This represents the reverse-mode adjoint for functions of complex variables, assuming a real-valued scalar loss function ( L \colon \mathbb{C} \to \mathbb{R} ). More...

Classes
struct	AllAwaiter
	Awaitable that waits for all provided awaiters to complete. More...

class	Async
	Async<T> is a container for coroutine-safe asynchronous read/write. More...

class	async_cancellation
	Base class for cancellation-related async exceptions. More...

struct	async_do_not_start_t
	Tag type to construct an Async without starting the queue object. More...

class	async_error
	Base class for all async subsystem exceptions. More...

class	async_state_error
	Base class for invalid async state transitions or accesses. More...

class	async_storage_missing
	Thrown when Async storage metadata is missing unexpectedly. More...

struct	async_value_type
	Strip `Async<T>` to `T` for assignment and expression type deduction. More...

struct	async_value_type< Async< T > >

struct	async_value_type< ReverseValue< T > >

class	async_value_uninitialized
	Thrown when an Async value is accessed before initialization. More...

class	AsyncTask
	Canonical host-oriented coroutine task type. More...

class	AsyncTaskPromise
	Promise type for ordinary AsyncTask coroutines. More...

class	BasicTask
	Move-only promise-neutral ownership claim for a Uni20 coroutine. More...

class	buffer_cancelled
	Raised when a buffer read path is cancelled intentionally. More...

class	buffer_error
	Base class for buffer-related failures. More...

class	buffer_read_uninitialized
	Raised when a reader attempts to access an uninitialized buffer. More...

class	buffer_uninitialized
	Base class for access to a buffer without a constructed value. More...

class	buffer_write_uninitialized
	Raised when a writer requests a mutable reference before construction. More...

class	CudaDeviceAwaiter
	Awaiter that establishes or changes a CUDA task's device affinity. More...

class	CudaResourceAcquireAwaiter
	CUDA-task awaiter for one provider resource from a fixed-capacity pool. More...

class	CudaStreamAcquireAwaiter
	CUDA-task awaiter for an actually-idle stream-pool lease. More...

class	CudaTask
	CUDA-constrained public task type used for typed initial admission. More...

struct	CudaTaskAwaiterTag
	Marker base for awaiters that require a `CudaTaskPromise` handle. More...

class	CudaTaskPromise
	Promise type for CUDA-constrained CudaTask coroutines. More...

class	DebugCudaScheduler
	Configurable debug scheduler that admits ordinary and CUDA tasks. More...

class	DebugScheduler
	Single-threaded scheduler with configurable runnable-batch ordering. More...

struct	DebugSchedulerOptions
	Configuration for deterministic debug scheduling. More...

class	Defer
	Move-only deferred writer helper for a single `WriteBuffer<T>`. More...

class	EpochContext

class	EpochContextReader
	RAII handle representing the writer of a value in an EpochContext. More...

class	EpochContextWriter
	RAII-scoped representation of a writer’s participation in an epoch. More...

class	EpochQueue
	Coordinates multiple epochs of read/write gates. More...

class	FutureValue
	Write-once async value wrapper that retains a pending write buffer. More...

class	IAsyncScheduler
	Scheduler interface for initial `AsyncTask` submission and host-side progress. More...

class	ICudaScheduler
	Scheduler interface for initial `CudaTask` admission. More...

struct	is_async_alias
	Identifies payload descriptors requiring shared async alias identity. More...

class	IScheduler
	Internal route for resuming an already-bound coroutine task. More...

class	LightweightTaskBatch
	Borrowed type-erased description of a synchronous lightweight task batch. More...

class	NodeInfo
	Records node information for DAG/debugging visualization. More...

class	OwningReadAccessProxy
	Owning proxy returned by `co_await read_buffer.transfer()`. More...

class	OwningReadAwaiter
	Awaiter that transfers a read handle into an owning read proxy. More...

class	OwningStorageAccessProxy
	Owning proxy providing access to writer storage. More...

class	OwningStorageAwaiter
	Awaiter that yields an owning storage-access proxy. More...

class	OwningTakeAwaiter
	Awaiter that takes and releases a value from an owning writer. More...

class	OwningWriteAccessProxy
	Owning proxy returned by `co_await write_buffer.transfer()`. More...

class	OwningWriteAwaiter
	Awaiter that transfers a writer handle into an owning write proxy. More...

class	PropagateExceptionsAwaiter
	Awaiter that registers explicit exception sinks and then resumes immediately. More...

class	ReadBuffer
	RAII handle for reading the value from an Async container at a given epoch. More...

class	ReadMaybeAwaiter
	Awaiter returning an optional owning read proxy. More...

class	ReadMaybeAwaiter< T const & >
	Awaiter returning pointer-or-null read access for lvalue buffers. More...

class	ReadOrCancelAwaiter
	Awaiter returning owning read access or throwing `task_cancelled`. More...

class	ReadOrCancelAwaiter< T const & >
	Awaiter returning borrowed read access or throwing `task_cancelled`. More...

class	ReverseEpochQueue
	Reverse-order epoch queue used by reverse-mode gradient accumulation. More...

class	ReverseValue
	Reverse-mode accumulation endpoint with staged input/output gradient channels. More...

class	ScopedScheduler
	RAII helper that temporarily overrides the global scheduler. More...

class	shared_storage
	A lightweight, thread-safe, reference-counted storage for a single object. More...

class	StorageAwaiter
	Awaiter returning direct access to a writer's shared storage object. More...

class	TakeAwaiter
	Awaiter that takes the writer value without releasing the writer epoch. More...

class	TakeReleaseAwaiter
	Awaiter that takes a value and immediately releases the writer epoch. More...

class	task_cancelled
	Raised when an AsyncTask is cancelled before completion. More...

struct	TaskAwaiter
	Forwarding awaiter that transfers one task ownership claim to an awaitable. More...

class	TaskFactory
	Factory for producing multiple BasicTask ownership claims for one coroutine. More...

struct	TaskFactoryAwaiter
	Forwarding awaiter that transfers shared task ownership through a TaskFactory. More...

class	TaskHandle
	Non-owning erased identity for one Uni20 coroutine frame. More...

class	TaskPromiseBase
	Shared non-polymorphic implementation for Uni20 task promises. More...

struct	TaskRoute
	Execution requirements used to validate a scheduler route. More...

class	TbbCudaScheduler
	Unified oneTBB scheduler with one host arena and one arena per CUDA device. More...

struct	TbbCudaSchedulerOptions
	Arena configuration for the unified host/CUDA TBB scheduler. More...

class	TbbNumaScheduler
	NUMA-aware scheduler that balances work across per-node TBB arenas. More...

class	TbbScheduler

struct	TbbSchedulerBatchOptions
	Runtime policy for synchronous lightweight batches executed by a TBB scheduler. More...

struct	TryAwaiter
	Awaiter wrapper that returns an optional result. More...

struct	ValueAwaiter
	Always-ready awaitable owning an immediate operation value. More...

class	Var

class	WriteAccessProxy
	Non-owning proxy returned by `co_await` on an lvalue `WriteBuffer<T>`. More...

class	WriteAssignProxy
	Assignment-only proxy returned by `WriteBuffer::write()`. More...

class	WriteBuffer
	Exclusive mutable-access buffer for one `Async<T>` epoch. More...

