forked from icl-utk-edu/papi
-
Notifications
You must be signed in to change notification settings - Fork 0
/
ChangeLogP413.txt
598 lines (416 loc) · 21 KB
/
ChangeLogP413.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
2011-05-10
* src/Rules.pfm_pe: The --with-bitmode parameter was not being
passed along to libpfm3, so it was not possible to build
perf_event PAPI on non-default bitmodes. This change passes
along the $(BITFLAGS) value to the libpfm3 make invocation.
* src/: papi_pfm_events.c, papi_pfm_events.h, perf_events.c: The
perf_events code was using __u64 instead of uint64_t and this was
causing a warning when compiling for 64-bit Power.
* src/libpfm-3.y/lib/amd64_events_fam15h.h: Added Robert Richter's
patch with a few new events for AMD Family 15h.
2011-05-06
* INSTALL.txt: Load the 'gcc' module not 'gnu' module for Cray.
* INSTALL.txt: Update the install instructions for Cray XT and XE
systems.
* src/ctests/: multiattach.c, multiattach2.c: Make the multiattach
and multiattach2 failures into warnings.
I have a proposed fix that makes the failures go away, but it has
not been tested much and also causes some new fcntl() error
messages under perfctr.
So temporarily make the tests only warn for the release and I'll
work on a proper fix for after. The behavior in these tests has
been broken for a long time so it is not a recent regression.
* src/papi_memory.c: Band-aid for the leak debugging statement in
papi_memory.c on NO_VARARG_MACRO systems. (aix currently)
2011-05-05
* src/ctests/multiattach.c: Had the division backwards on the
validation.
* src/ctests/multiattach.c: Update the multiattach test to fail if
the results aren't in the proper ratio. This was failing on
perf_event kernels but since the results weren't checked it was
never reported as an error.
* delete_before_release.sh: delete cvs2cl.pl before release
* ChangeLogP413.txt: First cut change log for the 4.1.3 release.
Nothing's frozen yet...
* cvs2cl.pl: Perl script to generate change logs. Keeping it with
the project makes life easier.
* INSTALL.txt: Change INSTALL to reflect that we support power7.
* src/Makefile.in, src/configure, src/configure.in, src/papi.h,
doc/Doxyfile, doc/Doxyfile-everything, papi.spec: Modfy version
number for pending release: 4.1.3.0
2011-05-03
* src/: papi_internal.c, papi_internal.h, sys_perf_event_open.c,
ctests/attach2.c: Cleanup the _papi_hwi_cleanup_eventset()
function in papi_internal.c
This function was re-using existing functionality to remove one
event at a time before cleaning out the eventset. This is not
strictly necessary and was breaking on perf_event eventsets that
were attached to finished processes, as a call to
update_control_state() would close/reopen the perf_event fd,
failing when the finished process went away after the close.
The new code removes all events from the eventset in one go
before calling update_control_state.
The change here also updates code comments as necessary, as some
of the code in papi_internal.c can be a bit obscure.
It also updates some of the comments in ctests/attach2.c to give
better debugging info.
2011-04-28
* src/threads.c: Uncomment the actual signal passing functionality
in _papi_hwi_broadcast_signal
* src/papi_debug.h: Include files added to papi_debug.h
* src/components/README: Added detailed instructions on how to
build PAPI with the CUDA component
2011-04-27
* src/threads.c: Move an escape test to the outer loop in
_papi_hwi_broadcast_signal.
This cleans up an infinite loop where before we would only break
out of the component look, not the thread list walking loop.
* src/: papi.c, papi_internal.c, papi_internal.h, papi_protos.h:
Clean up papi_internal.c so that functions not used outside are
marked static.
* src/: papi_pfm_events.c, papi_preset.c, pmapi-ppc64_events.c:
papi: Fix some memory leaks
Signed-off-by: Robert Richter <[email protected]>
* src/perf_events.c: papi: Make functions and variables static in
perf_events.c
All this functions and variables are not used outside
perf_events.c. Making them static.
Signed-off-by: Robert Richter <[email protected]>
* src/papi_pfm_events.c: papi: Fix crash in error handler for
pfm_get_event_code_counter()
Signed-off-by: Robert Richter <[email protected]>
* src/utils/native_avail.c: papi: Fix error check in native_avail.c
Signed-off-by: Robert Richter <[email protected]>
2011-04-26
* src/libpfm-3.y/: include/perfmon/pfmlib_amd64.h,
lib/pfmlib_amd64.c: AMD architectural PMU could not be detected
for family 15h as there was a strict check for AMD family 10h.
