2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016,2017,2018,2019, 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 * CUDA non-bonded kernel used through preprocessor-based code generation
39 * of multiple kernel flavors, see nbnxn_cuda_kernels.cuh.
41 * NOTE: No include fence as it is meant to be included multiple times.
43 * \author Szilárd Páll <pall.szilard@gmail.com>
44 * \author Berk Hess <hess@kth.se>
45 * \ingroup module_nbnxm
48 #include "gromacs/gpu_utils/cuda_arch_utils.cuh"
49 #include "gromacs/gpu_utils/cuda_kernel_utils.cuh"
50 #include "gromacs/math/utilities.h"
51 #include "gromacs/pbcutil/ishift.h"
52 /* Note that floating-point constants in CUDA code should be suffixed
53 * with f (e.g. 0.5f), to stop the compiler producing intermediate
54 * code that is in double precision.
57 #if defined EL_EWALD_ANA || defined EL_EWALD_TAB
58 /* Note: convenience macro, needs to be undef-ed at the end of the file. */
62 #if defined EL_EWALD_ANY || defined EL_RF || defined LJ_EWALD \
63 || (defined EL_CUTOFF && defined CALC_ENERGIES)
64 /* Macro to control the calculation of exclusion forces in the kernel
65 * We do that with Ewald (elec/vdw) and RF. Cut-off only has exclusion
68 * Note: convenience macro, needs to be undef-ed at the end of the file.
70 # define EXCLUSION_FORCES
73 #if defined LJ_EWALD_COMB_GEOM || defined LJ_EWALD_COMB_LB
74 /* Note: convenience macro, needs to be undef-ed at the end of the file. */
78 #if defined LJ_COMB_GEOM || defined LJ_COMB_LB
83 Kernel launch parameters:
84 - #blocks = #pair lists, blockId = pair list Id
85 - #threads = NTHREAD_Z * c_clSize^2
86 - shmem = see nbnxn_cuda.cu:calc_shmem_required_nonbonded()
88 Each thread calculates an i force-component taking one pair of i-j atoms.
92 /*! \brief Compute capability dependent definition of kernel launch configuration parameters.
94 * NTHREAD_Z controls the number of j-clusters processed concurrently on NTHREAD_Z
95 * warp-pairs per block.
97 * - On CC 3.0-3.5, and >=5.0 NTHREAD_Z == 1, translating to 64 th/block with 16
98 * blocks/multiproc, is the fastest even though this setup gives low occupancy
100 * NTHREAD_Z > 1 results in excessive register spilling unless the minimum blocks
101 * per multiprocessor is reduced proportionally to get the original number of max
102 * threads in flight (and slightly lower performance).
103 * - On CC 3.7 there are enough registers to double the number of threads; using
104 * NTHREADS_Z == 2 is fastest with 16 blocks (TODO: test with RF and other kernels
105 * with low-register use).
107 * Note that the current kernel implementation only supports NTHREAD_Z > 1 with
108 * shuffle-based reduction, hence CC >= 3.0.
111 * NOTEs on Volta / CUDA 9 extensions:
113 * - While active thread masks are required for the warp collectives
114 * (we use any and shfl), the kernel is designed such that all conditions
115 * (other than the inner-most distance check) including loop trip counts
116 * are warp-synchronous. Therefore, we don't need ballot to compute the
117 * active masks as these are all full-warp masks.
119 * - TODO: reconsider the use of __syncwarp(): its only role is currently to prevent
120 * WAR hazard due to the cj preload; we should try to replace it with direct
121 * loads (which may be faster given the improved L1 on Volta).
124 /* Kernel launch bounds for different compute capabilities. The value of NTHREAD_Z
125 * determines the number of threads per block and it is chosen such that
126 * 16 blocks/multiprocessor can be kept in flight.
127 * - CC 3.0,3.5, and >=5.0: NTHREAD_Z=1, (64, 16) bounds
128 * - CC 3.7: NTHREAD_Z=2, (128, 16) bounds
130 * Note: convenience macros, need to be undef-ed at the end of the file.
