wgpu refuses to create a buffer with INDIRECT usage on a device without
DownlevelFlags::INDIRECT_EXECUTION (OpenGL before 4.3 without compute shaders), so the game
panicked creating 'Section draws'. Multi-draw now also needs indirect execution, and without it
the buffer has no INDIRECT usage: sections are drawn one direct draw each, as they already were
without INDIRECT_FIRST_INSTANCE.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SpriteAnimator wrote every changed sprite with a compute pass, which OpenGL 3.3 to 4.2 has no
equivalent for in wgpu: creating the pipeline failed. Without compute shaders and storage, the same
jobs now run on the CPU with animate.wgsl's integer mix, and each level is copied into its layer
through the upload belt. The GPU test runs both paths against the same CPU reference.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
wgpu's GL backend picks a texture's target from its layer count when it creates it: one layer makes
a 2D texture and a multiple of six on a square texture a cube map, and neither can be bound as
terrain.wgsl's texture_2d_array. On GL, the 1-layer stand-in for unused sprite arrays gets two
layers and a sprite array of 1 or 6n layers one empty layer more. Other backends are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
GLSL combines a texture with its sampler, and naga refuses an entry point that reads one texture
through two (ImageMultipleSamplers): the BLUR override left both branches in fs_screen and
fs_vignette, so creating the screen effect pipelines failed on wgpu's GL backend at startup. Blurred
sprites now have their own entry points (fs_screen_blur, fs_vignette_blur); each pipeline still
samples once per pixel.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
wgpu's GL backend (WGL on Windows, EGL on Linux). Auto tries Vulkan, then DirectX 12 on Windows,
then OpenGL; graphicsApi:opengl (or HORIZON_GRAPHICS_API=opengl) picks it. A Radeon HD 7400M
(TeraScale 2: DirectX 11, OpenGL 4.x) could not start Horizon at all.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: thank you @rdbtCVS! Merged after review (a surface failure now falls through to the next backend, as adapter and device failures do) and this head's verify. Follow-up: an OpenGL fallback for GPUs with neither Vulkan nor DirectX 12 builds on this.
Keep surface creation errors in the backend failure list so Windows Auto can try DirectX 12 after Vulkan fails.
Co-Authored-By: OpenAI Codex <noreply@openai.com>
- FlyTowardsPositionParticle puts vault_connection on the translucent
layer: its LifetimeAlpha(0, 0.6, 0.25, 1) is not opaque.
- The shadow draw started at world_quads - shadow_quads; with the new
bucket after it, that drew translucent particles with the shadow
pipeline and dropped as many shadows.
- One atlas lookup per named-sprite particle and frame: the engine keeps
each sprite's layer beside its rect instead of hashing the name twice.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Quad particles all went into the entity batch, drawn before translucent
terrain with depth writes on: a stained-glass crumb wrote its depth and
the sorted translucent pass then culled the glass behind it, so breaking
glass in front of glass cut holes through the wall.
Vanilla 26.3 submits quad particles to two phases: opaque layers draw
with the solids and translucent layers draw in an "afterTerrain" phase
right after the translucent section pass (the translucent particle
pipeline still writes depth, so ordering is what matters). Mirror that:
HudAtlas records which sprites have a 0 < alpha < 255 texel
(`SpriteContents.transparency().hasTranslucent`), particles classify
their layer the way `SingleQuadParticle.getLayer` does — the type's
hardcoded layer, or `Layer.bySprite` for terrain, item and block marker
sprites — and the entity batch routes those quads to a sixth bucket the
renderer draws immediately after translucent terrain.
Fixes#1679.
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Packs sample blocks from one 2D atlas; Horizon keeps them in sprite
arrays. While a pack is loaded, every layer and its mip levels are
copied into a grid, and the animator reports the layers it rewrote so
only those are copied again. The sprite arrays gain COPY_SRC.
The gbuffers pass binds it in the next slice.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The first slice that draws: with HORIZON_SHADER_PACK=<name>, the pack in
shaderpacks/<name> (folder or zip) runs its prepare, deferred, composite
and final programs over the world, under the hand and the HUD. Gbuffers
programs come later, so the composites read Horizon's own picture.
