2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2016,2017 by the GROMACS development team.
5 * Copyright (c) 2018,2019,2020, by the GROMACS development team, led by
6 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
7 * and including many others, as listed in the AUTHORS file in the
8 * top-level source directory and at http://www.gromacs.org.
10 * GROMACS is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public License
12 * as published by the Free Software Foundation; either version 2.1
13 * of the License, or (at your option) any later version.
15 * GROMACS is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with GROMACS; if not, see
22 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
23 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
25 * If you want to redistribute modifications to GROMACS, please
26 * consider that scientific software is very special. Version
27 * control is crucial - bugs must be traceable. We will be happy to
28 * consider code for inclusion in the official distribution, but
29 * derived work must not be called official GROMACS. Details are found
30 * in the README & COPYING files - if they are missing, get the
31 * official version at http://www.gromacs.org.
33 * To help us fund GROMACS development, we humbly ask that you cite
34 * the research papers on the package. Check out http://www.gromacs.org.
38 * Utility constant and function declaration for the OpenCL non-bonded kernels.
39 * This header should be included once at the top level, just before the
40 * kernels are included (has to be preceded by nbnxn_ocl_types.h).
42 * \author Szilárd Páll <pall.szilard@gmail.com>
43 * \ingroup module_nbnxm
48 #include "gromacs/gpu_utils/device_utils.clh"
49 #include "gromacs/gpu_utils/vectype_ops.clh"
50 #include "gromacs/pbcutil/ishift.h"
52 #include "nbnxm_ocl_consts.h"
54 #define CL_SIZE (c_nbnxnGpuClusterSize)
56 #define WARP_SIZE (CL_SIZE * CL_SIZE / 2) // Currently only c_nbnxnGpuClusterpairSplit=2 supported
58 #if defined _NVIDIA_SOURCE_ || defined _AMD_SOURCE_
59 /* Currently we enable CJ prefetch for AMD/NVIDIA and disable it for other vendors
60 * Note that this should precede the kernel_utils include.
62 # define USE_CJ_PREFETCH 1
64 # define USE_CJ_PREFETCH 0
67 #if defined cl_intel_subgroups || defined cl_khr_subgroups \
68 || (defined __OPENCL_VERSION__ && __OPENCL_VERSION__ >= 210)
69 # define HAVE_SUBGROUP 1
71 # define HAVE_SUBGROUP 0
74 #ifdef cl_intel_subgroups
75 # define HAVE_INTEL_SUBGROUP 1
77 # define HAVE_INTEL_SUBGROUP 0
80 #if defined _INTEL_SOURCE_
81 # define SUBGROUP_SIZE 8
82 #elif defined _AMD_SOURCE_
83 # define SUBGROUP_SIZE 64
85 # define SUBGROUP_SIZE 32
88 #define REDUCE_SHUFFLE (HAVE_INTEL_SUBGROUP && CL_SIZE == 4 && SUBGROUP_SIZE == WARP_SIZE)
89 #define USE_SUBGROUP_ANY (HAVE_SUBGROUP && SUBGROUP_SIZE == WARP_SIZE)
90 #define USE_SUBGROUP_PRELOAD HAVE_INTEL_SUBGROUP
92 /* 1.0 / sqrt(M_PI) */
93 #define M_FLOAT_1_SQRTPI 0.564189583547756F
97 #ifndef NBNXN_OPENCL_KERNEL_UTILS_CLH
98 # define NBNXN_OPENCL_KERNEL_UTILS_CLH
101 # define WARP_SIZE_LOG2 (5)
102 # define CL_SIZE_LOG2 (3)
104 # define WARP_SIZE_LOG2 (3)
105 # define CL_SIZE_LOG2 (2)
107 # error unsupported CL_SIZE
110 # define CL_SIZE_SQ (CL_SIZE * CL_SIZE)
111 # define FBUF_STRIDE (CL_SIZE_SQ)
113 # define ONE_SIXTH_F 0.16666667F
114 # define ONE_TWELVETH_F 0.08333333F
119 /* GCC, clang, and some ICC pretending to be GCC */
120 # define gmx_unused __attribute__((unused))
125 // Data structures shared between OpenCL device code and OpenCL host code
126 // TODO: review, improve
127 // Replaced real by float for now, to avoid including any other header
135 /* Used with potential switching:
136 * rsw = max(r - r_switch, 0)
137 * sw = 1 + c3*rsw^3 + c4*rsw^4 + c5*rsw^5
138 * dsw = 3*c3*rsw^2 + 4*c4*rsw^3 + 5*c5*rsw^4
139 * force = force*dsw - potential*sw
149 // Data structure shared between the OpenCL device code and OpenCL host code
150 // Must not contain OpenCL objects (buffers)
151 typedef struct cl_nbparam_params
154 //! type of electrostatics, takes values from #eelCu
156 //! type of VdW impl., takes values from #evdwCu
159 //! charge multiplication factor
161 //! Reaction-field/plain cutoff electrostatics const.
