benchtests/Makefile

   1 # Copyright (C) 2013-2021 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   ilogb \
  52   ilogbf \
  53   isfinite \
  54   isinf \
  55   isnan \
  56   j0 \
  57   j1 \
  58   lgamma \
  59   log \
  60   log10 \
  61   log1p \
  62   log2 \
  63   log2f \
  64   logb \
  65   logbf \
  66   logf \
  67   modf \
  68   pow \
  69   powf \
  70   rint \
  71   roundeven \
  72   roundevenf \
  73   sin \
  74   sincos \
  75   sincosf \
  76   sinf \
  77   sinh \
  78   sqrt \
  79   tan \
  80   tanh \
  81   tgamma \
  82   trunc \
  83   truncf \
  84   y0 \
  85   y1 \
  86
  87 ifneq (,$(filter yes,$(float96-fcts)))
  88 bench-math += cbrtl
  89 endif
  90
  91 ifneq (,$(filter yes,$(float128-fcts) $(float128-alias-fcts)))
  92 bench-math += expf128 powf128 sinf128 ilogbf128
  93 endif
  94
  95 bench-pthread := pthread_once thread_create pthread-locks
  96
  97 bench-string := ffs ffsll
  98
  99 ifeq (${BENCHSET},)
 100 bench := $(bench-math) $(bench-pthread) $(bench-string)
 101 else
 102 bench := $(foreach B,$(filter bench-%,${BENCHSET}), ${${B}})
 103 endif
 104
 105 # String function benchmarks.
 106 string-benchset := memccpy memchr memcmp memcpy memmem memmove \
 107                    mempcpy memset rawmemchr stpcpy stpncpy strcasecmp strcasestr \
 108                    strcat strchr strchrnul strcmp strcpy strcspn strlen \
 109                    strncasecmp strncat strncmp strncpy strnlen strpbrk strrchr \
 110                    strspn strstr strcpy_chk stpcpy_chk memrchr strsep strtok \
 111                    strcoll memcpy-large memcpy-random memmove-large memset-large \
 112                    memcpy-walk memset-walk memmove-walk
 113
 114 # Build and run locale-dependent benchmarks only if we're building natively.
 115 ifeq (no,$(cross-compiling))
 116 wcsmbs-benchset := wcslen wcsnlen wcscpy wcpcpy wcsncpy wcpncpy wcscat wcsncat \
 117                    wcscmp wcsncmp wcschr wcschrnul wcsrchr wcsspn wcspbrk wcscspn \
 118                    wmemchr wmemset wmemcmp
 119 else
 120 wcsmbs-benchset :=
 121 endif
 122
 123 string-benchset-all := $(string-benchset) ${wcsmbs-benchset}
 124
 125 ifeq (no,$(cross-compiling))
 126 # We have to generate locales
 127 LOCALES := en_US.UTF-8 tr_TR.UTF-8 cs_CZ.UTF-8 fa_IR.UTF-8 fr_FR.UTF-8 \
 128            ja_JP.UTF-8 si_LK.UTF-8 en_GB.UTF-8 vi_VN.UTF-8 ar_SA.UTF-8 \
 129            da_DK.UTF-8 pl_PL.UTF-8 pt_PT.UTF-8 el_GR.UTF-8 ru_RU.UTF-8 \
 130            he_IL.UTF-8 is_IS.UTF-8 es_ES.UTF-8 hi_IN.UTF-8 sv_SE.UTF-8 \
 131            hu_HU.UTF-8 it_IT.UTF-8 sr_RS.UTF-8 zh_CN.UTF-8
 132 include ../gen-locales.mk
 133 endif
 134
 135 stdlib-benchset := strtod
 136
 137 stdio-common-benchset := sprintf
 138
 139 math-benchset := math-inlines
 140
 141 ifeq (${BENCHSET},)
 142 benchset := $(string-benchset-all) $(stdlib-benchset) $(stdio-common-benchset) \
 143             $(math-benchset)
 144 else
 145 benchset := $(foreach B,$(filter %-benchset,${BENCHSET}), ${${B}})
 146 endif
 147
 148 CFLAGS-bench-ffs.c += -fno-builtin
 149 CFLAGS-bench-ffsll.c += -fno-builtin
 150 CFLAGS-bench-sqrt.c += -fno-builtin
