| 1 | /* -*-c-*- |
| 2 | * |
| 3 | * Benchmarking in the test-vector framework |
| 4 | * |
| 5 | * (c) 2023 Straylight/Edgeware |
| 6 | */ |
| 7 | |
| 8 | /*----- Licensing notice --------------------------------------------------* |
| 9 | * |
| 10 | * This file is part of the mLib utilities library. |
| 11 | * |
| 12 | * mLib is free software: you can redistribute it and/or modify it under |
| 13 | * the terms of the GNU Library General Public License as published by |
| 14 | * the Free Software Foundation; either version 2 of the License, or (at |
| 15 | * your option) any later version. |
| 16 | * |
| 17 | * mLib is distributed in the hope that it will be useful, but WITHOUT |
| 18 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
| 19 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU Library General Public |
| 20 | * License for more details. |
| 21 | * |
| 22 | * You should have received a copy of the GNU Library General Public |
| 23 | * License along with mLib. If not, write to the Free Software |
| 24 | * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, |
| 25 | * USA. |
| 26 | */ |
| 27 | |
| 28 | /*----- Header files ------------------------------------------------------*/ |
| 29 | |
| 30 | #include "bench.h" |
| 31 | #include "tvec.h" |
| 32 | |
| 33 | /*----- Data structures ---------------------------------------------------*/ |
| 34 | |
| 35 | struct benchrun { |
| 36 | struct tvec_state *tv; /* test vector state */ |
| 37 | const struct tvec_env *env; /* subordinate environment */ |
| 38 | void *ctx; /* subordinate env's context */ |
| 39 | unsigned long *n; /* iteration count address */ |
| 40 | const struct tvec_reg *in; struct tvec_reg *out; /* register vectors */ |
| 41 | tvec_testfn *fn; /* test function to run */ |
| 42 | }; |
| 43 | |
| 44 | /*----- Global variables --------------------------------------------------*/ |
| 45 | |
| 46 | struct bench_state *tvec_benchstate; /* common benchmarking state */ |
| 47 | |
| 48 | /*----- Utilities ---------------------------------------------------------*/ |
| 49 | |
| 50 | /* --- @normalize@ --- * |
| 51 | * |
| 52 | * Arguments: @double *x_inout@ = address of a value to normalize |
| 53 | * @const char **unit_out@ = address to store unit prefix |
| 54 | * @double scale@ = scale factor for unit steps |
| 55 | * |
| 56 | * Returns: --- |
| 57 | * |
| 58 | * Use: Adjust @*x_inout@ by a power of @scale@, and set @*unit_out@ |
| 59 | * so that printing the two reflects the original value with an |
| 60 | * appropriate SI unit scaling. The @scale@ should be 1024 for |
| 61 | * binary quantities, most notably memory sizes, or 1000 for |
| 62 | * other quantities. |
| 63 | */ |
| 64 | |
| 65 | static void normalize(double *x_inout, const char **unit_out, double scale) |
| 66 | { |
| 67 | static const char |
| 68 | *const nothing = "", |
| 69 | *const big[] = { "k", "M", "G", "T", "P", "E", 0 }, |
| 70 | *const little[] = { "m", "ยต", "n", "p", "f", "a", 0 }; |
| 71 | const char *const *u; |
| 72 | double x = *x_inout; |
| 73 | |
| 74 | if (x < 1) |
| 75 | for (u = little, x *= scale; x < 1 && u[1]; u++, x *= scale); |
| 76 | else if (x >= scale) |
| 77 | for (u = big, x /= scale; x >= scale && u[1]; u++, x /= scale); |
| 78 | else |
| 79 | u = ¬hing; |
| 80 | |
| 81 | *x_inout = x; *unit_out = *u; |
