Library structure

eagle is organised as a single header-only library under the eagle/ directory. The public entry point is eagle/eagle.h, which pulls in all modules.

For most users, including the top-level header is sufficient:

#include <eagle/eagle.h>

Individual module headers may be included independently when compile-time isolation is required.

Modules

Graph

CUDA-graph capture and launch: StreamCapturer, Graph, Launcher, plus RAII Stream / Event / host-callback wrappers.

→ API reference API reference — cuda

Module: cuda

Filtering

Prefix-scan, scatter, and array slicing for stream compaction.

→ API reference API reference — filtering

Module: filtering

Reduce

GPU (CUDA) and CPU (OpenMP) parallel reductions over aether scalar arrays.

→ API reference API reference — reduce

Module: reduce

Host

OpenMP + SIMD packet-batched host dispatch with masked / flag-aware launch variants for the pure-C++ path.

→ API reference API reference — host

Module: cpu

Utilities

Global typedefs, logging, error handling, observer-pattern primitives, array slicing, and OpenMP thread management.

→ API reference API reference — util

Module: util

Directory layout

eagle/
├── eagle.h                   ← single top-level include (umbrella)
├── typedefs.h                ← Real, idx_t, uint_t, Vec3R, Vec6R, VecNT, …
├── cuda.h                    ← shortcut: cuda (graph) module
├── filtering.h               ← shortcut: filtering module
├── reduce.h                  ← shortcut: reduce module
├── ScratchArena.h            ← device-neutral graph scratch arena
│
├── cuda/
│   ├── Graph.h               ← CUDA graph builder + dependency manager
│   ├── Launcher.h            ← instantiated graph executor
│   ├── StreamCapturer.h      ← records launches into a CUDA graph
│   ├── CapturedGraph.h       ← owning captured-graph wrapper
│   ├── ComputeBlocks.h       ← grid / block sizing helper
│   ├── NativeNode.h          ← native graph-node concept (addNative)
│   ├── Reduction.h           ← cuda::Reduction — GPU reduction engine
│   ├── Scan.h                ← cuda::Scan — GPU prefix-scan engine
│   ├── traits.h              ← graph type traits
│   ├── Stream.h              ← shortcut: stream primitives
│   ├── detail.h              ← graph implementation detail
│   ├── detail/Holder.h       ← graph node holder
│   └── stream/
│       ├── Stream.h          ← RAII cudaStream_t wrapper
│       ├── Event.h           ← RAII cudaEvent_t wrapper
│       └── HostCallback.h    ← host-callback graph node
│
├── cpu/
│   ├── Graph.h               ← cpu::Graph — dependency-ordered host executor
│   ├── Host.h                ← cpu::Host packet-batched dispatcher
│   ├── Reduction.h           ← cpu::Reduction — OpenMP reduction engine
│   └── Scan.h                ← cpu::Scan — OpenMP prefix-scan engine
│
├── filtering/
│   ├── Scanner.h             ← prefix-scan container (Scanner / RefScanner)
│   ├── Slice.h               ← FilteringSlice compaction helper
│   ├── ScanNode.h            ← native scan graph node
│   ├── FilteringSliceNode.h  ← native compaction graph node
│   └── scan_detail.h         ← low-level scan / scatter kernels
│
├── reduce/
│   ├── ReductionNode.h       ← native reduction graph node
│   └── detail.h              ← warp / block reduction kernels
│
└── util/
    ├── log.h                 ← EAGLE_TRACE/DEBUG/INFO/WARN/ERROR
    ├── throw.h               ← EAGLE_THROW, EAGLE_ASSERT
    ├── DeviceError.h         ← CUDA error tracking + EagleErrorTypes
    ├── Observable.h          ← CRTP observer-pattern base
    ├── Observer.h            ← observer interface
    ├── ObservableArray.h     ← observable array wrapper
    ├── Slice.h               ← general-purpose array slicing
    └── Threads.h             ← OpenMP thread-count management