Concepts
concept	read_buffer_awaitable
	concept for a valid type that behaves as a read buffer awaitable This requires that the awaiter has a value_type and await_resume() is available.

concept	write_buffer_awaitable
	concept for a valid type that behaves as a write buffer awaitable This requires that `value_type x; co_await a = x` is valid.

concept	emplace_buffer_awaitable
	concept for a valid type that supports in-place construction via emplace(...)

concept	emplace_buffer_awaitable_of

concept	read_buffer_awaitable_of
	concept for a valid type that behaves as a read buffer awaitable for type `T`.

concept	write_buffer_awaitable_of
	concept for a valid type that behaves as a write buffer awaitable This requires that `T x; co_await a = x` is valid.

concept	read_write_buffer_awaitable
	concept for a valid type that behaves as a write buffer that is also readable This requires that `value_type x = co_await a; co_await a = x` is valid

concept	read_write_buffer_awaitable_of
	concept for a valid type that behaves as a write buffer that is also readable This requires that `T x = co_await a; co_await a = x` is valid

concept	async_reader
	An async_reader is satisfied if `.read()` returns an async read buffer.

concept	async_writer
	An async_writer is satisfied if `.write()` returns a buffer supporting `emplace(...)`.

concept	async_reader_of

concept	async_writer_to
	Concept requiring an async writer that accepts values constructible from `T`.

concept	async_movable_to
	Concept for an Awaitable that yields a movable object that can be moved into T.

concept	async_mutable_writer_of

concept	async_read_writer
	Concept for a type that behaves like an asynchronous writer that is also readable: it has a write() function that returns an awaiter that can be assigned from the nested value_type, and also read from.

concept	async_like
	Concept for a type that behaves like an asynchronous reader and writer (like Async<T>).

concept	async_binary_applicable
	Concept to determine if Op(T,U) is a candidate for forwarding for Async.

concept	async_binary_assignable
	concept for validating a binary operation true if the operation result can be emplaced into the writer output type.

concept	is_async_compound_applicable

concept	TaskPromise
	Promise concept accepted by Uni20 task machinery.

concept	CudaTaskAwaitable
	Identifies an explicitly opted-in CUDA task awaiter.

concept	AwaitSuspendResult
	Concept for the valid return types of await_suspend.

concept	TaskAwaitable
	Concept for awaitables that accept promise-neutral task ownership.

concept	TaskFactoryChildAwaitable
	Awaitable that can hold one ownership claim in a multi-input join.

concept	TaskFactoryAwaitable
	Concept for awaitables that support shared promise-neutral ownership.

concept	HasMemberCoAwait
	Detects if `T` has a member `operator co_await()`.

concept	HasFreeCoAwait
	Detects if a free `operator co_await(t)` exists.

concept	exception_sink_buffer
	Concept for buffers that can register explicit exception sinks.

Typedefs
template<typename T , typename U , typename Op >
using	async_binary_result_t = decltype(std::declval< Op >()(std::declval< async_value_t< T > >(), std::declval< async_value_t< U > >()))
	Result type of applying binary op Op to async_value_t<T>, async_value_t

template<typename T >
using	async_value_t = typename async_value_type< std::remove_cvref_t< T > >::type
	Extract the underlying value type from an immediate value or `Async<T>`.

template<typename T >
using	OwningWriteProxy = OwningWriteAccessProxy< T >
	Alias for the owning write access proxy type.

template<typename AsyncLike >
using	take_awaiter_t = decltype(std::declval< write_buffer_t< AsyncLike > & >().take())
	Alias for the awaiter type returned by `write_buffer_t::take()`.

template<typename AsyncLike >
using	write_buffer_t = decltype(std::declval< AsyncLike & >().write())
	Alias for the write-buffer type produced by `write()` on an lvalue.

template<typename T >
using	WriteAwaitProxy = WriteAccessProxy< T >
	Alias for the non-owning write access proxy type.

template<typename T >
using	WriteProxy = WriteAssignProxy< T >
	Alias for assignment-only writer proxy.

Enumerations
enum class	DebugSchedulerOrder { fifo , reverse , random }
	Runnable-batch ordering policy for `DebugScheduler`. More...

enum class	NodeValueState { Invalid , Unconstructed , Constructed }
	Construction state reported for a debug DAG value node. More...

enum class	TaskDomain { none , host , cuda }
	Execution domain declared by a concrete Uni20 task promise. More...

Functions
template<TaskFactoryChildAwaitable... Aw> requires ((!std::is_void_v<decltype(std::declval<Aw>().await_resume())> && ...))
auto	all (Aw &&... aw) noexcept
	Build an awaitable that waits for all of the provided awaitables.

template<typename T , typename U >
AsyncTask	async_accumulate (ReadBuffer< T > a_, ReadBuffer< U > b_, WriteBuffer< T > out_)
	Compute `out = a + b` while tolerating cancellation of either input.

template<typename T , typename U >
AsyncTask	async_accumulate_minus (ReadBuffer< T > a_, ReadBuffer< U > b_, WriteBuffer< T > out_)
	Compute `out = a - b` while tolerating cancellation of either input.

template<typename T , typename U > requires (detail::current_value_assignment_source<T, U> \|\| (is_async_alias_v<T> && detail::async_alias_assignable_from<T, async_value_t<U> const&>))
void	async_assign (Async< T > &lhs, U &&rhs)
	Assigns an Async or scalar value into an Async destination.

template<typename T , typename U > requires (!is_async_alias_v<T> && read_buffer_awaitable_of<U, T> && detail::write_proxy_assignable_source<T, T>)
void	async_assign (WriteBuffer< T > destination, U &&source)
	Assign an awaitable value into an existing independent-value timeline.

template<typename T , typename U > requires (is_async_alias_v<T> && detail::async_alias_assignment_awaitable<U, T>)
void	async_assign (WriteBuffer< T > destination, U &&source)
	Assign through an existing async alias descriptor.

template<typename T , typename U > requires (!read_buffer_awaitable<std::remove_cvref_t<U>>) && (detail::current_value_assignment_source<T, U> \|\| (is_async_alias_v<T> && detail::async_alias_assignable_from<T, async_value_t<U> const&>))
void	async_assign (WriteBuffer< T > lhs, U &&rhs)
	Assigns an Async or scalar value into a write buffer destination.

template<typename A , typename B , async_writer Writer, typename Op > requires async_binary_applicable<A, B, Op>
void	async_binary_op (Writer &out, A &&a, B &&b, Op op)
	Launch a coroutine computing result = op(a, b)

template<typename T , typename U , typename Op >
void	async_compound_op (Async< T > &lhs, U &&rhs, Op op)
	Launch a coroutine applying an in-place operation `op(lhs, rhs)`.

template<typename T , typename U > requires (!is_async_alias_v<T> && detail::async_source<U> && std::constructible_from<T, async_value_t<U> const&>)
void	async_initialize (Async< T > &destination, U &&source)
	Initialize a fresh independent async value from another async source.