 151 CFLAGS-bench-fmin.c += -fno-builtin
 152 CFLAGS-bench-fminf.c += -fno-builtin
 153 CFLAGS-bench-fmax.c += -fno-builtin
 154 CFLAGS-bench-fmaxf.c += -fno-builtin
 155 CFLAGS-bench-trunc.c += -fno-builtin
 156 CFLAGS-bench-truncf.c += -fno-builtin
 157 CFLAGS-bench-roundeven.c += -fno-builtin
 158 CFLAGS-bench-roundevenf.c += -fno-builtin
 159 CFLAGS-bench-isnan.c += -fsignaling-nans
 160 CFLAGS-bench-isinf.c += -fsignaling-nans
 161 CFLAGS-bench-isfinite.c += -fsignaling-nans
 162
 163 ifeq (${BENCHSET},)
 164 bench-malloc := malloc-thread malloc-simple
 165 else
 166 bench-malloc := $(filter malloc-%,${BENCHSET})
 167 endif
 168
 169 ifeq (${STATIC-BENCHTESTS},yes)
 170 +link-benchtests = $(+link-static-tests)
 171 link-libc-benchtests = $(link-libc-static)
 172 libm-benchtests = $(common-objpfx)math/libm.a
 173 thread-library-benchtests = $(static-thread-library)
 174 else
 175 link-libc-benchtests = $(link-libc)
 176 +link-benchtests = $(+link-tests)
 177 thread-library-benchtests = $(shared-thread-library)
 178 libm-benchtests = $(libm)
 179 endif
 180
 181 $(addprefix $(objpfx)bench-,$(bench-math)): $(libm-benchtests)
 182 $(addprefix $(objpfx)bench-,$(math-benchset)): $(libm-benchtests)
 183 $(addprefix $(objpfx)bench-,$(bench-pthread)): $(thread-library-benchtests)
 184 $(addprefix $(objpfx)bench-,$(bench-malloc)): $(thread-library-benchtests)
 185 $(addprefix $(objpfx)bench-,pthread-locks): $(libm-benchtests)
 186
 187 \f
 188
 189 # Rules to build and execute the benchmarks.  Do not put any benchmark
 190 # parameters beyond this point.
 191
 192 # We don't want the benchmark programs to run in parallel since that could
 193 # affect their performance.
 194 .NOTPARALLEL:
 195
 196 bench-extra-objs = json-lib.o
 197
 198 extra-objs += $(bench-extra-objs)
 199 others-extras = $(bench-extra-objs)
 200
 201 include ../Rules
 202
 203 binaries-bench := $(addprefix $(objpfx)bench-,$(bench))
 204 binaries-benchset := $(addprefix $(objpfx)bench-,$(benchset))
 205 binaries-bench-malloc := $(addprefix $(objpfx)bench-,$(bench-malloc))
 206
 207 # The default duration: 1 seconds.
 208 ifndef BENCH_DURATION
 209 BENCH_DURATION := 1
 210 endif
 211
 212 CPPFLAGS-nonlib += -DDURATION=$(BENCH_DURATION) -D_ISOMAC
 213
 214 # Use clock_gettime to measure performance of functions.  The default is
 215 # to use the architecture-specific high precision timing instructions.
 216 ifdef USE_CLOCK_GETTIME
 217 CPPFLAGS-nonlib += -DUSE_CLOCK_GETTIME
 218 else
 219 # On x86 processors, use RDTSCP, instead of RDTSC, to measure performance
 220 # of functions.  All x86 processors since 2010 support RDTSCP instruction.
 221 ifdef USE_RDTSCP
 222 CPPFLAGS-nonlib += -DUSE_RDTSCP
 223 endif
 224 endif
 225
 226 DETAILED_OPT :=
 227
 228 ifdef DETAILED
 229 DETAILED_OPT := -d
 230 endif
 231
 232 bench-deps := bench-skeleton.c bench-timing.h Makefile
 233
 234 run-bench = $(test-wrapper-env) \
 235             $(run-program-env) \
 236             $($*-ENV) $(test-via-rtld-prefix) $${run}
 237
 238 timing-type := $(objpfx)bench-timing-type
 239
 240 # This makes sure CPPFLAGS-nonlib and CFLAGS-nonlib are passed
 241 # for all these modules.
