annotate src/fftw-3.3.8/tests/fftw-bench.c @ 82:d0c2a83c1364

Add FFTW 3.3.8 source, and a Linux build
author Chris Cannam
date Tue, 19 Nov 2019 14:52:55 +0000
parents
children
rev   line source
Chris@82 1 /* See bench.c. We keep a few common subroutines in this file so
Chris@82 2 that they can be re-used in the MPI test program. */
Chris@82 3
Chris@82 4 #include <math.h>
Chris@82 5 #include <stdio.h>
Chris@82 6 #include <string.h>
Chris@82 7 #include "tests/fftw-bench.h"
Chris@82 8
Chris@82 9 /* define to enable code that traps floating-point exceptions.
Chris@82 10 Disabled by default because I don't want to worry about the
Chris@82 11 portability of such code. feenableexcept() seems to be a GNU
Chris@82 12 thing */
Chris@82 13 #undef TRAP_FP_EXCEPTIONS
Chris@82 14
Chris@82 15 #ifdef TRAP_FP_EXCEPTIONS
Chris@82 16 # include <signal.h>
Chris@82 17 # include <fenv.h>
Chris@82 18 #endif
Chris@82 19
Chris@82 20 #ifdef _OPENMP
Chris@82 21 # include <omp.h>
Chris@82 22 #endif
Chris@82 23
Chris@82 24 #ifdef HAVE_SMP
Chris@82 25 int threads_ok = 1;
Chris@82 26 #endif
Chris@82 27
Chris@82 28 FFTW(plan) the_plan = 0;
Chris@82 29
Chris@82 30 static const char *wisdat = "wis.dat";
Chris@82 31 unsigned the_flags = 0;
Chris@82 32 int paranoid = 0;
Chris@82 33 int usewisdom = 0;
Chris@82 34 int havewisdom = 0;
Chris@82 35 int nthreads = 1;
Chris@82 36 int amnesia = 0;
Chris@82 37
Chris@82 38 extern void install_hook(void); /* in hook.c */
Chris@82 39 extern void uninstall_hook(void); /* in hook.c */
Chris@82 40
Chris@82 41 #ifdef FFTW_RANDOM_ESTIMATOR
Chris@82 42 extern unsigned FFTW(random_estimate_seed);
Chris@82 43 #endif
Chris@82 44
Chris@82 45 #ifdef TRAP_FP_EXCEPTIONS
Chris@82 46 static void sigfpe_handler(int sig, siginfo_t *info, void *context)
Chris@82 47 {
Chris@82 48 /* fftw code is not supposed to generate FP exceptions */
Chris@82 49 UNUSED(sig); UNUSED(info); UNUSED(context);
Chris@82 50 fprintf(stderr, "caught FPE, aborting\n");
Chris@82 51 abort();
Chris@82 52 }
Chris@82 53
Chris@82 54 static void setup_sigfpe_handler(void)
Chris@82 55 {
Chris@82 56 struct sigaction a;
Chris@82 57 feenableexcept(FE_DIVBYZERO | FE_INVALID | FE_OVERFLOW | FE_UNDERFLOW);
Chris@82 58 memset(&a, 0, sizeof(a));
Chris@82 59 a.sa_sigaction = sigfpe_handler;
Chris@82 60 a.sa_flags = SA_SIGINFO;
Chris@82 61 if (sigaction(SIGFPE, &a, NULL) == -1) {
Chris@82 62 fprintf(stderr, "cannot install sigfpe handler\n");
Chris@82 63 exit(1);
Chris@82 64 }
Chris@82 65 }
Chris@82 66 #else
Chris@82 67 static void setup_sigfpe_handler(void)
Chris@82 68 {
Chris@82 69 }
Chris@82 70 #endif
Chris@82 71
Chris@82 72 void useropt(const char *arg)
Chris@82 73 {
Chris@82 74 int x;
Chris@82 75 double y;
Chris@82 76
Chris@82 77 if (!strcmp(arg, "patient")) the_flags |= FFTW_PATIENT;
Chris@82 78 else if (!strcmp(arg, "estimate")) the_flags |= FFTW_ESTIMATE;
Chris@82 79 else if (!strcmp(arg, "estimatepat")) the_flags |= FFTW_ESTIMATE_PATIENT;
Chris@82 80 else if (!strcmp(arg, "exhaustive")) the_flags |= FFTW_EXHAUSTIVE;
Chris@82 81 else if (!strcmp(arg, "unaligned")) the_flags |= FFTW_UNALIGNED;
Chris@82 82 else if (!strcmp(arg, "nosimd")) the_flags |= FFTW_NO_SIMD;
Chris@82 83 else if (!strcmp(arg, "noindirectop")) the_flags |= FFTW_NO_INDIRECT_OP;
Chris@82 84 else if (!strcmp(arg, "wisdom-only")) the_flags |= FFTW_WISDOM_ONLY;
Chris@82 85 else if (sscanf(arg, "flag=%d", &x) == 1) the_flags |= x;
Chris@82 86 else if (sscanf(arg, "bflag=%d", &x) == 1) the_flags |= 1U << x;
Chris@82 87 else if (!strcmp(arg, "paranoid")) paranoid = 1;
Chris@82 88 else if (!strcmp(arg, "wisdom")) usewisdom = 1;
Chris@82 89 else if (!strcmp(arg, "amnesia")) amnesia = 1;
Chris@82 90 else if (sscanf(arg, "nthreads=%d", &x) == 1) nthreads = x;
Chris@82 91 #ifdef FFTW_RANDOM_ESTIMATOR
Chris@82 92 else if (sscanf(arg, "eseed=%d", &x) == 1) FFTW(random_estimate_seed) = x;
Chris@82 93 #endif
Chris@82 94 else if (sscanf(arg, "timelimit=%lg", &y) == 1) {
Chris@82 95 FFTW(set_timelimit)(y);
Chris@82 96 }
Chris@82 97
Chris@82 98 else fprintf(stderr, "unknown user option: %s. Ignoring.\n", arg);
Chris@82 99 }
Chris@82 100
Chris@82 101 void rdwisdom(void)
Chris@82 102 {
Chris@82 103 FILE *f;
Chris@82 104 double tim;
Chris@82 105 int success = 0;
Chris@82 106
Chris@82 107 if (havewisdom) return;
Chris@82 108
Chris@82 109 #ifdef HAVE_SMP
Chris@82 110 if (threads_ok) {
Chris@82 111 BENCH_ASSERT(FFTW(init_threads)());
Chris@82 112 FFTW(plan_with_nthreads)(nthreads);
Chris@82 113 FFTW(make_planner_thread_safe)();
Chris@82 114 #ifdef _OPENMP
Chris@82 115 omp_set_num_threads(nthreads);
Chris@82 116 #endif
Chris@82 117 }
Chris@82 118 else if (nthreads > 1 && verbose > 1) {
Chris@82 119 fprintf(stderr, "bench: WARNING - nthreads = %d, but threads not supported\n", nthreads);
Chris@82 120 nthreads = 1;
Chris@82 121 }
Chris@82 122 #endif
Chris@82 123
Chris@82 124 if (!usewisdom) return;
Chris@82 125
Chris@82 126 timer_start(USER_TIMER);
Chris@82 127 if ((f = fopen(wisdat, "r"))) {
Chris@82 128 if (!import_wisdom(f))
Chris@82 129 fprintf(stderr, "bench: ERROR reading wisdom\n");
Chris@82 130 else
Chris@82 131 success = 1;
Chris@82 132 fclose(f);
Chris@82 133 }
Chris@82 134 tim = timer_stop(USER_TIMER);
Chris@82 135
Chris@82 136 if (success) {
Chris@82 137 if (verbose > 1) printf("READ WISDOM (%g seconds): ", tim);
Chris@82 138
Chris@82 139 if (verbose > 3)
Chris@82 140 export_wisdom(stdout);
Chris@82 141 if (verbose > 1)
Chris@82 142 printf("\n");
Chris@82 143 }
Chris@82 144 havewisdom = 1;
Chris@82 145 }
Chris@82 146
Chris@82 147 void wrwisdom(void)
Chris@82 148 {
Chris@82 149 FILE *f;
Chris@82 150 double tim;
Chris@82 151 if (!havewisdom) return;
Chris@82 152
Chris@82 153 timer_start(USER_TIMER);
Chris@82 154 if ((f = fopen(wisdat, "w"))) {
Chris@82 155 export_wisdom(f);
Chris@82 156 fclose(f);
Chris@82 157 }
Chris@82 158 tim = timer_stop(USER_TIMER);
Chris@82 159 if (verbose > 1) printf("write wisdom took %g seconds\n", tim);
Chris@82 160 }
Chris@82 161
Chris@82 162 static unsigned preserve_input_flags(bench_problem *p)
Chris@82 163 {
Chris@82 164 /*
Chris@82 165 * fftw3 cannot preserve input for multidimensional c2r transforms.
