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.
39 #include "gromacs/gpu_utils/device_utils.clh"
40 #include "gromacs/gpu_utils/vectype_ops.clh"
41 #include "gromacs/nbnxm/constants.h"
42 #include "gromacs/pbcutil/ishift.h"
44 #include "nbnxm_ocl_consts.h"
46 #define CL_SIZE (NBNXN_GPU_CLUSTER_SIZE)
47 #define NCL_PER_SUPERCL c_nbnxnGpuNumClusterPerSupercluster
49 #define WARP_SIZE (CL_SIZE * CL_SIZE / 2) // Currently only c_nbnxnGpuClusterpairSplit=2 supported
51 #if defined _NVIDIA_SOURCE_ || defined _AMD_SOURCE_
52 /* Currently we enable CJ prefetch for AMD/NVIDIA and disable it for other vendors
53 * Note that this should precede the kernel_utils include.
55 # define USE_CJ_PREFETCH 1
57 # define USE_CJ_PREFETCH 0
60 #if defined cl_intel_subgroups || defined cl_khr_subgroups \
61 || (defined __OPENCL_VERSION__ && __OPENCL_VERSION__ >= 210)
62 # define HAVE_SUBGROUP 1
64 # define HAVE_SUBGROUP 0
67 #ifdef cl_intel_subgroups
68 # define HAVE_INTEL_SUBGROUP 1
70 # define HAVE_INTEL_SUBGROUP 0
73 #if defined _INTEL_SOURCE_
74 # define SUBGROUP_SIZE 8
75 #elif defined _AMD_SOURCE_
76 # define SUBGROUP_SIZE 64
78 # define SUBGROUP_SIZE 32
81 #define REDUCE_SHUFFLE (HAVE_INTEL_SUBGROUP && CL_SIZE == 4 && SUBGROUP_SIZE == WARP_SIZE)
82 #define USE_SUBGROUP_ANY (HAVE_SUBGROUP && SUBGROUP_SIZE == WARP_SIZE)
83 #define USE_SUBGROUP_PRELOAD HAVE_INTEL_SUBGROUP
85 /* 1.0 / sqrt(M_PI) */
86 #define M_FLOAT_1_SQRTPI 0.564189583547756F
90 #ifndef NBNXN_OPENCL_KERNEL_UTILS_CLH
91 # define NBNXN_OPENCL_KERNEL_UTILS_CLH
94 # define WARP_SIZE_LOG2 (5)
95 # define CL_SIZE_LOG2 (3)
97 # define WARP_SIZE_LOG2 (3)
98 # define CL_SIZE_LOG2 (2)
100 # error unsupported CL_SIZE
103 # define CL_SIZE_SQ (CL_SIZE * CL_SIZE)
104 # define FBUF_STRIDE (CL_SIZE_SQ)
106 # define ONE_SIXTH_F 0.16666667F
107 # define ONE_TWELVETH_F 0.08333333F
112 /* GCC, clang, and some ICC pretending to be GCC */
113 # define gmx_unused __attribute__((unused))
118 // Data structures shared between OpenCL device code and OpenCL host code
119 // TODO: review, improve
120 // Replaced real by float for now, to avoid including any other header
128 /* Used with potential switching:
129 * rsw = max(r - r_switch, 0)
130 * sw = 1 + c3*rsw^3 + c4*rsw^4 + c5*rsw^5
131 * dsw = 3*c3*rsw^2 + 4*c4*rsw^3 + 5*c5*rsw^4
132 * force = force*dsw - potential*sw
142 // Data structure shared between the OpenCL device code and OpenCL host code
143 // Must not contain OpenCL objects (buffers)
144 typedef struct cl_nbparam_params
147 int eeltype; /**< type of electrostatics, takes values from #eelCu */
148 int vdwtype; /**< type of VdW impl., takes values from #evdwCu */
150 float epsfac; /**< charge multiplication factor */
151 float c_rf; /**< Reaction-field/plain cutoff electrostatics const. */
152 float two_k_rf; /**< Reaction-field electrostatics constant */
153 float ewald_beta; /**< Ewald/PME parameter */
154 float sh_ewald; /**< Ewald/PME correction term substracted from the direct-space potential */
