FlatRecon: Flat Placement Reconstruction Full Legalizer #3193

haydar-c · 2025-07-08T11:34:47Z

This PR brings the FlatRecon full legalizer. It aims to reconstruct a given flat placement with minimum disturbance to the given solution. It can be used with a solution read from an '.fplace' file or with Global Placement output. However, it expects the given solution to be close to legal in both cases.

It has beed run on the MCNC, VTR Chain, Koios, and Titan benchmarks. It also been tested on the elfPlace placements generated from the aug-elfPlace on Titan benchmarks. [Results to be appended]

I have added 2 regression tests for that legalizer. The first into the vtr_reg_strong with fast checks on MCNC benchmarks and the second one into the vtr_reg_nightly_test7 with some of the Titan benchmarks.

The logged output of that legalizer looks like below for an example of the LU32PEEng.v from VTR Chain:

This works well for the reconstruction of provided flat placement. However, if used directly after GP output, if we have too many orphan molecules the runtime goes really high.

The legalizations strategy of cluster_legalizer is set to SKIP_LB_ROUTING for the first passs. After first pass, each checked and illegal ones destroyed and their molecules are passed to second pass. The legalization strategy is set to FULL after first pass in that commit. Be careful with the place where you compress the legalizer and extract the atom lookup for placement.

Before that, the strategy was converted to full after the first reconstruction pass. After that, I was doing a neighbour search with full strategy. Now, I am doing neighbour search with SKIP_LB_ROUTING as well. After that, it requires last neighbour pass with FULL strategy. Also added ugly timer prints to be cleaned.

Changed the way of processing mainly in the first pass. In this version, they are first sorted by external pin number and then grouped into tiles. Then processed similar to naive by tiles but this one still checks the compatibility and capacity. Cleaning after each tile clusters created in that pass. Then, checking each cluster legality after they are created in the first neighbour pass. Second is already done with FULL strategy. The memory usage is decreased by nearly by %35. We lost runtime but it seems near 1-2%. TODO for that one: I added an if in neighbour pass that checks if the molecule is already clustered before checking its compatibility. This works now but ideally my algorithm should not try to add an already clustered molecule to another one. When removed, it failed on vtr_chain_largest for only LU32PEEng.v.

This is the version that results are presented in 2 May Friday meeting. The memory improved version is presented with master merged. We have things to try on top of that version. Handling the illegal clusters right away. Also added a 0.5f buffer into fixed blocks in partial placement verification to pass the assert. This should be handled more explicitely later.

In this version, after doing my packing for reconstruction, I am calling the initial_placement for ap flow. The results seems promising for this version and being added to the presentation. Especially reduces the max displacement as expected.

The displacement for each atom was being calculated as distance to the head of the tile. Now it considers the offset of the tile as well. That affects the percent of atoms displaced, max atom displacemetn and average atom displacement. Did no touched the cluster error for now. This can also be rewritten. Instead of being calculated solely on cluster location and centroid, we can try to use cluster information.

…terId

first pass if any failing cluster occurs In the first pass of reconstruction, creating clusters at a tile and checking whether they are legal or not. If illegal, try to cluster with FULL strategy there before going to next tile. Molecules being still unclustered passed to neighbour pass.

…terId

Prioritizing chained molecules in the reconstruction pass. Ensuring initial placer placing the clusters created in reconstruction pass first. Its current sorting also results in the same ordering due to standart deviation and size ordering. However, if used in reconstruction legalizer, I want to ensure that clusters created in reconstruction pass processed first.

…terId

Prioritizing the long carry chains in the reconstruction pass.

vaughnbetz · 2025-08-21T19:39:11Z

The output looks nice, except the neighbour clustering step doesn't print detailed stats while the other two steps do. I suspect that is an oversight @haydar-c ?

haydar-c · 2025-08-21T22:30:29Z

The output looks nice, except the neighbour clustering step doesn't print detailed stats while the other two steps do. I suspect that is an oversight @haydar-c ?