Chris@82 166 * Enforce FFTW_DESTROY_INPUT
Chris@82 167 */
Chris@82 168 if (p->kind == PROBLEM_REAL &&
Chris@82 169 p->sign > 0 &&
Chris@82 170 !p->in_place &&
Chris@82 171 p->sz->rnk > 1)
Chris@82 172 p->destroy_input = 1;
Chris@82 173
Chris@82 174 if (p->destroy_input)
Chris@82 175 return FFTW_DESTROY_INPUT;
Chris@82 176 else
Chris@82 177 return FFTW_PRESERVE_INPUT;
Chris@82 178 }
Chris@82 179
Chris@82 180 int can_do(bench_problem *p)
Chris@82 181 {
Chris@82 182 double tim;
Chris@82 183
Chris@82 184 if (verbose > 2 && p->pstring)
Chris@82 185 printf("Planning %s...\n", p->pstring);
Chris@82 186 rdwisdom();
Chris@82 187
Chris@82 188 timer_start(USER_TIMER);
Chris@82 189 the_plan = mkplan(p, preserve_input_flags(p) | the_flags | FFTW_ESTIMATE);
Chris@82 190 tim = timer_stop(USER_TIMER);
Chris@82 191 if (verbose > 2) printf("estimate-planner time: %g s\n", tim);
Chris@82 192
Chris@82 193 if (the_plan) {
Chris@82 194 FFTW(destroy_plan)(the_plan);
Chris@82 195 return 1;
Chris@82 196 }
Chris@82 197 return 0;
Chris@82 198 }
Chris@82 199
Chris@82 200 void setup(bench_problem *p)
Chris@82 201 {
Chris@82 202 double tim;
Chris@82 203
Chris@82 204 setup_sigfpe_handler();
Chris@82 205
Chris@82 206 if (amnesia) {
Chris@82 207 FFTW(forget_wisdom)();
Chris@82 208 havewisdom = 0;
Chris@82 209 }
Chris@82 210
Chris@82 211 /* Regression test: check that fftw_malloc exists and links
Chris@82 212 * properly */
Chris@82 213 {
Chris@82 214 void *ptr = FFTW(malloc(42));
Chris@82 215 BENCH_ASSERT(FFTW(alignment_of)(ptr) == 0);
Chris@82 216 FFTW(free(ptr));
Chris@82 217 }
Chris@82 218
Chris@82 219 rdwisdom();
Chris@82 220 install_hook();
Chris@82 221
Chris@82 222 #ifdef HAVE_SMP
Chris@82 223 if (verbose > 1 && nthreads > 1) printf("NTHREADS = %d\n", nthreads);
Chris@82 224 #endif
Chris@82 225
Chris@82 226 timer_start(USER_TIMER);
Chris@82 227 the_plan = mkplan(p, preserve_input_flags(p) | the_flags);
Chris@82 228 tim = timer_stop(USER_TIMER);
Chris@82 229 if (verbose > 1) printf("planner time: %g s\n", tim);
Chris@82 230
Chris@82 231 BENCH_ASSERT(the_plan);
Chris@82 232
Chris@82 233 {
Chris@82 234 double add, mul, nfma, cost, pcost;
Chris@82 235 FFTW(flops)(the_plan, &add, &mul, &nfma);
Chris@82 236 cost = FFTW(estimate_cost)(the_plan);
Chris@82 237 pcost = FFTW(cost)(the_plan);
Chris@82 238 if (verbose > 1) {
Chris@82 239 FFTW(print_plan)(the_plan);
Chris@82 240 printf("\n");
Chris@82 241 printf("flops: %0.0f add, %0.0f mul, %0.0f fma\n",
Chris@82 242 add, mul, nfma);
Chris@82 243 printf("estimated cost: %f, pcost = %f\n", cost, pcost);
Chris@82 244 }
Chris@82 245 }
Chris@82 246 }
Chris@82 247
Chris@82 248
Chris@82 249 void doit(int iter, bench_problem *p)
Chris@82 250 {
Chris@82 251 int i;
Chris@82 252 FFTW(plan) q = the_plan;
Chris@82 253
Chris@82 254 UNUSED(p);
Chris@82 255 for (i = 0; i < iter; ++i)
Chris@82 256 FFTW(execute)(q);
Chris@82 257 }
Chris@82 258
Chris@82 259 void done(bench_problem *p)
Chris@82 260 {
Chris@82 261 UNUSED(p);
Chris@82 262
Chris@82 263 FFTW(destroy_plan)(the_plan);
Chris@82 264 uninstall_hook();
Chris@82 265 }
Chris@82 266
Chris@82 267 void cleanup(void)
Chris@82 268 {
Chris@82 269 initial_cleanup();
Chris@82 270
Chris@82 271 wrwisdom();
Chris@82 272 #ifdef HAVE_SMP
Chris@82 273 FFTW(cleanup_threads)();
Chris@82 274 #else
Chris@82 275 FFTW(cleanup)();
Chris@82 276 #endif
Chris@82 277
Chris@82 278 # ifdef FFTW_DEBUG_MALLOC
Chris@82 279 {
Chris@82 280 /* undocumented memory checker */
Chris@82 281 FFTW_EXTERN void FFTW(malloc_print_minfo)(int v);
Chris@82 282 FFTW(malloc_print_minfo)(verbose);
Chris@82 283 }
Chris@82 284 # endif
Chris@82 285
Chris@82 286 final_cleanup();
Chris@82 287 }