Skip to content


Subversion checkout URL

You can clone with HTTPS or Subversion.

Download ZIP
branch: master
Fetching contributors…

Cannot retrieve contributors at this time

477 lines (349 sloc) 17.255 kb
* src/: configure, When --with-OS=CLE is enabled,
check the kernel version and use perfmon2 for old kernels and
perf_events for new kernels.
* src/: configure, If no sources of perf counters are
available, then use the generic_platform substrate instead.
Currently the code would always fall back on perfctr even if no
perfctr support was available.
* src/: configure, If you specify --with-perf-events
or --with-pe-include but the required perf_event.h header is not
available, then have configure fail with an error.
* papi.spec: Bump version number to 4.1.1 in affected files. Also
bump requirement for kernel from 2.6.31 to 2.6.32. This in prep
for the pending release.
* src/: configure,,, papi.h: Bump version
number to 4.1.1 in affected files. This in prep for the pending
* INSTALL.txt: Hope this late commit doesn't interfere with
anything. This updates the INSTALL.txt to reflect all of the
improvements we've made to perf_event support since the last
* src/Rules.pfm: The -Werror problem was still occurring on
ia64/perfmon compiles, as I hadn't updated Rules.pfm
* src/: configure,, perf_events.c, perf_events.h,
sys_perf_counter_open.c, sys_perf_event_open.c, syscalls.h:
Remove support for the perf_counter interface in kernel 2.6.31.
Now supports only the perf_event interface in kernel 2.6.32 and
* src/perf_events.c: Attempt to add mmtimer support to perf_events
* src/: multiplex.c, papi.c, papi_protos.h: The multiplex code
currently does not make a final adjustment at the time of
MPX_read(). This is to avoid the case where counts could be
decreasing if you have multiple reads returning estimated values
before the next actual counter read.
While this code works to keep the results non-decreasing, it can
cause significant differences from expected results for final
reads, especially if many counters are being multiplexed.
This is seen in the sdsc-mpx test. It was failing occasionally
on some machines by having error of over 20% (the cutoff for a
test error) when multiplexing 11 events.
What this fix does is to special case the PAPI_stop() case when
multiplexing is enabled, having the PAPI_stop() do a final
adjustment. The intermediate PAPI_read() case is not changed.
This fixes the sdsc-mpx case, while still passing the mendes-alt
case (which checks for non-decreasing values).
There is a #define that can be set in multiplex.c to restore the
previous behavior.
* src/ctests/mendes-alt.c: This is our only test that checks to see
if multiplexed values are non-decreasing or not. Unfortunately
the test currently doesn't fail if values do go backward.
This change causes the test to fail if it finds multiplexed
counts that decrease.
* src/libpfm-3.y/:, lib/intel_wsm_events.h: Fix conflicts
from merge.
* src/:, Rules.perfctr-pfm, Rules.pfm_pe: Finally fix
the -WExtra problem.
The issue was -WExtra was being passed to libpfm, but only in the
case where the user had a CFLAGS env variable.
It turns out this is due to the following from section 5.7.2 of
the gmake manual: Except by explicit request, make exports a
variable only if it is either defined in the environment
initially or set on the command line,
And the fix is also described: If you want to prevent a
variable from being exported, use the unexport directive,
So I've added an "unexport CFLAGS" directive, which seems to be
the right thing as our Makefile explicitly passes CFLAGS to the
sub-Makefiles that need it. This seems to fix the build.
* src/libpfm-3.y/: docs/man3/libpfm_westmere.3,
lib/intel_wsm_events.h, lib/intel_wsm_unc_events.h,
lib/pfmlib_intel_nhm.c, lib/pfmlib_priv.h: Fix the missing files
from the import (CVS claims this as a "conflict")
* src/ Fixed the recipies for [c|f]tests and utils.
$(LIBRARY) => $(papiLIBS) (this way we don't build libpapi.a if
we won't want it)
* src/ctests/sdsc.c: Had a "%d" instead of "%lld" in that last
* src/ctests/sdsc.c: Give a more detailed error message on the
sdsc-mpx test.
We're seeing sporadic failures (probably due to results being
close to the threshold value) but it's hard to tell on buildbot
which counter is failing because the error message didn't print
the value.
* src/papi.c: Remove code that reported ENOSUPP if HW multiplexing
is not available.
PAPI can automatically perform SW multiplexing if HW is not
With this part of my previous multiplexing patch reverted,
multiplexing seems to work even on 2.6.32 perf_events (by
reverting to SW mode on those machines)
* src/perf_events.c: Explicitly set the disabled flag to zero in
perf_events for new events.
It was possible with an event set that if you removed an event
then added a new one that the disabled flag was obtaining the
value from the previously removed event.
This fix doesn't seem to break anything, but the code involved is
a bit tricky to follow.
This fixes the sdsc4-mpx test on sol.
* src/components/coretemp/: Rules.coretemp, linux-coretemp.c,
linux-coretemp.h: Initial stab at a coretemp component.
This component exposes every thing that looks like a useful file
under /sys/class/hwmon.
* src/perf_events.c: F_SETOWN_EX is not available until 2.6.32, so
don't use it unless we are running on a recent enough kernel.
* src/perf_events.c: Pentium 4 was not supported by perf_events
until version 2.6.35. Print an error if we attempt to use it on
an older kernel.
* src/ctests/overflow_allcounters.c: The "overflow_allcounters"
test failed on perfmon2 kernels because the behavior of a
counter on overflow differs between the various substrates.
Therefore detect if we're running on perfmon2 and print a
warning, but still pass the test.
* src/libpfm-3.y/lib/: intel_wsm_events.h, intel_wsm_unc_events.h,
pfmlib_intel_nhm.c, pfmlib_priv.h: updating
* src/libpfm-3.y/docs/man3/libpfm_westmere.3: removing westmere
* src/perf_events.c: Fix warning in compile due to missing
parameter in a debug statement.
* src/ctests/test_utils.c: In the ctests, test_skip() was
attempting a PAPI_shutdown() before exiting. On multithreaded
tests (that had already spawned threads before the decision to
skip) this really causes the programs to end up confused and
reports spurious memory errors.
So remove the PAPI_shutdown() from test_skip(). There's a
comment in test_fail() that indicates this was already done
there for similar reasons.
* src/ctests/byte_profile.c: byte_profile was failing on systems
where fp_ops is a derived event.
modify the test so it gives a warning instead of failing and
avoids using the derived event.
* src/perf_events.c: At PAPI_stop() time a counter with overflow
enabled is being adjusted by a value equal to the sampling
It looks like this isn't needed (and is generating an overcount
that breaks overflow_allcounters).
I'm still checking up on this code; if it turns out to be
necessary I may have ro revert this later.
* src/ctests/overflow_allcounters.c: Add validation check to
It turns out perf_event kernels overcount overflows for some
reason, while perfctr doesn't. I'm investigating.
* src/ctests/: overflow_allcounters.c, papi_test.h, test_utils.c:
On Power5 and Power6, hardware counters 5 and 6 cannot generate
This means the overflow_allcounters test was failing because
overflow could not be generated for events 5 and 6.
Add code that special cases Power5 and Power6 for this test (and
generate a warning)
* src/perf_events.c: Change some debug messages to be warnings
instead of errors.
* src/: papi.c, ctests/second.c: Fix ctests/second on bluegrass
The test was testing domains by trying
PAPI_DOM_ALL^PAPI_DOM_SUPERVISOR in an attempt to turn off the
SUPERVISOR bit. This fails on Power6 as it leaves the
PAPI_DOM_OTHER bit set, which isn't allowed.
How did the test earlier measure PAPI_DOM_ALL then, which has all
bits set? Well it turns out papi.c silently corrects
PAPI_DOM_ALL to be available_domains. But if you fiddle any of
the bits this correction is lost. This is probably not the
right thing to do, but the best way to fix it is not clear.