- The world renders into its own target (as with a resolution scale); its
colour becomes colortex0 and its reverse-Z depth becomes depthtex0 in
GL's window depth (1 - d for the same planes), rows flipped to GL's
order. depthtex1 and depthtex2 read the same depth for now.
- colortex buffers in the formats the pack asked for (the nearest one
wgpu renders to), created only when a program uses them, each with two
textures: programs read main and write alt, the buffers a program wrote
flip after it, flip.* keys and deferred_pre/composite_pre apply, and
the mapping carries over to the next frame. Clears at the start of the
frame with the pack's colours or the defaults.
- noisetex (256x256 white noise, repeat), a lit stand-in behind shadow
samplers until shadow maps exist, a 1x1 stand-in for other names; the
frame uniforms of the previous commit and the translator's own block.
- final draws onto the window; without it colortex0 is shown.
- Pipelines are built through validation scopes: a program wgpu refuses
is reported with the full error and skipped. The dimension's folder is
reloaded when the player changes dimension.
Nothing changes without the environment variable: players see no change.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The values a pack's post-processing programs read once per frame, in
GL's conventions: camera position and the previous one, eye altitude,
up/sun/moon/shadow-light positions in view space (the sky's own frame for
the sun and moon), sun and shadow angles, world time, day and moon phase,
frame counter, frame time and run time, view size and aspect, near and
far, rain, smoothed wetness, thunder, fog colour, start and end, sky
colour, eye in fluid, blindness, night vision, darkness, brightness, GUI
hidden, eye brightness and its smoothed value, and the model-view and
projection matrices with their inverses and last frame's. The projection
is GL's perspective with the world's near and far planes times vanilla's
view bobbing. Each value is written as the type the pack declared.
The game passes what only it knows (day time, weather levels, the fluid
at the camera, light at the camera, blindness, F1, brightness) in the
renderer's Frame when HORIZON_SHADER_PACK is set.
Fix (slice 4): programs that draw to buffers read gl_FragCoord through a
variable only final set; every target sets it now, with a test. Found by
the first frame drawn with a pack.
Nothing reads the uniforms yet: players see no change.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Both switched on past 4,096 section draws, measured when a section had
2.2 to 2.5 draws. A section's draw count now depends on the sides it shows
the eye, so the bars count sections in view: 1,600, where 4,096 draws fell.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The clientbound boss_event packet (add/update/remove/style/progress), a
BossOverlay that renders each bar, darkens the world under a wither, and
suppresses the create-the-world fog — vanilla's BossHealthOverlay and
GameRenderer behaviour on a real server.
Co-Authored-By: Devin <devin@cognition.ai>
Meshing sorts each section's and distant tile's solid quads into seven
buckets: one per axis direction (all four corners in one axis plane, both
triangles wound the same way) and one for everything else, and keeps each
axis bucket's extreme plane. A quad facing +X at x = p is a back face from
an eye at x <= p, which the solid pipeline culls anyway, so the frame draws
only the buckets that can face the eye, merged into runs of adjacent
buckets: one to three indirect draws per section on the existing page-run,
occlusion-culling and flat paths. The eye comes from the frame's
view-projection, so view bobbing is exact. The picture does not change;
`horizon-ctl section-sides on|off` switches it for A/B checks.
Written from scratch from the general back-face technique; no other
renderer's code or Horizon's removed code read.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#503 grouped each section's solid quads by the axis side they face and
skipped the groups facing away from the eye. Its code and PR named an
internal class of another project as the model. Horizon doesn't build on
another project's code, so the grouping goes and every section draws its
solid quads in one draw, as vanilla does; the GPU's back-face culling
discards the back faces, so the picture is unchanged.
This costs frame time, measured on the Mac mini in the PR. A clean-room
version can come back later, designed from vanilla and from first
principles only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A developer of Iris showed that Horizon's shader pack code was written from
Iris's source. This removes it: the horizon-shaderpack crate, Java pack
loading and its render stages, Horizon's own WGSL packs and the example
packs, the Shader Packs screen and its option, the launcher's shader pack
browsing and installing, and the docs about them. Shader packs may come
back later as a clean-room rewrite.
