Praxis

Bench history

How fast main renders each Reference scene, record by record.

Machine Scene Tagged records Experiment

Loading the Bench history…

How to read this page

Bench History Explained covers what follows: the source hash, Harness versions, quiet records and Suspects.
Each point is one Bench record of main's source, from the bench-history branch (ADR 0024). Points are spaced by commit, not by date.
A Baseline record, imported from main's history. A record benched since. A tagged record, kept off the Series (a busy machine, another configuration), shown with the Tagged records switch. A tagged record off the chart's scale. An Experiment's record, benched from a branch with praxis bench --experiment <name>, drawn beside the Series record of its merge base's source and compared with it.
A suspect: a CPU or GPU frame median outside the median of up to 8 records before it on its line, ± the larger of 4 median absolute deviations and 3%. Red when slower, green when faster. Only a step's first record is marked, and an imported Baseline starts a new line, so its step is none. lavapipe's times are not judged.
Shaded bands are Harness versions. A metric's definition can change from one to the next, so lines break between bands and no change is measured across them. A scene whose content changed breaks its line the same way.
Timing changes are coloured by praxis compare's warn and fail ratios and noise floors: amber past +5%, red past +10% when also past the noise floor (10 µs on the CPU, 20 µs on the GPU), green past −5%; the rest stay grey as noise. Counter changes are exact, so every one is marked. Changes across a Harness version or scene content break stay grey.
Budgets are ADR 0013's, judged on the M4 Pro at 2560×1440 at the newest Harness version, on the median as praxis compare judges them. Other machines are comparison adapters: their figures are not Budgets. A dashed line is a machine with tagged records only.

Where main stands

Each scene's latest record on the machine above. Pick a row to see its scene below.

Scene

Frame

CPU and GPU frame medians on the machine above, with the Budget where the M4 Pro is picked. Across machines, below, sets them side by side.

Passes

Each GPU pass's median on the machine above, one line per pass.

Counters and memory

What the machine above counted. Measured frames are the 1,800 frames a reference run times, the load ADR 0013's Budgets judge. Counter frames are frames 8 to 71, early in the warm-up, which every configuration renders, lavapipe's included. Aliasing is measured on its own captures. Counters do not depend on load, so every machine's are exact.

Record

Pick a point in any chart to see its record.

Across machines

Each scene's latest frame medians on every machine with a GPU, with the Harness version that measured them; lavapipe, CI's software renderer, times nothing. No Budget is drawn here. The CPU figures time each machine's host CPU: the RTX 2070 SUPER's time the Windows machine's CPU, not the GPU they are named after.

Release build on CI's lavapipe runner

The release binary CI builds from each commit of main on the runner that benches lavapipe, whatever machine or scene is picked. Its build time counts only what CI's cache did not restore, so a jump may be a cache miss rather than slower code.