132 #if GMX_PTX_ARCH == 370
133 # define NTHREAD_Z (2)
134 # define MIN_BLOCKS_PER_MP (16)
136 # define NTHREAD_Z (1)
137 # define MIN_BLOCKS_PER_MP (16)
138 #endif /* GMX_PTX_ARCH == 370 */
139 #define THREADS_PER_BLOCK (c_clSize * c_clSize * NTHREAD_Z)
141 #if GMX_PTX_ARCH >= 350
143 __launch_bounds__(THREADS_PER_BLOCK, MIN_BLOCKS_PER_MP)
145 __launch_bounds__(THREADS_PER_BLOCK)
146 #endif /* GMX_PTX_ARCH >= 350 */
148 # ifdef CALC_ENERGIES
149 __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _VF_prune_cuda)
151 __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _F_prune_cuda)
152 # endif /* CALC_ENERGIES */
154 # ifdef CALC_ENERGIES
155 __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _VF_cuda)
157 __global__ void NB_KERNEL_FUNC_NAME(nbnxn_kernel, _F_cuda)
158 # endif /* CALC_ENERGIES */
159 #endif /* PRUNE_NBL */
160 (const cu_atomdata_t atdat, const cu_nbparam_t nbparam, const cu_plist_t plist, bool bCalcFshift)
161 #ifdef FUNCTION_DECLARATION_ONLY
162 ; /* Only do function declaration, omit the function body. */
165 /* convenience variables */
166 const nbnxn_sci_t* pl_sci = plist.sci;
170 nbnxn_cj4_t* pl_cj4 = plist.cj4;
171 const nbnxn_excl_t* excl = plist.excl;
173 const int* atom_types = atdat.atom_types;
174 int ntypes = atdat.ntypes;
176 const float2* lj_comb = atdat.lj_comb;
177 float2 ljcp_i, ljcp_j;
179 const float4* xq = atdat.xq;
181 const float3* shift_vec = atdat.shift_vec;
182 float rcoulomb_sq = nbparam.rcoulomb_sq;
183 # ifdef VDW_CUTOFF_CHECK
184 float rvdw_sq = nbparam.rvdw_sq;
188 float lje_coeff2, lje_coeff6_6;
191 float two_k_rf = nbparam.two_k_rf;
194 float beta2 = nbparam.ewald_beta * nbparam.ewald_beta;
195 float beta3 = nbparam.ewald_beta * nbparam.ewald_beta * nbparam.ewald_beta;
198 float rlist_sq = nbparam.rlistOuter_sq;
201 # ifdef CALC_ENERGIES
203 float beta = nbparam.ewald_beta;
204 float ewald_shift = nbparam.sh_ewald;
206 float c_rf = nbparam.c_rf;
207 # endif /* EL_EWALD_ANY */
208 float* e_lj = atdat.e_lj;
209 float* e_el = atdat.e_el;
210 # endif /* CALC_ENERGIES */
212 /* thread/block/warp id-s */
213 unsigned int tidxi = threadIdx.x;
214 unsigned int tidxj = threadIdx.y;
215 unsigned int tidx = threadIdx.y * blockDim.x + threadIdx.x;
217 unsigned int tidxz = 0;
219 unsigned int tidxz = threadIdx.z;
221 unsigned int bidx = blockIdx.x;
222 unsigned int widx = tidx / warp_size; /* warp index */
224 int sci, ci, cj, ai, aj, cij4_start, cij4_end;
228 int i, jm, j4, wexcl_idx;
229 float qi, qj_f, r2, inv_r, inv_r2;
230 # if !defined LJ_COMB_LB || defined CALC_ENERGIES
231 float inv_r6, c6, c12;
234 float sigma, epsilon;
236 float int_bit, F_invr;
237 # ifdef CALC_ENERGIES
240 # if defined CALC_ENERGIES || defined LJ_POT_SWITCH
243 unsigned int wexcl, imask, mask_ji;
245 float3 xi, xj, rv, f_ij, fcj_buf;
246 float3 fci_buf[c_numClPerSupercl]; /* i force buffer */
249 /*! i-cluster interaction mask for a super-cluster with all c_numClPerSupercl=8 bits set */
250 const unsigned superClInteractionMask = ((1U << c_numClPerSupercl) - 1U);
252 /*********************************************************************
253 * Set up shared memory pointers.