| 82 | } |
| 83 | |
| 84 | /*----- Benchmark environment scaffolding ---------------------------------*/ |
| 85 | |
| 86 | /* --- @tvec_benchsetup@ --- * |
| 87 | * |
| 88 | * Arguments: @struct tvec_state *tv@ = test vector state |
| 89 | * @const struct tvec_env *env@ = environment description |
| 90 | * @void *pctx@ = parent context (ignored) |
| 91 | * @void *ctx@ = context pointer to initialize |
| 92 | * |
| 93 | * Returns: --- |
| 94 | * |
| 95 | * Use: Initialize a benchmarking environment context. |
| 96 | * |
| 97 | * The environment description must really be a @struct |
| 98 | * tvec_benchenv@. If the @bst@ slot is null, then a temporary |
| 99 | * benchmark state is allocated for the current test group and |
| 100 | * released at the end. Otherwise, it must be the address of a |
| 101 | * pointer to a benchmark state: if the pointer is null, then a |
| 102 | * fresh state is allocated and initialized and the pointer is |
| 103 | * updated; otherwise, the pointer is assumed to refer to an |
| 104 | * existing valid benchmark state. |
| 105 | */ |
| 106 | |
| 107 | void tvec_benchsetup(struct tvec_state *tv, const struct tvec_env *env, |
| 108 | void *pctx, void *ctx) |
| 109 | { |
| 110 | struct tvec_benchctx *bc = ctx; |
| 111 | const struct tvec_benchenv *be = (const struct tvec_benchenv *)env; |
| 112 | const struct tvec_env *subenv = be->env; |
| 113 | struct bench_timer *bt; |
| 114 | |
| 115 | /* Basic initialization. */ |
| 116 | bc->be = be; bc->bst = 0; bc->subctx = 0; |
| 117 | |
| 118 | /* Set up the benchmarking state if it hasn't been done before. */ |
| 119 | if (!be->bst || !*be->bst) { |
| 120 | bt = bench_createtimer(); if (!bt) goto fail_timer; |
| 121 | bc->bst = xmalloc(sizeof(*bc->bst)); bench_init(bc->bst, bt); |
| 122 | if (be->bst) *be->bst = bc->bst; |
| 123 | } else if (!(*be->bst)->tm) |
| 124 | goto fail_timer; |
| 125 | else |
| 126 | bc->bst = *be->bst; |
| 127 | |
| 128 | /* Set the default target time. */ |
| 129 | bc->dflt_target = bc->bst->target_s; |
| 130 | |
| 131 | /* Initialize the subordinate environment. */ |
| 132 | if (subenv && subenv->ctxsz) bc->subctx = xmalloc(subenv->ctxsz); |
| 133 | if (subenv && subenv->setup) subenv->setup(tv, subenv, bc, bc->subctx); |
| 134 | |
| 135 | /* All done. */ |
| 136 | end: |
| 137 | return; |
| 138 | fail_timer: |
| 139 | tvec_skipgroup(tv, "failed to create timer"); goto end; |
| 140 | } |
| 141 | |
| 142 | /* --- @tvec_benchset@ --- * |
| 143 | * |
| 144 | * Arguments: @struct tvec_state *tv@ = test vector state |
| 145 | * @const char *var@ = variable name to set |
| 146 | * @const struct tvec_env *env@ = environment description |
| 147 | * @void *ctx@ = context pointer |
| 148 | * |
| 149 | * Returns: Zero on success, @-1@ on failure. |
| 150 | * |
| 151 | * Use: Set a special variable. The following special variables are |
| 152 | * supported. |
| 153 | * |
| 154 | * * %|@target|% is the (approximate) number of seconds to run |
| 155 | * the benchmark. |
| 156 | * |
| 157 | * Unrecognized variables are passed to the subordinate |
| 158 | * environment, if there is one. |
| 159 | */ |
| 160 | |
| 161 | int tvec_benchset(struct tvec_state *tv, const char *var, |
| 162 | const struct tvec_env *env, void *ctx) |
| 163 | { |
| 164 | struct tvec_benchctx *bc = ctx; |