 242 cpp-srcs-left := $(binaries-benchset:=.c) $(binaries-bench:=.c) \
 243                  $(binaries-bench-malloc:=.c) $(timing-type:=.c)
 244 lib := nonlib
 245 include $(patsubst %,$(..)libof-iterator.mk,$(cpp-srcs-left))
 246
 247 bench-clean:
 248         rm -f $(binaries-bench) $(addsuffix .o,$(binaries-bench))
 249         rm -f $(binaries-benchset) $(addsuffix .o,$(binaries-benchset))
 250         rm -f $(binaries-bench-malloc) $(addsuffix .o,$(binaries-bench-malloc))
 251         rm -f $(timing-type) $(addsuffix .o,$(timing-type))
 252         rm -f $(addprefix $(objpfx),$(bench-extra-objs))
 253
 254 # Validate the passed in BENCHSET
 255 ifneq ($(strip ${BENCHSET}),)
 256 VALIDBENCHSETNAMES := bench-pthread bench-math bench-string string-benchset \
 257    wcsmbs-benchset stdlib-benchset stdio-common-benchset math-benchset \
 258    malloc-thread malloc-simple
 259 INVALIDBENCHSETNAMES := $(filter-out ${VALIDBENCHSETNAMES},${BENCHSET})
 260 ifneq (${INVALIDBENCHSETNAMES},)
 261 $(info The following values in BENCHSET are invalid: ${INVALIDBENCHSETNAMES})
 262 $(info The valid ones are: ${VALIDBENCHSETNAMES})
 263 $(error Invalid BENCHSET value)
 264 endif
 265 endif
 266
 267 bench: bench-build bench-set bench-func bench-malloc
 268
 269 # Target to only build the benchmark without running it.  We generate locales
 270 # only if we're building natively.
 271 ifeq (no,$(cross-compiling))
 272 bench-build: $(gen-locales) $(timing-type) $(binaries-bench) \
 273         $(binaries-benchset) $(binaries-bench-malloc)
 274 else
 275 bench-build: $(timing-type) $(binaries-bench) $(binaries-benchset) \
 276         $(binaries-bench-malloc)
 277 endif
 278
 279 bench-set: $(binaries-benchset)
 280         for run in $^; do \
 281           echo "Running $${run}"; \
 282           $(run-bench) > $${run}.out; \
 283         done
 284
 285 bench-malloc: $(binaries-bench-malloc)
 286         for run in $^; do \
 287           echo "$${run}"; \
 288           if [ `basename $${run}` = "bench-malloc-thread" ]; then \
 289                 for thr in 1 8 16 32; do \
 290                         echo "Running $${run} $${thr}"; \
 291                         $(run-bench) $${thr} > $${run}-$${thr}.out; \
 292                 done;\
 293           else \
 294                 for thr in 8 16 32 64 128 256 512 1024 2048 4096; do \
 295                   echo "Running $${run} $${thr}"; \
 296                   $(run-bench) $${thr} > $${run}-$${thr}.out; \
 297                 done;\
 298           fi;\
 299         done
 300
 301 # Build and execute the benchmark functions.  This target generates JSON
 302 # formatted bench.out.  Each of the programs produce independent JSON output,
 303 # so one could even execute them individually and process it using any JSON
 304 # capable language or tool.
 305 bench-func: $(binaries-bench)
 306         if [ -n '$^' ] ; then \
 307         { timing_type=$$($(test-wrapper-env) \
 308                          $(run-program-env) \
 309                          $(test-via-rtld-prefix) \
 310                          $(timing-type)); \
 311           echo "{\"timing_type\": \"$${timing_type}\","; \
 312           echo " \"functions\": {"; \
 313           for run in $^; do \
 314             if ! [ "x$${run}" = "x$<" ]; then \
 315               echo ","; \
 316             fi; \
 317             echo "Running $${run}" >&2; \
 318             $(run-bench) $(DETAILED_OPT); \
 319           done; \
 320           echo; \
 321           echo " }"; \
 322           echo "}"; \
 323           } > $(objpfx)bench.out-tmp; \
 324           if [ -f $(objpfx)bench.out ]; then \
 325             mv -f $(objpfx)bench.out $(objpfx)bench.out.old; \
 326           fi; \
 327           mv -f $(objpfx)bench.out-tmp $(objpfx)bench.out; \
 328           $(PYTHON) scripts/validate_benchout.py $(objpfx)bench.out \
 329           scripts/benchout.schema.json; \
 330         fi
 331
 332 ifeq ($(bind-now),yes)
 333 link-bench-bind-now = -Wl,-z,now
 334 endif
 335
 336 bench-link-targets = $(timing-type) $(binaries-bench) $(binaries-benchset) \
 337         $(binaries-bench-malloc)
 338
 339 $(bench-link-targets): %: %.o $(objpfx)json-lib.o \
 340         $(link-extra-libs-tests) \
 341   $(sort $(filter $(common-objpfx)lib%,$(link-libc-benchtests))) \
 342   $(addprefix $(csu-objpfx),start.o) $(+preinit) $(+postinit)
 343         $(+link-benchtests)
 344
 345 $(bench-link-targets): LDFLAGS += $(link-bench-bind-now)
 346
 347 $(objpfx)bench-%.c: %-inputs $(bench-deps)
 348         { if [ -n "$($*-INCLUDE)" ]; then \
 349           cat $($*-INCLUDE); \
 350         fi; \
 351         $(PYTHON) scripts/bench.py $(patsubst %-inputs,%,$<); } > $@-tmp
 352         mv -f $@-tmp $@