155 float sh_lj_ewald; /**< LJ-Ewald/PME correction term added to the correction potential */
156 float ewaldcoeff_lj; /**< LJ-Ewald/PME coefficient */
158 float rcoulomb_sq; /**< Coulomb cut-off squared */
160 float rvdw_sq; /**< VdW cut-off squared */
161 float rvdw_switch; /**< VdW switched cut-off */
162 float rlistOuter_sq; /**< Full, outer pair-list cut-off squared */
163 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 */
165 shift_consts_t dispersion_shift; /**< VdW shift dispersion constants */
166 shift_consts_t repulsion_shift; /**< VdW shift repulsion constants */
167 switch_consts_t vdw_switch; /**< VdW switch constants */
169 /* Ewald Coulomb force table data - accessed through texture memory */
170 float coulomb_tab_scale; /**< table scale/spacing */
171 } cl_nbparam_params_t;
175 int sci; /* i-super-cluster */
176 int shift; /* Shift vector index plus possible flags */
177 int cj4_ind_start; /* Start index into cj4 */
178 int cj4_ind_end; /* End index into cj4 */
183 unsigned int imask; /* The i-cluster interactions mask for 1 warp */
184 int excl_ind; /* Index into the exclusion array for 1 warp */
189 int cj[4]; /* The 4 j-clusters */
190 nbnxn_im_ei_t imei[2]; /* The i-cluster mask data for 2 warps */
196 unsigned int pair[CL_SIZE * CL_SIZE / 2]; /* Topology exclusion interaction bits for one warp,
197 * each unsigned has bitS for 4*8 i clusters
201 /*! i-cluster interaction mask for a super-cluster with all NCL_PER_SUPERCL bits set */
202 __constant unsigned supercl_interaction_mask = ((1U << NCL_PER_SUPERCL) - 1U);
204 gmx_opencl_inline void preloadCj4Generic(__local int* sm_cjPreload,
205 const __global int* gm_cj,
208 bool gmx_unused iMaskCond)
210 /* Pre-load cj into shared memory */
211 # if defined _AMD_SOURCE_ // TODO: fix by setting c_nbnxnGpuClusterpairSplit properly
212 if (tidxj == 0 & tidxi < c_nbnxnGpuJgroupSize)
214 sm_cjPreload[tidxi] = gm_cj[tidxi];
217 const int c_clSize = CL_SIZE;
218 const int c_nbnxnGpuClusterpairSplit = 2;
219 const int c_splitClSize = c_clSize / c_nbnxnGpuClusterpairSplit;
220 if ((tidxj == 0 | tidxj == c_splitClSize) & (tidxi < c_nbnxnGpuJgroupSize))
222 sm_cjPreload[tidxi + tidxj * c_nbnxnGpuJgroupSize / c_splitClSize] = gm_cj[tidxi];
228 # if USE_SUBGROUP_PRELOAD
229 gmx_opencl_inline int preloadCj4Subgroup(const __global int* gm_cj)
231 // loads subgroup-size # of elements (8) instead of the 4 required
232 // equivalent to *cjs = *gm_cj
233 return intel_sub_group_block_read((const __global uint*)gm_cj);
235 # endif // USE_SUBGROUP_PRELOAD
237 # if USE_SUBGROUP_PRELOAD
238 typedef size_t CjType;
240 typedef __local int* CjType;
243 /*! \brief Preload cj4
245 * - For AMD we load once for a wavefront of 64 threads (on 4 threads * NTHREAD_Z)
246 * - For NVIDIA once per warp (on 2x4 threads * NTHREAD_Z)
247 * - For Intel(/USE_SUBGROUP_PRELOAD) loads into private memory(/register) instead of local memory
249 * It is the caller's responsibility to make sure that data is consumed only when
250 * it's ready. This function does not call a barrier.
252 gmx_opencl_inline void preloadCj4(CjType gmx_unused* cjs,
253 const __global int gmx_unused* gm_cj,
254 int gmx_unused tidxi,
255 int gmx_unused tidxj,
256 bool gmx_unused iMaskCond)
258 # if USE_SUBGROUP_PRELOAD
259 *cjs = preloadCj4Subgroup(gm_cj);
260 # elif USE_CJ_PREFETCH
261 preloadCj4Generic(*cjs, gm_cj, tidxi, tidxj, iMaskCond);
267 gmx_opencl_inline int loadCjPreload(__local int* sm_cjPreload, int jm, int gmx_unused tidxi, int gmx_unused tidxj)
269 # if defined _AMD_SOURCE_
270 int warpLoadOffset = 0; // TODO: fix by setting c_nbnxnGpuClusterpairSplit properly
272 const int c_clSize = CL_SIZE;
273 const int c_nbnxnGpuClusterpairSplit = 2;
274 const int c_splitClSize = c_clSize / c_nbnxnGpuClusterpairSplit;
275 int warpLoadOffset = (tidxj & c_splitClSize) * c_nbnxnGpuJgroupSize / c_splitClSize;
277 return sm_cjPreload[jm + warpLoadOffset];
280 /* \brief Load a cj given a jm index.
282 * If cj4 preloading is enabled, it loads from the local memory, otherwise from global.
284 gmx_opencl_inline int
285 loadCj(CjType cjs, const __global int gmx_unused* gm_cj, int jm, int gmx_unused tidxi, int gmx_unused tidxj)
287 # if USE_SUBGROUP_PRELOAD
288 return sub_group_broadcast(cjs, jm);
289 # elif USE_CJ_PREFETCH
290 return loadCjPreload(cjs, jm, tidxi, tidxj);
296 /*! Convert LJ sigma,epsilon parameters to C6,C12. */
297 gmx_opencl_inline void convert_sigma_epsilon_to_c6_c12(const float sigma, const float epsilon, float* c6, float* c12)
299 float sigma2, sigma6;
301 sigma2 = sigma * sigma;
302 sigma6 = sigma2 * sigma2 * sigma2;
303 *c6 = epsilon * sigma6;
308 /*! Apply force switch, force + energy version. */
309 gmx_opencl_inline void calculate_force_switch_F(const cl_nbparam_params_t* nbparam,
318 /* force switch constants */
319 float disp_shift_V2 = nbparam->dispersion_shift.c2;
320 float disp_shift_V3 = nbparam->dispersion_shift.c3;
321 float repu_shift_V2 = nbparam->repulsion_shift.c2;
322 float repu_shift_V3 = nbparam->repulsion_shift.c3;
325 r_switch = r - nbparam->rvdw_switch;
326 r_switch = r_switch >= 0.0F ? r_switch : 0.0F;
328 *F_invr += -c6 * (disp_shift_V2 + disp_shift_V3 * r_switch) * r_switch * r_switch * inv_r
329 + c12 * (repu_shift_V2 + repu_shift_V3 * r_switch) * r_switch * r_switch * inv_r;
332 /*! Apply force switch, force-only version. */
333 gmx_opencl_inline void calculate_force_switch_F_E(const 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;
358 *F_invr += -c6 * (disp_shift_V2 + disp_shift_V3 * r_switch) * r_switch * r_switch * inv_r
359 + c12 * (repu_shift_V2 + repu_shift_V3 * r_switch) * r_switch * r_switch * inv_r;
360 *E_lj += 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. */
365 gmx_opencl_inline void calculate_potential_switch_F(const cl_nbparam_params_t* nbparam,
374 /* potential switch constants */
375 float switch_V3 = nbparam->vdw_switch.c3;
376 float switch_V4 = nbparam->vdw_switch.c4;
377 float switch_V5 = nbparam->vdw_switch.c5;
378 float switch_F2 = nbparam->vdw_switch.c3;
379 float switch_F3 = nbparam->vdw_switch.c4;
380 float switch_F4 = nbparam->vdw_switch.c5;
383 r_switch = r - nbparam->rvdw_switch;
385 /* Unlike in the F+E kernel, conditional is faster here */
388 sw = 1.0F + (switch_V3 + (switch_V4 + switch_V5 * r_switch) * r_switch) * r_switch * r_switch * r_switch;
389 dsw = (switch_F2 + (switch_F3 + switch_F4 * r_switch) * r_switch) * r_switch * r_switch;
391 *F_invr = (*F_invr) * sw - inv_r * (*E_lj) * dsw;
395 /*! Apply potential switch, force + energy version. */
396 gmx_opencl_inline void calculate_potential_switch_F_E(const cl_nbparam_params_t* nbparam,
405 /* potential switch constants */
406 float switch_V3 = nbparam->vdw_switch.c3;
407 float switch_V4 = nbparam->vdw_switch.c4;
408 float switch_V5 = nbparam->vdw_switch.c5;
409 float switch_F2 = nbparam->vdw_switch.c3;
410 float switch_F3 = nbparam->vdw_switch.c4;
411 float switch_F4 = nbparam->vdw_switch.c5;
414 r_switch = r - nbparam->rvdw_switch;
415 r_switch = r_switch >= 0.0F ? r_switch : 0.0F;
417 /* Unlike in the F-only kernel, masking is faster here */
418 sw = 1.0F + (switch_V3 + (switch_V4 + switch_V5 * r_switch) * r_switch) * r_switch * r_switch * r_switch;
419 dsw = (switch_F2 + (switch_F3 + switch_F4 * r_switch) * r_switch) * r_switch * r_switch;
421 *F_invr = (*F_invr) * sw - inv_r * (*E_lj) * dsw;
425 /*! Calculate LJ-PME grid force contribution with
426 * geometric combination rule.
428 gmx_opencl_inline void calculate_lj_ewald_comb_geom_F(__constant const float* nbfp_comb_climg2d,
437 float c6grid, inv_r6_nm, cr2, expmcr2, poly;
439 c6grid = nbfp_comb_climg2d[2 * typei] * nbfp_comb_climg2d[2 * typej];
441 /* Recalculate inv_r6 without exclusion mask */
442 inv_r6_nm = inv_r2 * inv_r2 * inv_r2;
443 cr2 = lje_coeff2 * r2;
445 poly = 1.0F + cr2 + HALF_F * cr2 * cr2;
447 /* Subtract the grid force from the total LJ force */
448 *F_invr += c6grid * (inv_r6_nm - expmcr2 * (inv_r6_nm * poly + lje_coeff6_6)) * inv_r2;
451 /*! Calculate LJ-PME grid force + energy contribution with
452 * geometric combination rule.
454 gmx_opencl_inline void calculate_lj_ewald_comb_geom_F_E(__constant const float* nbfp_comb_climg2d,
455 const cl_nbparam_params_t* nbparam,
466 float c6grid, inv_r6_nm, cr2, expmcr2, poly, sh_mask;
468 c6grid = nbfp_comb_climg2d[2 * typei] * nbfp_comb_climg2d[2 * typej];
470 /* Recalculate inv_r6 without exclusion mask */
471 inv_r6_nm = inv_r2 * inv_r2 * inv_r2;
472 cr2 = lje_coeff2 * r2;
474 poly = 1.0F + cr2 + HALF_F * cr2 * cr2;
476 /* Subtract the grid force from the total LJ force */
477 *F_invr += c6grid * (inv_r6_nm - expmcr2 * (inv_r6_nm * poly + lje_coeff6_6)) * inv_r2;
479 /* Shift should be applied only to real LJ pairs */
480 sh_mask = nbparam->sh_lj_ewald * int_bit;
481 *E_lj += ONE_SIXTH_F * c6grid * (inv_r6_nm * (1.0F - expmcr2 * poly) + sh_mask);
484 /*! Calculate LJ-PME grid force + energy contribution (if E_lj != NULL) with
485 * Lorentz-Berthelot combination rule.
486 * We use a single F+E kernel with conditional because the performance impact
487 * of this is pretty small and LB on the CPU is anyway very slow.
489 gmx_opencl_inline void calculate_lj_ewald_comb_LB_F_E(__constant const float* nbfp_comb_climg2d,
490 const cl_nbparam_params_t* nbparam,
502 float c6grid, inv_r6_nm, cr2, expmcr2, poly;
503 float sigma, sigma2, epsilon;
505 /* sigma and epsilon are scaled to give 6*C6 */
506 sigma = nbfp_comb_climg2d[2 * typei] + nbfp_comb_climg2d[2 * typej];
508 epsilon = nbfp_comb_climg2d[2 * typei + 1] * nbfp_comb_climg2d[2 * typej + 1];
510 sigma2 = sigma * sigma;
511 c6grid = epsilon * sigma2 * sigma2 * sigma2;
513 /* Recalculate inv_r6 without exclusion mask */
514 inv_r6_nm = inv_r2 * inv_r2 * inv_r2;
515 cr2 = lje_coeff2 * r2;
517 poly = 1.0F + cr2 + HALF_F * cr2 * cr2;
519 /* Subtract the grid force from the total LJ force */
520 *F_invr += c6grid * (inv_r6_nm - expmcr2 * (inv_r6_nm * poly + lje_coeff6_6)) * inv_r2;
526 /* Shift should be applied only to real LJ pairs */
527 sh_mask = nbparam->sh_lj_ewald * int_bit;
528 *E_lj += ONE_SIXTH_F * c6grid * (inv_r6_nm * (1.0F - expmcr2 * poly) + sh_mask);
532 /*! Interpolate Ewald coulomb force using the table through the tex_nbfp texture.
533 * Original idea: from the OpenMM project
535 gmx_opencl_inline float interpolate_coulomb_force_r(__constant const float* coulomb_tab_climg2d,
539 float normalized = scale * r;
540 int index = (int)normalized;
541 float fract2 = normalized - (float)index;
542 float fract1 = 1.0F - fract2;
544 return fract1 * coulomb_tab_climg2d[index] + fract2 * coulomb_tab_climg2d[index + 1];
547 /*! Calculate analytical Ewald correction term. */
548 gmx_opencl_inline float pmecorrF(float z2)
550 const float FN6 = -1.7357322914161492954e-8F;
551 const float FN5 = 1.4703624142580877519e-6F;
552 const float FN4 = -0.000053401640219807709149F;
553 const float FN3 = 0.0010054721316683106153F;
554 const float FN2 = -0.019278317264888380590F;
555 const float FN1 = 0.069670166153766424023F;
556 const float FN0 = -0.75225204789749321333F;
558 const float FD4 = 0.0011193462567257629232F;
559 const float FD3 = 0.014866955030185295499F;
560 const float FD2 = 0.11583842382862377919F;
561 const float FD1 = 0.50736591960530292870F;
562 const float FD0 = 1.0F;
565 float polyFN0, polyFN1, polyFD0, polyFD1;
569 polyFD0 = FD4 * z4 + FD2;
570 polyFD1 = FD3 * z4 + FD1;
571 polyFD0 = polyFD0 * z4 + FD0;
572 polyFD0 = polyFD1 * z2 + polyFD0;
574 polyFD0 = 1.0F / polyFD0;
576 polyFN0 = FN6 * z4 + FN4;
577 polyFN1 = FN5 * z4 + FN3;
578 polyFN0 = polyFN0 * z4 + FN2;
579 polyFN1 = polyFN1 * z4 + FN1;
580 polyFN0 = polyFN0 * z4 + FN0;
581 polyFN0 = polyFN1 * z2 + polyFN0;
583 return polyFN0 * polyFD0;
587 gmx_opencl_inline void
588 reduce_force_j_shfl(float3 fin, __global float* fout, int gmx_unused tidxi, int gmx_unused tidxj, int aidx)
590 /* Only does reduction over 4 elements in cluster. Needs to be changed
591 * for CL_SIZE>4. See CUDA code for required code */
592 fin.x += intel_sub_group_shuffle_down(fin.x, fin.x, 1);
593 fin.y += intel_sub_group_shuffle_up(fin.y, fin.y, 1);
594 fin.z += intel_sub_group_shuffle_down(fin.z, fin.z, 1);
595 if ((tidxi & 1) == 1)
599 fin.x += intel_sub_group_shuffle_down(fin.x, fin.x, 2);
600 fin.z += intel_sub_group_shuffle_up(fin.z, fin.z, 2);
607 atomicAdd_g_f(&fout[3 * aidx + tidxi], fin.x);
612 gmx_opencl_inline void
613 reduce_force_j_generic(__local float* f_buf, float3 fcj_buf, __global float* fout, int tidxi, int tidxj, int aidx)
615 int tidx = tidxi + tidxj * CL_SIZE;
616 f_buf[tidx] = fcj_buf.x;
617 f_buf[FBUF_STRIDE + tidx] = fcj_buf.y;
618 f_buf[2 * FBUF_STRIDE + tidx] = fcj_buf.z;
