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.h
38 * \author Anca Hamuraru <anca@streamcomputing.eu>
39 * \author Teemu Virolainen <teemu@streamcomputing.eu>
40 * \author Dimitrios Karkoulis <dimitris.karkoulis@gmail.com>
41 * \ingroup module_mdlib
54 #include "gromacs/gmxlib/ocl_tools/oclutils.h"
55 #include "gromacs/legacyheaders/types/force_flags.h"
56 #include "gromacs/legacyheaders/types/hw_info.h"
57 #include "gromacs/legacyheaders/types/simple.h"
58 #include "gromacs/mdlib/nb_verlet.h"
59 #include "gromacs/mdlib/nbnxn_consts.h"
60 #include "gromacs/mdlib/nbnxn_pairlist.h"
61 #include "gromacs/timing/gpu_timing.h"
64 #include "thread_mpi/atomic.h"
67 #include "gromacs/mdlib/nbnxn_gpu.h"
68 #include "gromacs/mdlib/nbnxn_gpu_data_mgmt.h"
69 #include "gromacs/pbcutil/ishift.h"
70 #include "gromacs/utility/cstringutil.h"
71 #include "gromacs/utility/fatalerror.h"
73 #include "nbnxn_ocl_types.h"
75 #if defined TEXOBJ_SUPPORTED && __CUDA_ARCH__ >= 300
79 /*! \brief Convenience defines */
81 #define NCL_PER_SUPERCL (NBNXN_GPU_NCLUSTER_PER_SUPERCLUSTER)
82 #define CL_SIZE (NBNXN_GPU_CLUSTER_SIZE)
85 /*! \brief Always/never run the energy/pruning kernels -- only for benchmarking purposes */
87 static bool always_ener = (getenv("GMX_GPU_ALWAYS_ENER") != NULL);
88 static bool never_ener = (getenv("GMX_GPU_NEVER_ENER") != NULL);
89 static bool always_prune = (getenv("GMX_GPU_ALWAYS_PRUNE") != NULL);
92 /* Uncomment this define to enable kernel debugging */
95 /*! \brief Specifies which kernel run to debug */
96 #define DEBUG_RUN_STEP 2
98 /*! \brief Validates the input global work size parameter.
100 static inline void validate_global_work_size(size_t *global_work_size, int work_dim, gmx_device_info_t *dinfo)
102 cl_uint device_size_t_size_bits;
103 cl_uint host_size_t_size_bits;
107 /* Each component of a global_work_size must not exceed the range given by the
108 sizeof(device size_t) for the device on which the kernel execution will
110 https://www.khronos.org/registry/cl/sdk/1.0/docs/man/xhtml/clEnqueueNDRangeKernel.html
112 device_size_t_size_bits = dinfo->adress_bits;
113 host_size_t_size_bits = (cl_uint)(sizeof(size_t) * 8);
115 /* If sizeof(host size_t) <= sizeof(device size_t)
116 => global_work_size components will always be valid
118 => get device limit for global work size and
119 compare it against each component of global_work_size.
121 if (host_size_t_size_bits > device_size_t_size_bits)
125 device_limit = (((size_t)1) << device_size_t_size_bits) - 1;
127 for (int i = 0; i < work_dim; i++)
129 if (global_work_size[i] > device_limit)
131 gmx_fatal(FARGS, "Watch out, the input system is too large to simulate!\n"
132 "The number of nonbonded work units (=number of super-clusters) exceeds the"
133 "device capabilities. Global work size limit exceeded (%d > %d)!",
134 global_work_size[i], device_limit);
140 /* Constant arrays listing non-bonded kernel function names. The arrays are
141 * organized in 2-dim arrays by: electrostatics and VDW type.
143 * Note that the row- and column-order of function pointers has to match the
144 * order of corresponding enumerated electrostatics and vdw types, resp.,
145 * defined in nbnxn_cuda_types.h.
148 /*! \brief Force-only kernel function names. */
149 static const char* nb_kfunc_noener_noprune_ptr[eelOclNR][evdwOclNR] =
151 { "nbnxn_kernel_ElecCut_VdwLJ_F_opencl", "nbnxn_kernel_ElecCut_VdwLJFsw_F_opencl", "nbnxn_kernel_ElecCut_VdwLJPsw_F_opencl", "nbnxn_kernel_ElecCut_VdwLJEwCombGeom_F_opencl", "nbnxn_kernel_ElecCut_VdwLJEwCombLB_F_opencl" },
152 { "nbnxn_kernel_ElecRF_VdwLJ_F_opencl", "nbnxn_kernel_ElecRF_VdwLJFsw_F_opencl", "nbnxn_kernel_ElecRF_VdwLJPsw_F_opencl", "nbnxn_kernel_ElecRF_VdwLJEwCombGeom_F_opencl", "nbnxn_kernel_ElecRF_VdwLJEwCombLB_F_opencl" },
153 { "nbnxn_kernel_ElecEwQSTab_VdwLJ_F_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJFsw_F_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJPsw_F_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_F_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_F_opencl" },
154 { "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_F_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_F_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_F_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_F_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_F_opencl" },
155 { "nbnxn_kernel_ElecEw_VdwLJ_F_opencl", "nbnxn_kernel_ElecEw_VdwLJFsw_F_opencl", "nbnxn_kernel_ElecEw_VdwLJPsw_F_opencl", "nbnxn_kernel_ElecEw_VdwLJEwCombGeom_F_opencl", "nbnxn_kernel_ElecEw_VdwLJEwCombLB_F_opencl" },
156 { "nbnxn_kernel_ElecEwTwinCut_VdwLJ_F_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_F_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_F_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_F_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_F_opencl" }
159 /*! \brief Force + energy kernel function pointers. */
160 static const char* nb_kfunc_ener_noprune_ptr[eelOclNR][evdwOclNR] =
162 { "nbnxn_kernel_ElecCut_VdwLJ_VF_opencl", "nbnxn_kernel_ElecCut_VdwLJFsw_VF_opencl", "nbnxn_kernel_ElecCut_VdwLJPsw_VF_opencl", "nbnxn_kernel_ElecCut_VdwLJEwCombGeom_VF_opencl", "nbnxn_kernel_ElecCut_VdwLJEwCombLB_VF_opencl" },
163 { "nbnxn_kernel_ElecRF_VdwLJ_VF_opencl", "nbnxn_kernel_ElecRF_VdwLJFsw_VF_opencl", "nbnxn_kernel_ElecRF_VdwLJPsw_VF_opencl", "nbnxn_kernel_ElecRF_VdwLJEwCombGeom_VF_opencl", "nbnxn_kernel_ElecRF_VdwLJEwCombLB_VF_opencl" },
164 { "nbnxn_kernel_ElecEwQSTab_VdwLJ_VF_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJFsw_VF_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJPsw_VF_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_VF_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_VF_opencl" },
165 { "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_VF_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_VF_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_VF_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_VF_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_VF_opencl" },
166 { "nbnxn_kernel_ElecEw_VdwLJ_VF_opencl", "nbnxn_kernel_ElecEw_VdwLJFsw_VF_opencl", "nbnxn_kernel_ElecEw_VdwLJPsw_VF_opencl", "nbnxn_kernel_ElecEw_VdwLJEwCombGeom_VF_opencl", "nbnxn_kernel_ElecEw_VdwLJEwCombLB_VF_opencl" },
167 { "nbnxn_kernel_ElecEwTwinCut_VdwLJ_VF_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_VF_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_VF_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_VF_opencl" }
170 /*! \brief Force + pruning kernel function pointers. */
171 static const char* nb_kfunc_noener_prune_ptr[eelOclNR][evdwOclNR] =
173 { "nbnxn_kernel_ElecCut_VdwLJ_F_prune_opencl", "nbnxn_kernel_ElecCut_VdwLJFsw_F_prune_opencl", "nbnxn_kernel_ElecCut_VdwLJPsw_F_prune_opencl", "nbnxn_kernel_ElecCut_VdwLJEwCombGeom_F_prune_opencl", "nbnxn_kernel_ElecCut_VdwLJEwCombLB_F_prune_opencl" },
174 { "nbnxn_kernel_ElecRF_VdwLJ_F_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJFsw_F_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJPsw_F_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJEwCombGeom_F_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJEwCombLB_F_prune_opencl" },
175 { "nbnxn_kernel_ElecEwQSTab_VdwLJ_F_prune_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJFsw_F_prune_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJPsw_F_prune_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_F_prune_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_F_prune_opencl" },
176 { "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_F_prune_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_F_prune_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_F_prune_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_F_prune_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_F_prune_opencl" },
177 { "nbnxn_kernel_ElecEw_VdwLJ_F_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJFsw_F_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJPsw_F_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJEwCombGeom_F_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJEwCombLB_F_prune_opencl" },
178 { "nbnxn_kernel_ElecEwTwinCut_VdwLJ_F_prune_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_F_prune_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_F_prune_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_F_prune_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_F_prune_opencl" }
181 /*! \brief Force + energy + pruning kernel function pointers. */
182 static const char* nb_kfunc_ener_prune_ptr[eelOclNR][evdwOclNR] =
184 { "nbnxn_kernel_ElecCut_VdwLJ_VF_prune_opencl", "nbnxn_kernel_ElecCut_VdwLJFsw_VF_prune_opencl", "nbnxn_kernel_ElecCut_VdwLJPsw_VF_prune_opencl", "nbnxn_kernel_ElecCut_VdwLJEwCombGeom_VF_prune_opencl", "nbnxn_kernel_ElecCut_VdwLJEwCombLB_VF_prune_opencl" },
185 { "nbnxn_kernel_ElecRF_VdwLJ_VF_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJFsw_VF_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJPsw_VF_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJEwCombGeom_VF_prune_opencl", "nbnxn_kernel_ElecRF_VdwLJEwCombLB_VF_prune_opencl" },
186 { "nbnxn_kernel_ElecEwQSTab_VdwLJ_VF_prune_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJFsw_VF_prune_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJPsw_VF_prune_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombGeom_VF_prune_opencl", "nbnxn_kernel_ElecEwQSTab_VdwLJEwCombLB_VF_prune_opencl" },
187 { "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJ_VF_prune_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJFsw_VF_prune_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJPsw_VF_prune_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombGeom_VF_prune_opencl", "nbnxn_kernel_ElecEwQSTabTwinCut_VdwLJEwCombLB_VF_prune_opencl" },
188 { "nbnxn_kernel_ElecEw_VdwLJ_VF_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJFsw_VF_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJPsw_VF_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJEwCombGeom_VF_prune_opencl", "nbnxn_kernel_ElecEw_VdwLJEwCombLB_VF_prune_opencl" },
189 { "nbnxn_kernel_ElecEwTwinCut_VdwLJ_VF_prune_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJFsw_VF_prune_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJPsw_VF_prune_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF_prune_opencl", "nbnxn_kernel_ElecEwTwinCut_VdwLJEwCombLB_VF_prune_opencl" }
192 /*! \brief Return a pointer to the kernel version to be executed at the current step.
193 * OpenCL kernel objects are cached in nb. If the requested kernel is not
194 * found in the cache, it will be created and the cache will be updated.
196 static inline cl_kernel select_nbnxn_kernel(gmx_nbnxn_ocl_t *nb,
202 const char* kernel_name_to_run;
203 cl_kernel *kernel_ptr;
206 assert(eeltype < eelOclNR);
207 assert(evdwtype < eelOclNR);
213 kernel_name_to_run = nb_kfunc_ener_prune_ptr[eeltype][evdwtype];
214 kernel_ptr = &(nb->kernel_ener_prune_ptr[eeltype][evdwtype]);
218 kernel_name_to_run = nb_kfunc_ener_noprune_ptr[eeltype][evdwtype];
219 kernel_ptr = &(nb->kernel_ener_noprune_ptr[eeltype][evdwtype]);
226 kernel_name_to_run = nb_kfunc_noener_prune_ptr[eeltype][evdwtype];
227 kernel_ptr = &(nb->kernel_noener_prune_ptr[eeltype][evdwtype]);
231 kernel_name_to_run = nb_kfunc_noener_noprune_ptr[eeltype][evdwtype];
232 kernel_ptr = &(nb->kernel_noener_noprune_ptr[eeltype][evdwtype]);
236 if (NULL == kernel_ptr[0])
238 *kernel_ptr = clCreateKernel(nb->dev_info->program, kernel_name_to_run, &cl_error);
239 assert(cl_error == CL_SUCCESS);
241 // TODO: handle errors
246 /*! \brief Calculates the amount of shared memory required by the OpenCL kernel in use.
248 static inline int calc_shmem_required()
252 /* size of shmem (force-buffers/xq/atom type preloading) */
253 /* NOTE: with the default kernel on sm3.0 we need shmem only for pre-loading */
254 /* i-atom x+q in shared memory */
255 //shmem = NCL_PER_SUPERCL * CL_SIZE * sizeof(float4);
256 shmem = NCL_PER_SUPERCL * CL_SIZE * sizeof(float) * 4; /* xqib */
257 /* cj in shared memory, for both warps separately */
258 shmem += 2 * NBNXN_GPU_JGROUP_SIZE * sizeof(int); /* cjs */
259 #ifdef IATYPE_SHMEM // CUDA ARCH >= 300
260 /* i-atom types in shared memory */
261 #pragma error "Should not be defined"
262 shmem += NCL_PER_SUPERCL * CL_SIZE * sizeof(int); /* atib */
264 /* force reduction buffers in shared memory */
265 shmem += CL_SIZE * CL_SIZE * 3 * sizeof(float); /* f_buf */
266 /* Warp vote. In fact it must be * number of warps in block.. */
267 shmem += sizeof(cl_uint) * 2; /* warp_any */
271 /*! \brief Initializes data structures that are going to be sent to the OpenCL device.
273 * The device can't use the same data structures as the host for two main reasons:
274 * - OpenCL restrictions (pointers are not accepted inside data structures)
275 * - some host side fields are not needed for the OpenCL kernels.
277 static void fillin_ocl_structures(cl_nbparam_t *nbp,
278 cl_nbparam_params_t *nbparams_params)
280 nbparams_params->coulomb_tab_scale = nbp->coulomb_tab_scale;
281 nbparams_params->coulomb_tab_size = nbp->coulomb_tab_size;
282 nbparams_params->c_rf = nbp->c_rf;
283 nbparams_params->dispersion_shift = nbp->dispersion_shift;
284 nbparams_params->eeltype = nbp->eeltype;
285 nbparams_params->epsfac = nbp->epsfac;
286 nbparams_params->ewaldcoeff_lj = nbp->ewaldcoeff_lj;
287 nbparams_params->ewald_beta = nbp->ewald_beta;
288 nbparams_params->rcoulomb_sq = nbp->rcoulomb_sq;
289 nbparams_params->repulsion_shift = nbp->repulsion_shift;
290 nbparams_params->rlist_sq = nbp->rlist_sq;
291 nbparams_params->rvdw_sq = nbp->rvdw_sq;
292 nbparams_params->rvdw_switch = nbp->rvdw_switch;
293 nbparams_params->sh_ewald = nbp->sh_ewald;
294 nbparams_params->sh_lj_ewald = nbp->sh_lj_ewald;
295 nbparams_params->two_k_rf = nbp->two_k_rf;
296 nbparams_params->vdwtype = nbp->vdwtype;
297 nbparams_params->vdw_switch = nbp->vdw_switch;
300 /*! \brief Waits for the commands associated with the input event to finish.
301 * Then it releases the event and sets it to 0.
302 * Don't use this function when more than one wait will be issued for the event.
304 void wait_ocl_event(cl_event *ocl_event)
306 cl_int gmx_unused cl_error;
308 /* Blocking wait for the event */
309 cl_error = clWaitForEvents(1, ocl_event);
310 assert(CL_SUCCESS == cl_error);
312 /* Release event and reset it to 0 */
313 cl_error = clReleaseEvent(*ocl_event);
314 assert(CL_SUCCESS == cl_error);
318 /*! \brief Enqueues a wait for event completion.
320 * Then it releases the event and sets it to 0.
321 * Don't use this function when more than one wait will be issued for the event.
322 * Equivalent to Cuda Stream Sync. */
323 void sync_ocl_event(cl_command_queue stream, cl_event *ocl_event)
325 cl_int gmx_unused cl_error;
328 #ifdef CL_VERSION_1_2
329 cl_error = clEnqueueBarrierWithWaitList(stream, 1, ocl_event, NULL);
331 cl_error = clEnqueueWaitForEvents(stream, 1, ocl_event);
334 assert(CL_SUCCESS == cl_error);
336 /* Release event and reset it to 0. It is ok to release it as enqueuewaitforevents performs implicit retain for events. */
337 cl_error = clReleaseEvent(*ocl_event);
338 assert(CL_SUCCESS == cl_error);
342 /*! \brief Returns the duration in miliseconds for the command associated with the event.
344 * It then releases the event and sets it to 0.
345 * Before calling this function, make sure the command has finished either by
346 * calling clFinish or clWaitForEvents.
347 * The function returns 0.0 if the input event, *ocl_event, is 0.
348 * Don't use this function when more than one wait will be issued for the event.
350 double ocl_event_elapsed_ms(cl_event *ocl_event)
352 cl_int gmx_unused cl_error;
353 cl_ulong start_ns, end_ns;
357 assert(NULL != ocl_event);
361 cl_error = clGetEventProfilingInfo(*ocl_event, CL_PROFILING_COMMAND_START,
362 sizeof(cl_ulong), &start_ns, NULL);
363 assert(CL_SUCCESS == cl_error);
365 cl_error = clGetEventProfilingInfo(*ocl_event, CL_PROFILING_COMMAND_END,
366 sizeof(cl_ulong), &end_ns, NULL);
367 assert(CL_SUCCESS == cl_error);
369 clReleaseEvent(*ocl_event);
372 elapsed_ms = (end_ns - start_ns) / 1000000.0;
378 /*! \brief Launch GPU kernel
380 As we execute nonbonded workload in separate queues, before launching
381 the kernel we need to make sure that he following operations have completed:
382 - atomdata allocation and related H2D transfers (every nstlist step);
383 - pair list H2D transfer (every nstlist step);
384 - shift vector H2D transfer (every nstlist step);
385 - force (+shift force and energy) output clearing (every step).
387 These operations are issued in the local queue at the beginning of the step
388 and therefore always complete before the local kernel launch. The non-local
389 kernel is launched after the local on the same device/context, so this is
390 inherently scheduled after the operations in the local stream (including the
392 However, for the sake of having a future-proof implementation, we use the
393 misc_ops_done event to record the point in time when the above operations
394 are finished and synchronize with this event in the non-local stream.
396 void nbnxn_gpu_launch_kernel(gmx_nbnxn_ocl_t *nb,
397 const struct nbnxn_atomdata_t *nbatom,
402 int adat_begin, adat_len; /* local/nonlocal offset and length used for xq and f */
403 /* OpenCL kernel launch-related stuff */
405 size_t local_work_size[3], global_work_size[3];
406 cl_kernel nb_kernel = NULL; /* fn pointer to the nonbonded kernel */
408 cl_atomdata_t *adat = nb->atdat;
409 cl_nbparam_t *nbp = nb->nbparam;
410 cl_plist_t *plist = nb->plist[iloc];
411 cl_timers_t *t = nb->timers;
412 cl_command_queue stream = nb->stream[iloc];
414 bool bCalcEner = flags & GMX_FORCE_ENERGY;
415 int bCalcFshift = flags & GMX_FORCE_VIRIAL;
416 bool bDoTime = nb->bDoTime;
419 cl_nbparam_params_t nbparams_params;
421 float * debug_buffer_h;
422 size_t debug_buffer_size;
425 /* turn energy calculation always on/off (for debugging/testing only) */
426 bCalcEner = (bCalcEner || always_ener) && !never_ener;
428 /* Don't launch the non-local kernel if there is no work to do.
429 Doing the same for the local kernel is more complicated, since the
430 local part of the force array also depends on the non-local kernel.
431 So to avoid complicating the code and to reduce the risk of bugs,
432 we always call the local kernel, the local x+q copy and later (not in
433 this function) the stream wait, local f copyback and the f buffer
434 clearing. All these operations, except for the local interaction kernel,
435 are needed for the non-local interactions. The skip of the local kernel
436 call is taken care of later in this function. */
437 if (iloc == eintNonlocal && plist->nsci == 0)
442 /* calculate the atom data index range based on locality */
446 adat_len = adat->natoms_local;
450 adat_begin = adat->natoms_local;
451 adat_len = adat->natoms - adat->natoms_local;
454 /* beginning of timed HtoD section */
457 ocl_copy_H2D_async(adat->xq, nbatom->x + adat_begin * 4, adat_begin*sizeof(float)*4,
458 adat_len * sizeof(float) * 4, stream, bDoTime ? (&(t->nb_h2d[iloc])) : NULL);
460 /* When we get here all misc operations issues in the local stream as well as
461 the local xq H2D are done,
462 so we record that in the local stream and wait for it in the nonlocal one. */
463 if (nb->bUseTwoStreams)
465 if (iloc == eintLocal)
467 #ifdef CL_VERSION_1_2
468 cl_error = clEnqueueMarkerWithWaitList(stream, 0, NULL, &(nb->misc_ops_and_local_H2D_done));
470 cl_error = clEnqueueMarker(stream, &(nb->misc_ops_and_local_H2D_done));
472 assert(CL_SUCCESS == cl_error);
476 sync_ocl_event(stream, &(nb->misc_ops_and_local_H2D_done));
480 if (plist->nsci == 0)
482 /* Don't launch an empty local kernel (is not allowed with OpenCL).
