2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2016,2017,2018, 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 #include "gromacs/gpu_utils/vectype_ops.clh"
37 #include "gromacs/gpu_utils/device_utils.clh"
38 #include "gromacs/mdlib/nbnxn_consts.h"
39 #include "gromacs/pbcutil/ishift.h"
41 #include "nbnxn_ocl_consts.h"
43 #define CL_SIZE (NBNXN_GPU_CLUSTER_SIZE)
44 #define NCL_PER_SUPERCL c_nbnxnGpuNumClusterPerSupercluster
46 #define WARP_SIZE (CL_SIZE*CL_SIZE/2) //Currently only c_nbnxnGpuClusterpairSplit=2 supported
48 #if defined _NVIDIA_SOURCE_ || defined _AMD_SOURCE_
49 /* Currently we enable CJ prefetch for AMD/NVIDIA and disable it for other vendors
50 * Note that this should precede the kernel_utils include.
52 #define USE_CJ_PREFETCH 1
54 #define USE_CJ_PREFETCH 0
57 #if (defined cl_intel_subgroups || defined cl_khr_subgroups || __OPENCL_VERSION__ >= 210)
58 #define HAVE_SUBGROUP 1
60 #define HAVE_SUBGROUP 0
63 #ifdef cl_intel_subgroups
64 #define HAVE_INTEL_SUBGROUP 1
66 #define HAVE_INTEL_SUBGROUP 0
70 #define SUBGROUP_SIZE 8
72 #define SUBGROUP_SIZE 64
74 #define SUBGROUP_SIZE 32
77 #define REDUCE_SHUFFLE (HAVE_INTEL_SUBGROUP && CL_SIZE == 4 && SUBGROUP_SIZE == WARP_SIZE)
78 #define USE_SUBGROUP_ANY (HAVE_SUBGROUP && SUBGROUP_SIZE == WARP_SIZE)
79 #define USE_SUBGROUP_PRELOAD HAVE_INTEL_SUBGROUP
81 /* 1.0 / sqrt(M_PI) */
82 #define M_FLOAT_1_SQRTPI 0.564189583547756f
86 #ifndef NBNXN_OPENCL_KERNEL_UTILS_CLH
87 #define NBNXN_OPENCL_KERNEL_UTILS_CLH
90 #define WARP_SIZE_LOG2 (5)
91 #define CL_SIZE_LOG2 (3)
93 #define WARP_SIZE_LOG2 (3)
94 #define CL_SIZE_LOG2 (2)
96 #error unsupported CL_SIZE
99 #define CL_SIZE_SQ (CL_SIZE * CL_SIZE)
100 #define FBUF_STRIDE (CL_SIZE_SQ)
102 #define ONE_SIXTH_F 0.16666667f
103 #define ONE_TWELVETH_F 0.08333333f
106 // Data structures shared between OpenCL device code and OpenCL host code
107 // TODO: review, improve
108 // Replaced real by float for now, to avoid including any other header
115 /* Used with potential switching:
116 * rsw = max(r - r_switch, 0)
117 * sw = 1 + c3*rsw^3 + c4*rsw^4 + c5*rsw^5
118 * dsw = 3*c3*rsw^2 + 4*c4*rsw^3 + 5*c5*rsw^4
119 * force = force*dsw - potential*sw
128 // Data structure shared between the OpenCL device code and OpenCL host code
129 // Must not contain OpenCL objects (buffers)
130 typedef struct cl_nbparam_params
133 int eeltype; /**< type of electrostatics, takes values from #eelCu */
134 int vdwtype; /**< type of VdW impl., takes values from #evdwCu */
136 float epsfac; /**< charge multiplication factor */
137 float c_rf; /**< Reaction-field/plain cutoff electrostatics const. */
138 float two_k_rf; /**< Reaction-field electrostatics constant */
139 float ewald_beta; /**< Ewald/PME parameter */
140 float sh_ewald; /**< Ewald/PME correction term substracted from the direct-space potential */
141 float sh_lj_ewald; /**< LJ-Ewald/PME correction term added to the correction potential */
142 float ewaldcoeff_lj; /**< LJ-Ewald/PME coefficient */
144 float rcoulomb_sq; /**< Coulomb cut-off squared */
146 float rvdw_sq; /**< VdW cut-off squared */
147 float rvdw_switch; /**< VdW switched cut-off */
148 float rlistOuter_sq; /**< Full, outer pair-list cut-off squared */
149 float rlistInner_sq; /**< 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 */