620 /* Split the reduction between the first 3 column threads
621 Threads with column id 0 will do the reduction for (float3).x components
622 Threads with column id 1 will do the reduction for (float3).y components
623 Threads with column id 2 will do the reduction for (float3).z components.
624 The reduction is performed for each line tidxj of f_buf. */
628 for (int j = tidxj * CL_SIZE; j < (tidxj + 1) * CL_SIZE; j++)
630 f += f_buf[FBUF_STRIDE * tidxi + j];
633 atomicAdd_g_f(&fout[3 * aidx + tidxi], f);
637 /*! Final j-force reduction
639 gmx_opencl_inline void reduce_force_j(__local float gmx_unused* f_buf,
641 __global float* fout,
647 reduce_force_j_shfl(fcj_buf, fout, tidxi, tidxj, aidx);
649 reduce_force_j_generic(f_buf, fcj_buf, fout, tidxi, tidxj, aidx);
654 gmx_opencl_inline void reduce_force_i_and_shift_shfl(float3* fci_buf,
655 __global float* fout,
661 __global float* fshift)
663 /* Only does reduction over 4 elements in cluster (2 per warp). Needs to be changed
665 float2 fshift_buf = 0;
666 for (int ci_offset = 0; ci_offset < NCL_PER_SUPERCL; ci_offset++)
668 int aidx = (sci * NCL_PER_SUPERCL + ci_offset) * CL_SIZE + tidxi;
669 float3 fin = fci_buf[ci_offset];
670 fin.x += intel_sub_group_shuffle_down(fin.x, fin.x, CL_SIZE);
671 fin.y += intel_sub_group_shuffle_up(fin.y, fin.y, CL_SIZE);
672 fin.z += intel_sub_group_shuffle_down(fin.z, fin.z, CL_SIZE);
678 /* Threads 0,1 and 2,3 increment x,y for their warp */
679 atomicAdd_g_f(&fout[3 * aidx + (tidxj & 1)], fin.x);
682 fshift_buf[0] += fin.x;
684 /* Threads 0 and 2 increment z for their warp */
685 if ((tidxj & 1) == 0)
687 atomicAdd_g_f(&fout[3 * aidx + 2], fin.z);
690 fshift_buf[1] += fin.z;
694 /* add up local shift forces into global mem */
697 // Threads 0,1 and 2,3 update x,y
698 atomicAdd_g_f(&(fshift[3 * shift + (tidxj & 1)]), fshift_buf[0]);
699 // Threads 0 and 2 update z
700 if ((tidxj & 1) == 0)
702 atomicAdd_g_f(&(fshift[3 * shift + 2]), fshift_buf[1]);
708 /*! Final i-force reduction; this implementation works only with power of two
711 gmx_opencl_inline void reduce_force_i_and_shift_pow2(volatile __local float* f_buf,
713 __global float* fout,
719 __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);
732 /* Reduce the initial CL_SIZE values for each i atom to half
733 * every step by using CL_SIZE * i threads.
734 * Can't just use i as loop variable because than nvcc refuses to unroll.
737 for (int j = CL_SIZE_LOG2 - 1; j > 0; j--)
742 f_buf[tidxj * CL_SIZE + tidxi] += f_buf[(tidxj + i) * CL_SIZE + tidxi];
743 f_buf[FBUF_STRIDE + tidxj * CL_SIZE + tidxi] +=
744 f_buf[FBUF_STRIDE + (tidxj + i) * CL_SIZE + tidxi];
745 f_buf[2 * FBUF_STRIDE + tidxj * CL_SIZE + tidxi] +=
746 f_buf[2 * FBUF_STRIDE + (tidxj + i) * CL_SIZE + tidxi];
751 * a) for CL_SIZE<8: id 2 (doing z in next block) is in 2nd warp
752 * b) for all CL_SIZE a barrier is needed before f_buf is reused by next reduce_force_i call
753 * TODO: Test on Nvidia for performance difference between having the barrier here or after the atomicAdd
755 barrier(CLK_LOCAL_MEM_FENCE);
757 /* i == 1, last reduction step, writing to global mem */
758 /* Split the reduction between the first 3 line threads
759 Threads with line id 0 will do the reduction for (float3).x components
760 Threads with line id 1 will do the reduction for (float3).y components
761 Threads with line id 2 will do the reduction for (float3).z components. */
764 float f = f_buf[tidxj * FBUF_STRIDE + tidxi] + f_buf[tidxj * FBUF_STRIDE + i * CL_SIZE + tidxi];
766 atomicAdd_g_f(&fout[3 * aidx + tidxj], f);
774 /* add up local shift forces into global mem */
777 /* Only threads with tidxj < 3 will update fshift.
