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.
36 * \brief Define OpenCL implementation of nbnxm_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>
53 // TODO We would like to move this down, but the way gmx_nbnxn_gpu_t
54 // is currently declared means this has to be before gpu_types.h
55 #include "nbnxm_ocl_types.h"
57 // TODO Remove this comment when the above order issue is resolved
58 #include "gromacs/gpu_utils/gpu_utils.h"
59 #include "gromacs/gpu_utils/oclutils.h"
60 #include "gromacs/hardware/gpu_hw_info.h"
61 #include "gromacs/math/vectypes.h"
62 #include "gromacs/mdlib/force_flags.h"
63 #include "gromacs/mdtypes/interaction_const.h"
64 #include "gromacs/mdtypes/md_enums.h"
65 #include "gromacs/nbnxm/gpu_data_mgmt.h"
66 #include "gromacs/nbnxm/gpu_jit_support.h"
67 #include "gromacs/nbnxm/nbnxm.h"
68 #include "gromacs/nbnxm/nbnxm_gpu.h"
69 #include "gromacs/pbcutil/ishift.h"
70 #include "gromacs/timing/gpu_timing.h"
71 #include "gromacs/utility/cstringutil.h"
72 #include "gromacs/utility/fatalerror.h"
73 #include "gromacs/utility/gmxassert.h"
74 #include "gromacs/utility/real.h"
75 #include "gromacs/utility/smalloc.h"
77 #include "nbnxm_ocl_internal.h"
82 /*! \brief Copies of values from cl_driver_diagnostics_intel.h,
83 * which isn't guaranteed to be available. */
85 #define CL_CONTEXT_SHOW_DIAGNOSTICS_INTEL 0x4106
86 #define CL_CONTEXT_DIAGNOSTICS_LEVEL_GOOD_INTEL 0x1
87 #define CL_CONTEXT_DIAGNOSTICS_LEVEL_BAD_INTEL 0x2
88 #define CL_CONTEXT_DIAGNOSTICS_LEVEL_NEUTRAL_INTEL 0x4
91 /*! \brief This parameter should be determined heuristically from the
92 * kernel execution times
94 * This value is best for small systems on a single AMD Radeon R9 290X
95 * (and about 5% faster than 40, which is the default for CUDA
96 * devices). Larger simulation systems were quite insensitive to the
97 * value of this parameter.
99 static unsigned int gpu_min_ci_balanced_factor = 50;
102 /*! \brief Returns true if LJ combination rules are used in the non-bonded kernels.
104 * Full doc in nbnxn_ocl_internal.h */
105 bool useLjCombRule(int vdwType)
107 return (vdwType == evdwOclCUTCOMBGEOM ||
108 vdwType == evdwOclCUTCOMBLB);
111 /*! \brief Tabulates the Ewald Coulomb force and initializes the size/scale
112 * and the table GPU array.
114 * If called with an already allocated table, it just re-uploads the
117 static void init_ewald_coulomb_force_table(const interaction_const_t *ic,
119 const gmx_device_runtime_data_t *runData)
125 if (nbp->coulomb_tab_climg2d != nullptr)
127 freeDeviceBuffer(&(nbp->coulomb_tab_climg2d));
130 /* Switched from using textures to using buffers */
131 // TODO: decide which alternative is most efficient - textures or buffers.
133 cl_image_format array_format;
135 array_format.image_channel_data_type = CL_FLOAT;
136 array_format.image_channel_order = CL_R;
138 coul_tab = clCreateImage2D(runData->context, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
139 &array_format, tabsize, 1, 0, ftmp, &cl_error);
142 coul_tab = clCreateBuffer(runData->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
143 ic->tabq_size*sizeof(cl_float), ic->tabq_coul_F, &cl_error);
144 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
145 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
147 nbp->coulomb_tab_climg2d = coul_tab;
148 nbp->coulomb_tab_scale = ic->tabq_scale;
152 /*! \brief Initializes the atomdata structure first time, it only gets filled at
155 static void init_atomdata_first(cl_atomdata_t *ad, int ntypes, gmx_device_runtime_data_t *runData)
161 /* An element of the shift_vec device buffer has the same size as one element
162 of the host side shift_vec buffer. */
163 ad->shift_vec_elem_size = sizeof(*nbnxn_atomdata_t::shift_vec.data());
165 ad->shift_vec = clCreateBuffer(runData->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
166 SHIFTS * ad->shift_vec_elem_size, nullptr, &cl_error);
167 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
168 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
169 ad->bShiftVecUploaded = CL_FALSE;
171 /* An element of the fshift device buffer has the same size as one element
172 of the host side fshift buffer. */
173 ad->fshift_elem_size = sizeof(*cl_nb_staging_t::fshift);
175 ad->fshift = clCreateBuffer(runData->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
