starpu-max/doc/devel/replicate_refcnt

#!/bin/bash
# StarPU --- Runtime system for heterogeneous multicore architectures.
#
# Copyright (C) 2012-2020  Université de Bordeaux, CNRS (LaBRI UMR 5800), Inria
#
# StarPU is free software; you can redistribute it and/or modify
# it under the terms of the GNU Lesser General Public License as published by
# the Free Software Foundation; either version 2.1 of the License, or (at
# your option) any later version.
#
# StarPU is distributed in the hope that it will be useful, but
# WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
#
# See the GNU Lesser General Public License in COPYING.LGPL for more details.
#
replicate.refcnt usage
======================

It is used to make sure the replicate still exists on some memory node.

It is thus used

- during task duration for each data
- during data requests, to make sure that the source and target still exist
- during data reduction, to keep the per-worker data available for reduction
- during application data acquisition

Function detail
---------------

* makes sure there is at least a reference:
  - copy_data_1_to_1_generic()
  - _starpu_driver_copy_data_1_to_1()
  - _starpu_post_data_request()
  - starpu_handle_data_request()

* creates a reference:
  * Released by starpu_handle_data_request_completion:
    - _starpu_create_data_request():
      - 1 on dst_replicate
      - 1 on src_replicate if (mode & STARPU_R)
    - _starpu_search_existing_data_request() on src_replicate when turning a request without STARPU_R into one that does.

  * Released by _starpu_data_end_reduction_mode_terminate:
    - _starpu_data_end_reduction_mode() for each initialized per-worker buffer, creates a reduction_tmp_handles for them

  * Released by _starpu_release_data_on_node or inside _starpu_prefetch_data_on_node_with_mode:
    - _starpu_fetch_data_on_node() when !detached

* indirectly creates a reference:
  * Released by _starpu_push_task_output():
    - fetch_data()
    - _starpu_fetch_task_input() through fetch_data() for each task data (one per unique data)
  * Released by starpu_data_release():
    - starpu_data_acquire_cb()

* releases a reference:
  - starpu_handle_data_request_completion()
    - 1 on dst_replicate
    - 1 on src_replicate if (mode & STARPU_R)
  - _starpu_data_end_reduction_mode_terminate() for each per-worker buffer which has a reduction_tmp_handles
  - _starpu_release_data_on_node()

* indirectly releases a reference:
  - starpu_handle_data_request() through starpu_handle_data_request_completion() when returning 0.
  - _starpu_push_task_output() for each task data (one per unique data)
  - starpu_data_release()
  - _starpu_handle_node_data_requests() through _starpu_handle_node_data_requests() for each completed request, which is not put back on any list.

* temporarily increases, and decreases after:
  - transfer_subtree_to_node()
  - _starpu_allocate_interface()
  - _starpu_prefetch_data_on_node_with_mode(), when produced by the call to _starpu_fetch_data_on_node (!detached)
  - starpu_data_prefetch_on_node(), 
  - starpu_data_idle_prefetch_on_node(), 
  - starpu_data_invalidate()
  - _starpu_data_unregister()
  - _starpu_benchmark_ping_pong()