Module gpu_backend 

Unified GPU backend trait for tile-based FDTD simulation.

This module provides an abstraction layer over CUDA and WGPU backends, enabling GPU-resident state with minimal host transfers.

Design

The key optimization is keeping simulation state on the GPU:

• Pressure buffers persist on GPU across steps (no per-step upload/download)
• Only halo data (64 bytes per tile edge) transfers for K2K messaging
• Full grid readback only when GUI needs to render

Buffer Layout

Each tile uses an 18x18 buffer (324 floats = 1,296 bytes):

+---+----------------+---+
| NW|   North Halo   |NE |  <- Row 0 (from neighbor)
+---+----------------+---+
|   |                |   |
| W |   16x16 Tile   | E |  <- Rows 1-16 (owned)
|   |    Interior    |   |
+---+----------------+---+
| SW|   South Halo   |SE |  <- Row 17 (from neighbor)
+---+----------------+---+

Structs

FdtdParams

Parameters for FDTD computation.

TileGpuBuffers

GPU buffers for a single tile.

Enums

BoundaryCondition

Boundary condition type for domain edges.

Edge

Edge direction for halo exchange.

Traits

TileGpuBackend

Unified trait for GPU backends (CUDA and WGPU).

Functions

buffer_index

Calculate buffer index for interior coordinates (0-indexed).

halo_col_index

Calculate halo column index.

halo_row_start

Calculate halo row index (row 0 or row buffer_width-1).