Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.5/tests/hook.c @ 42:2cd0e3b3e1fd

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Current fftw source
author Chris Cannam
date Tue, 18 Oct 2016 13:40:26 +0100
parents
children
rev   line source
Chris@42 1 /* fftw hook to be used in the benchmark program.
Chris@42 2
Chris@42 3 We keep it in a separate file because
Chris@42 4
Chris@42 5 1) bench.c is supposed to test the API---we do not want to #include
Chris@42 6 "ifftw.h" and accidentally use internal symbols/macros.
Chris@42 7 2) this code is a royal mess. The messiness is due to
Chris@42 8 A) confusion between internal fftw tensors and bench_tensor's
Chris@42 9 (which we want to keep separate because the benchmark
Chris@42 10 program tests other routines too)
Chris@42 11 B) despite A), our desire to recycle the libbench verifier.
Chris@42 12 */
Chris@42 13
Chris@42 14 #include <stdio.h>
Chris@42 15 #include "bench-user.h"
Chris@42 16
Chris@42 17 #define CALLING_FFTW /* hack for Windows DLL nonsense */
Chris@42 18 #include "api.h"
Chris@42 19 #include "dft.h"
Chris@42 20 #include "rdft.h"
Chris@42 21
Chris@42 22 extern int paranoid; /* in bench.c */
Chris@42 23 extern X(plan) the_plan; /* in bench.c */
Chris@42 24
Chris@42 25 /*
Chris@42 26 transform an fftw tensor into a bench_tensor.
Chris@42 27 */
Chris@42 28 static bench_tensor *fftw_tensor_to_bench_tensor(tensor *t)
Chris@42 29 {
Chris@42 30 bench_tensor *bt = mktensor(t->rnk);
Chris@42 31
Chris@42 32 if (FINITE_RNK(t->rnk)) {
Chris@42 33 int i;
Chris@42 34 for (i = 0; i < t->rnk; ++i) {
Chris@42 35 /* FIXME: 64-bit unclean because of INT -> int conversion */
Chris@42 36 bt->dims[i].n = t->dims[i].n;
Chris@42 37 bt->dims[i].is = t->dims[i].is;
Chris@42 38 bt->dims[i].os = t->dims[i].os;
Chris@42 39 BENCH_ASSERT(bt->dims[i].n == t->dims[i].n);
Chris@42 40 BENCH_ASSERT(bt->dims[i].is == t->dims[i].is);
Chris@42 41 BENCH_ASSERT(bt->dims[i].os == t->dims[i].os);
Chris@42 42 }
Chris@42 43 }
Chris@42 44 return bt;
Chris@42 45 }
Chris@42 46
Chris@42 47 /*
Chris@42 48 transform an fftw problem into a bench_problem.
Chris@42 49 */
Chris@42 50 static bench_problem *fftw_problem_to_bench_problem(planner *plnr,
Chris@42 51 const problem *p_)
Chris@42 52 {
Chris@42 53 bench_problem *bp = 0;
Chris@42 54 switch (p_->adt->problem_kind) {
Chris@42 55 case PROBLEM_DFT:
Chris@42 56 {
Chris@42 57 const problem_dft *p = (const problem_dft *) p_;
Chris@42 58
Chris@42 59 if (!p->ri || !p->ii)
Chris@42 60 abort();
Chris@42 61
Chris@42 62 bp = (bench_problem *) bench_malloc(sizeof(bench_problem));
Chris@42 63
Chris@42 64 bp->kind = PROBLEM_COMPLEX;
Chris@42 65 bp->sign = FFT_SIGN;
Chris@42 66 bp->split = 1; /* tensor strides are in R's, not C's */
Chris@42 67 bp->in = UNTAINT(p->ri);
Chris@42 68 bp->out = UNTAINT(p->ro);
Chris@42 69 bp->ini = UNTAINT(p->ii);
