2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013, 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.
35 #ifndef _nbnxn_kernel_simd_utils_h_
36 #define _nbnxn_kernel_simd_utils_h_
38 /*! \brief Provides hardware-specific utility routines for the SIMD kernels.
40 * Defines all functions, typedefs, constants and macros that have
41 * explicit dependencies on the j-cluster size, precision, or SIMD
42 * width. This includes handling diagonal, Newton and topology
45 * The functionality which depends on the j-cluster size is:
48 * energy group pair energy storage
51 #if !defined GMX_NBNXN_SIMD_2XNN && !defined GMX_NBNXN_SIMD_4XN
52 #error "Must define an NBNxN kernel flavour before including NBNxN kernel utility functions"
55 #ifdef GMX_SIMD_REFERENCE_PLAIN_C
57 /* Align a stack-based thread-local working array. */
58 static gmx_inline int *
59 prepare_table_load_buffer(const int *array)
64 #include "nbnxn_kernel_simd_utils_ref.h"
66 #else /* GMX_SIMD_REFERENCE_PLAIN_C */
69 /* Include x86 SSE2 compatible SIMD functions */
71 /* Set the stride for the lookup of the two LJ parameters from their
72 * (padded) array. We use the minimum supported SIMD memory alignment.
74 #if defined GMX_DOUBLE
75 static const int nbfp_stride = 2;
77 static const int nbfp_stride = 4;
80 /* Align a stack-based thread-local working array. Table loads on
81 * full-width AVX_256 use the array, but other implementations do
83 static gmx_inline int *
84 prepare_table_load_buffer(const int *array)
86 #if defined GMX_X86_AVX_256 && !defined GMX_USE_HALF_WIDTH_SIMD_HERE
87 return gmx_simd_align_int(array);
93 #if defined GMX_X86_AVX_256 && !defined GMX_USE_HALF_WIDTH_SIMD_HERE
95 /* With full AVX-256 SIMD, half SIMD-width table loads are optimal */
96 #if GMX_SIMD_WIDTH_HERE == 8
101 Berk, 2xnn.c had the following code, but I think it is safe to remove now, given the code immediately above.
103 #if defined GMX_X86_AVX_256 && !defined GMX_DOUBLE
104 / * AVX-256 single precision 2x(4+4) kernel,
105 * we can do half SIMD-width aligned FDV0 table loads.
112 #include "nbnxn_kernel_simd_utils_x86_256d.h"
113 #else /* GMX_DOUBLE */
114 #include "nbnxn_kernel_simd_utils_x86_256s.h"
115 #endif /* GMX_DOUBLE */
117 #else /* defined GMX_X86_AVX_256 && !defined GMX_USE_HALF_WIDTH_SIMD_HERE */
119 /* We use the FDV0 table layout when we can use aligned table loads */
120 #if GMX_SIMD_WIDTH_HERE == 4
125 #include "nbnxn_kernel_simd_utils_x86_128d.h"
126 #else /* GMX_DOUBLE */
127 #include "nbnxn_kernel_simd_utils_x86_128s.h"
128 #endif /* GMX_DOUBLE */
130 #endif /* defined GMX_X86_AVX_256 && !defined GMX_USE_HALF_WIDTH_SIMD_HERE */
132 #else /* GMX_X86_SSE2 */
134 #if GMX_SIMD_WIDTH_HERE > 4
135 static const int nbfp_stride = 4;
137 static const int nbfp_stride = GMX_SIMD_WIDTH_HERE;
140 /* We use the FDV0 table layout when we can use aligned table loads */
141 #if GMX_SIMD_WIDTH_HERE == 4
145 #ifdef GMX_CPU_ACCELERATION_IBM_QPX
146 #include "nbnxn_kernel_simd_utils_ibm_qpx.h"
147 #endif /* GMX_CPU_ACCELERATION_IBM_QPX */
149 #endif /* GMX_X86_SSE2 */
150 #endif /* GMX_SIMD_REFERENCE_PLAIN_C */
154 /* Add energy register to possibly multiple terms in the energy array */
155 static inline void add_ener_grp(gmx_mm_pr e_S, real *v, const int *offset_jj)
159 /* We need to balance the number of store operations with
160 * the rapidly increases number of combinations of energy groups.
161 * We add to a temporary buffer for 1 i-group vs 2 j-groups.
163 for (jj = 0; jj < (UNROLLJ/2); jj++)
167 v_S = gmx_load_pr(v+offset_jj[jj]+jj*GMX_SIMD_WIDTH_HERE);
168 gmx_store_pr(v+offset_jj[jj]+jj*GMX_SIMD_WIDTH_HERE, gmx_add_pr(v_S, e_S));
173 #if defined GMX_NBNXN_SIMD_2XNN && defined UNROLLJ
174 /* As add_ener_grp, but for two groups of UNROLLJ/2 stored in
175 * a single SIMD register.
178 add_ener_grp_halves(gmx_mm_pr e_S, real *v0, real *v1, const int *offset_jj)
180 gmx_mm_hpr e_S0, e_S1;
183 gmx_pr_to_2hpr(e_S, &e_S0, &e_S1);
185 for (jj = 0; jj < (UNROLLJ/2); jj++)
189 gmx_load_hpr(&v_S, v0+offset_jj[jj]+jj*GMX_SIMD_WIDTH_HERE/2);
190 gmx_store_hpr(v0+offset_jj[jj]+jj*GMX_SIMD_WIDTH_HERE/2, gmx_add_hpr(v_S, e_S0));
192 for (jj = 0; jj < (UNROLLJ/2); jj++)
196 gmx_load_hpr(&v_S, v1+offset_jj[jj]+jj*GMX_SIMD_WIDTH_HERE/2);
197 gmx_store_hpr(v1+offset_jj[jj]+jj*GMX_SIMD_WIDTH_HERE/2, gmx_add_hpr(v_S, e_S1));
202 #endif /* _nbnxn_kernel_simd_utils_h_ */