derek.holloway
7502bf64a4
CodeQL / Analyze (actions) (push) Canceled after 0s
CodeQL / Analyze (python) (push) Canceled after 0s
pre-commit / pre-commit (push) Canceled after 0s
Multi-Arch CI / setup (push) Canceled after 0s
Publish build_manylinux_x86_64 images / publish_build_manylinux_x86_64 (push) Canceled after 0s
Setup (Multi-Arch) / setup (push) Canceled after 0s
Multi-Arch CI / Linux::${{ fromJSON(needs.setup.outputs.linux_build_config || '{}').build_variant_label || 'skip' }} (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Publish no_rocm_image_ubi10 image / publish_no_rocm_image_ubi10 (push) Canceled after 0s
Multi-Arch CI / Windows::${{ fromJSON(needs.setup.outputs.windows_build_config || '{}').build_variant_label || 'skip' }} (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Multi-Arch CI / Manifest Diff (push) Canceled after 0s
TheRock Manifest Diff Report / Generate Manifest Diff Report (push) Canceled after 0s
Multi-Arch CI / CI Summary (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04 (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04_ocl_rt images / publish_no_rocm_image_ubuntu24_04_ocl_rt (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_rocgdb (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04_rocgdb images / publish_no_rocm_image_ubuntu24_04_rocgdb (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_ocl_rt (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Unit Tests / Unit Tests :: ubuntu-24.04 (push) Canceled after 0s
Unit Tests / Unit Tests :: windows-2022 (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_media (push) Canceled after 0s
Unit Tests / Unit Tests Summary (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_openmpi (push) Canceled after 0s
702 lines
34 KiB
Plaintext
702 lines
34 KiB
Plaintext
2023-03-09 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/Makefile_comp_tests.target.in: tests: fix order of
|
|
headers in makefile The local includedir now has precedence over
|
|
the install includedir.
|
|
|
|
2023-02-14 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm_smi/rocs.c: rocm_smi: add support for XGMI
|
|
events Added events for XGMI on MI50 and MI100. Also support P2P
|
|
internode min and max bandwidth monitoring.
|
|
|
|
2023-03-06 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm_smi/tests/Makefile: rocm_smi: fix test warning
|
|
|
|
2023-02-14 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm_smi/rocs.c: rocm_smi: refactor component to
|
|
support XGMI events (3/3) Add infrastructure for supporting XGMI
|
|
events.
|
|
|
|
2023-02-07 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm_smi/Rules.rocm_smi, src/components/rocm_smi
|
|
/linux-rocm-smi.c: rocm_smi: refactor rocm_smi frontend to use rocs
|
|
API (2/3) Replace old linux-rocm-smi.c logic with calls to the
|
|
rocs layer interface.
|
|
|
|
2023-01-18 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm_smi/htable.h, src/components/rocm_smi/rocs.c,
|
|
src/components/rocm_smi/rocs.h: rocm_smi: refactor rocm_smi logic
|
|
into rocs backend (1/3) Refactors most of the code originally in
|
|
linux-rocm-smi.c by moving it to a new layer name rocs (for
|
|
ROCmSmi) and simplifying the original event detection logic.
|
|
|
|
2023-03-02 Daniel Barry <dbarry@vols.utk.edu>
|
|
|
|
* src/ftests/Makefile: build system: accommodate Fortran compiler
|
|
absence These changes introduce clean/clobber targets in the
|
|
ftests/Makefile to remove ftests/Makefile.target in the case that
|
|
the Fortran tests were not built. These changes were tested on
|
|
platform containing the AMD Zen4 architecture.
|
|
|
|
Sat Feb 25 18:01:45 2023 -0800 John Linford <jlinford@nvidia.com>
|
|
|
|
* src/libpfm4/README, src/libpfm4/docs/Makefile,
|
|
src/libpfm4/docs/man3/libpfm_arm_neoverse_v1.3,
|
|
src/libpfm4/docs/man3/libpfm_arm_neoverse_v2.3,
|
|
src/libpfm4/include/perfmon/pfmlib.h, src/libpfm4/lib/Makefile,
|
|
src/libpfm4/lib/events/arm_neoverse_n2_events.h,
|
|
src/libpfm4/lib/events/arm_neoverse_v1_events.h,
|
|
src/libpfm4/lib/events/arm_neoverse_v2_events.h,
|
|
src/libpfm4/lib/events/intel_skl_events.h,
|
|
src/libpfm4/lib/pfmlib_arm_armv8.c,
|
|
src/libpfm4/lib/pfmlib_arm_armv9.c,
|
|
src/libpfm4/lib/pfmlib_arm_priv.h, src/libpfm4/lib/pfmlib_common.c,
|
|
src/libpfm4/lib/pfmlib_priv.h, src/libpfm4/tests/validate_arm64.c:
|
|
libpfm4: update to commit c676419 Original commits: commit
|
|
c676419047f240468efd63407cf5e3fefa71752a update Intel SKL/SKX/CLX
|
|
event table Based on github.com/Intel/perfmon/SKX version 1.29
|
|
commit 098a39459fa0d0ed1d81f4c269a3b0ece46f9f27 add ARM Neoverse
|
|
V2 core PMU support Based on information from: github.com/ARM-
|
|
software/data/blob/master/pmu/neoverse-v2.json commit
|
|
1307e234db0f3922d6854e9b84283c5f6c72d2d6 move ARM Neoverse N2 to
|
|
ARMv9 support Neoverse N2 is a ARMv9 implementation therefore it
|
|
needs to be moved to the pfmlib_arm_armv9.c support file.