176 SHIFTS * ad->fshift_elem_size, nullptr, &cl_error);
177 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
178 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
180 ad->e_lj = clCreateBuffer(runData->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
181 sizeof(float), nullptr, &cl_error);
182 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
183 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
185 ad->e_el = clCreateBuffer(runData->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
186 sizeof(float), nullptr, &cl_error);
187 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
188 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
190 /* initialize to nullptr pointers to data that is not allocated here and will
191 need reallocation in nbnxn_gpu_init_atomdata */
195 /* size -1 indicates that the respective array hasn't been initialized yet */
200 /*! \brief Copies all parameters related to the cut-off from ic to nbp
202 static void set_cutoff_parameters(cl_nbparam_t *nbp,
203 const interaction_const_t *ic,
204 const NbnxnListParameters *listParams)
206 nbp->ewald_beta = ic->ewaldcoeff_q;
207 nbp->sh_ewald = ic->sh_ewald;
208 nbp->epsfac = ic->epsfac;
209 nbp->two_k_rf = 2.0 * ic->k_rf;
210 nbp->c_rf = ic->c_rf;
211 nbp->rvdw_sq = ic->rvdw * ic->rvdw;
212 nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
213 nbp->rlistOuter_sq = listParams->rlistOuter * listParams->rlistOuter;
214 nbp->rlistInner_sq = listParams->rlistInner * listParams->rlistInner;
215 nbp->useDynamicPruning = listParams->useDynamicPruning;
217 nbp->sh_lj_ewald = ic->sh_lj_ewald;
218 nbp->ewaldcoeff_lj = ic->ewaldcoeff_lj;
220 nbp->rvdw_switch = ic->rvdw_switch;
221 nbp->dispersion_shift = ic->dispersion_shift;
222 nbp->repulsion_shift = ic->repulsion_shift;
223 nbp->vdw_switch = ic->vdw_switch;
226 /*! \brief Returns the kinds of electrostatics and Vdw OpenCL
227 * kernels that will be used.
229 * Respectively, these values are from enum eelOcl and enum
232 map_interaction_types_to_gpu_kernel_flavors(const interaction_const_t *ic,
237 if (ic->vdwtype == evdwCUT)
239 switch (ic->vdw_modifier)
242 case eintmodPOTSHIFT:
246 *gpu_vdwtype = evdwOclCUT;
249 *gpu_vdwtype = evdwOclCUTCOMBGEOM;
252 *gpu_vdwtype = evdwOclCUTCOMBLB;
255 gmx_incons("The requested LJ combination rule is not implemented in the OpenCL GPU accelerated kernels!");
258 case eintmodFORCESWITCH:
259 *gpu_vdwtype = evdwOclFSWITCH;
261 case eintmodPOTSWITCH:
262 *gpu_vdwtype = evdwOclPSWITCH;
265 gmx_incons("The requested VdW interaction modifier is not implemented in the GPU accelerated kernels!");
268 else if (ic->vdwtype == evdwPME)
270 if (ic->ljpme_comb_rule == ljcrGEOM)
272 *gpu_vdwtype = evdwOclEWALDGEOM;
276 *gpu_vdwtype = evdwOclEWALDLB;
281 gmx_incons("The requested VdW type is not implemented in the GPU accelerated kernels!");
284 if (ic->eeltype == eelCUT)
286 *gpu_eeltype = eelOclCUT;
288 else if (EEL_RF(ic->eeltype))
290 *gpu_eeltype = eelOclRF;
292 else if ((EEL_PME(ic->eeltype) || ic->eeltype == eelEWALD))
294 /* Initially rcoulomb == rvdw, so it's surely not twin cut-off. */
295 *gpu_eeltype = gpu_pick_ewald_kernel_type(false);
299 /* Shouldn't happen, as this is checked when choosing Verlet-scheme */
300 gmx_incons("The requested electrostatics type is not implemented in the GPU accelerated kernels!");
304 /*! \brief Initializes the nonbonded parameter data structure.
306 static void init_nbparam(cl_nbparam_t *nbp,
307 const interaction_const_t *ic,
308 const NbnxnListParameters *listParams,
309 const nbnxn_atomdata_t::Params &nbatParams,
310 const gmx_device_runtime_data_t *runData)
314 set_cutoff_parameters(nbp, ic, listParams);
316 map_interaction_types_to_gpu_kernel_flavors(ic,
317 nbatParams.comb_rule,
321 if (ic->vdwtype == evdwPME)
323 if (ic->ljpme_comb_rule == ljcrGEOM)
325 GMX_ASSERT(nbatParams.comb_rule == ljcrGEOM, "Combination rule mismatch!");
329 GMX_ASSERT(nbatParams.comb_rule == ljcrLB, "Combination rule mismatch!");
332 /* generate table for PME */
333 nbp->coulomb_tab_climg2d = nullptr;
334 if (nbp->eeltype == eelOclEWALD_TAB || nbp->eeltype == eelOclEWALD_TAB_TWIN)
336 init_ewald_coulomb_force_table(ic, nbp, runData);
339 // TODO: improvement needed.