For now this modifies the "second" test to clear the DOM_OTHER
bit too if the domain setting fails with it set.
* src/: papi.c, papi.h, perf_events.c, ctests/kufrin.c,
ctests/mendes-alt.c, ctests/multiplex1.c,
ctests/multiplex1_pthreads.c, ctests/multiplex2.c,
ctests/multiplex3_pthreads.c, ctests/sdsc.c, ctests/sdsc2.c,
ctests/sdsc4.c, ftests/fmultiplex1.F, ftests/fmultiplex2.F: Add
support for including the OS version in the component_info_t
Use this support under perf_events to disable multiplexing
support if the kernel is < 2.6.33
Modify the various multiplexing tests to "skip" if they get a
PAPI_ENOSUPP when attempting to set up multiplexing.
* src/ctests/all_native_events.c: Update all_native_events ctest to
print warning in the case where we skip events because they
aren't implemented yet (offcore and uncore mostly).
* src/ctests/: papi_test.h, profile.c, test_utils.c: Adds a new
"test_warn()" function for the ctests.
This allows you to let tests pass with a warning.
This is useful in cases where you don't want to forget that an
option needs implementing, but that the feature being missed
isn't important enough to fail the test.
The first user of this is the "profile" test. We warn that
PAPI_PROFIL_RANDOM is not supported on perf_events.
* src/perf_events.c: From what I can tell, on perf_events the
overflow PAPI_OVERFLOW_FORCE_SW case was improperly falling
through in _papi_pe_dispatch_timer() to also run the HARDWARE
This meant that we were attempting to read non-existant hardware
overflow data, causing a lot of errors to be printed to the
This shows up in the overflow_force_software test
* src/ctests/: ipc.c, multiplex2.c, multiplex3_pthreads.c,
test_utils.c: Some minor changes to the ctests.
+ ipc -- fail if the reported IPC value is zero + multiplex2 --
fail if all 32 counter values report as zero + multiplex3_pthread
-- give up sooner if each counter returns zero. otherwise
the test can take upwards of an hour to finish and makes the
fan on my laptop sound like it's going to explode in the
* src/ Disable CFLAGS += $(EXTRA_CFLAGS) (-Wextra) for
now. This will get buildbot running again, and if I can manage
to figure out exactly what the Makefiles are doing I'll re-enable
it again.
* src/perf_events.c: Add support for Pentium 4 under perf events.
This requires a 2.6.35 kernel. On p4 perf events requires a
special format for the raw event, so we modify the results from
libpfm3 to conform to what the kernel expects.
* release_procedure.txt: release_procedure updated to reflect files
to keep under /doc
* src/perf_events.c: Patch from Gary Mohr that allows PAPI on perf
events to catch permissions problems at the time of
configuration, rather than only appearing once papi_start() is
Quick summary of changes: + Adds a check_permissions() routine
PERF_COUNT_HW_INSTRUCTIONS is used as the test event. +
check_permissions() is called during PAPI_ATTACH,
PAPI_CPU_ATTACH and PAPI_DOMAIN + Various "ctl" structures
renamed "pe_ctl" + Some minor debug changes
* src/perf_events.c: Use F_SETOWN_EX instead of F_SETOWN in
This fixes a multi-thread overflow bug found with the Rice
F_SETOWN_EX doesn't exist until Linux 2.6.32. We really need
some infrastructure that detects the running kernel at init time
and warns that things like F_SETOWN_EX, multiplexing, etc., are
unavailable if the kernel is too old.
* src/:, cpus.c, cpus.h, genpapifdef.c, papi.c,
papi.h, papi_defines.h, papi_internal.c, papi_internal.h,
perf_events.c, perf_events.h, threads.h: This is the
PAPI_CPU_ATTACH patch from Gary Mohr that also fixes a problem
with multiple event sets on perf events.
Changes by file:
+ Add strutctures needed for CPU_ATTACH
+ include the new cpus.c file
+ add call to _papi_hwi_shutdown_cpu() in
+ make remap_event_position() non-static
+ add_native_events() and remove_native_events() use
+ _papi_hw_read() has some whitespace and debug message
and removes an extraneous loop index
+ a new CPUS_LOCK is added
+ cpuinfo struct added to various structures
+ an inline call called _papi_hwi_get_context() added
+ a cpu_num field added to control_state_t
+ open_pe_events() allows per-cpu counting,
additional debug was added
+ set_cpu() function added
+ new debug messages in set_granularity() and
+ _papi_pe_ctl() has PAPI_CPU_ATTACH code added
+ _papi_pe_update_control_state() has the default domain
set to be PAPI_DOM_USER instead of pe_ctl->domain
+ an ESI field added to ThreadInfo_t
+ many new ABIDBG() debug messages added
+ PAPI_start() updated to check for CPU_ATTACH conflicts,
has whitespace fixes, gets context now,
if dirty calls update_control_state()
+ PAPI_stop(), PAPI_reset(), PAPI_read(), PAPI_read_ts(),
PAPI_accum(), PAPI_write(), PAPI_cleanup_eventset(),
all use _papi_hwi_get_context() to get context
+ PAPI_read() has some braces added
+ PAPI_get_opt() and PAPI_set_opt() have CPU_ATTACHED code
+ PAPI_overflow() and PAPI_sprofil() now report errors if
CPU_ATTACH enabled
cpus.c, cpus.h
+ New files based on threads.c and threads.h
I made some additional changes, based on warnings given by gcc
+ Added a few missing function prototypes in cpus.h + Update
PAPI_MAX_LOCK as it wasn't increased to handle the new
addition of CPUS_LOCK + Removed various variables and functions
reported as being unused.
* src/: papi_internal.h, papi_lock.h: The option
--with-no-cpu-counters was not supported on AIX. This has been
fixed and works now. Also the get_{real|virt}_{cycles|usec}
implementations for AIX (checked in Jul 29) have now been tested
and work correctly.
* src/: configure,, papi_lock.h, papi_vector.c: Added
AIX support for the get_{real|virt}_{cycles|usec} functions +++
Fortran tests are now compiling on AIX. Wrong compiler flags were
used for the AIX compilers.
* src/papi_events.csv: add PAPI_L1_DCM for atom
* src/x86_cache_info.c: Update the x86 cache_info table.
The data from this table now comes from figure 3-17 in the
Intel Architectures Software Reference Manual 2A (cpuid
instruction section)
This fixes an issue on my Atom N270 machine where the L2 cache
was not reported.
* INSTALL.txt, src/perf_events.c, src/perf_events.h: Perf Events
now support attach and detach.
The patch for supporting this was written by Gary Mohr
* src/papi_events.csv: Add a few missing events to Nehalem, based
on reading Intel Volume 3b.
* src/papi_events.csv: Fix Westmere to not use L1D_ALL_REF:ANY
I tested this on a Nehalem which has the proper behavior,
unfortunately no Westmere here to test on.
* src/: papi_events.csv, papi_pfm_events.c, perfctr-x86.c: Enable
support for having more than one CPU block with the same name in
the .csv file. This allows easier support for sharing events
between similar architectures.
I *think* this is needed and *think* it shouldn't break anything,
but I might have to back it out.
Also fixes event support for Pentium Pro / Pentium III/ P6 on
perfmon2 and perf events kernels.
Also fixed some confusion where perfctr called chips "Intel Core"
meaning Core Duo wheras pfmon called "Intel Core" meaning
This was tested on actual Pentium Pro and PIII hardware (as well
as on a few Pentium 4 machines plus a Core2 machine)
* src/: papi_hl.c, ctests/api.c: Added remaining low-level api
Jump to Line
Something went wrong with that request. Please try again.