Enable backtrack mapping strategy

[ShangkunLi]

In this pr, we add the following features to enable backtrack mapping. And we can achieve CompiledII = 4 for branch_for.mlir now.

• Enable choosing different mapping strategies using the command line. e.g., --map-to-accelerator="mapping-strategy=backtrack_mapping"

• Support backtrack mapping. Users can specify some parameters in include/Mapping/BacktrackMapping/BacktrackMapping.h for this mapping strategy:

  1. max_location_to_try: The MAX number of candidate locations for each op to try.
  2. max_backtrack_depth: The MAX depth of backtrack.

TODO:

• Fix the bug when handling functions with arguments. My initial plan is to create a neura.constant for each function argument.

[tancheng]

Thanks for the awesome work~! I just have one concern:

The CGRA-mapper repo's heuristic is exactly your backtrackMap with max_location_to_try = All & max_backtrack_depth = 1, while its exhaustive is exactly your backtrackMap with max_location_to_try = All & max_backtrack_depth = All.

So for this repo, can we deprecate the current heuristic, but wrap your max_location_to_try = All & max_backtrack_depth = 1 as the new heuristic and treat your max_location_to_try = All & max_backtrack_depth = All as the new exhaustive?

[tancheng]

Hi @ShangkunLi, re: #59

We didn't touch the placement tile selection, or prioritize anything related to utilization. This PR just enables backtrack and everything works as expected, right?

How many minutes we can get a valid mapping solution with your new algorithm? And how many times the location_to_try and backtrack_depth are updated (i.e., #times for backtracking)?

If it is fast enough, we can close that issue, instead, we can create new ones if compilation time is bottleneck in the future.

[ShangkunLi]

Hi @ShangkunLi, re: #59

We didn't touch the placement tile selection, or prioritize anything related to utilization. This PR just enables backtrack and everything works as expected, right?

How many minutes we can get a valid mapping solution with your new algorithm? And how many times the location_to_try and backtrack_depth are updated (i.e., #times for backtracking)?

If it is fast enough, we can close that issue, instead, we can create new ones if compilation time is bottleneck in the future.

I only added the backtracking functionality in this pr. The mapper can only map the former ~10 operations (most are neura.constant and grant_once/always) without triggering backtrack. And nearly all operations following need backtracking. So I think it's crucial to properly design the cost function.

For now, I set max_location_to_try = 5, and max_backtrack_depth = 3 in the mapper. The running time of mapping branch_for.mlir is about 14 seconds. It could be faster if we remove those debugging outputs, as they consume a lot of time.

[ShangkunLi]

Thanks for the awesome work~! I just have one concern:

The CGRA-mapper repo's heuristic is exactly your backtrackMap with max_location_to_try = All & max_backtrack_depth = 1, while its exhaustive is exactly your backtrackMap with max_location_to_try = All & max_backtrack_depth = All.

So for this repo, can we deprecate the current heuristic, but wrap your max_location_to_try = All & max_backtrack_depth = 1 as the new heuristic and treat your max_location_to_try = All & max_backtrack_depth = All as the new exhaustive?

Sure, will add it soon!

[ShangkunLi]

Thanks for the awesome work~! I just have one concern:
The CGRA-mapper repo's heuristic is exactly your backtrackMap with max_location_to_try = All & max_backtrack_depth = 1, while its exhaustive is exactly your backtrackMap with max_location_to_try = All & max_backtrack_depth = All.
So for this repo, can we deprecate the current heuristic, but wrap your max_location_to_try = All & max_backtrack_depth = 1 as the new heuristic and treat your max_location_to_try = All & max_backtrack_depth = All as the new exhaustive?

Sure, will add it soon!

In fact, in current heuristic, the max_location_to_try is also 1. If we set max_location_to_try = All & max_backtrack_depth = 1 as heuristic, we can also get an II=4, but the mapping time is ~30s.

Anyway, I will set max_location_to_try = All & max_backtrack_depth = 1 for new heuristic.