Enabling it now for all families from 10h.
Signed-off-by: Robert Richter
* src/libpfm-3.y/lib/amd64_events_fam15h.h: There is no kernel
support for AMD family 15h northbridge events, disabling them in
libpfm3 to not report them as available native events.
Patch from Robert Richter
* src/: configure, configure.in, linux-common.c: Add some extra
debug messages for better tracking of the --with-assumed-kernel
configure option.
2011-04-25
* src/: configure, configure.in, linux-common.c: Add a new
configure option: --with-assumed-kernel=<ver> This allows you
to specify a kernel revision to (instead of being autodetected
with uname) for perf_event workaround purposes. With this you
can force PAPI to not use workarounds on kernels with
backported versions of perf_event features.
2011-04-19
* src/: Makefile.inc, configure, configure.in, papi_debug.h,
papi_internal.h, sys_perf_event_open.c: Add debugging to
sys_perf_event_open.c to show exactly what values are being
passed to the perf_event_open syscall.
2011-04-18
* src/: run_tests.sh, ctests/attach2.c, ctests/attach3.c: Fix for
finding attach_target with execlp to search the path.
2011-04-14
* src/: Rules.pfm, configure, configure.in, linux-ia64-pfm.h,
linux-ia64.c, linux-ia64.h, perfmon-ia64-pfm.h, perfmon-ia64.c,
perfmon-ia64.h, perfmon.h: Rename the linux-ia64-* files to be
called perfmon-ia64-*
This is a more descriptive name, and makes it more obvious what
the files are for.
* src/libpfm-3.y/: include/perfmon/pfmlib_amd64.h,
lib/pfmlib_amd64.c, lib/pfmlib_amd64_priv.h: Patch to have
libpfm3 use 6 counters on Interlagos.
Patch provided by Robert Richter
* src/linux-memory.c: Fix the POWER cache detection routines to
work properly on POWER7.
Patch provided by Corey Ashford
* src/: configure, configure.in: Have configure check for ifort if
gfortran, etc, not found.
Patch by Gary Mohr
* src/ctests/johnmay2.c: Update the validation message on the
ctests/johnmay2.c test to be less confusing. Also add some
comments to the source code.
Problem reported by Steve Kaufmann.
2011-04-13
* src/ctests/: multiattach2, multiattach2.c: Remove the
accidentally added ctests/multiattach2 and add instead the proper
ctests/multiattach2.c
* src/Makefile.inc: components_config.h is cleaned out with make
clobber, not make clean this should fix the build bot issues.
* src/ctests/: Makefile, attach3.c, multiattach.c, multiattach2,
zero_attach.c: Minor typos in comments. Discovered another bug in
attach code demonstrated by multiattach2. You cannot have an
eventset running that is self counting as well as one that is
attached. PAPI thinks that both are running and throws an error.
* src/perf_events.c: We must update the control state after
attaching for perf_events, zero_attach now passes
* src/ctests/: Makefile, attach2.c, attach3.c, do_loops.c: This
commit adds testing of attaching to fork/exec'd executables.
zero_attach and multiattach just test forks. This also modifies
do_loops.c to be able to generate a test driver when
-DDUMMY_DRIVER is defined so we can use it to generate flops as a
sub process.
Attach2 and attach3 have one important difference.
Attach3 does a 'assign component' before attaching and then
adding events. Attach2 does not assign a component and thus
should inherit the default component.
The current bug in PAPI is that: * The default component is not
assigned until you add an event. * However, attaching an
eventset without events is perfectly valid, but we get an error.
Possible solution is that the default component should be
assigned at create time.
2011-04-12
* src/ctests/multiattach.c: Make sure the two processes compute
different numbers of flops to test attach
2011-04-05
* src/power7_events.h: Turns out Maynard Johnson answered my
questions about the native_name enum back in December. ( this is
a correct version of the events file )
As I found out, the AIX substrates do not use the native_name
enum. But a hypothetical perfctr build would.
