2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2016,2017 by the GROMACS development team.
5 * Copyright (c) 2018,2019,2020, by the GROMACS development team, led by
6 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
7 * and including many others, as listed in the AUTHORS file in the
8 * top-level source directory and at http://www.gromacs.org.
10 * GROMACS is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public License
12 * as published by the Free Software Foundation; either version 2.1
13 * of the License, or (at your option) any later version.
15 * GROMACS is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with GROMACS; if not, see
22 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
23 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
25 * If you want to redistribute modifications to GROMACS, please
26 * consider that scientific software is very special. Version
27 * control is crucial - bugs must be traceable. We will be happy to
28 * consider code for inclusion in the official distribution, but
29 * derived work must not be called official GROMACS. Details are found
30 * in the README & COPYING files - if they are missing, get the
31 * official version at http://www.gromacs.org.
33 * To help us fund GROMACS development, we humbly ask that you cite
34 * the research papers on the package. Check out http://www.gromacs.org.
38 * Utility constant and function declaration for the OpenCL non-bonded kernels.
39 * This header should be included once at the top level, just before the
40 * kernels are included (has to be preceded by nbnxn_ocl_types.h).
42 * \author Szilárd Páll <pall.szilard@gmail.com>
43 * \ingroup module_nbnxm
48 #include "gromacs/gpu_utils/device_utils.clh"
49 #include "gromacs/gpu_utils/vectype_ops.clh"
50 #include "gromacs/nbnxm/constants.h"
51 #include "gromacs/pbcutil/ishift.h"
53 #include "nbnxm_ocl_consts.h"
55 #define CL_SIZE (NBNXN_GPU_CLUSTER_SIZE)
56 #define NCL_PER_SUPERCL c_nbnxnGpuNumClusterPerSupercluster
58 #define WARP_SIZE (CL_SIZE * CL_SIZE / 2) // Currently only c_nbnxnGpuClusterpairSplit=2 supported
60 #if defined _NVIDIA_SOURCE_ || defined _AMD_SOURCE_
61 /* Currently we enable CJ prefetch for AMD/NVIDIA and disable it for other vendors
62 * Note that this should precede the kernel_utils include.
64 # define USE_CJ_PREFETCH 1
66 # define USE_CJ_PREFETCH 0
69 #if defined cl_intel_subgroups || defined cl_khr_subgroups \
70 || (defined __OPENCL_VERSION__ && __OPENCL_VERSION__ >= 210)
71 # define HAVE_SUBGROUP 1
73 # define HAVE_SUBGROUP 0
76 #ifdef cl_intel_subgroups
77 # define HAVE_INTEL_SUBGROUP 1
79 # define HAVE_INTEL_SUBGROUP 0
82 #if defined _INTEL_SOURCE_
83 # define SUBGROUP_SIZE 8
84 #elif defined _AMD_SOURCE_
85 # define SUBGROUP_SIZE 64
87 # define SUBGROUP_SIZE 32
90 #define REDUCE_SHUFFLE (HAVE_INTEL_SUBGROUP && CL_SIZE == 4 && SUBGROUP_SIZE == WARP_SIZE)
91 #define USE_SUBGROUP_ANY (HAVE_SUBGROUP && SUBGROUP_SIZE == WARP_SIZE)
92 #define USE_SUBGROUP_PRELOAD HAVE_INTEL_SUBGROUP
94 /* 1.0 / sqrt(M_PI) */
95 #define M_FLOAT_1_SQRTPI 0.564189583547756F
99 #ifndef NBNXN_OPENCL_KERNEL_UTILS_CLH
100 # define NBNXN_OPENCL_KERNEL_UTILS_CLH
103 # define WARP_SIZE_LOG2 (5)
104 # define CL_SIZE_LOG2 (3)
106 # define WARP_SIZE_LOG2 (3)
107 # define CL_SIZE_LOG2 (2)
109 # error unsupported CL_SIZE
112 # define CL_SIZE_SQ (CL_SIZE * CL_SIZE)
113 # define FBUF_STRIDE (CL_SIZE_SQ)
115 # define ONE_SIXTH_F 0.16666667F
116 # define ONE_TWELVETH_F 0.08333333F
121 /* GCC, clang, and some ICC pretending to be GCC */
122 # define gmx_unused __attribute__((unused))
127 // Data structures shared between OpenCL device code and OpenCL host code
128 // TODO: review, improve
129 // Replaced real by float for now, to avoid including any other header
137 /* Used with potential switching:
138 * rsw = max(r - r_switch, 0)
139 * sw = 1 + c3*rsw^3 + c4*rsw^4 + c5*rsw^5
140 * dsw = 3*c3*rsw^2 + 4*c4*rsw^3 + 5*c5*rsw^4
141 * force = force*dsw - potential*sw
151 // Data structure shared between the OpenCL device code and OpenCL host code
152 // Must not contain OpenCL objects (buffers)
153 typedef struct cl_nbparam_params
156 //! type of electrostatics, takes values from #eelCu
158 //! type of VdW impl., takes values from #evdwCu
161 //! charge multiplication factor
163 //! Reaction-field/plain cutoff electrostatics const.
165 //! Reaction-field electrostatics constant
167 //! Ewald/PME parameter
169 //! Ewald/PME correction term substracted from the direct-space potential
171 //! LJ-Ewald/PME correction term added to the correction potential
173 //! LJ-Ewald/PME coefficient
176 //! Coulomb cut-off squared
179 //! VdW cut-off squared
181 //! VdW switched cut-off
183 //! Full, outer pair-list cut-off squared
185 //! Inner, dynamic pruned pair-list cut-off squared XXX: this is only needed in the pruning kernels, but for now we also pass it to the nonbondeds
188 //! VdW shift dispersion constants
189 shift_consts_t dispersion_shift;
190 //! VdW shift repulsion constants
191 shift_consts_t repulsion_shift;
192 //! VdW switch constants
193 switch_consts_t vdw_switch;
195 /* Ewald Coulomb force table data - accessed through texture memory */
196 //! table scale/spacing
197 float coulomb_tab_scale;
198 } cl_nbparam_params_t;
204 //! Shift vector index plus possible flags
206 //! Start index into cj4
208 //! End index into cj4
214 //! The i-cluster interactions mask for 1 warp
216 //! Index into the exclusion array for 1 warp
224 //! The i-cluster mask data for 2 warps
225 nbnxn_im_ei_t imei[2];
231 unsigned int pair[CL_SIZE * CL_SIZE / 2]; /* Topology exclusion interaction bits for one warp,
232 * each unsigned has bitS for 4*8 i clusters
236 /*! i-cluster interaction mask for a super-cluster with all NCL_PER_SUPERCL bits set */
237 __constant unsigned supercl_interaction_mask = ((1U << NCL_PER_SUPERCL) - 1U);
239 gmx_opencl_inline void preloadCj4Generic(__local int* sm_cjPreload,
240 const __global int* gm_cj,
243 bool gmx_unused iMaskCond)
245 /* Pre-load cj into shared memory */
246 # if defined _AMD_SOURCE_ // TODO: fix by setting c_nbnxnGpuClusterpairSplit properly
247 if (tidxj == 0 & tidxi < c_nbnxnGpuJgroupSize)
249 sm_cjPreload[tidxi] = gm_cj[tidxi];
252 const int c_clSize = CL_SIZE;
253 const int c_nbnxnGpuClusterpairSplit = 2;
254 const int c_splitClSize = c_clSize / c_nbnxnGpuClusterpairSplit;
255 if ((tidxj == 0 | tidxj == c_splitClSize) & (tidxi < c_nbnxnGpuJgroupSize))
257 sm_cjPreload[tidxi + tidxj * c_nbnxnGpuJgroupSize / c_splitClSize] = gm_cj[tidxi];
263 # if USE_SUBGROUP_PRELOAD
264 gmx_opencl_inline int preloadCj4Subgroup(const __global int* gm_cj)
266 // loads subgroup-size # of elements (8) instead of the 4 required
267 // equivalent to *cjs = *gm_cj
268 return intel_sub_group_block_read((const __global uint*)gm_cj);
270 # endif // USE_SUBGROUP_PRELOAD
272 # if USE_SUBGROUP_PRELOAD
273 typedef size_t CjType;
275 typedef __local int* CjType;
278 /*! \brief Preload cj4
280 * - For AMD we load once for a wavefront of 64 threads (on 4 threads * NTHREAD_Z)
281 * - For NVIDIA once per warp (on 2x4 threads * NTHREAD_Z)
282 * - For Intel(/USE_SUBGROUP_PRELOAD) loads into private memory(/register) instead of local memory
284 * It is the caller's responsibility to make sure that data is consumed only when
285 * it's ready. This function does not call a barrier.