template<typename T , typename U > requires (!is_async_alias_v<T> && detail::async_initialization_awaitable<U, T>)
void	async_initialize (WriteBuffer< T > destination, U &&source)
	Initialize empty independent-value storage from an awaitable source.

template<typename T , typename U > requires (!is_async_alias_v<T>)
void	async_move (Async< T > &lhs, U &&rhs)
	Move-assign from a source expression into an Async destination.

template<typename T , typename U > requires (!is_async_alias_v<T>) && async_mutable_writer_of<U, T> && async_movable_to<U, T>
void	async_move (WriteBuffer< T > lhs, U &&rhs)

template<typename T , typename U > requires (!is_async_alias_v<T>) && (!async_writer<std::remove_cvref_t<U>>) && std::constructible_from<T, U&&>
void	async_move (WriteBuffer< T > lhs, U &&rhs)

template<typename T , typename U > requires read_buffer_awaitable_of<U, T>
void	async_negate (WriteBuffer< T > lhs, U &&rhs)
	Negate a source value asynchronously and write into a destination buffer.

template<typename T >
void	async_print (std::string Format, Async< T > x)
	Schedules a formatted print once an async value is ready.

template<typename T >
void	async_read (std::string Prompt, Async< T > &x)
	Schedules reading a value from `std::cin` into an async slot.

template<typename T >
AsyncTask	co_conj (ReadBuffer< T > in, WriteBuffer< T > out)
	Coroutine implementation of complex conjugation over async buffers.

template<typename T >
AsyncTask	co_cos (ReadBuffer< T > in, WriteBuffer< T > out)
	Coroutine implementation of `cos` over async buffers.

template<typename T >
AsyncTask	co_herm (ReadBuffer< T > in, WriteBuffer< T > out)
	Coroutine implementation of Hermitian conjugation over async buffers.

template<typename T >
AsyncTask	co_imag (ReadBuffer< T > in, WriteBuffer< uni20::make_real_t< T > > out)
	Coroutine implementation of imaginary-part extraction over async buffers.

template<typename T >
AsyncTask	co_real (ReadBuffer< T > in, WriteBuffer< uni20::make_real_t< T > > out)
	Coroutine implementation of real-part extraction over async buffers.

template<typename T >
AsyncTask	co_sin (ReadBuffer< T > in, WriteBuffer< T > out)
	Coroutine implementation of `sin` over async buffers.

template<typename T >
Async< T >	conj (Async< T > const &x)
	Schedules asynchronous conjugation.

template<uni20::TensorView Tensor> requires uni20::Complex<uni20::tensor_element_t<Tensor>>
auto	conj (Async< Tensor > const &tensor)
	Return a lazy conjugating alias of an async complex tensor.

template<uni20::TensorView Tensor> requires (!uni20::Complex<uni20::tensor_element_t<Tensor>>)
auto	conj (Async< Tensor > const &tensor)
	Return a lazy read-only identity alias of an async real tensor.

template<typename T >
Async< T >	cos (Async< T > const &x)
	Schedules asynchronous cosine evaluation.

template<typename T >
Var< T >	cos (Var< T > x)
	Computes `cos(x)` and wires reverse-mode accumulation for the input gradient.

CudaTaskPromise &	cuda_promise (TaskHandle handle) noexcept
	Narrow an erased task promise after verifying that it is CUDA-specific.

template<typename T >
	Defer (Async< T > &) -> Defer< T >
	CTAD guide for constructing `Defer<T>` from `Async<T>&`.

template<typename T >
Defer< T >	defer_write (Async< T > &a)
	Create a deferred writer helper for an Async value.

template<class Function > requires std::invocable<Function&, std::size_t>
void	execute_batch (std::size_t size, Function &&function)
	Execute a synchronous lightweight batch on the global scheduler.

constexpr std::string_view	format_as (uni20::async::EpochContext::Phase p)

template<typename U >
decltype(auto)	get_awaiter (U &&u) noexcept
	for an awaitable `a`, return the actual Awaitable object returned by co_await

IAsyncScheduler *	get_global_scheduler ()
	Returns the currently-installed process-wide scheduler.

template<typename T >
Async< T >	herm (Async< T > const &x)
	Schedules asynchronous Hermitian conjugation.

template<typename T >
Async< uni20::make_real_t< T > >	imag (Async< T > const &x)
	Schedules asynchronous imaginary-part extraction.

template<typename T >
Var< uni20::make_real_t< T > >	imag (Var< T > z)
	Computes the imaginary part and propagates gradient to a complex source.

template<typename Alias , typename Parent , typename... Args>
Async< Alias >	make_async_alias (Async< Parent > const &parent, Args &&... args)
	Construct an async alias that retains a parent value and shares its epoch queue.

template<typename Alias , typename Parent , typename... Args>
Async< Alias >	make_async_alias_from_parent (Async< Parent > &parent, Args &&... args)
	Construct an async alias from a mutable parent descriptor.

template<typename Alias , typename Parent , typename... Args>
Async< Alias >	make_async_alias_from_parent (Async< Parent > const &parent, Args &&... args)
	Construct a read-only async alias from a parent descriptor.

template<typename T , typename... Args>
shared_storage< T >	make_shared_storage (Args &&... args)
	Create a new shared_storage<T> with T(args...) in-place.

template<typename T , typename Owner , typename... Args>
shared_storage< T >	make_shared_storage_alias (shared_storage< Owner > const &owner, Args &&... args)
	Create storage for a value that keeps another storage value alive.

template<typename T >
shared_storage< T >	make_unconstructed_shared_storage ()
	Create a new shared_storage<T> with an unconstructed T.

template<typename T >
Var< T >	operator* (T x, Var< T > y)
	Computes `x * y` with reverse-mode gradient propagation.

template<typename T >
Var< T >	operator* (Var< T > x, T y)
	Computes `x * y` with reverse-mode gradient propagation.

template<typename T >
Var< T >	operator* (Var< T > x, Var< T > y)
	Computes `x * y` with reverse-mode gradient propagation.

template<typename T >
Var< T >	operator+ (T x, Var< T > y)
	Computes `x + y` with reverse-mode gradient propagation.

template<typename T >
Var< T >	operator+ (Var< T > x, T y)
	Computes `x + y` with reverse-mode gradient propagation.

template<typename T >
Var< T >	operator+ (Var< T > x, Var< T > y)
	Computes `x + y` with reverse-mode gradient propagation.

template<typename T > requires requires(T t) { -t; }
auto	operator- (Async< T > const &x)
	Unary negation overload for Async values.

template<typename T >
Var< T >	operator- (T x, Var< T > y)
	Computes `x - y` with reverse-mode gradient propagation.

template<typename T >
Var< T >	operator- (Var< T > x, T y)
	Computes `x - y` with reverse-mode gradient propagation.

template<typename T >
Var< T >	operator- (Var< T > x, Var< T > y)
	Computes `x - y` with reverse-mode gradient propagation.

template<typename T >
void	ProcessCoroutineArgument (TaskPromiseBase *promise, ReadBuffer< T > const &x)
	Register coroutine debug dependencies for a read-buffer argument.

