Combined analysis
CPU and GPU context, with the clocks accounted for.
Open both recordings
Pass a perf.data recording and an Nsight Systems report to open both sources in one workspace. An exported .sqlite report can be used in place of .nsys-rep.
pulsario perf.data report.nsys-rep
pulsario perf.data report.sqliteCheck alignment first
Being open together does not mean two recordings share a trustworthy clock. Pulsario reports a synchronization verdict with evidence. Missing clock metadata, non-overlapping recordings, or captures from different hosts can prevent reliable alignment.
Inspect relationships
Use the unified workspace to move between CPU samples, host-side CUDA calls, and GPU streams. CUDA launch relationships come from recorded correlation data. Cross-source timing conclusions must respect the alignment verdict.
Make assumptions explicit
When the clocks cannot be related automatically, a manual offset or shared-marker alignment can be used. These are labeled as user assumptions. An offset is not proof that two independently captured workloads are the same execution.