2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2016,2017,2018,2019, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
38 #include "gromacs/gpu_utils/device_utils.clh"
39 #include "gromacs/gpu_utils/vectype_ops.clh"
40 #include "gromacs/nbnxm/constants.h"
41 #include "gromacs/pbcutil/ishift.h"
43 #include "nbnxm_ocl_consts.h"
45 #define CL_SIZE (NBNXN_GPU_CLUSTER_SIZE)
46 #define NCL_PER_SUPERCL c_nbnxnGpuNumClusterPerSupercluster
48 #define WARP_SIZE (CL_SIZE * CL_SIZE / 2) // Currently only c_nbnxnGpuClusterpairSplit=2 supported
50 #if defined _NVIDIA_SOURCE_ || defined _AMD_SOURCE_
51 /* Currently we enable CJ prefetch for AMD/NVIDIA and disable it for other vendors
52 * Note that this should precede the kernel_utils include.
54 # define USE_CJ_PREFETCH 1
56 # define USE_CJ_PREFETCH 0
59 #if defined cl_intel_subgroups || defined cl_khr_subgroups \
60 || (defined __OPENCL_VERSION__ && __OPENCL_VERSION__ >= 210)
61 # define HAVE_SUBGROUP 1
63 # define HAVE_SUBGROUP 0
66 #ifdef cl_intel_subgroups
67 # define HAVE_INTEL_SUBGROUP 1
69 # define HAVE_INTEL_SUBGROUP 0
72 #if defined _INTEL_SOURCE_
73 # define SUBGROUP_SIZE 8
74 #elif defined _AMD_SOURCE_
75 # define SUBGROUP_SIZE 64
77 # define SUBGROUP_SIZE 32
80 #define REDUCE_SHUFFLE (HAVE_INTEL_SUBGROUP && CL_SIZE == 4 && SUBGROUP_SIZE == WARP_SIZE)
81 #define USE_SUBGROUP_ANY (HAVE_SUBGROUP && SUBGROUP_SIZE == WARP_SIZE)
82 #define USE_SUBGROUP_PRELOAD HAVE_INTEL_SUBGROUP
84 /* 1.0 / sqrt(M_PI) */
85 #define M_FLOAT_1_SQRTPI 0.564189583547756f
89 #ifndef NBNXN_OPENCL_KERNEL_UTILS_CLH
90 # define NBNXN_OPENCL_KERNEL_UTILS_CLH
93 # define WARP_SIZE_LOG2 (5)
94 # define CL_SIZE_LOG2 (3)
96 # define WARP_SIZE_LOG2 (3)
97 # define CL_SIZE_LOG2 (2)
99 # error unsupported CL_SIZE
102 # define CL_SIZE_SQ (CL_SIZE * CL_SIZE)
103 # define FBUF_STRIDE (CL_SIZE_SQ)
105 # define ONE_SIXTH_F 0.16666667f
106 # define ONE_TWELVETH_F 0.08333333f
110 /* GCC, clang, and some ICC pretending to be GCC */
111 # define gmx_unused __attribute__((unused))
116 // Data structures shared between OpenCL device code and OpenCL host code
117 // TODO: review, improve
118 // Replaced real by float for now, to avoid including any other header
126 /* Used with potential switching:
127 * rsw = max(r - r_switch, 0)
128 * sw = 1 + c3*rsw^3 + c4*rsw^4 + c5*rsw^5
129 * dsw = 3*c3*rsw^2 + 4*c4*rsw^3 + 5*c5*rsw^4
130 * force = force*dsw - potential*sw
140 // Data structure shared between the OpenCL device code and OpenCL host code
141 // Must not contain OpenCL objects (buffers)
142 typedef struct cl_nbparam_params
145 int eeltype; /**< type of electrostatics, takes values from #eelCu */
146 int vdwtype; /**< type of VdW impl., takes values from #evdwCu */
148 float epsfac; /**< charge multiplication factor */
149 float c_rf; /**< Reaction-field/plain cutoff electrostatics const. */
150 float two_k_rf; /**< Reaction-field electrostatics constant */
151 float ewald_beta; /**< Ewald/PME parameter */
152 float sh_ewald; /**< Ewald/PME correction term substracted from the direct-space potential */
153 float sh_lj_ewald; /**< LJ-Ewald/PME correction term added to the correction potential */
154 float ewaldcoeff_lj; /**< LJ-Ewald/PME coefficient */
156 float rcoulomb_sq; /**< Coulomb cut-off squared */
158 float rvdw_sq; /**< VdW cut-off squared */
159 float rvdw_switch; /**< VdW switched cut-off */
160 float rlistOuter_sq; /**< Full, outer pair-list cut-off squared */
161 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 */