287 gmx_opencl_inline void preloadCj4(CjType gmx_unused* cjs,
288 const __global int gmx_unused* gm_cj,
289 int gmx_unused tidxi,
290 int gmx_unused tidxj,
291 bool gmx_unused iMaskCond)
293 # if USE_SUBGROUP_PRELOAD
294 *cjs = preloadCj4Subgroup(gm_cj);
295 # elif USE_CJ_PREFETCH
296 preloadCj4Generic(*cjs, gm_cj, tidxi, tidxj, iMaskCond);
302 gmx_opencl_inline int loadCjPreload(__local int* sm_cjPreload, int jm, int gmx_unused tidxi, int gmx_unused tidxj)
304 # if defined _AMD_SOURCE_
305 int warpLoadOffset = 0; // TODO: fix by setting c_nbnxnGpuClusterpairSplit properly
307 const int c_clSize = CL_SIZE;
308 const int c_nbnxnGpuClusterpairSplit = 2;
309 const int c_splitClSize = c_clSize / c_nbnxnGpuClusterpairSplit;
310 int warpLoadOffset = (tidxj & c_splitClSize) * c_nbnxnGpuJgroupSize / c_splitClSize;
312 return sm_cjPreload[jm + warpLoadOffset];
315 /* \brief Load a cj given a jm index.
317 * If cj4 preloading is enabled, it loads from the local memory, otherwise from global.
319 gmx_opencl_inline int
320 loadCj(CjType cjs, const __global int gmx_unused* gm_cj, int jm, int gmx_unused tidxi, int gmx_unused tidxj)
322 # if USE_SUBGROUP_PRELOAD
323 return sub_group_broadcast(cjs, jm);
324 # elif USE_CJ_PREFETCH
325 return loadCjPreload(cjs, jm, tidxi, tidxj);
331 /*! Convert LJ sigma,epsilon parameters to C6,C12. */
332 gmx_opencl_inline void convert_sigma_epsilon_to_c6_c12(const float sigma, const float epsilon, float* c6, float* c12)
334 float sigma2, sigma6;
336 sigma2 = sigma * sigma;
337 sigma6 = sigma2 * sigma2 * sigma2;
338 *c6 = epsilon * sigma6;
343 /*! Apply force switch, force + energy version. */
344 gmx_opencl_inline void calculate_force_switch_F(const cl_nbparam_params_t* nbparam,
353 /* force switch constants */
354 float disp_shift_V2 = nbparam->dispersion_shift.c2;
355 float disp_shift_V3 = nbparam->dispersion_shift.c3;
356 float repu_shift_V2 = nbparam->repulsion_shift.c2;
357 float repu_shift_V3 = nbparam->repulsion_shift.c3;
360 r_switch = r - nbparam->rvdw_switch;
361 r_switch = r_switch >= 0.0F ? r_switch : 0.0F;
363 *F_invr += -c6 * (disp_shift_V2 + disp_shift_V3 * r_switch) * r_switch * r_switch * inv_r
364 + c12 * (repu_shift_V2 + repu_shift_V3 * r_switch) * r_switch * r_switch * inv_r;
367 /*! Apply force switch, force-only version. */
368 gmx_opencl_inline void calculate_force_switch_F_E(const cl_nbparam_params_t* nbparam,
378 /* force switch constants */
379 float disp_shift_V2 = nbparam->dispersion_shift.c2;
380 float disp_shift_V3 = nbparam->dispersion_shift.c3;
