CUDA Backend Documentation

Uni20 has a tested low-level CUDA runtime foundation on the current main branch: explicit scoped process-wide initialization, canonical per-device resources, structured errors, device guards, reference-counted stream-pool leases, immutable completion tokens, actually-idle stream pools, and typed move-only device buffers with scoped ReadAccess<T>/WriteAccess<T> objects. Deterministic and oneTBB unified host/multi-device task schedulers are implemented, as are non-blocking resource awaiters, generic provider-resource leases, CUDA Tensor storage, and non-blocking async Tensor-to-cuBLAS matrix-product lowering. General CUDA Tensor kernel coverage, direct non-async CUDA Tensor operations, and storage-driven scheduler selection remain incomplete.

Start Here

• Getting Started shows CUDA configuration, process-wide initialization, and ordinary CUDA Tensor construction.
• CUDA Buffers introduces the implemented low-level allocation and stream-synchronized access API for kernel and provider authors.
• Runtime Model explains device selection, stream ownership, completion tokens, and structured errors.
• Deployment Environment explains the multi-GPU, MPI, accelerator-generation, and host-architecture constraints behind the CUDA runtime design.
• CUDA hello-world example reports the configured runtime and exercises the stream/completion foundation.

Detailed Reference

• CUDA Buffer Completion Lowering defines the exact access, completion-ledger, lifetime, and failure contract.
• Memory Allocation records the current allocator path and future allocation-policy work.

Active Design

• CUDA Kernel Dispatch and Device Scheduling
• Runtime Resolution
• cuSOLVER Architecture

Background

• Backend Library Compatibility records provider/version constraints.
• GPU Landscape surveys tensor-network GPU libraries and implementation choices.

Source Navigation

• CUDA runtime foundation
• cuBLAS provider infrastructure
• Provider-ready cuBLAS operations
• cuBLAS operation-tag backend
• CUDA task schedulers
• cuSOLVER provider scaffolding
• cuSOLVER operation-tag backend