benchtests/Makefile

   1 # Copyright (C) 2013-2022 Free Software Foundation, Inc.
   2 # This file is part of the GNU C Library.
   3
   4 # The GNU C Library is free software; you can redistribute it and/or
   5 # modify it under the terms of the GNU Lesser General Public
   6 # License as published by the Free Software Foundation; either
   7 # version 2.1 of the License, or (at your option) any later version.
   8
   9 # The GNU C Library is distributed in the hope that it will be useful,
  10 # but WITHOUT ANY WARRANTY; without even the implied warranty of
  11 # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  12 # Lesser General Public License for more details.
  13
  14 # You should have received a copy of the GNU Lesser General Public
  15 # License along with the GNU C Library; if not, see
  16 # <https://www.gnu.org/licenses/>.
  17
  18 # Makefile for benchmark tests.  The only useful target here is `bench`.
  19 # Add benchmark functions in alphabetical order.
  20
  21 subdir := benchtests
  22
  23 include ../Makeconfig
  24
  25 bench-math := \
  26   acos \
  27   acosh \
  28   asin \
  29   asinh \
  30   atan \
  31   atan2 \
  32   atanh \
  33   cbrt \
  34   cos \
  35   cosf \
  36   cosh \
  37   erf \
  38   erfc \
  39   exp \
  40   exp10 \
  41   exp10f \
  42   exp2 \
  43   exp2f \
  44   expf \
  45   expm1 \
  46   fmax \
  47   fmaxf \
  48   fmin \
  49   fminf \
  50   hypot \
  51   hypotf \
  52   ilogb \
  53   ilogbf \
  54   isfinite \
  55   isinf \
  56   isnan \
  57   j0 \
  58   j1 \
  59   lgamma \
  60   log \
  61   log10 \
  62   log1p \
  63   log2 \
  64   log2f \
  65   logb \
  66   logbf \
  67   logf \
  68   modf \
  69   pow \
  70   powf \
  71   rint \
  72   roundeven \
  73   roundevenf \
  74   sin \
  75   sincos \
  76   sincosf \
  77   sinf \
  78   sinh \
  79   sqrt \
  80   tan \
  81   tanh \
  82   tgamma \
  83   trunc \
  84   truncf \
  85   y0 \
  86   y1 \
  87
  88 ifneq (,$(filter yes,$(float96-fcts)))
  89 bench-math += \
  90   cbrtl \
  91 # bench-math
  92 endif
  93
  94 ifneq (,$(filter yes,$(float128-fcts) $(float128-alias-fcts)))
  95 bench-math += \
  96   expf128 \
  97   ilogbf128 \
  98   powf128 \
  99   sinf128 \
 100 # bench-math
 101 endif
 102
 103 bench-pthread := \
 104   pthread-locks \
 105   pthread_once \
 106   thread_create \
 107 # bench-pthread
 108
 109 bench-string := \
 110   ffs \
 111   ffsll \
 112 # bench-string
 113
 114 ifeq (${BENCHSET},)
 115 bench := $(bench-math) $(bench-pthread) $(bench-string)
 116 else
 117 bench := $(foreach B,$(filter bench-%,${BENCHSET}), ${${B}})
 118 endif
 119
 120 # String function benchmarks.
 121 string-benchset := \
 122   bzero \
 123   bzero-large \
 124   bzero-walk \
 125   memccpy \
 126   memchr \
 127   memcmp \
 128   memcmpeq \
 129   memcpy \
 130   memcpy-large \
 131   memcpy-random \
 132   memcpy-walk \
 133   memmem \
 134   memmove \
 135   memmove-large \
 136   memmove-walk \
 137   mempcpy \
 138   memrchr \
 139   memset \
 140   memset-large \
 141   memset-walk \
 142   memset-zero \
 143   memset-zero-large \
 144   memset-zero-walk \
 145   rawmemchr \
 146   stpcpy \
 147   stpcpy_chk \
 148   stpncpy \
 149   strcasecmp \
 150   strcasestr \
 151   strcat \
 152   strchr \
 153   strchrnul \
 154   strcmp \
 155   strcoll \
 156   strcpy \
 157   strcpy_chk \
 158   strcspn \
 159   strlen \
 160   strncasecmp \
 161   strncat \
 162   strncmp \
 163   strncpy \
 164   strnlen \
 165   strpbrk \
 166   strrchr \
 167   strsep \
 168   strspn \
 169   strstr \
 170   strtok \
 171 # string-benchset
 172
 173 # Build and run locale-dependent benchmarks only if we're building natively.