483 * TODO: Separate H2D and kernel launch into separate functions.
488 /* beginning of timed nonbonded calculation section */
490 /* get the pointer to the kernel flavor we need to use */
491 nb_kernel = select_nbnxn_kernel(nb,
495 plist->bDoPrune || always_prune);
497 /* kernel launch config */
498 local_work_size[0] = CL_SIZE;
499 local_work_size[1] = CL_SIZE;
500 local_work_size[2] = 1;
502 global_work_size[0] = plist->nsci * local_work_size[0];
503 global_work_size[1] = 1 * local_work_size[1];
504 global_work_size[2] = 1 * local_work_size[2];
506 validate_global_work_size(global_work_size, 3, nb->dev_info);
508 shmem = calc_shmem_required();
512 static int run_step = 1;
514 if (DEBUG_RUN_STEP == run_step)
516 debug_buffer_size = global_work_size[0] * global_work_size[1] * global_work_size[2] * sizeof(float);
517 debug_buffer_h = (float*)calloc(1, debug_buffer_size);
518 assert(NULL != debug_buffer_h);
520 if (NULL == nb->debug_buffer)
522 nb->debug_buffer = clCreateBuffer(nb->dev_info->context, CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
523 debug_buffer_size, debug_buffer_h, &cl_error);
525 assert(CL_SUCCESS == cl_error);
534 fprintf(debug, "GPU launch configuration:\n\tLocal work size: %dx%dx%d\n\t"
535 "Global work size : %dx%d\n\t#Super-clusters/clusters: %d/%d (%d)\n",
536 (int)(local_work_size[0]), (int)(local_work_size[1]), (int)(local_work_size[2]),
537 (int)(global_work_size[0]), (int)(global_work_size[1]), plist->nsci*NCL_PER_SUPERCL,
538 NCL_PER_SUPERCL, plist->na_c);
541 fillin_ocl_structures(nbp, &nbparams_params);
544 cl_error = clSetKernelArg(nb_kernel, arg_no++, sizeof(int), &(adat->ntypes));
545 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(nbparams_params), &(nbparams_params));
546 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(adat->xq));
547 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(adat->f));
548 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(adat->e_lj));
549 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(adat->e_el));
550 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(adat->fshift));
551 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(adat->atom_types));
552 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(adat->shift_vec));
553 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(nbp->nbfp_climg2d));
554 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(nbp->nbfp_comb_climg2d));
555 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(nbp->coulomb_tab_climg2d));
556 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(plist->sci));
557 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(plist->cj4));
558 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(plist->excl));
559 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(int), &bCalcFshift);
560 cl_error |= clSetKernelArg(nb_kernel, arg_no++, shmem, NULL);
561 cl_error |= clSetKernelArg(nb_kernel, arg_no++, sizeof(cl_mem), &(nb->debug_buffer));
563 assert(cl_error == CL_SUCCESS);
567 printf("ClERROR! %d\n", cl_error);
569 cl_error = clEnqueueNDRangeKernel(stream, nb_kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, bDoTime ? &(t->nb_k[iloc]) : NULL);
570 assert(cl_error == CL_SUCCESS);
574 static int run_step = 1;
576 if (DEBUG_RUN_STEP == run_step)
579 char file_name[256] = {0};
581 ocl_copy_D2H_async(debug_buffer_h, nb->debug_buffer, 0,
582 debug_buffer_size, stream, NULL);
584 // Make sure all data has been transfered back from device
587 printf("\nWriting debug_buffer to debug_buffer_ocl.txt...");
589 sprintf(file_name, "debug_buffer_ocl_%d.txt", DEBUG_RUN_STEP);
590 pf = fopen(file_name, "wt");
593 fprintf(pf, "%20s", "");
594 for (int j = 0; j < global_work_size[0]; j++)
597 sprintf(label, "(wIdx=%2d thIdx=%2d)", j / local_work_size[0], j % local_work_size[0]);
598 fprintf(pf, "%20s", label);
601 for (int i = 0; i < global_work_size[1]; i++)
604 sprintf(label, "(wIdy=%2d thIdy=%2d)", i / local_work_size[1], i % local_work_size[1]);
605 fprintf(pf, "\n%20s", label);
607 for (int j = 0; j < global_work_size[0]; j++)
609 fprintf(pf, "%20.5f", debug_buffer_h[i * global_work_size[0] + j]);
620 free(debug_buffer_h);
621 debug_buffer_h = NULL;
629 /*! \brief Debugging helper function */
630 void dump_compare_results_cj4(nbnxn_cj4_t* results, int cnt, char* out_file, char* ref_file)
634 pf = fopen(out_file, "wt");
637 fprintf(pf, "%20s%20s%20s%20s%20s%20s%20s%20s\n",
638 "cj[0]", "cj[1]", "cj[2]", "cj[3]",
639 "imei[0].excl_ind", "imei[0].imask",
640 "imei[1].excl_ind", "imei[1].imask");
642 for (int index = 0; index < cnt; index++)
644 fprintf(pf, "%20d%20d%20d%20d%20d%20u%20d%20u\n",
645 results[index].cj[0], results[index].cj[1], results[index].cj[2], results[index].cj[3],
