2012-08-08
* 4b4f87ff ChangeLogP5000.txt: Changelog for PAPI5
* 6f208c06 doc/Doxyfile-common papi.spec src/Makefile.in...: Bump version
numbers in prep for a 5.0 release.
* c6fdbd11 release_procedure.txt: Update release_procedure.txt Change the
order of when we branch git, so that the main dev branch gets some of the
release related changes.
2012-04-17
* 97d4687f ChangeLogP440.txt: Pickup the changelog from papi 4.4 This was
only included in the stable-4.4 branch.
2012-08-23
* 628c2b6e src/buildbot_configure_with_components.sh: Take debug out of the
with several components build test config. When built with PAPI's memory
wrapper routines, the threaded stress tests will sometimes get into poor
performing situations. See trac ticket 148 for discussion.
http://icl.utk.edu/trac/papi/ticket/148
2012-08-22
* 46faae8e src/ctests/overflow2.c src/ctests/overflow_single_event.c
src/ctests/overflow_twoevents.c...: Move find_nonderived_event() from
overflow_twoevents to test_utils and call it from overflow2 and
overflow_single_event to insure that we're not trying to overflow on a
derived event.
* 3e7d8455 src/ctests/zero_smp.c: Fix a memory leak reported on the aix
power7 machine. zero_smp.c did not unregister at the end of its thread
function.
* 3ad5782f src/perf_events.c: perf_events: fix segfault if DEBUG is enabled
Was incorrectly using "i" as an index where it should be "0" in a debug
statement.
2012-08-21
* a3cadbdb src/ftests/accum.F src/ftests/avail.F src/ftests/case1.F...: Take
#2. Changing len_trim function in ftests to last_char. This time, I respect
72 char line limit.
* c9db8fbf src/ctests/overflow_force_software.c: overflow_force_software was
the only test that used a different hard_tolerance value (0.25) than the
other overflow tests (0.75). This caused trouble on Power7/AIX. Now we are
using the same hard_tolerance value in all overflow tests.
* 70515343 src/ftests/accum.F src/ftests/avail.F src/ftests/case1.F...:
Changed name of function len_trim to last_char.
* 95168d79 src/components/cuda/linux-cuda.c: Cleanup cuda shutdown code.
* The
shutdown_thread code cleaned out the whole component's state. This has been
split into shutdown_global for the whole component, and shutdown_thread is
left to cleanup some control state info.
* 56284f81 src/ctests/multiplex1_pthreads.c: Fix memory leaks in pthread
multiplex tests.
* aeead8b6 src/threads.c: Remove an outdated comment about
_papi_hwi_free_EventSet holding INTERNAL lock
* e598647b src/perf_events.c: perf_events: fix issue where we dereference a
pointer before NULL check. Fix suggested by Will Cohen, based on a coverity
report.
* 4e0ed976 src/ctests/calibrate.c: Modify warning message to eliminate the
word "error" Hopefully this will suppress it in buildbot outputs.
* 50fbba18 src/ctests/api.c src/ftests/case2.F: Cleanup a few more warnings
from the PAPI_perror change.
* 1f06bf28 src/ftests/case2.F: Missed an instance of perror in the fortran
code.
* 93e6ae2c src/ftests/ftests_util.F: Fix warning in ftest_util.F
2012-08-20
* 60c6029e src/perf_events.c: perf_events: Update multiplexing code It
* turns
out the PERF_EVENT_IOC_RESET ioctl resets the count but not the multiplexing
info. This means that when we fiddle with the events then reset them in
check_scheduability(), we are not really resetting things to zero. The
effect might be small, but since the new multiplex code by definition is
always scheduable, then let's skip the test if multiplexing.
* 9079236c src/ctests/zero.c: Change error reporting so FLOPS > 100% above
theoretical FAIL and FLOPS > 25% above theoretical WARN.
2012-08-18
* 980558af src/papi_internal.c: papi_internal: fix memory leak When I made
some changes a while back I forgot to free ESI->NativeBits properly. This
was causing memory leak warnings on buildbot.
2012-08-17
* 83a14612 src/perf_events.c: perf_events: more cleanups and comments We
really need to go back and figure out in more detail what the
profile/sampling/overflow code is doing.
* 7cafb941 src/perf_events.c: perf_events: more cleanups and comments
* e9e39a4b src/perf_events.c: perf_events: disable kernel multiplexing
* before
2.6.34 It turns out even our simple multiplexing won't work on kernels
before 2.6.34, so fall back to sw multiplex in that case.
* 05801901 src/perf_events.c: perf_events: more cleanup and comments
* 268e31d7 src/perf_events.c: perf_events: more cleanup and commenting
* d62fc2bf src/perf_events.c: perf_events: more cleanup and comments
* fb0081bc src/perf_events.c: perf_events: more cleanups and comments
* a1142fc8 src/perf_events.c: perf_events: cleanup and comment the kernel
* bug
workarounds
* b8560369 src/perf_events.c: perf_events: minor cleanups and new comments
* 6c320bb2 src/perf_events.c: perf_events: fix some debug messages I forget
to test with --with-debug enabled
* f7a3cccf src/perf_events.c: perf_events: enable new read_code This makes
the read code much simpler. It finishes the multiplexing changes. To avoid
complication, we no longer enable PERF_FORMAT_ID as reading that extra info
is unnecessary with the current implementation. This passes all the tests
on
a recent kernel, but on 2.6.32 there are still a few issues.
* 15749cff src/ctests/all_events.c src/ctests/all_native_events.c: Fix
warning in all_events and all_native_events. In the perror semantic change,
several strings for use in the old interface were left.
2012-08-16
* afdd25fa src/perf_events.c: perf_events: always enable kernel multiplexing
The new code should work on any kernel version.
* 9f5e23ae src/perf_events.c: Rewrite multiplex support. Drop support for
the former "partitioned" multiplexing, as we could never use it. Instead
use
the simple/braindead model. This still needs more work, as sometimes reads
are failing.
* cdd29909 src/ftests/strtest.F: Fix strtest.F ftest It was still making
some assumptions about PAPI_perror() writing to a string rather than
directly
to standard error.
* 565f60b3 src/papi_internal.c: Missing code to set num_error_chunks to 0
The new _papi_hwi_cleanup_errors() function was not resetting
num_error_chunks to 0, leading to a segfault in the fmultiplex1 test.
2012-08-02
* bb85bafd src/genpapifdef.c src/papi.c src/papi_common_strings.h...: Remove
usage of _papi_hwi_err. Move PAPI over to storing errors in a runtime list.
* Functions to add/lookup errors. * Generate the list of PAPI_E* errors at
library_init time. * genpapifdef pulled the values for the PAPI_* error
return codes from the _papi_hwi_err structure at configure time. Since this
is now built at run-time, I added the appropriate values to genpapifdef's
builting describe_t table. See : _papi_hwi_publish_error
_papi_hwi_init_errors For usage hints.
2012-08-10
* e27af085 src/perf_events.c: perf_event: rename BRAINDEAD_MULTIPLEXING It
is now "simple_multiplexing" and is a variable not an #ifdef This is needed
before perf_event multiplexing can be sorted out. It's unclear if it
actually works anyway.
* 7f8e8c58 src/perf_events.c: perf_event: remove context "cookie" field It
was a bit of overkill, we just need an initialized field. Also revamp how
context and control are initialized.
* 8cb8ac6d src/perf_events.c: perf_event: move all event specific info to
* the
control state previously half was in the context state and half in the
control state perf_event has a strange architecture with each event being
created having its own fd, which is context wide. In PAPI though we usually
only have one eventset (control state) active at once, so there's no need to
have the context be aware of this.
2012-08-09
* 8d7782cb src/perf_events.c: perf_event: rename evt_t to perf_event_info_t
This just makes the code easier to follow.
* 349de05c src/perf_events.c: perf_event: remove the superfluous
per_event_info_t structure
2012-08-08
* da8ad0a2 src/ctests/all_native_events.c src/ctests/get_event_component.c
src/utils/native_avail.c: Fix warnings about PAPI_enum_cmp_event() return
not
being checked Reported by coverity checker via Will Cohen Harmless
warnings, and now the checker will likely complain about the value being
checked but ignored.
* b4719888 src/papi_user_events.c: Fix unused value in papi_user_events.c
Reported by Coverity checker by Will Cohen
* 6a8f255c src/utils/event_chooser.c: remove unused
* PAPI_get_component_info()
call in event_chooser Reported by Will Cohen from coverity checker
2012-08-06
* 62cda478 src/genpapifdef.c src/papi_common_strings.h
src/papi_internal.c...: Remove usage of _papi_hwi_err. genpapifdef pulled
the values for the PAPI_* error return codes from the _papi_hwi_err structure
at configure time. Since this is now built at run-time, I added the
appropriate values to genpapifdef's builting describe_t table.
2012-08-02
* d11259f3 src/papi.c src/papi_internal.c src/papi_internal.h...: Move over
to generating the list of PAPI errors at library_init time.
* 097ffc44 src/papi_internal.c: Functions to add/lookup errors.
2012-08-07
* 2530533f src/papi_events.csv: tests/zero fails on Power7 due to PAPI_FP_INS
Error of 50%. Preset definition has been redefined and test now passes.
* 8e17836f src/components/appio/Rules.appio src/components/appio/appio.c
src/components/appio/appio.h...: We now intercept recv(). The support for
recv() requires dynamic linkage. For static linkage, recv is not intercepted.
2012-08-06
* 8b1eb84c src/perf_events.c: perf_events: some whitespace cleanup and extra
comments
* f10edba6 src/perf_events.c: perf_events: MAX_READ was no longer being used,
remove it
* 08c06ed1 src/perf_events.c: perf_event event_id is actually 64-bit, so make
our copy match
* a33e8d9c src/perf_events.c: Rename context_t pe_context_t in perf_events.c
Makes the code a bit clearer and matches how other components name things.
* 96ce9dcd src/perf_events.c: Rename control_state_t pe_control_state_t This
makes the code a bit easier to follow and matches how other components name
things.
2012-08-03
* 4c5dce7f src/ctests/zero.c: Beef up error reporting.
* 83b5d28a src/ctests/cycle_ratio.c: Have the cycle_ratio test skip if
PAPI_REF_CYC event is not defined.
2012-08-02
* 25b1ba41 src/ctests/cycle_ratio.c: Removed all TESTS_QUIET blocks. They
aren't needed because tests_quiet() overloads printf. We should probably
remove TEST_QUIET blocks in ALL tests at some point for code clarity…
* 8777d7d4 src/ctests/zero.c: Fixed error reporting. The error computation
was inside a TESTS_QUIET block and wasn't getting executed when run quietly.
Thus this test always passed on buildbot, even when it didn't.
* 006fe8e9 src/ctests/Makefile: Fix typo in cycle_ratio make line.
* 88e6d6a4 src/aix.c src/aix.h: Setting number of multiplex counters back to
32 for AIX. Before it was set equal to number of max HW counters. This caused
ctests/sdsc-mpx to fail.
* ab78deda src/papi_events.csv: ctests/calibrate on Power7/AIX failed with a
50% error all the way through. Updated the preset FP_OPS with a more
appropriate definition. Now the calibrate errors range from 0.0002 to 0.0011%
for double and single precision
* fadce32f src/ctests/calibrate.c: Modify calibrate test in two ways: 1. add
a -f flag to suppress failures and allow test to run to completion; 2. change
error detection to allow warnings above MAX_WARN and failures above MAX_FAIL.
Currently set to 10% and 80% respectively. This allows speculative over
counting to pass with warning rather than fail completely.
* 8a39ac9d src/papi_events.csv: LST_INS for Power7 was defined from 3 native
events that cannot be counted together at the same time. Caused
ctests/all_events to fail. Updated the preset with a more appropriate
definition.
* cdc16e5d src/papi_events.csv: L1_DCA for Power7 was defined from 3 native
events that cannot be counted together at the same time. That caused
ctests/tenth to fail. Updated the preset with a more appropriate definition.
2012-08-01
* 2bf44d13 src/papi_internal.c src/perf_events.c: icc does not like
arithmetic on void pointers. Added cast to unsigned char* when arithmetic was
being performed on void pointers in papi_internal and perf_events.
* 7825ec14 src/ctests/api.c src/ctests/attach2.c src/ctests/attach3.c...:
Modify tests that FAIL if PAPI_FP_OPS or PAPI_FP_INS not implemented. Now
they will warn and continue. This is specifically to accommodate the
brain-dead IvyBridge implementation.
