Skip to content

Add alternative implementation of device timer to SyclTimer class #1872

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
oleksandr-pavlyk merged 5 commits into from
Nov 15, 2024
Merged

Add alternative implementation of device timer to SyclTimer class #1872

Changes from 1 commit
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
77e3649
Add utility to submit empty task into the queue specifying dependent …
oleksandr-pavlyk Oct 4, 2024
8139fec
Extend SyclTimer
oleksandr-pavlyk Oct 4, 2024
ec6ad3b
Add test_sycl_timer file
oleksandr-pavlyk Oct 22, 2024
d1011c5
Add change log entry for change to SyclTimer
oleksandr-pavlyk Oct 22, 2024
7752078
Expanded docstring as suggested in review
oleksandr-pavlyk Nov 15, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Expanded docstring as suggested in review
  • Loading branch information
@oleksandr-pavlyk
oleksandr-pavlyk committed Nov 15, 2024
commit 77520787ec005645a41baa66c1fc495db1b970ab
53 changes: 41 additions & 12 deletions dpctl/_sycl_timer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -84,8 +84,7 @@ def get_event(self):

class SyclTimer:
"""
Context to measure device time and host wall-time of execution
of commands submitted to :class:`dpctl.SyclQueue`.
Context to time execution of tasks submitted to :class:`dpctl.SyclQueue`.

:Example:
.. code-block:: python
Expand All @@ -99,13 +98,18 @@ class SyclTimer:
milliseconds_sc = 1e3
timer = dpctl.SyclTimer(time_scale = milliseconds_sc)

untimed_code_block_1
# use the timer
with timer(queue=q):
code_block1
timed_code_block1

untimed_code_block_2

# use the timer
with timer(queue=q):
code_block2
timed_code_block2

untimed_code_block_3

# retrieve elapsed times in milliseconds
wall_dt, device_dt = timer.dt
Expand All @@ -116,16 +120,41 @@ class SyclTimer:
associated with these submissions to perform the timing. Thus
:class:`dpctl.SyclTimer` requires the queue with ``"enable_profiling"``
property. In order to be able to collect the profiling information,
the ``dt`` property ensures that both submitted barriers complete their
execution and thus effectively synchronizes the queue.

`device_timer` keyword argument controls the type of tasks submitted.
With `device_timer="queue_barrier"`, queue barrier tasks are used. With
`device_timer="order_manager"`, a single empty body task is inserted
instead relying on order manager (used by `dpctl.tensor` operations) to
the ``dt`` property ensures that both tasks submitted by the timer
complete their execution and thus effectively synchronizes the queue.

Execution of the above example results in the following task graph,
where each group of tasks is ordered after the one preceding it,
``[tasks_of_untimed_block1]``, ``[timer_fence_start_task]``,
``[tasks_of_timed_block1]``, ``[timer_fence_finish_task]``,
``[tasks_of_untimed_block2]``, ``[timer_fence_start_task]``,
``[tasks_of_timed_block2]``, ``[timer_fence_finish_task]``,
``[tasks_of_untimed_block3]``.

``device_timer`` keyword argument controls the type of tasks submitted.
With ``"queue_barrier"`` value, queue barrier tasks are used. With
``"order_manager"`` value, a single empty body task is inserted
and order manager (used by all `dpctl.tensor` operations) is used to
order these tasks so that they fence operations performed within
timer's context.

Timing offloading operations that do not use the order manager with
the timer that uses ``"order_manager"`` as ``device_timer`` value
will be misleading becaused the tasks submitted by the timer will not
be ordered with respect to tasks we intend to time.

Note, that host timer effectively measures the time of task
submissions. To measure host timer wall-time that includes execution
of submitted tasks, make sure to include synchronization point in
the timed block.

:Example:
.. code-block:: python

with timer(q):
timed_block
q.wait()

Args:
host_timer (callable, optional):
A callable such that host_timer() returns current
Expand All @@ -134,7 +163,7 @@ class SyclTimer:
device_timer (Literal["queue_barrier", "order_manager"], optional):
Device timing method. Default: "queue_barrier".
time_scale (Union[int, float], optional):
Ratio of the unit of time of interest and one second.
Ratio of one second and the unit of time-scale of interest.
Default: ``1``.
"""

Expand Down