2011-04-04
* src/Makefile.inc: Clear out the components_config.h file on make
clobber
* src/: aix.c, power7_events.h: Initial support for power7 aix, the
events file is a copy of power6_events.h with the number of
groups changed. The native_name enum is unchanged, but unused?
2011-04-01
* src/configure.in: Commited wrong configure.in
* src/: configure, configure.in: Clean up setting bitmode flags for
non-gcc (xlc in this case) compilers.
* src/papi_events.csv: Change the Nehalem PAPI_FP_OPS event from
FP_COMP_OPS_EXE:SSE_SINGLE_PRECISION+FP_COMP_OPS_EXE:DOUBLE_PRECISION
to FP_COMP_OPS_EXE:SSE+FP_COMP_OPS_EXE:X87
The new event gives the same results as the previous one, with
the added benefit of also counting 32-bit compiled x87 fp ops
properly.
More detailed analysis can be found here:
http://web.eecs.utk.edu/~vweaver1/projects/nehalem-fp_ops/
2011-03-28
* src/utils/multiplex_cost.c: Turns out that getopt_long isn't as
standard as I had hoped.
Convert multiplex_cost to use only getopt. -s disables software
multiplexing -k disables kernel multiplexing
2011-03-25
* src/: configure, utils/Makefile, utils/multiplex_cost.c,
configure.in: Multiplex_cost utility.
* src/utils/: Makefile, cost.c, cost_utils.c, cost_utils.h: Split
off the statistics functions from cost.
2011-03-22
* src/: run_tests_exclude_cuda.txt, run_tests.sh: Exclude some
fork/thread tests from fulltest that won't run with CUDA (reason:
cannot invoke same GPU from different threads)
2011-03-21
* src/utils/cost.c: Add a test for DERIVED_[ADD | SUB ] events to
papi_cost.
2011-03-18
* src/components/cuda/linux-cuda.c: all_native_events ctest failed
when CUDA Component is used. Reason: removing cuda events from
the eventset is currently not supported. According to the NVIDIA
folks this is a bug in cuda 4.0rc and will be fixed in rc2. Note
also, several fork and thread tests fail since it's illegal to
invoke the same GPU device from different processes / threads. We
need a mechanism that allows us to run tests for the CPU
component only.
2011-03-15
* src/utils/cost.c: Add a test case to cost util, look for a
derived-postfix event and if found, give timing information for
read calls to it.
This is just a first run at the test, Core2 and AMD have
candidate events and the test runs, but that is the extent of my
testing so far.
2011-03-11
* src/components/: README, cuda/Makefile.cuda.in, cuda/Rules.cuda,
cuda/configure, cuda/configure.in, cuda/linux-cuda.c,
cuda/linux-cuda.h: Added CUDA component, a hardware performance
counter measurement technology for the NVIDIA CUDA platform which
provides access to the hardware counters inside the GPU. PAPI
CUDA is based on CUPTI support - shipped with CUDA 4.0rc - in the
NVIDIA driver library. In any environment where the CUPTI-enabled
driver is installed, the PAPI CUDA component can provide detailed
performance counter information regarding the execution of GPU
kernels.
* src/components/: coretemp/linux-coretemp.c,
lustre/linux-lustre.c: Add some missing includes to components.
Thanks to Will Cohen for reminding us warnings matter. :)
* src/: configure, configure.in, perf_events.c: The SYNC_READ
workaround in perf_events.c was being handled at compile time,
rather than at run-time like all of our other workarounds.
Change it to be like our other kernel-version related
workarounds.
2011-03-09
* src/ctests/multiplex1_pthreads.c: Between 4.0.0 and 4.1.0 a
pthread_exit() call was added to ctest/multiplex1_pthreads.c that
caused the test to exit partway through the test and without
doing a proper PASS/FAIL result.
This changeset backs out that change, though the original change
was marked as a memory leak fix so a different fix may be needed.
Reported by Steve Kaufmann
* src/linux-timer.c: Add missing header needed by
--with-virtualtimer=times build.
Reported by Steve Kaufmann
2011-03-01
* src/: papi_pfm_events.c, perf_events.c: Fix broken Linux/PPC
build caused by my pfm_events code movement changes.
2011-02-25
* src/: papi_pfm_events.c, papi_pfm_events.h, perfctr-x86.h: My
changes yesterday broke the perfctr build. This should fix it.