The vanilla renderer and its post-processing (resolution scale, FXAA) are
unchanged. Options files that name a pack load and drop the key; the
launcher reads manifests that list a shader pack, and files in
`shaderpacks/` are never deleted.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Vanilla 26.3 multiplies bobHurt/bobView into the projection
(GameRenderer.renderLevel); Iris's MixinModelViewBobbing moves them into
the model-view when a pack is in use. The comment said vanilla keeps the
bob in the PoseStack, which is Iris's behaviour, not vanilla's.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A pack rebuilds view positions from the depth buffer with
`inverse_projection` and then multiplies by `inverse_view`
(`gbufferModelViewInverse`), and Horizon's own packs do the same through
`world_offset` and `frame.inverse_view`. Passing
`world_projection(fov, aspect, far) * bob` as the projection while `view`
stayed unbobbed left the two halves disagreeing by the walk bob, so
everything a pack reconstructed — reflections, ambient occlusion,
godrays — turned with the bob. That reads as reflections wobbling in
step with the first-person hand, and changing the FOV made it worse.
Vanilla and Iris carry the bob in the world's pose stack (the model-view
matrix), not in the projection. Pass `bob * view` as the view and the
plain world projection instead: `projection * view` is unchanged, so
terrain, entities and the pack's own geometry programs draw as before,
while depth reconstruction is now consistent. The hand's uniforms
already carried the bob in its pose and are untouched.
Checked with Complementary Reimagined r5.9.3 over the depth-feedback
passes at 70 and 110 FOV.
Co-Authored-By: DeepSeek V4.1 Flash (OpenCode agent) <noreply@deepseek.com>
🤖 Agent: merged after verify and CI parity passed (shaders stack: Java packs draw their own terrain, water and shadow map, as in Iris). BSL shows real shadows on Vulkan and Metal; frame times and screenshots are on #1326. Without a pack, nothing changes.
Shadow quads were lifted 0.001 above the block top and drawn in the entity
batch with depth writes. LivingEntityRenderer's translate(0, -1.501, 0)
puts a model's feet 0.001 above the ground too, so a chicken's feet and the
shadow under it were coplanar and flickered (#1440).
ShadowFeatureRenderer puts the quads on the block top itself
(relativeY + shapeBelow.minY), and RenderPipelines.ENTITY_SHADOW tests depth
without writing it. Shadows are now their own entity material (overlay 14)
drawn by such a pipeline after the entities. A small depth bias, constant
plus 1.5% of the slope, keeps them over the terrain under them, whose
vertices snap to the subpixel grid differently (without it whole shadows
vanished at some camera positions), while staying under feet 0.001 higher.
The shader pack shadow map no longer draws entity shadows as casters.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: merged after verify and CI parity passed (review-0-survey quick win: dead pub items; the generated version.rs is kept as the generator writes it).
The shadow samplers bind with main's custom textures (texture.* first,
then the shadow map, then the rest); the shadow layers are drawn from the
shadow list on the flat path too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With main's Metal culling: the pack's depth test in OpenGL's layout runs
forced early on Vulkan (fs_box_early_depth) and in the shader on Metal
(fs_box_gl, the pack's depth attached read only and bound), and the flat
path culls too (phase 1 calls, flatten_phase2 after test_gl, then
draw_phase2 with each page's group 3).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Brings main's shader api 2 (#1414) up the stack. One path for blending:
the terrain programs' blend.* now go through Pass.blend/buffer_blends like
composite and final (java_pack::pass_blends). depthtex1/depthtex2 are one
source, DepthSolid: Horizon's depth before translucent geometry, then the
pack's solid terrain's. Horizon's world goes to the pack's first terrain
target (colortex1 for MakeUp).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The shared GUI atlas was always 4096 texels wide, so Faithful 64x needed
4096x32768: past the 16384 texels a side GPUs allow, and wgpu panicked
in create_texture at every launch until the pack was removed (#1442).
- The atlas widens (4096, 8192, 16384) until it fits within 16384 a
side, so Faithful 64x loads. Its height rounds up to a multiple of
1024 instead of a power of two: the default resources' atlas drops
from 4096x8192 to 4096x5120, Faithful 32x's from 4096x16384 to
4096x12288.
- Textures that still don't fit are an error, not a GPU panic. At
startup it rolls the packs back as Minecraft.rollbackResourcePacks
does: the game starts with its own resources, and the selection is
cleared and saved (clearResourcePacksOnError), so the next launch
doesn't fail again. The existing "Resource reload failed" screen
shows the reason.
- The renderer refuses an atlas larger than the device allows with an
error instead of a wgpu validation panic.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: merged after verify and CI parity passed. Metal GPU occlusion culling: on the M4 Pro at RD 32, FPS is +8% to +48% (interleaved A/B against main); the hidden sections match Vulkan's; parity --changed passed 104/104.
The world target made the opaque depth, normal and material textures
for every scaled or anti-aliased frame, packs or not: 12 bytes a pixel
nobody read. They now follow what the pack reads, and Post::generation
tells the stages to bind again when its targets change (a recreated
target of the same size, or a shadow map made after the stages bound).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Review fixes for #1414 (contribution gate):
- Java packs no longer get Horizon's shadow map: it was mirrored in y and
reversed in depth against the GL conventions the packs sample with, had
no distortion and only the camera's sections, so BSL drew no shadows
while the pass cost 0.65 ms (MakeUp, M4 Pro). shadowtex* stays a far
depth, as on main; the packs' own shadow programs are #1326.
- Invented uniforms removed (centerDepthSmooth fixed at 0.5,
ambientOcclusionLevel and shadowMapResolution as uniforms, constant
temp/rainfall/biome/held items); an unknown uniform stays zero as
before. wetness keeps its Iris smoothing. Unused directives removed.
- texture.<stage>.<sampler> is scoped to Iris's stages (final reads the
composite ones) instead of overriding the sampler in every pass;
texture.noise replaces noisetex. One noise generator and one RGBA
upload for both pack kinds.
- The G-buffer pass, the opaque-depth copy and the shadow pass run only
for a pack that reads them (naga's per-entry-point binding use, not a
text search): every pack paid all three before.
- Api 2's prelude additions go in their own file after api 1's, so api 1
packs keep the names they declare (aces_tonemap, material_at...).
- Stage kinds shadow, gbuffer, deferred and final ran as composite: only
composite is accepted; api 2 adds formats (validated), profiles,
option descriptions and a noise texture.
- Shader Pack Settings applies once when the page closes, as Iris's Done
(each click or slider step rebuilt the pack: 2.7 s per click, 12.7 s
for one drag with BSL). Labels follow Iris (.lang, else the name).
A pack whose options can't be read isn't read again every frame.
- Saved option values an option doesn't list are ignored, as OptiFine's
ShaderOption.setValue (a stale file broke BSL's compile).
- Import: the blocking file dialog (osascript/zenity/kdialog/PowerShell
on the event loop) is gone; a pack dropped on the Shader Packs page is
copied and picked, as Iris's onFilesDrop. Drops elsewhere do nothing,
an existing pack isn't overwritten, links aren't followed.
- Docs and the example say what the code does (SSAO, not GTAO).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On Apple's GPUs the passes the test adds cost a fixed share of the frame
(it is stored and reloaded at each), which the hidden sections repay only
at high render distances: at 12 culling cost an M4 Pro 0.05 ms of GPU time
a frame, at 32 it saved 0.25 to 1.6 ms. Other backends keep culling every
frame; HORIZON_GPU_CULL=all culls every frame on Metal too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
93 pub items (78 functions and methods, 14 constants, 1 struct) in horizon-client,
horizon-world, horizon-core, horizon-friends and horizon-server-files had
no caller anywhere in the repository: not in the crate, the other crates,
the examples, the tests or the binaries. rustc can't flag them because
horizon-client and the others are libraries, where a pub item counts as
used. The imports and the private helpers only they used go with them
(title.rs date_string, goals.rs flag_index).
