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_ia32_sse2.h"
32 /* Include ia32 SSE kernel headers in local directory */
33 #include "nb_kernel010_ia32_sse2.h"
34 #include "nb_kernel030_ia32_sse2.h"
35 #include "nb_kernel100_ia32_sse2.h"
36 #include "nb_kernel101_ia32_sse2.h"
37 #include "nb_kernel102_ia32_sse2.h"
38 #include "nb_kernel103_ia32_sse2.h"
39 #include "nb_kernel104_ia32_sse2.h"
40 #include "nb_kernel110_ia32_sse2.h"
41 #include "nb_kernel111_ia32_sse2.h"
42 #include "nb_kernel112_ia32_sse2.h"
43 #include "nb_kernel113_ia32_sse2.h"
44 #include "nb_kernel114_ia32_sse2.h"
45 #include "nb_kernel130_ia32_sse2.h"
46 #include "nb_kernel131_ia32_sse2.h"
47 #include "nb_kernel132_ia32_sse2.h"
48 #include "nb_kernel133_ia32_sse2.h"
49 #include "nb_kernel134_ia32_sse2.h"
50 #include "nb_kernel200_ia32_sse2.h"
51 #include "nb_kernel201_ia32_sse2.h"
52 #include "nb_kernel202_ia32_sse2.h"
53 #include "nb_kernel203_ia32_sse2.h"
54 #include "nb_kernel204_ia32_sse2.h"
55 #include "nb_kernel210_ia32_sse2.h"
56 #include "nb_kernel211_ia32_sse2.h"
57 #include "nb_kernel212_ia32_sse2.h"
58 #include "nb_kernel213_ia32_sse2.h"
59 #include "nb_kernel214_ia32_sse2.h"
60 #include "nb_kernel230_ia32_sse2.h"
61 #include "nb_kernel231_ia32_sse2.h"
62 #include "nb_kernel232_ia32_sse2.h"
63 #include "nb_kernel233_ia32_sse2.h"
64 #include "nb_kernel234_ia32_sse2.h"
65 #include "nb_kernel300_ia32_sse2.h"
66 #include "nb_kernel301_ia32_sse2.h"
67 #include "nb_kernel302_ia32_sse2.h"
68 #include "nb_kernel303_ia32_sse2.h"
69 #include "nb_kernel304_ia32_sse2.h"
70 #include "nb_kernel310_ia32_sse2.h"
71 #include "nb_kernel311_ia32_sse2.h"
72 #include "nb_kernel312_ia32_sse2.h"
73 #include "nb_kernel313_ia32_sse2.h"
74 #include "nb_kernel314_ia32_sse2.h"
75 #include "nb_kernel330_ia32_sse2.h"
76 #include "nb_kernel331_ia32_sse2.h"
77 #include "nb_kernel332_ia32_sse2.h"
78 #include "nb_kernel333_ia32_sse2.h"
79 #include "nb_kernel334_ia32_sse2.h"
80 #include "nb_kernel400_ia32_sse2.h"
81 #include "nb_kernel410_ia32_sse2.h"
82 #include "nb_kernel430_ia32_sse2.h"
87 /* Necessary headers for POSIX-style long jumps. */
94 #include "../nb_kerneltype.h"
95 #include "nb_kernel_ia32_sse2.h"
96 #include "nb_kernel_ia32_sse2_test_asm.h"
99 kernellist_ia32_sse2[eNR_NBKERNEL_NR] =
101 nb_kernel010_ia32_sse2,
103 nb_kernel030_ia32_sse2,
104 nb_kernel100_ia32_sse2,
105 nb_kernel101_ia32_sse2,
106 nb_kernel102_ia32_sse2,
107 nb_kernel103_ia32_sse2,
108 nb_kernel104_ia32_sse2,
109 nb_kernel110_ia32_sse2,
110 nb_kernel111_ia32_sse2,
111 nb_kernel112_ia32_sse2,
112 nb_kernel113_ia32_sse2,
113 nb_kernel114_ia32_sse2,