|
|
Attributes are also updated to point to the V9 specific version.
|
|
commit 61b49e0bbcc0906c54c17007faca91d0c62e6b38 add ARM v9
|
|
support basic infrastructure Adds the pmlib_arm_armv9.c support
|
|
file and a few macros definitions to enable ARMv9 PMU support.
|
|
commit 21895bae4e59936079b908c08787aa63fe485141 add Arm Neoverse
|
|
V1 core PMU support This patch adds support for Arm Neoverse V1
|
|
core PMU. Based on Arm information posted on github.com/ARM-
|
|
software/data/blob/master/pmu/neoverse-v1.json
|
|
|
|
2023-03-07 Anthony <adanalis@icl.utk.edu>
|
|
|
|
* src/components/sde/tests/Makefile,
|
|
src/components/sde/tests/Simple2/Simple2_Lib++.cpp,
|
|
src/components/sde/tests/Simple2/Simple2_Lib.c: Modified a non-
|
|
compliant type aliasing code that gcc-12.2 treats as undefined
|
|
behavior to conform to the C standard and be more portable.
|
|
|
|
2023-02-27 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/configure, src/configure.in: rocm: define PAPI_ROCM_PROF if
|
|
rocm component enabled
|
|
* src/components/sysdetect/tests/Makefile: sysdetect: enable fortran
|
|
tests rules only if F77 is set
|
|
|
|
Wed Feb 22 23:00:00 2023 -0800 Stephane Eranian <eranian@gmail.com>
|
|
|
|
* src/libpfm4/lib/events/intel_icl_events.h,
|
|
src/libpfm4/lib/events/intel_spr_events.h: libpfm4: update to
|
|
commit b4361ca Original commits: commit
|
|
b4361ca023198b9a96f1d824cfcd276f020bcac3 Update Intel
|
|
SapphireRapid event table Based on github.com/intel/perfmon
|
|
version 1.11 commit f31c0f5ff0792d547eff436c577eac82d99b4e8b
|
|
update Intel Icelake event table Based on githib.com: - v1.19 for
|
|
IcelakeX - v1.17 for Icelake
|
|
|
|
2023-02-08 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/rocp.c: rocm: fix intercept mode shutdown bug
|
|
(2/2) Shutdown hash table in intercept mode path.
|
|
* src/components/rocm/rocm.c: rocm: fix init_private return code bug
|
|
(1/2) Check init_private return state in delay initialization
|
|
functions (e.g. rocm_update_control_state).
|
|
|
|
2023-02-20 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/Rules.rocm, src/components/rocm/htable.c,
|
|
src/components/rocm/htable.h: rocm: refactor htable (28/28) The
|
|
htable data structure is used across multiple components and having
|
|
it in a C file causes multiple definition errors. Instead move the
|
|
implementation to the header and declare all function static
|
|
inline.
|
|
|
|
2023-01-26 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/Rules.rocm, src/components/rocm/common.h,
|
|
src/components/rocm/rocd.c, src/components/rocm/rocd.h,
|
|
src/components/rocm/rocm.c, src/components/rocm/rocp.c,
|
|
src/components/rocm/rocp.h, src/configure, src/configure.in: rocm:
|
|
add dispatch layer for future extensions (27/28) Add dispatch
|
|
layer for accomodating the integration of additional backend
|
|
profiling tools (e.g. rocmtools).
|
|
|
|
2023-01-23 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/common.h, src/components/rocm/rocp.h: rocm:
|
|
refactor rocp backend header (26/28) Move context state defines to
|
|
rocp.h
|
|
|
|
2023-01-10 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/rocp.c: rocm: add note to source code (25/28)
|
|
Add fixme note to intercept code
|
|
|
|
2023-01-09 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/rocp.c: rocm: remove same-event-number-per-
|
|
device limitation (24/28) The rocm component used to make some
|
|
assumption when working in sampling mode. More specifically, in the
|
|
case of a single eventset containing events from different devices,
|
|
the component assumed every device had the same number of events.
|
|
This is a reasonable assumption to make because in the typical case
|
|
the user has a SIMD workload split across available devices. It
|
|
makes sense therefore to monitor the same events on all devices, in
|
|
order to make apple to apple comparisons. This assumption, however,
|
|
does not allow for the case in which the user wants to monitor
|
|
different events on different devices. This might be the case for a
|
|
MIMD workload. This patch removes the limitation by allowing
|
|
different number of events to be monitored on different devices
|
|
using a single eventset.