151 shift_consts_t dispersion_shift; /**< VdW shift dispersion constants */
152 shift_consts_t repulsion_shift; /**< VdW shift repulsion constants */
153 switch_consts_t vdw_switch; /**< VdW switch constants */
155 /* Ewald Coulomb force table data - accessed through texture memory */
156 float coulomb_tab_scale; /**< table scale/spacing */
157 }cl_nbparam_params_t;
160 int sci; /* i-super-cluster */
161 int shift; /* Shift vector index plus possible flags */
162 int cj4_ind_start; /* Start index into cj4 */
163 int cj4_ind_end; /* End index into cj4 */
167 unsigned int imask; /* The i-cluster interactions mask for 1 warp */
168 int excl_ind; /* Index into the exclusion array for 1 warp */
172 int cj[4]; /* The 4 j-clusters */
173 nbnxn_im_ei_t imei[2]; /* The i-cluster mask data for 2 warps */
178 unsigned int pair[CL_SIZE*CL_SIZE/2]; /* Topology exclusion interaction bits for one warp,
179 * each unsigned has bitS for 4*8 i clusters
183 /*! i-cluster interaction mask for a super-cluster with all NCL_PER_SUPERCL bits set */
184 __constant unsigned supercl_interaction_mask = ((1U << NCL_PER_SUPERCL) - 1U);
187 void preloadCj4Generic(__local int *sm_cjPreload,
188 const __global int *gm_cj,
194 /* Pre-load cj into shared memory */
195 #if defined _AMD_SOURCE_ //TODO: fix by setting c_nbnxnGpuClusterpairSplit properly
196 if (tidxj == 0 & tidxi < c_nbnxnGpuJgroupSize)
198 sm_cjPreload[tidxi] = gm_cj[tidxi];
201 const int c_clSize = CL_SIZE;
202 const int c_nbnxnGpuClusterpairSplit = 2;
203 const int c_splitClSize = c_clSize/c_nbnxnGpuClusterpairSplit;
205 if ((tidxj == 0 | tidxj == c_splitClSize) & (tidxi < c_nbnxnGpuJgroupSize))
207 sm_cjPreload[tidxi + tidxj * c_nbnxnGpuJgroupSize/c_splitClSize] = gm_cj[tidxi];
213 #if USE_SUBGROUP_PRELOAD
215 int preloadCj4Subgroup(const __global int *gm_cj)
217 //loads subgroup-size # of elements (8) instead of the 4 required
218 //equivalent to *cjs = *gm_cj
219 return intel_sub_group_block_read((const __global uint *)gm_cj);
221 #endif //USE_SUBGROUP_PRELOAD
223 #if USE_SUBGROUP_PRELOAD
226 typedef __local int* CjType;
229 /*! \brief Preload cj4
231 * - For AMD we load once for a wavefront of 64 threads (on 4 threads * NTHREAD_Z)
232 * - For NVIDIA once per warp (on 2x4 threads * NTHREAD_Z)
233 * - For Intel(/USE_SUBGROUP_PRELOAD) loads into private memory(/register) instead of local memory
235 * It is the caller's responsibility to make sure that data is consumed only when
236 * it's ready. This function does not call a barrier.
239 void preloadCj4(CjType *cjs,
240 const __global int *gm_cj,
245 #if USE_SUBGROUP_PRELOAD
246 *cjs = preloadCj4Subgroup(gm_cj);
247 #elif USE_CJ_PREFETCH
248 preloadCj4Generic(*cjs, gm_cj, tidxi, tidxj, iMaskCond);
255 int loadCjPreload(__local int* sm_cjPreload,
260 #if defined _AMD_SOURCE_
261 int warpLoadOffset = 0; //TODO: fix by setting c_nbnxnGpuClusterpairSplit properly
263 const int c_clSize = CL_SIZE;
264 const int c_nbnxnGpuClusterpairSplit = 2;
265 const int c_splitClSize = c_clSize/c_nbnxnGpuClusterpairSplit;
267 int warpLoadOffset = (tidxj & c_splitClSize) * c_nbnxnGpuJgroupSize/c_splitClSize;
269 return sm_cjPreload[jm + warpLoadOffset];
272 /* \brief Load a cj given a jm index.
