1 /* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
4 * This source code is part of
8 * GROningen MAchine for Chemical Simulations
10 * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
11 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
12 * Copyright (c) 2001-2012, The GROMACS development team,
13 * check out http://www.gromacs.org for more information.
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License
17 * as published by the Free Software Foundation; either version 2
18 * of the License, or (at your option) any later version.
20 * If you want to redistribute modifications, please consider that
21 * scientific software is very special. Version control is crucial -
22 * bugs must be traceable. We will be happy to consider code for
23 * inclusion in the official distribution, but derived work must not
24 * be called official GROMACS. Details are found in the README & COPYING
25 * files - if they are missing, get the official version at www.gromacs.org.
27 * To help us fund GROMACS development, we humbly ask that you cite
28 * the papers on the package - you can find them in the top README file.
30 * For more info, check our website at http://www.gromacs.org
33 * Gallium Rubidium Oxygen Manganese Argon Carbon Silicon
36 /* Note that floating-point constants in CUDA code should be suffixed
37 * with f (e.g. 0.5f), to stop the compiler producing intermediate
38 * code that is in double precision.
41 #include "../../gmxlib/cuda_tools/vectype_ops.cuh"
43 #ifndef NBNXN_CUDA_KERNEL_UTILS_CUH
44 #define NBNXN_CUDA_KERNEL_UTILS_CUH
46 #define WARP_SIZE_POW2_EXPONENT (5)
47 #define CL_SIZE_POW2_EXPONENT (3) /* change this together with GPU_NS_CLUSTER_SIZE !*/
48 #define CL_SIZE_SQ (CL_SIZE * CL_SIZE)
49 #define FBUF_STRIDE (CL_SIZE_SQ)
51 /*! Interpolate Ewald coulomb force using the table through the tex_nbfp texture.
52 * Original idea: OpenMM
54 static inline __device__
55 float interpolate_coulomb_force_r(float r, float scale)
57 float normalized = scale * r;
58 int index = (int) normalized;
59 float fract2 = normalized - index;
60 float fract1 = 1.0f - fract2;
62 return fract1 * tex1Dfetch(tex_coulomb_tab, index)
63 + fract2 * tex1Dfetch(tex_coulomb_tab, index + 1);
66 /*! Final j-force reduction; this generic implementation works with
67 * arbitrary array sizes.
69 static inline __device__
70 void reduce_force_j_generic(float *f_buf, float3 *fout,
71 int tidxi, int tidxj, int aidx)
75 float3 f = make_float3(0.0f);
76 for (int j = tidxj * CL_SIZE; j < (tidxj + 1) * CL_SIZE; j++)
79 f.y += f_buf[ FBUF_STRIDE + j];
80 f.z += f_buf[2 * FBUF_STRIDE + j];
83 atomicAdd(&fout[aidx], f);
87 /*! Final j-force reduction; this implementation only with power of two
88 * array sizes and with sm >= 3.0
90 #if __CUDA_ARCH__ >= 300
91 static inline __device__
92 void reduce_force_j_warp_shfl(float3 f, float3 *fout,
98 for (i = 0; i < 3; i++)
100 f.x += __shfl_down(f.x, 1<<i);
101 f.y += __shfl_down(f.y, 1<<i);
102 f.z += __shfl_down(f.z, 1<<i);
105 /* Write the reduced j-force on one thread for each j */
108 atomicAdd(&fout[aidx], f);
113 /*! Final i-force reduction; this generic implementation works with
114 * arbitrary array sizes.
116 static inline __device__
117 void reduce_force_i_generic(float *f_buf, float3 *fout,
118 float3 *fshift_buf, bool bCalcFshift,
119 int tidxi, int tidxj, int aidx)
123 float3 f = make_float3(0.0f);
124 for (int j = tidxi; j < CL_SIZE_SQ; j += CL_SIZE)
127 f.y += f_buf[ FBUF_STRIDE + j];
128 f.z += f_buf[2 * FBUF_STRIDE + j];
131 atomicAdd(&fout[aidx], f);
140 /*! Final i-force reduction; this implementation works only with power of two
143 static inline __device__
144 void reduce_force_i_pow2(volatile float *f_buf, float3 *fout,
145 float3 *fshift_buf, bool bCalcFshift,
146 int tidxi, int tidxj, int aidx)
149 float3 f = make_float3(0.0f);
151 /* Reduce the initial CL_SIZE values for each i atom to half
152 * every step by using CL_SIZE * i threads.
