[Atomsk-debug]

Before: in 0.8.0, the shore, trees and hills mirrored in lakes and rivers break up into hundreds of scattered dark dots instead of keeping their shapes. Shoreline foam is a thick, glowing white band. Shallow water turns milky, and thin sheets of water flowing down a bank are covered in blotches. Shadows keep the same sharp edge however far they reach from what casts them, so long low-sun shadows look like hard black strips. Past the shadow casters' range, they stop at a hard line. In the first 0.8.1 test build, reflections and far shadows shimmered, and a shadow falling on water showed twice, the two copies cancelling each other. While walking, the dappled shade under trees and the shadow edges further out changed on almost every frame, and far tree shade twinkled even standing still.

After: reflections are solid shapes that wobble gently with the waves and stay steady from frame to frame. Foam is a thin, faint, bubbly line where still water meets the shore. Shallow water stays clear, and water flowing down a slope has no foam. Shadows are sharp where they touch their caster and blur as they stretch away from it, as in real life, without shimmering. At sunrise and sunset they are lighter. At the edge of the casters' range they fade out instead of ending. A shadow on water is one shadow: on clear shallow water it is the bottom's, on deep or murky water a soft shade inside the water. While walking, the shade under trees holds still and only changes slightly about twice a second, and the far end of the shadows stays at the same distance. Standing still, far tree shade is completely still. The shadow map warp and the water's Fresnel are Etna's own code now, not Photon's.

How:

