Placement Refactoring

[amin1377]

This is a work in progress. The full description will be written soon.

[vaughnbetz]

Looks promising!

1. Resolve conflicts
2. Run koios & VTR to get more CPU times. These ones are looking promising.
3. If there's a runtime gap, we should then profile to see if there are any obvious hotspots. But we don't need to get exact runtime parity if it proves really hard.

[vaughnbetz]

Please attach Koios and VTR run results.

[vaughnbetz]

Thanks for the big refactoring @amin1377 !
I have a lot of comments though. Here is a summary:

Things to definitely do before merging this PR:

• Doxygen comments: expand, describe parameters, unify comments where we have a function-scope comment both at the declaration and the definition (prefer declaration).
• A few time-critical asserts should become VTR_ASSERT_SAFE or VTR_ASSERT_DEBUG
• Various specific code comments.

Things to consider, but to do in a separate PR with QoR measurements on each:

• We could clean up / organize some of the file scope variables by putting them in structures.
• Could we use the .size() member of a ts_ vector instead of num_nets_affected?
• We can speed up and simplify the bounding box calculation a bit, and make it a little more accurate, by getting rid of the min/max code that keeps all bounding boxes in the [1..width-1] range.
• Where is the z-directed bounding box cost, and the computation of the z-directed channel normalization factors?
• Do we need the num_sinks_per_layer[] data structure or could we get rid of it (it leads to O(layer) computations).

[vaughnbetz]

All these noc cost checks can be shortened by using vtr::isclose with an absolute tolerance of MIN_EXPECTED_NOC_LATENCY_COST etc. (i.e. set the last argument to that value). So the if statements can be taken out, the check_and_print_cost lambda should take one more parameter (absolute cost threshold) and that parameter should be passed to vtr::isclose().

[vaughnbetz]

Can file an issue to @soheilshahrouz if you want him to clean this up.

[amin1377]

QoR on all koios circuits:

[vaughnbetz]

You have a warning that is making the various build types fail (we turn on -Werror):
[ 81%] Building CXX object vpr/CMakeFiles/libvpr.dir/src/place/net_cost_handler.cpp.o
/home/runner/work/vtr-verilog-to-routing/vtr-verilog-to-routing/vpr/src/place/net_cost_handler.cpp:243:13: error: "/" within comment [-Werror=comment]
243 | * @details / Updates the bounding box of a net by storing its coordinates in the bb_coord_new data structure and

[vaughnbetz]

Thanks @amin1377 ! Once you have as many suggestions resolved as you want for this PR (others can go to separate issues), you'll need to run full QoR sweeps:

• VTR, Titan, Koios on usual 2D architecture each
• 3D architecture with 3D BB
• 3D architecture with layer BB

[vaughnbetz]

Adding @soheilshahrouz to the conversation, as he is interested in trying parallel "population annealing", and to do that he would need non-global objects for the placement state and the placement cost.

[amin1377]

QoR after applying comments:
Baseline: Master branch (121a16c)
Placement Refac Branch (d000874)
Titan:

Large VTR: