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 "device_utils.clh"
37 #include "vectype_ops.clh"
39 #define CL_SIZE (NBNXN_GPU_CLUSTER_SIZE)
40 #define NCL_PER_SUPERCL (NBNXN_GPU_NCLUSTER_PER_SUPERCLUSTER)
42 #define WARP_SIZE (CL_SIZE*CL_SIZE/2) //Currently only c_nbnxnGpuClusterpairSplit=2 supported
44 #if defined _NVIDIA_SOURCE_ || defined _AMD_SOURCE_
45 /* Currently we enable CJ prefetch for AMD/NVIDIA and disable it for other vendors
46 * Note that this should precede the kernel_utils include.
48 #define USE_CJ_PREFETCH 1
50 #define USE_CJ_PREFETCH 0
53 #if (defined cl_intel_subgroups || defined cl_khr_subgroups || __OPENCL_VERSION__ >= 210)
54 #define HAVE_SUBGROUP 1
56 #define HAVE_SUBGROUP 0
59 #ifdef cl_intel_subgroups
60 #define HAVE_INTEL_SUBGROUP 1
62 #define HAVE_INTEL_SUBGROUP 0
66 #define SUBGROUP_SIZE 8
68 #define SUBGROUP_SIZE 64
70 #define SUBGROUP_SIZE 32
73 #define REDUCE_SHUFFLE (HAVE_INTEL_SUBGROUP && CL_SIZE == 4 && SUBGROUP_SIZE == WARP_SIZE)
74 #define USE_SUBGROUP_ANY (HAVE_SUBGROUP && SUBGROUP_SIZE == WARP_SIZE)
75 #define USE_SUBGROUP_PRELOAD HAVE_INTEL_SUBGROUP
77 /* 1.0 / sqrt(M_PI) */
78 #define M_FLOAT_1_SQRTPI 0.564189583547756f
82 #ifndef NBNXN_OPENCL_KERNEL_UTILS_CLH
83 #define NBNXN_OPENCL_KERNEL_UTILS_CLH
86 #define WARP_SIZE_LOG2 (5)
87 #define CL_SIZE_LOG2 (3)
89 #define WARP_SIZE_LOG2 (3)
90 #define CL_SIZE_LOG2 (2)
92 #error unsupported CL_SIZE
95 #define CL_SIZE_SQ (CL_SIZE * CL_SIZE)
96 #define FBUF_STRIDE (CL_SIZE_SQ)
98 #define ONE_SIXTH_F 0.16666667f
99 #define ONE_TWELVETH_F 0.08333333f
102 // Data structures shared between OpenCL device code and OpenCL host code
103 // TODO: review, improve
104 // Replaced real by float for now, to avoid including any other header
111 /* Used with potential switching:
112 * rsw = max(r - r_switch, 0)
113 * sw = 1 + c3*rsw^3 + c4*rsw^4 + c5*rsw^5
114 * dsw = 3*c3*rsw^2 + 4*c4*rsw^3 + 5*c5*rsw^4
115 * force = force*dsw - potential*sw
124 // Data structure shared between the OpenCL device code and OpenCL host code
125 // Must not contain OpenCL objects (buffers)
126 typedef struct cl_nbparam_params
129 int eeltype; /**< type of electrostatics, takes values from #eelCu */
130 int vdwtype; /**< type of VdW impl., takes values from #evdwCu */
132 float epsfac; /**< charge multiplication factor */
133 float c_rf; /**< Reaction-field/plain cutoff electrostatics const. */
134 float two_k_rf; /**< Reaction-field electrostatics constant */
135 float ewald_beta; /**< Ewald/PME parameter */
136 float sh_ewald; /**< Ewald/PME correction term substracted from the direct-space potential */
137 float sh_lj_ewald; /**< LJ-Ewald/PME correction term added to the correction potential */
138 float ewaldcoeff_lj; /**< LJ-Ewald/PME coefficient */
140 float rcoulomb_sq; /**< Coulomb cut-off squared */
142 float rvdw_sq; /**< VdW cut-off squared */
143 float rvdw_switch; /**< VdW switched cut-off */
144 float rlistOuter_sq; /**< Full, outer pair-list cut-off squared */
145 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 */
147 shift_consts_t dispersion_shift; /**< VdW shift dispersion constants */
148 shift_consts_t repulsion_shift; /**< VdW shift repulsion constants */