Good catch. This was intentional but I see how it reads like an omission.
In the neighbor clustering stage we never create new clusters; we only join molecules to already created clusters. So I included its contribution only in the “molecules clustered in each stage breakdown" and left “clusters created” blank.

To make this explicit, maybe I should update the summary to always show “0 clusters created” for this stage.

vaughnbetz

Looks good overall; some suggested changes though.
After resolving the comments, you should re-run the full tests / QoR runs (nothing should change much, but we should be sure).
I suggest adding a small VTR design (spree) to the flat_recon tests in the basic tests. spree is very small (1229 primitives) but has DSP and RAM as well as logic so it gives good code coverage. Putting it in the basic tests also means we'll run the sanitized tests on it. Check it is fast though, as the sanitized tests slow down by 10x or more.

vaughnbetz · 2025-08-22T15:42:08Z

doc/src/vpr/command_line_usage.rst

+
+    * The x, y, and sub_tile location of the cluster that contains this atom.
+    * The flat site index of this atom in its cluster. The flat site index is a
+      linearized ID of primitive locations in a cluster. This may be used as a


I think we should remove flat site index.
Add the file format (with a short example) to the file formats list in the appropriate .rst.

vaughnbetz · 2025-08-22T15:42:49Z

doc/src/vpr/command_line_usage.rst

@@ -1291,6 +1303,12 @@ Analytical Placement is generally split into three stages:

    * ``appack`` Use APPack, which takes the Packer in VPR and uses the flat atom placement to create better clusters.

+    * ``flat-recon`` Use the Flat Placement Reconstruction Full Legalizer which tries to reconstruct a clustered placement that is
+      as close to the incoming flat placement as possible. It can be used to read a flat placement from a ``.fplace`` file (see :option:`--read_flat_place`)


Add .fplace format description to the file format documentation.

vaughnbetz · 2025-08-22T15:47:48Z

doc/src/vpr/command_line_usage.rst

@@ -1291,6 +1303,12 @@ Analytical Placement is generally split into three stages:

    * ``appack`` Use APPack, which takes the Packer in VPR and uses the flat atom placement to create better clusters.

+    * ``flat-recon`` Use the Flat Placement Reconstruction Full Legalizer which tries to reconstruct a clustered placement that is
+      as close to the incoming flat placement as possible. It can be used to read a flat placement from a ``.fplace`` file (see :option:`--read_flat_place`)
+      or with Global Placement output. In both cases, it expects the given solution to be close to legal. If used with a ``.fplace`` file (see :option:`--read_flat_place`),


or on the (in memory) output of VTR's integrated Global Placement algorithm

vaughnbetz · 2025-08-22T15:49:37Z

doc/src/vpr/command_line_usage.rst

+    * ``flat-recon`` Use the Flat Placement Reconstruction Full Legalizer which tries to reconstruct a clustered placement that is
+      as close to the incoming flat placement as possible. It can be used to read a flat placement from a ``.fplace`` file (see :option:`--read_flat_place`)
+      or with Global Placement output. In both cases, it expects the given solution to be close to legal. If used with a ``.fplace`` file (see :option:`--read_flat_place`),
+      each atom of a molecule should share same location information. It is legal to leave some molecules unconstrained; the reconstruction phase will choose where


each atom of a molecule should share --> each atom in a molecule should have compatible location information.

vaughnbetz · 2025-08-22T15:51:10Z

libs/libarchfpga/src/physical_types.h

+     */
+    friend bool operator<(const t_physical_tile_loc& lhs, const t_physical_tile_loc& rhs) {
+        if (lhs.layer_num != rhs.layer_num) return lhs.layer_num < rhs.layer_num;
+        if (lhs.x != rhs.x) return lhs.x < rhs.x;


Split into two lines:
if (lhs.x != rhs.x)
return lhs.x < rhs.x

vaughnbetz · 2025-08-22T17:36:57Z

vpr/src/analytical_place/full_legalizer.cpp

+
+    // Cast the partial placement to flat placement here. So that it can be
+    // used to guide the initial placer and for logging results. This enables
+    // the flow to be used with direct output of GP as well.