646 results[index].imei[0].excl_ind, results[index].imei[0].imask,
647 results[index].imei[1].excl_ind, results[index].imei[1].imask);
652 printf("\nWrote results to %s", out_file);
654 pf = fopen(ref_file, "rt");
659 printf("\n%s file found. Comparing results...", ref_file);
661 /* Skip the first line */
665 if (1 != fscanf(pf, "%c", &c))
671 for (int index = 0; index < cnt; index++)
674 unsigned int u_ref_val;
676 for (int j = 0; j < 4; j++)
678 if (1 != fscanf(pf, "%20d", &ref_val))
683 if (ref_val != results[index].cj[j])
685 printf("\nDifference for cj[%d] at index %d computed value = %d reference value = %d",
686 j, index, results[index].cj[j], ref_val);
692 for (int j = 0; j < 2; j++)
694 if (1 != fscanf(pf, "%20d", &ref_val))
699 if (ref_val != results[index].imei[j].excl_ind)
701 printf("\nDifference for imei[%d].excl_ind at index %d computed value = %d reference value = %d",
702 j, index, results[index].imei[j].excl_ind, ref_val);
707 if (1 != fscanf(pf, "%20u", &u_ref_val))
712 if (u_ref_val != results[index].imei[j].imask)
714 printf("\nDifference for imei[%d].imask at index %d computed value = %u reference value = %u",
715 j, index, results[index].imei[j].imask, u_ref_val);
723 printf("\nFinished comparing results. Total number of differences: %d", diff);
728 printf("\n%s file not found. No comparison performed.", ref_file);
732 /*! \brief Debugging helper function */
733 void dump_compare_results_f(float* results, int cnt, char* out_file, char* ref_file)
736 float cmp_eps = 0.001f;
738 pf = fopen(out_file, "wt");
741 for (int index = 0; index < cnt; index++)
743 fprintf(pf, "%15.5f\n", results[index]);
748 printf("\nWrote results to %s", out_file);
750 pf = fopen(ref_file, "rt");
754 printf("\n%s file found. Comparing results...", ref_file);
755 for (int index = 0; index < cnt; index++)
758 if (1 != fscanf(pf, "%20f", &ref_val))
763 if (((ref_val - results[index]) > cmp_eps) ||
764 ((ref_val - results[index]) < -cmp_eps))
766 printf("\nDifference at index %d computed value = %15.5f reference value = %15.5f",
767 index, results[index], ref_val);
773 printf("\nFinished comparing results. Total number of differences: %d", diff);
778 printf("\n%s file not found. No comparison performed.", ref_file);
783 * Debug function for dumping cj4, f and fshift buffers.
784 * By default this function does nothing. To enable debugging for any of these
785 * buffers, uncomment the corresponding definition inside the function:
786 * DEBUG_DUMP_CJ4_OCL, DEBUG_DUMP_F_OCL, DEBUG_DUMP_FSHIFT_OCL.
789 void debug_dump_cj4_f_fshift(gmx_nbnxn_ocl_t gmx_unused *nb,
790 const struct nbnxn_atomdata_t gmx_unused *nbatom,
791 cl_command_queue gmx_unused stream,
792 int gmx_unused adat_begin,
793 int gmx_unused adat_len)
795 /* Uncomment this define to enable cj4 debugging for the first kernel run */
796 //#define DEBUG_DUMP_CJ4_OCL
797 #ifdef DEBUG_DUMP_CJ4_OCL
799 static int run_step = 1;
801 if (DEBUG_RUN_STEP == run_step)
803 nbnxn_cj4_t *temp_cj4;
806 char ocl_file_name[256] = {0};
807 char cuda_file_name[256] = {0};
809 cnt = nb->plist[0]->ncj4;
810 size = cnt * sizeof(nbnxn_cj4_t);
811 temp_cj4 = (nbnxn_cj4_t*)malloc(size);
813 ocl_copy_D2H_async(temp_cj4, nb->plist[0]->cj4, 0,
816 // Make sure all data has been transfered back from device
819 sprintf(ocl_file_name, "ocl_cj4_%d.txt", DEBUG_RUN_STEP);
820 sprintf(cuda_file_name, "cuda_cj4_%d.txt", DEBUG_RUN_STEP);
821 dump_compare_results_cj4(temp_cj4, cnt, ocl_file_name, cuda_file_name);
830 /* Uncomment this define to enable f debugging for the first kernel run */
831 //#define DEBUG_DUMP_F_OCL
832 #ifdef DEBUG_DUMP_F_OCL
834 static int run_step = 1;
836 if (DEBUG_RUN_STEP == run_step)
838 char ocl_file_name[256] = {0};
839 char cuda_file_name[256] = {0};
841 // Make sure all data has been transfered back from device
844 sprintf(ocl_file_name, "ocl_f_%d.txt", DEBUG_RUN_STEP);
845 sprintf(cuda_file_name, "cuda_f_%d.txt", DEBUG_RUN_STEP);
847 dump_compare_results_f(nbatom->out[0].f + adat_begin * 3, (adat_len) * 3,
848 ocl_file_name, cuda_file_name);
855 /* Uncomment this define to enable fshift debugging for the first kernel run */
856 //#define DEBUG_DUMP_FSHIFT_OCL
857 #ifdef DEBUG_DUMP_FSHIFT_OCL
859 static int run_step = 1;
861 if (DEBUG_RUN_STEP == run_step)
863 char ocl_file_name[256] = {0};
864 char cuda_file_name[256] = {0};
866 // Make sure all data has been transfered back from device
869 sprintf(ocl_file_name, "ocl_fshift_%d.txt", DEBUG_RUN_STEP);
870 sprintf(cuda_file_name, "cuda_fshift_%d.txt", DEBUG_RUN_STEP);
872 dump_compare_results_f((float*)(nb->nbst.fshift), SHIFTS * 3,
873 ocl_file_name, cuda_file_name);
882 * Launch asynchronously the download of nonbonded forces from the GPU
883 * (and energies/shift forces if required).
885 void nbnxn_gpu_launch_cpyback(gmx_nbnxn_ocl_t *nb,
886 const struct nbnxn_atomdata_t *nbatom,
890 cl_int gmx_unused cl_error;
891 int adat_begin, adat_len; /* local/nonlocal offset and length used for xq and f */
894 /* determine interaction locality from atom locality */
899 else if (NONLOCAL_A(aloc))
906 sprintf(stmp, "Invalid atom locality passed (%d); valid here is only "
907 "local (%d) or nonlocal (%d)", aloc, eatLocal, eatNonlocal);
912 cl_atomdata_t *adat = nb->atdat;
913 cl_timers_t *t = nb->timers;
914 bool bDoTime = nb->bDoTime;
915 cl_command_queue stream = nb->stream[iloc];
917 bool bCalcEner = flags & GMX_FORCE_ENERGY;
918 int bCalcFshift = flags & GMX_FORCE_VIRIAL;
921 /* don't launch non-local copy-back if there was no non-local work to do */
922 if (iloc == eintNonlocal && nb->plist[iloc]->nsci == 0)
927 /* calculate the atom data index range based on locality */
931 adat_len = adat->natoms_local;
935 adat_begin = adat->natoms_local;
936 adat_len = adat->natoms - adat->natoms_local;
939 /* beginning of timed D2H section */
941 /* With DD the local D2H transfer can only start after the non-local
942 has been launched. */
943 if (iloc == eintLocal && nb->bUseTwoStreams)
945 sync_ocl_event(stream, &(nb->nonlocal_done));
949 ocl_copy_D2H_async(nbatom->out[0].f + adat_begin * 3, adat->f, adat_begin*3*sizeof(float),
950 (adat_len)* adat->f_elem_size, stream, bDoTime ? &(t->nb_d2h_f[iloc]) : NULL);
952 /* After the non-local D2H is launched the nonlocal_done event can be
953 recorded which signals that the local D2H can proceed. This event is not
954 placed after the non-local kernel because we first need the non-local
956 if (iloc == eintNonlocal)
958 #ifdef CL_VERSION_1_2
959 cl_error = clEnqueueMarkerWithWaitList(stream, 0, NULL, &(nb->nonlocal_done));
961 cl_error = clEnqueueMarker(stream, &(nb->nonlocal_done));
963 assert(CL_SUCCESS == cl_error);
966 /* only transfer energies in the local stream */
972 ocl_copy_D2H_async(nb->nbst.fshift, adat->fshift, 0,
973 SHIFTS * adat->fshift_elem_size, stream, bDoTime ? &(t->nb_d2h_fshift[iloc]) : NULL);
979 ocl_copy_D2H_async(nb->nbst.e_lj, adat->e_lj, 0,
980 sizeof(float), stream, bDoTime ? &(t->nb_d2h_e_lj[iloc]) : NULL);
982 ocl_copy_D2H_async(nb->nbst.e_el, adat->e_el, 0,
983 sizeof(float), stream, bDoTime ? &(t->nb_d2h_e_el[iloc]) : NULL);
987 debug_dump_cj4_f_fshift(nb, nbatom, stream, adat_begin, adat_len);
991 * Wait for the asynchronously launched nonbonded calculations and data
992 * transfers to finish.
994 void nbnxn_gpu_wait_for_gpu(gmx_nbnxn_ocl_t *nb,
995 const nbnxn_atomdata_t gmx_unused *nbatom,
997 real *e_lj, real *e_el, rvec *fshift)
999 /* NOTE: only implemented for single-precision at this time */
1000 cl_int gmx_unused cl_error;
1003 /* determine interaction locality from atom locality */
1008 else if (NONLOCAL_A(aloc))
1010 iloc = eintNonlocal;
1015 sprintf(stmp, "Invalid atom locality passed (%d); valid here is only "
1016 "local (%d) or nonlocal (%d)", aloc, eatLocal, eatNonlocal);
1020 cl_plist_t *plist = nb->plist[iloc];
1021 cl_timers_t *timers = nb->timers;
1022 struct gmx_wallclock_gpu_t *timings = nb->timings;
1023 cl_nb_staging nbst = nb->nbst;
1025 bool bCalcEner = flags & GMX_FORCE_ENERGY;
1026 int bCalcFshift = flags & GMX_FORCE_VIRIAL;
1028 /* turn energy calculation always on/off (for debugging/testing only) */
1029 bCalcEner = (bCalcEner || always_ener) && !never_ener;
1031 /* Launch wait/update timers & counters, unless doing the non-local phase
1032 when there is not actually work to do. This is consistent with
1033 nbnxn_gpu_launch_kernel.
