Caffe2 python Conv op can not specify engine

[Guodonggogo]

I am testing the Conv with 'depthwise_3x3' engine in Caffe2. My caffe2 is installed from source. I constructed one layer network which contains only one group convolution layer with input size (1,100,600,600) and kernel size (100,1,3,3), group=100. However, when I specify the engine to be 'depthwise_3x3', the speed is the same with engine 'cudnn' (or '" ""' or anything others). It seems that the argument 'engine=' does not work.

[Guodonggogo]

Or, dose anyone test the speed of 11 conv and 33 depthwise conv ? Are they really faster than caffe and other platform? Can anyone provide one demo?