[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-77829":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":13,"subscribersCount":13,"size":13,"stars1d":16,"stars7d":17,"stars30d":18,"stars90d":13,"forks30d":13,"starsTrendScore":19,"compositeScore":20,"rankGlobal":10,"rankLanguage":10,"license":10,"archived":21,"fork":21,"defaultBranch":22,"hasWiki":23,"hasPages":21,"topics":24,"createdAt":10,"pushedAt":10,"updatedAt":36,"readmeContent":37,"aiSummary":38,"trendingCount":13,"starSnapshotCount":13,"syncStatus":39,"lastSyncTime":40,"discoverSource":41},77829,"DEEPFOLD-SOLVER","a9876543245\u002FDEEPFOLD-SOLVER","a9876543245","GPU-accelerated GTO poker solver · CPU fallback · 2550+ preflop charts · Trilingual UI · Windows desktop","https:\u002F\u002Fdeepfold.co",null,"C++",300,0,108,1,42,86,192,126,87,false,"main",true,[25,26,27,28,29,30,31,32,33,34,35],"cfr","cpp","cuda","desktop-app","game-theory","gto","poker","rust","solver","tauri","texas-holdem","2026-06-12 04:01:22","# DEEPFOLD-SOLVER\n\n> GPU-accelerated GTO poker solver · CPU fallback · Trilingual UI · One-click installer\n\n**[English](README.md) · [中文](README.zh.md) · [日本語](README.ja.md)**\n\n📘 **[User Guide (English)](USER_GUIDE.md)** · **[使用說明 (中文)](USER_GUIDE.zh.md)**\n\n![Platform](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fplatform-Windows%2010%2F11-blue)\n![Backend](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fbackend-CUDA%20%2B%20CPU%20%28AVX2%2BMulti--core%29-orange)\n\nDEEPFOLD-SOLVER is the desktop GTO solver from [DEEPFOLD](https:\u002F\u002Fdeepfold.co). It pairs a GPU-accelerated DCFR engine with a multi-core CPU backend that scales linearly across every available core, surrounded by **runout aggregation, per-combo blocker analysis, EV \u002F aggression heatmaps, and a 2,500+ preflop chart browser** — all in a single Windows installer.\n\n## What sets DEEPFOLD-SOLVER apart\n\n### Engine that uses every core you have\n\n- **Dual-backend DCFR** — GPU when you have one, CPU when you don't, identical numerical strategy either way. The GPU path ships native CUDA SASS for Turing \u002F Ampere \u002F Ada \u002F Hopper, with PTX-JIT forward-compat for Blackwell. The CPU path is a BFS-flat \"levelized\" CFR backend that scales linearly across physical cores instead of capping at two threads.\n- **Runtime CPUID dispatch** — AVX2 kernels on Haswell-and-newer, scalar fallback on older silicon. One binary, no separate build, no startup crashes on pre-2013 CPUs.\n- **Per-commit parity gate** — every build re-verifies that `reference vs levelized`, `scalar vs AVX2`, and `1-thread vs N-thread` paths emit bit-identical strategies. The fast path can never silently drift from the correctness oracle.\n\n### Solve-mode presets that actually stop on time\n\n- **Quick \u002F Standard \u002F Deep** preset pills bundle iteration cap + time budget + exploitability target into one click. The solver stops at whichever fires first — and because CFR is anytime, the running average at any iteration *is* the strategy, so a budget-stopped solve is a usable strategy, not a half-baked one.\n- **Pre-solve ETA banner** — clicking Solve fires a sub-second `--estimate-only` engine call that shows wall-clock time before iterations begin, calibrated against the standard benchmark. Surfaces AUTO fallback reasons (e.g. \"Pascal needs CUDA-12.x build\") right next to the estimate.\n- **Stop button + Quality badge** — pure abort with no partial result preserved (use the time budget for \"stop with what we have\"). Result panel shows 🟢 high \u002F 🟡 good \u002F 🟠 rough \u002F 🔴 low confidence based on final exploitability%.