119 nb_kernel130_ia32_sse2,
120 nb_kernel131_ia32_sse2,
121 nb_kernel132_ia32_sse2,
122 nb_kernel133_ia32_sse2,
123 nb_kernel134_ia32_sse2,
124 nb_kernel200_ia32_sse2,
125 nb_kernel201_ia32_sse2,
126 nb_kernel202_ia32_sse2,
127 nb_kernel203_ia32_sse2,
128 nb_kernel204_ia32_sse2,
129 nb_kernel210_ia32_sse2,
130 nb_kernel211_ia32_sse2,
131 nb_kernel212_ia32_sse2,
132 nb_kernel213_ia32_sse2,
133 nb_kernel214_ia32_sse2,
139 nb_kernel230_ia32_sse2,
140 nb_kernel231_ia32_sse2,
141 nb_kernel232_ia32_sse2,
142 nb_kernel233_ia32_sse2,
143 nb_kernel234_ia32_sse2,
144 nb_kernel300_ia32_sse2,
145 nb_kernel301_ia32_sse2,
146 nb_kernel302_ia32_sse2,
147 nb_kernel303_ia32_sse2,
148 nb_kernel304_ia32_sse2,
149 nb_kernel310_ia32_sse2,
150 nb_kernel311_ia32_sse2,
151 nb_kernel312_ia32_sse2,
152 nb_kernel313_ia32_sse2,
153 nb_kernel314_ia32_sse2,
159 nb_kernel330_ia32_sse2,
160 nb_kernel331_ia32_sse2,
161 nb_kernel332_ia32_sse2,
162 nb_kernel333_ia32_sse2,
163 nb_kernel334_ia32_sse2,
164 nb_kernel400_ia32_sse2,
165 nb_kernel410_ia32_sse2,
166 nb_kernel430_ia32_sse2
169 #ifdef GMX_THREAD_SHM_FDECOMP
170 static tMPI_Thread_mutex_t
171 nb_kernel_ia32_sse2_test_mutex = TMPI_THREAD_MUTEX_INITIALIZER;
174 /*! Posix long jump label */
176 nb_kernel_ia32_sse2_testprog;
178 /*! Result of ia32 SSE2 test */
180 nb_kernel_ia32_sse2_present;
184 nb_kernel_ia32_sse2_sigill_handler(int n)
186 nb_kernel_ia32_sse2_present=FALSE;
187 longjmp(nb_kernel_ia32_sse2_testprog,n);
194 /* Return 0 if SSE2 support is present, or
195 * or non-zero on failure.
198 nb_kernel_ia32_sse2_test(FILE * log)
201 * This should NOT be called from threads,
202 * but just in case you still try to do it...
204 #ifdef GMX_THREAD_SHM_FDECOMP
205 tMPI_Thread_mutex_lock(&nb_kernel_ia32_sse2_test_mutex);
209 fprintf(log,"Testing ia32 SSE2 support...");
211 nb_kernel_ia32_sse2_present = TRUE;
212 signal(SIGILL,nb_kernel_ia32_sse2_sigill_handler);
214 /* return to this point after executing the signal handler
215 * if we catch a SIGILL
217 setjmp(nb_kernel_ia32_sse2_testprog);
219 if(nb_kernel_ia32_sse2_present)
220 nb_kernel_ia32_sse2_test_asm();
222 /* If SSE2 worked, then success is still 1.
223 * If we got SIGILL, it was set to 0 in sigill_handler().
227 fprintf(log," %spresent.\n",
228 nb_kernel_ia32_sse2_present ? "":"not ");
230 #ifdef GMX_THREAD_SHM_FDECOMP
231 tMPI_Thread_mutex_unlock(&nb_kernel_ia32_sse2_test_mutex);
234 return ((nb_kernel_ia32_sse2_present) ? 0 : -1);
241 nb_kernel_setup_ia32_sse2(FILE *log,nb_kernel_t **list)
246 if(nb_kernel_ia32_sse2_test(log) != 0)
249 for(i=0;i<eNR_NBKERNEL_NR;i++)
251 p = kernellist_ia32_sse2[i];