149 switch_consts_t vdw_switch; /**< VdW switch constants */
151 /* Ewald Coulomb force table data - accessed through texture memory */
152 float coulomb_tab_scale; /**< table scale/spacing */
153 }cl_nbparam_params_t;
156 int sci; /* i-super-cluster */
157 int shift; /* Shift vector index plus possible flags */
158 int cj4_ind_start; /* Start index into cj4 */
159 int cj4_ind_end; /* End index into cj4 */
163 unsigned int imask; /* The i-cluster interactions mask for 1 warp */
164 int excl_ind; /* Index into the exclusion array for 1 warp */
168 int cj[4]; /* The 4 j-clusters */
169 nbnxn_im_ei_t imei[2]; /* The i-cluster mask data for 2 warps */
174 unsigned int pair[CL_SIZE*CL_SIZE/2]; /* Topology exclusion interaction bits for one warp,
175 * each unsigned has bitS for 4*8 i clusters
179 /*! i-cluster interaction mask for a super-cluster with all NCL_PER_SUPERCL bits set */
180 __constant unsigned supercl_interaction_mask = ((1U << NCL_PER_SUPERCL) - 1U);
183 void preloadCj4Generic(__local int *sm_cjPreload,
184 const __global int *gm_cj,
190 /* Pre-load cj into shared memory */
191 #if defined _AMD_SOURCE_ //TODO: fix by setting c_nbnxnGpuClusterpairSplit properly
192 if (tidxj == 0 & tidxi < NBNXN_GPU_JGROUP_SIZE)
194 sm_cjPreload[tidxi] = gm_cj[tidxi];
197 const int c_clSize = CL_SIZE;
198 const int c_nbnxnGpuJgroupSize = NBNXN_GPU_JGROUP_SIZE;
199 const int c_nbnxnGpuClusterpairSplit = 2;
200 const int c_splitClSize = c_clSize/c_nbnxnGpuClusterpairSplit;
202 if ((tidxj == 0 | tidxj == c_splitClSize) & (tidxi < c_nbnxnGpuJgroupSize))
204 sm_cjPreload[tidxi + tidxj * c_nbnxnGpuJgroupSize/c_splitClSize] = gm_cj[tidxi];
210 #if USE_SUBGROUP_PRELOAD
212 int preloadCj4Subgroup(const __global int *gm_cj)
214 //loads subgroup-size # of elements (8) instead of the 4 required
215 //equivalent to *cjs = *gm_cj
216 return intel_sub_group_block_read((const __global uint *)gm_cj);
218 #endif //USE_SUBGROUP_PRELOAD
220 #if USE_SUBGROUP_PRELOAD
223 typedef __local int* CjType;
226 /*! \brief Preload cj4
228 * - For AMD we load once for a wavefront of 64 threads (on 4 threads * NTHREAD_Z)
229 * - For NVIDIA once per warp (on 2x4 threads * NTHREAD_Z)
230 * - For Intel(/USE_SUBGROUP_PRELOAD) loads into private memory(/register) instead of local memory
232 * It is the caller's responsibility to make sure that data is consumed only when
233 * it's ready. This function does not call a barrier.
236 void preloadCj4(CjType *cjs,
237 const __global int *gm_cj,
242 #if USE_SUBGROUP_PRELOAD
243 *cjs = preloadCj4Subgroup(gm_cj);
244 #elif USE_CJ_PREFETCH
245 preloadCj4Generic(*cjs, gm_cj, tidxi, tidxj, iMaskCond);
252 int loadCjPreload(__local int* sm_cjPreload,
257 #if defined _AMD_SOURCE_
258 int warpLoadOffset = 0; //TODO: fix by setting c_nbnxnGpuClusterpairSplit properly
260 const int c_clSize = CL_SIZE;
261 const int c_nbnxnGpuJgroupSize = NBNXN_GPU_JGROUP_SIZE;
262 const int c_nbnxnGpuClusterpairSplit = 2;
263 const int c_splitClSize = c_clSize/c_nbnxnGpuClusterpairSplit;
265 int warpLoadOffset = (tidxj & c_splitClSize) * c_nbnxnGpuJgroupSize/c_splitClSize;
267 return sm_cjPreload[jm + warpLoadOffset];
270 /* \brief Load a cj given a jm index.