| 165 | const struct tvec_benchenv *be = (const struct tvec_benchenv *)env; |
| 166 | const struct tvec_env *subenv = be->env; |
| 167 | union tvec_regval rv; |
| 168 | static const struct tvec_floatinfo fi = { TVFF_NOMAX, 0.0, 0.0, 0.0 }; |
| 169 | static const struct tvec_regdef rd = |
| 170 | { "@target", -1, &tvty_float, 0, { &fi } }; |
| 171 | |
| 172 | if (STRCMP(var, ==, "@target")) { |
| 173 | if (tvty_float.parse(&rv, &rd, tv)) return (-1); |
| 174 | if (bc) bc->bst->target_s = rv.f; |
| 175 | return (1); |
| 176 | } else if (subenv && subenv->set) |
| 177 | return (subenv->set(tv, var, subenv, bc->subctx)); |
| 178 | else |
| 179 | return (0); |
| 180 | } |
| 181 | |
| 182 | /* --- @tvec_benchbefore@ --- * |
| 183 | * |
| 184 | * Arguments: @struct tvec_state *tv@ = test vector state |
| 185 | * @void *ctx@ = context pointer |
| 186 | * |
| 187 | * Returns: --- |
| 188 | * |
| 189 | * Use: Invoke the subordinate environment's @before@ function to |
| 190 | * prepare for the benchmark. |
| 191 | */ |
| 192 | |
| 193 | void tvec_benchbefore(struct tvec_state *tv, void *ctx) |
| 194 | { |
| 195 | struct tvec_benchctx *bc = ctx; |
| 196 | const struct tvec_benchenv *be = bc->be; |
| 197 | const struct tvec_env *subenv = be->env; |
| 198 | |
| 199 | /* Just call the subsidiary environment. */ |
| 200 | if (subenv && subenv->before) subenv->before(tv, bc->subctx); |
| 201 | } |
| 202 | |
| 203 | /* --- @tvec_benchafter@ --- * |
| 204 | * |
| 205 | * Arguments: @struct tvec_state *tv@ = test vector state |
| 206 | * @void *ctx@ = context pointer |
| 207 | * |
| 208 | * Returns: --- |
| 209 | * |
| 210 | * Use: Invoke the subordinate environment's @after@ function to |
| 211 | * clean up after the benchmark. |
| 212 | */ |
| 213 | |
| 214 | void tvec_benchafter(struct tvec_state *tv, void *ctx) |
| 215 | { |
| 216 | struct tvec_benchctx *bc = ctx; |
| 217 | const struct tvec_benchenv *be = bc->be; |
| 218 | const struct tvec_env *subenv = be->env; |
| 219 | |
| 220 | /* Restore the benchmark state's old target. */ |
| 221 | bc->bst->target_s = bc->dflt_target; |
| 222 | |
| 223 | /* Pass the call on to the subsidiary environment. */ |
| 224 | if (subenv && subenv->after) subenv->after(tv, bc->subctx); |
| 225 | } |
| 226 | |
| 227 | /* --- @tvec_benchteardown@ --- * |
| 228 | * |
| 229 | * Arguments: @struct tvec_state *tv@ = test vector state |
| 230 | * @void *ctx@ = context pointer |
| 231 | * |
| 232 | * Returns: --- |
| 233 | * |
| 234 | * Use: Tear down the benchmark environment. |
| 235 | */ |
| 236 | |
| 237 | void tvec_benchteardown(struct tvec_state *tv, void *ctx) |
| 238 | { |
| 239 | struct tvec_benchctx *bc = ctx; |
| 240 | const struct tvec_benchenv *be; |
| 241 | const struct tvec_env *subenv; |
| 242 | |
| 243 | if (!bc) return; |
| 244 | be = bc->be; subenv = be->env; |
| 245 | |
| 246 | /* Tear down any subsidiary environment. */ |
| 247 | if (subenv && subenv->teardown && bc->subctx) |
| 248 | subenv->teardown(tv, bc->subctx); |
| 249 | |
| 250 | /* If the benchmark state was temporary, then dispose of it. */ |
| 251 | if (bc->bst) { |
| 252 | if (be->bst) bc->bst->target_s = bc->dflt_target; |
| 253 | else { bench_destroy(bc->bst); xfree(bc->bst); } |
| 254 | } |
| 255 | } |
| 256 | |
| 257 | /*----- Measurement machinery ---------------------------------------------*/ |