template<typename T >
void	ProcessCoroutineArgument (TaskPromiseBase *promise, T const &value)
	Process a coroutine argument for debug metadata (default no-op).

template<typename T >
void	ProcessCoroutineArgument (TaskPromiseBase *promise, WriteBuffer< T > const &x)
	Register coroutine debug dependencies and exception sink for a write-buffer argument.

template<exception_sink_buffer... Buffers>
auto	propagate_exceptions_to (Buffers &... buffers)
	Register buffers that should receive unhandled coroutine exceptions.

void	propagate_unhandled_writer_exception (EpochContext *epoch, std::exception_ptr eptr) noexcept
	Inject an unhandled writer-side coroutine exception into an epoch.

template<typename T >
ReadBuffer< T >	read (Async< T > const &a)
	Convenience helper that forwards to Async<T>::read().

template<typename T >
ReadBuffer< T >	read (ReverseValue< T > const &x)
	Free-function read adapter for ReverseValue.

template<typename T > requires detail::AsyncReadProvider<T>
decltype(auto)	read (T &&value)
	Obtain the read awaiter exposed by an async-like value.

template<typename T > requires (!detail::AsyncReadProvider<T>)
auto	read (T &&value) -> ValueAwaiter< std::remove_cvref_t< T > >
	Wrap an immediate value in an always-ready read awaiter.

template<typename T >
Async< uni20::make_real_t< T > >	real (Async< T > const &x)
	Schedules asynchronous real-part extraction.

template<typename T >
Var< uni20::make_real_t< T > >	real (Var< T > z)
	Computes the real part and propagates gradient to a complex source.

template<typename T >
ReadBuffer< T > &&	release (ReadBuffer< T > &&in)
	Adaptor for forwarding an rvalue ReadBuffer into an await expression.

template<typename T >
ReadBuffer< T > &&	release (ReadBuffer< T > &in)
	Adaptor for transferring ReadBuffer ownership into an await expression.

void	reset_global_scheduler ()
	Restores the process-wide scheduler to the default debug scheduler.

template<uni20::MutableStridedTensorView Tensor, std::integral... Extents> requires uni20::detail::CanonicalReshapeLayout<typename uni20::tensor_mdspec_t<Tensor>::layout_type>
auto	reshape_view (Async< Tensor > &tensor, Extents... requested_extents)
	Return a mutable structural reshape alias of an async tensor.

template<uni20::StridedTensorView Tensor, std::integral... Extents> requires uni20::detail::CanonicalReshapeLayout<typename uni20::tensor_mdspec_t<Tensor>::layout_type>
auto	reshape_view (Async< Tensor > const &tensor, Extents... requested_extents)
	Return a read-only structural reshape alias of an async tensor.

template<uni20::MutableStridedTensorView Tensor, std::integral... Extents>
auto	reshape_view_left (Async< Tensor > &tensor, Extents... requested_extents)
	Return a mutable column-major structural reshape alias.

template<uni20::StridedTensorView Tensor, std::integral... Extents>
auto	reshape_view_left (Async< Tensor > const &tensor, Extents... requested_extents)
	Return a read-only column-major structural reshape alias.

template<uni20::MutableStridedTensorView Tensor, std::integral... Extents>
auto	reshape_view_right (Async< Tensor > &tensor, Extents... requested_extents)
	Return a mutable row-major structural reshape alias.

template<uni20::StridedTensorView Tensor, std::integral... Extents>
auto	reshape_view_right (Async< Tensor > const &tensor, Extents... requested_extents)
	Return a read-only row-major structural reshape alias.

void	schedule (AsyncTask &&task)
	Schedule a task on the currently configured global scheduler.

void	schedule (CudaTask &&task)
	Schedule a CUDA task on the currently configured global scheduler.

void	schedule (CudaTask &&task, int device)
	Schedule a CUDA task with initial device affinity.

void	set_global_scheduler (IAsyncScheduler *sched)
	Sets the process-wide scheduler used by async helpers.

template<typename T >
Async< T >	sin (Async< T > const &x)
	Schedules asynchronous sine evaluation.

template<typename T >
Var< T >	sin (Var< T > x)
	Computes `sin(x)` and wires reverse-mode accumulation for the input gradient.

TaskRoute	task_route (TaskHandle handle) noexcept
	Return the current scheduling requirements recorded by a task.

template<typename Aw >
constexpr auto	try_await (Aw &&aw) noexcept
	Build a non-blocking awaiter from an rvalue awaitable.

template<typename Aw >
constexpr auto	try_await (Aw &aw) noexcept
	Build a non-blocking awaiter that returns `optional<T>` instead of suspending.

template<typename T >
WriteBuffer< T >	write (Async< T > &a)
	Convenience helper that forwards to Async<T>::write().

template<typename T , typename Value >
auto	write_to (WriteBuffer< T > &&buffer, Value value)
	Create a coroutine awaiter that writes a value to a WriteBuffer<T>.

Variables
constexpr async_do_not_start_t	async_do_not_start {}
	Tag constant to construct an Async without starting the queue object.

template<typename T >
constexpr bool	is_async_alias_v = is_async_alias<std::remove_cvref_t<T>>::value
	True when `T` is an owner-bound async alias descriptor.

Detailed Description

This represents the reverse-mode adjoint for functions of complex variables, assuming a real-valued scalar loss function ( L \colon \mathbb{C} \to \mathbb{R} ).

Only the Wirtinger derivative with respect to ( z^* ) is propagated:

\[ \bar{z} := \frac{\partial L}{\partial z^*} \]

For an intermediate variable ( f(z) ), the chain rule is:

\[ \frac{\partial L}{\partial z^*} = \frac{\partial L}{\partial f} \cdot \frac{\partial f}{\partial z^*} + \frac{\partial L}{\partial f^*} \cdot \frac{\partial f^*}{\partial z^*} \]

Under the real-valued loss assumption:

\[ \frac{\partial L}{\partial f^*} = \overline{ \left( \frac{\partial L}{\partial f} \right) }, \quad \frac{\partial f^*}{\partial z^*} = \overline{ \left( \frac{\partial f}{\partial z} \right) } \]

so the full gradient becomes:

\[ \frac{\partial L}{\partial z^*} = \frac{\partial L}{\partial f} \cdot \frac{\partial f}{\partial z^*} + \overline{ \left( \frac{\partial L}{\partial f} \cdot \frac{\partial f}{\partial z} \right) } \]

Note: This implementation assumes the upstream gradient is available via grad, and all reverse-mode propagation targets the Wirtinger ( \partial / \partial z^* ) direction.

Typedef Documentation

◆ async_value_t

template<typename T >

using uni20::async::async_value_t = typedef typename async_value_type<std::remove_cvref_t<T> >::type

Extract the underlying value type from an immediate value or Async<T>.

◆ OwningWriteProxy

template<typename T >

using uni20::async::OwningWriteProxy = typedef OwningWriteAccessProxy<T>

Alias for the owning write access proxy type.

