1 /* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
4 * This file is part of Gromacs Copyright (c) 1991-2004
5 * David van der Spoel, Erik Lindahl, University of Groningen.
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * as published by the Free Software Foundation; either version 2
10 * of the License, or (at your option) any later version.
12 * To help us fund GROMACS development, we humbly ask that you cite
13 * the research papers on the package. Check out http://www.gromacs.org
16 * Gnomes, ROck Monsters And Chili Sauce
22 /* Must come directly after config.h */
23 #ifdef GMX_THREAD_SHM_FDECOMP
24 #include <thread_mpi.h>
27 #include <types/simple.h>
28 #include <types/nrnb.h>
30 #include "nb_kernel_x86_64_sse.h"
32 /* Include x86_64 SSE kernel headers in local directory */
33 #include "nb_kernel010_x86_64_sse.h"
34 #include "nb_kernel030_x86_64_sse.h"
35 #include "nb_kernel100_x86_64_sse.h"
36 #include "nb_kernel101_x86_64_sse.h"
37 #include "nb_kernel102_x86_64_sse.h"
38 #include "nb_kernel103_x86_64_sse.h"
39 #include "nb_kernel104_x86_64_sse.h"
40 #include "nb_kernel110_x86_64_sse.h"
41 #include "nb_kernel111_x86_64_sse.h"
42 #include "nb_kernel112_x86_64_sse.h"
43 #include "nb_kernel113_x86_64_sse.h"
44 #include "nb_kernel114_x86_64_sse.h"
45 #include "nb_kernel130_x86_64_sse.h"
46 #include "nb_kernel131_x86_64_sse.h"
47 #include "nb_kernel132_x86_64_sse.h"
48 #include "nb_kernel133_x86_64_sse.h"
49 #include "nb_kernel134_x86_64_sse.h"
50 #include "nb_kernel200_x86_64_sse.h"
51 #include "nb_kernel201_x86_64_sse.h"
52 #include "nb_kernel202_x86_64_sse.h"
53 #include "nb_kernel203_x86_64_sse.h"
54 #include "nb_kernel204_x86_64_sse.h"
55 #include "nb_kernel210_x86_64_sse.h"
56 #include "nb_kernel211_x86_64_sse.h"
57 #include "nb_kernel212_x86_64_sse.h"
58 #include "nb_kernel213_x86_64_sse.h"
59 #include "nb_kernel214_x86_64_sse.h"
60 #include "nb_kernel230_x86_64_sse.h"
61 #include "nb_kernel231_x86_64_sse.h"
62 #include "nb_kernel232_x86_64_sse.h"
63 #include "nb_kernel233_x86_64_sse.h"
64 #include "nb_kernel234_x86_64_sse.h"
65 #include "nb_kernel300_x86_64_sse.h"
66 #include "nb_kernel301_x86_64_sse.h"
67 #include "nb_kernel302_x86_64_sse.h"
68 #include "nb_kernel303_x86_64_sse.h"
69 #include "nb_kernel304_x86_64_sse.h"
70 #include "nb_kernel310_x86_64_sse.h"
71 #include "nb_kernel311_x86_64_sse.h"
72 #include "nb_kernel312_x86_64_sse.h"
73 #include "nb_kernel313_x86_64_sse.h"
74 #include "nb_kernel314_x86_64_sse.h"
75 #include "nb_kernel330_x86_64_sse.h"
76 #include "nb_kernel331_x86_64_sse.h"
77 #include "nb_kernel332_x86_64_sse.h"
78 #include "nb_kernel333_x86_64_sse.h"
79 #include "nb_kernel334_x86_64_sse.h"
80 #include "nb_kernel400_x86_64_sse.h"
81 #include "nb_kernel410_x86_64_sse.h"
82 #include "nb_kernel430_x86_64_sse.h"
89 /* Necessary headers for POSIX-style long jumps. */
95 #include "../nb_kerneltype.h"
96 #include "nb_kernel_x86_64_sse.h"
97 #include "nb_kernel_x86_64_sse_test_asm.h"
101 kernellist_x86_64_sse[eNR_NBKERNEL_NR] =