274 * If cj4 preloading is enabled, it loads from the local memory, otherwise from global.
277 int loadCj(CjType cjs, const __global int *gm_cj,
278 int jm, int tidxi, int tidxj)
280 #if USE_SUBGROUP_PRELOAD
281 return sub_group_broadcast(cjs, jm);
282 #elif USE_CJ_PREFETCH
283 return loadCjPreload(cjs, jm, tidxi, tidxj);
289 /*! Convert LJ sigma,epsilon parameters to C6,C12. */
291 void convert_sigma_epsilon_to_c6_c12(const float sigma,
296 float sigma2, sigma6;
298 sigma2 = sigma * sigma;
299 sigma6 = sigma2 *sigma2 * sigma2;
300 *c6 = epsilon * sigma6;
305 /*! Apply force switch, force + energy version. */
307 void calculate_force_switch_F(cl_nbparam_params_t *nbparam,
316 /* force switch constants */
317 float disp_shift_V2 = nbparam->dispersion_shift.c2;
318 float disp_shift_V3 = nbparam->dispersion_shift.c3;
319 float repu_shift_V2 = nbparam->repulsion_shift.c2;
320 float repu_shift_V3 = nbparam->repulsion_shift.c3;
323 r_switch = r - nbparam->rvdw_switch;
324 r_switch = r_switch >= 0.0f ? r_switch : 0.0f;
327 -c6*(disp_shift_V2 + disp_shift_V3*r_switch)*r_switch*r_switch*inv_r +
328 c12*(-repu_shift_V2 + repu_shift_V3*r_switch)*r_switch*r_switch*inv_r;
331 /*! Apply force switch, force-only version. */
333 void calculate_force_switch_F_E(cl_nbparam_params_t *nbparam,
343 /* force switch constants */
344 float disp_shift_V2 = nbparam->dispersion_shift.c2;
345 float disp_shift_V3 = nbparam->dispersion_shift.c3;
346 float repu_shift_V2 = nbparam->repulsion_shift.c2;
347 float repu_shift_V3 = nbparam->repulsion_shift.c3;
349 float disp_shift_F2 = nbparam->dispersion_shift.c2/3;
350 float disp_shift_F3 = nbparam->dispersion_shift.c3/4;
351 float repu_shift_F2 = nbparam->repulsion_shift.c2/3;
352 float repu_shift_F3 = nbparam->repulsion_shift.c3/4;
355 r_switch = r - nbparam->rvdw_switch;
356 r_switch = r_switch >= 0.0f ? r_switch : 0.0f;
359 -c6*(disp_shift_V2 + disp_shift_V3*r_switch)*r_switch*r_switch*inv_r +
360 c12*(-repu_shift_V2 + repu_shift_V3*r_switch)*r_switch*r_switch*inv_r;
362 c6*(disp_shift_F2 + disp_shift_F3*r_switch)*r_switch*r_switch*r_switch -
363 c12*(repu_shift_F2 + repu_shift_F3*r_switch)*r_switch*r_switch*r_switch;
366 /*! Apply potential switch, force-only version. */
368 void calculate_potential_switch_F(cl_nbparam_params_t *nbparam,
377 /* potential switch constants */
378 float switch_V3 = nbparam->vdw_switch.c3;
379 float switch_V4 = nbparam->vdw_switch.c4;
380 float switch_V5 = nbparam->vdw_switch.c5;
381 float switch_F2 = nbparam->vdw_switch.c3;
382 float switch_F3 = nbparam->vdw_switch.c4;
383 float switch_F4 = nbparam->vdw_switch.c5;
386 r_switch = r - nbparam->rvdw_switch;
388 /* Unlike in the F+E kernel, conditional is faster here */
391 sw = 1.0f + (switch_V3 + (switch_V4 + switch_V5*r_switch)*r_switch)*r_switch*r_switch*r_switch;
392 dsw = (switch_F2 + (switch_F3 + switch_F4*r_switch)*r_switch)*r_switch*r_switch;
394 *F_invr = (*F_invr)*sw - inv_r*(*E_lj)*dsw;
398 /*! Apply potential switch, force + energy version. */
400 void calculate_potential_switch_F_E(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;
419 r_switch = r_switch >= 0.0f ? r_switch : 0.0f;
421 /* Unlike in the F-only kernel, masking is faster here */
422 sw = 1.0f + (switch_V3 + (switch_V4 + switch_V5*r_switch)*r_switch)*r_switch*r_switch*r_switch;
423 dsw = (switch_F2 + (switch_F3 + switch_F4*r_switch)*r_switch)*r_switch*r_switch;
425 *F_invr = (*F_invr)*sw - inv_r*(*E_lj)*dsw;
429 /*! Calculate LJ-PME grid force contribution with
430 * geometric combination rule.