381 float repu_shift_V2 = nbparam->repulsion_shift.c2;
382 float repu_shift_V3 = nbparam->repulsion_shift.c3;
384 float disp_shift_F2 = nbparam->dispersion_shift.c2 / 3;
385 float disp_shift_F3 = nbparam->dispersion_shift.c3 / 4;
386 float repu_shift_F2 = nbparam->repulsion_shift.c2 / 3;
387 float repu_shift_F3 = nbparam->repulsion_shift.c3 / 4;
390 r_switch = r - nbparam->rvdw_switch;
391 r_switch = r_switch >= 0.0F ? r_switch : 0.0F;
393 *F_invr += -c6 * (disp_shift_V2 + disp_shift_V3 * r_switch) * r_switch * r_switch * inv_r
394 + c12 * (repu_shift_V2 + repu_shift_V3 * r_switch) * r_switch * r_switch * inv_r;
395 *E_lj += c6 * (disp_shift_F2 + disp_shift_F3 * r_switch) * r_switch * r_switch * r_switch
396 - c12 * (repu_shift_F2 + repu_shift_F3 * r_switch) * r_switch * r_switch * r_switch;
399 /*! Apply potential switch, force-only version. */
400 gmx_opencl_inline void calculate_potential_switch_F(const cl_nbparam_params_t* nbparam,
409 /* potential switch constants */
410 float switch_V3 = nbparam->vdw_switch.c3;
411 float switch_V4 = nbparam->vdw_switch.c4;
412 float switch_V5 = nbparam->vdw_switch.c5;
413 float switch_F2 = nbparam->vdw_switch.c3;
414 float switch_F3 = nbparam->vdw_switch.c4;
415 float switch_F4 = nbparam->vdw_switch.c5;
418 r_switch = r - nbparam->rvdw_switch;
420 /* Unlike in the F+E kernel, conditional is faster here */
423 sw = 1.0F + (switch_V3 + (switch_V4 + switch_V5 * r_switch) * r_switch) * r_switch * r_switch * r_switch;
424 dsw = (switch_F2 + (switch_F3 + switch_F4 * r_switch) * r_switch) * r_switch * r_switch;
426 *F_invr = (*F_invr) * sw - inv_r * (*E_lj) * dsw;
430 /*! Apply potential switch, force + energy version. */
431 gmx_opencl_inline void calculate_potential_switch_F_E(const cl_nbparam_params_t* nbparam,
440 /* potential switch constants */
441 float switch_V3 = nbparam->vdw_switch.c3;
442 float switch_V4 = nbparam->vdw_switch.c4;
443 float switch_V5 = nbparam->vdw_switch.c5;
444 float switch_F2 = nbparam->vdw_switch.c3;
445 float switch_F3 = nbparam->vdw_switch.c4;
446 float switch_F4 = nbparam->vdw_switch.c5;
449 r_switch = r - nbparam->rvdw_switch;
450 r_switch = r_switch >= 0.0F ? r_switch : 0.0F;
452 /* Unlike in the F-only kernel, masking is faster here */
453 sw = 1.0F + (switch_V3 + (switch_V4 + switch_V5 * r_switch) * r_switch) * r_switch * r_switch * r_switch;
454 dsw = (switch_F2 + (switch_F3 + switch_F4 * r_switch) * r_switch) * r_switch * r_switch;
456 *F_invr = (*F_invr) * sw - inv_r * (*E_lj) * dsw;
460 /*! Calculate LJ-PME grid force contribution with
461 * geometric combination rule.
