Skip to main content

Module gpu

Module gpu 

Source
Expand description

GPU integration for calculus modality.

Provides cross-platform GPU compute shader execution for numerical integration and differential equation solving via portable wgpu (Vulkan / DX12 / Metal / WebGPU). The engine is vendor-neutral: there is no CUDA/cuFile path. The former NVIDIA GPUDirect-Storage bridge was removed; heterogeneous dispatch, mixed-precision policy, and the vendor-neutral storage path (mmap + staging upload) live in super::hetero_dispatch.

§Architecture

  • WebGPU (wgpu): portable compute on every backend; the single GPU path.
  • Storage: NVMe → OS page cache (mmap, super::host) → GPU staging upload. (NVIDIA GDS’s true NVMe→VRAM DMA was deliberately declined — vendor-lock off the affordability critical path; see hetero_dispatch.rs.)
  • State Tracking: GPU results packed into Quin metadata field.

Structs§

PlatformGpuIntegrator
The platform GPU integrator — uniformly the portable wgpu integrator on every backend. (The former CUDA/cuFile CudaIntegrator alias was removed for vendor neutrality; heterogeneous routing/fallback now lives in super::hetero_dispatch.)
WebGpuIntegrator

Enums§

GpuError

Traits§

GpuIntegrator
Platform-agnostic GPU integration interface.

Functions§

create_gpu_job_quin
Creates a suspended Quin for GPU computation.
extract_gpu_result_from_quin
Extracts GPU computation result from Quin metadata field.
pack_gpu_result_into_quin
Packs GPU computation result into Quin metadata field.