Refactor the implementation logic of backtrack mapping

[ShangkunLi]

In this pr:

• Refactor the backtrack mapping logic to enable real backtrack mapping
• Specifically, the mapping logic for now is illustrated using an example below.

Suppose I want to map a simple DFG that has 3 operations, with the dependencies A->B->C.

1. I first get 3 available candidate locations for A based on calculateAward, and I map the one with the highest award.
2. Then I continue to get 2 available candidate locations for B based on calculateAward, and I also successfully map the one with the highest award.
3. Then I find 3 available candidate locations for C, but none of them can be mapped legally, so I need to backtrack to B to try to select another available location of B for mapping, and then try C's 3 locations again (maybe not 3, it depends on the result of calculateAward), to see which one can succeed, at this time we define the backtracking depth as 1.
4. If B's two locations also cannot work, we need to backtrack to A, try A's other locations, and then re-map B and C, at this time we define the backtracking depth as 2.
5. If if the max backtracking depth is reached, and we still haven't found any legal mapping, it will return false and try next II

[tancheng]

Plz explain the diff from existing code a little bit use your example.

[ShangkunLi]

Plz explain the diff from existing code a little bit use your example.

In current implementation,

dataflow/lib/NeuraDialect/Mapping/HeuristicMapping/HeuristicMapping.cpp

Line 71 in 3158147

for (int i = 0; i < locations_to_try; ++i) {

when p&r in current candidate location for C fails, it will restore to the mapping_state before mapping C and try different candidate locations for C

dataflow/lib/NeuraDialect/Mapping/HeuristicMapping/HeuristicMapping.cpp

Line 94 in 3158147

mappingstate_snapshot.restore(mapping_state);

But the backtracking_depth increases each time when I try new candidate locations for the current op. It is not correct, because when I try different locations for C, the backtracking depth should remain the same (In fact, this is not backtracking, because I just try different locations for C).

dataflow/lib/NeuraDialect/Mapping/HeuristicMapping/HeuristicMapping.cpp

Line 96 in 3158147

backtrack_depth++;

[tancheng]

Plz explain the diff from existing code a little bit use your example.

In current implementation,

dataflow/lib/NeuraDialect/Mapping/HeuristicMapping/HeuristicMapping.cpp

Line 71 in 3158147

for (int i = 0; i < locations_to_try; ++i) {

when p&r in current candidate location for C fails, it will restore to the mapping_state before mapping C and try different candidate locations for C

dataflow/lib/NeuraDialect/Mapping/HeuristicMapping/HeuristicMapping.cpp

Line 94 in 3158147

mappingstate_snapshot.restore(mapping_state);

But the backtracking_depth increases each time when I try new candidate locations for the current op. It is not correct, because when I try different locations for C, the backtracking depth should remain the same (In fact, this is not backtracking, because I just try different locations for C).

dataflow/lib/NeuraDialect/Mapping/HeuristicMapping/HeuristicMapping.cpp

Line 96 in 3158147

backtrack_depth++;

This sounds if we increase the backtracking steps to N * locations_to_try, it would be the same as your current PR's backtracking steps of N?

[ShangkunLi]

Plz explain the diff from existing code a little bit use your example.

In current implementation,

dataflow/lib/NeuraDialect/Mapping/HeuristicMapping/HeuristicMapping.cpp

Line 71 in 3158147

for (int i = 0; i < locations_to_try; ++i) {

when p&r in current candidate location for C fails, it will restore to the mapping_state before mapping C and try different candidate locations for C

dataflow/lib/NeuraDialect/Mapping/HeuristicMapping/HeuristicMapping.cpp

Line 94 in 3158147

mappingstate_snapshot.restore(mapping_state);

But the backtracking_depth increases each time when I try new candidate locations for the current op. It is not correct, because when I try different locations for C, the backtracking depth should remain the same (In fact, this is not backtracking, because I just try different locations for C).

dataflow/lib/NeuraDialect/Mapping/HeuristicMapping/HeuristicMapping.cpp

Line 96 in 3158147

backtrack_depth++;

This sounds if we increase the backtracking steps to N * locations_to_try, it would be the same as your current PR's backtracking steps of N?

No, in the previous implementation, there is no backtracking. We just try different locations for C without restoring the state to before mapping B. And the number of locations to try is min(locations_to_try, max_backtracking_depth).