163 //! Reaction-field electrostatics constant
165 //! Ewald/PME parameter
167 //! Ewald/PME correction term substracted from the direct-space potential
169 //! LJ-Ewald/PME correction term added to the correction potential
171 //! LJ-Ewald/PME coefficient
174 //! Coulomb cut-off squared
177 //! VdW cut-off squared
179 //! VdW switched cut-off
181 //! Full, outer pair-list cut-off squared
183 //! 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
186 //! VdW shift dispersion constants
187 shift_consts_t dispersion_shift;
188 //! VdW shift repulsion constants
189 shift_consts_t repulsion_shift;
190 //! VdW switch constants
191 switch_consts_t vdw_switch;
193 /* Ewald Coulomb force table data - accessed through texture memory */
194 //! table scale/spacing
195 float coulomb_tab_scale;
196 } cl_nbparam_params_t;
202 //! Shift vector index plus possible flags
204 //! Start index into cj4
206 //! End index into cj4
212 //! The i-cluster interactions mask for 1 warp
214 //! Index into the exclusion array for 1 warp
222 //! The i-cluster mask data for 2 warps
223 nbnxn_im_ei_t imei[2];
229 unsigned int pair[CL_SIZE * CL_SIZE / 2]; /* Topology exclusion interaction bits for one warp,
230 * each unsigned has bitS for 4*8 i clusters
234 /*! i-cluster interaction mask for a super-cluster with all c_nbnxnGpuNumClusterPerSupercluster bits set */
235 __constant unsigned supercl_interaction_mask = ((1U << c_nbnxnGpuNumClusterPerSupercluster) - 1U);
237 /*! Minimum single precision threshold for r^2 to avoid r^-12 overflow. */
238 __constant float c_nbnxnMinDistanceSquared = NBNXM_MIN_DISTANCE_SQUARED_VALUE_FLOAT;
240 gmx_opencl_inline void preloadCj4Generic(__local int* sm_cjPreload,
241 const __global int* gm_cj,
244 bool gmx_unused iMaskCond)
246 /* Pre-load cj into shared memory */
247 # if defined _AMD_SOURCE_ // TODO: fix by setting c_nbnxnGpuClusterpairSplit properly
248 if (tidxj == 0 & tidxi < c_nbnxnGpuJgroupSize)
250 sm_cjPreload[tidxi] = gm_cj[tidxi];
253 const int c_clSize = CL_SIZE;
254 const int c_nbnxnGpuClusterpairSplit = 2;
255 const int c_splitClSize = c_clSize / c_nbnxnGpuClusterpairSplit;
256 if ((tidxj == 0 | tidxj == c_splitClSize) & (tidxi < c_nbnxnGpuJgroupSize))
258 sm_cjPreload[tidxi + tidxj * c_nbnxnGpuJgroupSize / c_splitClSize] = gm_cj[tidxi];
264 # if USE_SUBGROUP_PRELOAD
265 gmx_opencl_inline int preloadCj4Subgroup(const __global int* gm_cj)
267 // loads subgroup-size # of elements (8) instead of the 4 required
268 // equivalent to *cjs = *gm_cj
269 return intel_sub_group_block_read((const __global uint*)gm_cj);
271 # endif // USE_SUBGROUP_PRELOAD
273 # if USE_SUBGROUP_PRELOAD
274 typedef size_t CjType;
276 typedef __local int* CjType;
279 /*! \brief Preload cj4
281 * - For AMD we load once for a wavefront of 64 threads (on 4 threads * NTHREAD_Z)
282 * - For NVIDIA once per warp (on 2x4 threads * NTHREAD_Z)
283 * - For Intel(/USE_SUBGROUP_PRELOAD) loads into private memory(/register) instead of local memory
285 * It is the caller's responsibility to make sure that data is consumed only when
286 * it's ready. This function does not call a barrier.