2011-02-24
* src/ctests/inherit.c: Make the inherit test respect TESTS_QUIET
so that it does not print extra output during a run_tests.sh run
* src/ctests/overflow.c: Fix missing newline in the overflow
output.
Reported by Gary Mohr
* src/: papi_pfm_events.c, papi_pfm_events.h, perf_events.c: Move
the libpfm3 specific functions from perf_events.c into
papi_pfm_events.c
* src/perf_events.c: Separate the libpfm3-specific code from
_papi_pe_init_substrate() and _papi_pe_update_control_state()
into their own functions. This will allow eventual code sharing
and also make the libpfm4 merge easier.
* src/perf_events.c: Some minor cleanups I found after reviewing
the inherit merge. + Add missing "static inline" to the new
kernel-version codes + Remove duplicated test for Pentium 4
+ Fix a warning only seen if --with-debug is enabled
* src/: papi.c, papi.h, papi_internal.h, perf_events.c,
perf_events.h, ctests/Makefile, ctests/inherit.c,
ctests/test_utils.c: Merging Gary Mohr's re-implementation of
inherit into the code base. Thanks, Gary!
2011-02-23
* src/: any-null.h, freebsd.h, linux-bgp.h, linux-common.c,
linux-common.h, linux-context.h, linux-ia64.c, linux-ia64.h,
linux-lock.h, linux-memory.c, linux-ppc64.h, linux-timer.c,
papi_internal.h, papi_pfm_events.c, perf_events.c, perf_events.h,
perfctr-x86.h, perfctr.c, perfmon.h, solaris-niagara2.h,
solaris-ultra.h, solaris.h, x86_cache_info.c: Move some more
duplicated OS common code (in this case the locking code and the
context accessing code) out of the various substrate include
files and into a common location.
2011-02-22
* src/perf_events.c: Separate out the kernel-version dependent
checks and group them together near the beginning of the code.
This not only allows us to easily see which routines are
kernel-version dependent, but it makes it easier to disable the
checks one-by-one when debugging kernel-version related issues
like those found with the inherit patches.
2011-02-21
* src/papi_internal.c: Extend _papi_hwi_cleanup_eventset to free
memory and better cleanup after us.
2011-02-18
* src/papi.c: PAPI_assign_eventset_component changed; refuses to
reassign components.
2011-02-17
* src/: papi_events.csv, libpfm-3.y/include/perfmon/pfmlib.h,
libpfm-3.y/lib/amd64_events.h,
libpfm-3.y/lib/amd64_events_fam10h.h,
libpfm-3.y/lib/amd64_events_fam15h.h,
libpfm-3.y/lib/pfmlib_amd64.c,
libpfm-3.y/lib/pfmlib_amd64_priv.h: Add support for AMD Family
15h processors. Also adds suport for Family 10h RevE
Patches provided by Robert Richter
* src/utils/native_avail.c: Modify papi_native_avail to properly
handle event names with libpfm4-style "::" separators in them.
2011-02-15
* src/Makefile.inc: make install-doxyman will build/install the
doxygen version of the manpages.
Note that these pages are very rough right now, much work is
needed to get them to be a drop in replacement for the current
man pages. (mostly formatting related/use related issues, eg man
PAPI_start will not work yet; the content is there.)
* doc/Makefile: Add install target for doxygen generated man pages.
2011-02-11
* src/: perfctr-x86.c, perfctr.c: perfctr-2.6.42 introduced
PERFCTR_X86_INTEL_WSTMR PAPI added support for
PERFCTR_X86_INTEL_WSMR notice the missing T
Fix PAPI to use the proper define. This should fix Westmere
support on perfctr kernels.
2011-02-09
* src/: papi_protos.h, papi_vector.c, papi_vector.h,
papi_vector_redefine.h: Added function pointer destroy_eventset
to the PAPI vector table. Needed for the CUDA Component to
disable CUDA eventGroups, to destroy floating CUDA context, and
to free perfmon hardware on the GPU. (Note: the CUDA Component
cannot be released yet since we are still under NDA with NVIDIA.
Stay tuned.)
2011-02-07
* src/x86_cache_info.c: The cpuid leaf2 code was printing a message
to stderr if leaf4 was needed (only happens on Westmere
currently). Change this to be a MEMDBG() debug message instead.