254 * sm_nextSlotPtr should always be updated to point to the "next slot",
255 * that is past the last point where data has been stored.
257 extern __shared__ char sm_dynamicShmem[];
258 char* sm_nextSlotPtr = sm_dynamicShmem;
259 static_assert(sizeof(char) == 1,
260 "The shared memory offset calculation assumes that char is 1 byte");
262 /* shmem buffer for i x+q pre-loading */
263 float4* xqib = (float4*)sm_nextSlotPtr;
264 sm_nextSlotPtr += (c_numClPerSupercl * c_clSize * sizeof(*xqib));
266 /* shmem buffer for cj, for each warp separately */
267 int* cjs = (int*)(sm_nextSlotPtr);
268 /* the cjs buffer's use expects a base pointer offset for pairs of warps in the j-concurrent execution */
269 cjs += tidxz * c_nbnxnGpuClusterpairSplit * c_nbnxnGpuJgroupSize;
270 sm_nextSlotPtr += (NTHREAD_Z * c_nbnxnGpuClusterpairSplit * c_nbnxnGpuJgroupSize * sizeof(*cjs));
273 /* shmem buffer for i atom-type pre-loading */
274 int* atib = (int*)sm_nextSlotPtr;
275 sm_nextSlotPtr += (c_numClPerSupercl * c_clSize * sizeof(*atib));
277 /* shmem buffer for i-atom LJ combination rule parameters */
278 float2* ljcpib = (float2*)sm_nextSlotPtr;
279 sm_nextSlotPtr += (c_numClPerSupercl * c_clSize * sizeof(*ljcpib));
281 /*********************************************************************/
283 nb_sci = pl_sci[bidx]; /* my i super-cluster's index = current bidx */
284 sci = nb_sci.sci; /* super-cluster */
285 cij4_start = nb_sci.cj4_ind_start; /* first ...*/
286 cij4_end = nb_sci.cj4_ind_end; /* and last index of j clusters */
290 /* Pre-load i-atom x and q into shared memory */
291 ci = sci * c_numClPerSupercl + tidxj;
292 ai = ci * c_clSize + tidxi;
294 float* shiftptr = (float*)&shift_vec[nb_sci.shift];
295 xqbuf = xq[ai] + make_float4(LDG(shiftptr), LDG(shiftptr + 1), LDG(shiftptr + 2), 0.0f);
296 xqbuf.w *= nbparam.epsfac;
297 xqib[tidxj * c_clSize + tidxi] = xqbuf;
300 /* Pre-load the i-atom types into shared memory */
301 atib[tidxj * c_clSize + tidxi] = atom_types[ai];
303 /* Pre-load the LJ combination parameters into shared memory */
304 ljcpib[tidxj * c_clSize + tidxi] = lj_comb[ai];
309 for (i = 0; i < c_numClPerSupercl; i++)
311 fci_buf[i] = make_float3(0.0f);
315 /* TODO: we are trading registers with flops by keeping lje_coeff-s, try re-calculating it later */
316 lje_coeff2 = nbparam.ewaldcoeff_lj * nbparam.ewaldcoeff_lj;
317 lje_coeff6_6 = lje_coeff2 * lje_coeff2 * lje_coeff2 * c_oneSixth;
321 # ifdef CALC_ENERGIES
325 # ifdef EXCLUSION_FORCES /* Ewald or RF */
326 if (nb_sci.shift == CENTRAL && pl_cj4[cij4_start].cj[0] == sci * c_numClPerSupercl)
328 /* we have the diagonal: add the charge and LJ self interaction energy term */
329 for (i = 0; i < c_numClPerSupercl; i++)