|
|
|
|
2023-01-05 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/rocm.c, src/components/rocm/rocp.c: rocm: sort
|
|
events in component frontend (23/28) The rocp layer needs events
|
|
sorted by device. The sorting used to happen in the rocp layer and
|
|
counters would be eventually assigned to the correct position in
|
|
the user counter array. The same mechanism is already available in
|
|
the frontend through the ntv_info ni_position attribute. Thus, do
|
|
the sorting the in the rocm layer and set the ni_position to the
|
|
remapped event/counter. This allows to simplify the rocp layer
|
|
logic.
|
|
|
|
2023-01-04 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/rocp.c, src/components/rocm/rocp.h: rocm: add
|
|
comments for backend functions (22/28) Explicitly separate
|
|
interfaces by functionality in rocp.h and add descriptions in
|
|
rocp.c
|
|
|
|
2022-12-23 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/rocp.c: rocm: refactor static string lengths
|
|
(21/28) Use PATH_MAX instead of PAPI_MAX_STR_LEN for paths
|
|
|
|
2022-12-19 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/rocp.c: rocm: refactor shutdown function names
|
|
in backend (20/28) Rename shutdown funcs in rocp layer
|
|
|
|
2022-12-16 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/rocp.c: rocm: refactor event verification logic
|
|
in intercept mode (19/28) The verify_events function makes sure
|
|
that, in intercept mode, all eventsets have the same events. This
|
|
is dictated by rocprofiler that, currently, does not allow
|
|
resetting intercept mode callbacks. The way this check is carried
|
|
out is by going through a list of intercept events (kept internally
|
|
by the rocp layer) and a list of user requested events. If the two
|
|
differ, then there is a conflict and the new eventset cannot be
|
|
monitored. Use the htable to log the name of the intercept events
|
|
and check the presence/absence for conflict in verify_events.
|
|
|
|
2022-12-15 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/rocm.c: rocm: refactor event duplication
|
|
verification logic (18/28) Currently, update_native_events removes
|
|
event duplication whenever that happens. This is the case for
|
|
events that differ by instance number. The rocp layer reports
|
|
events instances as separate native events. The logic to remove
|
|
duplicate events is messy and probably not even needed as the user
|
|
will normally add events without indicating the instance of the
|
|
event. This patch removes such logic.
|
|
|
|
2022-12-14 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/rocp.c: rocm: refactor dispatch_counter logic
|
|
(17/28) dispatch_counter is used to keep track of what kernel has
|
|
been dispatched by what thread. The current implementation relies
|
|
on an array to keep the tid and the counter value. Using the hash
|
|
table, currently also used for keeping events, simplifies the code
|
|
significantly.
|
|
|
|
2022-12-13 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/rocm.c, src/components/rocm/rocp.c,
|
|
src/components/rocm/rocp.h: rocm: refactor rocp_ctx_open function
|
|
name (16/28) Rename rocp_ctx_open_v2 to rocp_ctx_open
|
|
|
|
2022-12-09 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/common.h, src/components/rocm/rocm.c,
|
|
src/components/rocm/rocp.c, src/components/rocm/rocp.h: rocm:
|
|
handle event table in component backend (15/28) No longer
|
|
initialize and use the event table in the rocm layer. No longer
|
|
pass the ntv_table around in the rocp layer.
|
|
* src/components/rocm/rocm.c: rocm: add event counting function in
|
|
component frontend (14/28) Add evt_get_count function to count
|
|
events
|
|
* src/components/rocm/rocm.c: rocm: use rocp API to enumerate events
|
|
in component (13/28) Do not use ntv_table for enum events in
|
|
rocm.c Instead of using ntv_table for enum events use the new rocp
|
|
exposed interfaces such as rocp_evt_code_to_name, etc.
|
|
* src/components/rocm/rocm.c: rocm: add event name tokenizer (12/28)
|
|
Add tokenize_event_string function to extract device information
|
|
from the event name.
|
|
* src/components/rocm/rocm.c, src/components/rocm/rocp.c,
|
|
src/components/rocm/rocp.h: rocm: get errors through
|
|
rocp_err_get_last (11/28) Instead of returning error string
|
|
explicitly during rocp_init/_environment, use rocp_err_get_last.
|
|
* src/components/rocm/rocm.c, src/components/rocm/rocp.c,
|
|
src/components/rocm/rocp.h: rocm: add error code to string function
|
|
(10/28) Add error string return function: rocp_err_get_last().
|
|
|
|
2022-12-08 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/rocm.c, src/components/rocm/rocp.c,
|
|
src/components/rocm/rocp.h: rocm: refactor rocp_ctx_open function
|
|
(9/28) The goal of this patch is to remove the need for passing
|
|
the event table reference to the rocp_ctx_open function. To avoid
|
|
too many changes in the code here we introduce a new
|
|
rocp_ctx_open_v2 function and replace rocp_ctx_open with it in a
|
|
later commit.