272 * If cj4 preloading is enabled, it loads from the local memory, otherwise from global.
275 int loadCj(CjType cjs, const __global int *gm_cj,
276 int jm, int tidxi, int tidxj)
278 #if USE_SUBGROUP_PRELOAD
279 return sub_group_broadcast(cjs, jm);
280 #elif USE_CJ_PREFETCH
281 return loadCjPreload(cjs, jm, tidxi, tidxj);
287 /*! Convert LJ sigma,epsilon parameters to C6,C12. */
289 void convert_sigma_epsilon_to_c6_c12(const float sigma,
294 float sigma2, sigma6;
296 sigma2 = sigma * sigma;
297 sigma6 = sigma2 *sigma2 * sigma2;
298 *c6 = epsilon * sigma6;
303 /*! Apply force switch, force + energy version. */
305 void calculate_force_switch_F(cl_nbparam_params_t *nbparam,
314 /* force switch constants */
315 float disp_shift_V2 = nbparam->dispersion_shift.c2;
316 float disp_shift_V3 = nbparam->dispersion_shift.c3;
317 float repu_shift_V2 = nbparam->repulsion_shift.c2;
318 float repu_shift_V3 = nbparam->repulsion_shift.c3;
321 r_switch = r - nbparam->rvdw_switch;
322 r_switch = r_switch >= 0.0f ? r_switch : 0.0f;
325 -c6*(disp_shift_V2 + disp_shift_V3*r_switch)*r_switch*r_switch*inv_r +
326 c12*(-repu_shift_V2 + repu_shift_V3*r_switch)*r_switch*r_switch*inv_r;
329 /*! Apply force switch, force-only version. */
331 void calculate_force_switch_F_E(cl_nbparam_params_t *nbparam,
341 /* force switch constants */
342 float disp_shift_V2 = nbparam->dispersion_shift.c2;
343 float disp_shift_V3 = nbparam->dispersion_shift.c3;
344 float repu_shift_V2 = nbparam->repulsion_shift.c2;
345 float repu_shift_V3 = nbparam->repulsion_shift.c3;
347 float disp_shift_F2 = nbparam->dispersion_shift.c2/3;
348 float disp_shift_F3 = nbparam->dispersion_shift.c3/4;
349 float repu_shift_F2 = nbparam->repulsion_shift.c2/3;
350 float repu_shift_F3 = nbparam->repulsion_shift.c3/4;
353 r_switch = r - nbparam->rvdw_switch;
354 r_switch = r_switch >= 0.0f ? r_switch : 0.0f;
357 -c6*(disp_shift_V2 + disp_shift_V3*r_switch)*r_switch*r_switch*inv_r +
358 c12*(-repu_shift_V2 + repu_shift_V3*r_switch)*r_switch*r_switch*inv_r;
360 c6*(disp_shift_F2 + disp_shift_F3*r_switch)*r_switch*r_switch*r_switch -
361 c12*(repu_shift_F2 + repu_shift_F3*r_switch)*r_switch*r_switch*r_switch;
364 /*! Apply potential switch, force-only version. */
366 void calculate_potential_switch_F(cl_nbparam_params_t *nbparam,
375 /* potential switch constants */
376 float switch_V3 = nbparam->vdw_switch.c3;
377 float switch_V4 = nbparam->vdw_switch.c4;
378 float switch_V5 = nbparam->vdw_switch.c5;
379 float switch_F2 = nbparam->vdw_switch.c3;
380 float switch_F3 = nbparam->vdw_switch.c4;
381 float switch_F4 = nbparam->vdw_switch.c5;
384 r_switch = r - nbparam->rvdw_switch;
386 /* Unlike in the F+E kernel, conditional is faster here */
389 sw = 1.0f + (switch_V3 + (switch_V4 + switch_V5*r_switch)*r_switch)*r_switch*r_switch*r_switch;
390 dsw = (switch_F2 + (switch_F3 + switch_F4*r_switch)*r_switch)*r_switch*r_switch;
392 *F_invr = (*F_invr)*sw - inv_r*(*E_lj)*dsw;
396 /*! Apply potential switch, force + energy version. */
398 void calculate_potential_switch_F_E(cl_nbparam_params_t *nbparam,
407 /* potential switch constants */
408 float switch_V3 = nbparam->vdw_switch.c3;
409 float switch_V4 = nbparam->vdw_switch.c4;
410 float switch_V5 = nbparam->vdw_switch.c5;
411 float switch_F2 = nbparam->vdw_switch.c3;
412 float switch_F3 = nbparam->vdw_switch.c4;
413 float switch_F4 = nbparam->vdw_switch.c5;
416 r_switch = r - nbparam->rvdw_switch;
417 r_switch = r_switch >= 0.0f ? r_switch : 0.0f;
419 /* Unlike in the F-only kernel, masking is faster here */
420 sw = 1.0f + (switch_V3 + (switch_V4 + switch_V5*r_switch)*r_switch)*r_switch*r_switch*r_switch;
421 dsw = (switch_F2 + (switch_F3 + switch_F4*r_switch)*r_switch)*r_switch*r_switch;
423 *F_invr = (*F_invr)*sw - inv_r*(*E_lj)*dsw;
427 /*! Calculate LJ-PME grid force contribution with
428 * geometric combination rule.