2011-02-03
* src/: papi_events.csv, perfctr-x86.c: perfctr-x86 was reporting
"Core i7" instead of "Nehalem". i7 can mean Westmere or Sandy
Bridge too, so change the code to properly report Nehalem.
2011-01-27
* src/ctests/all_native_events.c:
Fix this ctest. It failed when the package was built with several
components because the eventset was reused and failed to add
events that were not from the first component.
In order to fix it, I recreate & destroy the eventset when the
current event does not belong to the previous component.
2011-01-26
* src/: configure, configure.in, linux-timer.c, perfmon.c: Fix Cray
CLE build.
* src/: configure, configure.in: Putting -Wall in cflags now
requires CC = gcc
* src/: aix.c, freebsd.c, linux-bgp.c, linux-common.c,
linux-memory.c, linux-memory.h, papi.c, papi_protos.h,
papi_vector.c, papi_vector.h, solaris-niagara2.c,
solaris-ultra.c, windows-common.c, windows-memory.c: Change the
paramaters passed to update_shlib_info() to match better with
those passed to get_system_info(). This only affects the
substrates, outside users of PAPI will not notice this change.
2011-01-25
* src/: configure, configure.in: Make sure that aix gets -g.
* src/: configure, configure.in: Give everyone else -g when
configuring with debug.
To wit, we pass gcc -g3 but neglected platforms where CC!=gcc.
* src/aix.c: First run at supporting power7. NOTE: this code is
only good for getting event listings eg papi_native_avail,
passing PM_GET_GROUPS causes our code to segfault later on, a
buffer overflow I'm still tracking down.
* src/perfctr-x86.c: Accidentally converted a function to _perfctr_
that should have stayed _linux_.
* src/: perfctr-x86.c, perfctr.c: Rename the various perfctr
functions to be _perfctr_ rather than _linux_. This way _linux_
is reserved for the common functions used by all.
* src/: linux-common.c, linux-memory.c, linux-timer.c,
perf_events.c, windows-common.c, windows-memory.c,
windows-timer.c: Split the WIN32 specific code out from the new
linux common code.
In most cases very little code was shared (it tended to be a big
#ifdef block) and it is confusing to have windows-specific code
in files named linux-*
2011-01-24
* src/linux-timer.c: Fix a compile error that only shows up on PPC.
* src/linux-timer.c: Fix compile warning if mmtimer is enabled.
* src/perfctr-x86.c: Missing comma in the perfctr code.
* src/: Makefile.inc, aix.c, configure, configure.in,
hwinfo_linux.c, linux-bgp.c, linux-common.c, linux-common.h,
linux-ia64.c, linux-timer.c, linux-timer.h, papi_vector.h,
perf_events.c, perfctr-x86.c, perfctr.c, perfmon.c,
solaris-niagara2.c, solaris-ultra.c: One last batch of
consolidation changes.
This one moves get_system_info and get_cpu_info into
linux-common.c, plus moves some other routines from perf_events.c
there that are shared by the future libpfm4 version.
Some non-linux substrates are touched here; these are just short
fixes to make sure the get_system_info() function pointed to by
the papi_vector has the same format on all substrates.
* src/: Makefile.inc, configure, configure.in, linux-memory.c,
linux-memory.h, perf_events.c, perfctr-x86.c, perfctr.c,
perfmon.c: Move the various Linux update_shlib_info() functions
into a common place.
* src/: Makefile.inc, linux-timer.c, linux-timer.h, perf_events.c,
perfctr-x86.c, perfctr.c, perfmon.c: Move the various
timer-related functions to linux-timer.c This gets rid of the
duplicated code spread throughout the substrates.
2011-01-21
* delete_before_release.sh, release_procedure.txt: Updated the
release docs with what I learned when making the 4.1.2.1 release.
* src/: configure, configure.in, freebsd-memory.c,
linux-ia64-memory.c, linux-memory.c, linux-memory.h,
linux-mx-memory.c, linux-ppc64-memory.c, perf_events.c,
perfctr-x86.c, perfmon-memory.c, perfmon.c: Currently there are
at least 3 identical copies of the linux memory detection code
spread throughout the PAPI source code.
This change puts them all in linux-memory.c, and then has all the
individual substrates use the common code.