| 258 | |
| 259 | /* --- @benchloop_...@ --- * |
| 260 | * |
| 261 | * Arguments: @unsigned long n@ = iteration count |
| 262 | * @void *ctx@ = benchmark running context |
| 263 | * |
| 264 | * Returns: --- |
| 265 | * |
| 266 | * Use: Run various kinds of benchmarking loops. |
| 267 | * |
| 268 | * * The @..._outer_...@ functions call the underlying |
| 269 | * function @n@ times in a loop; by contrast, the |
| 270 | * @..._inner_...@ functions set a register value to the |
| 271 | * chosen iteration count and expect the underlying function |
| 272 | * to perform the loop itself. |
| 273 | * |
| 274 | * * The @..._direct@ functions just call the underlying test |
| 275 | * function directly (though still through an `indirect |
| 276 | * jump' instruction); by contrast, the @..._indirect@ |
| 277 | * functions invoke a subsidiary environment's @run@ |
| 278 | * function, which adds additional overhead. |
| 279 | */ |
| 280 | |
| 281 | static void benchloop_outer_direct(unsigned long n, void *ctx) |
| 282 | { |
| 283 | struct benchrun *r = ctx; |
| 284 | tvec_testfn *fn = r->fn; void *tctx = r->ctx; |
| 285 | const struct tvec_reg *in = r->in; struct tvec_reg *out = r->out; |
| 286 | |
| 287 | while (n--) fn(in, out, tctx); |
| 288 | } |
| 289 | |
| 290 | static void benchloop_inner_direct(unsigned long n, void *ctx) |
| 291 | { struct benchrun *r = ctx; *r->n = n; r->fn(r->in, r->out, r->ctx); } |
| 292 | |
| 293 | static void benchloop_outer_indirect(unsigned long n, void *ctx) |
| 294 | { |
| 295 | struct benchrun *r = ctx; |
| 296 | struct tvec_state *tv = r->tv; |
| 297 | void (*run)(struct tvec_state *, tvec_testfn, void *) = r->env->run; |
| 298 | tvec_testfn *fn = r->fn; void *tctx = r->ctx; |
| 299 | |
| 300 | while (n--) run(tv, fn, tctx); |
| 301 | } |
| 302 | |
| 303 | static void benchloop_inner_indirect(unsigned long n, void *ctx) |
| 304 | { struct benchrun *r = ctx; *r->n = n; r->env->run(r->tv, r->fn, r->ctx); } |
| 305 | |
| 306 | /* --- @tvec_benchrun@ --- * |
| 307 | * |
| 308 | * Arguments: @struct tvec_state *tv@ = test vector state |
| 309 | * @tvec_testfn *fn@ = test function to run |
| 310 | * @void *ctx@ = context pointer for the test function |
| 311 | * |
| 312 | * Returns: --- |
| 313 | * |
| 314 | * Use: Measures and reports the performance of a test function. |
| 315 | */ |
| 316 | |
| 317 | void tvec_benchrun(struct tvec_state *tv, tvec_testfn *fn, void *ctx) |
| 318 | { |
| 319 | struct tvec_benchctx *bc = ctx; |
| 320 | const struct tvec_benchenv *be = bc->be; |
| 321 | const struct tvec_env *subenv = be->env; |
| 322 | const struct tvec_regdef *rd; |
| 323 | struct tvec_output *o = tv->output; |
| 324 | struct bench_timing tm; |
| 325 | struct benchrun r; |
| 326 | bench_fn *loopfn; |
| 327 | unsigned unit; |
| 328 | dstr d = DSTR_INIT; |
| 329 | double base; |
| 330 | unsigned f = 0; |
| 331 | #define f_any 1u |
| 332 | |
| 333 | /* Fill in the easy parts of the run context. */ |
| 334 | r.tv = tv; r.env = subenv; r.ctx = bc->subctx; |
| 335 | r.in = tv->in; r.out = tv->out; r.fn = fn; |
| 336 | |
| 337 | /* Decide on the run function to select. */ |
| 338 | if (be->riter >= 0) { |
| 339 | r.n = &TVEC_REG(tv, in, be->riter)->v.u; |
| 340 | loopfn = subenv && subenv->run ? |