◆ WriteAwaitProxy

template<typename T >

using uni20::async::WriteAwaitProxy = typedef WriteAccessProxy<T>

Alias for the non-owning write access proxy type.

◆ WriteProxy

template<typename T >

using uni20::async::WriteProxy = typedef WriteAssignProxy<T>

Alias for assignment-only writer proxy.

Enumeration Type Documentation

◆ DebugSchedulerOrder

enum class uni20::async::DebugSchedulerOrder

strong

Runnable-batch ordering policy for DebugScheduler.

Enumerator
fifo	Execute each runnable batch in submission order.
reverse	Execute each runnable batch in reverse submission order.
random	Execute each runnable batch in seeded pseudo-random order.

◆ NodeValueState

enum class uni20::async::NodeValueState

strong

Construction state reported for a debug DAG value node.

Enumerator
Invalid	No storage or fallback value is available.
Unconstructed	Storage exists but does not currently contain a value.
Constructed	A value is currently constructed.

◆ TaskDomain

enum class uni20::async::TaskDomain

strong

Execution domain declared by a concrete Uni20 task promise.

Enumerator
none
host
cuda

Function Documentation

◆ all()

template<TaskFactoryChildAwaitable... Aw>
requires ((!std::is_void_v<decltype(std::declval<Aw>().await_resume())> && ...))

auto uni20::async::all ( Aw &&... aw )

noexcept

Build an awaitable that waits for all of the provided awaitables.

Parameters

aw	Awaitable arguments.

Returns: An object supporting co_await.

◆ async_accumulate()

template<typename T , typename U >

AsyncTask uni20::async::async_accumulate	(	ReadBuffer< T >	a_,
		ReadBuffer< U >	b_,
		WriteBuffer< T >	out_
	)

Compute out = a + b while tolerating cancellation of either input.

Template Parameters

T	Left/output value type.
U	Right input value type.

Parameters

a_	Left input read buffer.
b_	Right input read buffer.
out_	Output write buffer.

◆ async_accumulate_minus()

template<typename T , typename U >

AsyncTask uni20::async::async_accumulate_minus	(	ReadBuffer< T >	a_,
		ReadBuffer< U >	b_,
		WriteBuffer< T >	out_
	)

Compute out = a - b while tolerating cancellation of either input.

Template Parameters

T	Left/output value type.
U	Right input value type.

Parameters

a_	Left input read buffer.
b_	Right input read buffer.
out_	Output write buffer.

◆ async_assign() [1/2]

template<typename T , typename U >
requires (is_async_alias_v<T> && detail::async_alias_assignment_awaitable<U, T>)

void uni20::async::async_assign	(	WriteBuffer< T >	destination,
		U &&	source
	)

Assign through an existing async alias descriptor.

The destination descriptor, lifetime owner, and epoch queue remain intact. The write proxy invokes the alias's ADL customization.

◆ async_assign() [2/2]

template<typename T , typename U >
requires (!read_buffer_awaitable<std::remove_cvref_t<U>>) && (detail::current_value_assignment_source<T, U> || (is_async_alias_v<T> && detail::async_alias_assignable_from<T, async_value_t<U> const&>))

void uni20::async::async_assign	(	WriteBuffer< T >	lhs,
		U &&	rhs
	)

Assigns an Async or scalar value into a write buffer destination.

Template Parameters

T	Destination value type.
U	Source type (scalar or Async-like).

Parameters

lhs	Destination write buffer.
rhs	Source expression.

◆ async_move()

template<typename T , typename U >
requires (!is_async_alias_v<T>) && (!async_writer<std::remove_cvref_t<U>>) && std::constructible_from<T, U&&>

void uni20::async::async_move	(	WriteBuffer< T >	lhs,
		U &&	rhs
	)

◆ async_print()

template<typename T >

void uni20::async::async_print	(	std::string	Format,
		Async< T >	x
	)

Schedules a formatted print once an async value is ready.

Template Parameters

T	Value type.

Parameters

Format	`fmt`-style format string.
x	Async value to print.

◆ async_read()

template<typename T >

void uni20::async::async_read	(	std::string	Prompt,
		Async< T > &	x
	)

Schedules reading a value from std::cin into an async slot.

Template Parameters

T	Value type.

Parameters

Prompt	Prompt text printed before reading.
x	Async destination value.

◆ co_conj()

template<typename T >

AsyncTask uni20::async::co_conj	(	ReadBuffer< T >	in,
		WriteBuffer< T >	out
	)

Coroutine implementation of complex conjugation over async buffers.

Template Parameters

T	Value type.

Parameters

in	Input read buffer.
out	Output write buffer.

◆ co_cos()

template<typename T >

AsyncTask uni20::async::co_cos	(	ReadBuffer< T >	in,
		WriteBuffer< T >	out
	)

Coroutine implementation of cos over async buffers.

Template Parameters

T	Value type.

Parameters

in	Input read buffer.
out	Output write buffer.

◆ co_herm()

template<typename T >

AsyncTask uni20::async::co_herm	(	ReadBuffer< T >	in,
		WriteBuffer< T >	out
	)

Coroutine implementation of Hermitian conjugation over async buffers.

Template Parameters

T	Value type.

Parameters

in	Input read buffer.
out	Output write buffer.

◆ co_imag()

template<typename T >

AsyncTask uni20::async::co_imag	(	ReadBuffer< T >	in,
		WriteBuffer< uni20::make_real_t< T > >	out
	)

Coroutine implementation of imaginary-part extraction over async buffers.

Template Parameters

T	Value type.

Parameters

in	Input read buffer.
out	Output write buffer.

◆ co_real()

template<typename T >

AsyncTask uni20::async::co_real	(	ReadBuffer< T >	in,
		WriteBuffer< uni20::make_real_t< T > >	out
	)

Coroutine implementation of real-part extraction over async buffers.

Template Parameters

T	Value type.

Parameters

in	Input read buffer.
out	Output write buffer.

◆ co_sin()

template<typename T >

AsyncTask uni20::async::co_sin	(	ReadBuffer< T >	in,
		WriteBuffer< T >	out
	)

Coroutine implementation of sin over async buffers.

Template Parameters

T	Value type.

Parameters

in	Input read buffer.
out	Output write buffer.

◆ conj() [1/3]

template<typename T >

Async< T > uni20::async::conj ( Async< T > const & x )

Schedules asynchronous conjugation.

Template Parameters

T	Value type.

Parameters

x	Input async value.

Returns: Async result of conj(x).

◆ conj() [2/3]

template<uni20::TensorView Tensor>
requires uni20::Complex<uni20::tensor_element_t<Tensor>>

auto uni20::async::conj ( Async< Tensor > const & tensor )

Return a lazy conjugating alias of an async complex tensor.

The result retains the parent tensor storage and shares its exact epoch queue. Awaiting the alias therefore observes the same causal timeline as awaiting the parent. Default precision is read from that parent epoch rather than copied when the alias is created.

◆ conj() [3/3]

template<uni20::TensorView Tensor>
requires (!uni20::Complex<uni20::tensor_element_t<Tensor>>)

auto uni20::async::conj ( Async< Tensor > const & tensor )

Return a lazy read-only identity alias of an async real tensor.