431 void calculate_lj_ewald_comb_geom_F(__constant float * nbfp_comb_climg2d,
440 float c6grid, inv_r6_nm, cr2, expmcr2, poly;
442 c6grid = nbfp_comb_climg2d[2*typei]*nbfp_comb_climg2d[2*typej];
444 /* Recalculate inv_r6 without exclusion mask */
445 inv_r6_nm = inv_r2*inv_r2*inv_r2;
448 poly = 1.0f + cr2 + 0.5f*cr2*cr2;
450 /* Subtract the grid force from the total LJ force */
451 *F_invr += c6grid*(inv_r6_nm - expmcr2*(inv_r6_nm*poly + lje_coeff6_6))*inv_r2;
454 /*! Calculate LJ-PME grid force + energy contribution with
455 * geometric combination rule.
458 void calculate_lj_ewald_comb_geom_F_E(__constant float *nbfp_comb_climg2d,
459 cl_nbparam_params_t *nbparam,
470 float c6grid, inv_r6_nm, cr2, expmcr2, poly, sh_mask;
472 c6grid = nbfp_comb_climg2d[2*typei]*nbfp_comb_climg2d[2*typej];
474 /* Recalculate inv_r6 without exclusion mask */
475 inv_r6_nm = inv_r2*inv_r2*inv_r2;
478 poly = 1.0f + cr2 + 0.5f*cr2*cr2;
480 /* Subtract the grid force from the total LJ force */
481 *F_invr += c6grid*(inv_r6_nm - expmcr2*(inv_r6_nm*poly + lje_coeff6_6))*inv_r2;
483 /* Shift should be applied only to real LJ pairs */
484 sh_mask = nbparam->sh_lj_ewald*int_bit;
485 *E_lj += ONE_SIXTH_F*c6grid*(inv_r6_nm*(1.0f - expmcr2*poly) + sh_mask);
488 /*! Calculate LJ-PME grid force + energy contribution (if E_lj != NULL) with
489 * Lorentz-Berthelot combination rule.
490 * We use a single F+E kernel with conditional because the performance impact
491 * of this is pretty small and LB on the CPU is anyway very slow.
494 void calculate_lj_ewald_comb_LB_F_E(__constant float *nbfp_comb_climg2d,
495 cl_nbparam_params_t *nbparam,
507 float c6grid, inv_r6_nm, cr2, expmcr2, poly;
508 float sigma, sigma2, epsilon;
510 /* sigma and epsilon are scaled to give 6*C6 */
511 sigma = nbfp_comb_climg2d[2*typei] + nbfp_comb_climg2d[2*typej];
513 epsilon = nbfp_comb_climg2d[2*typei+1]*nbfp_comb_climg2d[2*typej+1];
515 sigma2 = sigma*sigma;
516 c6grid = epsilon*sigma2*sigma2*sigma2;
518 /* Recalculate inv_r6 without exclusion mask */
519 inv_r6_nm = inv_r2*inv_r2*inv_r2;
522 poly = 1.0f + cr2 + 0.5f*cr2*cr2;
524 /* Subtract the grid force from the total LJ force */
525 *F_invr += c6grid*(inv_r6_nm - expmcr2*(inv_r6_nm*poly + lje_coeff6_6))*inv_r2;
531 /* Shift should be applied only to real LJ pairs */
532 sh_mask = nbparam->sh_lj_ewald*int_bit;
533 *E_lj += ONE_SIXTH_F*c6grid*(inv_r6_nm*(1.0f - expmcr2*poly) + sh_mask);
537 /*! Interpolate Ewald coulomb force using the table through the tex_nbfp texture.
