search

intel / pti-gpu

Profiling Tools Interfaces for GPU (PTI for GPU) is a set of Getting Started Documentation and Tools Library to start performance analysis on Intel(R) Processor Graphics easily
MIT License
202 stars 57 forks source link

how to enable ChromePerformanceTimingInStages in L0 #5

Closed guizili0 closed 3 years ago

guizili0 commented 3 years ago

Hi in opencl-intercept-layer, I can use the config ChromePerformanceTimingInStages to dump the the multi-kernel submitted, execution time in pipeline, how to enable this one L0 trace? Detail in: https://github.com/intel/opencl-intercept-layer/blob/master/docs/chrome_tracing.md

anton-v-gorshkov commented 3 years ago

Hi, right now ze_tracer does not support stages highlighting like opencl-intercept-layer. But you can collect Chrome trace that includes both host and device activities with --chrome-call-logging and --chrome-device-timeline flags, where each kernel execution has it's unique ID that can be matched with host API calls.

E.g. one may find zeCommandListAppendLaunchKernel call (append phase), zeCommandQueueExecuteCommandLists call (submit phase), and kernel execution (execution phase) with the same ID to retrieve similar information.

If you need exactly the same functionality as ChromePerformanceTimingInStages provides - yes, I believe we can do this as well, just let us know. But note that L0 works differently than OpenCL and has not quite the same stages.

guizili0 commented 3 years ago

I think the ChromePerformanceTimingInStages is very helpful in performance analysis, it is great if your team can support this one.

anton-v-gorshkov commented 3 years ago

Got it, we will start work on this capability. Will update you when it will be ready.

guizili0 commented 3 years ago

great, thank you!!!

anton-v-gorshkov commented 3 years ago

Done, --chrome-device-stages option. Please try it and let me know if there are any issues with this new feature.

guizili0 commented 3 years ago

@anton-v-gorshkov thank you for the quick response, I will check it. And close this issue now. Thanks.