\n- **`--time-budget-seconds`** is checked **every iter**, so even slow per-iter spots stop precisely at budget instead of overshooting.\n\n### Post-solve insight tools other solvers don't have\n\n- **Runout Report** — one click after a solve fans out every canonical turn card into a 13×4 grid colored by dominant action. Switch to **By Class** view and the 23+ turns get grouped into **Pair \u002F Flush \u002F Straight \u002F Overcard \u002F Brick** texture buckets with weighted strategy + EV per bucket. Sort by Best EV \u002F Worst EV \u002F Most aggressive. CSV export.\n- **1326 Combo Drill** — click any 169-class label to expand the 4 \u002F 6 \u002F 12 specific combos in that class with **per-combo blocker analysis**: how much of the opponent's range each specific hand removes, plus the top-5 most-blocked opponent classes. The standard tie-breaker for mixed strategies, finally first-class in the UI.\n- **Strategy grid view modes** — toolbar above the 169 grid switches between Strategy Mix (default multi-action gradient), **EV** (per-class heatmap, red→grey→green normalized to in-range EV span), **Aggression** (Bet\u002FRaise\u002FAll-in frequency cool→hot), and single-action Heatmap. EV mode surfaces \"which combos are profit centres vs which are losing\" at a glance.\n\n### Memory you can trust\n\n- **Memory Profile presets** — `safe \u002F balanced \u002F performance` pick host-RAM, JSON, and strategy-tree-node budgets up front, with a live preview of each. The solver respects the budget end-to-end — pre-backend gates evaluate CPU host \u002F GPU VRAM \u002F AUTO fallback **before** allocation, so OOM scenarios become structured errors with a UI badge instead of crashes.\n- **Common host budget gate** applies on GPU backend too — matchup tables, strategy-tree EV cache, and JSON response all live on host RAM regardless of backend, and all are checked. The diagnostic clarifies that switching to GPU won't fix common-host overflows.\n- **Chunked GPU matchup upload** eliminates host-side `flat_ev` \u002F `flat_valid` duplication; upload happens per-runout via `cudaMemset + cudaMemcpy`, lowering peak host RAM during GPU prep.\n\n### Built-in content\n\n- **2,550+ preflop scenarios** browsable in-app. One click applies as IP \u002F OOP range.\n- **120+ pre-solved flop spots** in a one-click library.\n- **Bet sizing presets** — Standard \u002F Polar \u002F Small Ball — flow through to both the solver tree and the UI buttons.\n- **Range editor + node locking** — override any combo frequency and re-solve.\n- **Training mode** — 10-question drills that score your answers against the equilibrium.\n\n### Operational polish\n\n- **Trilingual UI** — English \u002F 中文 \u002F 日本語, switchable at any time.\n- **Auto-update** — banner-driven one-click installer refresh, signed releases, install-mode `passive`.\n- **Suit isomorphism** delivers 3–7× speedup on monotone \u002F three-of-suit boards automatically; per-runout matchup tables on GPU give 6–10× over CPU on iso-engaged trees.\n- **Route A navigation cache** — O(1) action switching, no re-solve. **Path B runout selector** for PioSolver-style chance-aware navigation.\n- **Reproducible benchmarks** — `deepsolver_core --benchmark standard` runs an AsKd7c rainbow \u002F 100-iter scenario and emits compact perf-tracking JSON (`iterations_per_sec`, `nodes_per_sec`, `memory_estimate_mb`, full timing breakdown). Greppable for CI regression tracking.