Increase benchmark iterations for math and vector math functions to improve
timing accuracy. Vector math benchmarks now take 1-3 seconds on a modern CPU.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Few tests needed to properly check for asprintf and system calls return
values with _FORTIFY_SOURCE enabled.
Reviewed-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
Reflow all long lines adding comment terminators.
Sort all reflowed text using scripts/sort-makefile-lines.py.
No regressions running microbenchmarks.
No code generation changes observed in binary artifacts.
No regressions on x86_64 and i686.
This patch enables libmvec on AArch64. The proposed change is mainly
implementing build infrastructure to add the new routines to ABI,
tests and benchmarks. I have demonstrated how this all fits together
by adding implementations for vector cos, in both single and double
precision, targeting both Advanced SIMD and SVE.
The implementations of the routines themselves are just loops over the
scalar routine from libm for now, as we are more concerned with
getting the plumbing right at this point. We plan to contribute vector
routines from the Arm Optimized Routines repo that are compliant with
requirements described in the libmvec wiki.
Building libmvec requires minimum GCC 10 for SVE ACLE. To avoid raising
the minimum GCC by such a big jump, we allow users to disable libmvec
if their compiler is too old.
Note that at this point users have to manually call the vector math
functions. This seems to be acceptable to some downstream users.
Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
Adjust iteration counts so benchmarks don't run too slowly or quickly.
Ensure benchmarks take less than 10 seconds on older, slower cores and
more than 0.5 seconds on fast cores.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
1. Subnormals: 128 inputs.
2. Normal numbers with large exponent difference (|x/y| > 2^8):
1024 inputs between FLT_MIN and FLT_MAX;
3. Close exponents (ey >= -103 and |x/y| < 2^8): 1024 inputs with
exponents between -10 and 10.
Reviewed-by: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
Add three different dataset, from random floating point numbers:
1. Subnormals: 128 inputs.
2. Normal numbers with large exponent difference (|x/y| > 2^52):
1024 inputs between DBL_MIN and DBL_MAX;
3. Close exponents (ey >= -907 and |x/y| < 2^52): 1024 inputs with
exponents between -10 and 10.
Reviewed-by: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
This allows other targets to use the same inputs for their own libmvec
microbenchmarks without having to duplicate them in their own
subdirectory.
Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
Remove simple_strcspn/strpbrk/strsep which are significantly slower than the
generic implementations. Also remove oldstrsep and oldstrtok since they are
practically identical to the generic implementation. Adjust iteration count
to reduce benchmark time.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Remove memchr_strnlen since it is now the same as generic_strnlen. Adjust
iteration count to reduce benchmark time. Keep memchr_strlen since the
generic strlen does not use memchr.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Almost all uses of rawmemchr find the end of a string. Since most targets use
a generic implementation, replacing it with strchr is better since that is
optimized by compilers into strlen (s) + s. Also fix the generic rawmemchr
implementation to use a cast to unsigned char in the if statement.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
It shows both throughput (total bytes obtained in the test duration)
and latecy for both arc4random and arc4random_buf with different
sizes.
Checked on x86_64-linux-gnu, aarch64-linux, and powerpc64le-linux-gnu.
Add a second iteration for memrchr to set `pos` starting from the end
of the buffer.
Previously `pos` was only set relative to the beginning of the
buffer. This isn't really useful for memrchr because the beginning
of the search space is (buf + len).
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
1. Use json_ctx for output to help standardize format across all
benchtests.
2. Add some additional tests to strstr and memchr expanding alignments
and adding more small values.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
1. Output results in json format so its easier to parse
2. Increase max alignment to `getpagesize () - 1` to make it possible
to test page cross cases.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
Unroll slightly and enforce good instruction scheduling. This improves
performance on out-of-order machines. The unrolling allows for
pipelined multiplies.
As well, as an optional sysdep, reorder the operations and prevent
reassosiation for better scheduling and higher ILP. This commit
only adds the barrier for x86, although it should be either no
change or a win for any architecture.
Unrolling further started to induce slowdowns for sizes [0, 4]
but can help the loop so if larger sizes are the target further
unrolling can be beneficial.
Results for _dl_new_hash
Benchmarked on Tigerlake: 11th Gen Intel(R) Core(TM) i7-1165G7 @ 2.80GHz
Time as Geometric Mean of N=30 runs
Geometric of all benchmark New / Old: 0.674
type, length, New Time, Old Time, New Time / Old Time
fixed, 0, 2.865, 2.72, 1.053
fixed, 1, 3.567, 2.489, 1.433
fixed, 2, 2.577, 3.649, 0.706
fixed, 3, 3.644, 5.983, 0.609
fixed, 4, 4.211, 6.833, 0.616
fixed, 5, 4.741, 9.372, 0.506
fixed, 6, 5.415, 9.561, 0.566
fixed, 7, 6.649, 10.789, 0.616
fixed, 8, 8.081, 11.808, 0.684
fixed, 9, 8.427, 12.935, 0.651
fixed, 10, 8.673, 14.134, 0.614
fixed, 11, 10.69, 15.408, 0.694
fixed, 12, 10.789, 16.982, 0.635
fixed, 13, 12.169, 18.411, 0.661
fixed, 14, 12.659, 19.914, 0.636
fixed, 15, 13.526, 21.541, 0.628
fixed, 16, 14.211, 23.088, 0.616
fixed, 32, 29.412, 52.722, 0.558
fixed, 64, 65.41, 142.351, 0.459
fixed, 128, 138.505, 295.625, 0.469
fixed, 256, 291.707, 601.983, 0.485
random, 2, 12.698, 12.849, 0.988
random, 4, 16.065, 15.857, 1.013
random, 8, 19.564, 21.105, 0.927
random, 16, 23.919, 26.823, 0.892
random, 32, 31.987, 39.591, 0.808
random, 64, 49.282, 71.487, 0.689
random, 128, 82.23, 145.364, 0.566
random, 256, 152.209, 298.434, 0.51
Co-authored-by: Alexander Monakov <amonakov@ispras.ru>
Reviewed-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
Add a simple benchmark that measures wcrtomb performance with various
locales with 1-4 byte characters.
Signed-off-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
Reviewed-by: Florian Weimer <fweimer@redhat.com>
Improve libmvec benchmark integration so that in future other
architectures may be able to run their libmvec benchmarks as well. This
now allows libmvec benchmarks to be run with `make BENCHSET=bench-math`.
Signed-off-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
The libmvec benchmarks print a message indicating that a certain CPU
feature is unsupported and exit prematurelyi, which breaks the JSON in
bench.out.
Handle this more elegantly in the bench makefile target by adding
support for an UNSUPPORTED exit status (77) so that bench.out continues
to have output for valid tests.
Signed-off-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
Benchmark for testing pthread mutex locks performance with different
threads and critical sections.
The test configuration consists of 3 parts:
1. thread number
2. critical-section length
3. non-critical-section length
Thread number starts from 1 and increased by 2x until num of CPU cores
(nprocs). An additional over-saturation case (1.25 * nprocs) is also
included.
Critical-section is represented by a loop of shared do_filler(),
length can be determined by the loop iters.
Non-critical-section is similiar to the critical-section, except it's
based on non-shared do_filler().
Currently, adaptive pthread_mutex lock is tested.
1. Use json-lib for printing results.
2. Expose all parameters (before pos, seek_char, and max_char where
not printed).
3. Add benchmarks that test multiple occurence of seek_char in the
string.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
Add benchmark that randomizes whether return should be NULL or pointer
to CHAR. The rationale is on many architectures there is a choice
between a predicate execution option (i.e cmovcc on x86) or a branch.
On x86 the results for cmovcc vs branch are something along the lines
of the following:
perc-zero, Br On Result, Time Br / Time cmov
0.10, 1, ,0.983
0.10, 0, ,1.246
0.25, 1, ,1.035
0.25, 0, ,1.49
0.33, 1, ,1.016
0.33, 0, ,1.579
0.50, 1, ,1.228
0.50, 0, ,1.739
0.66, 1, ,1.039
0.66, 0, ,1.764
0.75, 1, ,0.996
0.75, 0, ,1.642
0.90, 1, ,1.071
0.90, 0, ,1.409
1.00, 1, ,0.937
1.00, 0, ,0.999
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
Use "=" instead of ":=" to allow sysdeps Makefiles to add more benches
to bench and benchset. This fixes BZ #28970.
Reviewed-by: Sunil K Pandey <skpgkp2@gmail.com>
Commit ac759b1fbf added attribute
"overlap" to bench-memmove-walk, whose value is a string. This change
makes compare_strings.py fail since benchout_strings.schema.json
requires the values of attributes to be number.
