← Blog
✎blogbericht

Fixing a sky in the hospital render cost 48 percent more time and 29 percent more memory

A tall window in a pale corridor wall reflecting a flat dark teal sky panel, with a brass frame and clean sill under a warm amber light, in teal and steel, the AMARBARO mark and wordmark centred over it.

After the fix, the hospital ward frame no longer showed street sky, and the correct render cost 48 percent more time and 29 percent more memory than the broken one.

A replay of episode one showed sky through a hospital window. The fix removed the sky, and the render got heavier. The cost went from 8.42 seconds and about 980 MB of memory to 12.48 seconds and about 1.2 GB. The correct render costs 48 percent more time and 29 percent more memory.

We keep that number in the post because the earlier render was cheap for the wrong reason.

The cheap render was missing rooms

The replay dropped the hospital's interior rooms and the height offset that places the scene. It also followed a clip ID from before a resume, a stale identifier that pointed at the wrong clip. So the ward was drawn with no interior, and the street sky showed where the walls should have been. On our sky measure, the ward frame scored 90.7 percent.

The fix had four parts

The fix seeded the interior rooms, kept the height offset and recovered the player's old height, and picked the clip ID at clip time instead of reading it early. We also constrained the camera. A movie comparison passed, and a person inspected the ward frame. The sky measure fell from 90.7 percent to 0 percent.

What the cost means

The correct render adds 4.06 seconds and 290,036 KB, about 283 MB. Our record does not split that between the rooms and the other changes in the fix.

Cost is a poor reason to keep a render, and a good reason to check the frame before the cost. We checked the frame first.

The other checks in the same pass

The same pass checked the rest of the output. The overlays used the actual ink bounds, with safe margins of 5 and 8 percent. A 28-card layout check passed with frames inspected, and so did a two-clip cut in long and short formats. An animation check ran 60 moving and idle frames and passed, with an arm travel of 0.589 radians.

The cost pair in this post is for the sky fix. Our record gives no cost figures for the other checks, so we do not attribute the cost to them.

A run that stalled, and what we accepted

The last repeat of the movie stalled after the replay finished and timed out. We did not get a clean repeat. We accepted the earlier movie and ward frame that had passed, and the quit hang and a port race moved to a later task. Those are open problems. A passing run from before is weaker evidence than a clean final run, and the post should say so. It is not a result for the final repeat, which never finished. Our record lists the commits for the fix, and those are the only place to see exactly which change was measured.

How we judged the frame

We judged the ward frame in two ways: a movie comparison that passed, and a person looking at the frame itself. The sky measure is a number, and a number can pass while the picture is still wrong. The frame is what a viewer sees, so the frame judgement was the one that counted for the sign-off.

The cost is the other half of the same result. A report that showed only the cheaper render would look like a speedup. In what the frame showed, it was a regression.

What we still do not know

  • How many runs stood behind the 8.42 and 12.48 second figures, and what load the machine carried. Our record gives one pair without either detail.
  • Whether the memory gap holds on another clip. We measured one ward.
  • The frame inspection report that the record points to was not on disk when we checked. The frame claim rests on the record's own description.

How to prove this wrong

Run the replay with the fix off and on, on the same clip, on an idle machine, three runs each. Compare medians of time and memory. If the gap falls well below 48 percent, the original number came from load on the machine, and the post should say so. Repeat on a second clip before any claim about the ward generalises.

Provenance

  • ~/Brain/street-sim-render/2026-11-07-render-defects.md: the hospital fix, the sky measure, the render cost pair, the stalled final run, and the commits.
  • ~/Brain/street-sim-render/verdicts.md: the replay question, the stale ID and dropped rooms, and the checks that answered it.

Reacties

Nog geen reacties.

Inloggen om te reageren.