288 gmx_opencl_inline void preloadCj4(CjType gmx_unused* cjs,
289 const __global int gmx_unused* gm_cj,
290 int gmx_unused tidxi,
291 int gmx_unused tidxj,
292 bool gmx_unused iMaskCond)
294 # if USE_SUBGROUP_PRELOAD
295 *cjs = preloadCj4Subgroup(gm_cj);
296 # elif USE_CJ_PREFETCH
297 preloadCj4Generic(*cjs, gm_cj, tidxi, tidxj, iMaskCond);
303 gmx_opencl_inline int loadCjPreload(__local int* sm_cjPreload, int jm, int gmx_unused tidxi, int gmx_unused tidxj)
305 # if defined _AMD_SOURCE_
306 int warpLoadOffset = 0; // TODO: fix by setting c_nbnxnGpuClusterpairSplit properly
308 const int c_clSize = CL_SIZE;
309 const int c_nbnxnGpuClusterpairSplit = 2;
310 const int c_splitClSize = c_clSize / c_nbnxnGpuClusterpairSplit;
311 int warpLoadOffset = (tidxj & c_splitClSize) * c_nbnxnGpuJgroupSize / c_splitClSize;
313 return sm_cjPreload[jm + warpLoadOffset];
316 /* \brief Load a cj given a jm index.
318 * If cj4 preloading is enabled, it loads from the local memory, otherwise from global.
320 gmx_opencl_inline int
321 loadCj(CjType cjs, const __global int gmx_unused* gm_cj, int jm, int gmx_unused tidxi, int gmx_unused tidxj)
323 # if USE_SUBGROUP_PRELOAD
324 return sub_group_broadcast(cjs, jm);
325 # elif USE_CJ_PREFETCH
326 return loadCjPreload(cjs, jm, tidxi, tidxj);
332 /*! Convert LJ sigma,epsilon parameters to C6,C12. */
333 gmx_opencl_inline float2 convert_sigma_epsilon_to_c6_c12(const float sigma, const float epsilon)
335 const float sigma2 = sigma * sigma;
336 const float sigma6 = sigma2 * sigma2 * sigma2;
337 const float c6 = epsilon * sigma6;
338 const float2 c6c12 = (float2)(c6, /* c6 */
339 c6 * sigma6); /* c12 */
344 /*! Apply force switch, force + energy version. */
345 gmx_opencl_inline void calculate_force_switch_F(const cl_nbparam_params_t* nbparam,
352 /* force switch constants */
353 const float disp_shift_V2 = nbparam->dispersion_shift.c2;
354 const float disp_shift_V3 = nbparam->dispersion_shift.c3;
355 const float repu_shift_V2 = nbparam->repulsion_shift.c2;
356 const float repu_shift_V3 = nbparam->repulsion_shift.c3;
358 const float r = r2 * inv_r;
359 float r_switch = r - nbparam->rvdw_switch;
360 r_switch = r_switch >= 0.0F ? r_switch : 0.0F;
362 *F_invr += -c6 * (disp_shift_V2 + disp_shift_V3 * r_switch) * r_switch * r_switch * inv_r
363 + c12 * (repu_shift_V2 + repu_shift_V3 * r_switch) * r_switch * r_switch * inv_r;
366 /*! Apply force switch, force-only version. */
367 gmx_opencl_inline void calculate_force_switch_F_E(const cl_nbparam_params_t* nbparam,
375 /* force switch constants */
376 const float disp_shift_V2 = nbparam->dispersion_shift.c2;
377 const float disp_shift_V3 = nbparam->dispersion_shift.c3;
378 const float repu_shift_V2 = nbparam->repulsion_shift.c2;