* fd70a015 src/testlib/test_utils.c: Re-writing of test_utils introduced new
bugs that caused ctests/tenth to fail. test_events struct lists the same
event twice (MASK_L1_DCW), hence PAPI_add_event() fails because it's forced
to add the same preset twice.
* 74ece3a0 src/run_tests.sh: run_tests.sh was clobbering $EXCLUDE variable if
$CUDA was defined. Changed to add entries from run_tests_exclude_cuda.txt to
$EXCLUDE which should already contain entries from run_tests_exclude.txt
instead of replacing the entries already contained.
* 11ed2364 src/libpfm4/config.mk: Added check in libpfm4/config.mk to check
if using icc. If so, the -Wno-unused-parameter flag will no longer be used
because icc does not provide it and provides no alternative.
* dedf73f6 src/papi_user_events.c: fget() returns an int it should be treated
as an int The coverity scan flagged that the int return by fget was stored
in a char. The main concern with this is the EOF that fget() could return is
-1. Do not want to mess up that value by typecasting to char and then back to
int.
* c4fcbe7e src/ctests/kufrin.c: Check return values of PAPI_get_cmp_opt() and
calloc A coverity scan showed that PAPI_get_cmp_opt() could potentially
return a negative number. Also it is good form to check the return value of
calloc to ensure it is a non-null pointer.
* e89d6ffa src/testlib/test_utils.c: Clean up test_print_event_header()
There were a couple warnings flagged by coverity on
test_print_events_header(). The function now checks for error conditions
flagged by PAPI_get_cmp_opt() and also frees memory allocated by a calloc()
function.
* c81d8b60 src/threads.h: Eliminate deadcode from threads.h If
HAVE_THREAD_LOCAL_STORAGE is defined, a portion of the
_papi_hwi_lookup_thread() will never be executed. This patch make either one
section or the other section of code be compiled. This will eliminate a
coverity scan warning about unreachable code.
* f70f3f56 src/ctests/all_native_events.c: Eliminate unused variable in
ctests/all_native_events.c Coverity identified a variable that was set but
never used in all_native_events.c. This patch removes the unused variable to
eliminate that warning.
* a9f29840 src/components/appio/appio.c: A couple places in appio.c used the
FD_SET() without initializing the variable. Coverity scan pointed out this
issue.
* 9e535ae2 src/components/rapl/linux-rapl.c: A Coverity scan pointed out that
read_msr() could potentially use an invalid value of fd for pread(). Need to
check the value of fd before using it.
* 7b55c675 src/components/rapl/linux-rapl.c: The arrays used for
initialization were hard coded to 1024 packages. Want to avoid hard coding
that so the day when machines with 1025 packages are available is a
non-event. Also changed the initialization code to avoid having the
initialization be O((number of packages)^2) in time complexity.
2012-07-27
* 3703995a src/papi_internal.c: Fix the component name predending code. When
presented with a NULL component .short_name, the code did the wrong thing.
* 5258db8b src/components/cuda/linux-cuda.c: Fix a warning in cuda.
2012-07-26
* ddd6f193 src/ctests/Makefile src/ctests/cycle_ratio.c: Add a test to
compute nominal CPU MHz from real cycles and use PAPI_TOT_CYC and
PAPI_REF_CYC to compute effective MHz. Warns if PAPI_REF_CYC is zero, which
can happen on kernels < ~3.3.
* fab5e9ef src/papiStdEventDefs.h src/papi_common_strings.h
src/papi_events.csv: Add PAPI_REF_CYC preset event. Define it as
UNHALTED_REFERENCE_CYCLES for all Intel platforms on which this native event
is supported.
2012-07-25
* 8b9b6bef src/papi_events.csv: Modify SandyBridge and IvyBridge tables:
SandyBridge FP_OPS only counts scalars; SP_OPS and DP_OPS now count
correctly, including SSE and AVX. IvyBridge can't count FP at all;
adjustments made to eliminate event differences with SandyBridge.
2012-07-26
* 5b11c982 src/components/cuda/linux-cuda.c: Fix the cuda component. The
cuda component prepended CUDA. to all its event names, this is no longer the
case.
2012-07-25
* db5b0857 src/papi_events.csv: Added 2 new preset definitions for BGQ. Note,
these presets use the new feature where a generic event mask together with an
ORed opcode string is used. This won't work until the new Q driver is
released (currently scheduled for end of August).
2012-07-24
* af7cd721 src/components/coretemp/linux-coretemp.c
src/components/coretemp/tests/coretemp_pretty.c
src/components/cuda/linux-cuda.c...: Enforce all our components to use the
same naming. We setteled on :'s as inter-event seperators. This also touches
a few of the components' tests, we changed the name field so their searches
needed help.
2012-07-23
* 57aeb9d4 src/papi_internal.c: Prepend component .short_name to each event
name. Use ::: as a sep.
2012-07-24
* 762e9584 src/ctests/multiplex2.c src/sw_multiplex.c: Fix multiplex2 test
It complained if it tried to add a multiplex event and PAPI properly told it
that it couldn't.
* 531870f1 src/papi_internal.c: Add sanity check at component init time
Looks for num_cntrs being larger than num_mpx_cntrs which doesn't make much
sense.
* 53ad0259 src/extras.c src/genpapifdef.c src/papi.c...: Rename
PAPI_MAX_HWCTRS to PAPI_EVENTS_IN_DERIVED_EVENT Hopefully this will make
things a little less confusing.
* 700af24b src/papi_internal.c: Change EventInfoArrayLength to always return
num_mpx_cntrs Things should be consistently using num_mpx_cntrs rather than
num_cntrs now. Issue reported by Steve Kaufmann
* d1570bec src/sw_multiplex.c: Fix sw_multiplex bug when max SW counters is
less than max HW counters this was causing kufrin and max_multiplex to fail
* f47f5d6a src/aix.c src/components/appio/appio.c
src/components/bgpm/CNKunit/linux-CNKunit.c...: Remove PAPI_MPX_DEF_DEG It
was not well documented and being used in confused ways all over the code.
Now there is a different define PAPI_MAX_SW_MPX_EVENTS used by the software
multiplex code. All other components have had the value replaced with just
the maximum number of counters. If a component can handle its own
(non-software) multiplexing it is up to it to set .num_mpx_cntrs to a value
that's different from .num_cntrs
* 0d83f5db src/papi_internal.c src/papi_internal.h: Split NativeBits off of
NativeInfoArray in EventSet previously we were doing some crazy thing where
we allocated both at once and then split them afterward. The new code is
easier to follow.
* 98f2ecbd src/papi_internal.c: Clean up EventSet creation Sort out which
sizes are used for allocating which structures.
* e1024579 src/Makefile.inc src/multiplex.c src/multiplex.h...: Rename the
multiplex files to be sw_multiplex That way it's more clear the stuff
included only relates to software multiplexing, not generic multiplexing.
* a6adc7ff src/multiplex.h src/papi_internal.c src/papi_internal.h: Move some
sw-multiplex specific terms out of papi_internal.h and into multiplex.h
2012-07-23
* 1ddbe117 src/components/README: Added note that lmsensors component
requires lmsensors version >=3.0.0
* 94676869 src/components/appio/appio.c
src/components/appio/tests/appio_test_pthreads.c: proper checking of return
codes in response to tests using coverity
* ea958b18 src/components/appio/tests/appio_list_events.c
src/components/appio/tests/appio_values_by_code.c: As component name in table
has been changed from appio.c to appio, we now use appio in the tests.
2012-07-20
* f212cc34 src/components/appio/appio.c
src/components/coretemp/linux-coretemp.c
src/components/coretemp_freebsd/coretemp_freebsd.c...: Add .short_name
entries to each component.
* 1e755836 src/papi_libpfm4_events.c src/perf_events.c: Fix use-after-free
bug in perf_events.c This turned up in the ctests/flops test, and Valgrind
made it easy to track down.
* 4580ed1d src/perf_events.c: Update perf_event.c rdpmc support Use the
libpfm4 definition for mmap rather than our custom one, now that libpfm4 has
been updated
* 47558b2c src/libpfm4/examples/showevtinfo.c
src/libpfm4/include/perfmon/perf_event.h
src/libpfm4/lib/pfmlib_intel_nhm_unc.c...: Import current libpfm4 from
libpfm4 git It has some minor uncore fixes plus the header changes needed
for rdpmc.
2012-07-17
* 65d4c06c src/linux-common.c: Reorder statements to ensure the fclose() are
performed Coverity pointed out that it was possible for resources to be
leaked in linux-common.c if the fscanf() encountered error. This reordering
of the statements ensures that the fclose() calls are done regards of the
results of the fscanf() functions.
2012-07-18
* 7bf071ff src/papi_user_events.c: Ensure that load_user_event_table() frees
files and memory on error A Coverity scan showed that an error condition in
load_user_event_table() function would exit the the function without closing
the table file or freeing allocated memory. This patch addresses those
issues.
2012-07-17
* 1ba52e35 src/components/stealtime/linux-stealtime.c: Ensure that
read_stealtime() closes the file in case of an error condition A Coverity
scan showed that an error condition could cause read_stealtime() to exit
without closing the file. This patch ensures that the file is closed
regardless of success or failure.
2012-07-18
* f37f22e5 src/papi_libpfm4_events.c: Fix warning in papi_libpfm4_events.c
We were setting a value but never using it.
* 8e8401bc src/testlib/test_utils.c: Remove unused variable in test_utils.c
Most of the machines in buildbot were complaining about this.
* 133ce6a9 src/linux-timer.c: Add missing papi_vector.h include to
linux-timer.c This was breaking on PPC Linux
2012-07-17
* 6fd3cedd src/perf_events.c: Fix perf_events.c warnings reported by ICC
* 21c8f932 src/perfctr-x86.c: Fix perfctr-x86 build with debug enabled
* 08f76743 src/configure src/configure.in src/linux-bgq.c: Attempt to fix
linux-bgq compilation error. It turns out BGQ uses the standard
linux-context.h header
* 43457f4f src/linux-bgq.c: Made check for opcodes more robust.
* d58116b4 src/perf_events.c: More cleanups of perf_events.c file
* 409438b7 src/freebsd-context.h src/freebsd.c src/freebsd.h: Fix FreeBSD
compile warnings Similar to the perfctr issues.
* 1e6dfb02 src/aix.c src/aix.h: Fix AIX build warnings They were similar in
cause to the perfctr issues.
* 3d0b5785 src/Rules.perfmon2 src/components/appio/appio.c
src/components/bgpm/CNKunit/linux-CNKunit.h...: Remove papi_vector.h include
from papi_internal.h There were some semi-circular dependencies that came up
with the context split changes. The easiest way to fix things for perfctr
was just move papi_vector.h out to be included explicitly. This touches a
lot of files because a lot of files include papi_internal.h This should also
fix the perfctr and perfmon2 builds that were broken yesterday.
2012-07-16
* a7a14a5b src/ctests/zero.c src/testlib/test_utils.c: Modify zero test to
warm up processor before measuring events, and report timing errors as signed
deviations. Modify test_utils add_two_events code to check for errors after
adding nominally valid events. This is a more rigorous test than just
counting available registers.
* de0860d3 src/perf_events.c: Remove perf_events.h module header It's no
longer needed, everything important is merged into the perf_events.c file.
* 22975f14 src/perf_events.c: Remove perf_event SYNCHRONIZED_RESET code This
was never defined and never used, just remove the code.
* 48750b8c src/perf_events.c: Remove papi_pe_allocate_registers On
perf_event this code wasn't really doing anything useful, as
update_control_state would end up re-doing any possible tests we could want
to do here.
* 1775566f src/Makefile.in src/Makefile.inc src/Rules.pfm4_pe...: Remove
"include CPUCOMPONENT" from papi_internal.h This was the last major
dependency on CPU component in common PAPI code. It was mostly necessary for
the ucontext definitions when trying to get the instruction pointer when
doing sampling. This change creates new OS-specific header files that handle
the ucontext related code, and has papi_internal.h include that instead.
* 969ce035 src/Rules.pfm4_pe src/Rules.pfm_pe src/configure...: Make
perf_event libpfm4 only perf_event libpfm3 support is not really needed
anymore and supporting it was cluttering up the perf_event component.
2012-07-13
* adad1d2a src/perf_events.c: Add init time error messages to perf_event
component
* 827ccc07 src/perf_events.c: Add perf_event rdpmc / fast_real_timer
detection Currently we need a custom copy of struct perf_event_mmap_page
because the version included with libpfm4 doesn't define the fields we need
yet.
