* Target arch only supports 64 bit arm and 64 bit x86. BSP (MACHINE)
incluced in above archs should be supported.
* Bazel build takes lots of time, it has own rules and builds everything
from scratch. Currently bazel could not reuse Yocto DEPENDS/RDEPENDS.
* In order to run tensorflow cases in a reasonable time, although it builds
successfully on qemuarm, qemuarm64, qemux86 and qemux86-64, only qemux86-64
with kvm for runtime test.
* It failed to use pre-build model to do predict/inference on big-endian
platform, since upstream does not support big-endian very well
https://github.com/tensorflow/tensorflow/issues/16364
* If host(build system) is not x86_64, please add meta-java to BBLAYERS in
conf/bblayers.conf (git://git.yoctoproject.org/meta-java)
* Due to tensorflow build requires lots of CPU and Memory, in order to
avoid out of memory issue, explicitly set the number of local CPU
threads available to 25% of total number of local CPU cores and
the amount of local host RAM (in MB) available to 25% of total amount
of local host RAM by default. If host is powerful enough, adjust BAZEL_JOBS
and BAZEL_MEM in local.conf to override default set. If set BAZEL_JOBS = ""
and BAZEL_MEM = "", there will be no limitation on the available CPU and RAM.