163 shift_consts_t dispersion_shift; /**< VdW shift dispersion constants */
164 shift_consts_t repulsion_shift; /**< VdW shift repulsion constants */
165 switch_consts_t vdw_switch; /**< VdW switch constants */
167 /* Ewald Coulomb force table data - accessed through texture memory */
168 float coulomb_tab_scale; /**< table scale/spacing */
169 } cl_nbparam_params_t;
173 int sci; /* i-super-cluster */
174 int shift; /* Shift vector index plus possible flags */
175 int cj4_ind_start; /* Start index into cj4 */
176 int cj4_ind_end; /* End index into cj4 */
181 unsigned int imask; /* The i-cluster interactions mask for 1 warp */
182 int excl_ind; /* Index into the exclusion array for 1 warp */
187 int cj[4]; /* The 4 j-clusters */
188 nbnxn_im_ei_t imei[2]; /* The i-cluster mask data for 2 warps */
194 unsigned int pair[CL_SIZE * CL_SIZE / 2]; /* Topology exclusion interaction bits for one warp,
195 * each unsigned has bitS for 4*8 i clusters
199 /*! i-cluster interaction mask for a super-cluster with all NCL_PER_SUPERCL bits set */
200 __constant unsigned supercl_interaction_mask = ((1U << NCL_PER_SUPERCL) - 1U);
202 gmx_opencl_inline void preloadCj4Generic(__local int* sm_cjPreload,
203 const __global int* gm_cj,
206 bool gmx_unused iMaskCond)
208 /* Pre-load cj into shared memory */
209 # if defined _AMD_SOURCE_ // TODO: fix by setting c_nbnxnGpuClusterpairSplit properly
210 if (tidxj == 0 & tidxi < c_nbnxnGpuJgroupSize)
212 sm_cjPreload[tidxi] = gm_cj[tidxi];
215 const int c_clSize = CL_SIZE;
216 const int c_nbnxnGpuClusterpairSplit = 2;
217 const int c_splitClSize = c_clSize / c_nbnxnGpuClusterpairSplit;
218 if ((tidxj == 0 | tidxj == c_splitClSize) & (tidxi < c_nbnxnGpuJgroupSize))
220 sm_cjPreload[tidxi + tidxj * c_nbnxnGpuJgroupSize / c_splitClSize] = gm_cj[tidxi];
226 # if USE_SUBGROUP_PRELOAD
227 gmx_opencl_inline int preloadCj4Subgroup(const __global int* gm_cj)
229 // loads subgroup-size # of elements (8) instead of the 4 required
230 // equivalent to *cjs = *gm_cj
231 return intel_sub_group_block_read((const __global uint*)gm_cj);
233 # endif // USE_SUBGROUP_PRELOAD
235 # if USE_SUBGROUP_PRELOAD
236 typedef size_t CjType;
238 typedef __local int* CjType;
241 /*! \brief Preload cj4
243 * - For AMD we load once for a wavefront of 64 threads (on 4 threads * NTHREAD_Z)
244 * - For NVIDIA once per warp (on 2x4 threads * NTHREAD_Z)
245 * - For Intel(/USE_SUBGROUP_PRELOAD) loads into private memory(/register) instead of local memory
247 * It is the caller's responsibility to make sure that data is consumed only when
248 * it's ready. This function does not call a barrier.
250 gmx_opencl_inline void preloadCj4(CjType gmx_unused* cjs,
251 const __global int gmx_unused* gm_cj,
252 int gmx_unused tidxi,
253 int gmx_unused tidxj,
254 bool gmx_unused iMaskCond)
256 # if USE_SUBGROUP_PRELOAD
257 *cjs = preloadCj4Subgroup(gm_cj);
258 # elif USE_CJ_PREFETCH
259 preloadCj4Generic(*cjs, gm_cj, tidxi, tidxj, iMaskCond);
265 gmx_opencl_inline int loadCjPreload(__local int* sm_cjPreload, int jm, int gmx_unused tidxi, int gmx_unused tidxj)
267 # if defined _AMD_SOURCE_
268 int warpLoadOffset = 0; // TODO: fix by setting c_nbnxnGpuClusterpairSplit properly
270 const int c_clSize = CL_SIZE;
271 const int c_nbnxnGpuClusterpairSplit = 2;
272 const int c_splitClSize = c_clSize / c_nbnxnGpuClusterpairSplit;
273 int warpLoadOffset = (tidxj & c_splitClSize) * c_nbnxnGpuJgroupSize / c_splitClSize;
275 return sm_cjPreload[jm + warpLoadOffset];
278 /* \brief Load a cj given a jm index.
