Skip to content

docs(roadmap): the upscaler shipped, and what it measured - #875

Merged
lobinuxsoft merged 2 commits into
developmentfrom
docs/upscaler-result
Aug 18, 2026
Merged

docs(roadmap): the upscaler shipped, and what it measured#875
lobinuxsoft merged 2 commits into
developmentfrom
docs/upscaler-result

Conversation

@lobinuxsoft

Copy link
Copy Markdown
Owner

Records the result of #874 in the roadmap: phase 1 of #481 built, and the handheld budget closing for the first time.

Three things worth having written down beyond the numbers:

  • Everything that costs per pixel fell 3.7x–4.5x; everything that does not depend on resolution stayed identical. That is the independent check that the size sweep was complete — a miss would show as something that did not fall, or something that fell and should not have.
  • TAAU's fetch count, which is why it lost: 19 per output pixel against SGSR 2's 7, because SGSR 2 reuses the gather's taps for its variance box and reads history with one bilinear tap. It is what turns transliterate rather than invent from an opinion into a measurement.
  • 🎯 The rule the size sweep exposed: what DERIVED its size from a texture was already correct, what RECEIVED it as a parameter was wrong. The froxel grid was found by the owner from the picture, not by the suite.

Also marks what is left of #481 — RCAS and feature locking, FSR 3.1 as the second backend, and MipBias, noted since the resolve was written and still not done.

Docs only.

lobinuxsoft and others added 2 commits August 18, 2026 04:59
Phase 1 of #481 is built and the handheld budget closes for the first
time: 40.7 ms to 13.9, GPU 24.8 to ~8.4, 72 fps at 76 % occupancy where
the device used to sit at 93-99 % and power-limited to 1141 MHz.

Records three things the plan could not have known:

- Everything that costs per pixel fell 3.7x to 4.5x and everything
  that does not depend on resolution stayed identical. That is the
  independent check that the size sweep was complete.
- TAAU, built and removed, with the fetch count that explains it: 19
  per output pixel against SGSR 2's 7, because SGSR 2 reuses the
  gather's taps for its variance box. It is what turns "transliterate
  rather than invent" from an opinion into a measurement.
- The rule the size sweep exposed: what DERIVED its size from a
  texture was already right, what RECEIVED it as a parameter was
  wrong. The froxel grid was found by the owner from the picture.

And what is left: RCAS and feature locking, FSR 3.1 as the second
backend, and MipBias, which has been noted since the resolve was
written and is still not done.
@lobinuxsoft
lobinuxsoft merged commit 5bae440 into development Aug 18, 2026
@lobinuxsoft
lobinuxsoft deleted the docs/upscaler-result branch August 18, 2026 22:58
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant