feat: add VLSI Global Placement benchmark (Frontier-Eng)

.gitattributes

@@ -0,0 +1 @@
+benchmarks/ElectronicDesignAutomation/VLSIGlobalPlacement/references/*.json.gz binary

benchmarks/ElectronicDesignAutomation/VLSIGlobalPlacement/.gitignore

@@ -0,0 +1,13 @@
+
+# Temp working directories
+temp/
+
+# Debug/fix scripts
+_fix_*.py
+_check_*.py
+_debug_*.py
+
+# Evaluation artifacts
+artifacts.json
+metrics.json
+

benchmarks/ElectronicDesignAutomation/VLSIGlobalPlacement/README.md

@@ -0,0 +1,117 @@
+# VLSI Global Placement
+
+Global placement is a critical stage in VLSI (Very Large Scale Integration) physical design.
+After logic synthesis and floorplanning, standard cells and macros must be placed on the chip
+such that wirelength is minimized while respecting physical constraints.
+
+This benchmark uses the **ISPD 2005** placement contest benchmarks, the industry-standard
+open-source benchmark suite for VLSI placement. The agent must implement a placement algorithm
+that minimizes Half-Perimeter Wirelength (HPWL) without violating hard constraints.
+
+## File Structure
+
+```text
+VLSIGlobalPlacement/
+├── .gitignore                        # Git ignore rules
+├── datasets/                         # Raw ISPD 2005 Bookshelf data
+│   └── ispd2005/                     # (empty; preprocessed JSONs are in references/)
+├── README.md                         # Navigation doc (this file)
+├── README_zh-CN.md                   # Navigation doc (Chinese)
+├── Task.md                           # Detailed task description
+├── Task_zh-CN.md                     # Detailed task description (Chinese)
+├── references/                       # Benchmark reference data
+│   ├── adaptec1.json.gz                 # Easy benchmark (~211k cells, gzip)
+│   ├── adaptec1_difficulty.json       # Difficulty metadata
+│   ├── adaptec3.json.gz                 # Medium benchmark (~451k cells, gzip)
+│   └── adaptec3_difficulty.json       # Difficulty metadata
+├── scripts/
+│   ├── init.py                        # [MODIFIABLE] Placement algorithm
+│   └── preprocess.py                  # Bookshelf -> JSON converter
+├── verification/
+│   ├── evaluator.py                   # Scoring and legality checks
+┬   ├── test_evaluator.py             # Unit tests
+│   ├── requirements.txt               # Python dependencies
+│   └── docker/
+│       └── Dockerfile                 # Containerized evaluation
+├── baseline/
+│   └── solution.py                    # Row-based placement baseline
+┬   └── result_log.txt                 # Baseline evaluation results
+└── frontier_eval/                     # Unified task metadata
+    ├── initial_program.txt
+    ├── eval_command.txt
+    ├── agent_files.txt
+    ├── artifact_files.txt
+    ├── readonly_files.txt
+    ├── copy_files.txt
+    ├── candidate_destination.txt
+    ├── eval_cwd.txt
+    ├── constraints.txt
+    └── run_eval.py
+```
+
+## Quick Start
+
+### 1. Install Dependencies
+
+```bash
+pip install -r verification/requirements.txt
+```
+
+### 2. Run the Baseline Solver
+
+```bash
+cd benchmarks/ElectronicDesignAutomation/VLSIGlobalPlacement
+python scripts/init.py
+# Outputs: temp/submission.json
+```
+
+### 3. Evaluate a Candidate Program
+
+```bash
+cd benchmarks/ElectronicDesignAutomation/VLSIGlobalPlacement
+python verification/evaluator.py scripts/init.py --benchmark adaptec1
+```
+
+### 4. Run with Unified Task Framework
+
+```bash
+python -m frontier_eval   task=unified   task.benchmark=ElectronicDesignAutomation/VLSIGlobalPlacement   algorithm=openevolve   algorithm.iterations=0
+```
+
+## Benchmarks
+
+| Name | Difficulty | Fixed Cells | Movable Cells | Nets | Pins | Die Size |
+|------|-----------|-------------|---------------|------|------|----------|
+| adaptec1 | Easy | 543 | 210,904 | 221,142 | 944,053 | 11589x11589 |
+| adaptec3 | Medium | 723 | 450,927 | 466,758 | 1,875,039 | 23190x23386 |
+
+## Task Summary
+
+- **Input**: Die dimensions, cell library, fixed/movable cells, netlist, initial placement
+- **Output**: (x, y) coordinates for every movable cell
+- **Hard Constraints**: No fixed cells moved, no cells out of bounds, no overlaps
+- **Optimization Objective**: Minimize Half-Perimeter Wirelength (HPWL)
+- **Editable File**: scripts/init.py (only place_components() function)
+
+## Dataset License
+
+The ISPD 2005 benchmarks were created by the ICCAD 2005 / ISPD 2006 placement contest committees
+and are freely available for academic use.
+
+## Compressed JSON Format
+
+The reference files are gzip-compressed JSON with a compact netlist representation.
+Each net is stored as a list of integer cell indices rather than full pin dictionaries:
+
+```json
+{"netlist": [[0, 1, 2], [3, 4], ...]}
+```
+
+The compact netlist reduces JSON size by approximately 65% compared to the verbose
+format, and gzip further reduces the on-disk size by about 84% (adaptec1: 22.6MB ->
+3.7MB, adaptec3: 48.2MB -> 7.9MB). The `_decompress_netlist()` function in
+scripts/init.py and verification/evaluator.py reconstructs the full pin
+dictionaries at load time. The transformation is lossless with respect to the
+HPWL computation. The scripts/preprocess.py script generates this compressed
+format directly from the original Bookshelf data.
+