Explain p_placement has been set by the GP (if using internal VTR AP), or by reading in the flat placement file. Cast / copy it to the flat_placement data structures so we can always use them.

vaughnbetz · 2025-08-22T17:37:29Z

vpr/src/analytical_place/full_legalizer.cpp

+    }
+
+    // Run the initial placer on the clusters created.
+    // TODO: Currently, the way initial placer sort the blocks to place is aligned


sort -> sorts

vaughnbetz · 2025-08-22T17:38:00Z

vpr/src/analytical_place/full_legalizer.cpp

+
+    // Run the initial placer on the clusters created.
+    // TODO: Currently, the way initial placer sort the blocks to place is aligned
+    //       how self clustering pass clusters created, so there is no need to explicitely


pass -> passes the
explicitely -> explicitly

vaughnbetz · 2025-08-22T17:43:29Z

vpr/src/base/read_options.cpp

@@ -1841,6 +1841,11 @@ argparse::ArgumentParser create_arg_parser(const std::string& prog_name, t_optio
            "VPR's (or reconstructed external) placement solution in flat placement file format; this file lists cluster and intra-cluster placement coordinates for each atom and can be used to reconstruct a clustering and placement solution.")
        .show_in(argparse::ShowIn::HELP_ONLY);

+    file_grp.add_argument(args.write_legalized_flat_place_file, "--write_legalized_flat_place")
+        .help(
+            "VPR's (or reconstructed external) placement solution after legalization and before anneal in flat placement file format; this file lists cluster and intra-cluster placement coordinates for each atom and can be used to reconstruct a clustering and placement solution.")


lists (x, y, layer) coordinates for each atom
(we aren't using intra-cluster coordinates)

vpr/src/pack/pack.h

site index from flat placement info data structure.

…n referenced.

…header.

…terId

CMakeLists.txt

…odin.

…terId

vaughnb-cerebras

Looks good, thanks! This is a great new feature.

haydar-c added 30 commits March 28, 2025 16:08

pass 1st, 2nd, 3rd with a rough implementation of grids

c27bacc

tile compability with destroying incompatible ones

1535d9a

tile compability with checking before creating cluster

ef8c354

instepting speed with simple pass

3d26bfb

simplified first pass and search grids in manhattan

cdff896

sorting before processing

c67da89

added placement step

db36eff

reporting total clusters

e737efa

corrected reporting

2c04563

added neigbour search for cluster creation as well

41229b5

added neighbour search for orphan clusters as well

924073f

This works well for the reconstruction of provided flat placement. However, if used directly after GP output, if we have too many orphan molecules the runtime goes really high.

just added parsin max_rss for FL and whole run

72760c0

Merged master into this branch, solved conflict in qor parsin for ap

37a42fa

Makes my changes compilable after master merge

5761983

Merge branch 'master' into reconstruction_grids_with_LegalizationClus…

c02d48c

…terId

Merge branch 'master' into reconstruction_grids_with_LegalizationClus…

942274a

…terId

[AP][FL] Added Stats for debugging block num increase

d7a95b9

Stat reporting for reconstruction

a2efbe3

[AP][FL] Corrected displacement tile and atom reporting

1183393

Debugging the discrepancy between initial_placement and first pass

28b5836

Merge branch 'master' into reconstruction_grids_with_LegalizationClus…

5ba4a95

…terId

[AP] Reconstruction Legalizer

5ad9f8d

Prioritizing the long carry chains in the reconstruction pass.

haydar-c added 2 commits August 18, 2025 17:04

FlatRecon: Update parsing for new summmary and results.

