tree c3ac8f6450bbf1b34a8527ecc4c6dbed67cdae86
parent 64242eff9504b060cb9ed84a61751e909f8ae9b6
author Ahmed S. Taei <ataei@google.com> 1585289182 -0700
committer Copybara-Service <copybara-worker@google.com> 1585329758 -0700

xla_hlo.conv -> vmla.conv lowering and runtime kernel.

This will convert all arguments down to vmla.conv but only legalize if :
 - 2D conv.
 - No group convolution.
 - No deconvolution.
 By doing this we can easily extend support with minimal changes to the lowering.

- The runtime for 2D conv is a simple (slow) loop nest matches refrence TF doc:
 https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/nn/convolution

PiperOrigin-RevId: 303266802