|
|
* src/components/rocm/rocp.c, src/components/rocm/rocp.h: rocm:
|
|
refactor component backend interface (8/28) This patch does some
|
|
preparatory work to make the rocp layer completely opaque to the
|
|
component layer (rocm.c). These include storing a reference to the
|
|
native table built at rocp_init time, and adding four new
|
|
interfaces for enumerating events, getting event descriptors,
|
|
converting event names to codes and viceversa.
|
|
|
|
2022-12-06 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/rocp.c: rocm: refactor user to component event
|
|
mapping (7/28) The rocp_ctx already contains a reference to the
|
|
events_id provided by the user. Remove explicit reference in
|
|
get_user_counter_id arg list.
|
|
* src/components/rocm/rocp.c: rocm: refactor event sorting function
|
|
name (6/28) Rename sort_events_by_device to sort_events. Events
|
|
are numbered starting from the first device to the last. Thus,
|
|
events are always ordered by device, and the name of the function
|
|
is a redundant statement of how events are sorted.
|
|
* src/components/rocm/rocp.c: rocm: refactor event collision
|
|
detection (5/28) Rename compare_events to verify_events to
|
|
indicate the change in functionality for the function. Used to
|
|
compare events, returning either 0 if they all matched or an
|
|
integer if the did not. Now verify_events returns PAPI_OK if the
|
|
events match and PAPI_ECNFLT if they don't. verify_events also
|
|
checks whether the rocprofiler callbacks are set or not. If not the
|
|
function exits immediately as there cannot be any event conflict.
|
|
* src/components/rocm/rocp.c: rocm: refactor intercept_ctx_open
|
|
(4/28) Move rocprofiler init_callbacks from intercept_ctx_open
|
|
into ctx_init.
|
|
* src/components/rocm/rocp.c: rocm: quick sort component events by id
|
|
(3/28) Events from multiple GPUs can appear in any order in the
|
|
eventset. However, rocprofiler needs events to be ordered by
|
|
device. Previously, we were sorting events by device using a brute
|
|
force approach. This is unnecessary because events are numbered
|
|
according to device order anyway. Doing a quick sort of the events
|
|
identifiers is sufficient.
|
|
|
|
2022-12-05 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/rocm.c, src/components/rocm/rocp.c,
|
|
src/components/rocm/rocp.h: rocm: refactor rocp_ctx_read (2/28)
|
|
The events_id array, containing the id of the events requested by
|
|
the user, is passed to rocp_ctx_open and can be saved in the
|
|
rocp_ctx returned by this function. It is not necessary to pass the
|
|
array again as argument to rocp_ctx_read. Thus, remove it from the
|
|
argument list of rocp_ctx_read.
|
|
|
|
2022-12-13 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/common.h, src/components/rocm/rocm.c,
|
|
src/components/rocm/rocp.c, src/components/rocm/rocp.h: rocm:
|
|
refactor variable types (1/28) Variable type refactoring. Use
|
|
unsigned int for ids (e.g. events_id and devs_id) and int for
|
|
counts (e.g. num_events, num_devs).
|
|
|
|
2023-02-21 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/perf_event/perf_helpers.h: perf_event: used unused
|
|
attribute in mmap_read_self
|
|
* src/components/perf_event/perf_helpers.h: perf_event: add missing
|
|
mmap_read_reset_count for non default cpus Power cpus do not have
|
|
a version of mmap_read_reset_count. Implement the missing function.
|
|
* src/components/perf_event/perf_helpers.h: perf_event: bug fix in
|
|
mmap_read_self Commit 9a1f2d897 broke the perf_event component for
|
|
power cpus. The mmap_read_self function is missing one argument.
|
|
This patch restores the missing argument in the function.
|
|
|
|
2023-01-26 Daniel Barry <dbarry@vols.utk.edu>
|
|
|
|
* src/components/nvml/README.md, src/components/nvml/linux-nvml.c:
|
|
nvml: fix support for multiple devices Replace each
|
|
cudaGetDevicePtr() call with a table lookup. This tracks the
|
|
correct device ID when counting events; whereas, cudaGetDevicePtr()
|
|
will only ever return a single ID with the way it was used. The
|
|
dependency on CUDA unnecessarily restricts multi-process jobs on
|
|
Summit. Removing this dependency was necessary to properly support
|
|
multiple devices. These changes were tested on the Summit
|
|
supercomputer, which contains the IBM POWER9 and NVIDIA Tesla V100
|
|
architectures.