292a54e

Merge master into FlatRecon

f63113d

github-actions bot added build Build system lang-make CMake/Make code labels Aug 18, 2025

FlatRecon: Update get_flat_placement_files.py for linting.

c2565c4

vaughnbetz requested changes Aug 22, 2025

View reviewed changes

haydar-c added 11 commits August 24, 2025 16:55

FlatRecon: Fix a small bug in neighbor pass and remove block

ddf7e9d

site index from flat placement info data structure.

FlatRecon: Added .fplace file format description and linked to it whe…

0dbada4

…n referenced.

FlatRecon: Addressing documenting comments.

87b7cd6

FlatRecon: Add detailed doxygen comments and fix issues on legalizer …

965bf79

…header.

FlatRecon: remaining comments in the full legalizer cpp code.

4ea6b57

FlatRecon: Clearing clusterID data after compress.

69ad840

FlatRecon: Added a test to basic test (spree).

386dcc5

FlatRecon: Updated strong test result and for omitting site idx.

85414c6

FlatRecon: Updated the nighhtly test7 flat files and results.

8de8d63

Merge branch 'master' into reconstruction_grids_with_LegalizationClus…

0cfbaba

…terId

FlatRecon: Continue if only 1 atom of a mol provided in fplace

b876cda

AlexandreSinger reviewed Aug 26, 2025

View reviewed changes

CMakeLists.txt Show resolved Hide resolved

FlatRecon: Added extraction of flat placements to CI

913aeb3

github-actions bot added the infra Project Infrastructure label Aug 26, 2025

haydar-c added 2 commits August 26, 2025 11:45

FlatRecon: Move spree basic test from vtr_reg_basic to vtr_reg_basic_…

615520f

…odin.

FlatRecon: Use blif file for the basic odin test

5f95301

github-actions bot added the lang-netlist label Aug 27, 2025

Merge branch 'master' into reconstruction_grids_with_LegalizationClus…

ed57b5d

…terId

vaughnb-cerebras approved these changes Aug 27, 2025

View reviewed changes

FlatRecon: Small commenting fix.

11b4e66

vaughnbetz merged commit 4eec9fe into master Aug 28, 2025
30 checks passed

vaughnbetz deleted the reconstruction_grids_with_LegalizationClusterId branch August 28, 2025 17:06

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

FlatRecon: Flat Placement Reconstruction Full Legalizer #3193

FlatRecon: Flat Placement Reconstruction Full Legalizer #3193

Uh oh!

haydar-c commented Jul 8, 2025 •

edited

Loading

Uh oh!

vaughnbetz commented Aug 21, 2025

Uh oh!

haydar-c commented Aug 21, 2025

Uh oh!

vaughnbetz left a comment

Uh oh!

vaughnbetz Aug 22, 2025

Uh oh!

vaughnbetz Aug 22, 2025

Uh oh!

vaughnbetz Aug 22, 2025

Uh oh!

vaughnbetz Aug 22, 2025

Uh oh!

vaughnbetz Aug 22, 2025

Uh oh!

vaughnbetz Aug 22, 2025

Uh oh!

vaughnbetz Aug 22, 2025

Uh oh!

vaughnbetz Aug 22, 2025

Uh oh!

vaughnbetz Aug 22, 2025

Uh oh!

Uh oh!

Uh oh!

vaughnb-cerebras left a comment

Uh oh!

Uh oh!

Uh oh!

FlatRecon: Flat Placement Reconstruction Full Legalizer #3193

FlatRecon: Flat Placement Reconstruction Full Legalizer #3193

Uh oh!

Conversation

haydar-c commented Jul 8, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

vaughnbetz commented Aug 21, 2025

Uh oh!

haydar-c commented Aug 21, 2025

Uh oh!

vaughnbetz left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

vaughnb-cerebras left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

haydar-c commented Jul 8, 2025 •

edited

Loading