This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH] Use HP_TIMING for benchmarks if available
- From: Siddhesh Poyarekar <siddhesh at redhat dot com>
- To: libc-alpha at sourceware dot org
- Date: Mon, 22 Apr 2013 10:31:24 +0530
- Subject: Re: [PATCH] Use HP_TIMING for benchmarks if available
- References: <20130415081936 dot GN9444 at spoyarek dot pnq dot redhat dot com>
Ping!
On Mon, Apr 15, 2013 at 01:49:36PM +0530, Siddhesh Poyarekar wrote:
> Hi,
>
> Here's a patch that adds support for using HP_TIMING in benchmark
> measurements when it is available. clock_gettime is still kept as a
> fallback when !HP_TIMING_AVAIL. Additionally, I've also added support
> to override HP_TIMING to use clock_gettime by executing:
>
> make USE_CLOCK_GETTIME=1 bench
>
> One would need a 'make bench-clean' to ensure that the sources are
> rebuilt whenever one needs to switch between clock_gettime and
> HP_TIMING. Another easy way is to just touch benchtests/Makefile.
>
> I have verified that the measurements on x6_64 are consistent over
> multiple runs, so my concern of scheduler overhead causing jitters was
> unfounded. Jitter due to high system load is unavoidable even with
> clock_gettime since (as Rich Felker pointed out in an earlier
> discussion) competition for cache will still affect the performance
> numbers.
>
> Siddhesh
>
> * Makeconfig (cflags): Add bench-cflags.
> * benchtests/Makefile: Define bench-cflags if
> USE_CLOCK_GETTIME is defined.
> * benchtests-bench-skeleton.c: Include bench-timing.h.
> (main): Use TIMING_* macros instead of clock_gettime.
> * benchtests/bench-timing.h: New file.
>
> diff --git a/Makeconfig b/Makeconfig
> index a3d3e70..a83485f 100644
> --- a/Makeconfig
> +++ b/Makeconfig
> @@ -755,6 +755,9 @@ ifeq "$(strip $(+cflags))" ""
> +cflags := $(default_cflags)
> endif # $(+cflags) == ""
>
> +# Add common benchmark CFLAGS
> ++cflags += $(bench-cflags)
> +
> +cflags += $(cflags-cpu) $(+gccwarn) $(+merge-constants) $(+math-flags)
> +gcc-nowarn := -w
>
> diff --git a/benchtests/Makefile b/benchtests/Makefile
> index 3e794d7..d330abb 100644
> --- a/benchtests/Makefile
> +++ b/benchtests/Makefile
> @@ -109,12 +109,16 @@ LDFLAGS-bench-slowatan = -lm
> # Rules to build and execute the benchmarks. Do not put any benchmark
> # parameters beyond this point.
>
> +ifdef USE_CLOCK_GETTIME
> +bench-cflags := -DUSE_CLOCK_GETTIME
> +endif
> +
> include ../Makeconfig
> include ../Rules
>
> binaries-bench := $(addprefix $(objpfx)bench-,$(bench))
>
> -bench-deps := bench-skeleton.c Makefile
> +bench-deps := bench-skeleton.c bench-timing.h Makefile
>
> run-bench = $(test-wrapper-env) \
> GCONV_PATH=$(common-objpfx)iconvdata LC_ALL=C \
> diff --git a/benchtests/bench-skeleton.c b/benchtests/bench-skeleton.c
> index 13f986d..99a316e 100644
> --- a/benchtests/bench-skeleton.c
> +++ b/benchtests/bench-skeleton.c
> @@ -17,59 +17,53 @@
> <http://www.gnu.org/licenses/>. */
>
> #include <string.h>
> -#include <stdint.h>
> #include <stdio.h>
> #include <time.h>
> #include <inttypes.h>
> +#include "bench-timing.h"
>
> int
> main (int argc, char **argv)
> {
> unsigned long i, j, k;
> - uint64_t total = 0, max = 0, min = 0x7fffffffffffffff;
> - struct timespec start, end;
> + timing_t total = 0, max = 0, min = 0x7fffffffffffffff;
> + timing_t start, end;
>
> memset (&start, 0, sizeof (start));
> memset (&end, 0, sizeof (end));
>
> - clock_getres (CLOCK_PROCESS_CPUTIME_ID, &start);
> + unsigned long iters;
> +
> + TIMING_INIT (iters);
>
> - /* Measure 1000 times the resolution of the clock. So for a 1ns resolution
> - clock, we measure 1000 iterations of the function call at a time.