This patch relaxes such constraint.
Reviewed-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
Small sizes (<= 64) represent large portion of memset usages with zero
value. Add sizes (<= 64) to bench-bzero-walk.c to cover small sizes.
Reviewed-by: Sunil K Pandey <skpgkp2@gmail.com>
memset with zero as the value to set is by far the majority value (99%+
for Python3 and GCC). Add bench-memset-zero-large.c,
bench-memset-zero-walk.c and bench-memset-zero.c to measure memset
implementations for zeroing.
Reviewed-by: Sunil K Pandey <skpgkp2@gmail.com>
Zero is a relevant size for some workloads (roughly 5% of uses for
GCC) so we should be testing it's performance as well.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
Add more small and medium sized tests for strcmp and strncmp.
As well for strcmp add option for more direct control of
alignment. Previously alignment was being pushed to the end of the
page. While this is the most difficult case to implement, it is far
from the common case and so shouldn't be the only benchmark.
Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
I used these shell commands:
../glibc/scripts/update-copyrights $PWD/../gnulib/build-aux/update-copyright
(cd ../glibc && git commit -am"[this commit message]")
and then ignored the output, which consisted lines saying "FOO: warning:
copyright statement not found" for each of 7061 files FOO.
I then removed trailing white space from math/tgmath.h,
support/tst-support-open-dev-null-range.c, and
sysdeps/x86_64/multiarch/strlen-vec.S, to work around the following
obscure pre-commit check failure diagnostics from Savannah. I don't
know why I run into these diagnostics whereas others evidently do not.
remote: *** 912-#endif
remote: *** 913:
remote: *** 914-
remote: *** error: lines with trailing whitespace found
...
remote: *** error: sysdeps/unix/sysv/linux/statx_cp.c: trailing lines
This commit adds a new partial overlap benchmark. This is generally
the most interesting performance case for memmove and was missing.
Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
This commit adds more benchmarks for the common memcpy/memmove
benchmarks. The most signifcant cases are the half page offsets. The
current versions leaves dst and src near page aligned which leads to
false 4k aliasing on x86_64. This can add noise due to false
dependencies from one run to the next. As well, this seems like more
of an edge case that common case so it shouldn't be the only thing
Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
Based on random input arguments. About 85% tuples have exponents
of the two arguments close together (+-1 range).
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
Instead of inputs based on the algorithm implementation details.
About 85% tuples have exponents of the two arguments close
together (+-1 range).
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
Add acosf function to bench-math and copy acosf-inputs to benchtests.
Motivation for this patch is to prepare for upcoming libmvec new
functions. Float and double version of libmvec functions stays
together.
acosf-inputs file generated from acos-inputs file using following
scaling formula:
f = d * (FLT_MAX/DBL_MAX)
Where d is input(double) and f is output(float). If scaled float value
is duplicate in new input file, nextafterf() function used to find next
float value, ensuring no duplicates.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
Improve the random memcpy benchmark. Double the number of tests and increase
the size of the memory region to test between 32KB and 1024KB. This improves
accuracy on modern cores. Clean up formatting of the frequency array.
Reviewed-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
No bug. This commit adds __memcmpeq benchmarks. The benchmarks just
use the existing ones in memcmp. This will be useful for testing
implementations of __memcmpeq that do not just alias memcmp.
No bug.
This commit adds new medium size cases for lengths in [512, 1024). As
well it increase the iters to INNER_LOOP_ITERS_LARGE for more reliable
results.
Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
Building benchmarks as static executables:
=========================================
To build benchmarks as static executables, on the build system, run:
$ make STATIC-BENCHTESTS=yes bench-build
You can copy benchmark executables to another machine and run them
without copying the source nor build directories.
No bug. Remove reallocation of bufs between implementation tests. Move
initialization outside of foreach implementation test loop. Increase
iteration count.
Generally before this commit was seeing a great deal of variability
between runs. The goal of this commit is to make the results more
reliable.
Benchtests build and bench-memcmp succeeding.
Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
This patch fixed validate_benchout.py two exceptions,
1) AttributeError
if benchout_strings.schema.json is specified, and
2) json.decoder.JSONDecodeError
if benchout file is not JSON.