\n\n## Download\n\n**Windows 10 \u002F 11 (x64)** — [Latest installer](https:\u002F\u002Fgithub.com\u002Fa9876543245\u002FDEEPFOLD-SOLVER\u002Freleases\u002Flatest)\n\nAfter install, the app self-updates: a banner appears in the top-left when a new release is available; one click installs and restarts.\n\n> ⚠️ **First-run Windows warning**: when you launch the installer, Windows\n> will show a \"**Windows protected your PC**\" (SmartScreen) prompt. This is\n> expected — DEEPFOLD-SOLVER doesn't yet ship with an EV code-signing\n> certificate, so Windows doesn't recognize the publisher. Click\n> **More info** → **Run anyway** to continue. Full walkthrough in the\n> [User Guide — SmartScreen warning](USER_GUIDE.md#appendix-first-install-windows-smartscreen-warning).\n\n## System requirements\n\n| | Minimum | Recommended |\n|---|---|---|\n| OS | Windows 10 64-bit | Windows 11 64-bit |\n| CPU | x86-64, dual-core (any year) | 4+ physical cores, AVX2 (Haswell 2013 \u002F Excavator 2015 or newer) |\n| RAM | 4 GB | 8 GB+ |\n| GPU | — *(CPU backend is fully featured)* | NVIDIA RTX 2000 series or newer, 4 GB+ VRAM |\n| Disk | 200 MB | 200 MB |\n\nGPU and SIMD detection are both automatic. The status pill in the top-right shows **CUDA** \u002F **CPU** at a glance, and the CPU backend prints an `AVX2` or `scalar` tag based on what your hardware supports — pre-Haswell CPUs run the scalar kernels and never see an AVX2 opcode.\n\n## Getting started\n\n1. Install and launch the app.\n2. Click **Sign in with Google** — your system browser opens for OAuth.\n3. DEEPFOLD PRO members land straight in the solver.\n\nNot a member yet? Upgrade at [deepfold.co](https:\u002F\u002Fdeepfold.co).\n\n## Feature reference\n\n| Feature | Description |\n|---|---|\n| **GTO Solver** | Discounted CFR with vectorized GPU kernels and a multi-core CPU backend. Sub-percent exploitability in seconds for typical turn spots. |\n| **Per-combo strategy grid** | 13×13 grid colored by action mix at the current decision node. Hover for suited-variant breakdown. |\n| **Acting ↔ Opponent view** | Toggle between your strategy and the opponent's reach-weighted range at the same node. |\n| **Grid view modes** | Toolbar above the 169 grid switches between Strategy Mix \u002F **EV** \u002F **Aggression** \u002F single-action heatmap. EV mode normalizes red→grey→green across in-range cells so profit centres jump out. |\n| **Memory Profile selector** | UI pills for `safe \u002F balanced \u002F performance` in advanced settings with live budget preview. Threads through to the engine via `--memory-profile`. |\n| **Benchmark CLI** | `deepsolver_core --benchmark standard` runs a reproducible AsKd7c+100iter scenario and emits compact perf-tracking JSON. |\n| **Runout Report** | After any solve, fan out all enumerated turn cards into a 13×4 grid + texture-bucket view + 4 sort modes + CSV export. See the [User Guide](USER_GUIDE.md#2-runout-report--see-every-turn-at-once). |\n| **1326 Combo Drill** | Expand any 169-class into its 4\u002F6\u002F12 specific combos with per-combo blocker analysis vs the opponent's range. See the [User Guide](USER_GUIDE.md#3-combo-drill--break-169-classes-into-specific-combos). |\n| **Memory Profile** | `safe \u002F balanced \u002F performance` presets to bound host-RAM, JSON, and strategy-tree-node budgets. No more silent OOM kills. |\n| **Runout picker** | When iso enumeration is engaged, click any canonical river card to switch subtrees. |\n| **GTO chart library** | 2,550+ bundled preflop scenarios browsable in-app. One click applies as IP \u002F OOP range. |\n| **Bet sizing presets** | Standard \u002F Polar \u002F Small Ball — flows through to the solver tree AND the UI buttons. |\n| **Training mode** | 10-question drills that score your answers against the equilibrium. |\n| **Pre-solved spot library** | 120+ common flop spots, one click to load. |\n| **Range editor + node locking** | Override any combo frequency and re-solve. |\n\n## Architecture\n\n```\n┌─────────────────────────────────────────────┐\n│  React + TypeScript UI (Tauri webview)      │\n│  ├── Strategy grid · Runout picker          │\n│  └── GTO chart browser                      │\n├─────────────────────────────────────────────┤\n│  Rust (Tauri) — IPC + chart loader          │\n├─────────────────────────────────────────────┤\n│  C++ engine (deepsolver_core)               │\n│  ├── DCFR (CPU)                             │\n│  ├── CUDA kernels (GPU)                     │\n│  └── Suit isomorphism + per-runout matchup  │\n└─────────────────────────────────────────────┘\n```\n\nThe engine is a standalone CLI (`deepsolver_core.exe`) shipped as a Tauri sidecar. Tauri spawns it per solve and parses the JSON result, including a full strategy tree for client-side navigation.\n\n## Building from source\n\nRequires:\n- **Node.js 20+** + **npm**\n- **Rust 1.78+** (`rustup`)\n- **CMake 3.20+** + **MSVC 2022** (Windows)\n- **CUDA Toolkit 12.x** (optional — CPU build works without)\n\n```sh\ngit clone https:\u002F\u002Fgithub.com\u002Fa9876543245\u002FDEEPFOLD-SOLVER.git\ncd DEEPFOLD-SOLVER\nnpm install\n\n# Build C++ engine\ncd core && mkdir -p build && cd build\ncmake .. -DBUILD_TESTS=ON\ncmake --build . --config Release\nctest\n\n# Run dev app (engine binary picked up from ..\u002Fcore\u002Fbuild\u002FRelease\u002F)\ncd ..\u002F..\nnpm run tauri dev\n```\n\n> **Note**: the bundled GTO preflop chart library (`gto_output\u002F`, ~31MB) and\n> the precompiled engine sidecar (`src-tauri\u002Fbinaries\u002F`) are NOT in this\n> repository — they ship inside the official installer. Builds from source\n> won't have a populated chart browser unless you provide your own\n> `gto_output\u002F` directory at the repo root in the same JSON schema. Sign-in\n> additionally requires `DEEPFOLD_GOOGLE_CLIENT_SECRET` set in the build\n> environment (without it, OAuth will fail at runtime).\n\n## Support\n\n- **Bug reports \u002F feature requests**: [open an issue](https:\u002F\u002Fgithub.com\u002Fa9876543245\u002FDEEPFOLD-SOLVER\u002Fissues)\n- **Membership questions**: [contact@deepfold.co](mailto:contact@deepfold.co)\n\nWhen filing a bug, please include:\n- App version (top-right of window, or **About**)\n- Backend pill state at the time: **CUDA** \u002F **CPU**\n- Windows version (Settings → About)\n- Screenshot or screen recording for UI issues\n\n## FAQ\n\n**Does it work without a GPU?**\nYes. Auto-detects and falls back to CPU. Slower but produces identical strategies.\n\n**macOS \u002F Linux support?**\nNot currently. On the roadmap.\n\n**Are solver strategies uploaded anywhere?**\nNo. Everything runs locally. The only network call is the OAuth sign-in check against deepfold.co.\n\n**How do updates work?**\nOn launch the app checks the latest GitHub release. If signed and newer, a banner offers one-click update.\n\n## License\n\nDEEPFOLD-SOLVER source is published for transparency. Installers are intended for DEEPFOLD PRO members. © DEEPFOLD — All rights reserved.\n\n[deepfold.co](https:\u002F\u002Fdeepfold.co)\n","DEEPFOLD-SOLVER 是一个GPU加速的德州扑克GTO求解器，支持CPU回退，并提供2550多种翻牌前图表和三语用户界面。该项目利用CUDA技术进行GPU加速计算，同时具备多核CPU后端支持，确保在任何可用核心上都能线性扩展性能。此外，它还集成了跑出聚合、每组合阻断分析、EV\u002F激进度热图等功能。适用于需要精确策略分析的德州扑克玩家或研究者，在Windows 10\u002F11系统下运行。该软件特别适合那些希望深入理解游戏理论并优化自己玩法的高级用户。",2,"2026-06-11 03:56:10","CREATED_QUERY"]