280 * If cj4 preloading is enabled, it loads from the local memory, otherwise from global.
282 gmx_opencl_inline int
283 loadCj(CjType cjs, const __global int gmx_unused* gm_cj, int jm, int gmx_unused tidxi, int gmx_unused tidxj)
285 # if USE_SUBGROUP_PRELOAD
286 return sub_group_broadcast(cjs, jm);
287 # elif USE_CJ_PREFETCH
288 return loadCjPreload(cjs, jm, tidxi, tidxj);
294 /*! Convert LJ sigma,epsilon parameters to C6,C12. */
295 gmx_opencl_inline void convert_sigma_epsilon_to_c6_c12(const float sigma, const float epsilon, float* c6, float* c12)
297 float sigma2, sigma6;
299 sigma2 = sigma * sigma;
300 sigma6 = sigma2 * sigma2 * sigma2;
301 *c6 = epsilon * sigma6;
306 /*! Apply force switch, force + energy version. */
307 gmx_opencl_inline void calculate_force_switch_F(const 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;
326 *F_invr += -c6 * (disp_shift_V2 + disp_shift_V3 * r_switch) * r_switch * r_switch * inv_r
327 + c12 * (repu_shift_V2 + repu_shift_V3 * r_switch) * r_switch * r_switch * inv_r;
330 /*! Apply force switch, force-only version. */
331 gmx_opencl_inline void calculate_force_switch_F_E(const 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;
356 *F_invr += -c6 * (disp_shift_V2 + disp_shift_V3 * r_switch) * r_switch * r_switch * inv_r
357 + c12 * (repu_shift_V2 + repu_shift_V3 * r_switch) * r_switch * r_switch * inv_r;
358 *E_lj += c6 * (disp_shift_F2 + disp_shift_F3 * r_switch) * r_switch * r_switch * r_switch
359 - c12 * (repu_shift_F2 + repu_shift_F3 * r_switch) * r_switch * r_switch * r_switch;
362 /*! Apply potential switch, force-only version. */
363 gmx_opencl_inline void calculate_potential_switch_F(const cl_nbparam_params_t* nbparam,
372 /* potential switch constants */
373 float switch_V3 = nbparam->vdw_switch.c3;
374 float switch_V4 = nbparam->vdw_switch.c4;
375 float switch_V5 = nbparam->vdw_switch.c5;
376 float switch_F2 = nbparam->vdw_switch.c3;
377 float switch_F3 = nbparam->vdw_switch.c4;
378 float switch_F4 = nbparam->vdw_switch.c5;
381 r_switch = r - nbparam->rvdw_switch;
383 /* Unlike in the F+E kernel, conditional is faster here */
386 sw = 1.0f + (switch_V3 + (switch_V4 + switch_V5 * r_switch) * r_switch) * r_switch * r_switch * r_switch;
387 dsw = (switch_F2 + (switch_F3 + switch_F4 * r_switch) * r_switch) * r_switch * r_switch;
389 *F_invr = (*F_invr) * sw - inv_r * (*E_lj) * dsw;
393 /*! Apply potential switch, force + energy version. */
394 gmx_opencl_inline void calculate_potential_switch_F_E(const cl_nbparam_params_t* nbparam,
403 /* potential switch constants */
404 float switch_V3 = nbparam->vdw_switch.c3;
405 float switch_V4 = nbparam->vdw_switch.c4;
406 float switch_V5 = nbparam->vdw_switch.c5;
407 float switch_F2 = nbparam->vdw_switch.c3;
408 float switch_F3 = nbparam->vdw_switch.c4;
409 float switch_F4 = nbparam->vdw_switch.c5;
412 r_switch = r - nbparam->rvdw_switch;
413 r_switch = r_switch >= 0.0f ? r_switch : 0.0f;
415 /* Unlike in the F-only kernel, masking is faster here */
416 sw = 1.0f + (switch_V3 + (switch_V4 + switch_V5 * r_switch) * r_switch) * r_switch * r_switch * r_switch;
417 dsw = (switch_F2 + (switch_F3 + switch_F4 * r_switch) * r_switch) * r_switch * r_switch;
419 *F_invr = (*F_invr) * sw - inv_r * (*E_lj) * dsw;
423 /*! Calculate LJ-PME grid force contribution with
424 * geometric combination rule.
