tf lite armeabi version is much slower than the nightly aar

[zjmirving]

   At first I was using  'org.tensorflow:tensorflow-lite:0.0.0-nightly'  aar in my own project. My tf lite model costs 900ms. But in my company project, the ndk abiFilters is armeabi.  I changed abiFilters from armeabi-v7a to armeabi. However,  I found the aar not support armeabi.

I compile the armeabi version with the following command, and import so and jar file, but my model costs 4000ms. Is there some wrong with the command ??
bazel build --cxxopt='--std=c++11' //tensorflow/lite/java:tensorflowlite
--crosstool_top=//external:android/crosstool
--host_crosstool_top=@bazel_tools//tools/cpp:toolchain
--cpu=armeabi

[zjmirving]

wrong project sorry

[psyhtest]

Phew, I don't even know what aar means :).