433 void calculate_lj_ewald_comb_geom_F(__constant float * nbfp_comb_climg2d,
442 float c6grid, inv_r6_nm, cr2, expmcr2, poly;
444 c6grid = nbfp_comb_climg2d[2*typei]*nbfp_comb_climg2d[2*typej];
446 /* Recalculate inv_r6 without exclusion mask */
447 inv_r6_nm = inv_r2*inv_r2*inv_r2;
450 poly = 1.0f + cr2 + 0.5f*cr2*cr2;
452 /* Subtract the grid force from the total LJ force */
453 *F_invr += c6grid*(inv_r6_nm - expmcr2*(inv_r6_nm*poly + lje_coeff6_6))*inv_r2;
456 /*! Calculate LJ-PME grid force + energy contribution with
457 * geometric combination rule.
460 void calculate_lj_ewald_comb_geom_F_E(__constant float *nbfp_comb_climg2d,
461 cl_nbparam_params_t *nbparam,
472 float c6grid, inv_r6_nm, cr2, expmcr2, poly, sh_mask;
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;
480 poly = 1.0f + cr2 + 0.5f*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;
485 /* Shift should be applied only to real LJ pairs */
486 sh_mask = nbparam->sh_lj_ewald*int_bit;
487 *E_lj += ONE_SIXTH_F*c6grid*(inv_r6_nm*(1.0f - expmcr2*poly) + sh_mask);
490 /*! Calculate LJ-PME grid force + energy contribution (if E_lj != NULL) with
491 * Lorentz-Berthelot combination rule.
492 * We use a single F+E kernel with conditional because the performance impact
493 * of this is pretty small and LB on the CPU is anyway very slow.
496 void calculate_lj_ewald_comb_LB_F_E(__constant float *nbfp_comb_climg2d,
497 cl_nbparam_params_t *nbparam,
509 float c6grid, inv_r6_nm, cr2, expmcr2, poly;
510 float sigma, sigma2, epsilon;
512 /* sigma and epsilon are scaled to give 6*C6 */
513 sigma = nbfp_comb_climg2d[2*typei] + nbfp_comb_climg2d[2*typej];
515 epsilon = nbfp_comb_climg2d[2*typei+1]*nbfp_comb_climg2d[2*typej+1];
517 sigma2 = sigma*sigma;
518 c6grid = epsilon*sigma2*sigma2*sigma2;
520 /* Recalculate inv_r6 without exclusion mask */
521 inv_r6_nm = inv_r2*inv_r2*inv_r2;
524 poly = 1.0f + cr2 + 0.5f*cr2*cr2;
526 /* Subtract the grid force from the total LJ force */
527 *F_invr += c6grid*(inv_r6_nm - expmcr2*(inv_r6_nm*poly + lje_coeff6_6))*inv_r2;
533 /* Shift should be applied only to real LJ pairs */
534 sh_mask = nbparam->sh_lj_ewald*int_bit;
535 *E_lj += ONE_SIXTH_F*c6grid*(inv_r6_nm*(1.0f - expmcr2*poly) + sh_mask);
539 /*! Interpolate Ewald coulomb force using the table through the tex_nbfp texture.