Chris@42 70 bp->outi = UNTAINT(p->io);
Chris@42 71 bp->inphys = bp->outphys = 0;
Chris@42 72 bp->iphyssz = bp->ophyssz = 0;
Chris@42 73 bp->in_place = p->ri == p->ro;
Chris@42 74 bp->sz = fftw_tensor_to_bench_tensor(p->sz);
Chris@42 75 bp->vecsz = fftw_tensor_to_bench_tensor(p->vecsz);
Chris@42 76 bp->k = 0;
Chris@42 77 break;
Chris@42 78 }
Chris@42 79 case PROBLEM_RDFT:
Chris@42 80 {
Chris@42 81 const problem_rdft *p = (const problem_rdft *) p_;
Chris@42 82 int i;
Chris@42 83
Chris@42 84 if (!p->I || !p->O)
Chris@42 85 abort();
Chris@42 86
Chris@42 87 for (i = 0; i < p->sz->rnk; ++i)
Chris@42 88 switch (p->kind[i]) {
Chris@42 89 case R2HC01:
Chris@42 90 case R2HC10:
Chris@42 91 case R2HC11:
Chris@42 92 case HC2R01:
Chris@42 93 case HC2R10:
Chris@42 94 case HC2R11:
Chris@42 95 return bp;
Chris@42 96 default:
Chris@42 97 ;
Chris@42 98 }
Chris@42 99
Chris@42 100 bp = (bench_problem *) bench_malloc(sizeof(bench_problem));
Chris@42 101
Chris@42 102 bp->kind = PROBLEM_R2R;
Chris@42 103 bp->sign = FFT_SIGN;
Chris@42 104 bp->split = 0;
Chris@42 105 bp->in = UNTAINT(p->I);
Chris@42 106 bp->out = UNTAINT(p->O);
Chris@42 107 bp->ini = bp->outi = 0;
Chris@42 108 bp->inphys = bp->outphys = 0;
Chris@42 109 bp->iphyssz = bp->ophyssz = 0;
Chris@42 110 bp->in_place = p->I == p->O;
Chris@42 111 bp->sz = fftw_tensor_to_bench_tensor(p->sz);
Chris@42 112 bp->vecsz = fftw_tensor_to_bench_tensor(p->vecsz);
Chris@42 113 bp->k = (r2r_kind_t *) bench_malloc(sizeof(r2r_kind_t) * p->sz->rnk);
Chris@42 114 for (i = 0; i < p->sz->rnk; ++i)
Chris@42 115 switch (p->kind[i]) {
Chris@42 116 case R2HC: bp->k[i] = R2R_R2HC; break;
Chris@42 117 case HC2R: bp->k[i] = R2R_HC2R; break;
Chris@42 118 case DHT: bp->k[i] = R2R_DHT; break;
Chris@42 119 case REDFT00: bp->k[i] = R2R_REDFT00; break;
Chris@42 120 case REDFT01: bp->k[i] = R2R_REDFT01; break;
Chris@42 121 case REDFT10: bp->k[i] = R2R_REDFT10; break;
Chris@42 122 case REDFT11: bp->k[i] = R2R_REDFT11; break;
Chris@42 123 case RODFT00: bp->k[i] = R2R_RODFT00; break;
Chris@42 124 case RODFT01: bp->k[i] = R2R_RODFT01; break;
Chris@42 125 case RODFT10: bp->k[i] = R2R_RODFT10; break;
Chris@42 126 case RODFT11: bp->k[i] = R2R_RODFT11; break;
Chris@42 127 default: CK(0);
Chris@42 128 }
Chris@42 129 break;
Chris@42 130 }
Chris@42 131 case PROBLEM_RDFT2:
Chris@42 132 {
Chris@42 133 const problem_rdft2 *p = (const problem_rdft2 *) p_;
Chris@42 134 int rnk = p->sz->rnk;
Chris@42 135
Chris@42 136 if (!p->r0 || !p->r1 || !p->cr || !p->ci)
Chris@42 137 abort();
Chris@42 138
Chris@42 139 /* give up verifying rdft2 R2HCII */
Chris@42 140 if (p->kind != R2HC && p->kind != HC2R)
Chris@42 141 return bp;
Chris@42 142
Chris@42 143 if (rnk > 0) {
Chris@42 144 /* can't verify separate even/odd arrays for now */
Chris@42 145 if (2 * (p->r1 - p->r0) !=
Chris@42 146 ((p->kind == R2HC) ?