 174 ifeq (no,$(cross-compiling))
 175 wcsmbs-benchset := \
 176   wcpcpy \
 177   wcpncpy \
 178   wcscat \
 179   wcschr \
 180   wcschrnul \
 181   wcscmp \
 182   wcscpy \
 183   wcscspn \
 184   wcslen \
 185   wcsncat \
 186   wcsncmp \
 187   wcsncpy \
 188   wcsnlen \
 189   wcspbrk \
 190   wcsrchr \
 191   wcsspn \
 192   wmemchr \
 193   wmemcmp \
 194   wmemset \
 195 # wcsmbs-benchset
 196 else
 197 wcsmbs-benchset :=
 198 endif
 199
 200 string-benchset-all := $(string-benchset) ${wcsmbs-benchset}
 201
 202 ifeq (no,$(cross-compiling))
 203 # We have to generate locales
 204 LOCALES := \
 205   ar_SA.UTF-8 \
 206   cs_CZ.UTF-8 \
 207   da_DK.UTF-8 \
 208   el_GR.UTF-8 \
 209   en_GB.UTF-8 \
 210   en_US.UTF-8 \
 211   es_ES.UTF-8 \
 212   fa_IR.UTF-8 \
 213   fr_FR.UTF-8 \
 214   he_IL.UTF-8 \
 215   hi_IN.UTF-8 \
 216   hu_HU.UTF-8 \
 217   is_IS.UTF-8 \
 218   it_IT.UTF-8 \
 219   ja_JP.UTF-8 \
 220   pl_PL.UTF-8 \
 221   pt_PT.UTF-8 \
 222   ru_RU.UTF-8 \
 223   si_LK.UTF-8 \
 224   sr_RS.UTF-8 \
 225   sv_SE.UTF-8 \
 226   tr_TR.UTF-8 \
 227   vi_VN.UTF-8 \
 228   zh_CN.UTF-8 \
 229 # LOCALES
 230 include ../gen-locales.mk
 231 endif
 232
 233 stdlib-benchset := strtod
 234
 235 stdio-common-benchset := sprintf
 236
 237 math-benchset := math-inlines
 238
 239 ifeq (${BENCHSET},)
 240 benchset := $(string-benchset-all) $(stdlib-benchset) $(stdio-common-benchset) \
 241             $(math-benchset)
 242 else
 243 benchset := $(foreach B,$(filter %-benchset,${BENCHSET}), ${${B}})
 244 endif
 245
 246 CFLAGS-bench-ffs.c += -fno-builtin
 247 CFLAGS-bench-ffsll.c += -fno-builtin
 248 CFLAGS-bench-sqrt.c += -fno-builtin
 249 CFLAGS-bench-fmin.c += -fno-builtin
 250 CFLAGS-bench-fminf.c += -fno-builtin
 251 CFLAGS-bench-fmax.c += -fno-builtin
 252 CFLAGS-bench-fmaxf.c += -fno-builtin
 253 CFLAGS-bench-trunc.c += -fno-builtin
 254 CFLAGS-bench-truncf.c += -fno-builtin
 255 CFLAGS-bench-roundeven.c += -fno-builtin
 256 CFLAGS-bench-roundevenf.c += -fno-builtin
 257 CFLAGS-bench-isnan.c += -fsignaling-nans
 258 CFLAGS-bench-isinf.c += -fsignaling-nans
 259 CFLAGS-bench-isfinite.c += -fsignaling-nans
 260
 261 ifeq (${BENCHSET},)
 262 bench-malloc := malloc-thread malloc-simple
 263 else
 264 bench-malloc := $(filter malloc-%,${BENCHSET})
 265 endif
 266
 267 ifeq (${STATIC-BENCHTESTS},yes)
 268 +link-benchtests = $(+link-static-tests)
 269 link-libc-benchtests = $(link-libc-static)
 270 libm-benchtests = $(common-objpfx)math/libm.a
 271 thread-library-benchtests = $(static-thread-library)
 272 else
 273 link-libc-benchtests = $(link-libc)
 274 +link-benchtests = $(+link-tests)
 275 thread-library-benchtests = $(shared-thread-library)
 276 libm-benchtests = $(libm)
 277 endif
 278
 279 $(addprefix $(objpfx)bench-,$(bench-math)): $(libm-benchtests)
 280 $(addprefix $(objpfx)bench-,$(math-benchset)): $(libm-benchtests)
 281 $(addprefix $(objpfx)bench-,$(bench-pthread)): $(thread-library-benchtests)
 282 $(addprefix $(objpfx)bench-,$(bench-malloc)): $(thread-library-benchtests)
 283 $(addprefix $(objpfx)bench-,pthread-locks): $(libm-benchtests)
 284
 285 \f
 286
 287 # Rules to build and execute the benchmarks.  Do not put any benchmark
 288 # parameters beyond this point.
 289
 290 # We don't want the benchmark programs to run in parallel since that could
 291 # affect their performance.
 292 .NOTPARALLEL:
 293
 294 bench-extra-objs = json-lib.o
 295
 296 extra-objs += $(bench-extra-objs)
 297 others-extras = $(bench-extra-objs)
 298
 299 # NB: Use "=" instead of ":=" since sysdeps Makefiles may add more
 300 # benches.
 301 binaries-bench = $(addprefix $(objpfx)bench-,$(bench))
 302 extra-objs += $(addsuffix .o,$(addprefix bench-,$(bench)))
 303 binaries-benchset = $(addprefix $(objpfx)bench-,$(benchset))
 304 extra-objs += $(addsuffix .o,$(addprefix bench-,$(benchset)))
 305 binaries-bench-malloc := $(addprefix $(objpfx)bench-,$(bench-malloc))
 306 extra-objs += $(addsuffix .o,$(addprefix bench-,$(bench-malloc)))
 307
 308 # The default duration: 1 seconds.
 309 ifndef BENCH_DURATION
 310 BENCH_DURATION := 1
 311 endif
 312
 313 CPPFLAGS-nonlib += -DDURATION=$(BENCH_DURATION) -D_ISOMAC
 314
 315 # Use clock_gettime to measure performance of functions.  The default is
 316 # to use the architecture-specific high precision timing instructions.
 317 ifdef USE_CLOCK_GETTIME
 318 CPPFLAGS-nonlib += -DUSE_CLOCK_GETTIME
 319 else
 320 # On x86 processors, use RDTSCP, instead of RDTSC, to measure performance
 321 # of functions.  All x86 processors since 2010 support RDTSCP instruction.
 322 ifdef USE_RDTSCP
 323 CPPFLAGS-nonlib += -DUSE_RDTSCP
 324 endif
 325 endif
 326
 327 DETAILED_OPT :=
 328
 329 ifdef DETAILED
 330 DETAILED_OPT := -d
 331 endif
 332
 333 bench-deps := bench-skeleton.c bench-timing.h Makefile
 334
 335 run-bench = $(test-wrapper-env) \
 336             $(run-program-env) \
 337             $($*-ENV) $(test-via-rtld-prefix) $${run}
 338
 339 timing-type := $(objpfx)bench-timing-type
 340 extra-objs += bench-timing-type.o
 341
 342 include ../Rules
 343
 344 # This makes sure CPPFLAGS-nonlib and CFLAGS-nonlib are passed
 345 # for all these modules.
 346 cpp-srcs-left := $(binaries-benchset:=.c) $(binaries-bench:=.c) \
 347                  $(binaries-bench-malloc:=.c) $(timing-type:=.c)
 348 lib := nonlib
 349 include $(patsubst %,$(..)libof-iterator.mk,$(cpp-srcs-left))
 350
 351 bench-clean:
 352         rm -f $(binaries-bench) $(addsuffix .o,$(binaries-bench))
 353         rm -f $(binaries-benchset) $(addsuffix .o,$(binaries-benchset))
 354         rm -f $(binaries-bench-malloc) $(addsuffix .o,$(binaries-bench-malloc))
 355         rm -f $(timing-type) $(addsuffix .o,$(timing-type))
 356         rm -f $(addprefix $(objpfx),$(bench-extra-objs))
 357
 358 # Validate the passed in BENCHSET
 359 ifneq ($(strip ${BENCHSET}),)
 360 VALIDBENCHSETNAMES := bench-pthread bench-math bench-string string-benchset \
 361    wcsmbs-benchset stdlib-benchset stdio-common-benchset math-benchset \
 362    malloc-thread malloc-simple
 363 INVALIDBENCHSETNAMES := $(filter-out ${VALIDBENCHSETNAMES},${BENCHSET})
 364 ifneq (${INVALIDBENCHSETNAMES},)
 365 $(info The following values in BENCHSET are invalid: ${INVALIDBENCHSETNAMES})
 366 $(info The valid ones are: ${VALIDBENCHSETNAMES})
 367 $(error Invalid BENCHSET value)
 368 endif
 369 endif
 370
 371 bench: bench-build bench-set bench-func bench-malloc
 372
 373 # Target to only build the benchmark without running it.  We generate locales
 374 # only if we're building natively.
