ICCAD04 benchmark legalization troubles

[profmadden]

I'm running some experiments with the ICCAD04 benchmark set, and am seeing some overlap between macro blocks. Is there a configuration setting to ensure legal placements? I'm using JSON configurations adapted from the ISPD05 benchmark set (no additional processing with ntuplace3 is done). I had to edit the ICCAD04 Bookshelf files a bit to get everything to parse cleanly; the changes were all minor formatting issues.

....
[ERROR  ] row 135 (2160, 2176), overlap node 12714 (1546, 2080, 1642, 2164) with node 5211 (1585, 2160, 1665, 2261) site (1585, 1665), gap 57
[INFO   ] Legality check takes 0.807194 ms
[ERROR  ] DREAMPlace - legality check failed in greedy legalization, return illegal results after greedy legalization.
[INFO   ] DREAMPlace - legalization takes 0.026 seconds
[INFO   ] DREAMPlace - iteration  458, wHPWL 1.956688E+06, time 0.166ms
[INFO   ] DREAMPlace - Start ABCDPlace for refinement
[ERROR  ] row 5 (80, 96), overlap node 12690 (1260, 80, 1356, 164) with node 12715 (1279, 0, 1375, 84) site (1279, 1375), gap 77
[ERROR  ] row 8 (128, 144), overlap node 4886 (1595, 32, 1675, 133) with node 12635 (1648, 128, 1711, 256) site (1648, 1711), gap 27
.....

Here's the JSON configuration. Thanks for any help!

{
"aux_input" : "benchmarks/iccad04/ibm01/ibm01.aux", 
"gpu" : 0,
"num_bins_x" : 512,
"num_bins_y" : 512,
"global_place_stages" : [
    {"num_bins_x" : 512, "num_bins_y" : 512, "iteration" : 1000, "learning_rate" : 0.01, "wirelength" : "weighted_average", "optimizer" : "nesterov", "Llambda_density_weight_itera
tion" : 1, "Lsub_iteration" : 1}
],
"target_density" : 1.0,
"density_weight" : 8e-5,
"gamma" : 4.0,
"random_seed" : 1000,
"scale_factor" : 1.0,
"ignore_net_degree" : 100,
"enable_fillers" : 1,
"gp_noise_ratio" : 0.025,
"global_place_flag" : 1,
"legalize_flag" : 1,
"detailed_place_flag" : 1,
"detailed_place_engine" : "", 
"detailed_place_command" : "", 
"stop_overflow" : 0.07, 
"dtype" : "float32", 
"plot_flag" : 0, 
"random_center_init_flag" : 1, 
"gift_init_flag" : 0, 
"sort_nets_by_degree" : 0, 
"num_threads" : 8,
"deterministic_flag" : 1
}

[limbo018]

Could you share a case with overlap problem? I cannot find ICCAD04 benchmarks online.

[profmadden]

I've put my edited ibm01 in https://www.cs.binghamton.edu/~pmadden/ibm01.tar.gz
I edited the ".pl" file to add "/FIXED" to the pads, to keep them in place, and clamped the pin positions in the ".nets" file to one decimal place of precision (the original sometimes used scientific notation -- pin positions were randomly assigned, and some of the random numbers were almost but not quite zero). There's a PDF of the placement result from my run.

The GSRC Bookshelf files can be found here: http://vlsicad.eecs.umich.edu/BK/Slots/
Check under the "wirelength driven standard cell placement" entry.

-- edit -- just realized I put the wrong PDF in the tar file; it's one from before I added the "FIXED" to the pads.

[limbo018]

I have figured out that the problem comes from non-integral-row-height movable macros. Forcing these macros align to rows and sites may cause overlaps. I may need some time to fix it.

[limbo018]

The bug should be fixed in the latest develop branch with commit (a834d72). Please let me know if it works on your side.

[profmadden]

Fantastic -- looks like that fixed it.

