Skip to content

Better compositor tracking and live statistics

Jakob Bornecrantz requested to merge wallbraker/monado:jakob/submit-timing-changes into main

Builds on !2101 (merged)

If you use the fake compositor pacer and set U_PACING_LIVE_STATS=true the pacing code will now generate lives statistics about every 1024 frames. It looks something like the below. This MR also include improvements in frame tracking so we can see the overhead of the kernel and any delays before the GPU picks up the work we scheduled. If you see the first message that the worst gpu_delay is 0.963ms, and it is a major contributor to the frame taking a long time.

 INFO [print_and_reset] Compositor frame timing:
            name          median            mean           worst
             cpu         0.008ms         0.007ms         0.014ms
            draw         0.107ms         0.108ms         0.385ms
          submit         0.049ms         0.049ms         0.077ms
             gpu         0.175ms         0.180ms         0.386ms
       gpu_delay         0.089ms         0.089ms         0.963ms
     total_frame         0.380ms         0.388ms         1.509ms
 WARN [wait_for_scheduled_free] 277549598.581ms: Dropping old missed frame in favour for completed new frame
 INFO [print_and_reset] Compositor frame timing:
            name          median            mean           worst
             cpu         0.008ms         0.007ms         0.013ms
            draw         0.108ms         0.107ms         0.166ms
          submit         0.049ms         0.050ms         0.118ms
             gpu         0.175ms         0.175ms         0.180ms
       gpu_delay         0.092ms         0.092ms         0.196ms
     total_frame         0.385ms         0.384ms         0.499ms

A really large values looks like this:

            name          median            mean           worst
             cpu         0.008ms    79'650.000ms         0.013ms
            draw         0.105ms 1'096'130.000ms         0.166ms
          submit         0.052ms   509'260.000ms         0.118ms
             gpu         0.174ms 1'745'250.000ms         0.180ms
       gpu_delay         0.089ms   903'380.000ms         0.196ms
     total_frame         0.382ms 3'840'680.000ms         0.499ms
Edited by Jakob Bornecrantz

Merge request reports