likwid

• Fix pin string parsing in pinning library
• Make SBIN path configurable in build system
• Add PKGBUILD for ArchLinux package builds
• Remove accessDaemon double-fork in systemd environments
• Group updates for L2/L3 (mainly AMD Zen)
• Fix multi-initialization in MarkerAPI
• Add energy event scaling for Fujitsu A64FX
• Nvmon: Use Cupti error string to get better warning/error messages
• Nvmon: Store events internally to re-use event strings in stopCounters
• AccessLayer: Catch SIGCHLD to stop sending requests to accessDaemon if it was killed
• likwid-genTopoCfg: Update writing and reading of topology file
• Add INST_RETIRED_NOP event for Intel Icelake (desktop & server)
• Removed some memory leaks
• Improved checks for RDPMC availability
• Add TOPDOWN_SLOTS for perf_event
• Fix for systems with CPU sockets without hwthreads (A64FX FX1000)
• Fix if HOME environment variable is not set (systemd)
• Reader function for perf_event_paranoid in Lua to get state early
• likwid-mpirun: Sanitize np and ppn values to avoid crashes

Note: The groups MEM_DP and MEM_SP use only 6 of 8 memory controllers for Intel Icelake SP. The attached patch fixes both groups.

We are happy to release a new bugfix version of the LIKWID tool suite.

• Add support for Intel Rocketlake and AMD Zen3 variant (Family 19, Model 0x50)
• Fix for perf_event multiplexing (important!)
• Fix for potential deadlock in MarkerAPI (thx @jenny-cheung)
• Build and runtime fixes for Nvidia GPU backend, updates for CUDA test codes
• peakflops kernel for ARMv8
• Updates for AMD Zen1/2/3 event lists and groups
• Support spaces in MarkerAPI region tags (thx @jrmadsen)
• Use 'online' cpulist instead of 'present'
• Switch CI from Travis-CI to NHR@FAU Cx services
• Document -reset and -ureset for likwid-setFrequencies
• Reset cpuset in unpinned runs
• Remove destructor in frequency module
• Check PID if given through --perfpid
• Intel Icelake: OFFCORE_RESPONSE events
• AccessDaemon: Check PCI init state before using it
• likwid-mpirun: Set mpi type for SLURM automatically
• likwid-mpirun: Fix for skip mask for OpenMPI
• Fix for triad_sve* benchmarks

Note: The groups MEM_DP and MEM_SP use only 6 of 8 memory controllers for Intel Icelake SP. The attached patch fixes both groups.

We are happy to release a new major update of the LIKWID tool suite.

• Support for AMD Zen3 (Core + Uncore)
• Support for Intel IcelakeSP (Core + Uncore)
• New affinity code
• Fix for Ivybridge uncore code
• Bypass accessdaemon by using rdpmc instruction on x86_64
• Introduce notion of CPU die in topology module
• Use CPU dies for socket-lock for Intel CascadelakeAP
• Add environment variable LIKWID_IGNORE_CPUSET to break out of current CPUset
• Fixes for affinity module CPUlist sorting
• Build against system-installed hwloc
• Update for Intel SkylakeX/CascadelakeX L3 group
• Rename DataFabric events for all generations of AMD Zen
• Add static cache configuration for Fujitsu A64FX
• Add multiplexing checks for perf_event backend
• Fix for table width of likwid-topology after adding CPU die column
• Adding RasPi 4 with 32 bit OS as ARMv7
• Add default groups for Intel Icelake desktop
• Fix for likwid-setFrequencies to not apply minFreq when setting governor
• likwid-powermeter: Fix hwthread selection when run with -p
• likwid-setFrequencies: Get measured base frequency if register is not readable
• CLOCK group for all AMD Zen
• Fixes in Nvidia GPU support in NvMarkerAPI and topology module

WARNING: This version has bugs in the perf_event backend. The multiplexing checks cause problems.
WARNING: The benchmarks triad_sve* for ARM8 chips use only 3 instead of 4 streams.
Note: The groups MEM_DP and MEM_SP use only 6 of 8 memory controllers for Intel Icelake SP. The attached patch fixes both groups.