| 341 | benchloop_inner_indirect : benchloop_inner_direct; |
| 342 | } else { |
| 343 | r.n = 0; |
| 344 | loopfn = subenv && subenv->run ? |
| 345 | benchloop_outer_indirect : benchloop_outer_direct; |
| 346 | } |
| 347 | |
| 348 | /* Decide on the kind of unit and the base count. */ |
| 349 | base = be->niter; |
| 350 | if (be->rbuf < 0) unit = TVBU_OP; |
| 351 | else { unit = TVBU_BYTE; base *= TVEC_REG(tv, in, be->rbuf)->v.bytes.sz; } |
| 352 | |
| 353 | /* Construct a description of the test using the identifier registers. */ |
| 354 | for (rd = tv->test->regs; rd->name; rd++) |
| 355 | if (rd->f&TVRF_ID) { |
| 356 | if (f&f_any) dstr_puts(&d, ", "); |
| 357 | else f |= f_any; |
| 358 | dstr_putf(&d, "%s = ", rd->name); |
| 359 | rd->ty->dump(&TVEC_REG(tv, in, rd->i)->v, rd, |
| 360 | TVSF_COMPACT, &dstr_printops, &d); |
| 361 | } |
| 362 | |
| 363 | /* Run the benchmark. */ |
| 364 | o->ops->bbench(o, d.buf, unit); |
| 365 | if (bench_measure(&tm, bc->bst, base, loopfn, &r)) |
| 366 | o->ops->ebench(o, d.buf, unit, 0); |
| 367 | else |
| 368 | o->ops->ebench(o, d.buf, unit, &tm); |
| 369 | |
| 370 | dstr_destroy(&d); |
| 371 | |
| 372 | #undef f_any |
| 373 | } |
| 374 | |
| 375 | /*----- Output utilities --------------------------------------------------*/ |
| 376 | |
| 377 | /* --- @tvec_benchreport@ --- * |
| 378 | * |
| 379 | * Arguments: @const struct gprintf_ops *gops@ = print operations |
| 380 | * @void *go@ = print destination |
| 381 | * @unsigned unit@ = the unit being measured (~TVBU_...@) |
| 382 | * @const struct bench_timing *tm@ = the benchmark result |
| 383 | * |
| 384 | * Returns: --- |
| 385 | * |
| 386 | * Use: Formats a report about the benchmark performance. This |
| 387 | * function is intended to be called on by an output |
| 388 | * @ebench@ function. |
| 389 | */ |
| 390 | |
| 391 | void tvec_benchreport(const struct gprintf_ops *gops, void *go, |
| 392 | unsigned unit, const struct bench_timing *tm) |
| 393 | { |
| 394 | double scale, x, n = tm->n; |
| 395 | const char *u, *what, *whats; |
| 396 | |
| 397 | if (!tm) { gprintf(gops, go, "benchmark FAILED"); return; } |
| 398 | |
| 399 | assert(tm->f&BTF_TIMEOK); |
| 400 | |
| 401 | switch (unit) { |
| 402 | case TVBU_OP: |
| 403 | gprintf(gops, go, "%.0f iterations ", n); |
| 404 | what = "op"; whats = "ops"; scale = 1000; |
| 405 | break; |
| 406 | case TVBU_BYTE: |
| 407 | x = n; normalize(&x, &u, 1024); gprintf(gops, go, "%.3f %sB ", x, u); |
| 408 | what = whats = "B"; scale = 1024; |
| 409 | break; |
| 410 | default: |
| 411 | abort(); |
| 412 | } |
| 413 | |
| 414 | x = tm->t; normalize(&x, &u, 1000); |
| 415 | gprintf(gops, go, "in %.3f %ss", x, u); |
| 416 | if (tm->f&BTF_CYOK) { |
| 417 | x = tm->cy; normalize(&x, &u, 1000); |
| 418 | gprintf(gops, go, " (%.3f %scy)", x, u); |
| 419 | } |
| 420 | gprintf(gops, go, ": "); |
| 421 | |
| 422 | x = n/tm->t; normalize(&x, &u, scale); |
| 423 | gprintf(gops, go, "%.3f %s%s/s", x, u, whats); |
| 424 | x = tm->t/n; normalize(&x, &u, 1000); |
| 425 | gprintf(gops, go, ", %.3f %ss/%s", x, u, what); |
| 426 | if (tm->f&BTF_CYOK) { |
| 427 | x = tm->cy/n; normalize(&x, &u, 1000); |
| 428 | gprintf(gops, go, " (%.3f %scy/%s)", x, u, what); |
| 429 | } |
| 430 | } |
| 431 | |
| 432 | /*----- That's all, folks -------------------------------------------------*/ |