Connection to third-party tools¶
Caliper provides bindings to export Caliper annotations to third-party tools. Currently, Nvidia’s NVProf/NSight and Intel VTune Amplifier are supported.
NVidia NVProf/NSight¶
The nvtx service translates Caliper annotations into NVTX ranges (Nvidia’s source-code annoatation interface). Caliper-annotated regions will then be visible in the “Markers and Ranges” section of the NVprof or NSight timeline display.
To use the NVprof bindings, enable the nvtx service and run the target application under nvprof:
$ CALI_SERVICES_ENABLE=nvtx nvprof <nvprof-options> app <app-options>
Then, load the measurement result file created by nvprof in nvvp (use the -o flag for nvprof to write output into a file).
By default, Caliper exports all regions with the NESTED attribute property. This includes the regions created by Caliper’s annotation macros (CALI_CXX_MARK_FUNCTION, CALI_MARK_BEGIN/END etc.). Alternatively, the region attributes to export can be specified explicitly in the CALI_NVTX_TRIGGER_ATTRIBUTES configuration variable.
- CALI_NVTX_TRIGGER_ATTRIBUTES
Specify which attributes should be exported as NVTX ranges. Comma-separated list.
Intel VTune¶
The vtune service translates Caliper annotations into Intel “itt” markups that can be shown as “tasks” in the Intel VTune GUI.
To use the vtune bindings, run the target application in VTune with the vtune service enabled. To do so in the VTune GUI, do the following:
In the “Analysis Target” tab, go to the “User-defined environment variables” section, and add an entry setting “CALI_SERVICES_ENABLE” to “vtune”
In the “Analysis Type” tab, check the “Analyze user tasks, events, and counters” checkbox.
Caliper-annotated regions will then be visible as “tasks” in the VTune analysis views.
By default, Caliper exports all regions with the NESTED attribute property. Alternatively, the region attributes to export can be specified explicitly in the CALI_VTUNE_TRIGGER_ATTRIBUTES configuration variable.
CrayPAT¶
The craypat service translates Caliper annotations into CrayPAT regions. This API is only available in recent (2022+) CrayPAT versions. It also only works with the “trace” call stack mode in CrayPAT.
By default, the craypat service will attempt to initialize CrayPAT internally
and record a CrayPAT region profile. Alternatively, you can use the craypat
service in conjunction with pat_run to set additional recording options. In
this case, make sure to set PAT_RT_CALLSTACK_MODE=trace
:
$ PAT_RT_CALLSTACK_MODE=trace CALI_SERVICES_ENABLE=craypat pat_run -w -E PAPI_TOT_CYC ./examples/apps/cxx-example
CrayPat/X: Version 22.06.0 Revision 4b5ab6256 05/21/22 02:04:43
Experiment data directory written:
/<path>/cxx-example+91423-10t
$ pat_report ./cxx-example+91423-10t
[...]
Table 1: Profile by Function Group and Function
Time% | Time | Imb. | Imb. | Calls | Group
| | Time | Time% | | Function
100.0% | 0.000716 | -- | -- | 7.0 | Total
|------------------------------------------------------
| 100.0% | 0.000716 | -- | -- | 7.0 | USER
||-----------------------------------------------------
|| 89.2% | 0.000639 | -- | -- | 4.0 | [R]foo
|| 4.7% | 0.000034 | -- | -- | 1.0 | [R]main
|| 3.9% | 0.000028 | -- | -- | 1.0 | [R]mainloop
|| 2.1% | 0.000015 | -- | -- | 1.0 | [R]init
|======================================================
[...]