* 4f82fe94 src/perf_events.c: Read in paranoid info on perf_events This
indicates whether a regular user can read CPU-specific or system-wide
measurements.
* 03080450 src/perf_events.c: Add proper perf_event detection Using the
official /proc/sys/kernel/perf_event_paranoid file
* 6e71d3f7 src/linux-bgq.c: Updated BGQ opcode stuff; cleaned up code.
2012-07-11
* 3114d3dc src/multiplex.c src/papi_internal.c src/perf_events.c: Minor
documentation improvements Plus fixes some typos
2012-07-09
* b60c0f0c src/perf_events.c: Minor cleanups to perf_events.c
* 075278a0 src/aix.c src/freebsd.c src/linux-bgp.c...: Change return value
for .allocate_registers For some reason it returned 1 on success and 0 on
failure. Change it so you return PAPI_OK on success and a PAPI error on
failure, to better match all of the other component vectors.
* 29d9e62b src/testlib/test_utils.c: Fixed the print_header routine to report
an error message if counters are not found, instead of a negative counter
number. Tested by forcing the return value negative; not by running on a Mac,
where the error first appeared.
* 74257334 src/ctests/Makefile src/ctests/remove_events.c: Add remove_events
test This just makes sure EventSets still work after an event has been
removed. This is probably covered by other more elaborate tests, but I
needed a simple test to make sure I wasn't breaking anything.
* 1372714f src/papi.c src/papi_internal.c src/papi_internal.h: Clean up,
rename, and comment _papi_hwi_remap_event_position I've found this section
of code to be confusing for a long time. I think I finally have it mostly
figured out. I've renamed it _papi_hwi_map_events_to_native() to better
describe what it does. The issue is that the native event list in an
EventSet can change at various times. At event add, event remove, and
somewhat unexpectedly, whenever ->update_control_state is called (a component
can re-arrange native events if it wants, to handle conflicts, etc.) Once
the native event list has been changed, _papi_hwi_map_events_to_native() has
to be called to make sure the events all map to the proper native_event
again. Currently we have the _papi_hwi_map_events_to_native() calls in odd
places. It seems to cover all possible needed locations, but analyzing that
we do takes a lot of analysis...
* f1b837d8 src/papi.c: Remove unused variable in papi.c
* 541bcf44 src/papi_internal.h: Update commens in papi_internal.h Some of
the EventSetInfo comments were out of date.
* e6587847 src/papi.c src/papi_internal.c src/papi_internal.h: Remove unused
paramater from _papi_hwi_remap_event_position The mysterious
_papi_hwi_remap_event_position function had a "thisindex" field that was
ignored. This will possibly speed PAPI_start() time as it was running a loop
over num_native_events on _papi_hwi_remap_event_position even though each
call did the same thing since the value being passed was ignored.
* 3ad3d14b src/papi_internal.c: Clean up and comment add_native_events in
papi_internal.c I'm chasing some weird perf_events behavior with the
papi_event_chooser. The add_native_events code is very hard to understand,
working on commenting it more.
* 4e5e7664 src/utils/event_chooser.c: Fix coverity warning in
papi_event_chooser
* 666249a8 src/jni/EventSet.java src/jni/FlipInfo.java
src/jni/FlopInfo.java...: RIP Java. Java PAPI wrappers have not been
supported for years (2005?). They are being removed to declutter the source.
* e18561fc src/papi_preset.c: Update cmpinfo->num_preset_events properly
This value wasn't being set if we were reading the presets directly from the
CSV file.
* 290ab7c3 src/utils/component.c: Have papi_component_avail report counter
and event info
* 7c421b9c src/testlib/test_utils.c src/utils/native_avail.c: Remove counter
number from the testlib header. The header was only reporting number of
counters for the CPU component, even though the header is printed for many
utils and the CPU component might not even be involved. This could be a bit
confusing, so remove it.
* 26432359 src/darwin-common.c src/darwin-memory.c: Improve OSX support This
properly detects CPU information now. You can get results like this:
Available native events and hardware information.
- PAPI Version : 4.9.0.0 Vendor string and code : GenuineIntel
(1) Model string and code : Intel(R) Core(TM) i5-2435M CPU @ 2.40GHz (42)
CPU Revision : 7.000000 CPUID Info : Family: 6
Model: 42 Stepping: 7 CPU Max Megahertz : 2400 CPU Min Megahertz
: 2400 CPUs per Node : 0 Total CPUs : 4 Running in
a VM : no Number Hardware Counters : -4 Max Multiplex Counters :
-4
-
2012-07-08
* 845d9ecb src/Makefile.inc src/configure src/configure.in...: Add Mac OSX
support This is enough that things compile and simple utilities run. No CPU
perf counter support.
2012-07-06
* ff6f9ab4 src/linux-bgq.c: missed to delete a debug output.
2012-04-17
* 12e9a11a RELEASENOTES.txt: Release notes for the 4.4 release.
2012-07-06
* ac2eac56 src/papi.c src/papi.h: Add a PAPI_disable_component_by_name entry
point.
* 8c490849 src/components/coretemp_freebsd/coretemp_freebsd.c src/freebsd.c:
Fix FreeBSD to work I'm not sure how it ever worked in the past. With these
changes I can at least do a papi_component_avail and a papi_native_avail and
get sane results
* 108b5ce6 src/freebsd.c src/freebsd.h src/freebsd/map-atom.c...: Fix FreeBSD
build some of the recent changes broke the FreeBSD build
* 40a60f0a src/linux-bgq.c src/linux-bgq.h: Added BGQ's opcode and generic
event functionality to PAPI. For BGQ there are multiple ways to define
presets. The naive way is to derive from multiple events. This eats up
multiple counters and we lose sample capability as well as overflow
capability. On the other side, some events come with multiple InstrGroup
derived in the field. If that's the case we can use a generic event and
opcodes to filter multiple groups in a single counter. This is not working
properly yet due to a known error in BGPM. Bgpm_AddEvent() does not work
properly when multiple generic events are added to an EventSet. The BGPM
folks have been made aware of this issue, they confirmed the error, and they
are currently working on a fix.
* 6f72b70f src/papi_events_table.sh: Make this script robust enough to handle
any line ending, including CR (Mac), CRLF (Windows), and LF (Unix). It
appears that google mail now automagically converts attached files to CRLF
format.
* 765ed0d2 src/papi_internal.c: Fix a type warning in the UE code.
* 94bc1b15 src/MACROS: Remove the MACROS file it held out of date info and
hasn't been touched since 2004
* d19e73ba src/ctests/Makefile src/ctests/clockcore.c
src/testlib/Makefile...: Move the clockcore.c file from ctests to testlib
it's common code used by multiple tests, including some in the utils
directory also add a function definition to fix a build-time warning
* 1101a6aa src/aix-lock.h src/aix.h src/configure...: Make papi_lock.h
changes for non Linux architectures
2012-07-05
* 3b82b03d src/Makefile.in src/Makefile.inc src/aix.c...: Make the PAPI locks
be tied to OS, not to CPU There is not a papi_lock.h file that when included
gets the proper lock include for the OS. This fixes a lot of previous build
hacks where a CPU component was needed in order for locks to work.
* 0632ef42 src/threads.c: Fix spurious init_thread call in threads.c
threads.c was calling init_thread() on all components, even ones that were
disabled Fix it to honor the disable bit, as well as for shutdown_thread().
This was causing perfctr disable code to not work.
* 19d9de7f src/Makefile.in src/Makefile.inc src/Rules.pfm4_pe...: Replace
SUBSTRATE with CPUCOMPONENT in build This was mostly a configure/build
change but it also cleaned up some cases where we were including SUBSTRATE
where we didn't have to.
* 829780db src/solaris-common.c src/solaris-common.h
src/solaris-niagara2.c...: Move some common solaris code to solaris-common
* 681ef027 src/configure src/configure.in src/solaris-memory.c...: Merge
solaris-memory.c and solaris-niagara2-memory.c
* bbd41743 src/solaris-ultra/get_tick.S src/solaris.h: Remove
solaris-ultra/get_tick.S Nothing was using it.
* dc3b6920 src/papi_sys_headers.h src/solaris.h: Remove papi_sys_headers.h
Solaris was the only thing including it, and it wasn't really using it.
* 7ccfa9df src/Makefile.in src/Makefile.inc src/configure...: Move move OS
specific code into the new OSFILESSRC Linux in particular was using MISC for
this.
* 6f16c0c5 src/configure: Re-run autoconf to pickup the substrate=>component
change.
* cfff1ede src/Makefile.in src/Makefile.inc src/configure...: Remove
MEMSUBSTR In reality, what we want instead of a Memory Substrate is an idea
of the OS-specific common code that includes the memory substrate. This
change adds OSFILESSRC and OSFILESOBJ to handle this case in configure
* ca4729e6 src/configure.in: Separate out MEMSUBSTR and make it per-OS
* 3148cba5 src/Matlab/PAPI_Matlab.dsp src/ctests/calibrate.c
src/ctests/flops.c...: RIP Windows, remove the windows support code. Windows
has not been activly supported since the transition to Component PAPI (4.0)
This cleans up the code-base.
2012-07-03
* a366adf7 src/papi.c src/utils/error_codes.c: Change PAPI_strerror and
PAPI_perror to behave more like thir POSIX namesakes. PAPI_error_descr is
made redundant and removed as a result.
2012-07-05
* 7df46f81 src/Rules.pfm src/aix.c
src/components/coretemp/linux-coretemp.c...: Move uses of PAPI_ESBSTR to
PAPI_ECMP I left PAPI_ESBSTR defined too for backward compatability. Also
some of the changes update PAPI_ESBSTR to be a more relevant error code, it
one is available.
2012-07-03
* fdb348ad src/components/coretemp_freebsd/coretemp_freebsd.c
src/components/example/example.c src/components/net/linux-net.c...: A few
more substrate removals
* 791747c1 src/cpus.c src/papi.h src/perf_events.c...: Fix bugs introduced by
substrate -> component change Fix some stupid compile bugs that I missed.
* 79b01a47 src/aix.c src/components/appio/appio.c
src/components/bgpm/CNKunit/linux-CNKunit.c...: More substrate -> component
changes This changes the vectors .init_substrate -> .init_component
.shutdown_substrate -> .shutdown_substrate .init -> .init_thread .shutdown ->
.shutdown_thread hopefully this will make the code clearer.
* 02a10d71 src/Makefile.inc src/aix.c src/cpus.c...: Rename "substrate" to
"component" This first pass only re-names things in comments.
2012-07-02
* c4bbff1c src/papi.c src/papi.h: Minor documentation fixes Found when
writing up the PAPI 5.0 changes document
2012-06-30
* f9cb7346 src/components/vmware/vmware.c: Fix vmware component apparently I
forgot to test the build with the vmguestlib support disabled.
2012-06-22
* 599040d1 src/components/coretemp/linux-coretemp.c
src/components/rapl/linux-rapl.c
src/components/stealtime/linux-stealtime.c...: Fix libpfm4 ntv_event_to_info
event_info_t on other components This was actually a widespead problem due
to cut-and-paste.
* 2b51b439 src/papi_libpfm4_events.c: Properly fix libpfm4 ntv_event_to_info
event_info_t event value The previous fix was subtly wrong. This is the
proper fix, which is to do nothing inside of papi_libpfm4_events.c because
papi_internal.c does the right thing for us and we were overwriting that with
the wrong value.
* a4f576bf src/ctests/overflow_allcounters.c src/testlib/papi_test.h
src/testlib/test_utils.c: Clean up overflow_allcounters code While tracking
down a previous issue I also cleaned up the overflow_allcounters test code to
use some of the new interfaces.
* 6903e053 src/papi_libpfm4_events.c: Fix libpfm4 ntv_event_to_info
event_info_t event value The recently added libpfm4 ntv_event_to_info
function was not properly oring PAPI_NATIVE_MASK to the event value in the
event_info_t struct. That means if you tried to use that event value to add
an event it would fail. The overflow_allcounters test broke because of this.
* 420c3d11 src/ctests/Makefile src/ctests/disable_component.c src/papi.c...:
Add PAPI_get_component_index() and PAPI_disable_component()
PAPI_get_component_index() will return a component index if you give it the
name of a component to match. This saves you having to iterate the entire
component list looking. PAPI_disable_component() will manually mark a
component as disabled. It has to be run before PAPI_library_init() is called.