340 // The image2d is created here even if eeltype is not eelCuEWALD_TAB or eelCuEWALD_TAB_TWIN because the OpenCL kernels
341 // don't accept nullptr values for image2D parameters.
343 /* Switched from using textures to using buffers */
344 // TODO: decide which alternative is most efficient - textures or buffers.
346 cl_image_format array_format;
348 array_format.image_channel_data_type = CL_FLOAT;
349 array_format.image_channel_order = CL_R;
351 nbp->coulomb_tab_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE,
352 &array_format, 1, 1, 0, nullptr, &cl_error);
355 nbp->coulomb_tab_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY, sizeof(cl_float), nullptr, &cl_error);
356 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
357 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
360 const int nnbfp = 2*nbatParams.numTypes*nbatParams.numTypes;
361 const int nnbfp_comb = 2*nbatParams.numTypes;
364 /* Switched from using textures to using buffers */
365 // TODO: decide which alternative is most efficient - textures or buffers.
367 cl_image_format array_format;
369 array_format.image_channel_data_type = CL_FLOAT;
370 array_format.image_channel_order = CL_R;
372 nbp->nbfp_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
373 &array_format, nnbfp, 1, 0, nbat->nbfp, &cl_error);
377 clCreateBuffer(runData->context,
378 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
379 nnbfp*sizeof(cl_float),
380 const_cast<float *>(nbatParams.nbfp.data()),
382 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
383 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
385 if (ic->vdwtype == evdwPME)
387 /* Switched from using textures to using buffers */
388 // TODO: decide which alternative is most efficient - textures or buffers.
389 /* nbp->nbfp_comb_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
390 &array_format, nnbfp_comb, 1, 0, nbat->nbfp_comb, &cl_error);*/
391 nbp->nbfp_comb_climg2d =
392 clCreateBuffer(runData->context,
393 CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY | CL_MEM_COPY_HOST_PTR,
394 nnbfp_comb*sizeof(cl_float),
395 const_cast<float *>(nbatParams.nbfp_comb.data()),
397 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
398 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
402 // TODO: improvement needed.
403 // The image2d is created here even if vdwtype is not evdwPME because the OpenCL kernels
404 // don't accept nullptr values for image2D parameters.
405 /* Switched from using textures to using buffers */
406 // TODO: decide which alternative is most efficient - textures or buffers.
407 /* nbp->nbfp_comb_climg2d = clCreateImage2D(runData->context, CL_MEM_READ_WRITE,
408 &array_format, 1, 1, 0, nullptr, &cl_error);*/
409 nbp->nbfp_comb_climg2d = clCreateBuffer(runData->context, CL_MEM_READ_ONLY, sizeof(cl_float), nullptr, &cl_error);
410 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
411 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
416 //! This function is documented in the header file
417 void gpu_pme_loadbal_update_param(const nonbonded_verlet_t *nbv,
418 const interaction_const_t *ic,
419 const NbnxnListParameters *listParams)
421 if (!nbv || !nbv->useGpu())
425 gmx_nbnxn_ocl_t *nb = nbv->gpu_nbv;
426 cl_nbparam_t *nbp = nb->nbparam;
428 set_cutoff_parameters(nbp, ic, listParams);
430 nbp->eeltype = gpu_pick_ewald_kernel_type(ic->rcoulomb != ic->rvdw);
432 init_ewald_coulomb_force_table(ic, nbp, nb->dev_rundata);
435 /*! \brief Initializes the pair list data structure.
437 static void init_plist(cl_plist_t *pl)
439 /* initialize to nullptr pointers to data that is not allocated here and will
440 need reallocation in nbnxn_gpu_init_pairlist */
446 /* size -1 indicates that the respective array hasn't been initialized yet */
453 pl->imask_nalloc = -1;
455 pl->excl_nalloc = -1;
456 pl->haveFreshList = false;
459 /*! \brief Initializes the timings data structure.
461 static void init_timings(gmx_wallclock_gpu_nbnxn_t *t)
470 for (i = 0; i < 2; i++)
472 for (j = 0; j < 2; j++)
474 t->ktime[i][j].t = 0.0;
475 t->ktime[i][j].c = 0;
480 t->pruneTime.t = 0.0;
481 t->dynamicPruneTime.c = 0;
482 t->dynamicPruneTime.t = 0.0;
486 //! OpenCL notification callback function
487 static void CL_CALLBACK
488 ocl_notify_fn( const char *pErrInfo, const void *, size_t, void *)
490 if (pErrInfo != NULL)
492 printf("%s\n", pErrInfo ); // Print error/hint
496 /*! \brief Creates context for OpenCL GPU given by \p mygpu
498 * A fatal error results if creation fails.
