TheRock/rocm-systems/projects/rocprofiler-systems/external/papi/ChangeLogs/ChangeLogP411.txt

2010-09-30

	* src/: configure, configure.in: When --with-OS=CLE is enabled,
	  check the kernel version and use perfmon2 for old kernels and
	  perf_events for new kernels.

	* src/: configure, configure.in: 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, configure.in: 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, Makefile.in, configure.in, papi.h: Bump version
	  number to 4.1.1 in affected files.  This in prep for the pending
	  release.

	* 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
	  release.

2010-09-29

	* src/Rules.pfm: The -Werror problem was still occurring on
	  ia64/perfmon compiles, as I hadn't updated Rules.pfm

	* src/: configure, configure.in, 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
	  above.

2010-09-22

	* src/perf_events.c: Attempt to add mmtimer support to perf_events
	  substrate.

	* 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.

2010-09-17

	* src/libpfm-3.y/: config.mk, lib/intel_wsm_events.h: Fix conflicts
	  from merge.

2010-09-15

	* src/: Makefile.inc, 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.

2010-09-13

	* 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")

2010-09-08

	* src/Makefile.inc: 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)

2010-09-03

	* src/ctests/sdsc.c: Had a "%d" instead of "%lld" in that last
	  commit.

	* 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.

2010-09-02

	* src/papi.c: Remove code that reported ENOSUPP if HW multiplexing
	  is not available.

	  PAPI can automatically perform SW multiplexing if HW is not
	  available.

	  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)

2010-08-31

	* 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.

2010-08-30

	* 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.

2010-08-27

	* 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
	  documentation

	* 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.

2010-08-26

	* 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
	  period.

	  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
	  overflow_allcounters

	  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
	  interrupts.

	  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
	  (POWER6)

	  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.

2010-08-25

	* 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
	  struct.

	  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).

2010-08-24

	* 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
	  code.

	  This meant that we were attempting to read non-existant hardware
	  overflow data, causing a lot of errors to be printed to the
	  screen.

	  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
	  process

2010-08-20

	* src/Makefile.inc: 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

2010-08-18

	* 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
	  called.

	  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

2010-08-05

	* src/perf_events.c: Use F_SETOWN_EX instead of F_SETOWN in
	  tune_up_fd()

	  This fixes a multi-thread overflow bug found with the Rice
	  test-suite.

	  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.

2010-08-04

	* src/: Makefile.inc, 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:

	    papi.h
	       +  Add PAPI_CPU_ATTACHED
	       +  Add strutctures needed for CPU_ATTACH
	    Makefile.in
	       + include the new cpus.c file
	    papi_internal.c
	       + add call to _papi_hwi_shutdown_cpu() in
	  _papi_hwi_free_EventSet()
	       + make remap_event_position() non-static
	       + add_native_events() and remove_native_events() use
		 _papi_hwi_get_context()
	       + _papi_hw_read() has some whitespace and debug message
	  changes,
		 and removes an extraneous loop index
	    papi_internal.h
	       + a new CPUS_LOCK is added
	       + cpuinfo struct added to various structures
	       + an inline call called _papi_hwi_get_context() added
	    perf_events.h
	       + a cpu_num field added to control_state_t
	    perf_events.c
	       + open_pe_events() allows per-cpu counting,
		 additional debug was added
	       + set_cpu() function added
	       + new debug messages in set_granularity() and
	  _papi_pe_read()
	       + _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
	    genpapifdef.c
	       + PAPI_CPU_ATTACHED added
	    threads.h
	       + an ESI field added to ThreadInfo_t
	    papi.c
	       + 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
	  added.
	       + 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.

2010-08-03

	* 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.

2010-07-29

	* src/: configure, configure.in, 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.

2010-07-26

	* 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.

2010-07-16

	* 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
	  Core2.

	  This was tested on actual Pentium Pro and PIII hardware (as well
	  as on a few	Pentium 4 machines plus a Core2 machine)

2010-07-02

	* src/: papi_hl.c, ctests/api.c: Added remaining low-level api
	  tests