$ ~/glibc/benchtests/scripts/validate_benchout.py bench-memset.out \
~/glibc/benchtests/scripts/benchout_strings.schema.json
Traceback (most recent call last):
File "/home/naohirot/glibc/benchtests/scripts/validate_benchout.py", line 86, in <module>
sys.exit(main(sys.argv[1:]))
File "/home/naohirot/glibc/benchtests/scripts/validate_benchout.py", line 69, in main
bench.parse_bench(args[0], args[1])
File "/home/naohirot/glibc/benchtests/scripts/import_bench.py", line 139, in parse_bench
do_for_all_timings(bench, lambda b, f, v:
File "/home/naohirot/glibc/benchtests/scripts/import_bench.py", line 107, in do_for_all_timings
if 'timings' not in bench['functions'][func][k].keys():
AttributeError: 'str' object has no attribute 'keys'
$ ~/glibc/benchtests/scripts/validate_benchout.py bench-math-inlines.out \
~/glibc/benchtests/scripts/benchout_strings.schema.json
Traceback (most recent call last):
File "/home/naohirot/glibc/benchtests/scripts/validate_benchout.py", line 86, in <module>
sys.exit(main(sys.argv[1:]))
File "/home/naohirot/glibc/benchtests/scripts/validate_benchout.py", line 69, in main
bench.parse_bench(args[0], args[1])
File "/home/naohirot/glibc/benchtests/scripts/import_bench.py", line 137, in parse_bench
bench = json.load(benchfile)
File "/usr/lib/python3.6/json/__init__.py", line 299, in load
parse_constant=parse_constant, object_pairs_hook=object_pairs_hook, **kw)
File "/usr/lib/python3.6/json/__init__.py", line 354, in loads
return _default_decoder.decode(s)
File "/usr/lib/python3.6/json/decoder.py", line 342, in decode
raise JSONDecodeError("Extra data", s, end)
json.decoder.JSONDecodeError: Extra data: line 1 column 17 (char 16)
Reviewed-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
This patch removed redundant "#include <assert.h>" from
bench-memset-large.c and bench-memset-walk.c.
Reviewed-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
This patch enables scripts/plot_strings.py to read a benchmark result
file from stdin.
To keep backward compatibility, that is to keep accepting multiple of
benchmark result files in argument, blank argument doesn't mean stdin,
but '-' does.
Therefore nargs parameter of ArgumentParser.add_argument() method is
not changed to '?', but keep '+'.
ex:
$ jq '.' bench-memset.out | plot_strings.py -
$ jq '.' bench-memset.out | plot_strings.py - bench-memset-large.out
$ plot_strings.py bench-memset.out bench-memset-large.out
error ex:
$ jq '.' bench-memset.out | plot_strings.py
Reviewed-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
They provide TLS_GD/TLS_LD/TLS_IE/TLS_IE macros for TLS testing. Now
that we have migrated to __thread and tls_model attributes, these macros
are unused and the tls-macros.h files can retire.
Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
These comments refer to slow paths that were removed in
glibc 2.34 or earlier. The corresponding "names" that yield
separate workload traces for "make bench" are thus obsolete.
We are however keeping the corresponding inputs.
Reviewed-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
The benchmark and tests must fail in case of allocation failure in the
implementation array. Also annotate the x* allocators in support.h so
that the compiler has more information about them.
Reviewed-by: Florian Weimer <fweimer@redhat.com>
This patch fixed mprotect system call failure on AArch64.
This failure happened on not only A64FX but also ThunderX2.
Also this patch updated a JSON key from "max-size" to "length" so that
'plot_strings.py' can process 'bench-memcpy-random.out'
This patch adds support for testing memcpy with both dst > src and dst
< src. Since memcpy is implemented as memmove which has seperate
control flows for certain sizes depending on dst > src it seems like
1) information that should be provided in the benchtest output and a
variable that can be controlled for the benchmarks.
Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
No bug. This commit adds some additional performance test cases to
bench-memcmp.c and test-memcmp.c. The new benchtests include some
medium range sizes, as well as small sizes near page cross. The new
correctness tests correspond with the new benchtests though add some
additional cases for checking the page cross logic.
Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
Convert the output of benchtests/bench-rawmemchr to JSON like other string
benchmarks. This makes the output more parseable and allows usage of
compare_strings.py, for example.
Reviewed-by: Lucas A. M. Magalhaes <lamm@linux.ibm.com>
These workload traces cover the whole "long double" range.