463 gmx_opencl_inline void calculate_lj_ewald_comb_geom_F(__constant const float* nbfp_comb_climg2d,
472 float c6grid, inv_r6_nm, cr2, expmcr2, poly;
474 c6grid = nbfp_comb_climg2d[2 * typei] * nbfp_comb_climg2d[2 * typej];
476 /* Recalculate inv_r6 without exclusion mask */
477 inv_r6_nm = inv_r2 * inv_r2 * inv_r2;
478 cr2 = lje_coeff2 * r2;
480 poly = 1.0F + cr2 + HALF_F * cr2 * cr2;
482 /* Subtract the grid force from the total LJ force */
483 *F_invr += c6grid * (inv_r6_nm - expmcr2 * (inv_r6_nm * poly + lje_coeff6_6)) * inv_r2;
486 /*! Calculate LJ-PME grid force + energy contribution with
487 * geometric combination rule.
489 gmx_opencl_inline void calculate_lj_ewald_comb_geom_F_E(__constant const float* nbfp_comb_climg2d,
490 const cl_nbparam_params_t* nbparam,
501 float c6grid, inv_r6_nm, cr2, expmcr2, poly, sh_mask;
503 c6grid = nbfp_comb_climg2d[2 * typei] * nbfp_comb_climg2d[2 * typej];
505 /* Recalculate inv_r6 without exclusion mask */
506 inv_r6_nm = inv_r2 * inv_r2 * inv_r2;
507 cr2 = lje_coeff2 * r2;
509 poly = 1.0F + cr2 + HALF_F * cr2 * cr2;
511 /* Subtract the grid force from the total LJ force */
512 *F_invr += c6grid * (inv_r6_nm - expmcr2 * (inv_r6_nm * poly + lje_coeff6_6)) * inv_r2;
514 /* Shift should be applied only to real LJ pairs */
515 sh_mask = nbparam->sh_lj_ewald * int_bit;
516 *E_lj += ONE_SIXTH_F * c6grid * (inv_r6_nm * (1.0F - expmcr2 * poly) + sh_mask);
519 /*! Calculate LJ-PME grid force + energy contribution (if E_lj != NULL) with
520 * Lorentz-Berthelot combination rule.
521 * We use a single F+E kernel with conditional because the performance impact
522 * of this is pretty small and LB on the CPU is anyway very slow.
524 gmx_opencl_inline void calculate_lj_ewald_comb_LB_F_E(__constant const float* nbfp_comb_climg2d,
525 const cl_nbparam_params_t* nbparam,
537 float c6grid, inv_r6_nm, cr2, expmcr2, poly;
538 float sigma, sigma2, epsilon;
540 /* sigma and epsilon are scaled to give 6*C6 */
541 sigma = nbfp_comb_climg2d[2 * typei] + nbfp_comb_climg2d[2 * typej];
543 epsilon = nbfp_comb_climg2d[2 * typei + 1] * nbfp_comb_climg2d[2 * typej + 1];
545 sigma2 = sigma * sigma;
546 c6grid = epsilon * sigma2 * sigma2 * sigma2;
548 /* Recalculate inv_r6 without exclusion mask */
549 inv_r6_nm = inv_r2 * inv_r2 * inv_r2;
550 cr2 = lje_coeff2 * r2;
552 poly = 1.0F + cr2 + HALF_F * cr2 * cr2;
554 /* Subtract the grid force from the total LJ force */
555 *F_invr += c6grid * (inv_r6_nm - expmcr2 * (inv_r6_nm * poly + lje_coeff6_6)) * inv_r2;
561 /* Shift should be applied only to real LJ pairs */
562 sh_mask = nbparam->sh_lj_ewald * int_bit;
563 *E_lj += ONE_SIXTH_F * c6grid * (inv_r6_nm * (1.0F - expmcr2 * poly) + sh_mask);
567 /*! Interpolate Ewald coulomb force using the table through the tex_nbfp texture.
