Skip to content

Enhanced 3D Placement Cost and Router Lookahead with Cross-Die Delay #2348

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Merged
merged 274 commits into from
Nov 11, 2023
Merged

Enhanced 3D Placement Cost and Router Lookahead with Cross-Die Delay #2348

merged 274 commits into from
Nov 11, 2023

Conversation

amin1377
Copy link
Contributor

In this PR, I have introduced a new move to the placement stage, enabling the modification of a block's layer (in the second state of the agent). Additionally, the first three moves of the placement stage, which do not target the critical paths, have been enhanced in the following way. After calculating the x-y location, they now check whether the current layer has available space. If not, they search across layers to identify any free blocks at the designated x-y location. In case a free block is found, it is relocated to that layer; otherwise, the block remains in its current layer.

Furthermore, the router lookahead mechanism has been updated to account for the cost associated with crossing layers.

amin1377 and others added 24 commits July 6, 2023 12:54
@amin1377
decrease NUM_PL_NONTIMING_MOVE_TYPES NUM_PL_1ST_STATE_MOVE_TYPES by o…
974eeb5 
…ne since inter layer uniform is now performed in state
@amin1377
related to makeing unifrom inter die move being performed in state 2
a13e38d
@amin1377
pass a vild to_layer to find_loc_uniform
a55af92
@amin1377
remove redundant parts from inter layer uniform
c0e553f
@amin1377
check whether rt node has a connection to other layer
74fec5f
@amin1377
comment unused parameter - debug get_min_cross_layer_delay
8ad1c01
@amin1377
for the non-timing-driven moves, find a free layer at the given x-y l…
847aed9 
…ocation and move block to that layer
@amin1377
add inter layer uniform only if several layers are on FPGA
4eec20d
@amin1377
if multiple layers are not available set num_2nd_state_avail_moves to…
53193f3 
… NUM_PL_MOVE_TYPES - 1
@amin1377
add layer_num to t_pl_offset
f313ded
@amin1377
delete the assertion in has_path_to_sink that check sink type
6b83d54
@amin1377
if node is of type sink, add it to heap even if it is not on the same…
7f25d98 
… layer as sink
set layer num when high fanout bb is set
0a7d67a
fix move idx to move name index for uniform inter layer move
b270a01
@amin1377
if there is no switch assigned for inter-lie connection for connectio…
8e0e147 
…n box, assing the infinit delay in router lookahead
@amin1377
debug: node type was used instead of node layer
cacd06c
@amin1377
multiply the area of fpga die by the number of layers to get the corr…
690c41e 
…ect router area per tile
@amin1377
add swap op numbers to parsed file
74c61d1
@amin1377
put assertion to assert if uniform_inter_layer is called when there i…
ad11a17 
…sn't multiple layers
@amin1377
fix the white space for place stats
915be19
@amin1377
add layer check to is_legal_swap_to_location
3e6a734
@amin1377
get ref const
e8b593b
@amin1377
incorporate layer num in hashing t_pl_loc
335d3bd
@amin1377
fix the bug in UniformInterLayerMoveGenerator
476b879
@github-actions github-actions bot added VPR VPR FPGA Placement & Routing Tool libarchfpga Library for handling FPGA Architecture descriptions labels Jul 12, 2023
amin1377 added 20 commits November 3, 2023 11:15
@amin1377
fix a bug - comment on weighted_move layer
93fb9d7
@amin1377
comment on why we need to decreament number of sinks
977d98e
@amin1377
update has_path_to_sink to accomodate for sw block inter-layer connec…
015de7b 
…tions
@amin1377
forward declaration of static functions
dca4108
@amin1377
comment on has_path_to_sink
0a1c499
@amin1377
comment on the case that BB should cover all layers
b8617f4
@amin1377
comment on router delay profiler calculate_delay
78be619
@amin1377
comment on get_cost_from_src_opin
8135d80
@amin1377
comment on compute_router_src_opin_lookahead
1879400
@amin1377
remove redundant min func
122ab4d
@amin1377
comment on get_min_cross_layer_delay
17074d3
@amin1377
centroid loc is not always valid
f45c4d2
@amin1377
make format
b2c9e71
@amin1377
experiment: change std vec to array for num sink on each layer per net
9b1c60d
@amin1377
initialize the num_sink ds to number of nets
65fb007
@amin1377
initialize num_sink per layer to OPEN instead of zero
aa0b5e9
@amin1377
define max_num_layers and change place.cpp to use that
d18fc4f
@amin1377
change placer context to use max_num_layer const
c5178e4
@amin1377
add an assertion to check the maximum number of layers
114012a
@amin1377
fix a typo in number of layers
eb71943
@vaughnbetz
Copy link
Contributor

Needs format.
Try matrix.
Get QoR.
Post this PR: could try grabbing some cpu time by tuning SMALL_NET for incremental BB, and by defining a new fanout threshold at which we ignore bb costs/computations (would be a background task).

@amin1377
Copy link
Contributor Author

amin1377 commented Nov 10, 2023 •
edited
Loading

Titan QoR:
The first column represents the master branch. In the second column, the data structure storing the number of sinks on each layer per net is implemented as a vector of arrays. The third column represents the case where this structure is a 2D matrix. Considering that a 2D matrix offers slight improvements, I have modified the implementation to use it.
Screenshot from 2023-11-10 09-55-47

@vaughnbetz
Copy link
Contributor

Great! Matrix it is (more maintainable as its a dynamic data structure, and faster too).

@amin1377
Copy link
Contributor Author

@vaughnbetz This PR is ready to merge.

@vaughnbetz vaughnbetz merged commit d4bed03 into master Nov 11, 2023
@vaughnbetz vaughnbetz deleted the 3d_link_cost_in_placement branch November 11, 2023 01:25
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@vaughnbetz vaughnbetz vaughnbetz requested changes

@duck2 duck2 duck2 left review comments

Assignees
No one assigned
Labels
VPR VPR FPGA Placement & Routing Tool
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
3 participants
@amin1377 @vaughnbetz @duck2