2D Dynamic Texture Atlas Caching Algorithm

While algorithms like Skyline handle 2D packing (space management) and S3-FIFO handle cache eviction (time management), few solutions address both simultaneously. Biped fills this gap by managing tile placement, recycling, and eviction based on "hot/cold" access patterns within a 2D Atlas.

Inspired by the Buddy algorithm, Biped supports node splitting and merging for fast texture cache space management.

Features:

• Power-of-2 sized Atlas textures
• Size Class Level system for tile classification
• Efficient allocation and recycling via MAP/BASE/LEVEL structures

Limitations:

• Relatively lower space utilization compared to some packing algorithms

Documentation: Algorithm Details (English) | 算法说明 (中文)

Header-only library. Include biped.h in your project.

In one compilation unit, define BIPED_C_IMPLEMENTATION before including the header:

#define BIPED_C_IMPLEMENTATION
#include "biped.h"

Override these macros before including biped.h to customize memory allocation and hashing:

Example:

#define biped_malloc my_malloc
#define biped_free my_free
#define biped_realloc my_realloc
#define biped_hash my_hash_function

#define BIPED_C_IMPLEMENTATION
#include "biped.h"

• Uses fixed power-of-2 sized textures
• Target scenario: 2048×2048 RGBA Atlas storing text glyphs ranging from 8×16 to 64×64, with thousands of nodes
• RGB channels store MSDF or subpixel data; A channel stores grayscale
• Two Biped caches can jointly manage a single RGBA Atlas

Test environment: MSVC, small object allocation with free(malloc(211)). Below are typical path timings and relative speeds.

• Overall performance is comparable to malloc/free operations
• No external fragmentation
• Internal fragmentation depends on data (no fragmentation when using only 32/64 MSDF)
• Cache Hit: Only queries MAP and marks access — minimal overhead
• Miss (A): Evicts and reuses same-level cold node — slightly slower than hit
• Miss (B): Finds coldest node across levels and splits — involves LEVEL scanning and splitting
• Miss (Fail): Attempts A/B/C paths before failing — significantly higher latency

目前已有天际线(Skyline)等 2D pack 类算法负责空间管理, 也有 S3 FIFO 等负责时间管理(如缓存淘汰), 但少有同时兼顾空间与时间的方案. 这里即针对这一缺口:在 2D Atlas 上既做图块摆放与回收, 又按"冷热"做淘汰与复用.

本算法受 Buddy(伙伴)算法启发, 支持节点的分裂与合并, 用于快速管理纹理缓存空间.

特点: 基于 2 幂次尺寸的 Atlas 纹理, 通过 Size Class Level 将图块分级, 结合 MAP/BASE/LEVEL 等结构实现分配与回收.

局限: 空间利用率相对不高.

文档: 算法说明 (中文) | Algorithm Details (English)

Header-only 风格库. 在项目中包含 biped.h 即可.

在一个编译单元中, 在包含头文件前定义 BIPED_C_IMPLEMENTATION:

#define BIPED_C_IMPLEMENTATION
#include "biped.h"

在包含 biped.h 之前定义以下宏以自定义内存分配和哈希函数:

示例:

#define biped_malloc my_malloc
#define biped_free my_free
#define biped_realloc my_realloc
#define biped_hash my_hash_function

#define BIPED_C_IMPLEMENTATION
#include "biped.h"

• 使用固定 2 幂次大小的纹理.
• 预设场景: 在 2048x2048, RGBA 的 Atlas 上, 存放 8x16 至 64x64 的文本, 节点数量为数千量级.
• RGB 通道存 Msdf 或次像素信息, A 通道存灰度信息; 因此设计上可用两个 Biped 缓存共同管理一个 RGBA Atlas.

测试环境: MSVC, free(malloc(211)) 小对象申请操作. 以下为典型路径的耗时与相对速度.

• 总体来看: 可以视为malloc/free相同代价的操作.
• 但是不存在外部碎片
• 内部碎片则视数据而定(只用32/64 Msdf将不会有碎片)
• Cache Hit: 仅查 MAP 并 MARK, 开销最小.
• Miss (A): 同 Level 冷节点驱逐与复用, 略慢于 Hit.
• Miss (B): 跨 Level 取最冷节点再分裂, 涉及 LEVEL 扫描与分裂.
• Miss (失败): 需尝试 A/B/C 均失败后的路径, 耗时显著升高.