Hi, When converting a tensorflow graph to tensorrt, we see the some reformatter stages while running the tensorrt model:
I0815 16:02:00.449360 19849 rt_profiler.cc:17] conv1_1_1/Conv2D + (Unnamed Layer* 2) [Activation] input reformatter 0 0.375616
I0815 16:02:00.450865 19849 rt_profiler.cc:17] conv2_2_1/Conv2D + (Unnamed Layer* 14) [Activation] input reformatter 0 0.076256
I0815 16:02:00.451476 19849 rt_profiler.cc:17] conv2_3_1/Conv2D + (Unnamed Layer* 17) [Activation] input reformatter 0 0.12176
Could anyone help explaining what are these reformatters and if they could be optimized away?
We also tried profile the program and saw some genericReformat::copyPackedKernel calls, not sure if that’s related with the reformatter invocation?