[Question]: Inspector: why exec time is the max of channel exec times?

[RicoloveFeng]

Question

static uint64_t calculateMaxKernelExecTimeUsecs(struct inspectorCollInfo *collInfo,
                                                inspectorTimingSource_t *timingSource) {
  // ...
  for (uint32_t i = 0; i < collInfo->nChannels; i++) {
    struct inspectorKernelChInfo *kernelCh = &collInfo->kernelCh[i];  // <-- get channel[i] info
    uint64_t gpuExecTimeUsecs = calculateKernelGpuExecTimeUsecs(kernelCh);  // <-- stopClk - startClk

    if (gpuExecTimeUsecs > 0) {
      if (gpuExecTimeUsecs > maxKernelExecTimeUsecs) {
        maxKernelExecTimeUsecs = gpuExecTimeUsecs;  // <-- max of all time
        bestTimingSource = inspectorTimingSourceKernelGpu;
      }
    } else {
    // ...

Let's say there is something wrong with a kernel, making a channel later than all others.

time point        0          1          2         ...       98        99
channel 0      start                    end
channel 1      start                    end           
channel 2      start                    end
channel 3                                                   start      end

In this case, the kernel exec time should be like 99 - 0 = 99, but in current algorithm we will say it is max(2-0, 99-98) = 2.

Could you elaborate why the inspector picks the latter algorithm? I'm not sure if this is working as intended or there are some mechanisms that make it reasonable than the former one. Thanks! 😃