Optimized TF compilation to use local CPU supported instructions

[gkorland]

Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA"

[lantiga]

The TF binaries are built without vectorization instructions, in order to support the broadest set of CPU architectures possible [1]. It is possible to enable vectorizations at build time at the bazel level, see [2]. We'll need to ship custom libtensorflow in this case, but the speedup will likely be significant.

References

[1] https://stackoverflow.com/questions/47068709/your-cpu-supports-instructions-that-this-tensorflow-binary-was-not-compiled-to-u
[2] https://stackoverflow.com/questions/41293077/how-to-compile-tensorflow-with-sse4-2-and-avx-instructions/46426188#46426188