540 * Original idea: from the OpenMM project
542 gmx_opencl_inline float
543 interpolate_coulomb_force_r(__constant float *coulomb_tab_climg2d,
547 float normalized = scale * r;
548 int index = (int) normalized;
549 float fract2 = normalized - index;
550 float fract1 = 1.0f - fract2;
552 return fract1*coulomb_tab_climg2d[index] +
553 fract2*coulomb_tab_climg2d[index + 1];
556 /*! Calculate analytical Ewald correction term. */
558 float pmecorrF(float z2)
560 const float FN6 = -1.7357322914161492954e-8f;
561 const float FN5 = 1.4703624142580877519e-6f;
562 const float FN4 = -0.000053401640219807709149f;
563 const float FN3 = 0.0010054721316683106153f;
564 const float FN2 = -0.019278317264888380590f;
565 const float FN1 = 0.069670166153766424023f;
566 const float FN0 = -0.75225204789749321333f;
568 const float FD4 = 0.0011193462567257629232f;
569 const float FD3 = 0.014866955030185295499f;
570 const float FD2 = 0.11583842382862377919f;
571 const float FD1 = 0.50736591960530292870f;
572 const float FD0 = 1.0f;
575 float polyFN0, polyFN1, polyFD0, polyFD1;
579 polyFD0 = FD4*z4 + FD2;
580 polyFD1 = FD3*z4 + FD1;
581 polyFD0 = polyFD0*z4 + FD0;
582 polyFD0 = polyFD1*z2 + polyFD0;
584 polyFD0 = 1.0f/polyFD0;
586 polyFN0 = FN6*z4 + FN4;
587 polyFN1 = FN5*z4 + FN3;
588 polyFN0 = polyFN0*z4 + FN2;
589 polyFN1 = polyFN1*z4 + FN1;
590 polyFN0 = polyFN0*z4 + FN0;
591 polyFN0 = polyFN1*z2 + polyFN0;
593 return polyFN0*polyFD0;
598 void reduce_force_j_shfl(float3 fin, __global float *fout,
599 int tidxi, int tidxj, int aidx)
601 /* Only does reduction over 4 elements in cluster. Needs to be changed
602 * for CL_SIZE>4. See CUDA code for required code */
603 fin.x += intel_sub_group_shuffle_down(fin.x, fin.x, 1);
604 fin.y += intel_sub_group_shuffle_up (fin.y, fin.y, 1);
605 fin.z += intel_sub_group_shuffle_down(fin.z, fin.z, 1);
606 if ((tidxi & 1) == 1)
610 fin.x += intel_sub_group_shuffle_down(fin.x, fin.x, 2);
611 fin.z += intel_sub_group_shuffle_up (fin.z, fin.z, 2);
618 atomicAdd_g_f(&fout[3 * aidx + tidxi], fin.x);
624 void reduce_force_j_generic(__local float *f_buf, float3 fcj_buf, __global float *fout,
625 int tidxi, int tidxj, int aidx)
627 int tidx = tidxi + tidxj*CL_SIZE;
628 f_buf[ tidx] = fcj_buf.x;
629 f_buf[ FBUF_STRIDE + tidx] = fcj_buf.y;
630 f_buf[2 * FBUF_STRIDE + tidx] = fcj_buf.z;