◆ cos() [1/2]

template<typename T >

Async< T > uni20::async::cos ( Async< T > const & x )

Schedules asynchronous cosine evaluation.

Template Parameters

T	Value type.

Parameters

x	Input async value.

Returns: Async result of cos(x).

◆ cos() [2/2]

template<typename T >

Var< T > uni20::async::cos ( Var< T > x )

Computes cos(x) and wires reverse-mode accumulation for the input gradient.

Template Parameters

T	Value type.

Parameters

x	Input variable.

Returns: Output variable representing cos(x).

◆ cuda_promise()

CudaTaskPromise & uni20::async::cuda_promise ( TaskHandle handle )

inlinenoexcept

Narrow an erased task promise after verifying that it is CUDA-specific.

Parameters

handle Erased task identity produced from its original typed coroutine handle.

Returns: The concrete CUDA promise.

◆ Defer()

template<typename T >

uni20::async::Defer ( Async< T > & ) -> Defer< T >

CTAD guide for constructing Defer<T> from Async<T>&.

◆ defer_write()

template<typename T >

Defer< T > uni20::async::defer_write ( Async< T > & a )

Create a deferred writer helper for an Async value.

Template Parameters

T	Stored value type.

Parameters

a	Async value to write.

Returns: Deferred writer helper.

◆ execute_batch()

template<class Function >
requires std::invocable<Function&, std::size_t>

void uni20::async::execute_batch	(	std::size_t	size,
		Function &&	function
	)

Execute a synchronous lightweight batch on the global scheduler.

Template Parameters

Function Callable accepting one std::size_t index.

Parameters

size	Number of work items.
function	Work-item callable borrowed until the call returns.

◆ format_as()

constexpr std::string_view uni20::async::format_as ( uni20::async::EpochContext::Phase p )

constexpr

◆ get_awaiter()

template<typename U >

decltype(auto) uni20::async::get_awaiter ( U && u )

noexcept

for an awaitable a, return the actual Awaitable object returned by co_await

Template Parameters

U	Awaitable or awaiter type.

Parameters

u	The object to transform.

Returns: The awaiter object to use.

Note: Within a coroutine, this type may be transformed into something else via the promise_type::await_transform() function.

◆ get_global_scheduler()

IAsyncScheduler * uni20::async::get_global_scheduler ( )

inline

Returns the currently-installed process-wide scheduler.

Returns: Pointer to the active scheduler.

◆ herm()

template<typename T >

Async< T > uni20::async::herm ( Async< T > const & x )

Schedules asynchronous Hermitian conjugation.

Template Parameters

T	Value type.

Parameters

x	Input async value.

Returns: Async result of herm(x).

◆ imag() [1/2]

template<typename T >

Async< uni20::make_real_t< T > > uni20::async::imag ( Async< T > const & x )

Schedules asynchronous imaginary-part extraction.

Template Parameters

T	Value type.

Parameters

x	Input async value.

Returns: Async result containing imag(x).

◆ imag() [2/2]

template<typename T >

Var< uni20::make_real_t< T > > uni20::async::imag ( Var< T > z )

Computes the imaginary part and propagates gradient to a complex source.

Template Parameters

T	Complex-like value type.

Parameters

z	Input variable.

Returns: Output variable representing imag(z).

◆ make_async_alias()

template<typename Alias , typename Parent , typename... Args>

Async< Alias > uni20::async::make_async_alias	(	Async< Parent > const &	parent,
		Args &&...	args
	)

Construct an async alias that retains a parent value and shares its epoch queue.

◆ make_async_alias_from_parent() [1/2]

template<typename Alias , typename Parent , typename... Args>

Async< Alias > uni20::async::make_async_alias_from_parent	(	Async< Parent > &	parent,
		Args &&...	args
	)

Construct an async alias from a mutable parent descriptor.

The alias constructor receives the address reserved for the parent value followed by args. The address may identify unconstructed storage and must not be dereferenced before the shared epoch is readable.

◆ make_async_alias_from_parent() [2/2]

template<typename Alias , typename Parent , typename... Args>

Async< Alias > uni20::async::make_async_alias_from_parent	(	Async< Parent > const &	parent,
		Args &&...	args
	)

Construct a read-only async alias from a parent descriptor.

◆ make_shared_storage()

template<typename T , typename... Args>

shared_storage< T > uni20::async::make_shared_storage ( Args &&... args )

inline

Create a new shared_storage<T> with T(args...) in-place.

Template Parameters

T	Stored value type.
Args	Constructor argument types.

Parameters

args	Constructor arguments forwarded to `T`.

Returns: New shared_storage<T> with a constructed value.

◆ make_shared_storage_alias()

template<typename T , typename Owner , typename... Args>

shared_storage< T > uni20::async::make_shared_storage_alias	(	shared_storage< Owner > const &	owner,
		Args &&...	args
	)

inline

Create storage for a value that keeps another storage value alive.

The alias has its own local strong count. Its control block retains one reference to owner, which is released after the alias value is destroyed. Copies of the alias increment only the local count.

Template Parameters

T	Alias value type.
Owner	Lifetime-owner value type.
Args	Alias constructor argument types.

Parameters

owner	Storage whose lifetime must cover the alias value.
args	Arguments forwarded to the alias value constructor.

Returns: Constructed alias storage sharing the owner's lifetime.

◆ make_unconstructed_shared_storage()

template<typename T >

shared_storage< T > uni20::async::make_unconstructed_shared_storage ( )

inline

Create a new shared_storage<T> with an unconstructed T.

◆ operator*() [1/3]

template<typename T >

Var< T > uni20::async::operator*	(	T	x,
		Var< T >	y
	)

Computes x * y with reverse-mode gradient propagation.

Template Parameters

T	Value type.

Parameters

x	Left scalar operand.
y	Right input variable.

Returns: Output variable representing x * y.

◆ operator*() [2/3]

template<typename T >

Var< T > uni20::async::operator*	(	Var< T >	x,
		T	y
	)

Computes x * y with reverse-mode gradient propagation.

Template Parameters

T	Value type.

Parameters

x	Left input variable.
y	Right scalar operand.

Returns: Output variable representing x * y.

◆ operator*() [3/3]

template<typename T >

Var< T > uni20::async::operator*	(	Var< T >	x,
		Var< T >	y
	)

Computes x * y with reverse-mode gradient propagation.

Template Parameters

T	Value type.

Parameters

x	Left input variable.
y	Right input variable.

Returns: Output variable representing x * y.

◆ operator+() [1/3]

template<typename T >

Var< T > uni20::async::operator+	(	T	x,
		Var< T >	y
	)

Computes x + y with reverse-mode gradient propagation.

Template Parameters

T	Value type.

Parameters

x	Left scalar operand.
y	Right input variable.

Returns: Output variable representing x + y.

◆ operator+() [2/3]

template<typename T >

Var< T > uni20::async::operator+	(	Var< T >	x,
		T	y
	)

Computes x + y with reverse-mode gradient propagation.

Template Parameters

T	Value type.

Parameters

x	Left input variable.
y	Right scalar operand.