Changelog for version 5.1.1:

• Support for Intel Cometlake desktop (Core + Uncore)
• Fix for topology module of Fujitsu A64FX
• Fix for Intel Skylake SP in SNC mode
• Fix for likwid-perfscope
• Fix for CLI argument parsing
• Updated group and data file checkers
• Vector sum benchmark in SVE
• FP_PIPE group for Fujitsu A64FX
• Maximal number of CLI arguments configurable in config.mk (currently 16384)
• Fix for cpulist_sort function
• Fix for Intel SkylakeSP/CascadelakeSP CBOX devices in perf_event mode
• Multiplexing-Fix for perf_event (with warning)
• Adjust CUDA function pointer names in topology_gpu to avoid name clashes
• Fix for Lua 5.1
• Fix for likwid-setFrequency when reading CPU base frequency

Note: This version does not contain any updates for AMD Zen3 and Intel IcelakeSP.
Note: Uncore measurements on Intel Cascadelake AP systems require an update of the topology module which will come in 5.2.0
WARNING: The benchmarks triad_sve* for ARM8 chips use only 3 instead of 4 streams.

Changelog for version 5.1.0:

• Support for Intel Icelake desktop (Core + Uncore)
• Support for Intel Icelake server (Core only)
• Support for Intel Tigerlake desktop (Core only)
• Support for Intel Cannonlake (Core only)
• Support for Nvidia GPUs with compute capability >= 7.0 (CUpti Profiling API)
• Initial support for Fujitsu A64FX (Core) including SVE assembly benchmarks
• Support for ARM Neoverse N1 (AWS Graviton 2)
• Support for AMD Zen3 (Core + Uncore but without any events)
• Check for Intel HWP
• Fix for TID filter of Skylake SP LLC filter0 register
• Fix for Lua 5.1
• Fix for likwid-mpirun skip masks
• Fortran90 interface for NvMarkerAPI (update)
• CPU_is_online check to filter non-usable CPU cores
• Fix for freeMemory in NUMA module (with hwloc backend)
• Fix for likwid-setFrequencies

We want to thank Intel, AMD, AWS and the University of Regensburg for their support.

If you want to use this release in a publication, please cite: https://doi.org/10.5281/zenodo.4282696

Changelog for 5.0.2:

• Fix memory leak in calc_metric()
• New peakflops benchmarks in likwid-bench
• Fix for NUMA domain handling properly
• Improvements for perf_event backend
• Fix for perfctr and powermeter with perf_event backend
• Fix for likwid-mpirun for SLURM with cpusets
• Fix for likwid-setFrequencies in cpusets
• Update for POWER9 event list
• Updates for AMD Zen, Zen+ and Zen2 (events, groups)
• Fix for Intel Uncore events with same name for different devices
• Fix for file descriptor handling
• Fix for compilation with GCC10
• Remove sleep timer warning
• Update examples C-markerAPI and C-internalMarkerAPI

Note: If you want to use LIKWID 5.0.2 with Lua 5.1, please apply this patch

I'm happy to announce a new bugfix release of LIKWID 5.

• Some fixes for likwid-mpirun
  • Fix for hybrid pinning with multiple hosts
  • Fix for perf.groups without core-local events (switch to likwid-pin)
  • Fix for command line parser
  • For for mpiopts parameter
  • Add UPMC as Uncore counter to splitUncoreEvents()
  • Expand user-given input to abspath if possible
  • Check for at least one executable in user-given command
  • Add skip mask for SLURM + Intel OpenMP
  • Check if user-given MPI type is available
• Fix for perf_event backend when used as root
• Include likwid-marker.h in likwid.h to not break old MarkerAPI code
• Enable build with ARM HPC compiler (ARMCLANG compiler setting)
• Fix creation of likwid-bench benchmarks on POWER platforms
• Fix for build system in NVIDIA_INTERFACE=BUILD_APPDAEMON=true
• Update for executable tester
• Update for MPI+X test (X: OpenMP or Pthreads)

Merry Christmas

New version LIKWID 5.0.0

Changelog:

• Support for ARM architectures. Special support for Marvell Thunder X2
• Support for IBM POWER architectures. Support for POWER8 and POWER9.
• Support for AMD Zen2 microarchitecture.
• Support for data fabric counters of AMD Zen microarchitecture
• Support for Nvidia GPU monitoring (with NvMarkerAPI)
• New clock frequency backend (with less overhead)
• Generation of benchmarks for likwid-bench on-the-fly from ptt files
• Switch back to C-based metric calculator (less overhead)
• Interface function to performance groups, create your own.
• Integration of GOTCHA for hooking into client application at runtime
• Thread-local initialization of streams for likwid-bench
• Enhanced support for SLURM with likwid-mpirun
• New MPI and Hybrid pinning features for likwid-mpirun
• Interface to enable the membind kernel memory policy
• JSON output filter file (use -o output.json)
• Update of internal HWLOC to 2.1.0

Note: The MarkerAPI Macros have been moved to a separate header "likwid-marker.h"

New bugfix release:

• Fix for detecting PCI devices if system can split up LLC and memory channels (Intel CoD or SNC)
• Don't pin accessDaemon to threads to avoid long access latencies due to busy hardware thread
• Fix for calculations in likwid-bench if streams are used for input and output
• Fix for LIKWID_MARKER_REGISTER with perf_event backend
• Support for Intel Atom (Tremont) (nothing new, same as Intel Atom (Goldmont Plus))
• Workaround for topology detection if LLC and memory channels are split up. Kernel does not detect it properly sometimes. (Intel CoD or SNC)
• Minor updates for build system
• Minor updates for documentation

Notice: If you want to compile likwid-4.3.4 with ACCESSMODE=perf_event, please apply the attached patch before compiling.

• Fixes for likwid-mpirun
• Fixes for events of Intel Skylake SP and Intel Broadwell
• Support for Intel CascadeLake X (only new eventlist, uses code from Intel Skylake SP)
• Fix for bitmask creation in Lua
• Event options for perf_event backend
• New assembly benchmarks in likwid-bench
• MarkerAPI: Function to reset regions
• Some new performance groups (DIVIDE and TMA)
• Fixes for AMD Zen performance groups
• Fix when using topology input file
• Minor bugfixes

• Fix in internal metric calculator
• Support for Intel Knights Mill (core, rapl, uncore)
• Intel Skylake X: Some fixes for events and perf. groups
• Set KMP_INIT_AT_FORK to bypass bug in Intel OpenMP memory allocator
• AMD Zen: Use RETIRED_INSTRUCTION instead of fixed-purpose counter for metric calculation
• All FLOPS_* groups now have vectorization ratio
• Fix for MarkerAPI with perf_event backend
• Fix for maximal/minimal uncore frequency
• Skip counters that are already in use, don't exit
• likwid-mpirun: minor fix when overloading a host
• Improved detection of PCI devices

Minor fixes for Intel Skylake and frequency module

• Support for Intel Skylake SP architecture (core, uncore, energy)
• Support for AMD Zen architecture (core, l2, energy)
• Support for Intel Goldmont Plus architecture
• Pinning strategy 'balanced'
• New Lua based calculator
• Support for Intel PState CPU frequency daemon

Minor:

• Fixed MCDRAM measurements on Intel Xeon Phi (KNL) with perf_event back end

Merry Christmas

• Fix for logical selection strings
• likwid-agent: general update
• likwid-mpirun: Improved SLURM support
• likwid-mpirun: Print metrics sorted as they are listen in perf. group
• likwid-perfctr: Print metrics/events as header in timeline mode
  Redirect to file when -o switch is used
• likwid-setFrequency: Commandline options to set min, max and current frequency
• Pinning-Library: Automatically detect and skip shepard threads
• Intel Broadwell: Added support for E3 (like Desktop), Fix for L3 group
• Intel IvyBridge: Fix for PCU fixed-purpose counters
• Intel Skylake: Fix for events CYCLE_ACTIVITY, new event L2_LINES_OUT
• Intel Xeon Phi (KNL): Fix for overflow register, Update for ENERGY group
  OFFCORE_RESPONSE events are now tile-specific
• Intel SandyBridge: Fix for L3CACHE group
• Event/Counter list contains only usable counters and events
• Fix and warning message for static library builds