632 /* Split the reduction between the first 3 column threads
633 Threads with column id 0 will do the reduction for (float3).x components
634 Threads with column id 1 will do the reduction for (float3).y components
635 Threads with column id 2 will do the reduction for (float3).z components.
636 The reduction is performed for each line tidxj of f_buf. */
640 for (int j = tidxj * CL_SIZE; j < (tidxj + 1) * CL_SIZE; j++)
642 f += f_buf[FBUF_STRIDE * tidxi + j];
645 atomicAdd_g_f(&fout[3 * aidx + tidxi], f);
649 /*! Final j-force reduction
652 void reduce_force_j(__local float *f_buf, float3 fcj_buf, __global float *fout,
653 int tidxi, int tidxj, int aidx)
656 reduce_force_j_shfl(fcj_buf, fout, tidxi, tidxj, aidx);
658 reduce_force_j_generic(f_buf, fcj_buf, fout, tidxi, tidxj, aidx);
664 void reduce_force_i_and_shift_shfl(float3* fci_buf, __global float *fout,
665 bool bCalcFshift, int tidxi, int tidxj,
666 int sci, int shift, __global float *fshift)
668 /* Only does reduction over 4 elements in cluster (2 per warp). Needs to be changed
670 float2 fshift_buf = 0;
671 for (int ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
673 int aidx = (sci * NCL_PER_SUPERCL + ci_offset) * CL_SIZE + tidxi;
674 float3 fin = fci_buf[ci_offset];
675 fin.x += intel_sub_group_shuffle_down(fin.x, fin.x, CL_SIZE);
676 fin.y += intel_sub_group_shuffle_up (fin.y, fin.y, CL_SIZE);
677 fin.z += intel_sub_group_shuffle_down(fin.z, fin.z, CL_SIZE);
683 /* Threads 0,1 and 2,3 increment x,y for their warp */
684 atomicAdd_g_f(&fout[3*aidx + (tidxj & 1)], fin.x);
687 fshift_buf[0] += fin.x;
689 /* Threads 0 and 2 increment z for their warp */
690 if ((tidxj & 1) == 0)
692 atomicAdd_g_f(&fout[3*aidx+2], fin.z);
695 fshift_buf[1] += fin.z;
699 /* add up local shift forces into global mem */
702 //Threads 0,1 and 2,3 update x,y
703 atomicAdd_g_f(&(fshift[3 * shift + (tidxj&1)]), fshift_buf[0]);
704 //Threads 0 and 2 update z
705 if ((tidxj & 1) == 0)
707 atomicAdd_g_f(&(fshift[3 * shift + 2]), fshift_buf[1]);
713 /*! Final i-force reduction; this implementation works only with power of two
717 void reduce_force_i_and_shift_pow2(volatile __local float *f_buf, float3* fci_buf,
718 __global float *fout,
720 int tidxi, int tidxj,
721 int sci, int shift, __global float *fshift)
723 float fshift_buf = 0;
724 for (int ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
726 int aidx = (sci * NCL_PER_SUPERCL + ci_offset) * CL_SIZE + tidxi;
727 int tidx = tidxi + tidxj*CL_SIZE;
728 /* store i forces in shmem */
729 f_buf[ tidx] = fci_buf[ci_offset].x;
730 f_buf[ FBUF_STRIDE + tidx] = fci_buf[ci_offset].y;
731 f_buf[2 * FBUF_STRIDE + tidx] = fci_buf[ci_offset].z;
732 barrier(CLK_LOCAL_MEM_FENCE);
735 /* Reduce the initial CL_SIZE values for each i atom to half
736 * every step by using CL_SIZE * i threads.
737 * Can't just use i as loop variable because than nvcc refuses to unroll.
740 for (j = CL_SIZE_LOG2 - 1; j > 0; j--)
745 f_buf[ tidxj * CL_SIZE + tidxi] += f_buf[ (tidxj + i) * CL_SIZE + tidxi];
746 f_buf[ FBUF_STRIDE + tidxj * CL_SIZE + tidxi] += f_buf[ FBUF_STRIDE + (tidxj + i) * CL_SIZE + tidxi];
747 f_buf[2 * FBUF_STRIDE + tidxj * CL_SIZE + tidxi] += f_buf[2 * FBUF_STRIDE + (tidxj + i) * CL_SIZE + tidxi];
752 * a) for CL_SIZE<8: id 2 (doing z in next block) is in 2nd warp
753 * b) for all CL_SIZE a barrier is needed before f_buf is reused by next reduce_force_i call
754 * TODO: Test on Nvidia for performance difference between having the barrier here or after the atomicAdd
756 barrier(CLK_LOCAL_MEM_FENCE);
758 /* i == 1, last reduction step, writing to global mem */
759 /* Split the reduction between the first 3 line threads
760 Threads with line id 0 will do the reduction for (float3).x components
761 Threads with line id 1 will do the reduction for (float3).y components
762 Threads with line id 2 will do the reduction for (float3).z components. */
765 float f = f_buf[tidxj * FBUF_STRIDE + tidxi] + f_buf[tidxj * FBUF_STRIDE + i * CL_SIZE + tidxi];
767 atomicAdd_g_f(&fout[3 * aidx + tidxj], f);
775 /* add up local shift forces into global mem */
778 /* Only threads with tidxj < 3 will update fshift.