Returns: Output variable representing x + y.

◆ operator+() [3/3]

template<typename T >

Var< T > uni20::async::operator+	(	Var< T >	x,
		Var< T >	y
	)

Computes x + y with reverse-mode gradient propagation.

Template Parameters

T	Value type.

Parameters

x	Left input variable.
y	Right input variable.

Returns: Output variable representing x + y.

◆ operator-() [1/3]

template<typename T >

Var< T > uni20::async::operator-	(	T	x,
		Var< T >	y
	)

Computes x - y with reverse-mode gradient propagation.

Template Parameters

T	Value type.

Parameters

x	Left scalar operand.
y	Right input variable.

Returns: Output variable representing x - y.

◆ operator-() [2/3]

template<typename T >

Var< T > uni20::async::operator-	(	Var< T >	x,
		T	y
	)

Computes x - y with reverse-mode gradient propagation.

Template Parameters

T	Value type.

Parameters

x	Left input variable.
y	Right scalar operand.

Returns: Output variable representing x - y.

◆ operator-() [3/3]

template<typename T >

Var< T > uni20::async::operator-	(	Var< T >	x,
		Var< T >	y
	)

Computes x - y with reverse-mode gradient propagation.

Template Parameters

T	Value type.

Parameters

x	Left input variable.
y	Right input variable.

Returns: Output variable representing x - y.

◆ ProcessCoroutineArgument() [1/3]

template<typename T >

void uni20::async::ProcessCoroutineArgument	(	TaskPromiseBase *	promise,
		ReadBuffer< T > const &	x
	)

Register coroutine debug dependencies for a read-buffer argument.

Template Parameters

T	Stored value type.

Parameters

promise	Promise collecting dependency metadata.
x	Read buffer argument.

◆ ProcessCoroutineArgument() [2/3]

template<typename T >

void uni20::async::ProcessCoroutineArgument	(	TaskPromiseBase *	promise,
		T const &	value
	)

Process a coroutine argument for debug metadata (default no-op).

Template Parameters

T	Argument type.

Parameters

promise	Promise receiving metadata.
value	Argument ignored by default.

◆ ProcessCoroutineArgument() [3/3]

template<typename T >

void uni20::async::ProcessCoroutineArgument	(	TaskPromiseBase *	promise,
		WriteBuffer< T > const &	x
	)

Register coroutine debug dependencies and exception sink for a write-buffer argument.

Template Parameters

T	Stored value type.

Parameters

promise	Promise collecting dependency metadata.
x	Write buffer argument.

◆ propagate_exceptions_to()

template<exception_sink_buffer... Buffers>

auto uni20::async::propagate_exceptions_to ( Buffers &... buffers )

Register buffers that should receive unhandled coroutine exceptions.

The registration remains active for as long as each buffer object remains alive.

Note: If a registered explicit sink is destroyed during stack unwinding, the runtime aborts.

◆ propagate_unhandled_writer_exception()

void uni20::async::propagate_unhandled_writer_exception	(	EpochContext *	epoch,
		std::exception_ptr	eptr
	)

inlinenoexcept

Inject an unhandled writer-side coroutine exception into an epoch.

Parameters

epoch	Target epoch context associated with a writer buffer.
eptr	Exception captured by the coroutine promise.

◆ read() [1/4]

template<typename T >

ReadBuffer< T > uni20::async::read ( Async< T > const & a )

Convenience helper that forwards to Async<T>::read().

Template Parameters

T	Stored value type.

Parameters

a	Async container providing the read buffer.

Returns: Read buffer obtained from the Async instance.

◆ read() [2/4]

template<typename T >

ReadBuffer< T > uni20::async::read ( ReverseValue< T > const & x )

Free-function read adapter for ReverseValue.

Template Parameters

T	Stored value type.

Parameters

x	Reverse value.

Returns: Input read buffer for the reverse channel.

◆ read() [3/4]

template<typename T >
requires detail::AsyncReadProvider<T>

decltype(auto) uni20::async::read ( T && value )

Obtain the read awaiter exposed by an async-like value.

◆ read() [4/4]

template<typename T >
requires (!detail::AsyncReadProvider<T>)

auto uni20::async::read ( T && value ) -> ValueAwaiter<std::remove_cvref_t<T>>

Wrap an immediate value in an always-ready read awaiter.

◆ real() [1/2]

template<typename T >

Async< uni20::make_real_t< T > > uni20::async::real ( Async< T > const & x )

Schedules asynchronous real-part extraction.

Template Parameters

T	Value type.

Parameters

x	Input async value.

Returns: Async result containing real(x).

◆ real() [2/2]

template<typename T >

Var< uni20::make_real_t< T > > uni20::async::real ( Var< T > z )

Computes the real part and propagates gradient to a complex source.

Template Parameters

T	Complex-like value type.

Parameters

z	Input variable.

Returns: Output variable representing real(z).

◆ release() [1/2]

template<typename T >

ReadBuffer< T > && uni20::async::release ( ReadBuffer< T > && in )

Adaptor for forwarding an rvalue ReadBuffer into an await expression.

This is a synonym for std::move.

◆ release() [2/2]

template<typename T >

ReadBuffer< T > && uni20::async::release ( ReadBuffer< T > & in )

Adaptor for transferring ReadBuffer ownership into an await expression.

This is a synonym for std::move.

◆ reset_global_scheduler()

void uni20::async::reset_global_scheduler ( )

inline

Restores the process-wide scheduler to the default debug scheduler.

◆ reshape_view() [1/2]

template<uni20::MutableStridedTensorView Tensor, std::integral... Extents>
requires uni20::detail::CanonicalReshapeLayout<typename uni20::tensor_mdspec_t<Tensor>::layout_type>

auto uni20::async::reshape_view	(	Async< Tensor > &	tensor,
		Extents...	requested_extents
	)

Return a mutable structural reshape alias of an async tensor.

The descriptor retains the parent storage and shares its exact epoch queue. Shape and layout validation occurs when the shared parent epoch first becomes readable. Default precision follows the parent state observed by each operation through the alias.

◆ reshape_view() [2/2]

template<uni20::StridedTensorView Tensor, std::integral... Extents>
requires uni20::detail::CanonicalReshapeLayout<typename uni20::tensor_mdspec_t<Tensor>::layout_type>

auto uni20::async::reshape_view	(	Async< Tensor > const &	tensor,
		Extents...	requested_extents
	)

Return a read-only structural reshape alias of an async tensor.

◆ reshape_view_left() [1/2]

template<uni20::MutableStridedTensorView Tensor, std::integral... Extents>

auto uni20::async::reshape_view_left	(	Async< Tensor > &	tensor,
		Extents...	requested_extents
	)

Return a mutable column-major structural reshape alias.

◆ reshape_view_left() [2/2]

template<uni20::StridedTensorView Tensor, std::integral... Extents>

auto uni20::async::reshape_view_left	(	Async< Tensor > const &	tensor,
		Extents...	requested_extents
	)

Return a read-only column-major structural reshape alias.