|
|
|
|
2023-02-21 Masahiko, Yamada <yamada.masahiko@fujitsu.com>
|
|
|
|
* src/components/perf_event/perf_event.c,
|
|
src/components/perf_event/perf_event_lib.h,
|
|
src/components/perf_event/perf_helpers.h: PAPI_read performance
|
|
improvement for the arm64 processor We developed PAPI_read
|
|
performance improvements for the arm64 processor with a plan to
|
|
port direct user space PMU register access processing from libperf
|
|
to the papi library without using libperf. The workaround has been
|
|
implemented that stores the counter value at the time of reset and
|
|
subtracts the counter value at the time of reset from the read
|
|
counter value at the next read. When reset processing is called,
|
|
the value of pc->offset is cleared to 0, and only the counter value
|
|
read from the PMU counter is referenced. There was no problem with
|
|
the counters FAILED with negative values during the multiplex+reset
|
|
test, except for sdsc2-mpx and sdsc4-mpx. To apply the workaround
|
|
only during reset, the _pe_reset function call sets the reset_flag
|
|
and the next _pe_start function call clears the reset_flag. The
|
|
workaround works if the mmap_read_self function is called between
|
|
calls to the _pe_reset function and the next call to the _pe_start
|
|
function. Switching PMU register direct access from user space
|
|
from OFF to ON is done by changing the setting of the kernel
|
|
variable "/proc/sys/kernel/perf_user_access". Setting PMU Register
|
|
Direct Access from User Space Off $ echo 0 >
|
|
/proc/sys/kernel/perf_user_access $ cat
|
|
/proc/sys/kernel/perf_user_access 0 Setting PMU Register Direct
|
|
Access from User Space ON $ echo 1 >
|
|
/proc/sys/kernel/perf_user_access $ cat
|
|
/proc/sys/kernel/perf_user_access 1 Performance of PAPI_read has
|
|
been improved as expected from the execution result of the
|
|
papi_cost command. Improvement effect of switching PMU register
|
|
direct access from user space from OFF to ON Total cost for
|
|
PAPI_read (2 counters) over 1000000 iterations min cycles: 689
|
|
-> 28 max cycles: 3876 -> 1323 mean cycles: 724.471979 ->
|
|
28.888076 Total cost for PAPI_read_ts (2 counters) over 1000000
|
|
iterations min cycles: 693 -> 29 max cycles: 4066
|
|
-> 3718 mean cycles: 726.753003 -> 29.977226 Total cost for
|
|
PAPI_read (1 derived_[add|sub] counter) over 1000000 iterations min
|
|
cycles: 698 -> 28 max cycles: 7406 -> 2346 mean
|
|
cycles: 728.527079 -> 28.880691
|
|
|
|
Sun Feb 5 22:56:09 2023 -0800 Stephane Eranian <eranian@gmail.com>
|
|
|
|
* src/libpfm4/lib/events/amd64_events_fam19h_zen4.h: libpfm4: update
|
|
to commit 678bca9 Original commits: commit
|
|
678bca9bf803b089c089629661d457533a7705b0 Update AMD Zen4 event
|
|
table - Fix wrong encodings in for event RETIRED_FP_OPS_BY_TYPE -
|
|
Fix INT256_OTHER bogus name - Add missing
|
|
RETIRED_UCODE_INSTRUCTIONS - Fix Name and descripiton for event
|
|
RETIRED_UNCONDITIONAL_INDIRECT_BRANCH_INSTRUCTIONS_MISPREDICTED
|
|
commit dcb2f5e73d0343c87995919495c3c10252a7b0ca remove useless
|
|
combination in AMD Zen4 packed_int_ops_retired event This
|
|
combination is useless and does not match the rest of the logic for
|
|
this event. libpfm4 allows one umask at a time.
|
|
|
|
2023-01-26 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/ctests/all_native_events.c: ctests/all_native_events: bug
|
|
workaround Sampling mode fails in the presence of more than one
|
|
rocm GPU. This is due to a bug in the rocprofiler library. To avoid
|
|
the failure in the all_native_events test we skip rocm tests.
|
|
|
|
2023-02-07 Daniel Barry <dbarry@vols.utk.edu>
|
|
|
|
* src/components/Makefile_comp_tests.target.in,
|
|
src/components/cuda/sampling/Makefile,
|
|
src/components/cuda/tests/BlackScholes/Makefile,
|
|
src/components/cuda/tests/Makefile,
|
|
src/components/nvml/tests/Makefile,
|
|
src/components/nvml/utils/Makefile, src/configure,
|
|
src/configure.in: build system: workaround for GCC8+CUDA10 bug on
|
|
POWER9 When using CUDA 10 with GCC 8 on IBM POWER9, the following
|
|
compile-time error occurs with 'nvcc': > error: identifier
|
|
"__ieee128" is undefined We work around this issue by passing the
|
|
flags "-Xcompiler -mno-float128" to 'nvcc' invocations. These
|
|
changes have been tested on the ppc64le architecture and NVIDIA
|
|
Tesla V100 GPUs.
|
|
* src/components/sysdetect/nvidia_gpu.c,
|
|
src/components/sysdetect/nvidia_gpu.h: sysdetect: account for older
|
|
CUDA versions In CUDA 11.0 or greater, the macro
|
|
"NVML_DEVICE_UUID_V2_BUFFER_SIZE" is defined. Older versions of
|
|
CUDA define "NVML_DEVICE_UUID_BUFFER_SIZE." In order to support
|
|
older versions of CUDA, these changes apply the appropriate macro.