331 # if defined EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF
332 qi = xqib[i * c_clSize + tidxi].w;
337 # if DISABLE_CUDA_TEXTURES
339 &nbparam.nbfp[atom_types[(sci * c_numClPerSupercl + i) * c_clSize + tidxi] * (ntypes + 1) * 2]);
341 E_lj += tex1Dfetch<float>(
343 atom_types[(sci * c_numClPerSupercl + i) * c_clSize + tidxi] * (ntypes + 1) * 2);
348 /* divide the self term(s) equally over the j-threads, then multiply with the coefficients. */
350 E_lj /= c_clSize * NTHREAD_Z;
351 E_lj *= 0.5f * c_oneSixth * lje_coeff6_6;
354 # if defined EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF
355 /* Correct for epsfac^2 due to adding qi^2 */
356 E_el /= nbparam.epsfac * c_clSize * NTHREAD_Z;
357 # if defined EL_RF || defined EL_CUTOFF
358 E_el *= -0.5f * c_rf;
360 E_el *= -beta * M_FLOAT_1_SQRTPI; /* last factor 1/sqrt(pi) */
362 # endif /* EL_EWALD_ANY || defined EL_RF || defined EL_CUTOFF */
364 # endif /* EXCLUSION_FORCES */
366 # endif /* CALC_ENERGIES */
368 # ifdef EXCLUSION_FORCES
369 const int nonSelfInteraction = !(nb_sci.shift == CENTRAL & tidxj <= tidxi);
372 /* loop over the j clusters = seen by any of the atoms in the current super-cluster;
373 * The loop stride NTHREAD_Z ensures that consecutive warps-pairs are assigned
374 * consecutive j4's entries.
376 for (j4 = cij4_start + tidxz; j4 < cij4_end; j4 += NTHREAD_Z)
378 wexcl_idx = pl_cj4[j4].imei[widx].excl_ind;
379 imask = pl_cj4[j4].imei[widx].imask;
380 wexcl = excl[wexcl_idx].pair[(tidx) & (warp_size - 1)];
386 /* Pre-load cj into shared memory on both warps separately */
387 if ((tidxj == 0 | tidxj == 4) & (tidxi < c_nbnxnGpuJgroupSize))
389 cjs[tidxi + tidxj * c_nbnxnGpuJgroupSize / c_splitClSize] = pl_cj4[j4].cj[tidxi];
391 __syncwarp(c_fullWarpMask);
393 /* Unrolling this loop
394 - with pruning leads to register spilling;
395 - on Kepler and later it is much slower;
396 Tested with up to nvcc 7.5 */
397 for (jm = 0; jm < c_nbnxnGpuJgroupSize; jm++)
399 if (imask & (superClInteractionMask << (jm * c_numClPerSupercl)))
401 mask_ji = (1U << (jm * c_numClPerSupercl));
403 cj = cjs[jm + (tidxj & 4) * c_nbnxnGpuJgroupSize / c_splitClSize];
404 aj = cj * c_clSize + tidxj;
406 /* load j atom data */
408 xj = make_float3(xqbuf.x, xqbuf.y, xqbuf.z);
411 typej = atom_types[aj];
413 ljcp_j = lj_comb[aj];
416 fcj_buf = make_float3(0.0f);
418 # if !defined PRUNE_NBL
421 for (i = 0; i < c_numClPerSupercl; i++)
425 ci = sci * c_numClPerSupercl + i; /* i cluster index */
427 /* all threads load an atom from i cluster ci into shmem! */
428 xqbuf = xqib[i * c_clSize + tidxi];
429 xi = make_float3(xqbuf.x, xqbuf.y, xqbuf.z);
431 /* distance between i and j atoms */
