Hi everyone,
I am measuring the FPS for the TensorRT Imagenet application by computing the loop time. The measurements are however providing me with a FPS larger than the actual video frame rate. What could be the reason?
I thank you very much.
Hi maycondouglasd, measuring the CPU wall clock may not account for GPU time depending on synchronization (or vice-versa). TensorRT has hooks for layer profiling built-in which will provide you with more accurate timing information.
You can see example code of utilizing TensorRT’s profiler API in this example code, using the tensorNet::EnableProfiler() function:
- https://github.com/dusty-nv/jetson-inference/blob/e12e6e64365fed83e255800382e593bf7e1b1b1a/tensorNet.cpp#L82
- https://github.com/dusty-nv/jetson-inference/blob/e12e6e64365fed83e255800382e593bf7e1b1b1a/tensorNet.h#L155
- https://github.com/dusty-nv/jetson-inference/blob/e12e6e64365fed83e255800382e593bf7e1b1b1a/imagenet-console/imagenet-console.cpp#L60