500 * \param[inout] runtimeData runtime data including program and context
501 * \param[in] devInfo device info struct
502 * \param[in] rank MPI rank (for error reporting)
505 nbnxn_gpu_create_context(gmx_device_runtime_data_t *runtimeData,
506 const gmx_device_info_t *devInfo,
509 cl_context_properties context_properties[5];
510 cl_platform_id platform_id;
511 cl_device_id device_id;
515 assert(runtimeData != nullptr);
516 assert(devInfo != nullptr);
518 platform_id = devInfo->ocl_gpu_id.ocl_platform_id;
519 device_id = devInfo->ocl_gpu_id.ocl_device_id;
522 context_properties[i++] = CL_CONTEXT_PLATFORM;
523 context_properties[i++] = reinterpret_cast<cl_context_properties>(platform_id);
524 if (getenv("GMX_OCL_SHOW_DIAGNOSTICS"))
526 context_properties[i++] = CL_CONTEXT_SHOW_DIAGNOSTICS_INTEL;
527 context_properties[i++] = CL_CONTEXT_DIAGNOSTICS_LEVEL_BAD_INTEL | CL_CONTEXT_DIAGNOSTICS_LEVEL_NEUTRAL_INTEL;
529 context_properties[i++] = 0; /* Terminates the list of properties */
531 context = clCreateContext(context_properties, 1, &device_id, ocl_notify_fn, nullptr, &cl_error);
532 if (CL_SUCCESS != cl_error)
534 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s:\n OpenCL error %d: %s",
536 devInfo->device_name,
537 cl_error, ocl_get_error_string(cl_error).c_str());
540 runtimeData->context = context;
543 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
544 static cl_kernel nbnxn_gpu_create_kernel(gmx_nbnxn_ocl_t *nb,
545 const char *kernel_name)
550 kernel = clCreateKernel(nb->dev_rundata->program, kernel_name, &cl_error);
551 if (CL_SUCCESS != cl_error)
553 gmx_fatal(FARGS, "Failed to create kernel '%s' for GPU #%s: OpenCL error %d",
555 nb->dev_info->device_name,
562 /*! \brief Clears nonbonded shift force output array and energy outputs on the GPU.
565 nbnxn_ocl_clear_e_fshift(gmx_nbnxn_ocl_t *nb)
569 cl_atomdata_t * adat = nb->atdat;
570 cl_command_queue ls = nb->stream[InteractionLocality::Local];
572 size_t local_work_size[3] = {1, 1, 1};
573 size_t global_work_size[3] = {1, 1, 1};
575 cl_int shifts = SHIFTS*3;
579 cl_kernel zero_e_fshift = nb->kernel_zero_e_fshift;
581 local_work_size[0] = 64;
582 // Round the total number of threads up from the array size
583 global_work_size[0] = ((shifts + local_work_size[0] - 1)/local_work_size[0])*local_work_size[0];
586 cl_error = clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->fshift));
587 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_lj));
588 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_mem), &(adat->e_el));
589 cl_error |= clSetKernelArg(zero_e_fshift, arg_no++, sizeof(cl_uint), &shifts);
590 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
592 cl_error = clEnqueueNDRangeKernel(ls, zero_e_fshift, 3, nullptr, global_work_size, local_work_size, 0, nullptr, nullptr);
593 GMX_ASSERT(cl_error == CL_SUCCESS, ocl_get_error_string(cl_error).c_str());
596 /*! \brief Initializes the OpenCL kernel pointers of the nbnxn_ocl_ptr_t input data structure. */
597 static void nbnxn_gpu_init_kernels(gmx_nbnxn_ocl_t *nb)
599 /* Init to 0 main kernel arrays */
600 /* They will be later on initialized in select_nbnxn_kernel */
601 // TODO: consider always creating all variants of the kernels here so that there is no
602 // need for late call to clCreateKernel -- if that gives any advantage?
603 memset(nb->kernel_ener_noprune_ptr, 0, sizeof(nb->kernel_ener_noprune_ptr));
604 memset(nb->kernel_ener_prune_ptr, 0, sizeof(nb->kernel_ener_prune_ptr));
605 memset(nb->kernel_noener_noprune_ptr, 0, sizeof(nb->kernel_noener_noprune_ptr));
606 memset(nb->kernel_noener_prune_ptr, 0, sizeof(nb->kernel_noener_prune_ptr));
608 /* Init pruning kernels
610 * TODO: we could avoid creating kernels if dynamic pruning is turned off,
611 * but ATM that depends on force flags not passed into the initialization.
613 nb->kernel_pruneonly[epruneFirst] = nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_opencl");
614 nb->kernel_pruneonly[epruneRolling] = nbnxn_gpu_create_kernel(nb, "nbnxn_kernel_prune_rolling_opencl");
616 /* Init auxiliary kernels */
617 nb->kernel_zero_e_fshift = nbnxn_gpu_create_kernel(nb, "zero_e_fshift");
620 /*! \brief Initializes simulation constant data.
622 * Initializes members of the atomdata and nbparam structs and
623 * clears e/fshift output buffers.
625 static void nbnxn_ocl_init_const(gmx_nbnxn_ocl_t *nb,
626 const interaction_const_t *ic,
627 const NbnxnListParameters *listParams,
628 const nbnxn_atomdata_t::Params &nbatParams)
630 init_atomdata_first(nb->atdat, nbatParams.numTypes, nb->dev_rundata);
631 init_nbparam(nb->nbparam, ic, listParams, nbatParams, nb->dev_rundata);
635 //! This function is documented in the header file
637 gpu_init(const gmx_device_info_t *deviceInfo,
638 const interaction_const_t *ic,
639 const NbnxnListParameters *listParams,
640 const nbnxn_atomdata_t *nbat,
642 const gmx_bool bLocalAndNonlocal)
646 cl_command_queue_properties queue_properties;
652 snew(nb->nbparam, 1);
653 snew(nb->plist[InteractionLocality::Local], 1);
654 if (bLocalAndNonlocal)
656 snew(nb->plist[InteractionLocality::NonLocal], 1);
659 nb->bUseTwoStreams = static_cast<cl_bool>(bLocalAndNonlocal);
661 nb->timers = new cl_timers_t();
662 snew(nb->timings, 1);
664 /* set device info, just point it to the right GPU among the detected ones */
665 nb->dev_info = deviceInfo;
666 snew(nb->dev_rundata, 1);
669 pmalloc(reinterpret_cast<void**>(&nb->nbst.e_lj), sizeof(*nb->nbst.e_lj));
670 pmalloc(reinterpret_cast<void**>(&nb->nbst.e_el), sizeof(*nb->nbst.e_el));
671 pmalloc(reinterpret_cast<void**>(&nb->nbst.fshift), SHIFTS * sizeof(*nb->nbst.fshift));
673 init_plist(nb->plist[InteractionLocality::Local]);
675 /* OpenCL timing disabled if GMX_DISABLE_GPU_TIMING is defined. */
676 nb->bDoTime = static_cast<cl_bool>(getenv("GMX_DISABLE_GPU_TIMING") == nullptr);
678 /* Create queues only after bDoTime has been initialized */
681 queue_properties = CL_QUEUE_PROFILING_ENABLE;
685 queue_properties = 0;
688 nbnxn_gpu_create_context(nb->dev_rundata, nb->dev_info, rank);
690 /* local/non-local GPU streams */
691 nb->stream[InteractionLocality::Local] =
692 clCreateCommandQueue(nb->dev_rundata->context, nb->dev_info->ocl_gpu_id.ocl_device_id, queue_properties, &cl_error);
693 if (CL_SUCCESS != cl_error)
695 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s: OpenCL error %d",
697 nb->dev_info->device_name,
701 if (nb->bUseTwoStreams)
703 init_plist(nb->plist[InteractionLocality::NonLocal]);
705 nb->stream[InteractionLocality::NonLocal] =
706 clCreateCommandQueue(nb->dev_rundata->context, nb->dev_info->ocl_gpu_id.ocl_device_id, queue_properties, &cl_error);
707 if (CL_SUCCESS != cl_error)
709 gmx_fatal(FARGS, "On rank %d failed to create context for GPU #%s: OpenCL error %d",
711 nb->dev_info->device_name,
718 init_timings(nb->timings);
721 nbnxn_ocl_init_const(nb, ic, listParams, nbat->params());
723 /* Enable LJ param manual prefetch for AMD or Intel or if we request through env. var.
724 * TODO: decide about NVIDIA
726 nb->bPrefetchLjParam =
727 (getenv("GMX_OCL_DISABLE_I_PREFETCH") == nullptr) &&
728 ((nb->dev_info->vendor_e == OCL_VENDOR_AMD) || (nb->dev_info->vendor_e == OCL_VENDOR_INTEL)
729 || (getenv("GMX_OCL_ENABLE_I_PREFETCH") != nullptr));
731 /* NOTE: in CUDA we pick L1 cache configuration for the nbnxn kernels here,
732 * but sadly this is not supported in OpenCL (yet?). Consider adding it if
733 * it becomes supported.
735 nbnxn_gpu_compile_kernels(nb);
736 nbnxn_gpu_init_kernels(nb);
738 /* clear energy and shift force outputs */
739 nbnxn_ocl_clear_e_fshift(nb);
743 fprintf(debug, "Initialized OpenCL data structures.\n");
749 /*! \brief Clears the first natoms_clear elements of the GPU nonbonded force output array.