379 const float repu_shift_V3 = nbparam->repulsion_shift.c3;
381 const float disp_shift_F2 = nbparam->dispersion_shift.c2 / 3;
382 const float disp_shift_F3 = nbparam->dispersion_shift.c3 / 4;
383 const float repu_shift_F2 = nbparam->repulsion_shift.c2 / 3;
384 const float repu_shift_F3 = nbparam->repulsion_shift.c3 / 4;
386 const float r = r2 * inv_r;
387 float r_switch = r - nbparam->rvdw_switch;
388 r_switch = r_switch >= 0.0F ? r_switch : 0.0F;
390 *F_invr += -c6 * (disp_shift_V2 + disp_shift_V3 * r_switch) * r_switch * r_switch * inv_r
391 + c12 * (repu_shift_V2 + repu_shift_V3 * r_switch) * r_switch * r_switch * inv_r;
392 *E_lj += c6 * (disp_shift_F2 + disp_shift_F3 * r_switch) * r_switch * r_switch * r_switch
393 - c12 * (repu_shift_F2 + repu_shift_F3 * r_switch) * r_switch * r_switch * r_switch;
396 /*! Apply potential switch, force-only version. */
397 gmx_opencl_inline void calculate_potential_switch_F(const cl_nbparam_params_t* nbparam,
403 /* potential switch constants */
404 const float switch_V3 = nbparam->vdw_switch.c3;
405 const float switch_V4 = nbparam->vdw_switch.c4;
406 const float switch_V5 = nbparam->vdw_switch.c5;
407 const float switch_F2 = nbparam->vdw_switch.c3;
408 const float switch_F3 = nbparam->vdw_switch.c4;
409 const float switch_F4 = nbparam->vdw_switch.c5;
411 const float r = r2 * inv_r;
412 const float r_switch = r - nbparam->rvdw_switch;
414 /* Unlike in the F+E kernel, conditional is faster here */
417 const float sw = 1.0F
418 + (switch_V3 + (switch_V4 + switch_V5 * r_switch) * r_switch) * r_switch
419 * r_switch * r_switch;
420 const float dsw = (switch_F2 + (switch_F3 + switch_F4 * r_switch) * r_switch) * r_switch * r_switch;
422 *F_invr = (*F_invr) * sw - inv_r * (*E_lj) * dsw;
426 /*! Apply potential switch, force + energy version. */
427 gmx_opencl_inline void calculate_potential_switch_F_E(const cl_nbparam_params_t* nbparam,
433 /* potential switch constants */
434 const float switch_V3 = nbparam->vdw_switch.c3;
435 const float switch_V4 = nbparam->vdw_switch.c4;
436 const float switch_V5 = nbparam->vdw_switch.c5;
437 const float switch_F2 = nbparam->vdw_switch.c3;
438 const float switch_F3 = nbparam->vdw_switch.c4;
439 const float switch_F4 = nbparam->vdw_switch.c5;
441 const float r = r2 * inv_r;
442 float r_switch = r - nbparam->rvdw_switch;
443 r_switch = r_switch >= 0.0F ? r_switch : 0.0F;
445 /* Unlike in the F-only kernel, masking is faster here */
447 1.0F + (switch_V3 + (switch_V4 + switch_V5 * r_switch) * r_switch) * r_switch * r_switch * r_switch;
448 const float dsw = (switch_F2 + (switch_F3 + switch_F4 * r_switch) * r_switch) * r_switch * r_switch;
450 *F_invr = (*F_invr) * sw - inv_r * (*E_lj) * dsw;
454 /*! Calculate LJ-PME grid force contribution with
455 * geometric combination rule.
