WEBVTT

1
00:00:00.000 --> 00:00:08.167
This overview explains how workload, host, output, lifecycle,
and timed stages define what a measurement answers.

2
00:00:08.167 --> 00:00:14.560
Results that look contradictory can all be correct
because they describe different tasks or boundaries.

3
00:00:14.560 --> 00:00:14.567
Results that look contradictory can all be correct
because they describe different tasks or boundaries.

4
00:00:14.867 --> 00:00:20.967
A useful comparison card identifies the input family and support, the required artifact,

5
00:00:20.967 --> 00:00:26.700
the lifetime of prepared objects, and the timed boundary on a specified host.

6
00:00:26.700 --> 00:00:32.300
These details determine which requests the evidence actually covers.

7
00:00:32.300 --> 00:00:37.733
Suppose the candidate kernel costs two units and the control kernel costs three.

8
00:00:37.733 --> 00:00:42.667
Within this explicitly limited kernel boundary, the candidate is cheaper.

9
00:00:42.667 --> 00:00:46.167
We have not yet counted how either kernel is prepared or how its

10
00:00:46.167 --> 00:00:49.133
answer reaches the caller.

11
00:00:49.133 --> 00:00:54.800
Add five units of candidate overhead and one unit of control overhead.

12
00:00:54.800 --> 00:00:58.033
Complete totals become seven and four.

13
00:00:58.033 --> 00:01:04.333
The kernel win and the request loss coexist because they refer to different scopes.

14
00:01:04.333 --> 00:01:09.833
Report both without silently moving the claim between them.

15
00:01:09.833 --> 00:01:09.837
Scope identifies inputs, host, versions, output and lifecycle.

16
00:01:09.837 --> 00:01:15.567
Scope identifies inputs, host, versions, output and lifecycle.

17
00:01:15.567 --> 00:01:19.300
Boundary identifies the stages timed.

18
00:01:19.300 --> 00:01:27.367
Before pooling values, ask whether both studies estimate
the same quantity and use a compatible sampling design.

19
00:01:27.367 --> 00:01:33.633
The corrected B2/B4 and accepted EPFL
results occupy separate workload rows.

20
00:01:33.633 --> 00:01:38.333
Their compiled-kernel ratios answer different scoped questions.

21
00:01:38.333 --> 00:01:44.300
Preparation and public-wrapper results occupy
separate columns from the kernel measurements.

22
00:01:44.300 --> 00:01:47.900
A cell's location now tells us what it measured.

23
00:01:47.900 --> 00:01:55.233
Adding a scalar-count result creates a different output
category, not another sample of complete-vector execution.

24
00:01:55.233 --> 00:02:02.567
The retained program includes kernel gains, parity, slower
wrappers and task-specific counting improvements.

25
00:02:02.567 --> 00:02:06.933
They can coexist because they answer different scoped questions.

26
00:02:06.933 --> 00:02:15.131
Cold, warm and reload observations must retain their lifecycle
labels; a warm count is not a cold full-vector comparison.

27
00:02:15.131 --> 00:02:15.133
Cold, warm and reload observations must retain their lifecycle
labels; a warm count is not a cold full-vector comparison.

28
00:02:15.433 --> 00:02:20.233
Imagine evidence for one family producing a fresh full vector.

29
00:02:20.233 --> 00:02:23.633
That cell in the comparison table is supported.

30
00:02:23.633 --> 00:02:30.400
It does not automatically fill a neighboring cell for point queries or resident queries.

31
00:02:30.400 --> 00:02:34.733
Changing any of those dimensions may change the work and the winner.

32
00:02:34.733 --> 00:02:38.933
An unmeasured cell should remain unmeasured rather than inheriting

33
00:02:38.933 --> 00:02:41.533
the nearest favorable result.

34
00:02:41.533 --> 00:02:46.167
Lack of evidence is not a measured loss either.

35
00:02:46.167 --> 00:02:50.933
A defensible conclusion repeats the scope that was actually tested.

36
00:02:50.933 --> 00:02:56.633
Broader application needs a matching experiment or proof, depending on the claim.

37
00:02:56.633 --> 00:03:03.685
Keeping that boundary visible makes the result more useful to someone choosing a method.

38
00:03:03.685 --> 00:03:03.700
Keeping that boundary visible makes the result more useful to someone choosing a method.

39
00:03:03.700 --> 00:03:07.433
Keep each result's scope and lifecycle attached.

40
00:03:07.433 --> 00:03:13.100
Combine estimates only under a justified
common target and sampling design.