568 * Original idea: from the OpenMM project
570 gmx_opencl_inline float interpolate_coulomb_force_r(__constant const float* coulomb_tab_climg2d,
574 float normalized = scale * r;
575 int index = (int)normalized;
576 float fract2 = normalized - (float)index;
577 float fract1 = 1.0F - fract2;
579 return fract1 * coulomb_tab_climg2d[index] + fract2 * coulomb_tab_climg2d[index + 1];
582 /*! Calculate analytical Ewald correction term. */
583 gmx_opencl_inline float pmecorrF(float z2)
585 const float FN6 = -1.7357322914161492954e-8F;
586 const float FN5 = 1.4703624142580877519e-6F;
587 const float FN4 = -0.000053401640219807709149F;
588 const float FN3 = 0.0010054721316683106153F;
589 const float FN2 = -0.019278317264888380590F;
590 const float FN1 = 0.069670166153766424023F;
591 const float FN0 = -0.75225204789749321333F;
593 const float FD4 = 0.0011193462567257629232F;
594 const float FD3 = 0.014866955030185295499F;
595 const float FD2 = 0.11583842382862377919F;
596 const float FD1 = 0.50736591960530292870F;
597 const float FD0 = 1.0F;
600 float polyFN0, polyFN1, polyFD0, polyFD1;
604 polyFD0 = FD4 * z4 + FD2;
605 polyFD1 = FD3 * z4 + FD1;
606 polyFD0 = polyFD0 * z4 + FD0;
607 polyFD0 = polyFD1 * z2 + polyFD0;
609 polyFD0 = 1.0F / polyFD0;
611 polyFN0 = FN6 * z4 + FN4;
612 polyFN1 = FN5 * z4 + FN3;
613 polyFN0 = polyFN0 * z4 + FN2;
614 polyFN1 = polyFN1 * z4 + FN1;
615 polyFN0 = polyFN0 * z4 + FN0;
616 polyFN0 = polyFN1 * z2 + polyFN0;
618 return polyFN0 * polyFD0;
622 gmx_opencl_inline void
623 reduce_force_j_shfl(float3 fin, __global float* fout, int gmx_unused tidxi, int gmx_unused tidxj, int aidx)
625 /* Only does reduction over 4 elements in cluster. Needs to be changed
626 * for CL_SIZE>4. See CUDA code for required code */
627 fin.x += intel_sub_group_shuffle_down(fin.x, fin.x, 1);
628 fin.y += intel_sub_group_shuffle_up(fin.y, fin.y, 1);
629 fin.z += intel_sub_group_shuffle_down(fin.z, fin.z, 1);
630 if ((tidxi & 1) == 1)
634 fin.x += intel_sub_group_shuffle_down(fin.x, fin.x, 2);
635 fin.z += intel_sub_group_shuffle_up(fin.z, fin.z, 2);
642 atomicAdd_g_f(&fout[3 * aidx + tidxi], fin.x);
647 gmx_opencl_inline void
648 reduce_force_j_generic(__local float* f_buf, float3 fcj_buf, __global float* fout, int tidxi, int tidxj, int aidx)
650 int tidx = tidxi + tidxj * CL_SIZE;
651 f_buf[tidx] = fcj_buf.x;
652 f_buf[FBUF_STRIDE + tidx] = fcj_buf.y;
653 f_buf[2 * FBUF_STRIDE + tidx] = fcj_buf.z;
