Performance Analysis with Intel Vtunes

[Abhishek Tyagi]

Hi everyone,

This is my first post on Caffe Users group so bear with me.

I am using Intel-caffe to observe the performance on an Intel Xeon machine. To analyze the performance, there are two methods:

1. I have used 

caffe time -model path/to/deploy/prototxt -iterations 100

this gives the numbers for each layer and is really handy. But the drawback is that it does not give out any indication about what operations (talking about inference) take the most time and are the bottleneck in the entire process. And hence approach 2.

2. For such analysis, Intel VTunes seems to be the way to go with their Hotspot analysis. Now, they do have a tutorial on how to use Intel Vtunes with Tensorflow but they have no information on using it with Caffe.

So, I am wondering if anyone has used the Intel VTunes performance profiler with caffe? If yes, your help would be really appreciated if you can share the method for doing so