* 11946525 src/aix.c src/components/cuda/linux-cuda.c
src/components/example/example.c...: Standardize component names to not end
in .c We were being inconsistent; the time to make them all be the same is
now before 5.0 gets out.
2012-06-21
* 274e1ad8 src/components/vmware/tests/Makefile: Fix cut-and-paste error in
the vmware component Makefile
* 85d6438d src/utils/event_chooser.c: Update papi_event_chooser to work with
components Now you can specify events from components and it will tell you
all the other events on that component that can run with it. Previously the
utility was limited to the CPU component (0) only.
* 3c2fcc83 src/papi_libpfm3_events.c src/papi_libpfm4_events.c
src/perf_events.c: Hook up .ntv_code_to_info on perf_event
* 36e864b3 src/papi_libpfm4_events.c src/papi_libpfm_events.h
src/perf_events.c: Enable support for showing extended umasks on perf_event
With this change, papi_native_avail now shows event umasks such as :u, :k,
:c, :e, and :i. (user, kernel, cmask, edge-trigger, invert) Thes are boolean
or integer events. They were supported by previous PAPI but they were never
enumerated.
* 8f3e305e src/components/coretemp/linux-coretemp.c: Fix cut-and-paste error
in linux-coretemp.c that could lead to wrong size being copied
* 0eedd562 src/libpfm4/lib/events/intel_atom_events.h
src/libpfm4/lib/events/intel_core_events.h
src/libpfm4/lib/events/intel_coreduo_events.h...: Import most recent libpfm4
git This fixes an issue where there can be confusion between :i and :i=1
type events. It also has initial support for Uncore, though you need a
specially patched kernel and PAPI does not support it yet.
* 2f86ec78 src/components/appio/tests/Makefile
src/components/appio/tests/appio_test_blocking.c
.../appio/tests/appio_test_fread_fwrite.c...: - Fixed tests verbosity by
using TESTS_QUIET macro - Fixed Makefile to only include necessary tests for
automatic builds (skip blocking tests that read from stdin)
* 6936b955 src/components/appio/README src/components/appio/appio.c
src/components/appio/appio.h...: Added polling of read/write descriptor to
check which operations would block.
* 48cacccf src/papi.h: Add back PAPI_COMPONENT_INDEX() for backward
compatability It turns out some people were using this for cases other than
enumeration. The proper way to do things now is to use
PAPI_get_event_component() which is what this PAPI_COMPONENT_INDEX() now maps
to.
* d1ed12b7 src/ctests/Makefile src/ctests/get_event_component.c
src/papi.c...: Add PAPI_get_event_component() This function returns the
component an event belongs to. It also adds a test to test this
functionality.
2012-06-20
* ffccf633 src/papi.h: Add component_type field to .cmp_info The idea is
we'll specify CPU, I/O, GPU, hardware, etc.
* 9998eecc src/components/lmsensors/Rules.lmsensors: Another lmsensors build
fix
* caa94d64 src/components/lmsensors/linux-lmsensors.c: Update lmsensors
component to actually compile. I finally found a machine with lmsensors
installed.
* fbcde325 src/components/lmsensors/linux-lmsensors.c
src/components/lmsensors/linux-lmsensors.h: Update lmsensor component Unlike
the other components it hadn't been updated to PAPI 5 standards. Also, it
was wrongly de-allocating all state in "_shutdown" rather than
"_shutdown_substrate" which was causing double-frees during tests.
* 0d3c0ae2 src/papi_internal.c: Add some extra debugging to
_papi_hwi_get_native_event_info
* 5961c03d src/aix.c src/components/nvml/linux-nvml.c
src/ctests/subinfo.c...: Remove cntr_groups from .cmp_info This information
is better exposed by enumeration.
* 2b4193fd src/utils/event_chooser.c: Cleanup and comment event_chooser code
* 7f9fab2b src/ctests/all_native_events.c: Cleanup and add comments to
all_native_events.c
* a245b502 src/components/nvml/linux-nvml.c src/ctests/subinfo.c
src/freebsd.c...: Remove profile_ear from .cmp_info The CPU components
should handle this internally.
* bca07f3c src/papi.c: Add comments to the PAPI_sprofil code.
* b1e2090c src/papi.c: Minor papi.c cleanups Fix some minor cosmetic things,
including a typo in a comment.
* 8f3aef4a src/ctests/subinfo.c src/papi.h: Remove opcode_match_width field
from .cmp_info This should be exposed via enumeration and not by a field in
the generic cmp_info structure.
* 047af629 src/components/nvml/linux-nvml.c src/ctests/subinfo.c
src/papi.h...: Remove cntr_OPCM_events field from .cmp_info This should be
exposed via enumeration and not by a field in the generic cmp_info structure.
* 3f1f9e10 src/components/nvml/linux-nvml.c src/ctests/subinfo.c
src/papi.h...: Remove cntr_DEAR_events field from .cmp_info This should be
exposed via enumeration and not by a field in the generic cmp_info structure.
* 962c642a src/components/nvml/linux-nvml.c src/ctests/subinfo.c
src/papi.h...: Remove cntr_IEAR_events field from .cmp_info This should be
exposed via enumeration and not by a field in the generic cmp_info structure.
* 5aa7eac1 src/components/nvml/linux-nvml.c src/ctests/subinfo.c
src/papi.h...: Remove instr_address_range from .cmp_info This feature should
be deteced via enumeration, not via a flag in the generic .cmp_info
structure.
* 1bf68d5d src/components/nvml/linux-nvml.c src/ctests/subinfo.c
src/papi.h...: Remove data_address_range field from .cmp_info The proper way
to detect this feature is via enumeration.
2012-06-19
* 90037307 src/linux-context.h: Change Linux from using "struct siginfo" to
"siginfo_t" This conforms to POSIX, and fixes newer Fedora where struct
siginfo is no longer supported. This can in theory break on older setups
(possibly kernel 2.4). If that happens, we need to somehow detect this using
autoconf.
2012-06-18
* ad48b4fa src/Rules.perfctr-pfm: Fix the perfctr-pfm build; for buildbot,
mostly. Have the perfctr-pfm build only build libpfm, like the perfevents
builds. The icc build was choking on warnings (-Werror => errors) in the
example programs with libpfm, this is not something we depend upon.
2012-06-17
* 358b14f9 src/papi_events.csv: Update BGQ presets
* cf26fc87 src/components/bgpm/CNKunit/linux-CNKunit.c
src/components/bgpm/IOunit/linux-IOunit.c
src/components/bgpm/L2unit/linux-L2unit.c...: Update bgpm components
according to the papi5 changes
* a7b08a91 src/configure src/configure.in src/linux-bgq.c: Merging the BG/Q
stuff from stable_4.2 to PAPI 5 did break it. It's corrected now; also
predefined events are now working.)
2012-06-15
* 2d5a4205 src/components/bgpm/IOunit/linux-IOunit.c
src/components/bgpm/L2unit/linux-L2unit.c src/configure...: Merging the BG/Q
stuff from stable_4.2 to PAPI 5 did break it. It's corrected now (almost);
predefined events are not working yet.)
* 1b034920 src/papi.c: Re-enable PAPI_event_name_to_code() optimization In
PAPI_event_name_to_code() there was a commented-out optimization where we
would check if an event name begins with "PAPI_" before searching the entire
preset list for an event name. The comment says we had to disable this due
to "user events", but a check shows that was introduced in e7bd768850ecf90
and that the "user events" it means is not the current support, but the
now-removed PAPI_set_event_info() function where you could change the names
of presets on the fly (even to something not starting with PAPI_). Since we
don't support that anymore, we can re-enable the optimization.
2012-06-14
* 9a26b43d src/papi_internal.c src/papi_internal.h src/papi_preset.c: Remove
the 16-component limit This turned out to be easier than I thought it would
be. Now determining which component an event is in is a two step process.
Before, the code shifted and masked to find the component from bits 26-30.
Now, _papi_hwi_component_index() is used. There's a new native event table
which maps all native events (which are allocated incrementally at first use
starting with 0x4000000) to two values, a component number and an "internal"
event number.
2012-06-13
* d5c50353 src/papi_internal.c: Fix for the PAPI_COMPONENT_MASK change I
missed two cases in papi_internal.c This was causing the overflow_allcounters
test to fail
* 46fd84ce src/components/bgpm/CNKunit/linux-CNKunit.c
src/components/bgpm/CNKunit/linux-CNKunit.h
src/components/bgpm/IOunit/linux-IOunit.c...: Updating the Q substrate
according to the PAPI 5 changes
* 05a8dcbf src/components/appio/appio.c
src/components/appio/tests/appio_list_events.c
src/components/appio/tests/appio_values_by_code.c...: First steps of removing
16-component limit This change removes PAPI_COMPONENT_INDEX(),
PAPI_COMPONENT_MASK and PAPI_COMPONENT_AND_MASK. It adds the new functions
_papi_hwi_component_index() _papi_hwi_native_to_eventcode()
_papi_hwi_eventcode_to_native() By replacing all of the former macros with
the equivelant of the latter functions, it allows all of the future
16-component limit changes to be made in the functions. Components now
receive as events a plain 32-bit value as their internal native event; the
high bits are not set. This may break some external components. This change
should not break things, but a lot of testing is needed.
* af4cbb86 src/run_tests_exclude.txt: Exclude iozone helper scripts from
run_tests. run_tests.sh looks for executible files under components/*/tests
Some of the plotting scripts in appio/iozone were getting picked up.
2012-06-12
* c10c7ccb src/configure src/configure.in: Configure does not work on BGQ due
to missing subcomp feature. It worked for stable-4.2 but got lost in current
git origin.
* d9a58148 src/aix.c src/ctests/hwinfo.c src/ctests/overflow.c...: Update
hw_info_t CPU frequency reporting. Previously PAPI reported "float mhz" and
"int clock_mhz". In theory the first was the current CPU speed, and the
latter was the resolution of the cycle counter. In practice they were both
set to the same value (on Linux, read from /proc/cpuinfo) and not very useful
when DVFS was enabled, as the value reported was usually lower than the
actual frequency running once CPU started being used. This change adds two
new values "cpu_max_mhz" and "cpu_min_mhz" which are read from
/sys/devices/system/cpu/cpu0/cpufreq/cpuinfo_max_freq and
/sys/devices/system/cpu/cpu0/cpufreq/cpuinfo_min_freq if they are available,
and falls back to /proc/cpuinfo otherwise. All of the tests were updated to
use cpu_max_mhz. The old mhz and clock_mhz values are left for compatability
reasons (and set to cpu_max_mhz) but are currently unused otherwise.
2012-06-11
* 0f124891 src/papi_events.csv: Initial PAPI Ivy Bridge Support for now try
to re-use the sandy bridge event presets
* a1f46077 src/libpfm4/docs/man3/libpfm_intel_ivb.3
src/libpfm4/include/perfmon/err.h
src/libpfm4/lib/events/intel_ivb_events.h...: Import libpfm4 git snapshot
This adds IvyBridge support
* 3bb983cc src/libpfm-3.y/examples_v2.x/self_smpl_multi.c: Fix a libpfm3
example program for icc, local fix because libpfm3 is deprecated. icc does
have more enjoyable warnings than gcc, : error 186: pointless comparison of
unsigned integer with zero on this: unsigned int foo; ... if ( foo < 0 )
2012-06-06
* d28adccf src/papi_user_events.c: The user events code had a call to exit,
this was bad...
2012-06-04
* 6bf43022 src/testlib/test_utils.c: Further the hack for testing for perf SW
events. Events like
- | perf::CPU-CLOCK
| | PERF_COUNT_SW_CPU_CLOCK
|
- were passing the check, now we also check the event_info_struct.long_descr
field for PERF_COUNT_SW....
* fa4b1a28 src/components/nvml/linux-nvml.c: Cleanup nvml code a little. A
few print statements were left over from debugging. Also check errors from
nvml and cuda pciinfo functions, disabling the component in a few more cases.
2012-06-01
* da144a94 src/components/nvml/Makefile.nvml.in src/components/nvml/README
src/components/nvml/Rules.nvml...: Rewrite and merge of the nVidia Management
library component. This component attempts to expose all supported
'performance counters' on each card cuda knows about at runtime. Much like
the cuda component reads happen on the card you're executing on at PAPI_read
time. The test included is a copy of the cuda helloworld test, but it
attempts to start/stop the event on each gpgpu. If you select an event that
is not supported on the card you're running on we should fail gracefully but
this has not been tested.
