//TU190
In this paper, the author presents a detailed placement algorithm that
can handle both row-based standard cell placement and placement in
presence of fixed macros. This detailed placer works on legalized
placement to further improve the solution quality. It consists of a
Global Swap technique to swap cells based on their optimal region, a
Vertical Swap technique to fix the local errors vertically, a fast
Local Re-ordering technique to fix the local order horizontally, and
a Single-Segment Clustering technique that can place the cells in a
segment optimally when cells not in this segment are fixed. Experiments
on two sets of benchmarks show that this detailed placer can achieve
better wirelength and runtime compared to Fengshui 5.0, rowIroning
and Domino.
1. The author claims four main contributions in this paper: Global
Swap, Vertical Swap, Local Re-ordering, and Single-Segment Clustering.
However, three of these ideas (Global Swap, Local Re-ordering, and
Single-Segment Clustering) are very similar to those in the detailed
placement of "FastPlace: An Analytical Placer for Mixed-Mode Designs"
published in ISPD05. This paper just contributes more implementation
details about these ideas and provides a faster algorithm for
Fixed-Order Single Segment Placement Problem.
2. The experimental results in this paper may not be an appropriate
comparison. The author used the Placement Utilities in Capo9.1 to
legalize MPL5's global placement result. The placement result was
too much disturbed by the legalizer and shouldn't be used to compare
the performance of detailed placer. The legalized wirelength by
MPL5's own legalizer is 1.70e+06, which is even better then the
wirelength (1.78e+06) result generated by the detailed place stated
in this paper. The claim in the abstract (a legalized state-of-the-art
global placement solution may still have a lot of room for
improvement in wirelength) is not appropriate.
3. The paper claims that Global Swap gives the result most profit,
but there isn't any experimental result in this. The readers cannot
distinguish the individual contributions of the algorithms proposed
in this paper.
//HI112
A three-step mixed-size detailed placement is described in this paper.
A combination of constraint graph and linear programming is used to
remove the overlap between macros. Standard cells are legalized by
an enhanced greedy method. In the end, a slidingwindow-based cell
swapping is applied to further reduce wirelength. The experimental
results show that it can produce legal placement for all the circuits
in ICCAD04-MS, while Fengshui fails for 1/3 of them, and the
wirelength produced using its algorithm is shorter by 3% on average.
1. The paper proposes a macro legalization method with a combination
of constraint graph and linear programming. However, the experimental
results show very little wirelength improvement compared with a
simple greedy legalizer. The author claimed the method one of the
two main contributions, which does not seem an efficient way to macro
legalization.
2. The paper claims that it can produce legal placement for all the
circuits in ICCAD04-MS, while Fengshui fails for 1/3 of them. This
is not a fair comparison while the global placement results are
generated by an unknown analytical placer. Why not just use Fengshui's
global placement result for comparison? The experimental results will
be much more convincing if the proposed method could generate better
result on Fengshui's global placement compared with Fengshui's own
detailed placer.
--
103 Jao-Chang Xsi,
Department of Fool Engineering
National TigerLand University,
Taipei, Taiwan.
--
※ 發信站: 批踢踢實業坊(ptt.cc)
◆ From: 140.112.48.60
※ 編輯: tellux 來自: 140.112.48.60 (05/29 00:51)