Options - Options - 57368

-o| --output-dir <directory-path>

Predefined sampling configuration to be used to collect samples.

Use the command info --list collect-configs to get the list of supported configs. Multiple occurrences of --config are allowed.

-e | --event or <predefined-event>

A predefined event can directly be used with -e, --event which has predefined arguments.

Alternatively, for providing more granular parameters, specify Timer, PMU, IBS event, or a predefined event with arguments in the form of comma separated key=value pairs. The supported keys are:

• event=<timer | ibs-fetch | ibs-op> or <PMU-event> or <predefined-event>
• mask=<unit-mask>
• user=<0 | 1>
• os=<0 | 1>
• cmask=<count-mask> (Value should be in the range 0x0 to 0x7f)
• inv=<0 | 1>
• interval=<sampling-interval>
• frequency=<frequency (n)> (Supported only for Core PMC events. Frequency should be provided in Hz)
• ibsop-count-control=<0 | 1> (for ibs-op event)
• loadstore (for ibs-op event, only on Windows platform)
• ibsop-l3miss (for ibs-op event, supported only on AMD “Zen4” processors)
• ibsfetch-l3miss (for ibs-fetch event, supported only on AMD “Zen4” processors)
• call-graph

Argument details:

• user – Enable(1) or disable(0) user space samples collection
• os - Enable(1) or disable(0) kernel space samples collection
• interval – Sample collection interval. For timer, it is the time interval in milliseconds. For PMU and predefined events, it is the count of the event occurrences. For IBS FETCH, it is the fetch count. For IBS OP, it is the cycle count or the dispatch count.
• op-count-control – Choose IBS OP sampling by cycle(0) count or dispatch(1) count.
• loadstore – Enable only the IBS OP load/store samples collection, other IBS OP samples are not collected.
• ibsop-l3miss – Don’t filter out any IBS OP samples (0), or filter out all IBS OP samples except loads/stores that miss in the L3 cache (1).
• ibsfetch-l3miss – Don’t filter out any IBS FETCH samples (0), or filter out all IBS FETCH samples except loads/stores that miss in the L3 cache (1).

Profile the existing processes by attaching to a running process. The process IDs are separated by comma.

System Wide Profile (SWP)

If this flag is not set, the command line tool will profile only the launched application or the Process IDs attached with -p option.

Comma separated list of CPUs to profile. The ranges of CPUs can be specified with ‘-’, for example, 0-3. This option is not supported with MPI profiling.

Sampling interval for the PMC events.

--affinity <core-id...>

--thread <thread=concurrency>

Profiling paused indefinitely. The target application resumes the profiling using the profile control APIs. This option must be used only when the launched application is instrumented to control the profile data collection using the resume and pause APIs (see AMDPowerProfileAPI Library for definitions).

-w| --working-dir <path>

Specify the path where the log file should be created. If this option is not provided, the log file will be created either in path set by AMDUPROF_LOGDIR environment variable or $TEMP path (Linux, FreeBSD) or %TEMP% path (on Windows) by default.

The log file name will be of the format $USER-AMDuProfCLI.log (on Linux,FreeBSD) or %USERNAME%-AMDuProfCLI.log (on Windows).

Capture the timestamp of the log records. It should be used with --enable-log option

Enable data collection at the specified frequency 'n' (in Hz) for Core PMC events.

Specify the report to be generated. The supported report options are:

• process: Report process details
• module: Report module details
• thread: Report thread details

This option is applicable only with the --detail option. The default is group-by process.

Report only the events present in the given view file. Use the command info --list view-configs to get the list of supported view-configs.

Show inline functions for C, C++ executables.

--src-path <path1;...>

Source file directories (semicolon separated paths). Multiple use of --src-path is allowed.

--disasm-style <att |intel>

-s| --sort-by <EVENT>

Specify the Timer, PMC, or IBS event on which the reported profile data will be sorted with arguments in the form of comma separated key=value pairs. The supported keys are:

• event=<timer| ibs-fetch | ibs-op | pmcxNNN>, where NNN is hexadecimal Core PMC event id.
• umask=<unit-mask>
• cmask=<count-mask>
• inv=<0 | 1>
• user=<0 | 1>
• os=<0 | 1>
• metric=<cpu_time | total_cpu_time | self_time | total_time>

If both event and metric are enabled, event will take priority over metric.

