layout: true class: typo, typo-selection --- count: false class: nord-dark, middle, center # 🧱 Arena-Based Tree & Elongation in DME ### Arena storage, tapping points & elongation @luk036 👨‍💻 · 2026 📅 --- ### 📋 Agenda .pull-left[ **Part 1: Arena Tree Design** 🏗️ - From Pointers to Indices - The Tree Arena Class **Part 2: Delay Models** ⏱️ - Pure Function Approach - Linear & Elmore Delay **Part 3: Elongation** 📐 - Mechanics & Visualization - The `raw_extend_left` Secret **Part 4: Testing** ✅ - Unit Test Values - Cross-Language Consistency ] .pull-right[ **Part 5: SVG Visualization** 🎨 - Clock Tree Visualizer **Part 6: C++ → Rust Porting** 🔄 - Architecture Comparison - Rust Implementation Highlights **Part 7: Results & Summary** 📈 - What Changed - Key Takeaways - Test Results ] --- ## 🎯 Motivation **Problem**: The DME algorithm used `std::shared_ptr<TreeNode>` for tree connectivity. .pull-left[ ### ❌ Shared Pointer Pain - 🔗 **Reference counting overhead** — atomic increments on every traversal - 💥 **Cache-unfriendly** — nodes scattered across the heap - 🚫 **No simultaneous mutation** — can't get `&mut left` and `&mut right` at once - 🧵 **No arena reuse** — allocation per node, no memory locality ] .pull-right[ ### ✅ Arena-Based Tree - 🗂️ **Contiguous `Vec<TreeNode>`** — cache-friendly linear storage - 🔢 **`NodeIdx` indices** (`usize`/`size_t`) — no pointers - ✌️ **`get_pair_mut(a, b)`** — safe dual-mutable access - ⚡ **Zero allocation overhead** — single `push_back` per node ] --- class: nord-light, middle, center ## Part 1: Arena Tree Design 🏗️ --- ### From Pointers to Indices 🔄 **Before** (`shared_ptr`): ```cpp class TreeNode { std::shared_ptr<TreeNode> left; std::shared_ptr<TreeNode> right; std::shared_ptr<TreeNode> parent; // ... }; ``` **After** (arena indices): ```cpp using NodeIdx = std::size_t; // SIZE_MAX = null class TreeNode { NodeIdx left = SIZE_MAX; NodeIdx right = SIZE_MAX; NodeIdx parent = SIZE_MAX; // ... }; ``` --- ### The `Tree` Arena Class 🌳 ```cpp class Tree { std::vector<TreeNode> nodes; public: NodeIdx root = SIZE_MAX; NodeIdx add(TreeNode node); // push into arena const TreeNode& get(NodeIdx idx) const; // read TreeNode& get_mut(NodeIdx idx); // write std::pair<TreeNode&, TreeNode&> get_pair_mut(NodeIdx a, NodeIdx b); // dual mutation }; ``` .pull-left[ ### 🦀 Rust version ```rust pub struct Tree { nodes: Vec<TreeNode>, pub root: Option<NodeIdx>, } ``` ] .pull-right[ ### ⚡ Key advantage `get_pair_mut(a, b)` enables **safe simultaneous mutation** of two distinct children — impossible with `shared_ptr` in C++ and requires `split_at_mut` in Rust. ] --- class: nord-light, middle, center ## Part 2: Delay Models ⏱️ --- ### Pure Function Approach ✨ **Before**: `calculate_tapping_point` took `TreeNode&` and mutated them as side effects. ```cpp // 😬 OLD: Side-effect heavy virtual std::pair<int, double> calculate_tapping_point(TreeNode& left, TreeNode& right, int distance) = 0; // Inside: mutates left.wire_length, right.wire_length, // left.need_elongation, right.need_elongation ``` **After**: Pure function returning `TappingResult`. ```cpp struct TappingResult { int extend_left; // clamped to [0, distance] int raw_extend_left; // pre-clamp value (for elongation) double delay_left; // delay at tapping point }; virtual TappingResult calculate_tapping_point(int distance, double