426 gmx_opencl_inline void calculate_lj_ewald_comb_geom_F(__constant const float* nbfp_comb_climg2d,
435 float c6grid, inv_r6_nm, cr2, expmcr2, poly;
437 c6grid = nbfp_comb_climg2d[2 * typei] * nbfp_comb_climg2d[2 * typej];
439 /* Recalculate inv_r6 without exclusion mask */
440 inv_r6_nm = inv_r2 * inv_r2 * inv_r2;
441 cr2 = lje_coeff2 * r2;
443 poly = 1.0f + cr2 + 0.5f * cr2 * cr2;
445 /* Subtract the grid force from the total LJ force */
446 *F_invr += c6grid * (inv_r6_nm - expmcr2 * (inv_r6_nm * poly + lje_coeff6_6)) * inv_r2;
449 /*! Calculate LJ-PME grid force + energy contribution with
450 * geometric combination rule.
452 gmx_opencl_inline void calculate_lj_ewald_comb_geom_F_E(__constant const float* nbfp_comb_climg2d,
453 const cl_nbparam_params_t* nbparam,
464 float c6grid, inv_r6_nm, cr2, expmcr2, poly, sh_mask;
466 c6grid = nbfp_comb_climg2d[2 * typei] * nbfp_comb_climg2d[2 * typej];
468 /* Recalculate inv_r6 without exclusion mask */
469 inv_r6_nm = inv_r2 * inv_r2 * inv_r2;
470 cr2 = lje_coeff2 * r2;
472 poly = 1.0f + cr2 + 0.5f * cr2 * cr2;
474 /* Subtract the grid force from the total LJ force */
475 *F_invr += c6grid * (inv_r6_nm - expmcr2 * (inv_r6_nm * poly + lje_coeff6_6)) * inv_r2;
477 /* Shift should be applied only to real LJ pairs */
478 sh_mask = nbparam->sh_lj_ewald * int_bit;
479 *E_lj += ONE_SIXTH_F * c6grid * (inv_r6_nm * (1.0f - expmcr2 * poly) + sh_mask);
482 /*! Calculate LJ-PME grid force + energy contribution (if E_lj != NULL) with
483 * Lorentz-Berthelot combination rule.
484 * We use a single F+E kernel with conditional because the performance impact
485 * of this is pretty small and LB on the CPU is anyway very slow.
487 gmx_opencl_inline void calculate_lj_ewald_comb_LB_F_E(__constant const float* nbfp_comb_climg2d,
488 const cl_nbparam_params_t* nbparam,
500 float c6grid, inv_r6_nm, cr2, expmcr2, poly;
501 float sigma, sigma2, epsilon;
503 /* sigma and epsilon are scaled to give 6*C6 */
504 sigma = nbfp_comb_climg2d[2 * typei] + nbfp_comb_climg2d[2 * typej];
506 epsilon = nbfp_comb_climg2d[2 * typei + 1] * nbfp_comb_climg2d[2 * typej + 1];
508 sigma2 = sigma * sigma;
509 c6grid = epsilon * sigma2 * sigma2 * sigma2;
511 /* Recalculate inv_r6 without exclusion mask */
512 inv_r6_nm = inv_r2 * inv_r2 * inv_r2;
513 cr2 = lje_coeff2 * r2;
515 poly = 1.0f + cr2 + 0.5f * cr2 * cr2;
517 /* Subtract the grid force from the total LJ force */
518 *F_invr += c6grid * (inv_r6_nm - expmcr2 * (inv_r6_nm * poly + lje_coeff6_6)) * inv_r2;
524 /* Shift should be applied only to real LJ pairs */
525 sh_mask = nbparam->sh_lj_ewald * int_bit;
526 *E_lj += ONE_SIXTH_F * c6grid * (inv_r6_nm * (1.0f - expmcr2 * poly) + sh_mask);
530 /*! Interpolate Ewald coulomb force using the table through the tex_nbfp texture.