|
|
These changes have been tested on the NVIDIA Tesla V100
|
|
architecture.
|
|
|
|
2023-01-26 Daniel Barry <dbarry@vols.utk.edu>
|
|
|
|
* src/components/nvml/README.md: nvml: fix small typo in README
|
|
Remove extra underscore in the README.md file.
|
|
|
|
2023-01-13 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/tests/sample_multi_thread_monitoring.cpp: rocm:
|
|
skip sampling multi-thread tests Sampling mode tests fail because
|
|
of a still unresolved bug in rocm-5.3. Skip them until the bug is
|
|
resolved.
|
|
|
|
2023-01-09 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/rocp.c: rocm: fix bug in sampling_ctx_open The
|
|
function creates one profiling context per device. The way the
|
|
agent corresponding to the device is selected was erroneous. This
|
|
caused different threads monitoring different devices with
|
|
different eventsets to all access the counters from the first
|
|
device. The fix is not to select the agent using a for loop index
|
|
but instead to use that index to get the device id from the devs_id
|
|
array.
|
|
|
|
Thu Jan 5 12:29:51 2023 -0800 Giuseppe Congiu <gcongi@icl.utk.edu>
|
|
|
|
* src/libpfm4/lib/pfmlib_amd64_fam19h.c: libpfm4: update to commit
|
|
dd42292 Original commits: commit
|
|
dd422923f79a6c160e499f484212020ca2398f90 Fix AMD Zen4 cpu_family
|
|
used in detection code AMD Zen4 was expecting Zen3 CPU family
|
|
number.
|
|
|
|
2022-12-22 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/net/linux-net.c: net: fix warning in strncpy The
|
|
source and target string have the same length. If the source is
|
|
null terminated (as expected in the absence of bugs) there is no
|
|
need to null terminate the target manually.
|
|
* src/components/net/linux-net.c: net: fix warning in snprintf
|
|
Compute the length of the source string instead of copying
|
|
PAPI_MAX_STR_LEN characters regardless.
|
|
* src/components/coretemp/linux-coretemp.c: coretemp: fix warning in
|
|
strncpy Source and target are the same length. No need to null
|
|
terminate if the source is already null terminated.
|
|
* src/components/powercap_ppc/tests/powercap_basic.c: powercap_ppc:
|
|
fix warning in powercap_basic test Copy the whole source string to
|
|
the target.
|
|
* src/components/powercap_ppc/tests/powercap_basic.c: powercap_ppc:
|
|
fix bug in powercap_basic test Make target and source strings the
|
|
same size.
|
|
|
|
2022-12-20 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/libpfm4/docs/man3/libpfm_amd64_fam19h_zen4.3,
|
|
src/libpfm4/lib/events/amd64_events_fam19h_zen4.h: libpfm4: add
|
|
missing zen 4 files Commit 2fe62da left out additional zen 4
|
|
files. This patch adds them.
|
|
|
|
2022-12-15 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/sensors_ppc/linux-sensors-ppc.c: sensors_ppc: fix
|
|
typo in fall through comment The compiler throws a warning because
|
|
it does not recognize fallthrough as valid indication that the code
|
|
can fall through in the case statement.
|
|
* src/components/powercap_ppc/linux-powercap-ppc.c: powercap-ppc: fix
|
|
warning in strncpy strncpy causes the following warning in linux-
|
|
powercap-ppc.c: warning: '__builtin_strncpy' specified bound 1024
|
|
equals destination size 62 | char *retval = strncpy( dst, src,
|
|
size ); | ^~~~~~~ the problem is that size is
|
|
the same size as dst. If src is also the same length, the null
|
|
termination character will not be copied over. Instead copy only
|
|
size - 1 and terminate the string manually.
|
|
|
|
Fri Dec 2 00:01:47 2022 -0800 Stephane Eranian <eranian@gmail.com>
|
|
|
|
* src/libpfm4/README, src/libpfm4/docs/Makefile,
|
|
src/libpfm4/include/perfmon/pfmlib.h, src/libpfm4/lib/Makefile,
|
|
src/libpfm4/lib/events/intel_icl_events.h,
|
|
src/libpfm4/lib/pfmlib_amd64.c,
|
|
src/libpfm4/lib/pfmlib_amd64_fam19h.c,
|
|
src/libpfm4/lib/pfmlib_common.c, src/libpfm4/lib/pfmlib_priv.h,
|
|
src/libpfm4/tests/validate_x86.c: libpfm4: update libpfm4 to commit
|
|
c0116f9 Original commits: commit
|
|
c0116f9433f34e5953407036243af998c00fcc1f Add AMD Zen4 core PMU
|
|
support Based on AMD PPR for Fam19h model 11 B1 rec 0.25 commit
|
|
11f94169598b84a71fb9da4357baf3673f83038b Correctly detect all AMD
|
|
Zen3 processors Fixes commit 79031f76f8a1 ("fix amd_get_revision()
|
|
to identify AMD Zen3 uniquely") The commit above broke the
|
|
detection of certain AMD Zen3, such as: Vendor ID:
|
|
AuthenticAMD BIOS Vendor ID: Advanced Micro Devices, Inc.