778 The threads performing the update, must be the same as the threads
779 storing the reduction result above.
783 atomicAdd_g_f(&(fshift[3 * shift + tidxj]), fshift_buf);
788 /*! Final i-force reduction
790 gmx_opencl_inline void reduce_force_i_and_shift(__local float gmx_unused* f_buf,
798 __global float* fshift)
801 reduce_force_i_and_shift_shfl(fci_buf, f, bCalcFshift, tidxi, tidxj, sci, shift, fshift);
803 reduce_force_i_and_shift_pow2(f_buf, fci_buf, f, bCalcFshift, tidxi, tidxj, sci, shift, fshift);
809 gmx_opencl_inline void reduce_energy_shfl(float E_lj,
811 volatile __global float* e_lj,
812 volatile __global float* e_el,
815 E_lj = sub_group_reduce_add(E_lj);
816 E_el = sub_group_reduce_add(E_el);
817 /* Should be get_sub_group_local_id()==0. Doesn't work with Intel Classic driver.
818 * To make it possible to use REDUCE_SHUFFLE with single subgroup per i-j pair
819 * (e.g. subgroup size 16 with CL_SIZE 4), either this "if" needs to be changed or
820 * the definition of WARP_SIZE (currently CL_SIZE*CL_SIZE/2) needs to be changed
821 * (by supporting c_nbnxnGpuClusterpairSplit=1). */
822 if (tidx == 0 || tidx == WARP_SIZE)
824 atomicAdd_g_f(e_lj, E_lj);
825 atomicAdd_g_f(e_el, E_el);
830 /*! Energy reduction; this implementation works only with power of two
833 gmx_opencl_inline void reduce_energy_pow2(volatile __local float* buf,
834 volatile __global float* e_lj,
835 volatile __global float* e_el,
840 int i = WARP_SIZE / 2;
842 /* Can't just use i as loop variable because than nvcc refuses to unroll. */
843 for (j = WARP_SIZE_LOG2 - 1; j > 0; j--)
847 buf[tidx] += buf[tidx + i];
848 buf[FBUF_STRIDE + tidx] += buf[FBUF_STRIDE + tidx + i];
853 /* last reduction step, writing to global mem */
856 float e1 = buf[tidx] + buf[tidx + i];
857 float e2 = buf[FBUF_STRIDE + tidx] + buf[FBUF_STRIDE + tidx + i];
859 atomicAdd_g_f(e_lj, e1);
860 atomicAdd_g_f(e_el, e2);
864 gmx_opencl_inline void reduce_energy(volatile __local float gmx_unused* buf,
867 volatile __global float* e_lj,
868 volatile __global float* e_el,
872 reduce_energy_shfl(E_lj, E_el, e_lj, e_el, tidx);
874 /* flush the energies to shmem and reduce them */
876 buf[FBUF_STRIDE + tidx] = E_el;
877 reduce_energy_pow2(buf + (tidx & WARP_SIZE), e_lj, e_el, tidx & ~WARP_SIZE);
881 gmx_opencl_inline bool gmx_sub_group_any_localmem(volatile __local int* warp_any, int widx, bool pred)
888 bool ret = warp_any[widx];
895 //! Returns a true if predicate is true for any work item in warp
896 gmx_opencl_inline bool gmx_sub_group_any(volatile __local int gmx_unused* warp_any, int gmx_unused widx, bool pred)
898 # if USE_SUBGROUP_ANY
899 return sub_group_any(pred);
901 return gmx_sub_group_any_localmem(warp_any, widx, pred);
905 #endif /* NBNXN_OPENCL_KERNEL_UTILS_CLH */