531 * Original idea: from the OpenMM project
533 gmx_opencl_inline float interpolate_coulomb_force_r(__constant const float* coulomb_tab_climg2d,
537 float normalized = scale * r;
538 int index = (int)normalized;
539 float fract2 = normalized - index;
540 float fract1 = 1.0f - fract2;
542 return fract1 * coulomb_tab_climg2d[index] + fract2 * coulomb_tab_climg2d[index + 1];
545 /*! Calculate analytical Ewald correction term. */
546 gmx_opencl_inline float pmecorrF(float z2)
548 const float FN6 = -1.7357322914161492954e-8f;
549 const float FN5 = 1.4703624142580877519e-6f;
550 const float FN4 = -0.000053401640219807709149f;
551 const float FN3 = 0.0010054721316683106153f;
552 const float FN2 = -0.019278317264888380590f;
553 const float FN1 = 0.069670166153766424023f;
554 const float FN0 = -0.75225204789749321333f;
556 const float FD4 = 0.0011193462567257629232f;
557 const float FD3 = 0.014866955030185295499f;
558 const float FD2 = 0.11583842382862377919f;
559 const float FD1 = 0.50736591960530292870f;
560 const float FD0 = 1.0f;
563 float polyFN0, polyFN1, polyFD0, polyFD1;
567 polyFD0 = FD4 * z4 + FD2;
568 polyFD1 = FD3 * z4 + FD1;
569 polyFD0 = polyFD0 * z4 + FD0;
570 polyFD0 = polyFD1 * z2 + polyFD0;
572 polyFD0 = 1.0f / polyFD0;
574 polyFN0 = FN6 * z4 + FN4;
575 polyFN1 = FN5 * z4 + FN3;
576 polyFN0 = polyFN0 * z4 + FN2;
577 polyFN1 = polyFN1 * z4 + FN1;
578 polyFN0 = polyFN0 * z4 + FN0;
579 polyFN0 = polyFN1 * z2 + polyFN0;
581 return polyFN0 * polyFD0;
585 gmx_opencl_inline void
586 reduce_force_j_shfl(float3 fin, __global float* fout, int gmx_unused tidxi, int gmx_unused tidxj, int aidx)
588 /* Only does reduction over 4 elements in cluster. Needs to be changed
589 * for CL_SIZE>4. See CUDA code for required code */
590 fin.x += intel_sub_group_shuffle_down(fin.x, fin.x, 1);
591 fin.y += intel_sub_group_shuffle_up(fin.y, fin.y, 1);
592 fin.z += intel_sub_group_shuffle_down(fin.z, fin.z, 1);
593 if ((tidxi & 1) == 1)
597 fin.x += intel_sub_group_shuffle_down(fin.x, fin.x, 2);
598 fin.z += intel_sub_group_shuffle_up(fin.z, fin.z, 2);
605 atomicAdd_g_f(&fout[3 * aidx + tidxi], fin.x);
610 gmx_opencl_inline void
611 reduce_force_j_generic(__local float* f_buf, float3 fcj_buf, __global float* fout, int tidxi, int tidxj, int aidx)
613 int tidx = tidxi + tidxj * CL_SIZE;
614 f_buf[tidx] = fcj_buf.x;
615 f_buf[FBUF_STRIDE + tidx] = fcj_buf.y;
616 f_buf[2 * FBUF_STRIDE + tidx] = fcj_buf.z;
