Add FlattenAtrousConv transformation

[Icemist]

This transformation flattens atrous convolution, i.e. finds a sequence of space_to_batch_nd->conv2d->batch_to_space_nd operations and convert them into subgraphs with a convolution with the modified "dilation" and recalculated "padding" parameters

  x     w      x     w
  |     |      |     |
  s2b   |       \   /
   \   /         \ /
   conv2d   ->  conv2d
     |            |
     b2s          |

[elvin-n]

I propose to add invocation of this pass into standard optimize function OptimizeImpl here. This will allow to avoid situation when users are not familiar with this pass and will not know when and how to call it

[Icemist]

@masahi @jwfromm

[masahi]

Interesting, I think we can safely enable this pass in the default pipeline as @elvin-n suggested, @Icemist do you want to do that?

You can put the new pass in

tvm/src/relay/backend/utils.cc

Line 206 in 4c608be

Array<Pass> GetPassPrefix(bool is_homegeneous, bool is_vm) {

[elvin-n]

It's pity not to include this pass into standard optimization flows. @Icemist please create another PR with adding it to regular optimization passes.

[Icemist]

Interesting, I think we can safely enable this pass in the default pipeline as @elvin-n suggested, @Icemist do you want to do that?

You can put the new pass in

tvm/src/relay/backend/utils.cc

Line 206 in 4c608be

Array<Pass> GetPassPrefix(bool is_homegeneous, bool is_vm) {

Yes, I posted it as a separate commit #11077. I apologize for the long response.

[TejashShah]

@Icemist Excellent piece of contribution, regardless of target! Appreciate your efforts in mainlining it

cc @elvin-n