457 gmx_opencl_inline void calculate_lj_ewald_comb_geom_F(__constant const float* nbfp_comb_climg2d,
466 const float c6grid = nbfp_comb_climg2d[2 * typei] * nbfp_comb_climg2d[2 * typej];
468 /* Recalculate inv_r6 without exclusion mask */
469 const float inv_r6_nm = inv_r2 * inv_r2 * inv_r2;
470 const float cr2 = lje_coeff2 * r2;
471 const float expmcr2 = exp(-cr2);
472 const float poly = 1.0F + cr2 + HALF_F * 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;
478 /*! Calculate LJ-PME grid force + energy contribution with
479 * geometric combination rule.
481 gmx_opencl_inline void calculate_lj_ewald_comb_geom_F_E(__constant const float* nbfp_comb_climg2d,
482 const cl_nbparam_params_t* nbparam,
493 const float c6grid = nbfp_comb_climg2d[2 * typei] * nbfp_comb_climg2d[2 * typej];
495 /* Recalculate inv_r6 without exclusion mask */
496 const float inv_r6_nm = inv_r2 * inv_r2 * inv_r2;
497 const float cr2 = lje_coeff2 * r2;
498 const float expmcr2 = exp(-cr2);
499 const float poly = 1.0F + cr2 + HALF_F * cr2 * cr2;
501 /* Subtract the grid force from the total LJ force */
502 *F_invr += c6grid * (inv_r6_nm - expmcr2 * (inv_r6_nm * poly + lje_coeff6_6)) * inv_r2;
504 /* Shift should be applied only to real LJ pairs */
505 const float sh_mask = nbparam->sh_lj_ewald * int_bit;
506 *E_lj += ONE_SIXTH_F * c6grid * (inv_r6_nm * (1.0F - expmcr2 * poly) + sh_mask);
509 /*! Calculate LJ-PME grid force + energy contribution (if E_lj != NULL) with
510 * Lorentz-Berthelot combination rule.
511 * We use a single F+E kernel with conditional because the performance impact
512 * of this is pretty small and LB on the CPU is anyway very slow.
514 gmx_opencl_inline void calculate_lj_ewald_comb_LB_F_E(__constant const float* nbfp_comb_climg2d,
515 const cl_nbparam_params_t* nbparam,
527 /* sigma and epsilon are scaled to give 6*C6 */
528 const float sigma = nbfp_comb_climg2d[2 * typei] + nbfp_comb_climg2d[2 * typej];
530 const float epsilon = nbfp_comb_climg2d[2 * typei + 1] * nbfp_comb_climg2d[2 * typej + 1];
532 const float sigma2 = sigma * sigma;
533 const float c6grid = epsilon * sigma2 * sigma2 * sigma2;
535 /* Recalculate inv_r6 without exclusion mask */
536 const float inv_r6_nm = inv_r2 * inv_r2 * inv_r2;
537 const float cr2 = lje_coeff2 * r2;
538 const float expmcr2 = exp(-cr2);
539 const float poly = 1.0F + cr2 + HALF_F * cr2 * cr2;
541 /* Subtract the grid force from the total LJ force */
542 *F_invr += c6grid * (inv_r6_nm - expmcr2 * (inv_r6_nm * poly + lje_coeff6_6)) * inv_r2;
547 /* Shift should be applied only to real LJ pairs */
548 const float sh_mask = nbparam->sh_lj_ewald * int_bit;
549 *E_lj += ONE_SIXTH_F * c6grid * (inv_r6_nm * (1.0F - expmcr2 * poly) + sh_mask);
553 /*! Interpolate Ewald coulomb force using the table through the tex_nbfp texture.
554 * Original idea: from the OpenMM project
556 gmx_opencl_inline float interpolate_coulomb_force_r(__constant const float* coulomb_tab_climg2d,
560 float normalized = scale * r;
561 int index = (int)normalized;
562 float fract2 = normalized - (float)index;
563 float fract1 = 1.0F - fract2;
565 return fract1 * coulomb_tab_climg2d[index] + fract2 * coulomb_tab_climg2d[index + 1];
568 /*! Calculate analytical Ewald correction term. */
569 gmx_opencl_inline float pmecorrF(float z2)
571 const float FN6 = -1.7357322914161492954e-8F;
572 const float FN5 = 1.4703624142580877519e-6F;
573 const float FN4 = -0.000053401640219807709149F;
574 const float FN3 = 0.0010054721316683106153F;
575 const float FN2 = -0.019278317264888380590F;
576 const float FN1 = 0.069670166153766424023F;
577 const float FN0 = -0.75225204789749321333F;
579 const float FD4 = 0.0011193462567257629232F;
580 const float FD3 = 0.014866955030185295499F;
581 const float FD2 = 0.11583842382862377919F;
582 const float FD1 = 0.50736591960530292870F;
583 const float FD0 = 1.0F;
585 const float z4 = z2 * z2;
587 float polyFD0 = FD4 * z4 + FD2;
588 float polyFD1 = FD3 * z4 + FD1;
589 polyFD0 = polyFD0 * z4 + FD0;
590 polyFD0 = polyFD1 * z2 + polyFD0;
592 polyFD0 = 1.0F / polyFD0;
594 float polyFN0 = FN6 * z4 + FN4;
595 float polyFN1 = FN5 * z4 + FN3;
596 polyFN0 = polyFN0 * z4 + FN2;
597 polyFN1 = polyFN1 * z4 + FN1;
598 polyFN0 = polyFN0 * z4 + FN0;
599 polyFN0 = polyFN1 * z2 + polyFN0;
601 return polyFN0 * polyFD0;
605 gmx_opencl_inline void
606 reduce_force_j_shfl(float3 fin, __global float* fout, int gmx_unused tidxi, int gmx_unused tidxj, int aidx)
608 /* Only does reduction over 4 elements in cluster. Needs to be changed
609 * for CL_SIZE>4. See CUDA code for required code */
610 fin.x += intel_sub_group_shuffle_down(fin.x, fin.x, 1);
611 fin.y += intel_sub_group_shuffle_up(fin.y, fin.y, 1);
612 fin.z += intel_sub_group_shuffle_down(fin.z, fin.z, 1);
613 if ((tidxi & 1) == 1)
617 fin.x += intel_sub_group_shuffle_down(fin.x, fin.x, 2);
618 fin.z += intel_sub_group_shuffle_up(fin.z, fin.z, 2);
625 atomicAdd_g_f(&fout[3 * aidx + tidxi], fin.x);
630 gmx_opencl_inline void
631 reduce_force_j_generic(__local float* f_buf, float3 fcj_buf, __global float* fout, int tidxi, int tidxj, int aidx)
633 int tidx = tidxi + tidxj * CL_SIZE;
634 f_buf[tidx] = fcj_buf.x;
635 f_buf[FBUF_STRIDE + tidx] = fcj_buf.y;
636 f_buf[2 * FBUF_STRIDE + tidx] = fcj_buf.z;
