src/uni20/async

This directory contains Uni20's coroutine-based async runtime. The runtime owns task lifetime, dependency ordering, and buffer access discipline; library code should use these abstractions instead of raw threads or ad hoc synchronization.

Contents

• async.hpp, async_impl.hpp, async_node.hpp, async_ops.hpp, async_traits.hpp: async value wrappers, payload capabilities, and dependency graph plumbing.
• async_errors.hpp: async cancellation and deadlock exception types.
• async_task.hpp, async_task_impl.hpp, async_task_promise.hpp, awaiters.hpp: implemented coroutine task wrappers and await support.
• cuda_task.hpp: runtime-neutral CUDA task-domain type and typed scheduler admission interface. CudaTaskPromise stores optional device affinity while sharing the common promise state and internal task representation with AsyncTask, so this header remains usable when CUDA is not configured. CUDA runtime awaiters live under backend/cuda/ and opt into the concrete CUDA-promise pass-through with CudaTaskAwaiterTag.
• debug_cuda_scheduler.hpp: deterministic scheduler with one shared queue for ordinary tasks and CUDA tasks on any validated device. It establishes and restores the CUDA device around each CUDA activation.
• debug_scheduler.hpp: calling-thread scheduler with configurable FIFO, reverse, or seeded pseudo-random runnable-batch ordering. Reverse order is the default.
• tbb_cuda_scheduler.hpp: unified oneTBB scheduler with one host arena and one worker-only arena per enrolled CUDA device. Device-arena observers establish and restore the required device around every CUDA activation.
• tbb_task_submission.hpp: internal task-group registration and non-blocking arena-admission helper shared by host and CUDA TBB schedulers.
• scheduler.hpp, debug_scheduler.hpp, tbb_scheduler.hpp, tbb_numa_scheduler.hpp: scheduler interfaces and implementations.
• epoch_context.hpp, epoch_queue.hpp: epoch ordering and causal execution support.
• buffers.hpp, storage_buffer.hpp, shared_storage.hpp: read and exclusive-mutable capabilities, their await paths, and storage lifetime helpers.
• var.hpp, future_value.hpp, reverse_value.hpp: async dataflow values.
• task_registry_*: optional debug task registry support.
• async_toys.hpp, var_toys.hpp: experiments and examples; do not treat these as stable API.

Notes

• Coroutine lambdas that return Uni20 async tasks must be captureless and use the C++23 static lambda modifier. See /AGENTS.md for the rule and rationale.
• Shared data should flow through the buffer/value abstractions in this module so scheduler ordering remains explicit.
• is_async_alias_v<T> controls independent-value versus shared-alias identity. Assignment to an alias is available only through an ADL-visible assign_through operation. WriteBuffer<T> remains the single exclusive epoch capability, with proxy operations constrained by the payload.

Related Documentation

• Source tree map
• Async documentation index
• Async storage and identity
• Exceptions and cancellation
• Scheduler routing and task domains
• Task registry diagnostics