tree 025be61e7ed4ca7a0dd4baceea8f7a50d372f9dc
parent 851626e0c6645bcbb90554ced6c02be6fe52fbbe
author rsuderman <suderman@google.com> 1603327614 -0700
committer GitHub <noreply@github.com> 1603327614 -0700
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJfkNZ+CRBK7hj4Ov3rIwAAdHIIAAwQwJTVzPZ2lM/WNQi84Z+8
 p5PJcCRD+K8XJAdmi2HFiZVorsNgWH4kWzKaX/vWy12IsYdYoTH7gclub6Xpfz2E
 eVBYggDN/NknvpuOdrBJQpAwBGTiDCbIlWTyd9KnO/wOGHZd56023494G8R9y5UV
 fUhwen0dbDH8Rv7ee5A2ZDxQHUogi0ubPIN44qN3ct3zOmlD8mfVZMGawu6n1dDW
 sQ11R4xRUxhCpEESQ4ocp6HofA3sN7v2jz2R3J17ZZXvuSyzKlMkCBehGZoJvzTl
 lF7UyTmmsCbyl14B1q3ZD5vboGjufvzI8LiJIthpGhSxVwtNOpT7kYKPES2ittU=
 =/OHl
 -----END PGP SIGNATURE-----
 

Added a TF2XLA Lowering that includes XLA folders

Without Tensorflow folders we were not able to progressively lower TF to XLA in cases where constant evaluation is required (e.g. constant axis for tf.Concatenate). Including all TF folders is not an option due to kernel build times, so by creating a custom legalize-tf pass for XLA that includes the XLA folders and alternates folding/lowering, we can legalize these cases without the TF folders.

Includes re-enabling linspace for VMLA.