638 /* Split the reduction between the first 3 column threads
639 Threads with column id 0 will do the reduction for (float3).x components
640 Threads with column id 1 will do the reduction for (float3).y components
641 Threads with column id 2 will do the reduction for (float3).z components.
642 The reduction is performed for each line tidxj of f_buf. */
646 for (int j = tidxj * CL_SIZE; j < (tidxj + 1) * CL_SIZE; j++)
648 f += f_buf[FBUF_STRIDE * tidxi + j];
651 atomicAdd_g_f(&fout[3 * aidx + tidxi], f);
655 /*! Final j-force reduction
657 gmx_opencl_inline void reduce_force_j(__local float gmx_unused* f_buf,
659 __global float* fout,
665 reduce_force_j_shfl(fcj_buf, fout, tidxi, tidxj, aidx);
667 reduce_force_j_generic(f_buf, fcj_buf, fout, tidxi, tidxj, aidx);
672 gmx_opencl_inline void reduce_force_i_and_shift_shfl(float3* fci_buf,
673 __global float* fout,
679 __global float* fshift)
681 /* Only does reduction over 4 elements in cluster (2 per warp). Needs to be changed
683 float2 fshift_buf = 0;
684 for (int ci_offset = 0; ci_offset < c_nbnxnGpuNumClusterPerSupercluster; ci_offset++)
686 int aidx = (sci * c_nbnxnGpuNumClusterPerSupercluster + ci_offset) * CL_SIZE + tidxi;
687 float3 fin = fci_buf[ci_offset];
688 fin.x += intel_sub_group_shuffle_down(fin.x, fin.x, CL_SIZE);
689 fin.y += intel_sub_group_shuffle_up(fin.y, fin.y, CL_SIZE);
690 fin.z += intel_sub_group_shuffle_down(fin.z, fin.z, CL_SIZE);
696 /* Threads 0,1 and 2,3 increment x,y for their warp */
697 atomicAdd_g_f(&fout[3 * aidx + (tidxj & 1)], fin.x);
700 fshift_buf[0] += fin.x;
702 /* Threads 0 and 2 increment z for their warp */
703 if ((tidxj & 1) == 0)
705 atomicAdd_g_f(&fout[3 * aidx + 2], fin.z);
708 fshift_buf[1] += fin.z;
712 /* add up local shift forces into global mem */
715 // Threads 0,1 and 2,3 update x,y
716 atomicAdd_g_f(&(fshift[3 * shift + (tidxj & 1)]), fshift_buf[0]);
717 // Threads 0 and 2 update z
718 if ((tidxj & 1) == 0)
720 atomicAdd_g_f(&(fshift[3 * shift + 2]), fshift_buf[1]);
726 /*! Final i-force reduction; this implementation works only with power of two
729 gmx_opencl_inline void reduce_force_i_and_shift_pow2(volatile __local float* f_buf,
731 __global float* fout,
737 __global float* fshift)
739 float fshift_buf = 0;
740 for (int ci_offset = 0; ci_offset < c_nbnxnGpuNumClusterPerSupercluster; ci_offset++)
742 int aidx = (sci * c_nbnxnGpuNumClusterPerSupercluster + ci_offset) * CL_SIZE + tidxi;
743 int tidx = tidxi + tidxj * CL_SIZE;
744 /* store i forces in shmem */
745 f_buf[tidx] = fci_buf[ci_offset].x;
746 f_buf[FBUF_STRIDE + tidx] = fci_buf[ci_offset].y;
747 f_buf[2 * FBUF_STRIDE + tidx] = fci_buf[ci_offset].z;
748 barrier(CLK_LOCAL_MEM_FENCE);
750 /* Reduce the initial CL_SIZE values for each i atom to half
751 * every step by using CL_SIZE * i threads.
752 * Can't just use i as loop variable because than nvcc refuses to unroll.
755 for (int j = CL_SIZE_LOG2 - 1; j > 0; j--)
760 f_buf[tidxj * CL_SIZE + tidxi] += f_buf[(tidxj + i) * CL_SIZE + tidxi];
761 f_buf[FBUF_STRIDE + tidxj * CL_SIZE + tidxi] +=
762 f_buf[FBUF_STRIDE + (tidxj + i) * CL_SIZE + tidxi];
763 f_buf[2 * FBUF_STRIDE + tidxj * CL_SIZE + tidxi] +=
764 f_buf[2 * FBUF_STRIDE + (tidxj + i) * CL_SIZE + tidxi];
769 * a) for CL_SIZE<8: id 2 (doing z in next block) is in 2nd warp
770 * b) for all CL_SIZE a barrier is needed before f_buf is reused by next reduce_force_i call
771 * TODO: Test on Nvidia for performance difference between having the barrier here or after the atomicAdd
773 barrier(CLK_LOCAL_MEM_FENCE);
775 /* i == 1, last reduction step, writing to global mem */
776 /* Split the reduction between the first 3 line threads
777 Threads with line id 0 will do the reduction for (float3).x components
778 Threads with line id 1 will do the reduction for (float3).y components
779 Threads with line id 2 will do the reduction for (float3).z components. */
782 float f = f_buf[tidxj * FBUF_STRIDE + tidxi] + f_buf[tidxj * FBUF_STRIDE + i * CL_SIZE + tidxi];
784 atomicAdd_g_f(&fout[3 * aidx + tidxj], f);
792 /* add up local shift forces into global mem */
795 /* Only threads with tidxj < 3 will update fshift.