◆ reshape_view_right() [1/2]

template<uni20::MutableStridedTensorView Tensor, std::integral... Extents>

auto uni20::async::reshape_view_right	(	Async< Tensor > &	tensor,
		Extents...	requested_extents
	)

Return a mutable row-major structural reshape alias.

◆ reshape_view_right() [2/2]

template<uni20::StridedTensorView Tensor, std::integral... Extents>

auto uni20::async::reshape_view_right	(	Async< Tensor > const &	tensor,
		Extents...	requested_extents
	)

Return a read-only row-major structural reshape alias.

◆ schedule() [1/3]

void uni20::async::schedule ( AsyncTask && task )

inline

Schedule a task on the currently configured global scheduler.

Parameters

task	Task to schedule.

◆ schedule() [2/3]

void uni20::async::schedule ( CudaTask && task )

inline

Schedule a CUDA task on the currently configured global scheduler.

The installed scheduler must implement CUDA task admission.

Parameters

task	CUDA task to schedule without establishing device affinity.

◆ schedule() [3/3]

void uni20::async::schedule	(	CudaTask &&	task,
		int	device
	)

inline

Schedule a CUDA task with initial device affinity.

Initial admission to the correct device avoids a later scheduler migration when the task first selects that device.

Parameters

task	CUDA task to schedule.
device	CUDA runtime device ordinal.

◆ set_global_scheduler()

void uni20::async::set_global_scheduler ( IAsyncScheduler * sched )

inline

Sets the process-wide scheduler used by async helpers.

Parameters

sched Scheduler pointer to install.

◆ sin() [1/2]

template<typename T >

Async< T > uni20::async::sin ( Async< T > const & x )

Schedules asynchronous sine evaluation.

Template Parameters

T	Value type.

Parameters

x	Input async value.

Returns: Async result of sin(x).

◆ sin() [2/2]

template<typename T >

Var< T > uni20::async::sin ( Var< T > x )

Computes sin(x) and wires reverse-mode accumulation for the input gradient.

Template Parameters

T	Value type.

Parameters

x	Input variable.

Returns: Output variable representing sin(x).

◆ task_route()

TaskRoute uni20::async::task_route ( TaskHandle handle )

inlinenoexcept

Return the current scheduling requirements recorded by a task.

◆ try_await() [1/2]

template<typename Aw >

constexpr auto uni20::async::try_await ( Aw && aw )

constexprnoexcept

Build a non-blocking awaiter from an rvalue awaitable.

Template Parameters

Aw	Awaitable type.

Parameters

aw	Awaitable value moved into the try-await wrapper.

Returns: A TryAwaiter that owns the awaitable value.

◆ try_await() [2/2]

template<typename Aw >

constexpr auto uni20::async::try_await ( Aw & aw )

constexprnoexcept

Build a non-blocking awaiter that returns optional<T> instead of suspending.

If you pass an lvalue awaiter, it’s stored by reference. If you pass a prvalue awaiter (e.g. a proxy type you’ve designed), it’s stored by value. Exceptions from the readiness probe or result extraction propagate normally.

Example:

ReadBuffer<int> rbuf = I.read();
// ok: Aw = ReadBuffer<int>&, builder holds rbuf by reference
auto maybe_i = co_await try_await(rbuf);
 
// also ok if AsyncProxyAwaiter is a small, copyable awaiter
auto proxy = AsyncProxyAwaiter{/*…*&zwj;/};
auto maybe_v = co_await try_await(proxy);

Parameters

aw	An awaitable (must outlive the coroutine for references).

Returns: A TryAwaiter wrapper around the supplied awaitable.

◆ write()

template<typename T >

WriteBuffer< T > uni20::async::write ( Async< T > & a )

Convenience helper that forwards to Async<T>::write().

Template Parameters

T	Stored value type.

Parameters

a	Async container providing the write buffer.

Returns: Write buffer obtained from the Async instance.

◆ write_to()

template<typename T , typename Value >

auto uni20::async::write_to	(	WriteBuffer< T > &&	buffer,
		Value	value
	)

Create a coroutine awaiter that writes a value to a WriteBuffer<T>.

Constructs an awaitable object that can be co_awaited to write a given value into an Async<T> via its WriteBuffer. Ensures lifetime safety and correct causal ordering when the buffer is a temporary.

Parameters

buffer	A WriteBuffer<T>, moved into the coroutine.
value	The value to assign, moved into the awaiter frame.

Returns: An awaitable object that performs the write upon resumption.

Variable Documentation

◆ async_do_not_start

constexpr async_do_not_start_t uni20::async::async_do_not_start {}

inlineconstexpr

Tag constant to construct an Async without starting the queue object.

◆ is_async_alias_v

template<typename T >

constexpr bool uni20::async::is_async_alias_v = is_async_alias<std::remove_cvref_t<T>>::value

inlineconstexpr

True when T is an owner-bound async alias descriptor.

Classes

Concepts

Typedefs

Enumerations

Functions

Variables

Detailed Description

Typedef Documentation

◆ async_value_t

◆ OwningWriteProxy

◆ WriteAwaitProxy

◆ WriteProxy

Enumeration Type Documentation

◆ DebugSchedulerOrder

◆ NodeValueState

◆ TaskDomain

Function Documentation

◆ all()

◆ async_accumulate()

◆ async_accumulate_minus()

◆ async_assign() [1/2]

◆ async_assign() [2/2]

◆ async_move()

◆ async_print()

◆ async_read()

◆ co_conj()

◆ co_cos()

◆ co_herm()

◆ co_imag()

◆ co_real()

◆ co_sin()

◆ conj() [1/3]

◆ conj() [2/3]

◆ conj() [3/3]

◆ cos() [1/2]

◆ cos() [2/2]

◆ cuda_promise()

◆ Defer()

◆ defer_write()

◆ execute_batch()

◆ format_as()

◆ get_awaiter()

◆ get_global_scheduler()

◆ herm()

◆ imag() [1/2]

◆ imag() [2/2]

◆ make_async_alias()

◆ make_async_alias_from_parent() [1/2]

◆ make_async_alias_from_parent() [2/2]

◆ make_shared_storage()

◆ make_shared_storage_alias()

◆ make_unconstructed_shared_storage()

◆ operator*() [1/3]

◆ operator*() [2/3]

◆ operator*() [3/3]

◆ operator+() [1/3]

◆ operator+() [2/3]

◆ operator+() [3/3]

◆ operator-() [1/3]

◆ operator-() [2/3]

◆ operator-() [3/3]

◆ ProcessCoroutineArgument() [1/3]

◆ ProcessCoroutineArgument() [2/3]

◆ ProcessCoroutineArgument() [3/3]

◆ propagate_exceptions_to()

◆ propagate_unhandled_writer_exception()

◆ read() [1/4]

◆ read() [2/4]

◆ read() [3/4]

◆ read() [4/4]

◆ real() [1/2]

◆ real() [2/2]

◆ release() [1/2]

◆ release() [2/2]

◆ reset_global_scheduler()

◆ reshape_view() [1/2]

◆ reshape_view() [2/2]

◆ reshape_view_left() [1/2]

◆ reshape_view_left() [2/2]

◆ reshape_view_right() [1/2]