• Support for Intel Xeon Phi (Knights Landing): Core, Uncore, RAPL
• Support for Uncore counters of some desktop chips (SandyBridge, IvyBridge, Haswell, Broadwell and Skylake)
• Basic support for Linux perf_event interface instead of native access. Currently only core-local counters working, Uncore is experimental
• Support to build against a existing Lua installation (5.1 - 5.3 tested)
• Support for CPU frequency manipulation, Lua interface updated
• Access module checks for LLNL's msr_safe kernel module
• Support for counter registers that are only available when HyperThreading is off
• Socket measurements can be used for all cores on the socket in metric formulas.

The LIKWID team wishes Merry Christmas to everyone.

• Fix for likwid-powermeter: Use proper energy unit
• Fix for performance groups for Intel Broadwell (D/EP): DATA and FALSE_SHARE
• Reduce number of started access daemons
• Clean Uncore unit local control registers (needed for simultaneous use of LIKWID 3 and 4)
• Clean config, filter and counter registers at *_finalize function
• Fix for likwid-features and likwid-perfctr

• Fix for Uncore handling for EP/EN/EX systems
• Minor fix for Uncore handling on Intel desktop systems
• Fix in generic readCounters function
• Support for Intel Goldmont (untested)
• Fixes for likwid-mpirun

• Support for Intel Skylake (Core + Uncore)
• Support for Intel Broadwell (Core + Uncore)
• Support for Intel Broadwell D (Core + Uncore)
• Support for Intel Broadwell EP/EN/EX (Core + Uncore)
• Support for Intel Airmont (Core)
• Uncore support for Intel SandyBridge, IvyBridge and Haswell
• Performance group and event set handling in library
• Internal calculator for derived metrics
• Improvement of Marker API
• Get results/metrics of last measurement cycle
• Fixed most memory leaks
• Respect 'Intel PMU sharing guide'
• Update of internal Lua to 5.3
• More examples (C++11 threads,Cilk+, TBB)
• Test suite for executables and library
• Accuracy checker supports multiple CPUs
• Security checked access daemon
• Likwid-bench supports Integer benchmarks
• Likwid-bench selects interation count automatically
• Likwid-bench has new FMA related benchmarks
• Likwid-mpirun supports SLURM job scheduler
• Reintroduced tool likwid-features

• likwid-bench: Iteration determination is done serially
• likwid-bench: Manual selection of iterations possible
• likwid-perfctr: Set cpuset to all CPUs not only the first
• likwid-pin: Set cpuset to all CPUs not only the first
• likwid-accuracy.py: Enhanced plotting functions, use only instrumented likwid-bench
• likwid-accessD: Check for allowed register for PCI accesses
• Add models HASWELL_M1 (0x45) and HASWELL_M2 (0x46) to likwid-powermeter and likwid-accessD
• New test application using Cilk and Marker API
• New test application using C++11 threads and Marker API
• likwid-agent: gmetric version check for --group option and s/\s*/_/ in metric names
• likwid-powermeter: Print RAPL domain name
• Marker API: Initialize access already at likwid_markerInit()
• Marker API: likwid_markerThreadInit() only pins if not already pinned

After about one year of development, We are happy to announce the newest version of LIKWID.

The features of LIKWID 4.0.0:

• Support for Intel Broadwell
• Uncore support for all Uncore-aware architectures
  • Nehalem (EX)
  • Westmere (EX)
  • SandyBridge EP
  • IvyBridge EP
  • Haswell EP
• Measure multiple event sets in a round-robin fashion (no multiplexing!)
• Event options to filter the counter increments
• Whole LIKWID functionality is exposed as API for C/C++ and Lua
• New functions in the Marker API to switch event sets and get intermediate results
• Topology code relies on hwloc. CPUID is still included but only as fallback
• Most LIKWID applications are written in Lua (only exception likwid-bench)
• Monitoring daemon likwid-agent with multiple output backends
• More performance groups