[Atomsk-debug]

Before: Etna always shaded every screen pixel. Its temporal anti-aliasing and the light shafts' 2x2 order were ported from Photon Shaders, and the jar carried the Photon Shaders License, which blocks Modrinth and CurseForge. Mobs, chests and the hand kept vanilla's flat lighting, so they didn't match the shaded world around them.

After: a new World Resolution setting (50 to 100%) draws the world at fewer pixels, and Etna's own anti-aliasing builds the full-size picture back up. Auto Resolution (off by default) picks that size by itself to hold a chosen frame rate. The anti-aliasing is new, and it is steadier and sharper than the old one even at full size. Mobs, animals, chests, signs, item frames and the first-person hand now take Etna's sun, sky and block light and sit in its shadows; this is the new Entity Lighting toggle on the Lighting page. No code from other shader packs is left, so LICENSE-PHOTON.txt is gone and the licence is plain MIT.

This branch includes PR #3 (0.8.1, merged in up to 10f1337) because the licence cleanup sits on top of that branch's own shadow warp and Fresnel. Merge #3 first; this PR's diff then shrinks to its own changes. Not tested in game yet, so don't merge until Cosimo has tested it: a merge publishes a release.

How

• Render scale. While the level is drawn, GameRenderer.mainRenderTarget() hands out a smaller target of Etna's (EtnaScale), so the whole level is drawn into it.

  • The scaled world chain (etna_world_scaled) runs on that target.
  • Its anti-aliasing history stays screen-sized, and its last pass writes Minecraft's own picture.
  • The hand, the final chain and the HUD stay at full size. The final chain reads the level's depth through EtnaScale.LevelDepth.
  • Main's depth is cleared on scaled frames, so the bloom's hand mask sees only the hand.
  • Texture mip selection follows the screen's pixels.

• Auto Resolution (EtnaAutoScale). It picks the scale from fixed steps (100% down to 50%) by the GPU's own time per frame, up to the end of Etna's final pass, timed on one frame in eight. Using GPU time means VSync and frame limiters don't hide spare room.

  • Two timed frames over budget step down; one far over budget steps down two at once.
  • It steps up only after two seconds in which every timed frame would still fit with headroom at the bigger size. Only the level and world passes are counted as growing.
  • It never goes below World Resolution (70% when that is at 100%).
  • It is off during the benchmark and without GPU timestamps.
  • Each step change restarts the TAA history.

• Anti-aliasing (taa.fsh), clean-room from the cited papers.

  • Each frame is rebuilt from 3x3 level pixels with Gaussian weights that account for the jitter. The history's weight is kept in alpha.
  • Reprojection uses the nearest depth.
  • The history is read with Catmull-Rom from five bilinear reads.
  • The history is clipped to a YCoCg variance box, along the line to this frame's colour.
  • Frames are blended with 1 / (1 + luma) weights, so bright glints don't linger.

• Jitter. 8 Halton steps at full size; at smaller sizes, 8 / scale² rounded up to a power of two. The terrain now reads the current jitter from its block instead of keeping its own copy of the table.

• Entity Lighting.

  • A mixin on the render type's draw swaps vanilla entity and item pipelines for Etna copies: etna:lit/... while the level is drawn, etna:hand/... for the first-person hand. It also binds Etna's light block and shadow maps next to the default uniforms.
  • The copies compile two per frame as the level starts, so a new mob or the hand never stalls a frame. After that, each draw is one identity-map lookup with no allocation.
  • GUI drawing and the screen effects (the totem) are never swapped, and Etna off or the toggle off draws exactly vanilla.
  • The lighting functions moved out of terrain.fsh into include/etna_light.glsl, shared by terrain, entities and items, so a mob in shade darkens like the grass next to it.
  • Order-independent translucent pipelines, and fog for entities under water, stay vanilla for now.

• Benchmark. Two new steps: "shadow map fade" holds a fade at its middle using PR Etna 0.8.1: solid water reflections, thin shoreline foam, shadows that soften with distance #3's hook, and "entity lighting off". That makes 23 steps, about 150 seconds.

• Post target sizes. Fixed sizes are now written as REFERENCE_SIZED + size. Before, a size that happened to equal the chain's own size was taken for "no size", which made the half-size passes four times too big at 1280x720 and 75%.

• Robustness from review.

  • The frame state resets at render HEAD and renderLevel RETURN.
  • The smaller target uses main's own formats and is freed without a level.
  • A failed scaled pass clears the screen instead of compositing over last frame's picture.
  • The TAA's centre pixel is clamped at the screen edges.
  • validate_shaders.py rejects option-like arguments.

• Licence cleanup. Removed:

  • LICENSE-PHOTON.txt and its jar entry;
  • the Photon rows in CREDITS and the Modrinth note;
  • the LICENSE footer;
  • the Photon licence in fabric.mod.json.

  The techniques are credited in CREDITS.md. Version 0.9.0.

Offline TAA test bench (64 frames against a 16x supersampled reference), old vs new at full size:

Measure	Old	New
Still error	0.0138	0.0101
Flicker	0.0081	0.0053
Edge sharpness (of the reference)	0.69	0.82
Ghosting	0.057	0.054

At 75% it is still steadier than the old one at full size; in motion it is slightly softer.

🤖 Generated with Claude Code

https://claude.ai/code/session_01F7MajtknSApr1T68EGCPtx