Use the command info--list pmu-events for the list of supported PMC events.

Argument details are:

• umask: Unit mask in decimal or hexadecimal, applicable only to the PMC events.
• cmask: Count mask in decimal or hexadecimal, applicable only to the PMC events.
• user, os: User and OS mode. Applicable only to the PMC events.
• inv: Invert Count Mask, applicable only to the PMC events.
• metric:
  • cpu_time is applicable only if CPU_TIME event is collected.
  • total_cpu_time is applicable only with hotspots (or) threading analysis, if callstack collection (-g) is enabled for dynamically linked launch application.
  • self_time and total_time are applicable only if function tracing is collected.
• Multiple occurrences of --sort-by(-s) are not allowed.

--agg-interval <low | medium | high | INTERVAL>

Use this option to configure the sample aggregation interval which is useful when the session gets imported to GUI.

low level of aggregation interval generates better timeline view in GUI, but increases the database size.

Aggregation INTERVAL can also be specified as numeric value in milliseconds.

--time-filter <T1:T2>

IMIX report generation. Supported group-by options are:

• module: Report module-wise IMIX.
• thread: Report thread-wise IMIX.
• function: Report function-wise IMIX.

Show all cachelines in report sections for cache analysis. By default, only cachelines accessed by more than one process/thread are listed. Use this option for false cache sharing analysis.

--symbol-path

<path1;...>

--report-output <path>

Use this option to generate ASCII dump of IBS OP profile samples.

--call-graph <I:D:S:F>

Enables Callstack Sampling. Specify the Unwind Interval (I) in milliseconds and Unwind Depth (D) value. Specify the Scope (S) by choosing one of the following:

• user: Collect only for the user space code.
• kernel: Collect only for the kernel space code.
• all: Collect for the code executed in the user and kernel space code. Specify to collect missing frames due to Frame Pointer Omission (F) by compiler:
• fpo: If frame pointers are not available, collect callstack information using unwind information.
• fp: Use frame pointers to collect callstack information.

Enables callstack sampling. Specify (F) to collect/ignore missing frames due to omission of frame pointers by compiler:

• fpo | dwarf: Collect process call stack during sample collection and use DWARF information to reconstruct the call stack.
• fp: Use Frame pointers to collect call stack information.

When F = fpo, (N) specifies the max stack-size in bytes to collect per sample collection. Valid range of the stack size: 16 - 32768. If (N) is not a multiple of 8, then it is aligned down to the nearest value multiple of 8. The default value is 1024 bytes.

When F= fp, the value for N is ignored and hence, there is no need to pass it.

Same as passing --call-graph 1:128:user:fp (Windows)

Same as passing --call-graph fp (Linux)

Set callstack collection mode.

• fpo - If frame pointers are not available, collect call-stack information using unwind information.
• fp - Use Frame pointers to collect callstack information.

Default mode is fp.

Callstack collection mode. Default mode is fp.

• fp: Use Frame pointers to collect call stack information.
• fpo | dwarf: Collect process call stack during sample collection and use DWARF information to reconstruct the call stack.

Set callstack scope type.

Scope type should contain one of these options:

• user - Collect only for user space code.
• kernel - Collect only for kernel space code.
• all - Collect for code executed in user and kernel space.

Default scope type is user.

Set callstack unwind interval. Interval must be within the range [1 - 100].

Default interval is 1 ms.

Set callstack unwind depth. Depth must be within the range [2 - 392].

Default depth is 128.

Set callstack unwind depth. Depth must be within the range [2 - 1024].

Default depth is 32. This option is applicable for Hotspots and Threading configurations, for any other configurations this option will be ignored.

Callstack Size. Default size is 1024 bytes.

When mode = fpo | dwarf; size must be within [16 - 32768] and specifies the max stack-size (in bytes) to collect per call stack sample.

When mode = fp; the size is ignored, hence no need to pass it.

Collects the runtime thread details.

Profile existing threads by attaching to a running thread. The thread IDs are separated by comma.

Size (number of pages per core) of the buffer used for data collection by the driver. The default size is 512 pages per core.