779 The threads performing the update, must be the same as the threads
780 storing the reduction result above.
784 atomicAdd_g_f(&(fshift[3 * shift + tidxj]), fshift_buf);
789 /*! Final i-force reduction
792 void reduce_force_i_and_shift(__local float *f_buf, float3* fci_buf, __global float *f,
793 bool bCalcFshift, int tidxi, int tidxj, int sci,
794 int shift, __global float *fshift)
797 reduce_force_i_and_shift_shfl(fci_buf, f, bCalcFshift, tidxi, tidxj,
800 reduce_force_i_and_shift_pow2(f_buf, fci_buf, f, bCalcFshift, tidxi, tidxj,
809 void reduce_energy_shfl(float E_lj, float E_el,
810 volatile __global float *e_lj,
811 volatile __global float *e_el,
814 E_lj = sub_group_reduce_add(E_lj);
815 E_el = sub_group_reduce_add(E_el);
816 /* Should be get_sub_group_local_id()==0. Doesn't work with Intel Classic driver.
817 * To make it possible to use REDUCE_SHUFFLE with single subgroup per i-j pair
818 * (e.g. subgroup size 16 with CL_SIZE 4), either this "if" needs to be changed or
819 * the definition of WARP_SIZE (currently CL_SIZE*CL_SIZE/2) needs to be changed
820 * (by supporting c_nbnxnGpuClusterpairSplit=1). */
821 if (tidx == 0 || tidx == WARP_SIZE)
823 atomicAdd_g_f(e_lj, E_lj);
824 atomicAdd_g_f(e_el, E_el);
829 /*! Energy reduction; this implementation works only with power of two
833 void reduce_energy_pow2(volatile __local float *buf,
834 volatile __global float *e_lj,
835 volatile __global float *e_el,
843 /* Can't just use i as loop variable because than nvcc refuses to unroll. */
844 for (j = WARP_SIZE_LOG2 - 1; j > 0; j--)
848 buf[ tidx] += buf[ tidx + i];
849 buf[FBUF_STRIDE + tidx] += buf[FBUF_STRIDE + tidx + i];
854 /* last reduction step, writing to global mem */
857 e1 = buf[ tidx] + buf[ tidx + i];
858 e2 = buf[FBUF_STRIDE + tidx] + buf[FBUF_STRIDE + tidx + i];
860 atomicAdd_g_f(e_lj, e1);
861 atomicAdd_g_f(e_el, e2);
866 void reduce_energy(volatile __local float *buf,
867 float E_lj, float E_el,
868 volatile __global float *e_lj,
869 volatile __global float *e_el,
873 reduce_energy_shfl(E_lj, E_el, e_lj, e_el, tidx);
875 /* flush the energies to shmem and reduce them */
877 buf[FBUF_STRIDE + tidx] = E_el;
878 reduce_energy_pow2(buf + (tidx & WARP_SIZE), e_lj, e_el, tidx & ~WARP_SIZE);
882 bool gmx_sub_group_any_localmem(volatile __local uint *warp_any, int widx, bool pred)
889 bool ret = warp_any[widx];
896 //! Returns a true if predicate is true for any work item in warp
897 bool gmx_sub_group_any(volatile __local uint *warp_any, int widx, bool pred)
900 return sub_group_any(pred);
902 return gmx_sub_group_any_localmem(warp_any, widx, pred);
906 #endif /* NBNXN_OPENCL_KERNEL_UTILS_CLH */