538 * Original idea: from the OpenMM project
540 gmx_opencl_inline float
541 interpolate_coulomb_force_r(__constant float *coulomb_tab_climg2d,
545 float normalized = scale * r;
546 int index = (int) normalized;
547 float fract2 = normalized - index;
548 float fract1 = 1.0f - fract2;
550 return fract1*coulomb_tab_climg2d[index] +
551 fract2*coulomb_tab_climg2d[index + 1];
554 /*! Calculate analytical Ewald correction term. */
556 float pmecorrF(float z2)
558 const float FN6 = -1.7357322914161492954e-8f;
559 const float FN5 = 1.4703624142580877519e-6f;
560 const float FN4 = -0.000053401640219807709149f;
561 const float FN3 = 0.0010054721316683106153f;
562 const float FN2 = -0.019278317264888380590f;
563 const float FN1 = 0.069670166153766424023f;
564 const float FN0 = -0.75225204789749321333f;
566 const float FD4 = 0.0011193462567257629232f;
567 const float FD3 = 0.014866955030185295499f;
568 const float FD2 = 0.11583842382862377919f;
569 const float FD1 = 0.50736591960530292870f;
570 const float FD0 = 1.0f;
573 float polyFN0, polyFN1, polyFD0, polyFD1;
577 polyFD0 = FD4*z4 + FD2;
578 polyFD1 = FD3*z4 + FD1;
579 polyFD0 = polyFD0*z4 + FD0;
580 polyFD0 = polyFD1*z2 + polyFD0;
582 polyFD0 = 1.0f/polyFD0;
584 polyFN0 = FN6*z4 + FN4;
585 polyFN1 = FN5*z4 + FN3;
586 polyFN0 = polyFN0*z4 + FN2;
587 polyFN1 = polyFN1*z4 + FN1;
588 polyFN0 = polyFN0*z4 + FN0;
589 polyFN0 = polyFN1*z2 + polyFN0;
591 return polyFN0*polyFD0;
596 void reduce_force_j_shfl(float3 fin, __global float *fout,
597 int tidxi, int tidxj, int aidx)
599 /* Only does reduction over 4 elements in cluster. Needs to be changed
600 * for CL_SIZE>4. See CUDA code for required code */
601 fin.x += intel_sub_group_shuffle_down(fin.x, fin.x, 1);
602 fin.y += intel_sub_group_shuffle_up (fin.y, fin.y, 1);
603 fin.z += intel_sub_group_shuffle_down(fin.z, fin.z, 1);
604 if ((tidxi & 1) == 1)
608 fin.x += intel_sub_group_shuffle_down(fin.x, fin.x, 2);
609 fin.z += intel_sub_group_shuffle_up (fin.z, fin.z, 2);
616 atomicAdd_g_f(&fout[3 * aidx + tidxi], fin.x);
622 void reduce_force_j_generic(__local float *f_buf, float3 fcj_buf, __global float *fout,
623 int tidxi, int tidxj, int aidx)
625 int tidx = tidxi + tidxj*CL_SIZE;
626 f_buf[ tidx] = fcj_buf.x;
627 f_buf[ FBUF_STRIDE + tidx] = fcj_buf.y;
628 f_buf[2 * FBUF_STRIDE + tidx] = fcj_buf.z;
