Add GroupBinding synthesis process decision

[ryukzak]

Add a new type of synthesis process decision: GroupBinding.

Currently, binding on the Network level can be implemented only one by one, which is inefficient and creates many duplicated synthesis nodes. E.g.: Bind "mul" "a*b=c"; Bind "accum" "d + e = f" is equal to Bind "accum" "d + e = f"; Bind "mul" "a*b=c".

Or more complicated example (rows - bind options, columns - function, cells - bind to PUs):

|   | buf1         | buf2        |
|---+--------------+-------------|
| 1 | fram1 fram2  |             |
| 2 | fram1        | fram2       |
| 3 | fram2        | fram1       |
| 4 |              | fram1 fram2 |

You can see four variants, but really we have only two. 1 is equal to 4, 2 to 3.

We can resolve this problem by introducing GroupBinding, which allows making several binding at once. Possible subtypes:

• NonAlternativeBinds - all bind without alternative (if we have only one multiplier PU, we know who will do all multiplications).
• AllBinds - make all available binds at once. For example, from the table, we will have two options.
• FirstWaveBinds - make all bindings for the first wave functions (first wave - functions which can be executed right now).

Additional, we need to design some metrics for comparing AllBinds and FirstWaveBinds options.

Metrics ideas:

• standard deviation between similar PU loading in functions, in receiving variables, in sending variables; their average value for all PU.
• number of self sending for each PU, and average value.
  • How to calculate it with optimization into account?

[ryukzak]

Future work:

• Group binding: reduce combinatorial complexity and improve quality #255