Skip to main content

class dry_run_executor

Declaration

class dry_run_executor : public executor { /* full declaration omitted */ };

Description

Executor implementation selected when Celerity performs a dry run, that is, graph generation for debugging purposes without actually allocating memory, launching kernels or issuing data transfers. dry_run_executor still executes horizon-, epoch- and fence instructions to guarantee forward progress in the user application.

Inherits from: executor

Member Variables

double_buffered_queue<submission> m_submission_queue
std::thread m_thread

Member Aliases

  • using host_object_transfer = std::pair< host_object_id, std::unique_ptr<host_object_instance>>;
  • using submission = std::variant<std::vector<const instruction*>, host_object_transfer>;

Member Function Overview

Inherited from executor:

Member Functions

explicit dry_run_executor(delegate* dlg)

Description

dlg (optional) receives notifications about reached horizons and epochs from the executor thread.

Parameters

delegate* dlg

dry_run_executor(const dry_run_executor&)

Parameters

const dry_run_executor& 

dry_run_executor(dry_run_executor&&)

Parameters

dry_run_executor&& 

dry_run_executor& operator=(
    const dry_run_executor&)

Parameters

const dry_run_executor& 

dry_run_executor& operator=(dry_run_executor&&)

Parameters

dry_run_executor&& 

void submit(
    std::vector<const instruction*> instructions,
    std::vector<outbound_pilot> pilots)

Description

Submits a list of instructions to execute once their dependencies have been fulfilled, and a list of outbound pilots to be transmitted to their recipients as soon as possible. Instructions must be in topological order of dependencies, as must be the concatenation of all vectors passed to subsequent invocations of this function.

Parameters

std::vector<const instruction*> instructions
std::vector<outbound_pilot> pilots

void track_host_object_instance(
    host_object_id hoid,
    std::unique_ptr<host_object_instance>
        instance)

Description

Transfer ownership of a host object instance to the executor. The executor will later destroy this instance when executing a matching destroy_host_object_instruction.

Parameters

host_object_id hoid
std::unique_ptr<host_object_instance> instance

void track_reducer(
    reduction_id rid,
    std::unique_ptr<reducer> reducer)

Description

Informs the executor about the runtime behavior of a reduction. Will be used by any fill_identity_instruction and reduce_instruction later submitted on the same reduction_id. Reducer instances are removed from executor tracking and destroyed when they later appear in an instruction_garbage list attached to a horizon or epoch instruction.

Parameters

reduction_id rid
std::unique_ptr<reducer> reducer

void track_user_allocation(allocation_id aid,
                           void* ptr)

Description

Informs the executor about the runtime address of an allocation on user_memory_id. Must be called before submitting any instruction referring to the allocation id in question. User allocations are later removed from executor tracking as they appear in an instruction_garbage list attached to a horizon or epoch instruction.

Parameters

allocation_id aid
void* ptr

~dry_run_executor()

Description

Waits until an epoch with epoch_action::shutdown has executed and the executor thread has exited.

void thread_main(delegate* dlg)

Parameters

delegate* dlg