Also drops the #[allow(dead_code)]/#[allow(unused)] items that only hid
dead code: chat_debug_fill, merchant_title and Goals::locked_by had no
caller, and four stubs (_keep, _unused, _rule_kind, const _) existed only
to keep an import used. LightTables::from_parts is used by its tests, so
it is #[cfg(test)] instead of allowed.
Protocol id tables and entity data serializer ids stay: they mirror
vanilla's registries in full. The launcher's three unused items are left
for a launcher PR.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Flat draws (one index list per frame, the default on Apple GPUs at high
render distances) turned occlusion culling off. Now phase 2's draws are
flattened after the test, by a compute pass that copies only the sections
the test saw, and each page run is drawn indirectly with the count it
reached.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
GPU occlusion culling ran on Vulkan only: its test needs early fragment
tests, which naga's Metal backend does not emit. On Metal the boxes are
tested in a pass of their own: phase 1's depth is bound read-only as the
attachment and as a texture, and the fragment shader compares its own
depth with it (reversed Z, greater or equal), as the fixed-function test
would. HORIZON_GPU_CULL=shader takes this path on Vulkan too, for checks.
Frames drawn flat (the default on Apple GPUs at high render distances)
still skip culling here.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The vignette, drawn over every frame, went from one texture sample to up to
four texel loads and per-pixel address arithmetic: +0.85 ms of GPU a frame on
an HD 620 (integer modulo, removed in the commit before), still +0.4 ms after
it.
- A blurred sprite gets a one-texel border in the atlas, copied from its
opposite edges: the hardware's bilinear sample of the atlas with the
entity group's linear sampler is then GL_LINEAR with repeat, edges
included. Clamp-to-edge keeps the sample point half a texel inside the
sprite, which is what the hardware's clamp does for both filters.
- Each sampling has its own pipeline, the shader specialised by override
constants: a repeat sprite runs main's arithmetic and one sample, nearest
or linear; only clamp adds a clamp. The sampling moves from the vertices
to the quad, so the vertex layout is main's again.
- A black vignette changes no pixel (it darkens by texture x colour), and is
no longer drawn: in daylight its brightness fades to 0 within half a
minute, so most frames outdoors skip it, which main never could.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: merged after verify passed and the wave-3 integration on the desktop by integ-w3: main + 17 PRs, fmt, clippy 0, 1929 workspace tests, parity --changed (only known flakes failed, and those passed on rerun), lan_e2e in both modes, and the perf numbers back to back against main posted on each PR.
On Apple GPUs timestamps go only at pass boundaries, and the frame was
bracketed by two one-thread compute passes. Two things made it read
about 0.01 ms:
- Metal resolves counter samples without waiting for the passes that
write them. Resolved in the frame's own command buffer, the end
timestamp was mostly the one an earlier frame left in the shared
query. Each readback slot now has its own queries, and with markers a
slot is resolved in a later frame, once on_submitted_work_done says the
GPU finished the frame it timed.
- A compute pass depends on nothing, so the end marker ran alongside the
frame's render passes. The end is now a render pass that loads the
frame's target, shading one pixel with every write masked: it runs
after every pass that drew into the target. The start stays a compute
pass (cheaper than a load and store of the target).
M4 Pro, render distance 32, offscreen: GPU p50 0.002 ms before, now
1.95 ms against a 1.95 ms frame (village) and 4.46 ms against 4.51 ms
(mountain), where the frame is GPU-bound. Vulkan and DirectX 12, which
write timestamps between commands, are unchanged but for the per-slot
queries.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
🤖 Agent: merged after verify passed and the wave-3 integration on the desktop by integ-w3: main + 17 PRs, fmt, clippy 0, 1929 workspace tests, parity --changed (only known flakes failed, and those passed on rerun), lan_e2e in both modes, and the perf numbers back to back against main posted on each PR.
🤖 Agent: merged after verify passed and the wave-3 integration on the desktop by integ-w3: main + 17 PRs, fmt, clippy 0, 1929 workspace tests, parity --changed (only known flakes failed, and those passed on rerun), lan_e2e in both modes, and the perf numbers back to back against main posted on each PR.