153 * Can't just use i as loop variable because than nvcc refuses to unroll.
157 for (j = CL_SIZE_POW2_EXPONENT - 1; j > 0; j--)
162 f_buf[ tidxj * CL_SIZE + tidxi] += f_buf[ (tidxj + i) * CL_SIZE + tidxi];
163 f_buf[ FBUF_STRIDE + tidxj * CL_SIZE + tidxi] += f_buf[ FBUF_STRIDE + (tidxj + i) * CL_SIZE + tidxi];
164 f_buf[2 * FBUF_STRIDE + tidxj * CL_SIZE + tidxi] += f_buf[2 * FBUF_STRIDE + (tidxj + i) * CL_SIZE + tidxi];
169 /* i == 1, last reduction step, writing to global mem */
172 f.x = f_buf[ tidxj * CL_SIZE + tidxi] + f_buf[ (tidxj + i) * CL_SIZE + tidxi];
173 f.y = f_buf[ FBUF_STRIDE + tidxj * CL_SIZE + tidxi] + f_buf[ FBUF_STRIDE + (tidxj + i) * CL_SIZE + tidxi];
174 f.z = f_buf[2 * FBUF_STRIDE + tidxj * CL_SIZE + tidxi] + f_buf[2 * FBUF_STRIDE + (tidxj + i) * CL_SIZE + tidxi];
176 atomicAdd(&fout[aidx], f);
185 /*! Final i-force reduction wrapper; calls the generic or pow2 reduction depending
186 * on whether the size of the array to be reduced is power of two or not.
188 static inline __device__
189 void reduce_force_i(float *f_buf, float3 *f,
190 float3 *fshift_buf, bool bCalcFshift,
191 int tidxi, int tidxj, int ai)
193 if ((CL_SIZE & (CL_SIZE - 1)))
195 reduce_force_i_generic(f_buf, f, fshift_buf, bCalcFshift, tidxi, tidxj, ai);
199 reduce_force_i_pow2(f_buf, f, fshift_buf, bCalcFshift, tidxi, tidxj, ai);
203 /*! Final i-force reduction; this implementation works only with power of two
204 * array sizes and with sm >= 3.0
206 #if __CUDA_ARCH__ >= 300
207 static inline __device__
208 void reduce_force_i_warp_shfl(float3 fin, float3 *fout,
209 float3 *fshift_buf, bool bCalcFshift,
215 for (j = 0; j < 2; j++)
217 fin.x += __shfl_down(fin.x, CL_SIZE<<j);
218 fin.y += __shfl_down(fin.y, CL_SIZE<<j);
219 fin.z += __shfl_down(fin.z, CL_SIZE<<j);
222 /* The first thread in the warp writes the reduced force */
223 if (tidxj == 0 || tidxj == 4)
225 atomicAdd(&fout[aidx], fin);
229 fshift_buf->x += fin.x;
230 fshift_buf->y += fin.y;
231 fshift_buf->z += fin.z;
237 /*! Energy reduction; this implementation works only with power of two
240 static inline __device__
241 void reduce_energy_pow2(volatile float *buf,
242 float *e_lj, float *e_el,
250 /* Can't just use i as loop variable because than nvcc refuses to unroll. */
252 for (j = WARP_SIZE_POW2_EXPONENT - 1; j > 0; j--)
256 buf[ tidx] += buf[ tidx + i];
257 buf[FBUF_STRIDE + tidx] += buf[FBUF_STRIDE + tidx + i];
262 /* last reduction step, writing to global mem */
265 e1 = buf[ tidx] + buf[ tidx + i];
266 e2 = buf[FBUF_STRIDE + tidx] + buf[FBUF_STRIDE + tidx + i];
273 /*! Energy reduction; this implementation works only with power of two
274 * array sizes and with sm >= 3.0
276 #if __CUDA_ARCH__ >= 300
277 static inline __device__
278 void reduce_energy_warp_shfl(float E_lj, float E_el,
279 float *e_lj, float *e_el,
286 for (i = 0; i < 5; i++)
288 E_lj += __shfl_down(E_lj,sh);
289 E_el += __shfl_down(E_el,sh);
293 /* The first thread in the warp writes the reduced energies */
294 if (tidx == 0 || tidx == WARP_SIZE)
296 atomicAdd(e_lj,E_lj);
297 atomicAdd(e_el,E_el);
300 #endif /* __CUDA_ARCH__ */
302 #endif /* NBNXN_CUDA_KERNEL_UTILS_CUH */