655 /* Split the reduction between the first 3 column threads
656 Threads with column id 0 will do the reduction for (float3).x components
657 Threads with column id 1 will do the reduction for (float3).y components
658 Threads with column id 2 will do the reduction for (float3).z components.
659 The reduction is performed for each line tidxj of f_buf. */
663 for (int j = tidxj * CL_SIZE; j < (tidxj + 1) * CL_SIZE; j++)
665 f += f_buf[FBUF_STRIDE * tidxi + j];
668 atomicAdd_g_f(&fout[3 * aidx + tidxi], f);
672 /*! Final j-force reduction
674 gmx_opencl_inline void reduce_force_j(__local float gmx_unused* f_buf,
676 __global float* fout,
682 reduce_force_j_shfl(fcj_buf, fout, tidxi, tidxj, aidx);
684 reduce_force_j_generic(f_buf, fcj_buf, fout, tidxi, tidxj, aidx);
689 gmx_opencl_inline void reduce_force_i_and_shift_shfl(float3* fci_buf,
690 __global float* fout,
696 __global float* fshift)
698 /* Only does reduction over 4 elements in cluster (2 per warp). Needs to be changed
700 float2 fshift_buf = 0;
701 for (int ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
703 int aidx = (sci * NCL_PER_SUPERCL + ci_offset) * CL_SIZE + tidxi;
704 float3 fin = fci_buf[ci_offset];
705 fin.x += intel_sub_group_shuffle_down(fin.x, fin.x, CL_SIZE);
706 fin.y += intel_sub_group_shuffle_up(fin.y, fin.y, CL_SIZE);
707 fin.z += intel_sub_group_shuffle_down(fin.z, fin.z, CL_SIZE);
713 /* Threads 0,1 and 2,3 increment x,y for their warp */
714 atomicAdd_g_f(&fout[3 * aidx + (tidxj & 1)], fin.x);
717 fshift_buf[0] += fin.x;
719 /* Threads 0 and 2 increment z for their warp */
720 if ((tidxj & 1) == 0)
722 atomicAdd_g_f(&fout[3 * aidx + 2], fin.z);
725 fshift_buf[1] += fin.z;
729 /* add up local shift forces into global mem */
732 // Threads 0,1 and 2,3 update x,y
733 atomicAdd_g_f(&(fshift[3 * shift + (tidxj & 1)]), fshift_buf[0]);
734 // Threads 0 and 2 update z
735 if ((tidxj & 1) == 0)
737 atomicAdd_g_f(&(fshift[3 * shift + 2]), fshift_buf[1]);
743 /*! Final i-force reduction; this implementation works only with power of two
746 gmx_opencl_inline void reduce_force_i_and_shift_pow2(volatile __local float* f_buf,
748 __global float* fout,
754 __global float* fshift)
756 float fshift_buf = 0;
757 for (int ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
759 int aidx = (sci * NCL_PER_SUPERCL + ci_offset) * CL_SIZE + tidxi;
760 int tidx = tidxi + tidxj * CL_SIZE;
761 /* store i forces in shmem */
762 f_buf[tidx] = fci_buf[ci_offset].x;
763 f_buf[FBUF_STRIDE + tidx] = fci_buf[ci_offset].y;
764 f_buf[2 * FBUF_STRIDE + tidx] = fci_buf[ci_offset].z;
765 barrier(CLK_LOCAL_MEM_FENCE);
767 /* Reduce the initial CL_SIZE values for each i atom to half
768 * every step by using CL_SIZE * i threads.
769 * Can't just use i as loop variable because than nvcc refuses to unroll.
772 for (int j = CL_SIZE_LOG2 - 1; j > 0; j--)
777 f_buf[tidxj * CL_SIZE + tidxi] += f_buf[(tidxj + i) * CL_SIZE + tidxi];
778 f_buf[FBUF_STRIDE + tidxj * CL_SIZE + tidxi] +=
779 f_buf[FBUF_STRIDE + (tidxj + i) * CL_SIZE + tidxi];
780 f_buf[2 * FBUF_STRIDE + tidxj * CL_SIZE + tidxi] +=
781 f_buf[2 * FBUF_STRIDE + (tidxj + i) * CL_SIZE + tidxi];
786 * a) for CL_SIZE<8: id 2 (doing z in next block) is in 2nd warp
787 * b) for all CL_SIZE a barrier is needed before f_buf is reused by next reduce_force_i call
788 * TODO: Test on Nvidia for performance difference between having the barrier here or after the atomicAdd
790 barrier(CLK_LOCAL_MEM_FENCE);
792 /* i == 1, last reduction step, writing to global mem */
793 /* Split the reduction between the first 3 line threads
794 Threads with line id 0 will do the reduction for (float3).x components
795 Threads with line id 1 will do the reduction for (float3).y components
796 Threads with line id 2 will do the reduction for (float3).z components. */
799 float f = f_buf[tidxj * FBUF_STRIDE + tidxi] + f_buf[tidxj * FBUF_STRIDE + i * CL_SIZE + tidxi];
801 atomicAdd_g_f(&fout[3 * aidx + tidxj], f);
809 /* add up local shift forces into global mem */
812 /* Only threads with tidxj < 3 will update fshift.