630 /* Split the reduction between the first 3 column threads
631 Threads with column id 0 will do the reduction for (float3).x components
632 Threads with column id 1 will do the reduction for (float3).y components
633 Threads with column id 2 will do the reduction for (float3).z components.
634 The reduction is performed for each line tidxj of f_buf. */
638 for (int j = tidxj * CL_SIZE; j < (tidxj + 1) * CL_SIZE; j++)
640 f += f_buf[FBUF_STRIDE * tidxi + j];
643 atomicAdd_g_f(&fout[3 * aidx + tidxi], f);
647 /*! Final j-force reduction
650 void reduce_force_j(__local float *f_buf, float3 fcj_buf, __global float *fout,
651 int tidxi, int tidxj, int aidx)
654 reduce_force_j_shfl(fcj_buf, fout, tidxi, tidxj, aidx);
656 reduce_force_j_generic(f_buf, fcj_buf, fout, tidxi, tidxj, aidx);
662 void reduce_force_i_and_shift_shfl(float3* fci_buf, __global float *fout,
663 bool bCalcFshift, int tidxi, int tidxj,
664 int sci, int shift, __global float *fshift)
666 /* Only does reduction over 4 elements in cluster (2 per warp). Needs to be changed
668 float2 fshift_buf = 0;
669 for (int ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
671 int aidx = (sci * NCL_PER_SUPERCL + ci_offset) * CL_SIZE + tidxi;
672 float3 fin = fci_buf[ci_offset];
673 fin.x += intel_sub_group_shuffle_down(fin.x, fin.x, CL_SIZE);
674 fin.y += intel_sub_group_shuffle_up (fin.y, fin.y, CL_SIZE);
675 fin.z += intel_sub_group_shuffle_down(fin.z, fin.z, CL_SIZE);
681 /* Threads 0,1 and 2,3 increment x,y for their warp */
682 atomicAdd_g_f(&fout[3*aidx + (tidxj & 1)], fin.x);
685 fshift_buf[0] += fin.x;
687 /* Threads 0 and 2 increment z for their warp */
688 if ((tidxj & 1) == 0)
690 atomicAdd_g_f(&fout[3*aidx+2], fin.z);
693 fshift_buf[1] += fin.z;
697 /* add up local shift forces into global mem */
700 //Threads 0,1 and 2,3 update x,y
701 atomicAdd_g_f(&(fshift[3 * shift + (tidxj&1)]), fshift_buf[0]);
702 //Threads 0 and 2 update z
703 if ((tidxj & 1) == 0)
705 atomicAdd_g_f(&(fshift[3 * shift + 2]), fshift_buf[1]);
711 /*! Final i-force reduction; this implementation works only with power of two
715 void reduce_force_i_and_shift_pow2(volatile __local float *f_buf, float3* fci_buf,
716 __global float *fout,
718 int tidxi, int tidxj,
719 int sci, int shift, __global float *fshift)
721 float fshift_buf = 0;
722 for (int ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
724 int aidx = (sci * NCL_PER_SUPERCL + ci_offset) * CL_SIZE + tidxi;
725 int tidx = tidxi + tidxj*CL_SIZE;
726 /* store i forces in shmem */
727 f_buf[ tidx] = fci_buf[ci_offset].x;
728 f_buf[ FBUF_STRIDE + tidx] = fci_buf[ci_offset].y;
729 f_buf[2 * FBUF_STRIDE + tidx] = fci_buf[ci_offset].z;
730 barrier(CLK_LOCAL_MEM_FENCE);
733 /* Reduce the initial CL_SIZE values for each i atom to half
734 * every step by using CL_SIZE * i threads.
735 * Can't just use i as loop variable because than nvcc refuses to unroll.
738 for (j = CL_SIZE_LOG2 - 1; j > 0; j--)
743 f_buf[ tidxj * CL_SIZE + tidxi] += f_buf[ (tidxj + i) * CL_SIZE + tidxi];
744 f_buf[ FBUF_STRIDE + tidxj * CL_SIZE + tidxi] += f_buf[ FBUF_STRIDE + (tidxj + i) * CL_SIZE + tidxi];
745 f_buf[2 * FBUF_STRIDE + tidxj * CL_SIZE + tidxi] += f_buf[2 * FBUF_STRIDE + (tidxj + i) * CL_SIZE + tidxi];
750 * a) for CL_SIZE<8: id 2 (doing z in next block) is in 2nd warp
751 * b) for all CL_SIZE a barrier is needed before f_buf is reused by next reduce_force_i call
752 * TODO: Test on Nvidia for performance difference between having the barrier here or after the atomicAdd
754 barrier(CLK_LOCAL_MEM_FENCE);
756 /* i == 1, last reduction step, writing to global mem */
757 /* Split the reduction between the first 3 line threads
758 Threads with line id 0 will do the reduction for (float3).x components
759 Threads with line id 1 will do the reduction for (float3).y components
760 Threads with line id 2 will do the reduction for (float3).z components. */
763 float f = f_buf[tidxj * FBUF_STRIDE + tidxi] + f_buf[tidxj * FBUF_STRIDE + i * CL_SIZE + tidxi];
765 atomicAdd_g_f(&fout[3 * aidx + tidxj], f);
773 /* add up local shift forces into global mem */
776 /* Only threads with tidxj < 3 will update fshift.