618 /* Split the reduction between the first 3 column threads
619 Threads with column id 0 will do the reduction for (float3).x components
620 Threads with column id 1 will do the reduction for (float3).y components
621 Threads with column id 2 will do the reduction for (float3).z components.
622 The reduction is performed for each line tidxj of f_buf. */
626 for (int j = tidxj * CL_SIZE; j < (tidxj + 1) * CL_SIZE; j++)
628 f += f_buf[FBUF_STRIDE * tidxi + j];
631 atomicAdd_g_f(&fout[3 * aidx + tidxi], f);
635 /*! Final j-force reduction
637 gmx_opencl_inline void reduce_force_j(__local float gmx_unused* f_buf,
639 __global float* fout,
645 reduce_force_j_shfl(fcj_buf, fout, tidxi, tidxj, aidx);
647 reduce_force_j_generic(f_buf, fcj_buf, fout, tidxi, tidxj, aidx);
652 gmx_opencl_inline void reduce_force_i_and_shift_shfl(float3* fci_buf,
653 __global float* fout,
659 __global float* fshift)
661 /* Only does reduction over 4 elements in cluster (2 per warp). Needs to be changed
663 float2 fshift_buf = 0;
664 for (int ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
666 int aidx = (sci * NCL_PER_SUPERCL + ci_offset) * CL_SIZE + tidxi;
667 float3 fin = fci_buf[ci_offset];
668 fin.x += intel_sub_group_shuffle_down(fin.x, fin.x, CL_SIZE);
669 fin.y += intel_sub_group_shuffle_up(fin.y, fin.y, CL_SIZE);
670 fin.z += intel_sub_group_shuffle_down(fin.z, fin.z, CL_SIZE);
676 /* Threads 0,1 and 2,3 increment x,y for their warp */
677 atomicAdd_g_f(&fout[3 * aidx + (tidxj & 1)], fin.x);
680 fshift_buf[0] += fin.x;
682 /* Threads 0 and 2 increment z for their warp */
683 if ((tidxj & 1) == 0)
685 atomicAdd_g_f(&fout[3 * aidx + 2], fin.z);
688 fshift_buf[1] += fin.z;
692 /* add up local shift forces into global mem */
695 // Threads 0,1 and 2,3 update x,y
696 atomicAdd_g_f(&(fshift[3 * shift + (tidxj & 1)]), fshift_buf[0]);
697 // Threads 0 and 2 update z
698 if ((tidxj & 1) == 0)
700 atomicAdd_g_f(&(fshift[3 * shift + 2]), fshift_buf[1]);
706 /*! Final i-force reduction; this implementation works only with power of two
709 gmx_opencl_inline void reduce_force_i_and_shift_pow2(volatile __local float* f_buf,
711 __global float* fout,
717 __global float* fshift)
719 float fshift_buf = 0;
720 for (int ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
722 int aidx = (sci * NCL_PER_SUPERCL + ci_offset) * CL_SIZE + tidxi;
723 int tidx = tidxi + tidxj * CL_SIZE;
724 /* store i forces in shmem */
725 f_buf[tidx] = fci_buf[ci_offset].x;
726 f_buf[FBUF_STRIDE + tidx] = fci_buf[ci_offset].y;
727 f_buf[2 * FBUF_STRIDE + tidx] = fci_buf[ci_offset].z;
728 barrier(CLK_LOCAL_MEM_FENCE);
731 /* Reduce the initial CL_SIZE values for each i atom to half
732 * every step by using CL_SIZE * i threads.
733 * Can't just use i as loop variable because than nvcc refuses to unroll.
736 for (j = CL_SIZE_LOG2 - 1; j > 0; j--)
741 f_buf[tidxj * CL_SIZE + tidxi] += f_buf[(tidxj + i) * CL_SIZE + tidxi];
742 f_buf[FBUF_STRIDE + tidxj * CL_SIZE + tidxi] +=
743 f_buf[FBUF_STRIDE + (tidxj + i) * CL_SIZE + tidxi];
744 f_buf[2 * FBUF_STRIDE + tidxj * CL_SIZE + tidxi] +=
745 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
789 gmx_opencl_inline void reduce_force_i_and_shift(__local float gmx_unused* f_buf,
797 __global float* fshift)
800 reduce_force_i_and_shift_shfl(fci_buf, f, bCalcFshift, tidxi, tidxj, sci, shift, fshift);
802 reduce_force_i_and_shift_pow2(f_buf, fci_buf, f, bCalcFshift, tidxi, tidxj, sci, shift, fshift);
808 gmx_opencl_inline void reduce_energy_shfl(float E_lj,
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
832 gmx_opencl_inline void reduce_energy_pow2(volatile __local float* buf,
833 volatile __global float* e_lj,
834 volatile __global float* e_el,
839 unsigned int i = WARP_SIZE / 2;
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 float e1 = buf[tidx] + buf[tidx + i];
856 float e2 = buf[FBUF_STRIDE + tidx] + buf[FBUF_STRIDE + tidx + i];
858 atomicAdd_g_f(e_lj, e1);
859 atomicAdd_g_f(e_el, e2);
863 gmx_opencl_inline void reduce_energy(volatile __local float gmx_unused* buf,
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 gmx_opencl_inline 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 gmx_opencl_inline bool gmx_sub_group_any(volatile __local uint gmx_unused* warp_any,
899 # if USE_SUBGROUP_ANY
900 return sub_group_any(pred);
902 return gmx_sub_group_any_localmem(warp_any, widx, pred);
906 #endif /* NBNXN_OPENCL_KERNEL_UTILS_CLH */