751 static void nbnxn_ocl_clear_f(gmx_nbnxn_ocl_t *nb, int natoms_clear)
753 if (natoms_clear == 0)
758 cl_int gmx_used_in_debug cl_error;
760 cl_atomdata_t *atomData = nb->atdat;
761 cl_command_queue ls = nb->stream[InteractionLocality::Local];
762 cl_float value = 0.0f;
764 cl_error = clEnqueueFillBuffer(ls, atomData->f, &value, sizeof(cl_float),
765 0, natoms_clear*sizeof(rvec), 0, nullptr, nullptr);
766 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS, ("clEnqueueFillBuffer failed: " +
767 ocl_get_error_string(cl_error)).c_str());
770 //! This function is documented in the header file
772 gpu_clear_outputs(gmx_nbnxn_ocl_t *nb,
775 nbnxn_ocl_clear_f(nb, nb->atdat->natoms);
776 /* clear shift force array and energies if the outputs were
777 used in the current step */
778 if (flags & GMX_FORCE_VIRIAL)
780 nbnxn_ocl_clear_e_fshift(nb);
783 /* kick off buffer clearing kernel to ensure concurrency with constraints/update */
784 cl_int gmx_unused cl_error;
785 cl_error = clFlush(nb->stream[InteractionLocality::Local]);
786 assert(CL_SUCCESS == cl_error);
789 //! This function is documented in the header file
790 void gpu_init_pairlist(gmx_nbnxn_ocl_t *nb,
791 const NbnxnPairlistGpu *h_plist,
792 const InteractionLocality iloc)
795 // Timing accumulation should happen only if there was work to do
796 // because getLastRangeTime() gets skipped with empty lists later
797 // which leads to the counter not being reset.
798 bool bDoTime = ((nb->bDoTime == CL_TRUE) && !h_plist->sci.empty());
799 cl_command_queue stream = nb->stream[iloc];
800 cl_plist_t *d_plist = nb->plist[iloc];
802 if (d_plist->na_c < 0)
804 d_plist->na_c = h_plist->na_ci;
808 if (d_plist->na_c != h_plist->na_ci)
810 sprintf(sbuf, "In cu_init_plist: the #atoms per cell has changed (from %d to %d)",
811 d_plist->na_c, h_plist->na_ci);
816 gpu_timers_t::Interaction &iTimers = nb->timers->interaction[iloc];
820 iTimers.pl_h2d.openTimingRegion(stream);
821 iTimers.didPairlistH2D = true;
824 // TODO most of this function is same in CUDA and OpenCL, move into the header
825 Context context = nb->dev_rundata->context;
827 reallocateDeviceBuffer(&d_plist->sci, h_plist->sci.size(),
828 &d_plist->nsci, &d_plist->sci_nalloc, context);
829 copyToDeviceBuffer(&d_plist->sci, h_plist->sci.data(), 0, h_plist->sci.size(),
830 stream, GpuApiCallBehavior::Async,
831 bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
833 reallocateDeviceBuffer(&d_plist->cj4, h_plist->cj4.size(),
834 &d_plist->ncj4, &d_plist->cj4_nalloc, context);
835 copyToDeviceBuffer(&d_plist->cj4, h_plist->cj4.data(), 0, h_plist->cj4.size(),
836 stream, GpuApiCallBehavior::Async,
837 bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
839 reallocateDeviceBuffer(&d_plist->imask, h_plist->cj4.size()*c_nbnxnGpuClusterpairSplit,
840 &d_plist->nimask, &d_plist->imask_nalloc, context);
842 reallocateDeviceBuffer(&d_plist->excl, h_plist->excl.size(),
843 &d_plist->nexcl, &d_plist->excl_nalloc, context);
844 copyToDeviceBuffer(&d_plist->excl, h_plist->excl.data(), 0, h_plist->excl.size(),
845 stream, GpuApiCallBehavior::Async,
846 bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
850 iTimers.pl_h2d.closeTimingRegion(stream);
853 /* need to prune the pair list during the next step */
854 d_plist->haveFreshList = true;
857 //! This function is documented in the header file
858 void gpu_upload_shiftvec(gmx_nbnxn_ocl_t *nb,
859 const nbnxn_atomdata_t *nbatom)
861 cl_atomdata_t *adat = nb->atdat;
862 cl_command_queue ls = nb->stream[InteractionLocality::Local];
864 /* only if we have a dynamic box */
865 if (nbatom->bDynamicBox || !adat->bShiftVecUploaded)
867 ocl_copy_H2D_async(adat->shift_vec, nbatom->shift_vec.data(), 0,
868 SHIFTS * adat->shift_vec_elem_size, ls, nullptr);