 375 ifeq (no,$(cross-compiling))
 376 bench-build: $(gen-locales) $(timing-type) $(binaries-bench) \
 377         $(binaries-benchset) $(binaries-bench-malloc)
 378 else
 379 bench-build: $(timing-type) $(binaries-bench) $(binaries-benchset) \
 380         $(binaries-bench-malloc)
 381 endif
 382
 383 bench-set: $(binaries-benchset)
 384         for run in $^; do \
 385           echo "Running $${run}"; \
 386           $(run-bench) > $${run}.out; \
 387         done
 388
 389 bench-malloc: $(binaries-bench-malloc)
 390         for run in $^; do \
 391           echo "$${run}"; \
 392           if [ `basename $${run}` = "bench-malloc-thread" ]; then \
 393                 for thr in 1 8 16 32; do \
 394                         echo "Running $${run} $${thr}"; \
 395                         $(run-bench) $${thr} > $${run}-$${thr}.out; \
 396                 done;\
 397           else \
 398                 for thr in 8 16 32 64 128 256 512 1024 2048 4096; do \
 399                   echo "Running $${run} $${thr}"; \
 400                   $(run-bench) $${thr} > $${run}-$${thr}.out; \
 401                 done;\
 402           fi;\
 403         done
 404
 405 # Build and execute the benchmark functions.  This target generates JSON
 406 # formatted bench.out.  Each of the programs produce independent JSON output,
 407 # so one could even execute them individually and process it using any JSON
 408 # capable language or tool.
 409 bench-func: $(binaries-bench)
 410         if [ -n '$^' ] ; then \
 411         { timing_type=$$($(test-wrapper-env) \
 412                          $(run-program-env) \
 413                          $(test-via-rtld-prefix) \
 414                          $(timing-type)); \
 415           echo "{\"timing_type\": \"$${timing_type}\","; \
 416           echo " \"functions\": {"; \
 417           for run in $^; do \
 418             if ! [ "x$${run}" = "x$<" ]; then \
 419               echo ","; \
 420             fi; \
 421             echo "Running $${run}" >&2; \
 422             $(run-bench) $(DETAILED_OPT); \
 423           done; \
 424           echo; \
 425           echo " }"; \
 426           echo "}"; \
 427           } > $(objpfx)bench.out-tmp; \
 428           if [ -f $(objpfx)bench.out ]; then \
 429             mv -f $(objpfx)bench.out $(objpfx)bench.out.old; \
 430           fi; \
 431           mv -f $(objpfx)bench.out-tmp $(objpfx)bench.out; \
 432           $(PYTHON) scripts/validate_benchout.py $(objpfx)bench.out \
 433           scripts/benchout.schema.json; \
 434         fi
 435
 436 ifeq ($(bind-now),yes)
 437 link-bench-bind-now = -Wl,-z,now
 438 endif
 439
 440 bench-link-targets = $(timing-type) $(binaries-bench) $(binaries-benchset) \
 441         $(binaries-bench-malloc)
 442
 443 $(bench-link-targets): %: %.o $(objpfx)json-lib.o \
 444         $(link-extra-libs-tests) \
 445   $(sort $(filter $(common-objpfx)lib%,$(link-libc-benchtests))) \
 446   $(addprefix $(csu-objpfx),start.o) $(+preinit) $(+postinit)
 447         $(+link-benchtests)
 448
 449 $(bench-link-targets): LDFLAGS += $(link-bench-bind-now)
 450
 451 $(objpfx)bench-%.c: %-inputs $(bench-deps)
 452         { if [ -n "$($*-INCLUDE)" ]; then \
 453           cat $($*-INCLUDE); \
 454         fi; \
 455         $(PYTHON) scripts/bench.py $(patsubst %-inputs,%,$<); } > $@-tmp
 456         mv -f $@-tmp $@