Modify the L3->L2 channels to distribute data over scf.for

[erwei-xilinx]

• This pattern implies a chain of bds performing time-multiplexed movement of multiple memrefs, operating on one shared DMA channel.
• Depends on AIRDependency: Improved areAsyncDependent method #766 and AIRSpecializeChannelWrapAndStride: Unroll any remaining data movement loops #767 to pass.

[hunhoffe]

Looks good to me once CI passes! Interesting error currently in the CI, not sure what would cause it.

[fifield]

• This pattern implies a chain of bds performing time-multiplexed movement of multiple memrefs, operating on one shared DMA channel.

You did this based on your deep knowledge of the compiler. How can we document this and other such patterns so that others can access them?

[erwei-xilinx]

Looks good to me once CI passes! Interesting error currently in the CI, not sure what would cause it.

I think this is reproducing the issue that you previously mentioned in #709. I'm working on a number of pass fixups to generalize some AIR passes to cover this pattern.

[erwei-xilinx]

• This pattern implies a chain of bds performing time-multiplexed movement of multiple memrefs, operating on one shared DMA channel.

You did this based on your deep knowledge of the compiler. How can we document this and other such patterns so that others can access them?

Yes it would be very useful to have proper documentation for users who wish to write designs in AIR using python bindings. I created an issue to keep track of the documentation progress: #770.