436 /* If _none_ of the atoms pairs are in cutoff range,
437 the bit corresponding to the current
438 cluster-pair in imask gets set to 0. */
439 if (!__any_sync(c_fullWarpMask, r2 < rlist_sq))
445 int_bit = (wexcl & mask_ji) ? 1.0f : 0.0f;
447 /* cutoff & exclusion check */
448 # ifdef EXCLUSION_FORCES
449 if ((r2 < rcoulomb_sq) * (nonSelfInteraction | (ci != cj)))
451 if ((r2 < rcoulomb_sq) * int_bit)
454 /* load the rest of the i-atom parameters */
458 /* LJ 6*C6 and 12*C12 */
459 typei = atib[i * c_clSize + tidxi];
460 fetch_nbfp_c6_c12(c6, c12, nbparam, ntypes * typei + typej);
462 ljcp_i = ljcpib[i * c_clSize + tidxi];
464 c6 = ljcp_i.x * ljcp_j.x;
465 c12 = ljcp_i.y * ljcp_j.y;
467 /* LJ 2^(1/6)*sigma and 12*epsilon */
468 sigma = ljcp_i.x + ljcp_j.x;
469 epsilon = ljcp_i.y * ljcp_j.y;
470 # if defined CALC_ENERGIES || defined LJ_FORCE_SWITCH || defined LJ_POT_SWITCH
471 convert_sigma_epsilon_to_c6_c12(sigma, epsilon, &c6, &c12);
473 # endif /* LJ_COMB_GEOM */
474 # endif /* LJ_COMB */
476 // Ensure distance do not become so small that r^-12 overflows
477 r2 = max(r2, NBNXN_MIN_RSQ);
480 inv_r2 = inv_r * inv_r;
481 # if !defined LJ_COMB_LB || defined CALC_ENERGIES
482 inv_r6 = inv_r2 * inv_r2 * inv_r2;
483 # ifdef EXCLUSION_FORCES
484 /* We could mask inv_r2, but with Ewald
485 * masking both inv_r6 and F_invr is faster */
487 # endif /* EXCLUSION_FORCES */
489 F_invr = inv_r6 * (c12 * inv_r6 - c6) * inv_r2;
490 # if defined CALC_ENERGIES || defined LJ_POT_SWITCH
492 * (c12 * (inv_r6 * inv_r6 + nbparam.repulsion_shift.cpot) * c_oneTwelveth
493 - c6 * (inv_r6 + nbparam.dispersion_shift.cpot) * c_oneSixth);
495 # else /* !LJ_COMB_LB || CALC_ENERGIES */
496 float sig_r = sigma * inv_r;
497 float sig_r2 = sig_r * sig_r;
498 float sig_r6 = sig_r2 * sig_r2 * sig_r2;
499 # ifdef EXCLUSION_FORCES
501 # endif /* EXCLUSION_FORCES */
503 F_invr = epsilon * sig_r6 * (sig_r6 - 1.0f) * inv_r2;
504 # endif /* !LJ_COMB_LB || CALC_ENERGIES */
506 # ifdef LJ_FORCE_SWITCH
507 # ifdef CALC_ENERGIES
508 calculate_force_switch_F_E(nbparam, c6, c12, inv_r, r2, &F_invr, &E_lj_p);
510 calculate_force_switch_F(nbparam, c6, c12, inv_r, r2, &F_invr);
511 # endif /* CALC_ENERGIES */
512 # endif /* LJ_FORCE_SWITCH */
516 # ifdef LJ_EWALD_COMB_GEOM
517 # ifdef CALC_ENERGIES
518 calculate_lj_ewald_comb_geom_F_E(nbparam, typei, typej, r2, inv_r2,
519 lje_coeff2, lje_coeff6_6, int_bit,
522 calculate_lj_ewald_comb_geom_F(nbparam, typei, typej, r2, inv_r2,
523 lje_coeff2, lje_coeff6_6, &F_invr);
524 # endif /* CALC_ENERGIES */
525 # elif defined LJ_EWALD_COMB_LB
526 calculate_lj_ewald_comb_LB_F_E(nbparam, typei, typej, r2, inv_r2,
527 lje_coeff2, lje_coeff6_6,
528 # ifdef CALC_ENERGIES
529 int_bit, &F_invr, &E_lj_p
532 # endif /* CALC_ENERGIES */
534 # endif /* LJ_EWALD_COMB_GEOM */
535 # endif /* LJ_EWALD */
537 # ifdef LJ_POT_SWITCH
538 # ifdef CALC_ENERGIES
539 calculate_potential_switch_F_E(nbparam, inv_r, r2, &F_invr, &E_lj_p);
541 calculate_potential_switch_F(nbparam, inv_r, r2, &F_invr, &E_lj_p);
542 # endif /* CALC_ENERGIES */
543 # endif /* LJ_POT_SWITCH */
545 # ifdef VDW_CUTOFF_CHECK
546 /* Separate VDW cut-off check to enable twin-range cut-offs
547 * (rvdw < rcoulomb <= rlist)
549 vdw_in_range = (r2 < rvdw_sq) ? 1.0f : 0.0f;
550 F_invr *= vdw_in_range;
551 # ifdef CALC_ENERGIES
552 E_lj_p *= vdw_in_range;
554 # endif /* VDW_CUTOFF_CHECK */
556 # ifdef CALC_ENERGIES
562 # ifdef EXCLUSION_FORCES
563 F_invr += qi * qj_f * int_bit * inv_r2 * inv_r;
565 F_invr += qi * qj_f * inv_r2 * inv_r;
569 F_invr += qi * qj_f * (int_bit * inv_r2 * inv_r - two_k_rf);
571 # if defined EL_EWALD_ANA
573 * (int_bit * inv_r2 * inv_r + pmecorrF(beta2 * r2) * beta3);
574 # elif defined EL_EWALD_TAB
577 - interpolate_coulomb_force_r(nbparam, r2 * inv_r))
579 # endif /* EL_EWALD_ANA/TAB */
581 # ifdef CALC_ENERGIES
583 E_el += qi * qj_f * (int_bit * inv_r - c_rf);
586 E_el += qi * qj_f * (int_bit * inv_r + 0.5f * two_k_rf * r2 - c_rf);
589 /* 1.0f - erff is faster than erfcf */
591 * (inv_r * (int_bit - erff(r2 * inv_r * beta)) - int_bit * ewald_shift);
592 # endif /* EL_EWALD_ANY */
596 /* accumulate j forces in registers */
599 /* accumulate i forces in registers */
604 /* shift the mask bit by 1 */
608 /* reduce j forces */
609 reduce_force_j_warp_shfl(fcj_buf, f, tidxi, aj, c_fullWarpMask);
613 /* Update the imask with the new one which does not contain the
614 out of range clusters anymore. */
615 pl_cj4[j4].imei[widx].imask = imask;
618 // avoid shared memory WAR hazards between loop iterations
619 __syncwarp(c_fullWarpMask);
622 /* skip central shifts when summing shift forces */
623 if (nb_sci.shift == CENTRAL)
628 float fshift_buf = 0.0f;
630 /* reduce i forces */
631 for (i = 0; i < c_numClPerSupercl; i++)
633 ai = (sci * c_numClPerSupercl + i) * c_clSize + tidxi;
634 reduce_force_i_warp_shfl(fci_buf[i], f, &fshift_buf, bCalcFshift, tidxj, ai, c_fullWarpMask);
637 /* add up local shift forces into global mem, tidxj indexes x,y,z */
638 if (bCalcFshift && (tidxj & 3) < 3)
640 atomicAdd(&(atdat.fshift[nb_sci.shift].x) + (tidxj & 3), fshift_buf);
643 # ifdef CALC_ENERGIES
644 /* reduce the energies over warps and store into global memory */
645 reduce_energy_warp_shfl(E_lj, E_el, e_lj, e_el, tidx, c_fullWarpMask);
648 #endif /* FUNCTION_DECLARATION_ONLY */
651 #undef MIN_BLOCKS_PER_MP
652 #undef THREADS_PER_BLOCK
655 #undef EXCLUSION_FORCES