> - Measurements close to the minimum clock resolution won't make much sense,
> - but it's better than having nothing at all. */
> - unsigned long iters = 1000 * start.tv_nsec;
> unsigned long total_iters = ITER / iters;
>
> for (i = 0; i < NUM_SAMPLES; i++)
> {
> for (j = 0; j < total_iters; j ++)
> {
> - clock_gettime (CLOCK_PROCESS_CPUTIME_ID, &start);
> + int64_t cur;
> +
> + TIMING_NOW (start);
> for (k = 0; k < iters; k++)
> BENCH_FUNC(i);
> - clock_gettime (CLOCK_PROCESS_CPUTIME_ID, &end);
> -
> - uint64_t cur = (end.tv_nsec - start.tv_nsec
> - + ((end.tv_sec - start.tv_sec)
> - * (uint64_t) 1000000000));
> + TIMING_NOW (end);
>
> + TIMING_DIFF (cur, start, end);
> if (cur > max)
> max = cur;
>
> if (cur < min)
> min = cur;
>
> - total += cur;
> + TIMING_ACCUM (total, cur);
> }
> }
>
> - double d_total_s = total * 1e-9;
> + double d_total_s = total;
> double d_iters = iters;
> double d_total_i = (double)ITER * NUM_SAMPLES;
> - printf (FUNCNAME ": ITERS:%g: TOTAL:%gs, MAX:%gns, MIN:%gns, %g iter/s\n",
> - d_total_i, d_total_s, max / d_iters, min / d_iters,
> - d_total_i / d_total_s);
> + TIMING_PRINT_STATS (d_total_s, d_iters, d_total_i, max, min);
>
> return 0;
> }
> diff --git a/benchtests/bench-timing.h b/benchtests/bench-timing.h
> new file mode 100644
> index 0000000..e67a88d
> --- /dev/null
> +++ b/benchtests/bench-timing.h
> @@ -0,0 +1,72 @@
> +/* Define timing macros.
> + Copyright (C) 2013 Free Software Foundation, Inc.
> + This file is part of the GNU C Library.
> +
> + The GNU C Library is free software; you can redistribute it and/or
> + modify it under the terms of the GNU Lesser General Public
> + License as published by the Free Software Foundation; either
> + version 2.1 of the License, or (at your option) any later version.
> +
> + The GNU C Library is distributed in the hope that it will be useful,
> + but WITHOUT ANY WARRANTY; without even the implied warranty of
> + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
> + Lesser General Public License for more details.
> +
> + You should have received a copy of the GNU Lesser General Public
> + License along with the GNU C Library; if not, see
> + <http://www.gnu.org/licenses/>. */
> +
> +#include <hp-timing.h>
> +#include <stdint.h>
> +
> +#if HP_TIMING_AVAIL && !defined USE_CLOCK_GETTIME
> +# define GL(x) _##x
> +# define GLRO(x) _##x
> +hp_timing_t _dl_hp_timing_overhead;
> +typedef hp_timing_t timing_t;
> +
> +# define TIMING_INIT(iters) \
> +({ \
> + HP_TIMING_DIFF_INIT(); \
> + (iters) = 1000; \
> +})
> +
> +# define TIMING_NOW(var) HP_TIMING_NOW (var)
> +# define TIMING_DIFF(diff, start, end) HP_TIMING_DIFF ((diff), (start), (end))
> +# define TIMING_ACCUM(sum, diff) HP_TIMING_ACCUM_NT ((sum), (diff))
> +
> +# define TIMING_PRINT_STATS(d_total_s, d_iters, d_total_i, max, min) \
> + printf (FUNCNAME ": ITERS:%g: TOTAL:%gC, MAX:%gC, MIN:%gC, %g calls/MC\n", \
> + (d_total_i), (d_total_s), (max) / (d_iters), (min) / (d_iters), \
> + 1e6 * (d_total_i) / (d_total_s));
> +
> +#else
> +typedef uint64_t timing_t;
> +
> +/* Measure 1000 times the resolution of the clock. So for a 1ns
> + resolution clock, we measure 1000 iterations of the function call at a
> + time. Measurements close to the minimum clock resolution won't make
> + much sense, but it's better than having nothing at all. */
> +# define TIMING_INIT(iters) \
> +({ \
> + struct timespec start; \
> + clock_getres (CLOCK_PROCESS_CPUTIME_ID, &start); \
> + (iters) = 1000 * start.tv_nsec; \
> +})
> +
> +# define TIMING_NOW(var) \
> +({ \
> + struct timespec tv; \
> + clock_gettime (CLOCK_PROCESS_CPUTIME_ID, &tv); \
> + (var) = (uint64_t) (tv.tv_nsec + (uint64_t) 1000000000 * tv.tv_sec); \
> +})
> +
> +# define TIMING_DIFF(diff, start, end) (diff) = (end) - (start)
> +# define TIMING_ACCUM(sum, diff) (sum) += (diff)
> +
> +# define TIMING_PRINT_STATS(d_total_s, d_iters, d_total_i, max, min) \
> + printf (FUNCNAME ": ITERS:%g: TOTAL:%gs, MAX:%gs, MIN:%gs, %g iter/s\n", \
> + (d_total_i), (d_total_s) * 1e9, (max) / (d_iters), \
> + (min) / (d_iters), (d_total_i) / ((d_total_s) * 1e9))
> +
> +#endif