I'm trying to get all sorts of results organized for a paper on current-state-of-the-art mixed size placement. DREAMPlace gets used quite a bit in AI/ML placement research, and I'm not sure authors from that area know when results are incorrect, or that they've run into pathological behavior in a stock EDA tool. There are also a bunch of weird things in the ICCAD04 suite, and I'm planning on posting a "cleaned up" version that should work correctly with modern mixed-size placement tools.

A heads-up on some things I ran into during the build -- I'm using NumPy 2.0, and they've removed np.string_, replacing it with nm.bytes_.

  File "/home/pmadden/dreamplace/dreamplace/PlaceDB.py", line 512, in initialize_from_rawdb
    self.node_names = np.array(pydb.node_names, dtype=np.string_)
                                                      ^^^^^^^^^^
  File "/home/pmadden/venv/lib/python3.12/site-packages/numpy/__init__.py", line 781, in __getattr__
    raise AttributeError(
AttributeError: `np.string_` was removed in the NumPy 2.0 release. Use `np.bytes_` instead.

I made the edits to my PlaceDB.py, and it runs w/o trouble. I also ran into a bit of compile error with OpenTimer, and hacked unittest/doctest.h. I'm not sure which g++ flag needs to be set, but it didn't like the static variable sized altStackMem. The stand-alone OpenTimer compiles, and I have not dug into the CMake files to figure out where the flags differ.

    struct FatalConditionHandler
    {
        static bool             isSet;
        static struct sigaction oldSigActions[DOCTEST_COUNTOF(signalDefs)];
        static stack_t          oldSigStack;
        static char             altStackMem[4 * 1024];
      //static char             altStackMem[4 * SIGSTKSZ];              

Thanks very much for the help!

[profmadden]

Aaah! I just realized something (must be a lot of dust in my brain -- these are some very old benchmarks!).

For the ICCAD04 benchmarks, all of the macros have integer row heights (in the non-rotated orientations). Row height is 16, but site width is 1. I think that Igor Markov and Saurabh Adya were not expecting the macro blocks to rotate. Is there a way to turn off macro block rotation?

[limbo018]

I found the problem. DREAMPlace itself does not support rotation, but it will follow the rotation given by the input. Take ibm01 as an example.

In ibm01.nodes:

a10461     101    80

In ibm01.pl:

a10461   2051.79    1890.33  :  W

DREAMPlace will rotate this macro according to the rotation in the .pl file . If I simply ignore the rotations, it will change the behavior when running DREAMPlace for incremental placement.

The latest commit (67a191e) in the develop branch applies this change.

[profmadden]

Good catch! I've been tinkering with the benchmarks forever, and never realized that some blocks were not oriented N (my placer ignores any orientation for movable objects -- pads keep their orientation). I'll strip the orientations off in the "cleaned up" version, and I'll ask Igor and Saurabh what they were thinking (on the chance that they might remember why they did it.)

Thanks very much -- very very helpful!

[Ultra0312]

@profmadden Hello, I've been trying to do some experiments with the ICCAD04 benchmark set recently, but the bookshelf format I got doesn't seem to work with DREAMPlace, and I haven't been able to run DREAMPlace successfully on ICCAD04 benchmark. I would like to ask what changes you have made to the ICCAD04 benchmark? Thank you very much for your reply!

[limbo018]

I've put my edited ibm01 in https://www.cs.binghamton.edu/~pmadden/ibm01.tar.gz I edited the ".pl" file to add "/FIXED" to the pads, to keep them in place, and clamped the pin positions in the ".nets" file to one decimal place of precision (the original sometimes used scientific notation -- pin positions were randomly assigned, and some of the random numbers were almost but not quite zero). There's a PDF of the placement result from my run.

The GSRC Bookshelf files can be found here: http://vlsicad.eecs.umich.edu/BK/Slots/ Check under the "wirelength driven standard cell placement" entry.

-- edit -- just realized I put the wrong PDF in the tar file; it's one from before I added the "FIXED" to the pads.

@Ultra0312 You can refer to the modification in this sample case.