1035 NOTE: if timing with multiple GPUs (streams) becomes possible, the
1036 counters could end up being inconsistent due to not being incremented
1037 on some of the nodes! */
1038 if (iloc == eintNonlocal && nb->plist[iloc]->nsci == 0)
1043 /* Actual sync point. Waits for everything to be finished in the command queue. TODO: Find out if a more fine grained solution is needed */
1044 cl_error = clFinish(nb->stream[iloc]);
1045 assert(CL_SUCCESS == cl_error);
1047 /* timing data accumulation */
1050 /* only increase counter once (at local F wait) */
1054 timings->ktime[plist->bDoPrune ? 1 : 0][bCalcEner ? 1 : 0].c += 1;
1057 /* kernel timings */
1059 timings->ktime[plist->bDoPrune ? 1 : 0][bCalcEner ? 1 : 0].t +=
1060 ocl_event_elapsed_ms(timers->nb_k + iloc);
1062 /* X/q H2D and F D2H timings */
1063 timings->nb_h2d_t += ocl_event_elapsed_ms(timers->nb_h2d + iloc);
1064 timings->nb_d2h_t += ocl_event_elapsed_ms(timers->nb_d2h_f + iloc);
1065 timings->nb_d2h_t += ocl_event_elapsed_ms(timers->nb_d2h_fshift + iloc);
1066 timings->nb_d2h_t += ocl_event_elapsed_ms(timers->nb_d2h_e_el + iloc);
1067 timings->nb_d2h_t += ocl_event_elapsed_ms(timers->nb_d2h_e_lj + iloc);
1069 /* only count atdat and pair-list H2D at pair-search step */
1070 if (plist->bDoPrune)
1072 /* atdat transfer timing (add only once, at local F wait) */
1075 timings->pl_h2d_c++;
1076 timings->pl_h2d_t += ocl_event_elapsed_ms(&(timers->atdat));
1079 timings->pl_h2d_t +=
1080 ocl_event_elapsed_ms(timers->pl_h2d_sci + iloc) +
1081 ocl_event_elapsed_ms(timers->pl_h2d_cj4 + iloc) +
1082 ocl_event_elapsed_ms(timers->pl_h2d_excl + iloc);
1087 /* add up energies and shift forces (only once at local F wait) */
1092 *e_lj += *nbst.e_lj;
1093 *e_el += *nbst.e_el;
1098 for (i = 0; i < SHIFTS; i++)
1100 fshift[i][0] += (nbst.fshift)[i][0];
1101 fshift[i][1] += (nbst.fshift)[i][1];
1102 fshift[i][2] += (nbst.fshift)[i][2];
1107 /* turn off pruning (doesn't matter if this is pair-search step or not) */
1108 plist->bDoPrune = false;
1112 /*! \brief Selects the Ewald kernel type, analytical or tabulated, single or twin cut-off. */
1113 int nbnxn_gpu_pick_ewald_kernel_type(bool bTwinCut)
1115 bool bUseAnalyticalEwald, bForceAnalyticalEwald, bForceTabulatedEwald;
1118 /* Benchmarking/development environment variables to force the use of
1119 analytical or tabulated Ewald kernel. */
1120 bForceAnalyticalEwald = (getenv("GMX_OCL_NB_ANA_EWALD") != NULL);
1121 bForceTabulatedEwald = (getenv("GMX_OCL_NB_TAB_EWALD") != NULL);
1123 if (bForceAnalyticalEwald && bForceTabulatedEwald)
1125 gmx_incons("Both analytical and tabulated Ewald OpenCL non-bonded kernels "
1126 "requested through environment variables.");
1129 /* CUDA: By default, on SM 3.0 and later use analytical Ewald, on earlier tabulated. */
1130 /* OpenCL: By default, use analytical Ewald, on earlier tabulated. */
1131 // TODO: decide if dev_info parameter should be added to recognize NVIDIA CC>=3.0 devices.
1132 //if ((dev_info->prop.major >= 3 || bForceAnalyticalEwald) && !bForceTabulatedEwald)
1133 if ((1 || bForceAnalyticalEwald) && !bForceTabulatedEwald)
1135 bUseAnalyticalEwald = true;
1139 fprintf(debug, "Using analytical Ewald OpenCL kernels\n");
1144 bUseAnalyticalEwald = false;
1148 fprintf(debug, "Using tabulated Ewald OpenCL kernels\n");
1152 /* Use twin cut-off kernels if requested by bTwinCut or the env. var.
1153 forces it (use it for debugging/benchmarking only). */
1154 if (!bTwinCut && (getenv("GMX_OCL_NB_EWALD_TWINCUT") == NULL))
1156 kernel_type = bUseAnalyticalEwald ? eelOclEWALD_ANA : eelOclEWALD_TAB;
1160 kernel_type = bUseAnalyticalEwald ? eelOclEWALD_ANA_TWIN : eelOclEWALD_TAB_TWIN;