796 The threads performing the update, must be the same as the threads
797 storing the reduction result above.
801 atomicAdd_g_f(&(fshift[3 * shift + tidxj]), fshift_buf);
806 /*! Final i-force reduction
808 gmx_opencl_inline void reduce_force_i_and_shift(__local float gmx_unused* f_buf,
816 __global float* fshift)
819 reduce_force_i_and_shift_shfl(fci_buf, f, bCalcFshift, tidxi, tidxj, sci, shift, fshift);
821 reduce_force_i_and_shift_pow2(f_buf, fci_buf, f, bCalcFshift, tidxi, tidxj, sci, shift, fshift);
827 gmx_opencl_inline void reduce_energy_shfl(float E_lj,
829 volatile __global float* e_lj,
830 volatile __global float* e_el,
833 E_lj = sub_group_reduce_add(E_lj);
834 E_el = sub_group_reduce_add(E_el);
835 /* Should be get_sub_group_local_id()==0. Doesn't work with Intel Classic driver.
836 * To make it possible to use REDUCE_SHUFFLE with single subgroup per i-j pair
837 * (e.g. subgroup size 16 with CL_SIZE 4), either this "if" needs to be changed or
838 * the definition of WARP_SIZE (currently CL_SIZE*CL_SIZE/2) needs to be changed
839 * (by supporting c_nbnxnGpuClusterpairSplit=1). */
840 if (tidx == 0 || tidx == WARP_SIZE)
842 atomicAdd_g_f(e_lj, E_lj);
843 atomicAdd_g_f(e_el, E_el);
848 /*! Energy reduction; this implementation works only with power of two
851 gmx_opencl_inline void reduce_energy_pow2(volatile __local float* buf,
852 volatile __global float* e_lj,
853 volatile __global float* e_el,
856 int i = WARP_SIZE / 2;
858 /* Can't just use i as loop variable because than nvcc refuses to unroll. */
859 for (int j = WARP_SIZE_LOG2 - 1; j > 0; j--)
863 buf[tidx] += buf[tidx + i];
864 buf[FBUF_STRIDE + tidx] += buf[FBUF_STRIDE + tidx + i];
869 /* last reduction step, writing to global mem */
872 float e1 = buf[tidx] + buf[tidx + i];
873 float e2 = buf[FBUF_STRIDE + tidx] + buf[FBUF_STRIDE + tidx + i];
875 atomicAdd_g_f(e_lj, e1);
876 atomicAdd_g_f(e_el, e2);
880 gmx_opencl_inline void reduce_energy(volatile __local float gmx_unused* buf,
883 volatile __global float* e_lj,
884 volatile __global float* e_el,
888 reduce_energy_shfl(E_lj, E_el, e_lj, e_el, tidx);
890 /* flush the energies to shmem and reduce them */
892 buf[FBUF_STRIDE + tidx] = E_el;
893 reduce_energy_pow2(buf + (tidx & WARP_SIZE), e_lj, e_el, tidx & ~WARP_SIZE);
897 gmx_opencl_inline bool gmx_sub_group_any_localmem(volatile __local int* warp_any, int widx, bool pred)
904 bool ret = warp_any[widx];
911 //! Returns a true if predicate is true for any work item in warp
912 gmx_opencl_inline bool gmx_sub_group_any(volatile __local int gmx_unused* warp_any, int gmx_unused widx, bool pred)
914 # if USE_SUBGROUP_ANY
915 return sub_group_any(pred);
917 return gmx_sub_group_any_localmem(warp_any, widx, pred);
921 #endif /* NBNXN_OPENCL_KERNEL_UTILS_CLH */