left_delay, double right_delay, double left_cap, double right_cap) const = 0; ``` --- ### Linear Delay Model 📏 $$ {\color{green} \text{extend\_left}} = \text{round}\left(\frac{({\color{orange} R} - {\color{blue} L}) / {\color{darkmagenta} \alpha} + {\color{coral} d}}{2}\right) $$ $$ \text{TappingResult} = \begin{cases} (0, {\color{blue} L}) & \text{if raw} < 0 \quad \text{(clamped)} \\[2pt] ({\color{coral} d}, {\color{orange} R}) & \text{if raw} > {\color{coral} d} \quad \text{(clamped)} \\[2pt] (\text{raw}, {\color{blue} L} + \text{raw} \cdot {\color{darkmagenta} \alpha}) & \text{otherwise} \end{cases} $$ where ${\color{darkmagenta} \alpha} = \text{delay\_per\_unit}$, ${\color{blue} L} = \text{left\_delay}$, ${\color{orange} R} = \text{right\_delay}$, ${\color{coral} d} = \text{distance}$. The **raw** value is preserved so the caller can implement elongation. --- ### Elmore Delay Model ⚡ $$ z = \frac{({\color{orange} R} - {\color{blue} L}) + {\color{green} r} \cdot ({\color{red} C_r} + {\color{coral} c_w}/2)}{{\color{green} r} \cdot ({\color{coral} c_w} + {\color{red} C_r} + {\color{darkmagenta} C_l})} $$ $$ \text{raw} = \text{round}(z \cdot {\color{coral} d}), \qquad \text{delay\_left} = {\color{blue} L} + {\color{green} r_l} \cdot \left(\frac{{\color{coral} c_l}}{2} + {\color{darkmagenta} C_l}\right) $$ with ${\color{green} r} = {\color{coral} d} \cdot R_u$, ${\color{coral} c_w} = {\color{coral} d} \cdot C_u$, ${\color{green} r_l} = \text{raw} \cdot R_u$, ${\color{coral} c_l} = \text{raw} \cdot C_u$. Same clamping rule: raw < 0 → 0, raw > d → d. --- class: nord-light, middle, center ## Part 3: Elongation — What and Why 📐 --- ### What Is Elongation? 🤔 When two subtrees have **imbalanced delays**, the ideal merge point falls "beyond" the faster child. .mermaid[ <pre> graph TD subgraph Normal["Normal Merge ✅"] L1["Left (delay=1)"] --- P1["Parent at midpoint"] --- R1["Right (delay=1)"] end subgraph Elongation["Elongation Needed ⚠️"] L2["Left (delay=10)\n🟥 Slower"] --- P2["Parent at Left\n(extend=0)"] -.-> R2["Right (delay=1)\n🟢 Faster"] R2 -.->|"elongated\nwire=14 (> 10)"| P2 end style L1 fill:#c8e6c9,stroke:#2e7d32,stroke-width:3px style P1 fill:#c8e6c9,stroke:#2e7d32,color:#2e3440,stroke-width:3px style R1 fill:#c8e6c9,stroke:#2e7d32,stroke-width:3px style L2 fill:#ffcdd2,stroke:#c62828,stroke-width:3px style P2 fill:#ffcdd2,stroke:#c62828,color:#2e3440,stroke-width:3px style R2 fill:#c8e6c9,stroke:#2e7d32,stroke-width:3px </pre> ] The merge point **snaps** to the slower child, and the **faster** child's wire is made longer than the segment distance to add delay. --- ### Elongation Mechanics ⚙️ **When `raw_extend_left < 0`**: - Merge point snaps to left child: `extend_left = 0` - Left wire = 0 (merge at left's position) - Right wire = `distance - raw` = $10 - (-4) = 14$ (🐍 snakes past distance) - `right.need_elongation = true` **When `raw_extend_left > distance`**: - Merge point snaps to right child: `extend_left = distance` - Right wire = 0 (merge at right's position) - Left wire = `raw` = $14$ (🐍 snakes past distance) - `left.need_elongation = true` --- ### Elongation Visualization 📊 | Scenario | extend_left | raw | left_wl | right_wl | need_el | |---|---|---|---|---|---| | Balanced | 5 | 5 | 5 | 5 | none | | Right elongated | 0 | **-4** | 0 | **14** (>10) | right✅ | | Left elongated | 10 | **14** | **14** (>10) | 0 | left✅ | **Key insight**: The elongated branch gets `wire_length > distance`. The wire doesn't go straight — it takes a longer Manhattan path 🐍 to add delay on the faster side, balancing the skew. --- ### The `raw_extend_left` Secret 🗝️ The `TappingResult` preserves the **raw** pre-clamp value. The caller uses it: ```cpp auto tp = calc.calculate_tapping_point(distance, ...); // Caller-side elongation logic: l_node.wire_length = tp.extend_left; r_node.wire_length = distance - tp.raw_extend_left; // ← raw! if (tp.raw_extend_left < 0) { l_node.wire_length = 0; // snap to left r_node.wire_length = distance - tp.raw_extend_left; // > distance! r_node.need_elongation = true; } else if (tp.raw_extend_left > distance) { r_node.wire_length = 0; // snap to right l_node.wire_length = tp.raw_extend_left; // > distance! l_node.need_elongation = true; } ``` --- class: nord-light, middle, center ## Part 4: Testing ✅ --- ### Unit Test Values Match Exactly ✅ **Linear model, right elongation** (`left.delay=10, right.delay=1, d=10`): ``` raw = round((-9/0.5 + 10)/2) = round(-4) = -4 ↓ clamped extend_left = 0, delay_left = 10.0 ↓ caller-side logic left_wl = 0, right_wl = 14, right.need_elongation = true ``` **C++ test** (original, restored): ```cpp CHECK_EQ(extend_left, 0); CHECK(delay_left == doctest::Approx(10.0)); CHECK_EQ(left.wire_length, 0); CHECK_EQ(right.wire_length, 14); // > distance! CHECK_EQ(left.need_elongation, false); CHECK_EQ(right.need_elongation, true); ``` --- ### Cross-Language Consistency 🌐 Same struct, same semantics, **same numerical values in both** ✅ | Language | Tests | Assertions | Status | |---|---|---|---| | C++ | 149 | 748 | ✅ All pass | | Rust | 346 | (281 unit + 65 doc) | ✅ All pass | .pull-left[ **C++ `TappingResult`** (physdes-cpp): ```cpp struct TappingResult { int extend_left; int raw_extend_left; double delay_left; }; ``` ] .pull-right[ **Rust `TappingResult`** (physdes-rs): ```rust pub struct TappingResult { pub extend_left: i32, pub raw_extend_left: i32, pub delay_left: f64, } ``` ] --- class: nord-light, middle, center ## Part 5: SVG Visualization 🎨 --- ### Clock Tree Visualizer 🖼️ The `ClockTreeVisualizer` generates SVG output with: - 🔴 **Root** (red, larger circle) — clock source - 🟢 **Sinks** (green) — leaf nodes with `d:delay` and `c:capacitance` labels - 🔵 **Internal** (blue, smaller) — merge points - 📏 Wire lengths annotated on edges - 📊 Analysis panel: max/min delay, skew, wirelength Generated SVGs: | File | Sinks | Model | Skew | WL | |---|---|---|---|---| | `dme_circle_linear.svg` | 8 ring | Linear | 1.0 | 358 | | `dme_random_linear.svg` | 12 random | Linear | 1.0 | 692 | | `dme_comparison.svg` | L-shape | **Linear vs Elmore** | 0.0 / 0.6 | 210 | --- class: nord-light, middle, center ## Part 6: C++ → Rust Porting 🔄 --- ### Architecture Comparison 🏗️ .mermaid[ <pre> graph LR subgraph cpp["physdes-cpp (C++20)"] CPP_TN["TreeNode\nleft/right: size_t"] CPP_T["Tree\nVec<TreeNode>"] CPP_TR["TappingResult"] CPP_EL["elongation: _compute_merging_segment"] end subgraph rs["physdes-rs (Rust)"] RS_TN["TreeNode\nleft/right: Option<usize>"] RS_T["Tree\nVec<TreeNode>"] RS_TR["TappingResult"] RS_EL["elongation: compute_merging_segment"] end cpp -.