This patch was prepared with the help of Adhemerval Zanella.
Reviewed-by: Carlos O'Donell <carlos@redhat.com>
No bug. This commit adds some additional cases for bench-memchr.c
including testing medium sizes and testing short length with both an
inbound match and out of bound match.
Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
No bug. This commit adds tests cases and benchmarks for page cross and
for memset to the end of the page without crossing. As well in
test-memset.c this commit adds sentinel on start/end of tstbuf to test
for overwrites
Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
The benchtests json allows {function {variant}} categorization of
results whereas the pthread-locks tests had {function {variant
{subvariant}}}, which broke validation. Fix that by serializing the
subvariants as variant-subvariant. Also update the schema to
recognize the new benchmark attributes after fixing the naming
conventions.
No Bug. This commit expanding the range of tests / benchmarks for
memmove and memcpy. The test expansion is mostly in the vein of
increasing the maximum size, increasing the number of unique
alignments tested, and testing both source < destination and vice
versa. The benchmark expansaion is just to increase the number of
unique alignments. test-memcpy, test-memccpy, test-mempcpy,
test-memmove, and tst-memmove-overflow all pass.
Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
This patch adds workload traces for all double format functions where such
files are missing. For each function, a set of 1000 random values is
generated at random using SageMath, such that the output values are
meaningful (for example avoiding too large inputs for exp10 where the
output would be +Inf). More details about the generated values are
given at the beginning of each file.
Reviewed-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
This patch updates json "bench-variant" attribute of "bench-memset.c"
to "default" so that the script "benchtests/scripts/plot_strings.py"
can generate a file "memset_time_default_linear.png".
Without this patch, the script "benchtests/scripts/plot_strings.py"
generates a file "memset_time__linear.png" which has inconsistent form
with "memcpy_time_default_linear.png" and
"memmove_time_default_linear.png".
Since commit 2682695e5c, `make bench-build' with `--enable-static-pie'
fails due to bench-timing-type being incorrectly built with MODULE_NAME
set to `libc'. This commit sets MODULE_NAME to nonlib, thus fixing the
build failure.
Reviewed-by: Carlos O'Donell <carlos@redhat.com>
GNU ld and gold have supported --print-output-format since 2011. glibc
requires binutils>=2.25 (2015), so if LD is GNU ld or gold, we can
assume the option is supported.
lld is by default a cross linker supporting multiple targets. It auto
detects the file format and does not need OUTPUT_FORMAT. It does not
support --print-output-format.
By parsing objdump -f, we can support all the three linkers.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
I used these shell commands:
../glibc/scripts/update-copyrights $PWD/../gnulib/build-aux/update-copyright
(cd ../glibc && git commit -am"[this commit message]")
and then ignored the output, which consisted lines saying "FOO: warning:
copyright statement not found" for each of 6694 files FOO.
I then removed trailing white space from benchtests/bench-pthread-locks.c
and iconvdata/tst-iconv-big5-hkscs-to-2ucs4.c, to work around this
diagnostic from Savannah:
remote: *** pre-commit check failed ...
remote: *** error: lines with trailing whitespace found
remote: error: hook declined to update refs/heads/master
Performance benchmarks for various posix locks: mutex, rwlock,
spinlock, condvar, and semaphore. Each test is performed with
an empty loop body or with a computationally "interesting" (i.e.
difficult to optimize away, and used just to allow lock code to
be "hidden" in the filler's CPU cycles).
__float128 is a non-standard name and is not available on some architectures
(like aarch64 or s390x) even though they may support the standard _Float128
type. Other architectures (like armv7) don't support quad-precision
floating-point operations at all.
This commit replaces benchtests references to __float128 with _Float128 and
runs the corresponding tests only on architectures that support it.
Wilco Dijkstra
Paul Eggert
Frederic Berat
Paul Pluzhnikov
Carlos O'Donell
Joe Ramsay
Nisha Menon
Adhemerval Zanella Netto
Joseph Myers
Noah Goldstein
Siddhesh Poyarekar
Wangyang Guo
H.J. Lu
Su Lifan
Sunil K Pandey
Wilco Dijkstra
Naohiro Tamura
Fangrui Song
Paul Zimmermann
Naohiro Tamura
Matheus Castanho
Raphael Moreira Zinsly
Arjun Shankar
DJ Delorie