Symbol Server directories (semicolon separated paths). For example, Microsoft Symbol Server. Multiple use of --symbol-server is allowed.

The path to store the symbol files downloaded from the Symbol Servers.

Use this option to download symbols using the Microsoft Symsrv. By default, AMD symbol downloader will be used to download symbols.

Use this option to stop the profiling once the collected data file size (in MBs) crosses the limit. When used with the --overwrite option, the limit is before the collection is terminated. Size can be specified with a suffix Mega bytes (M/ m), Giga Bytes (G/g), and Seconds (secs).

Specify the profile-data collection mode as a ring buffer. Collection limit can be set using --limit-data option. Default --limit-data is to restrict the raw data file size to 512 pages per core.

To trace a target domain. TARGET can be one or more of the following:

• osrt - to enable tracing of os runtime. Use command info --list trace-events for the list of trace events.
• func - to enable tracing of functions. Use --func, --func-size and --func-threshold to configure additional options.
• memory - to enable tracing of dynamic memory allocations. Use --memory-threshold to configure threshold.
• openmp - to enable tracing of OpenMP application. Use --openmp-impl and --openmp-scope to configure additional options.

Provide OpenMP implementation type:

Provide tracing scope: full for complete tracing, basic for basic tracing, where synchronization related OpenMP events are not traced to reduce the disk space usage. Default selection is basic.

This option is only applicable with --openmp-impl ompt.

Provide event names. Use command info --list trace-events for the list of trace events.

Provide event name and threshold value.

Specify functions to trace from the library or executable.

• Function-pattern can be a function name or partial name ending with '*'. Use only '*' to trace all the functions of a module.
• Module can be absolute path to library or executable. This option will be deprecated in a future release.
• Recommended to use --func.

Specify functions to exclude from the library or executable.

• Function-pattern can be a function name or partial name ending with '*'. Use only '*' to trace all the functions of a module.
• Module can be absolute path to library or executable. This option will be deprecated in a future release.
• Recommended to use --exclude-func.

Provide minimum function size to trace. Default function size is 128 bytes.

This option will be deprecated in a future release. Recommended to use --func-size.

Specify functions to trace from the library, executable, or kernel: function- pattern can be a function name or partial name ending with '*' or only '*' to trace all the functions of a module.

Module can be a library or executable. To trace the kernel functions, replace the module with ‘kernel’.

Specify functions to exclude from the library, executable, or kernel:

• function-pattern can be a function name or partial name ending with '*' or only '*' to trace all the functions of a module.
• Module can be a library or executable. To trace the kernel functions, replace the module with ‘kernel’.

Set the kernel memory mapped data buffer to size. The size can be specified in pages or with a suffix Bytes (B/b), Kilo bytes (K/k), Megabytes (M/m), and Gigabytes (G/g).

Specify the PID of qemu-kvm process to be profiled to collect guest-side performance profile.

Specify the path of guest /proc/kallsyms copied on the local host. AMD uProf reads it to get the guest kernel symbols.

Specify the path of guest/proc/modules copied to the local host. AMD uProf reads it to get the guest kernel module information.

Specify the path of guest vmlinux and kernel sources copied on the local host. AMD uProf reads it to resolve the guest kernel module information.

This option is used along with the --input-dir option. Generates report belonging to a specific host. The supported options are:

• <hostname>: Report process belonging to a specific host.
• all: Report all the processes.

Generate report only for specific profiling category. Comma separated multiple categories can be specified. If this option is not used, the report for all categories is generated. Multiple instances of --category is allowed.

Example:

• --category cpu, mpi, trace, gputrace, gpuprof
• --category mpi --category cpu --category trace --category gputrace
• --category gpuprof

Specify the time interval in seconds to list the function count detail report. If this option is not specified, function count will be generated for the entire profile duration.

Use this option to capture LBR data. Specify the branch filter type:

• u: user branches
• k: kernel branches
• any: any branch type
• any_call: any call branch
• any_ret: any return branch
• ind_call: indirect calls
• ind_jmp: indirect jumps
• cond: conditional branches
• call: direct calls

When the above filters are not set, the default filter type will be any.

Path to the Linux kernel debug info file. If no path provided, it searches for the debug info file in the default download path.