Chris@42 147 p->sz->dims[rnk-1].is : p->sz->dims[rnk-1].os))
Chris@42 148 return bp;
Chris@42 149 }
Chris@42 150
Chris@42 151 bp = (bench_problem *) bench_malloc(sizeof(bench_problem));
Chris@42 152
Chris@42 153 bp->kind = PROBLEM_REAL;
Chris@42 154 bp->sign = p->kind == R2HC ? FFT_SIGN : -FFT_SIGN;
Chris@42 155 bp->split = 1; /* tensor strides are in R's, not C's */
Chris@42 156 if (p->kind == R2HC) {
Chris@42 157 bp->sign = FFT_SIGN;
Chris@42 158 bp->in = UNTAINT(p->r0);
Chris@42 159 bp->out = UNTAINT(p->cr);
Chris@42 160 bp->ini = 0;
Chris@42 161 bp->outi = UNTAINT(p->ci);
Chris@42 162 }
Chris@42 163 else {
Chris@42 164 bp->sign = -FFT_SIGN;
Chris@42 165 bp->out = UNTAINT(p->r0);
Chris@42 166 bp->in = UNTAINT(p->cr);
Chris@42 167 bp->outi = 0;
Chris@42 168 bp->ini = UNTAINT(p->ci);
Chris@42 169 }
Chris@42 170 bp->inphys = bp->outphys = 0;
Chris@42 171 bp->iphyssz = bp->ophyssz = 0;
Chris@42 172 bp->in_place = p->r0 == p->cr;
Chris@42 173 bp->sz = fftw_tensor_to_bench_tensor(p->sz);
Chris@42 174 if (rnk > 0) {
Chris@42 175 if (p->kind == R2HC)
Chris@42 176 bp->sz->dims[rnk-1].is /= 2;
Chris@42 177 else
Chris@42 178 bp->sz->dims[rnk-1].os /= 2;
Chris@42 179 }
Chris@42 180 bp->vecsz = fftw_tensor_to_bench_tensor(p->vecsz);
Chris@42 181 bp->k = 0;
Chris@42 182 break;
Chris@42 183 }
Chris@42 184 default:
Chris@42 185 abort();
Chris@42 186 }
Chris@42 187
Chris@42 188 bp->userinfo = 0;
Chris@42 189 bp->pstring = 0;
Chris@42 190 bp->destroy_input = !NO_DESTROY_INPUTP(plnr);
Chris@42 191
Chris@42 192 return bp;
Chris@42 193 }
Chris@42 194
Chris@42 195 static void hook(planner *plnr, plan *pln, const problem *p_, int optimalp)
Chris@42 196 {
Chris@42 197 int rounds = 5;
Chris@42 198 double tol = SINGLE_PRECISION ? 1.0e-3 : 1.0e-10;
Chris@42 199 UNUSED(optimalp);
Chris@42 200
Chris@42 201 if (verbose > 5) {
Chris@42 202 printer *pr = X(mkprinter_file)(stdout);
Chris@42 203 pr->print(pr, "%P:%(%p%)\n", p_, pln);
Chris@42 204 X(printer_destroy)(pr);
Chris@42 205 printf("cost %g \n\n", pln->pcost);
Chris@42 206 }
Chris@42 207
Chris@42 208 if (paranoid) {
Chris@42 209 bench_problem *bp;
Chris@42 210
Chris@42 211 bp = fftw_problem_to_bench_problem(plnr, p_);
Chris@42 212 if (bp) {
Chris@42 213 X(plan) the_plan_save = the_plan;
Chris@42 214
Chris@42 215 the_plan = (apiplan *) MALLOC(sizeof(apiplan), PLANS);
Chris@42 216 the_plan->pln = pln;
Chris@42 217 the_plan->prb = (problem *) p_;
Chris@42 218
Chris@42 219 X(plan_awake)(pln, AWAKE_SQRTN_TABLE);
Chris@42 220 verify_problem(bp, rounds, tol);
Chris@42 221 X(plan_awake)(pln, SLEEPY);
Chris@42 222
Chris@42 223 X(ifree)(the_plan);
Chris@42 224 the_plan = the_plan_save;
Chris@42 225
Chris@42 226 problem_destroy(bp);
Chris@42 227 }
Chris@42 228
Chris@42 229 }
Chris@42 230 }
Chris@42 231
Chris@42 232 static void paranoid_checks(void)
Chris@42 233 {
Chris@42 234 /* FIXME: assumes char = 8 bits, which is false on at least one
Chris@42 235 DSP I know of. */
Chris@42 236 #if 0
Chris@42 237 /* if flags_t is not 64 bits i want to know it. */
Chris@42 238 CK(sizeof(flags_t) == 8);
Chris@42 239
Chris@42 240 CK(sizeof(md5uint) >= 4);
Chris@42 241 #endif
Chris@42 242
Chris@42 243 CK(sizeof(uintptr_t) >= sizeof(R *));
Chris@42 244
Chris@42 245 CK(sizeof(INT) >= sizeof(R *));
Chris@42 246 }
Chris@42 247
Chris@42 248 void install_hook(void)
Chris@42 249 {
Chris@42 250 planner *plnr = X(the_planner)();
Chris@42 251 plnr->hook = hook;
Chris@42 252 paranoid_checks();
Chris@42 253 }
Chris@42 254
Chris@42 255 void uninstall_hook(void)
Chris@42 256 {
Chris@42 257 planner *plnr = X(the_planner)();
Chris@42 258 plnr->hook = 0;
Chris@42 259 }