2012-05-23
* b2d414dc src/components/stealtime/linux-stealtime.c: At units to stealtime
component Added the function but forgot to add a function vector for it.
* ce9d4500 src/components/stealtime/linux-stealtime.c: Add units to stealtime
Properly report that the units are in micro seconds.
* 149948c8 src/components/rapl/linux-rapl.c: Minor cleanup of RAPL code
missing "void" paramter in init_substrate function
* 6a7e22fa src/components/vmware/vmware.c: More vmware component fixes. This
makes the component thread-safe. Also makes it fail more gracefully if the
guestlib SDK is installed but does not support our hypervisor (for example,
if we are running under VM Workstation). Still need to test on ESX.
* 072d6473 src/components/appio/tests/appio_test_select.c: added code to
intercept and time select() calls.
2012-05-22
* 12b6d0d7 src/components/vmware/vmware.c: Some more minor fixes to VMware
component Try to handle things better if VMguest SDK not working
* 6e015bc5 src/components/vmware/Rules.vmware src/components/vmware/vmware.c:
More vmware component fixups Now works with the events from the VMguest SDK
library
* 5fc0f646 src/components/vmware/vmware.c: More cleanup of vmware component
The pseudo-performance counters work again. Now they behave in accumulate
mode, like all other PAPI counters.
* f72b0967 src/components/vmware/tests/vmware_basic.c: Make vmware test a bit
more complete
* 070e5481 src/components/vmware/tests/Makefile
src/components/vmware/tests/vmware_basic.c: Add a test for the vmware
component
* 7cf62498 src/components/vmware/Makefile.vmware.in
src/components/vmware/Rules.vmware src/components/vmware/configure...: Clean
up the vmware component. bring it up to date with other components. make it
possible to build it without the vmguest library being installed
* b32ae1ae src/components/stealtime/Rules.stealtime
src/components/stealtime/linux-stealtime.c
src/components/stealtime/tests/Makefile...: Add a stealtime component When
running in a VM, this provides information on how much time was "stolen" by
the hypervisor due to the VM being disabled. This info is gathered from
column 8 of /proc/stat This currently only works on KVM.
* 9e95b480 src/components/appio/tests/appio_test_blocking.c: Use a
non-blocking select to determine which reads and writes would block
2012-05-19
* f60d991f src/components/appio/README src/components/appio/appio.c
src/components/appio/tests/appio_test_read_write.c...: Interception of
close() implemented. This allows us to correctly determine the number of
currently open descriptors.
2012-05-17
* 7cd8b5a3 src/libpfm4/.gitignore src/libpfm4/config.mk
src/libpfm4/lib/Makefile...: Update libpfm4 to current git tree
* ebffdb7e src/components/rapl/tests/rapl_overflow.c: Skip rapl_overflow test
if RAPL not available
* 98d21ef3 src/components/example/example.c src/components/rapl/linux-rapl.c:
Fix some component warnings.
* 0447f373 src/configure src/configure.in src/linux-generic.h: Make build not
stall if PAPI_EVENTS_CSV not set This is some fallout from the FreeBSD
changes. PAPI_EVENTS_CSV could not be set, which would make the event
creation script hang forever. Also catch various fallthroughs in the code
where SUBSTR wasn't being set, which is how the above problem can happen.
* ef484c00 src/linux-timer.h: Fix typo in linux-timer.h
2012-04-14
* 7c3385f4 src/components/bgpm/CNKunit/CVS/Entries
src/components/bgpm/CNKunit/CVS/Repository
src/components/bgpm/CNKunit/CVS/Root...: Removed CVS stuff from Q code.
* 2cf4aeb2 src/configure src/configure.in src/linux-bgq.c...: Removed
papi_events.csv parsing from Q code. (CVS stuff still needs to be taken care
of.)
2012-04-12
* 153c2bb1 INSTALL.txt: Updated INSTALL notes for Q
2012-05-17
* ff6a43fb src/Makefile.in src/Makefile.inc src/components/README...: Added
missing files for Q merge. Conflicts: src/configure src/configure.in
src/freq.c
2012-04-12
* 0e142630 src/Rules.bgpm src/components/bgpm/CNKunit/CVS/Entries
src/components/bgpm/CNKunit/CVS/Repository...: Added PAPI support for Blue
Gene/Q.
2012-05-14
* ad7e3fa0 src/components/rapl/linux-rapl.c: Properly accumulate RAPL results
Previously it was resetting the counts on read, instead of continuing to
count as per other PAPI events.
* c79e3018 src/components/rapl/tests/rapl_overflow.c: Fix some warnings in
rapl_overflow test
* 731afd1a src/components/rapl/tests/Makefile
src/components/rapl/tests/rapl_overflow.c: Add rapl_overflow test This test
sees if we can measure RAPL in conjunction with overflow CPU performance
events.
* b0e201bb src/components/rapl/utils/Makefile
src/components/rapl/utils/rapl_plot.c: Remove derived "uncore" values from
rapl tool They weren't really measuring uncore, but were just TOTAL - PP0.
It was causing some confusion.
2012-05-09
* 547e9379 doc/Doxyfile-common papi.spec src/Makefile.in...: Bump the version
number to 4.9.0.0 Read 4.9 as pre-5.0 master was at version number 4.2.1,
this was archaic... Sorry for the confusion Tushar, master is the correct
branch for the latest development code.
* 133e3d67 src/configure src/configure.in: Fix perfctr build In the FreeBSD
changes I removed the CPU determination by reading /proc/cpuinfo as that was
prone to failure and non-portable. This broke perfctr as it was doing a huge
CPU name lookup to determine if it was on an x86 system or not. This change
fixes that.
2012-05-08
* 42b21d67 src/papi_libpfm4_events.c: Fix PAPI event enumeration inside of
VMware VMware disables the availability of architectural events when
virtualized PMU is not available. libpfm4 was checkign this when enumerating
events, and we would end up in the situation where ix86arch was marked active
but 0 events were available. We didn't check for this error condition and
thus end up thoroughly confused.
2012-05-07
* fd79a584 src/freebsd.c: Fix event enumeration on FreeBSD It was passing
PAPI_OK in all cases, causing papi_native_avail to try to do things like
report groups even when groups weren't available.
* 53732c2e src/freebsd.c: Add Virtual Machine detection support to FreeBSD
again, support for this on x86 is OS Neutral
* 7b4d7c96 src/configure src/configure.in src/freebsd-memory.c...: Add
x86_cacheinfo support to FreeBSD The x86 cache and memory info is
OS-independent, so add support for it to FreeBSD.
* 91033df6 src/Makefile.in src/Makefile.inc src/configure...: Re-enable
predefined events on FreeBSD
* 36f6dc1b src/freebsd.c src/freebsd/map.c src/freebsd/map.h: Modify FreeBSD
to use _papi_load_preset_table
* 45651746 src/freebsd.c: Cleanup the freebsd code a bit.
* e1554ed8 src/configure: re-run autoconf for updated configure
* 1deb2f5d src/Makefile.inc: Make sure a proper dependency for
papi_events_table.h exists Our Makefile code that builds a shared library is
way broken; it will fail to rebuild in many cases where the static library
properly detects thing.
* 28e28006 src/configure.in: Make papi_events_table.h build normally, not by
configure.
* 9a66dfa5 src/configure.in: Another place papi_events_table.sh is called
* 12e4a934 src/Makefile.inc src/papi_events_table.sh: Make
papi_events_table.sh take a command line argument This way we can use it on
any .csv file, not just papi_events.csv
* 7018528f src/freebsd/memory.c: Remove unused freebsd/memory.c file
* 819e5826 src/freebsd_events.csv: Make freebsd_events.csv a valid PAPI event
file
* 9cc4a468 src/freebsd.c src/freebsd/map-atom.c src/freebsd/map-core.c...:
Fix FreeBSD build on head. This temporarily disables preset events. There
are also a few other minor fixes.
2012-05-01
* ab36c0a2 src/Makefile.inc src/configure src/configure.in: Update build
system for FreeBSD
* 2b61d8b7 src/freebsd.c src/freebsd.h: Fix various compiler warnings on
FreeBSD
* 2c0bcc84 src/freebsd.c: Enable new Westmere events on FreeBSD
* b0499663 src/freebsd/map-i7.c src/freebsd/map-i7.h
src/freebsd/map-westmere.c...: Add Westmere event support for FreeBSD
* e54cabc6 src/ctests/inherit.c: Fix the inherit ctest to compile on FreeBSD
* d9dbdd31 src/components/appio/appio.c: - change in appio component
(appio.c): removed reference to .ntv_bits_to_info as it doesn't exist in the
PAPI component interface.
2012-04-27
* 5d661b2d src/Rules.pfm src/Rules.pfm_pe: Add the libpfm -Wno-override-init
bandaid to the other rules files. In
b33331b66137668155c02e52c98a7e389fad402e we test if gcc -Wextra complains
about some structure initialization that libpfm does. This was incoperated
into Rules.pfm4_pe only. Jim Galarowicz noticed the other Rules files didn't
have it.
* 4349b6fd src/Rules.pfm4_pe src/Rules.pfm_pe: Cleanup the perf events Rules
files. Steve Kaufmann reported that CONFIG_PFMLIB_OLD_PFMV2 is only used for
libpfm3 builds targeting old versions of perfmon2.
2012-04-26
* 8a7fef68 src/mb.h: Add memory barries for ia64
2012-04-24
* 9af4dd4a src/libpfm4/README src/libpfm4/config.mk
src/libpfm4/include/perfmon/perf_event.h...: Import libpfm4 git snapshot
This brings libpfm4 up to 9ffc45e048661a29c2a8ba4bfede78d3feb828f4 The
important change is support for Intel Atom Cedarview.
2012-04-20
* fac6aec0 src/linux-bgp-memory.c src/linux-bgp.c: Some BG/P cleanups.
Removed a lot of dead code, noticed when looking for any potential BG/P
issues.
* 977709f6 src/linux-bgp-preset-events.c src/linux-bgp.c: Fix PAPI compile on
BG/P Thanks to Harald Servat <harald.servat@bsc.es>
2012-04-19
* 5207799e release_procedure.txt: Modified release_procedure.txt to push
tags.
2012-04-18
* b248ae80 doc/Makefile: Have clean remove the doxygen error file.
* 1d4f75a3 doc/Doxyfile-man1 doc/Doxyfile-man3: Fix an error in the Doxygen
config files. Doxygen includes things with @INCLUDE not @include. The html
file had this, the man page files did not...
2012-04-17
* 979cda20 cvs2cl.pl delete_before_release.sh gitlog2changelog.py...: Update
the release machinery for git. gitlog2changelog.py takes the output of git
log and parses it to something like a changelog.
* 67bdd45f doc/Doxyfile-html: Cover up an instance of doxygen using full
paths. Doxygen ( up to 1.8.0, the most recent at this writing ) would use
full paths in directory dependencies ignoring the use relative paths config
option.
2012-04-13
* c38eb0b7 src/libpfm-3.y/lib/intel_corei7_events.h
src/libpfm-3.y/lib/intel_wsm_events.h src/libpfm-3.y/lib/pfmlib_intel_nhm.c:
Add missing update to libpfm3 Somehow during all of the troubles we had with
importing libpfm3 into CVS, we lost some Nehalem/Westmere updates. Tested on
a Nehalem machine to make sure this doesn't break anything.
* 193d8d06 src/papi_libpfm3_events.c: Fix max_multiplex case on
perf_event/libpfm3 num_mpx_cntrs was being set to 512 even though the real
maximum is 32, causing a buffer overflow and segfault.
2012-04-12
* f1f7fb5b src/threads.h: Fix minor typo in a comment
* 0373957d src/linux-timer.c: Fix potential fd leak Noticed by coverity
checker.
* 71727e38 src/ctests/max_multiplex.c: Improve max_multiplex ctest on
perfmon2, this test was failing because the maximum number of multiplexed
counters was much more than the available counters we could test with. This
change modifies the test to not fail in this case.
2012-04-11
* fdbdac9f src/perfmon.c: Fix the perfmon substrate. It was missing a
_papi_libpfm_init() call, which meant the number of events was being left at
0.
2012-04-09
* 2a44df97 src/libpfm-3.y/examples_v2.x/multiplex.c
src/libpfm-3.y/examples_v2.x/pfmsetup.c
src/libpfm-3.y/examples_v2.x/rtop.c...: Catch a few libpfm-3.y files up to
libpfm-3.10. More skeletons keep falling out of the cvs closet. This is just
what diff -q -r catches.
2012-04-04
* 0e05da68 src/components/rapl/utils/Makefile
src/components/rapl/utils/README src/components/rapl/utils/rapl_plot.c: Add
the rapl_plot utility to the RAPL component. This utility uses PAPI to
periodicly poll the RAPL counters and generate average power results suitable
for plotting. There's been a lot of interest in this utility so it's
probably useful to include it with the RAPL component.
* 2daa03ac src/papi_internal.c: Check if a component is disabled at init
time. This change modifies the code so that at PAPI_library_init() time we
check the component disable field, and we don't call the init routines for
components the user has disabled. This allows code like the following to
happen _before_ PAPI_library_init(): numcmp = PAPI_num_components();
for(cid=0; cid<numcmp; cid++) { cmpinfo = PAPI_get_component_info(cid); if
(strstr(cmpinfo->name,"cuda")) { cmpinfo->disabled=1;
strncpy(cmpinfo->disabled_reason,"Disabled by user",PAPI_MAX_STR_LEN); } }
We might want to add a specific PAPI_disable_component(int cid) call of maybe
even a PAPI_disable_component(char *name) as the above code causes compiler
warnings since cmpinfo is returned as a const pointer. This all works
because currently PAPI currently statically allocates all of the components
at compile time, so we can view and modify the cmp_info structure before
PAPI_library_init() is called.
* 3fd2b21e src/components/appio/README src/components/appio/appio.c
src/components/appio/appio.h...: Added support to count reads that are
interrupted or would block.
2012-04-03
* dd3a192f release_procedure.txt: Change chmod flags for doxygen stuff from
755 to 775 to allow group write permissions.
2012-03-30
* deac54cc src/components/coretemp/linux-coretemp.c
src/components/coretemp/tests/coretemp_basic.c
src/components/coretemp/tests/coretemp_pretty.c...: Add new
PAPI_enum_cmp_event() function This will be needed when we remove the
16-component limit. Currently in PAPI_enum_event() the component number is
gathered from bits 29-26 of the eventcode. This won't work anymore once we
remove those bits. Also update the various components to not use
PAPI_COMPONENT_MASK() as this too will go away in the transition.
* 48331cc9 src/configure src/configure.in src/papi.c...: Place all
compiled-in components in the _papi_hwd[] array. Previously we had separate
compiled_in[] and _papi_hwd[] arrays. At init time a pointer to the
compiled_in[] was copied to _papi_hwd[] if initialization passed. This kind
of code setup makes enumerating components hard, and finding info from
non-available components would require additional function entry points.
This change leaves all compiled in components to _papi_hwd[]. Availability of
the component can be checked with the new "disabled" field. This will make
enumeration support a lot easier to add. It can possibly cause user confusion
if they try to access component structures directly without checking the
"disabled" field first. This change should also make any eventual support
for run-time component enabling/disabling a lot easier.
* 66a72f44 src/papi.c: Documentation was referring to nonexistent
"PAPI_enum_events()" The actual function we have is PAPI_enum_event()
* 0f2c2593 src/components/coretemp/linux-coretemp.c
src/components/lustre/linux-lustre.c src/components/mx/linux-mx.c...: Add
support for reporting reason for failed component initialization. This
change adds the fields "disabled" and "disabled_reason" to the
component_info_t structure. At initialization time, PAPI will set the
"disabled" field to the value returned by component init (that is PAPI_OK if
OK or an error otherwise). This can be checked later to find why component
init failed. Also provided is the "disabled_reason" string. The components
can set this at failure time, and this can be printed later. For example,
this is sample output of the updated papi_component_avail routine:
- Compiled-in components: Name: perf_events.c Linux perf_event
CPU counters Name: linux-rapl Linux SandyBridge RAPL energy
measurements \-> Disabled: Not a SandyBridge processor Name: example.c
A simple example component Name: linux-coretemp Linux
hwmon temperature and other info \-> Disabled: No coretemp events found Name:
linux-net.c Linux network driver statistics Name: linux-mx.c
Myricom MX (Myrinet Express) statistics \-> Disabled: No MX
utilities found Name: linux-lustre.c Lustre filesystem statistics
\-> Disabled: No lustre filesystems found Active components: Name:
perf_events.c Linux perf_event CPU counters Name: example.c
A simple example component Name: linux-net.c Linux
network driver statistics
2012-03-29
* d84b144e src/components/rapl/Rules.rapl src/components/rapl/linux-rapl.c
src/components/rapl/tests/Makefile...: Add a SandyBridge RAPL (Running
Average Power Level) Component This component allows energy measurement at
the package-level on Sandybridge machines. To run, you need the Linux
x86-msr kernel module installed and read permissions to /dev/cpu/*/msr The
output from the rapl_busy test looks like this on a SandyBridge-EP machine:
Trying all RAPL events Found rapl component at cid 2 Starting
measurements... Doing a naive 1024x1024 MMM... Matrix multiply sum:
s=1016404871450364.375000 Stopping measurements, took 3.979s, gathering
results... Energy measurements: PACKAGE_ENERGY:PACKAGE0 175.786011J (Average
Power 44.2W) PACKAGE_ENERGY:PACKAGE1 73.451096J (Average Power 18.5W)
DRAM_ENERGY:PACKAGE0 11.663467J (Average Power 2.9W) DRAM_ENERGY:PACKAGE1
8.055389J (Average Power 2.0W) PP0_ENERGY:PACKAGE0 119.215500J (Average
Power 30.0W) PP0_ENERGY:PACKAGE1 16.315216J (Average Power 4.1W) Fixed
values: THERMAL_SPEC:PACKAGE0 135.000W THERMAL_SPEC:PACKAGE1 135.000W
MINIMUM_POWER:PACKAGE0 51.000W MINIMUM_POWER:PACKAGE1 51.000W
MAXIMUM_POWER:PACKAGE0 215.000W MAXIMUM_POWER:PACKAGE1 215.000W
MAXIMUM_TIME_WINDOW:PACKAGE0 0.046s MAXIMUM_TIME_WINDOW:PACKAGE1 0.046s
rapl_basic.c PASSED
2012-03-26
* b44d60ca src/components/appio/appio.c src/components/appio/appio.h
src/components/appio/tests/appio_test_read_write.c: Added support for
intercepting open calls.
2012-03-23
* 9e9fac4b src/Makefile.in src/Rules.pfm4_pe src/configure...: Fix the test
case in configure at 0cea1848 Make use of the structure we're using for the
override-init test case.
* 0cea1848 src/configure src/configure.in: Doctor CFLAGS when testing for a
gcc warning. -Wextra was not in CFLAGS when I attempted to check for the
initialized field overwritten warning. So we set -Wall -Wextra -Werror when
running the test code.
2012-03-22
* b33331b6 src/Makefile.in src/Rules.pfm4_pe src/configure...: Fix
initialized field overwritten warning when building libpfm4 on some gcc
versions. In gcc 4.2 or so, -Woverride-init was added to -Wextra causing
issues with code like struct foo { int a; int b;}; struct foo bar = { .a=0,
.b=0, .b=5; }; --Wno-override-init allows us to keep -Werror for libpfm4
compiles.
2012-03-21
* ae149766 src/papi_internal.h: Delete an old comment. Yes, Dan in 2003, we
should and do use MAX_COUNTER_TERMS as the size of the event position array.
2012-03-20
* b937cdd8 src/papi_user_events.c: Move the user events code over to using
the new preset event data structure.
2012-03-14
* 6ca599e2 src/papi_internal.c: Fix a small memory leak. We weren't freeing
_papi_hwd, causing a lot of MEM_LEAK warnings in buildbot.
2012-03-13
* 473b8203 src/aix.h src/configure src/configure.in...: Remove last MY_VECTOR
usage. Have configure explicitly set the name of the perf counter substrate
vector in the components_config.h file This removes one more special case,
and gets us slightly closer to being able to have multiple CPU substrates
compiled in at once.
* 360c3003 src/papi.c src/papi_libpfm3_events.c src/papi_libpfm_events.h...:
Clean up the papi_libpfm3_events.c code. Move code that was perfctr specific
into perfctr-x86.c
* 03de65e3 src/libpfm-3.y/examples_v2.x/multiplex.c
src/libpfm-3.y/examples_v2.x/pfmsetup.c
src/libpfm-3.y/examples_v2.x/rtop.c...: Fix some libpfm3 warnings. libpfm3
is not maintained anymore, so applied these changes locally. libpfm3 is
compiled with -Werror so they broke the build with newer gcc even though they
are just warnings in example programs.
* ad490353 src/ctests/zero_named.c src/utils/multiplex_cost.c: Fix a few
compiler warnings in the tests.
* a0fec783 src/linux-timer.c: Fix another linux-timer.c compile problem. I
hadn't tested with debug enabled, so all of buildbot failed last night.
2012-03-12
* a3733ecd src/linux-timer.h: Fix typo in the linux-timer.h header
_linux_get_virt_usec_timess should have been _linux_get_virt_usec_times
Thanks to Steve Kaufmann for noticing this.
* 785db5ae src/linux-common.c src/linux-timer.c: Fix timer compile on Power
machines Power, ARM, and MIPS have no get_cycles() call so provide a dummy
function on these architectures.
* 708090ee src/linux-common.c src/linux-timer.h: Another fix for non-POSIX
timers The recent changes had the name of the fallback usec method wrong.
* 88e8d355 src/papi_libpfm3_events.c: Fix a warning in the libpfm3 code.
* 8ca63705 src/configure src/configure.in src/linux-common.c...: Fix build
when not using POSIX timers The PAPI build system was being overly clever
with how it defined what kind of wall clock timers were to be used, so of
course I broke things when breaking the timer code out to make it a bit more
understandable. This patch breaks out the timer define into two pieces; one
saying it's a POSIX timer and one saying whether to use HR timers or not.
2012-03-09
* b69ad727 src/linux-common.c src/linux-timer.c src/linux-timer.h: Add Linux
posix gettime() nanosecond functions
* af2c9a49 src/papi.c src/papi_vector.c src/papi_vector.h: Add
->get_virt_nsec() and ->get_real_nsec() OS vectors Currently PAPI was just
cheating and running the usec functions and multiplying by 1000. Make this
the default, but allow the OS code to override if they have timers capable of
returning nsec percision.
* 24c68dbe src/aix.c src/freebsd.c src/linux-bgp.c...: Clean up
->get_virt_usec() It no longer needs to be passed a context, so remove that
from all callers. Also, ->get_virt_cycles() was just a get_virt_usec()*MHz
on most platforms. While this is a bit dubious (especially as MHz can't be
relied on) make this a common routine that will be added at innoculate time
if ->get_virt_cycles() is set to NULL.
* a3ef7cef src/linux-common.c src/linux-timer.c src/linux-timer.h: Cleanup
the Linux timer code. Split things up a bit to make the code more readable.
* 50ce8ea0 src/papi_internal.c: Change a strcpy() to strncpy() just to be a
bit safer.
* 0526b125 src/components/lmsensors/linux-lmsensors.c: Fix buffer overrun in
lmsensors component
* b088db70 src/libpfm4/config.mk
src/libpfm4/docs/man3/pfm_get_os_event_encoding.3
src/libpfm4/examples/showevtinfo.c...: Update to current git libpfm4 snapshot
* ccb45f61 src/aix.c src/extras.c: Fix segfault on AIX During some of the
cleanups, the extras.h header was not added to aix.c This made some of the
functions (silently) use default data types for the function parameters,
leading to segfaults in some of the tests.
2012-03-08
* 1cb22d0b src/components/coretemp/linux-coretemp.c src/utils/native_avail.c:
Make "native_avail -d" report units if available Add units support to the
coretemp component, have native_avail -d (detailed mode) print it to make
sure it works.
* 9c54840e src/extras.c src/extras.h src/papi_internal.c...: Add new
ntv_code_to_info vector This will allow components to return the extended
event_info data for native events. If a component doesn't implement
ntv_code_to_info then get_event_info falls back to the old way of just
reporting symbol name and long description.
* c4579559 src/papi.h: Add new event_info fields New fields are added to
event_info that allow passing on extended information. This includes things
such as measurement units, data type, location, timescope, etc.
* 17533e4e src/ctests/all_events.c src/ctests/derived.c
src/ctests/kufrin.c...: Restore fields to event_info structure The changes
made were probably too ambitious, even for a 5.0 release. In the end it
looks like we can remain API compatible while just using up a little more
memory. We can still save space by shrinking preset_t behind the scenes.
