Note:
(1) This section provides program-level analysis.
(2) A tooltip with extra information will show up if you mouse over a point in the roofline chart.
(3) You can choose whether to include infeed and outfeed ops for the analysis.
(4) "Total" aggregates all operations throughout the entire profiling session. It includes incomplete steps.
(5) "Total (HW)" is based on the hardware performance counters while the others are based on the XLA's cost analysis. It is always calculated including infeed and outfeed ops regardless of the option. The gap between "Total" and "Total (HW)" is due to hardware limitation (e.g., padding).
(6) "Average" shows the average step information by aggregating the operations in the complete steps only.

Note:
(1) This section provides operation-level analysis.
(2) A tooltip with extra information will show up if you mouse over a point in the roofline chart.
(3) To avoid sluggishness, only the 1000 most time-consuming operations are shown.
(4) You can choose whether to include infeed and outfeed ops for the analysis.
(5) You can filter data by HLO category, bottleneck resource or HLO name.
(6) "IDLE" represents the portion of the total execution time on device that is idle.
(7) Ops with zero FLOP (e.g., data formatting ops like reshape, IDLE, etc.) do not show up in the roofline chart.