These charts follow the 20 paths that every release's comparison has measured
since 1.12.0, as time relative to 1.12.0. Lower is faster. Point at a chart,
or tab to it, to read each release's value. The table under each chart lists every
measured value.
How it is measured
Each release's notes compare it with the release before it on the
development laptop, a Lenovo Legion Pro 7 16AFR10H, with diagnostic logging off and both builds
measured in the same session. A chart multiplies those comparisons from 1.12.0 on, so the 2.1.0 point is
2.0.0's ratio to 1.12.0, times 2.0.1's ratio to 2.0.0, times 2.0.2's ratio to 2.0.1, times 2.1.0's ratio to 2.0.3. Each ratio compares medians.
1.12.0 to 2.0.0: 30 runs per build, the two builds one after the other for each path.
2.0.0 to 2.0.1: 30 runs per build in each of two passes, with the order of the builds flipped between passes, three passes for start, create and Dispose, and six for the per-call rows.
2.0.1 to 2.0.2: 30 runs per build in each of two passes, with the order of the builds flipped between passes, three passes for start, create and Dispose, and eight rounds alternating the builds for the per-call rows.
2.0.3 to 2.1.0: six rounds alternating the two builds in one session, each round installing each build's driver once for the device paths. 2.0.3 changed only the full removal's usbip-win2 step, which no charted path runs, so this step continues from the 2.0.2 point.
Values from different sessions never meet in one ratio. Between two
sessions, the same release has measured up to 31 percent apart on a per-call SubmitState
row and 86 percent apart on a few-millisecond USB/IP row. Within one session, rows measured in two
passes have moved several percent from one pass to the next, so a step of a few percent is within
that spread.
From 2.1.0 on, a release's notes read a path faster only when every
round of a session read it faster and either a second session agreed or a change had named the
path as its target before the measurement. They read it slower only when a second session agreed.
A chart moves only where the notes read the median faster or slower. Any other step is plotted
flat, and the table under the chart keeps its measured values.
Start and create
Process start to the first controller, DualShock 4