* 6f13a5f6 src/aix.c src/components/coretemp/linux-coretemp.c
src/components/coretemp_freebsd/coretemp_freebsd.c...: Remove
->ntv_bits_to_info vector from component interface We weren't using it
anymore, and many of the components were just setting it to NULL
unncessarily. We'll be replacing the functionality soon with
ntv_code_to_info
* 401f37bc src/components/example/example.c src/ctests/subinfo.c src/papi.h:
Remove invert and edge_detect fields from component info These fields were
there to indicate if a CPU component supported these attributes (for Intel
processors) but in the end we never used these. The proper way to export
this info is during event enumeration.
* f32fe481 src/papi_events.csv: We had the PAPI_VEC_INS preset wrong on amd
fam12h llano
* 38a8d8a7 src/ctests/multiplex2.c src/papi_preset.c: Fix preset adding code
to be more robust. If an invalid event is in a preset definition, we'd
currently add it with an eventcode of 0 to the preset, which would break if
you tried to use the event. This change properly prints a warning in this
case, and sets the preset to be unavailable.
* 2591a546 src/ctests/val_omp.c src/ctests/zero_omp.c: Remove the hw_info
field from add_two_events calls. Two ctests missed the bus when Vince
reworked the add_two_events call.
* 358a2e32 src/papi_internal.c src/papi_preset.c: Fix segfault seen on an AMD
fusion machine With the recent preset and component hanges, we were not
properly resetting papi_num_components if PAPI_library_init()/PAPI_shutdown()
was called multiple times.
2012-03-07
* 7751f5d8 src/ftests/zeronamed.F: Fix a compile error on aix. Dan ran over
72 characters on a single line. xlf actually enforced that part the Fortran
spec.
2012-03-06
* 1c87d89c src/ftests/Makefile src/ftests/zeronamed.F src/papi_fwrappers.c:
Add support for {add, remove, query}_named to Fortran interface; add zero
named.F test case; modify ftests Makefile to support "all" tag.
* 71bd4fdd src/configure src/configure.in: Modify configure to define the
default FTEST_TARGETS as "all"
* 54e39855 src/components/vmware/vmware.c: Changed tri8ggering environment
variable to PAPI_VMWARE_PSEUDOPERFORMANCE per Vince's earlier email. This
should complete all the VMware component changes.
2012-03-05
* 845503fb src/Makefile.inc: Add missing MISCSRCS line to Makefile.inc This
was breaking the shared library build
2012-02-01
* 11be8e4b .../appio/tests/appio_test_fread_fwrite.c
src/components/appio/tests/appio_test_pthreads.c
src/components/appio/tests/appio_test_read_write.c: updated these tests to
print timing information
* 9ad62ab1 src/components/appio/README src/components/appio/appio.c
src/components/appio/appio.h...: Added support for timing I/O calls. Updated
tests and README.
2012-01-31
* beaa5ff0 src/components/appio/tests/iozone/Changes.txt
src/components/appio/tests/iozone/Generate_Graphs
src/components/appio/tests/iozone/Gnuplot.txt...: added the latest stable
iozone to the appio tests.
* 4af58174 src/components/appio/README src/components/appio/tests/Makefile
src/components/appio/tests/init_fini.c: added a hook to run the appio test
for iozone.
2012-01-21
* 15c733cf src/components/appio/CHANGES src/components/appio/README
src/components/appio/appio.c...: Removed stray 'net' references. All
remaining references are only for the purpose of giving credit. Updated
change log.
2012-01-20
* ca4b6785 src/components/appio/README src/components/appio/appio.c
src/components/appio/tests/appio_list_events.c...: - general cleanup -
improved tests to be quiet and be conform to other PAPI tests - replaced
hardwire constants in appio.c with symbolic ones - tests will now write to
/dev/null to avoid filling the terminal screen with useless text - more
comments added - @author added to files - updated README
2012-01-18
* bb22ed9f src/components/appio/README src/components/appio/Rules.appio
src/components/appio/appio.c...: - Added support to measure
bytes/calls/eof/short calls for read/write calls. - Interception of
read/write and fread/fwrite calls. - Works for static and dynamic linkage
(without need for LD_PRELOAD) - Tested OK on 32-bit i686 Linux 2.6.38.
Tushar
2011-12-03
* d58b34b6 src/components/appio/tests/Makefile
src/components/appio/tests/appio_list_events.c
src/components/appio/tests/appio_values_by_code.c...: *** empty log message
***
* cd7d7acc src/components/appio/tests/appio_values_by_name.c: file
appio_values_by_name.c was added on branch appio on 2011-12-03 05:22:06 +0000
* 425e4d09 src/components/appio/tests/appio_values_by_code.c: file
appio_values_by_code.c was added on branch appio on 2011-12-03 05:22:06 +0000
* 596ad9bb src/components/appio/tests/appio_list_events.c: file
appio_list_events.c was added on branch appio on 2011-12-03 05:22:06 +0000
* 119543dc src/components/appio/tests/Makefile: file Makefile was added on
branch appio on 2011-12-03 05:22:06 +0000
2012-03-05
* ba748a41 src/components/vmware/configure: Remove old configuration
parameters from vmware/configure
2012-03-02
* 2b7e2abb src/ctests/Makefile src/ctests/max_multiplex.c: Add a new
max_multiplex test This tries to use the maximum number of multiplexed
events. This was written in response to the 32/64 perf_event multiplexed
event limit reported by Mohammad j. Ranji
* a0985ff5 src/multiplex.c src/papi_internal.c src/papi_libpfm4_events.c...:
Fix issue when using more than 32 multiplexed events on perf_event On
perf_event we were setting num_mpx_cntrs to 64. This broke, as the
MPX_EventSet struct only allocates room for PAPI_MPX_DEF_DEG events, which is
32. This patch makes perf_event use a value of 32 for num_mpx_cntrs,
especially as 64 was arbitrarily chosen at some point (the actual value
perf_event can support is static, but I'm pretty sure it is higher than 64).
* 331c516c src/ctests/acpi.c: Remove the acpi.c file from ctests It wasn't
being built, and we removed the ACPI component a while ago.
* 73e7d191 src/components/vmware/vmware.c: Removed all old references to
#define VMWARE_PSEUDO_PPERF and switched over to getenv
2012-03-01
* 969b8aa9 src/ctests/Makefile src/ctests/zero_named.c src/papi.c: Three new
APIs: PAPI_query_named_event PAPI_add_named_event PAPI_remove_named_event and
a new test: zero_named Still to do: maybe test named native events and
support Fortran
* 97bf9bf8 src/papi.c src/papi.h: First pass implementation of {add, remove,
query}_named_event
* 2416af88 src/components/vmware/vmware.c: Add functionality to getenv
selectors
* 297f9cd6 src/papi.c: Fix possible race in _papi_hwi_gather_all_thrspec_data
The valgrind helgrind tool noticed this with the thrspecific test
* be599976 src/papi_internal.c: Add some locking in
_papi_hwi_shutdown_global_internal This caused a glibc double-free warning,
and was caught by the Valgrind helgrind tool in krentel_pthreads There are
some other potential locking issues in PAPI_shutdown, especially when debug
is enabled.
* 8444d577 src/utils/clockres.c src/utils/command_line.c: Cleanup the oxygen
markup for the utilities.
* 7144394f doc/Doxyfile-html: Missed a recursive tag for the html config
file.
* 63b2efc4 src/papi_preset.c: Fix segfaults in tests on AMD machines The
papi_preset code was wrongly calling papi_free() on some code that was
allocated with strdup() (not with papi_malloc). We were only noticing this
on AMD machines because it was the code for freeing developer notes in
presets, and currently only AMD events have developer notes.
* 0b1350df src/linux-common.c: Touch 'virtual_vendor_name' to cleanup a
warning on bluegrass.
2012-02-29
* 1f17b571 src/Makefile.inc src/Rules.perfctr-pfm src/Rules.pfm4_pe...: Merge
the contents of papi_libpfm_presets.c into papi_preset.c The code isn't
libpfm specific at all anymore, it's the generic "read presets from a file"
code. It makes more sense to find it in papi_presets.c
* d087d49f src/papi_fwrappers.c: Fix Fortran breakage after the preset event
changes
* 156141ec src/papi_libpfm_presets.c src/papi_preset.c src/papi_preset.h:
Simplify papi_libpfm_presets.c Previously adding presets from
papi_events.csv was a three step process. 1. Load the presets from the
file, put in temporary structure. 2. Convert this temporary structure to a
"findem" dense structure 3. Pass this dense structure to
_papi_hwi_setup_all_presets for final assignment. This change creates the
final assignment directly without the intermediate two steps.
* 8bc2bafd src/papi.c src/papi.h src/papi_common_strings.h...: Make the
internal preset_info match the one exported by papi.h There were a lot of
cases where the same structure fields were available, just with different
names. That was confusing. Also, this allows using a pointer to the preset
info instead of having to copy values out of the structure when gathering
event info for presets.
* 8fda68cb src/genpapifdef.c src/papi.c src/papi_common_strings.h...: Merge
the 4 separate preset structs into one. _papi_hwi_presets was a structure
containing pointers to 4 other arrays of structures which held the actual
preset data. This change merges all of these into one big structure.
2012-02-28
* e69815d7 src/linux-bgp.c src/papi_internal.c src/papi_internal.h...:
Removing remaining vestiges of references to bipartite routines. Now the only
references are in papi_bipartite.h, perfctr-x86.c and winpmc-p3.c.
* 5766b641 src/papi_bipartite.h src/perfctr-x86.c
src/win2k/substrate/winpmc-p3.c: These changes implement the bipartite
allocation routine as a header file to be included in whatever cpu component
needs it. Right now, that's just perfecter-x86 and windows. Both components
have been modified and perfecter-86 compiles cleanly. Neither has been tested
since I don't have access to a test bed.
* 7f444b76 src/papi_libpfm_presets.c src/papi_preset.c src/papi_preset.h:
Merge the hwi_dev_notes structure into hwi_preset_data
* 21a1d197 src/components/vmware/vmware.c: add getenv
* 08c1b474 src/perfctr-x86.c: Merge bipartite routine into perfecter-x86
component, since this is effectively the only place it is used.
* 9ed9b1f5 src/papi.c: Remove a reference to PAPI_set_event_info() which was
removed for PAPI 4
* c626f064 src/ctests/all_events.c src/ctests/derived.c
src/ctests/kufrin.c...: Convert PAPI_event_info_t to separate preset event
info This moves the preset event info to its own separate structure, which
reduces greatly the large string overhead that is not used by the native
events.
* 787d6822 src/perfctr-x86.c: Move bipartite stuff to perfctr_x86 since
that's really the only place it's currently used.
* 229c8b41 src/components/vmware/vmware.h: Add env_var definition to vmware.h
* 46aaf6ca src/components/vmware/vmware.c: Remove all unneeded cases
* 874a5718 src/freebsd.c src/perfctr-ppc64.c: Remove more unused references
to .bpt_ routines in preparation for refactoring.
* 74e5a5fd src/components/vmware/vmware.h: Remove uneeded defines from
vmware.h header
* 58b51367 src/components/coretemp_freebsd/coretemp_freebsd.c
src/components/vmware/vmware.c src/solaris-niagara2.c...: Remove unused
references to .bpt_ routines in preparation for refactoring.
2012-02-27
* 6b184158 src/Makefile.inc src/components/coretemp/linux-coretemp.c
src/configure...: Have separate concept of "compiled-in" versus "active"
components With this change, the _papi_hwd[] component info array only
contains a null-terminated list of _active_ components. The
_papi_compiled_components[] array has the original full list. At
init_substrate[] time a pointer to a component is only put in the _papi_hwd[]
list if it is successfully initialized. In addition the
PAPI_num_compiled_components() and PAPI_get_compiled_component_info() calls
have been added, but this is probably a confusing interface so they might
only be temporary additions.
* 042bfd5b src/Makefile.inc src/papi.c src/papi_data.c...: Split the contents
of papi_data.c to various other files. The data declarations in papi_data.c
were mostly used in other files. Move these into more relevant locations.
* 1877862c src/papiStdEventDefs.h src/papi_common_strings.h: Remove the BGL
and BGP specific pre-defined events. They can be better replaced by
user-events, and we also had already removed BGL support completely a while
back. This removes some ifdefs from the pre-defined event list and keeps
future pre-defined events from having different eventcodes on different
platforms.