[profmadden]

@profmadden Hello, I've been trying to do some experiments with the ICCAD04 benchmark set recently, but the bookshelf format I got doesn't seem to work with DREAMPlace, and I haven't been able to run DREAMPlace successfully on ICCAD04 benchmark. I would like to ask what changes you have made to the ICCAD04 benchmark? Thank you very much for your reply!

IBM01 is up on the web, with the link above. I'll do the conversion on all of the benchmarks, and will post them -- I should be able to get that done this weekend. I'm in the middle of a fairly big benchmarking effort, and getting these benchmarks "right" is part of that.

The original files from the UMich site have some weird formatting. Pin positions were generated randomly, and in some cases the floating point locations of the pins are printed in scientific notation(!), which breaks some Bookshelf parsers. The /FIXED indicator lets Dreamplace keep the pads fixed (my code just checked to see if something was marked as a terminal -- some variation in how different placement groups handled things).

[Ultra0312]

I've put my edited ibm01 in https://www.cs.binghamton.edu/~pmadden/ibm01.tar.gz I edited the ".pl" file to add "/FIXED" to the pads, to keep them in place, and clamped the pin positions in the ".nets" file to one decimal place of precision (the original sometimes used scientific notation -- pin positions were randomly assigned, and some of the random numbers were almost but not quite zero). There's a PDF of the placement result from my run.我将编辑后的 ibm01 放入 https://www.cs.binghamton.edu/~pmadden/ibm01.tar.gz 我编辑了“.pl”文件以将“/FIXED”添加到焊盘中，将它们固定到位，并将“.nets”文件中的引脚位置夹在小数点后一位精度（原始有时使用科学记数法 - 引脚位置是随机分配的，一些随机数几乎但不完全为零）。有一个我跑步的放置结果的 PDF。
The GSRC Bookshelf files can be found here: http://vlsicad.eecs.umich.edu/BK/Slots/ Check under the "wirelength driven standard cell placement" entry.GSRC 书架文件可在此处找到：http://vlsicad.eecs.umich.edu/BK/Slots/ 在“线长驱动标准单元放置”条目下进行检查。
-- edit -- just realized I put the wrong PDF in the tar file; it's one from before I added the "FIXED" to the pads.-- 编辑 -- 刚刚意识到我在 tar 文件中放错了 PDF;这是我在垫子上添加“FIXED”之前的一个。

@Ultra0312 You can refer to the modification in this sample case.您可以参考此示例案例中的修改。

Thanks for your reply. I am trying to make some modifications based on this.

[Ultra0312]

IBM01 is up on the web, with the link above. I'll do the conversion on all of the benchmarks, and will post them -- I should be able to get that done this weekend. I'm in the middle of a fairly big benchmarking effort, and getting these benchmarks "right" is part of that.IBM01 已上架，链接如下。我将对所有基准测试进行转换，并发布它们——我应该能够在本周末完成。我正在进行一项相当大的基准测试工作，让这些基准测试“正确”是其中的一部分。

The original files from the UMich site have some weird formatting. Pin positions were generated randomly, and in some cases the floating point locations of the pins are printed in scientific notation(!), which breaks some Bookshelf parsers. The /FIXED indicator lets Dreamplace keep the pads fixed (my code just checked to see if something was marked as a terminal -- some variation in how different placement groups handled things).来自 UMich 站点的原始文件有一些奇怪的格式。引脚位置是随机生成的，在某些情况下，引脚的浮点位置以科学记数法 （！） 打印，这破坏了一些 Bookshelf 解析器。/FIXED 指示器让 Dreamplace 保持打击垫固定（我的代码刚刚检查了某些内容是否被标记为终端——不同放置组处理事物的方式有所不同）。

Thanks for the reply. I did find the problem you said while trying to run DREAMPlace on IBM01. And I also found that in another version of ICCAD 04, i.e. benckmark in the form of a DEF/LEF file, there is also some information that is missing from the Bookshelf format. For example, about the type of cell instances, which cells are standard cells (CORE) and which are macro cells (BLOCK). I hope this discovery of mine can be helpful in your work, looking forward to your work!