Module gpu 

GPU acceleration for graph algorithms

Based on research from:

• Gunrock (Wang et al., ACM ToPC 2017) - GPU graph traversal primitives
• cuGraph (Bader et al., 2022) - GPU-accelerated graph analytics
• GraphBLAST (Yang et al., 2022) - GPU linear algebra for graphs
• Umbra (Neumann & Freitag, CIDR 2020) - Morsel-driven parallelism
• Funke et al. (SIGMOD 2018) - GPU memory paging

Architecture

• device: GPU device initialization and management
• buffer: GPU buffer management for CSR data
• memory: VRAM detection and memory limits
• cache: LRU cache for graph tiles
• paging: Graph tiling and paging coordinator
• kernels: WGSL compute shaders for BFS, PageRank, etc.

Feature Flag

This module is only available with the gpu feature flag:

cargo build --features gpu

Structs

GpuBfsResult

GPU BFS result

GpuCsrBuffers

GPU buffers for CSR graph representation

GpuDevice

GPU device wrapper for graph operations

GpuMemoryLimits

GPU memory limits and statistics

GpuPageRankResult

GPU PageRank result

GraphTile

Graph tile (subset of nodes and their edges)

LruTileCache

LRU cache for GPU buffers

PagingCoordinator

Paging coordinator for managing graph tiles

Enums

GpuDeviceError

GPU device initialization errors

Constants

DEFAULT_MORSEL_SIZE

Default morsel size (128MB chunks as per DuckDB/Umbra)

Functions

gpu_bfs

Run GPU BFS from source node

gpu_bfs_paged

Run GPU breadth-first search over a graph that may exceed VRAM.

gpu_pagerank

Run GPU PageRank algorithm