|
|
Model name: AMD Ryzen 9 5950X 16-Core Processor BIOS
|
|
Model name: AMD Ryzen 9 5950X 16-Core Processor BIOS CPU
|
|
family: 107 CPU family: 25 Model: 33
|
|
commit c5100b69add67172366e897cef5b854c5348dc91 fix
|
|
CPU_CLK_UNHALTED.REF_DISTRIBUTED on Intel Icelake Had the wrong
|
|
encoding of 0x8ec instead of 0x083c.
|
|
|
|
2022-12-19 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/utils/papi_hardware_avail.c: sysdetect: fix typo in
|
|
papi_hardware_avail
|
|
|
|
2022-12-01 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/rocp.c: rocm: give precendence to new dir
|
|
structure for metrics file
|
|
|
|
2022-11-28 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/rocp.c: rocm: account for new directory tree
|
|
structure in rocm librocprofiler64.so is being moved from
|
|
<rocm_root>/rocprofiler/lib to <rocm_root>/lib. This patch allow
|
|
the rocm component to search in the new location if the old is
|
|
empty/non-existant.
|
|
|
|
2022-11-30 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/ftests/clockres.F, src/ftests/fmatrixlowpapi.F: ftest: fix
|
|
warning in fortran tests Arrays that are statically allocated
|
|
cannot be placed in the stack if the exceed a certain size (-fmax-
|
|
stack-var-size). The compilers moves them into static storage which
|
|
might cause problems if the function is called recursively. Solve
|
|
the warning by allocating the arrays dynamically.
|
|
|
|
2022-12-08 Daniel Barry <dbarry@vols.utk.edu>
|
|
|
|
* src/components/infiniband/linux-infiniband.c: infiniband: fix
|
|
compiler warnings Recent versions of GCC (9.3.0 in this case)
|
|
threw the following warnings: components/infiniband/linux-
|
|
infiniband.c: In function '_infiniband_ntv_code_to_info':
|
|
components/infiniband/linux-infiniband.c:937:9: warning: 'strncpy'
|
|
specified bound depends on the length of the source argument
|
|
[-Wstringop-overflow=] 937 | strncpy(info->symbol,
|
|
infiniband_native_events[index].name, len); |
|
|
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
|
components/infiniband/linux-infiniband.c:935:28: note: length
|
|
computed here 935 | unsigned int len =
|
|
strlen(infiniband_native_events[index].name); |
|
|
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ components/infiniband
|
|
/linux-infiniband.c:944:9: warning: 'strncpy' specified bound
|
|
depends on the length of the source argument [-Wstringop-overflow=]
|
|
944 | strncpy(info->long_descr,
|
|
infiniband_native_events[index].description, len); | ^~~~~~
|
|
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
|
~~ components/infiniband/linux-infiniband.c:942:28: note: length
|
|
computed here 942 | unsigned int len =
|
|
strlen(infiniband_native_events[index].description); |
|
|
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ The changes in
|
|
this commit fix these warnings by using the maximum possible length
|
|
of the source arguments. These changes were tested on Summit,
|
|
which has the following IB device listing from 'lspci': Infiniband
|
|
controller: Mellanox Technologies MT28800 Family [ConnectX-5 Ex].
|
|
|
|
2022-11-29 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/sysdetect/Rules.sysdetect: sysdetect: fix order in
|
|
include dirs
|
|
* src/components/sysdetect/Rules.sysdetect: sysdetect: fix typo in
|
|
makefile Rules
|
|
* src/components/sysdetect/amd_gpu.c: sysdetect: explicit cast AMD
|
|
hsa attributes to hsa_agent_info_t
|
|
|
|
2022-11-28 Florian Weimer <fweimer@redhat.com>
|
|
|
|
* src/configure, src/configure.in: configure: Avoid implicit ints and
|
|
implicit function declarations Implicit ints and implicit function
|
|
declarations were removed from the C language in 1999. Relying on
|
|
them can cause spurious autoconf check failures with compilers that
|
|
do not support them in the default language mode.
|
|
|
|
2022-12-02 Daniel Barry <dbarry@vols.utk.edu>
|
|
|
|
* src/components/infiniband/linux-infiniband.c: infiniband: increase
|
|
max number of events The maximum number of events
|
|
('INFINIBAND_MAX_COUNTERS') was hard-coded to be 128. However,
|
|
some Infiniband devices provide more than 128 events, causing the
|
|
component test 'infiniband_values_by_code' to seg fault. The IB
|
|
devices on Summit nodes provide 188 events, so the macro needs to
|
|
be greater than or equal to this number. These changes were tested
|
|
on Summit, which has the following IB device listing from 'lspci':
|
|
Infiniband controller: Mellanox Technologies MT28800 Family
|
|
[ConnectX-5 Ex].