777 The threads performing the update, must be the same as the threads
778 storing the reduction result above.
782 atomicAdd_g_f(&(fshift[3 * shift + tidxj]), fshift_buf);
787 /*! Final i-force reduction
790 void reduce_force_i_and_shift(__local float *f_buf, float3* fci_buf, __global float *f,
791 bool bCalcFshift, int tidxi, int tidxj, int sci,
792 int shift, __global float *fshift)
795 reduce_force_i_and_shift_shfl(fci_buf, f, bCalcFshift, tidxi, tidxj,
798 reduce_force_i_and_shift_pow2(f_buf, fci_buf, f, bCalcFshift, tidxi, tidxj,
807 void reduce_energy_shfl(float E_lj, float E_el,
808 volatile __global float *e_lj,
809 volatile __global float *e_el,
812 E_lj = sub_group_reduce_add(E_lj);
813 E_el = sub_group_reduce_add(E_el);
814 /* Should be get_sub_group_local_id()==0. Doesn't work with Intel Classic driver.
815 * To make it possible to use REDUCE_SHUFFLE with single subgroup per i-j pair
816 * (e.g. subgroup size 16 with CL_SIZE 4), either this "if" needs to be changed or
817 * the definition of WARP_SIZE (currently CL_SIZE*CL_SIZE/2) needs to be changed
818 * (by supporting c_nbnxnGpuClusterpairSplit=1). */
819 if (tidx == 0 || tidx == WARP_SIZE)
821 atomicAdd_g_f(e_lj, E_lj);
822 atomicAdd_g_f(e_el, E_el);
827 /*! Energy reduction; this implementation works only with power of two
831 void reduce_energy_pow2(volatile __local float *buf,
832 volatile __global float *e_lj,
833 volatile __global float *e_el,
841 /* Can't just use i as loop variable because than nvcc refuses to unroll. */
842 for (j = WARP_SIZE_LOG2 - 1; j > 0; j--)
846 buf[ tidx] += buf[ tidx + i];
847 buf[FBUF_STRIDE + tidx] += buf[FBUF_STRIDE + tidx + i];
852 /* last reduction step, writing to global mem */
855 e1 = buf[ tidx] + buf[ tidx + i];
856 e2 = buf[FBUF_STRIDE + tidx] + buf[FBUF_STRIDE + tidx + i];
858 atomicAdd_g_f(e_lj, e1);
859 atomicAdd_g_f(e_el, e2);
864 void reduce_energy(volatile __local float *buf,
865 float E_lj, float E_el,
866 volatile __global float *e_lj,
867 volatile __global float *e_el,
871 reduce_energy_shfl(E_lj, E_el, e_lj, e_el, tidx);
873 /* flush the energies to shmem and reduce them */
875 buf[FBUF_STRIDE + tidx] = E_el;
876 reduce_energy_pow2(buf + (tidx & WARP_SIZE), e_lj, e_el, tidx & ~WARP_SIZE);
880 bool gmx_sub_group_any_localmem(volatile __local uint *warp_any, int widx, bool pred)
887 bool ret = warp_any[widx];
894 //! Returns a true if predicate is true for any work item in warp
895 bool gmx_sub_group_any(volatile __local uint *warp_any, int widx, bool pred)
898 return sub_group_any(pred);
900 return gmx_sub_group_any_localmem(warp_any, widx, pred);
904 #endif /* NBNXN_OPENCL_KERNEL_UTILS_CLH */