• post/water_ssr.fsh:
  • 24 fixed steps plus 6 halvings, with no per-pixel noise and one fixed pixel per 2x2 block.
  • A crossing counts as a hit only when the surface is no nearer than the ray was just before it. A ray that passes behind something in front of it doesn't count.
  • Misses are written with their distance and no colour change.
  • The ray follows a quarter of the swell's slope. The sky reflection that gets removed keeps the full swell.
  • Each ray starts from the point the jittered pixel shows, worked out without the jitter, and reads depth and colour where the jittered frame holds each point it wants.
  • With anti-aliasing on, each texel blends with its own history (30 % of the new frame, more while the camera moves). The history is reprojected through the water point's position and PrevViewProj. A history texel is only used where last frame held water at the same distance, allowing for how fast that distance changes across a few pixels on grazing far water.
• tools/gen_post_effects.py, etna_world.json, etna_world_plain.json, PostChainMixin: two persistent half-size targets, water_ssr0/1, hold that history and are swapped together with the TAA history. The plain chain binds a placeholder it never reads.
• include/water_reflection.glsl: the upsampler uses the same 16 texels every frame, and misses count. In 0.8.0 it skipped them, so one lone hit grew into a speck.
• include/terrain_water.glsl:
  • The foam over shallow floors is gone. The line is narrower and fainter, broken by fine bubbles, and pale instead of bright white. It only appears on flat water.
  • fresnelDielectric is rewritten in the exact Cook-Torrance form.
  • New waterBodyShadow: the water's sun share is the shadow at four points down the refracted view, placed at equal shares of the light scattered back above the bottom (green extinction down along the sun and back up along the view). Each point is looked up where its way to the sun crosses the surface, so only what stands above the water shades it, and a bottom rising ahead of the view doesn't. This replaces the surface's own sharp shadow, which crossed out the bottom's shifted one.
  • The thickness search has a fourth ring at 8 blocks. Where the shadow map shows no bottom at all, the depth model assumes 4 blocks of water instead of falling back to the plain, opaque look.
• core/terrain.fsh: new etnaWaterLevel flag, taken from the face normal before it is snapped to an axis. Water faces take waterBodyShadow.
• include/shadow_distortion.glsl: Etna's own warp, factor = 0.85 n + 0.15 with n = sqrt(1 - (1 - x^2)(1 - y^2)). It fills the whole map and keeps far texels two to five times finer towards the corners.
• core/terrain.vsh: in the shadow pass, leaves and plants sway only near the camera (fading out from 20 to 44 blocks). Further out a map texel is wider than the sway, which only flipped leaf texels between lit and shaded every frame. The leaves still sway in the main pass, by less than their shadow lookup's offset and bias.
• include/terrain_shadow.glsl:
  • Contact hardening, the PCSS idea reimplemented, now deterministic. The area-weighted box filter also reports how much it found blocked and how far the casters are. Six fixed bilinear reads search the ring beyond it.
  • The blur radius is the softness plus 0.025 x the casters' average distance along the light, capped at 1.2 blocks or 16 texels. Where everything read is blocked or everything is open, the box result is final. Otherwise it blends into a 16-point bilinear disc. Only discs wider than about 8 texels are turned, by more the wider they are, so near shadows never change from frame to frame.
  • New shadowSoftAt: a bilinear 2x2 test of any point, used by the water.
  • Shadows get up to 35% lighter when the sun is below about 20 degrees.
  • Receivers beyond the casters' chunk box fade to lit.
  • Points beyond the map's far depth plane count as lit. This also applies in the unfiltered test, leaves, and post/shafts.fsh.
  • New etnaMapOut: how far out a point lies, counted from the camera (shadowMat[3].xy) over EtnaShadowInfo.w, the share of the radius the map always covers round the camera. Every reader (filter, single and bilinear tests, leaf light, water, post/shafts.fsh) fades and stops by it, so where the shadows end doesn't move when the map's middle steps. They end about 3.5 blocks short of the shadow distance (the hold plus 2 blocks of slack). A point off the map itself also counts as past the edge (a map read a frame late behind a fast-flying camera).
• EtnaShadows.java:
  • Caster culling allows for the blur's reach, both sideways and away from the light, in the view test's margins and the receiver floor, and its bounds follow the new warp.
  • New benchmarkFade, off by default: the benchmark can hold a sun-step fade at its middle, to measure the cost of reading both maps while the sun moves.
  • The map's middle no longer moves a texel at a time (most frames while walking). The warp is centred on the middle, so every move re-samples every texel at new points of the world: leaf holes and far edges came out different each time. followMiddle keeps each map's middle fixed in the world and steps it by 2 blocks (whole texels) once the camera is 1.5 blocks from it on an axis. Near the camera texels are at most 7% wider. The caster tests count from the middle (the square, the caster floor) or widen by the camera's lead (the columns, the view margins). EtnaShadowInfo.w (unread since the shafts became a post pass) now carries the reach.
• CREDITS and README list the warp and the Fresnel as Etna's own. Version 0.8.1, CHANGELOG updated.

The dot analysis was cross-checked against the look thread's own patch: the calmer ray normal and the steady 16-texel upsample come from it.

Testing so far:

• Reflections and shadows were checked offline with headless GL renders and a model of Etna's TAA. After anti-aliasing, shadow flicker is 35-65 % below the first test build, and below 0.8.0 with a low sun.
• Caster culling was checked at FOV 70, 110 and 126: none of 80,000 casters was wrongly dropped.
• The map's middle steps were checked offline on a swaying canopy of leaf blocks with holes, under the area-weighted filter, walking at 4.3 blocks a second: the mean change from one frame to the next drops from 0.05-0.09 to 0.004-0.008 of full shade, 16 to 150 blocks away, and the frame the map steps changes no more than any frame did before. With the shadow sway faded out, a still camera sees no change at all from 32 to 150 blocks (0.003-0.004 before), and walking about 40% less.
• A second review read the water-shadow and reflection-history code, and its two findings are fixed.
• All shader variants pass tools/validate_shaders.py.
• The jar builds on a Mac mini M4, where the game's post-chain format accepted the persistent, sized history targets.

The latest build has not been tested in game yet.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Xzo7zbTpyU7a3y3XQNbw1n

---

Generated by Claude Code