benchmarks/ElectronicDesignAutomation/VLSIGlobalPlacement/README_zh-CN.md

@@ -0,0 +1,116 @@
+# VLSI 全局布局
+
+全局布局是 VLSI（超大规模集成电路）物理设计中的关键阶段。
+在逻辑综合和布图规划之后，标准单元和宏单元必须放置在芯片上，
+以最小化线长，同时满足物理约束。
+
+本基准测试使用 **ISPD 2005** 布局竞赛基准，这是 VLSI 布局领域
+行业标准的开源基准套件。Agent 必须实现一个最小化半周长线长（HPWL）
+的布局算法，同时不违反硬约束。
+
+## 文件结构
+
+```text
+VLSIGlobalPlacement/
+├── .gitignore                        # Git 忽略规则
+├── datasets/                         # 原始 ISPD 2005 Bookshelf 数据
+│   └── ispd2005/                     # （空目录；预处理后的 JSON 在 references/ 中）
+├── README.md                         # 导航文档（英文）
+├── README_zh-CN.md                   # 导航文档（中文，本文件）
+├── Task.md                           # 详细任务描述（英文）
+├── Task_zh-CN.md                     # 详细任务描述（中文）
+├── references/                       # 基准参考数据
+│   ├── adaptec1.json.gz                  # 简单基准（约21万单元，gzip）
+│   ├── adaptec1_difficulty.json       # 难度元数据
+│   ├── adaptec3.json.gz                  # 中等基准（约45万单元）
+│   └── adaptec3_difficulty.json       # 难度元数据
+├── scripts/
+│   ├── init.py                        # [可修改] 布局算法
+│   └── preprocess.py                  # Bookshelf 格式转 JSON
+├── verification/
+│   ├── evaluator.py                   # 评分和合法性检查
+│   ├── requirements.txt               # Python 依赖
+│   └── docker/
+│       └── Dockerfile                 # 容器化评测
+├── baseline/
+│   └── solution.py                    # 行式放置基线
+└── frontier_eval/                     # Unified task 元数据
+    ├── initial_program.txt
+    ├── eval_command.txt
+    ├── agent_files.txt
+    ├── artifact_files.txt
+    ├── readonly_files.txt
+    ├── copy_files.txt
+    ├── candidate_destination.txt
+    ├── eval_cwd.txt
+    ├── constraints.txt
+    └── run_eval.py
+```
+
+## 快速开始
+
+### 1. 安装依赖
+
+```bash
+pip install -r verification/requirements.txt
+```
+
+### 2. 运行基线求解器
+
+```bash
+cd benchmarks/ElectronicDesignAutomation/VLSIGlobalPlacement
+python scripts/init.py
+# 输出: temp/submission.json
+```
+
+### 3. 评估候选程序
+
+```bash
+cd benchmarks/ElectronicDesignAutomation/VLSIGlobalPlacement
+python verification/evaluator.py scripts/init.py --benchmark adaptec1
+```
+
+### 4. 使用 Unified Task 框架运行
+
+```bash
+python -m frontier_eval \
+  task=unified \
+  task.benchmark=ElectronicDesignAutomation/VLSIGlobalPlacement \
+  algorithm=openevolve \
+  algorithm.iterations=0
+```
+
+## 基准数据
+
+| 名称 | 难度 | 固定单元 | 可移动单元 | 网络数 | 引脚数 | 芯片尺寸 |
+|------|------|---------|-----------|--------|--------|---------|
+| adaptec1 | 简单 | 543 | 210,904 | 221,142 | 944,053 | 11589x11589 |
+| adaptec3 | 中等 | 723 | 450,927 | 466,758 | 1,875,039 | 23190x23386 |
+
+## 任务概要
+
+- **输入**：芯片尺寸、单元库、固定/可移动单元、网表、初始布局
+- **输出**：每个可移动单元的 (x, y) 坐标
+- **硬约束**：不移动固定单元、所有单元在芯片内、无重叠
+- **优化目标**：最小化半周长线长（HPWL）
+- **可编辑文件**：scripts/init.py（仅 place_components() 函数）
+
+## 数据集许可
+
+ISPD 2005 基准由 ICCAD 2005 / ISPD 2006 布局竞赛委员会创建，
+可免费用于学术用途。
+
+## 压缩 JSON 格式
+
+参考文件为 gzip 压缩的 JSON，并采用紧凑的网表表示。
+每个网表存储为整数单元索引列表，而非完整的引脚字典：
+
+```json
+{"netlist": [[0, 1, 2], [3, 4], ...]}
+```
+
+紧凑网表相比详细格式可减少约 65% 的 JSON 大小，
+gzip 进一步将磁盘占用减少约 84%（adaptec1：22.6MB -> 3.7MB，adaptec3：48.2MB -> 7.9MB）。
+scripts/init.py 和 verification/evaluator.py 中的 _decompress_netlist() 函数
+在加载时重建完整的引脚字典。该转换相对于 HPWL 计算是无损的。
+scripts/preprocess.py 脚本直接从原始 Bookshelf 数据生成此压缩格式。