|
|
|
|
2022-05-26 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/cuda/Rules.cuda: cuda: remove untested dependency
|
|
linux-cuda.o depends on cuda_sampling (directory). This contains
|
|
untested code and does not seem to be an indispensable dependency
|
|
for the cuda component. This patch removes the cuda_sampling
|
|
dependency for now.
|
|
|
|
2022-11-29 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/sysdetect/sysdetect.c: sysdetect: add missing numa
|
|
memsize attr
|
|
* src/components/cuda/linux-cuda.c: cuda: use appropriate macro for
|
|
perfworks API calls Two perfworks API calls were made using
|
|
CUPTI_CALL macro instead of NVPW_CALL. This patch uses the
|
|
appropriate call macro.
|
|
* src/components/rocm_smi/Rules.rocm_smi: rsmi: fix warning in
|
|
Makefile Rules This warning shows up for rocm version 5.2 and
|
|
later, which changed the directory structure and deprecated headers
|
|
for the old structure. This patch prioritizes the new structure
|
|
when looking for rocm_smi headers at build time.
|
|
* src/components/rocm_smi/linux-rocm-smi.c: rsmi: fix warning in
|
|
strncpy strncpy warning was caused by the len of the copy being
|
|
equal to the target string len. Increasing the target string by one
|
|
character leaves space for a termination character and fixes the
|
|
warning.
|
|
* src/components/rocm/Rules.rocm: rocm: fix dependency priority in
|
|
Makefile Recent versions of ROCM have deprecated the old directory
|
|
tree in favour of a different organization of headers and
|
|
libraries. This patch gives priority to the new directory structure
|
|
when searching for headers.
|
|
* src/components/rocm/tests/Makefile: rocm: fix deprecated warning in
|
|
tests
|
|
|
|
2022-11-28 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/rocm/rocp.c: rocm: fix miscellaneous warnings in
|
|
rocp.c Fix following warnings: - implicit cast from enum
|
|
<anonymous> to rocprofiler_feature_kind_t: this is caused by
|
|
rocprofiler and is solved by explicit cast of
|
|
ROCPROFILER_FEATURE_KIND_METRIC to rocprofiler_feature_kind_t; -
|
|
casting 'getpid' to (unsigned long (*)(void)) incompatible type:
|
|
this is solved by using '_papi_getpid' instead of 'getpid'.
|
|
|
|
2022-12-01 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/nvml/linux-nvml.c: nvml: also copy null termination
|
|
character in strncpy String literals are null terminated. When
|
|
using strncpy the number of characters to be copied has to be the
|
|
string length plus 1 in order to include the null termination
|
|
character. Since the null termination is already included in the
|
|
string literal, manually terminating the string is superflous.
|
|
|
|
2022-11-29 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/nvml/linux-nvml.c: nvml: fix warning in strncpy The
|
|
code defines the source string twice as long as the target string
|
|
in number of characters. This causes the warning. The warning is
|
|
removed by making the target string PAPI_MAX_STR_LEN long and the
|
|
source string PAPI_MIN_STR_LEN long.
|
|
* src/components/nvml/linux-nvml.c: nvml: trim excessive long
|
|
description string An excessive long description string for power
|
|
managemenent upper bound limit does not fit into the 128 characters
|
|
of PAPI_MAX_STR_LEN. This patch trims the description string to
|
|
make it fit in the description string limit.
|
|
* src/components/nvml/linux-nvml.c: nvml: fix cuInit returned
|
|
variable type cuInit return result is of type CUresult and not
|
|
cudaError_t. Fix the error to resolve the warning: implicit
|
|
conversion from CUresult to cudaError_t.
|
|
|
|
2022-11-22 Giuseppe Congiu <gcongiu@icl.utk.edu>
|
|
|
|
* src/components/cuda/linux-cuda.c: cuda: fix compile error with gcc
|
|
10 Some of the symbols used by the cuda component clash with the
|
|
nvml component as they are not defined static. This problem only
|
|
affects latest versions of the gcc compiler (>= 10). This is due to
|
|
how gcc places global variables that do not have an initializer
|
|
(tentative definition variables in the C standard). These variables
|
|
are placed in the .BSS section of the object file. This avoids the
|
|
merging of tentative definition variables by the linker, which
|
|
causes multiple definition errors. (This happens by default in gcc
|
|
and corresponds to the -fno-common option).
|
|
* src/components/nvml/linux-nvml.c: nvml: fix compile error with gcc
|
|
10 Some of the symbols used by the nvml component clash with the
|
|
cuda component as they are not defined static. This problem only
|
|
affects latest versions of the gcc compiler (>= 10). This is due to
|
|
how gcc places global variables that do not have an initializer
|
|
(tentative definition variables in the C standard). These variables
|
|
are placed in the .BSS section of the object file. This avoids the
|
|
merging of tentative definition variables by the linker, which
|
|
causes multiple definition errors. (This happens by default in gcc
|
|
and corresponds to the -fno-common option).
|