869 adat->bShiftVecUploaded = CL_TRUE;
873 //! This function is documented in the header file
874 void gpu_init_atomdata(gmx_nbnxn_ocl_t *nb,
875 const nbnxn_atomdata_t *nbat)
880 bool bDoTime = nb->bDoTime == CL_TRUE;
881 cl_timers_t *timers = nb->timers;
882 cl_atomdata_t *d_atdat = nb->atdat;
883 cl_command_queue ls = nb->stream[InteractionLocality::Local];
885 natoms = nbat->numAtoms();
890 /* time async copy */
891 timers->atdat.openTimingRegion(ls);
894 /* need to reallocate if we have to copy more atoms than the amount of space
895 available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
896 if (natoms > d_atdat->nalloc)
898 nalloc = over_alloc_small(natoms);
900 /* free up first if the arrays have already been initialized */
901 if (d_atdat->nalloc != -1)
903 freeDeviceBuffer(&d_atdat->f);
904 freeDeviceBuffer(&d_atdat->xq);
905 freeDeviceBuffer(&d_atdat->lj_comb);
906 freeDeviceBuffer(&d_atdat->atom_types);
909 d_atdat->f_elem_size = sizeof(rvec);
911 d_atdat->f = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_WRITE | CL_MEM_HOST_READ_ONLY,
912 nalloc * d_atdat->f_elem_size, nullptr, &cl_error);
913 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
914 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
916 d_atdat->xq = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
917 nalloc * sizeof(cl_float4), nullptr, &cl_error);
918 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
919 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
921 if (useLjCombRule(nb->nbparam->vdwtype))
923 d_atdat->lj_comb = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
924 nalloc * sizeof(cl_float2), nullptr, &cl_error);
925 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
926 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
930 d_atdat->atom_types = clCreateBuffer(nb->dev_rundata->context, CL_MEM_READ_ONLY | CL_MEM_HOST_WRITE_ONLY,
931 nalloc * sizeof(int), nullptr, &cl_error);
932 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
933 ("clCreateBuffer failed: " + ocl_get_error_string(cl_error)).c_str());
936 d_atdat->nalloc = nalloc;
940 d_atdat->natoms = natoms;
941 d_atdat->natoms_local = nbat->natoms_local;
943 /* need to clear GPU f output if realloc happened */
946 nbnxn_ocl_clear_f(nb, nalloc);
949 if (useLjCombRule(nb->nbparam->vdwtype))
951 ocl_copy_H2D_async(d_atdat->lj_comb, nbat->params().lj_comb.data(), 0,
952 natoms*sizeof(cl_float2), ls, bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
956 ocl_copy_H2D_async(d_atdat->atom_types, nbat->params().type.data(), 0,
957 natoms*sizeof(int), ls, bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
963 timers->atdat.closeTimingRegion(ls);
966 /* kick off the tasks enqueued above to ensure concurrency with the search */
967 cl_error = clFlush(ls);
968 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
969 ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
972 /*! \brief Releases an OpenCL kernel pointer */
973 static void free_kernel(cl_kernel *kernel_ptr)
975 cl_int gmx_unused cl_error;
977 assert(nullptr != kernel_ptr);
981 cl_error = clReleaseKernel(*kernel_ptr);
982 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS, ("clReleaseKernel failed: " + ocl_get_error_string(cl_error)).c_str());
984 *kernel_ptr = nullptr;
988 /*! \brief Releases a list of OpenCL kernel pointers */
989 static void free_kernels(cl_kernel *kernels, int count)
993 for (i = 0; i < count; i++)
995 free_kernel(kernels + i);
999 /*! \brief Free the OpenCL runtime data (context and program).
1001 * The function releases the OpenCL context and program assuciated with the
1002 * device that the calling PP rank is running on.
1004 * \param runData [in] porinter to the structure with runtime data.