* c3986b79 src/components/coretemp/linux-coretemp.c
src/components/cuda/linux-cuda.c
src/components/infiniband/linux-infiniband.c...: Add names and descriptions
for components. Also fixes cuda and lmsensors build issues introduced by
vector.h cleanup
* 2c84f920 src/aix.c src/freebsd.c src/perf_events.c...: Add names and
descriptions to all of the CPU substrates.
* 9f3e634a src/components/example/example.c src/papi.h src/utils/component.c:
Add new "description" and "short_name" fields to .cmp_info structure This
description field allows components to provide extra information on what they
do. The short_name field will eventually be used to pre-pend event names.
The papi_component_avail utility has been updated to print the description.
The example component was updated to fill in these values.
* ab61c9a7 src/Makefile.inc src/genpapifdef.c src/papi_common_strings.h...:
Split papi_data.c into two parts papi_data.c was half data structure
definitions for all of PAPI and half string definitions used by both PAPI
*and* genpapifdef This splits the common string definitions to
papi_common_strings.h so that genpapifdef can still be built w/o linking
libpapi.a while making the code a lot easier to follow.
* b8e6294c src/solaris-ultra.c: Remove unncessary extern declarations from
solaris-ultra.c.
* 5ddaff91 src/sys_perf_event_open.c: Remove unncessary extern declarations
from sys_perf_event_open.c
* a6c463b7 doc/Doxyfile-common.config: Create a common config file for
doxygen. As part of streamlining the doxygen process, this is a new template
doxygen config files. This is a blank template file generated by doxygen
1.7.4 (the version currently mandated by the release procedure )
* dc2c11fa src/aix.c src/aix.h src/perfmon.c...: The vector pre-definition
should be in the .c file, not the .h file
* 0b3c83c3 src/perf_events.c: Remove unnecessary extern declarations in
perf_events.c
* b93efca0 src/perfmon.c src/perfmon.h: Remove unnecessary extern
declarations in perfmon.c
* 7f7a2359 src/papi_preset.c: Remove unnecessary extern declarations from
papi_preset.c
* ecec03ad src/papi_libpfm_presets.c: Remove extraneous extern declarations
from papi_libpfm_presets.c
* 7b5f3991 src/extras.c: remove extraneous extern declarations from extras.c
* f6470e4d src/aix-memory.c src/aix.c src/aix.h: Remove unncessary extern
declarations from aix.c
* f197d4ab src/papi_data.h src/papi_internal.c: Remove unncessary extern
declarations in papi_internal.c
* e7b39d48 src/papi.c src/papi_data.c src/papi_data.h...: remove unnecessary
extern definitions from papi.c
2012-02-24
* 92689f62 src/configure src/configure.in src/linux-common.c...: Add a
--with-pthread-mutexes option to enable using pthread mutexes rather than
PAPI custom locks This is useful when running on new architectures that
don't have a custom PAPI lock coded yet, and also for running valgrind
deadlock detection utilities that only work with pthread based locking.
* ca51ae67 src/papi_events.csv: Fix broken Pentium 4 Prescott support We
were missing the netbusrt_p declaration in papi_events.csv
* f6460736 src/linux-common.c: Fix build on POWER, broken by the
virtualization change.
* 91d32585 src/perfctr-x86.c src/perfmon.c: Fix some warnings that have
appeared due to recent changes.
* ae0cf00f src/linux-common.c src/papi_libpfm3_events.c
src/papi_libpfm4_events.c...: Clean up the Linux lock files The locking
primitives for some reason were spread among the libpfm code and the
substrate codes. This change moves them into linux-common and has them part
of the OS code. This way they will get properly initialized even if the perf
counter or libpfm code isn't being used.
2012-02-23
* 88847e52 src/papi.c src/papi_memory.h: Remove _papi_cleanup_all_memory
define from papi_memory.h The code in papi_memory.h said: /* define an
alternate entry point for compatibility with papi.c for 3.1.x*/ /* this line
should be deleted for the papi 4.0 head */ Since we are post papi-4.0 I
thought it was time to act on this. Of course papi.c was still using the old
name in one place.
* 1d29dfc6 src/papi_libpfm_presets.c src/perfctr.c src/perfmon.c: Fix some
missing includes found after the header cleanup.
* b425a9f4 src/Makefile.inc src/extras.c src/extras.h...: Header file cleanup
The papi_protos.h file contained a lot of no-longer in use exports. I split
up the ones that are still relevant to header files corresponding to the C
file that the functions are defined in.
* 07199b41 src/extras.c src/papi_vector.c src/papi_vector.h: Clean up the
papi_vector code. Remove things no longer being used, mark static functions
as static.
* d7496311 src/linux-common.c src/x86_cpuid_info.c src/x86_cpuid_info.h: Fix
a missing "return 1" which meant that the virtualization flag wasn't being
set right. With this fix, on saturn-vm1 we now get: Running in a VM
: yes VM Vendor: : VMwareVMware in the papi_native_avail
header
* 8da36222 src/freebsd.c src/linux-bgp.c src/papi.c...: Remove the
->add_prog_event function vector As far as I can tell this is a PAPI 2.0
remnant that was never properly removed. This also removes
PAPI_add_pevent(), PAPI_save(), and PAPI_restore(), none of which were
exported in papi.h so in theory no one could have been using them. Also
removes _papi_hwi_add_pevent()
* a5f3c8b5 src/aix.c src/freebsd.c src/linux-timer.c...: Reduce the usage of
MY_VECTOR whenever possible. This is an attempt to make the cpu-counter
components to be as similar as possible to external components.
* abbcbf29 src/any-null.h: Missed removing any-null.h during the any-null
removal.
* 665d4c5c src/linux-common.c: Somehow missed an include during the
virtualization addition.
* 0c06147b src/perfctr-2.6.x/usr.lib/event_set_centaur.os
src/perfctr-2.6.x/usr.lib/event_set_p5.os
src/perfctr-2.6.x/usr.lib/event_set_p6.os: Removes the last of the binary
files from perfctr2.6.x Some binary files were left out in the cold after a
mishap trying to configure perfctr for the build test.
* 3acb7d57 src/Makefile.inc src/configure src/configure.in...: Add support
for reporting if we are running in a virtualized environment to the
PAPI_hw_info_t structure. This currently only works on x86. it works by
looking at bit 31 of ecx after a cpuid (the "in a VM" bit) and then using
leaf 0x40000000 to get the name of the VM software (this works for VMware and
Xen at least) x86_cache_info.c was renamed to x86_cpuid_info.c to better
reflect what goes on in that file (it does various things based on the cpuid
instruction). the testlib header was updated to report virtualization status
in the papi header (printed for things like papi_native_avail).
2012-02-22
* 9c7659b5 src/Makefile.inc src/freq.c: Remove the freq.c file as nothing
seemed to be using it.
* d205e2d3 src/perfctr-x86.c: Made a stupid typo when converting perfctr to
call libpfm functions with the component id.
* 25b41779 src/papi_libpfm3_events.c src/papi_libpfm4_events.c
src/papi_libpfm_events.h...: When updating the preset code to take a
component index I missed a few callers.
* a713ffb1 src/papi_internal.c src/papi_vector.c: Remove any-null component
* 27e1c2c5 src/any-null-memory.c src/any-null.c src/any-proc-null.c...:
Remove the any-null component.
* 25779ae0 PAPI_FAQ.html: Saving another version of the FAQ after adding a
git section, and removing several obsolete sections. These questions still
need detailed review for relevance and timeliness.
* 449a1a61 src/ctests/overflow_allcounters.c: Fix overflow_allcounters which
was making assumptions about component 0 existing.
* f21be742 src/ctests/hwinfo.c: Make the hwinfo test not bail out if no
counters are available.
* ebc675e6 src/ctests/memory.c: Make sure the memory ctest runs even if no
components are available.
* 9b3de551 src/linux-common.c src/perf_events.c src/perfmon-ia64.c...: Make
sure the system info init happens at os init time. Otherwise the system info
never gets set if a perfcounter component isn't available.
* 59e47e12 src/papi_internal.c: Make sure that _papi_hwi_assign_eventset()
does the right thing if no components are available.
* dd51e5d6 src/ctests/api.c: The api test would fail in the no cpu component
case. Fix it to properly check for errors before attempting to run
high-level PAPI tests.
* 069e9d2f src/aix.c src/papi.c src/papi_internal.h...: Fix code that was
depending on _papi_hwd[0] existing. Most of this was in the presets code.
The preset code had many assumptions so that you can only code presets with
component[0]. This fixes some of them by passing the component index around.
* 7259eaec src/papi_vector.c: Fix up papi_vector to get rid of some warnings
introduced on AIX.
* 16fe0a61 src/aix.c src/solaris-ultra.c: Fix two last substrates where I
missed some fields in the OS structure conversion.
* 625871ec src/perfmon.c: Missed a cmp_info field in perfmon.c
* 680919d9 PAPI_FAQ.html: Saving the latest version of the FAQ before
undertaking major revisions.
* 3d4fa2e5 src/linux-timer.c src/perfctr-x86.h: Fix the perfctr code to
compile if configured with --with-virtualtimer=perfctr
* bbd7871f src/perfctr.c: Missed two OS vector calls in the perfctr code
during the conversion.
* bc6d1713 src/Makefile.inc: Removed one of the two instances of MISCOBJS
listed in Makefile.inc.
2012-02-21
* 40bc4c57 src/papi_vector.c src/papi_vector.h: Remove now-unused OS vectors
from the main papi vector table.
* 3c6a0f7b src/aix.c src/freebsd.c src/linux-bgp.c...: Convert PAPI to use
the _papi_os_vector for the operating-system specific function vectors.
* 568abad5 src/papi_vector.h: Add new _papi_os_vector structure to hold
operating-system specific function vectors.
* a39d2373 src/ctests/subinfo.c: Missed removing a field from the subinfo
ctest.
* 1d930868 src/papi.h: Remove fields now in PAPI_os_info_t from the
component_info_t struct.
* d397d74a src/components/example/example.c: Remove fields now in
PAPI_os_info_t from the example component.
* 8cd5c8e0 src/aix.c src/freebsd.c src/linux-bgp.c...: Modify all the
substrates to use _papi_os_info. instead of _papi_hwd[0]->cmp_info for the
values moved to the OS struct
* 58855d3a src/papi_internal.h: Add padding for future expansion to
PAPI_os_info_t Add _papi_hwi_init_os(void); definition
* ea1930e1 src/papi_internal.h: Add new PAPI_os_info_t structure to
papi_internal.h
* 0eac1b29 src/utils/multiplex_cost.c: Modify multiplex_cost to properly use
the API_get_opt() interface to get itimer data, rather than directly
accessing the fields from the cmp_info structure. This would have broken
after the OS split.
* 87c2aa2f src/ctests/subinfo.c: subinfo was printing itimer data from the
cmpinfo structure. These values will not be in cmpinfo once the OS split
happens.
* f2c62d50 src/components/vmware/vmware.h: Clean up the VMware Header a bit
2012-02-17
* 6f0c1230 src/aix.c src/components/coretemp/linux-coretemp.c
src/components/coretemp_freebsd/coretemp_freebsd.c...: The git conversion
reset all of the CVS $Id$ lines to just $Id$ Since we depend on the $Id$
lines for the component names, I had to go back and fix all of them to be the
component names again.
* 2d208d0e src/perfctr-2.6.x/usr.lib/event_set_centaur.o
src/perfctr-2.6.x/usr.lib/event_set_p5.o
src/perfctr-2.6.x/usr.lib/event_set_p6.o: Remove a few binary files in
perfctr-2.6.x
* f78bf1af src/libpfm-3.y/Makefile src/libpfm-3.y/README
src/libpfm-3.y/docs/Makefile...: More cleanups from the migration, latest
version of libpfm-3.y perfctr-2.[6,7] Version numbers got really confused in
cvs and the git cvsimport didn't know that eg 1.1.1.28 > 1.1 ( see
perfctr-2.6.x/CHANGES revision 1.1.1.28.6.1 :~)
* e8aa2e61 INSTALL.txt: Explicitly state that 3.7 was the last version of
PAPI with good windows support.
* 546901fa src/components/cuda/linux-cuda.c: Modified CUDA component so that
a PAPI version - that was configured with CUDA - will successfully build on a
machine that does not have GPUs.
2012-02-16
* 49d9f71c src/.gitignore: Add a .gitignore file with the files that PAPI
autogenerates. This way they won't clutter up "git status" messages