Skip to content

Profiling: Add GPU Time column to -log_view, stop censoring host time when GPUs are used

Toby Isaacrequested to merge
tisaac/feature-log-gpu-time into main

Before this MR the alternatives are

  • -log_view_gpu_time 0 And have most timings censored, even if they are not measuring events that are sensitive to the asynchronous nature of GPU computations, or
  • -log_view_gpu_time 1 And insert lots of event synchronizations into a program, slowing it down (I see event synchronizations that are on the order of a millisecond in Nsight, which is often longer than the event being timed).

This MR adds a GPU Time (s) column to -log_view. These times are censored unless -log_view_gpu_time is passed. The timings in the Time (sec) columns are not censored anymore.

This MR also adds documentation about interpreting -log_view output with GPUs to the manual.

Merge request reports