813 The threads performing the update, must be the same as the threads
814 storing the reduction result above.
818 atomicAdd_g_f(&(fshift[3 * shift + tidxj]), fshift_buf);
823 /*! Final i-force reduction
825 gmx_opencl_inline void reduce_force_i_and_shift(__local float gmx_unused* f_buf,
833 __global float* fshift)
836 reduce_force_i_and_shift_shfl(fci_buf, f, bCalcFshift, tidxi, tidxj, sci, shift, fshift);
838 reduce_force_i_and_shift_pow2(f_buf, fci_buf, f, bCalcFshift, tidxi, tidxj, sci, shift, fshift);
844 gmx_opencl_inline void reduce_energy_shfl(float E_lj,
846 volatile __global float* e_lj,
847 volatile __global float* e_el,
850 E_lj = sub_group_reduce_add(E_lj);
851 E_el = sub_group_reduce_add(E_el);
852 /* Should be get_sub_group_local_id()==0. Doesn't work with Intel Classic driver.
853 * To make it possible to use REDUCE_SHUFFLE with single subgroup per i-j pair
854 * (e.g. subgroup size 16 with CL_SIZE 4), either this "if" needs to be changed or
855 * the definition of WARP_SIZE (currently CL_SIZE*CL_SIZE/2) needs to be changed
856 * (by supporting c_nbnxnGpuClusterpairSplit=1). */
857 if (tidx == 0 || tidx == WARP_SIZE)
859 atomicAdd_g_f(e_lj, E_lj);
860 atomicAdd_g_f(e_el, E_el);
865 /*! Energy reduction; this implementation works only with power of two
868 gmx_opencl_inline void reduce_energy_pow2(volatile __local float* buf,
869 volatile __global float* e_lj,
870 volatile __global float* e_el,
875 int i = WARP_SIZE / 2;
877 /* Can't just use i as loop variable because than nvcc refuses to unroll. */
878 for (j = WARP_SIZE_LOG2 - 1; j > 0; j--)
882 buf[tidx] += buf[tidx + i];
883 buf[FBUF_STRIDE + tidx] += buf[FBUF_STRIDE + tidx + i];
888 /* last reduction step, writing to global mem */
891 float e1 = buf[tidx] + buf[tidx + i];
892 float e2 = buf[FBUF_STRIDE + tidx] + buf[FBUF_STRIDE + tidx + i];
894 atomicAdd_g_f(e_lj, e1);
895 atomicAdd_g_f(e_el, e2);
899 gmx_opencl_inline void reduce_energy(volatile __local float gmx_unused* buf,
902 volatile __global float* e_lj,
903 volatile __global float* e_el,
907 reduce_energy_shfl(E_lj, E_el, e_lj, e_el, tidx);
909 /* flush the energies to shmem and reduce them */
911 buf[FBUF_STRIDE + tidx] = E_el;
912 reduce_energy_pow2(buf + (tidx & WARP_SIZE), e_lj, e_el, tidx & ~WARP_SIZE);
916 gmx_opencl_inline bool gmx_sub_group_any_localmem(volatile __local int* warp_any, int widx, bool pred)
923 bool ret = warp_any[widx];
930 //! Returns a true if predicate is true for any work item in warp
931 gmx_opencl_inline bool gmx_sub_group_any(volatile __local int gmx_unused* warp_any, int gmx_unused widx, bool pred)
933 # if USE_SUBGROUP_ANY
934 return sub_group_any(pred);
936 return gmx_sub_group_any_localmem(warp_any, widx, pred);
940 #endif /* NBNXN_OPENCL_KERNEL_UTILS_CLH */