->|"same design"| rs style CPP_TN fill:#e3f2fd,stroke:#1565c0,stroke-width:3px style CPP_T fill:#e3f2fd,stroke:#1565c0,color:#2e3440,stroke-width:3px style CPP_TR fill:#e1f5fe,stroke:#01579b,color:#2e3440,stroke-width:3px style CPP_EL fill:#e3f2fd,stroke:#1565c0,stroke-width:3px style RS_TN fill:#fff3e0,stroke:#e65100,stroke-width:3px style RS_T fill:#fff3e0,stroke:#e65100,color:#2e3440,stroke-width:3px style RS_TR fill:#fff3e0,stroke:#e65100,color:#2e3440,stroke-width:3px style RS_EL fill:#fff3e0,stroke:#e65100,stroke-width:3px </pre> ] **Identical architecture**: - `NodeIdx` / `usize` indices - `TappingResult` struct with `raw_extend_left` - Pure `calculate_tapping_point` + caller-side elongation - `get_pair_mut` for safe dual mutation --- ### Rust Implementation Highlights 🦀 ```rust pub struct TappingResult { pub extend_left: i32, pub raw_extend_left: i32, pub delay_left: f64, } ``` ```rust // compute_merging_segment with elongation: let tp = self.delay_calculator.calculate_tapping_point( distance, left_delay, right_delay, left_cap, right_cap, ); let (l_node, r_node) = self.tree.get_pair_mut(left, right); l_node.wire_length = tp.extend_left; r_node.wire_length = distance - tp.raw_extend_left; if tp.raw_extend_left < 0 { l_node.wire_length = 0; r_node.wire_length = distance - tp.raw_extend_left; r_node.need_elongation = true; log::warn!("Right node needs elongation..."); } ``` --- class: nord-light, middle, center ## Part 7: Results & Summary 📈 --- ### What Changed 📋 | File | Change | |---|---| | `include/recti/dme_algorithm.hpp` | `Tree` class, `TappingResult`, pure `DelayCalculator` | | `source/dme_algorithm.cpp` | Arena-based operations, elongation with raw values | | `include/recti/visualize_clock_tree.hpp` | `Tree&`+`NodeIdx` API | | `source/visualize_clock_tree.cpp` | Nested `<g>` depth tracking fix | | `test/source/test_dme_algorithm.cpp` | Pure calc tests + elongation integration | | `src/dme_algorithm.rs` (Rust) | `TappingResult`, same elongation logic | | `dme_*.svg` (8 files) | Generated clock tree visualizations | --- ### Key Takeaways 🎯 - 🏗️ **Arena-based trees** replace `shared_ptr` with indices — zero overhead, cache-friendly - ✨ **Pure `DelayCalculator`** — no side effects, testable, composable - 🔌 **Elongation = wire beyond segment distance** — the faster child's wire snakes 🐍 to add delay and balance skew - 🗝️ **`raw_extend_left`** preserves the pre-clamp value, enabling caller-side elongation - 🌐 **C++ ↔ Rust parity** — same `TappingResult`, same numerical values, same tests - 🎨 **SVG visualizer** — depth-tracked extraction fixes broken comparison XML --- ### Test Results 🏆 ``` C++: 149 test cases | 748 assertions | 0 failed ✅ Rust: 281 unit + 65 doc-tests = 346 | 0 failed ✅ ``` **Elongation numerical values verified in both:** | Test | extend_left | raw | left_wl | right_wl | left_el | right_el | |---|---|---|---|---|---|---| | Linear right | 0 | -4 | 0 | **14** | no | **yes** | | Linear left | 10 | 14 | **14** | 0 | **yes** | no | | Elmore right | 0 | -18 | 0 | **28** | no | **yes** | | Elmore left | 10 | 28 | **28** | 0 | **yes** | no | --- count: false class: nord-dark, middle, center # 🙋 Q&A ### Questions? Discussion? 💬 --- count: false class: nord-dark, middle, center # 👏 Thank You ### 🧱 Arena-Based Tree & Elongation in DME https://github.com/luk036/physdes-cpp https://github.com/luk036/physdes-rs Slides built with Remark.js 🎉 | KaTeX 📐 | Mermaid 🧩 | Nord Theme 🌙