103 nb_kernel010_x86_64_sse,
105 nb_kernel030_x86_64_sse,
106 nb_kernel100_x86_64_sse,
107 nb_kernel101_x86_64_sse,
108 nb_kernel102_x86_64_sse,
109 nb_kernel103_x86_64_sse,
110 nb_kernel104_x86_64_sse,
111 nb_kernel110_x86_64_sse,
112 nb_kernel111_x86_64_sse,
113 nb_kernel112_x86_64_sse,
114 nb_kernel113_x86_64_sse,
115 nb_kernel114_x86_64_sse,
121 nb_kernel130_x86_64_sse,
122 nb_kernel131_x86_64_sse,
123 nb_kernel132_x86_64_sse,
124 nb_kernel133_x86_64_sse,
125 nb_kernel134_x86_64_sse,
126 nb_kernel200_x86_64_sse,
127 nb_kernel201_x86_64_sse,
128 nb_kernel202_x86_64_sse,
129 nb_kernel203_x86_64_sse,
130 nb_kernel204_x86_64_sse,
131 nb_kernel210_x86_64_sse,
132 nb_kernel211_x86_64_sse,
133 nb_kernel212_x86_64_sse,
134 nb_kernel213_x86_64_sse,
135 nb_kernel214_x86_64_sse,
141 nb_kernel230_x86_64_sse,
142 nb_kernel231_x86_64_sse,
143 nb_kernel232_x86_64_sse,
144 nb_kernel233_x86_64_sse,
145 nb_kernel234_x86_64_sse,
146 nb_kernel300_x86_64_sse,
147 nb_kernel301_x86_64_sse,
148 nb_kernel302_x86_64_sse,
149 nb_kernel303_x86_64_sse,
150 nb_kernel304_x86_64_sse,
151 nb_kernel310_x86_64_sse,
152 nb_kernel311_x86_64_sse,
153 nb_kernel312_x86_64_sse,
154 nb_kernel313_x86_64_sse,
155 nb_kernel314_x86_64_sse,
161 nb_kernel330_x86_64_sse,
162 nb_kernel331_x86_64_sse,
163 nb_kernel332_x86_64_sse,
164 nb_kernel333_x86_64_sse,
165 nb_kernel334_x86_64_sse,
166 nb_kernel400_x86_64_sse,
167 nb_kernel410_x86_64_sse,
168 nb_kernel430_x86_64_sse
172 #ifdef GMX_THREAD_SHM_FDECOMP
173 static tMPI_Thread_mutex_t
174 nb_kernel_x86_64_sse_test_mutex = TMPI_THREAD_MUTEX_INITIALIZER;
178 /*! Posix long jump label */
180 nb_kernel_x86_64_sse_testprog;
182 /*! Result of x86_64 SSE test */
184 nb_kernel_x86_64_sse_present;
188 nb_kernel_x86_64_sse_sigill_handler(int n)
190 nb_kernel_x86_64_sse_present=FALSE;
191 longjmp(nb_kernel_x86_64_sse_testprog,n);
196 /* Return 0 if SSE support is present, or
197 * non-zero on failure.
200 nb_kernel_x86_64_sse_test(FILE * log)
203 * This should NOT be called from threads,
204 * but just in case you still try to do it...
206 #ifdef GMX_THREAD_SHM_FDECOMP
207 tMPI_Thread_mutex_lock(&nb_kernel_x86_64_sse_test_mutex);
211 fprintf(log,"Testing x86_64 SSE support...");
213 nb_kernel_x86_64_sse_present = TRUE;
214 signal(SIGILL,nb_kernel_x86_64_sse_sigill_handler);
216 /* return to this point after executing the signal handler
217 * if we catch a SIGILL
219 setjmp(nb_kernel_x86_64_sse_testprog);
221 if(nb_kernel_x86_64_sse_present)
222 nb_kernel_x86_64_sse_test_asm();
224 /* If SSE worked, then success is still 1.
225 * If we got SIGILL, it was set to 0 in sigill_handler().
229 fprintf(log," %spresent.\n",
230 nb_kernel_x86_64_sse_present ? "":"not ");
232 #ifdef GMX_THREAD_SHM_FDECOMP
233 tMPI_Thread_mutex_unlock(&nb_kernel_x86_64_sse_test_mutex);
236 return ((nb_kernel_x86_64_sse_present) ? 0 : -1);
243 nb_kernel_setup_x86_64_sse(FILE *log,nb_kernel_t **list)
248 if(nb_kernel_x86_64_sse_test(log) != 0)
251 for(i=0;i<eNR_NBKERNEL_NR;i++)
253 p = kernellist_x86_64_sse[i];