv->gpu_nbv && nbv->gpu_nbv->bDoTime)
1096 init_timings(nbv->gpu_nbv->timings);
1100 //! This function is documented in the header file
1101 int nbnxn_gpu_min_ci_balanced(gmx_nbnxn_ocl_t *nb)
1104 gpu_min_ci_balanced_factor * nb->dev_info->compute_units : 0;
1107 //! This function is documented in the header file
1108 gmx_bool nbnxn_gpu_is_kernel_ewald_analytical(const gmx_nbnxn_ocl_t *nb)
1110 return ((nb->nbparam->eeltype == eelOclEWALD_ANA) ||
1111 (nb->nbparam->eeltype == eelOclEWALD_ANA_TWIN));