1006 static void free_gpu_device_runtime_data(gmx_device_runtime_data_t *runData)
1008 if (runData == nullptr)
1013 cl_int gmx_unused cl_error;
1015 if (runData->context)
1017 cl_error = clReleaseContext(runData->context);
1018 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS, ("clReleaseContext failed: " + ocl_get_error_string(cl_error)).c_str());
1019 runData->context = nullptr;
1022 if (runData->program)
1024 cl_error = clReleaseProgram(runData->program);
1025 GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS, ("clReleaseProgram failed: " + ocl_get_error_string(cl_error)).c_str());
1026 runData->program = nullptr;
1031 //! This function is documented in the header file
1032 void gpu_free(gmx_nbnxn_ocl_t *nb)
1040 int kernel_count = sizeof(nb->kernel_ener_noprune_ptr) / sizeof(nb->kernel_ener_noprune_ptr[0][0]);
1041 free_kernels(nb->kernel_ener_noprune_ptr[0], kernel_count);
1043 kernel_count = sizeof(nb->kernel_ener_prune_ptr) / sizeof(nb->kernel_ener_prune_ptr[0][0]);
1044 free_kernels(nb->kernel_ener_prune_ptr[0], kernel_count);
1046 kernel_count = sizeof(nb->kernel_noener_noprune_ptr) / sizeof(nb->kernel_noener_noprune_ptr[0][0]);
1047 free_kernels(nb->kernel_noener_noprune_ptr[0], kernel_count);
1049 kernel_count = sizeof(nb->kernel_noener_prune_ptr) / sizeof(nb->kernel_noener_prune_ptr[0][0]);
1050 free_kernels(nb->kernel_noener_prune_ptr[0], kernel_count);
1052 free_kernel(&(nb->kernel_zero_e_fshift));
1055 freeDeviceBuffer(&(nb->atdat->xq));
1056 freeDeviceBuffer(&(nb->atdat->f));
1057 freeDeviceBuffer(&(nb->atdat->e_lj));
1058 freeDeviceBuffer(&(nb->atdat->e_el));
1059 freeDeviceBuffer(&(nb->atdat->fshift));
1060 freeDeviceBuffer(&(nb->atdat->lj_comb));
1061 freeDeviceBuffer(&(nb->atdat->atom_types));
1062 freeDeviceBuffer(&(nb->atdat->shift_vec));
1066 freeDeviceBuffer(&(nb->nbparam->nbfp_climg2d));
1067 freeDeviceBuffer(&(nb->nbparam->nbfp_comb_climg2d));
1068 freeDeviceBuffer(&(nb->nbparam->coulomb_tab_climg2d));
1072 auto *plist = nb->plist[InteractionLocality::Local];
1073 freeDeviceBuffer(&plist->sci);
1074 freeDeviceBuffer(&plist->cj4);
1075 freeDeviceBuffer(&plist->imask);
1076 freeDeviceBuffer(&plist->excl);
1078 if (nb->bUseTwoStreams)
1080 auto *plist_nl = nb->plist[InteractionLocality::NonLocal];
1081 freeDeviceBuffer(&plist_nl->sci);
1082 freeDeviceBuffer(&plist_nl->cj4);
1083 freeDeviceBuffer(&plist_nl->imask);
1084 freeDeviceBuffer(&plist_nl->excl);
1089 pfree(nb->nbst.e_lj);
1090 nb->nbst.e_lj = nullptr;
1092 pfree(nb->nbst.e_el);
1093 nb->nbst.e_el = nullptr;
1095 pfree(nb->nbst.fshift);
1096 nb->nbst.fshift = nullptr;
1098 /* Free command queues */
1099 clReleaseCommandQueue(nb->stream[InteractionLocality::Local]);
1100 nb->stream[InteractionLocality::Local] = nullptr;
1101 if (nb->bUseTwoStreams)
1103 clReleaseCommandQueue(nb->stream[InteractionLocality::NonLocal]);
1104 nb->stream[InteractionLocality::NonLocal] = nullptr;
1106 /* Free other events */
1107 if (nb->nonlocal_done)
1109 clReleaseEvent(nb->nonlocal_done);
1110 nb->nonlocal_done = nullptr;
1112 if (nb->misc_ops_and_local_H2D_done)
1114 clReleaseEvent(nb->misc_ops_and_local_H2D_done);
1115 nb->misc_ops_and_local_H2D_done = nullptr;
1118 free_gpu_device_runtime_data(nb->dev_rundata);
1119 sfree(nb->dev_rundata);
1121 /* Free timers and timings */
1128 fprintf(debug, "Cleaned up OpenCL data structures.\n");
1132 //! This function is documented in the header file
1133 gmx_wallclock_gpu_nbnxn_t *gpu_get_timings(gmx_nbnxn_ocl_t *nb)
1135 return (nb != nullptr && nb->bDoTime) ? nb->timings : nullptr;
1138 //! This function is documented in the header file
1139 void gpu_reset_timings(nonbonded_verlet_t* nbv)
1141 if (nbv->gpu_nbv && nbv->gpu_nbv->bDoTime)
1143 init_timings(nbv->gpu_nbv->timings);
1147 //! This function is documented in the header file
1148 int gpu_min_ci_balanced(gmx_nbnxn_ocl_t *nb)
1150 return nb != nullptr ?
1151 gpu_min_ci_balanced_factor * nb->dev_info->compute_units : 0;
1154 //! This function is documented in the header file
1155 gmx_bool gpu_is_kernel_ewald_analytical(const gmx_nbnxn_ocl_t *nb)
1157 return ((nb->nbparam->eeltype == eelOclEWALD_ANA) ||
1158 (nb->nbparam->eeltype == eelOclEWALD_ANA_TWIN));
1161 } // namespace Nbnxm