Skip to content

vulkan_implementation_roadmap.md

Latest commit

 

History

History
366 lines (283 loc) · 26.3 KB

vulkan_implementation_roadmap.md

File metadata and controls

366 lines (283 loc) · 26.3 KB

Distant Horizons → VulkanMod: Implementation Roadmap

Status as of 2026-03-15: Phases 1-7, extension mod port, multi-version support, MC depth comparison, cloud rendering, and DH 3.0 compatibility complete. LODs render with correct colors, depth, lightmap, transparency, compositing, SSAO, fog, noise, earth curvature, overdraw prevention, cloud depth, weather ordering, and MC depth occlusion via an extension mod (mixin-based, no DH source modifications). Supports MC 1.20.6 (VulkanMod 0.4.2) and MC 1.21.11 (VulkanMod 0.6.1) from a single codebase. Compatible with both DH 2.4.x and 3.0.x.

Architecture Overview

The Vulkan backend is implemented as a Fabric extension mod (dh-vulkanmod) in the vulkan/ subproject. It requires Distant Horizons to be installed and uses mixins to intercept DH's rendering pipeline without modifying DH's source code. All version-specific API differences are centralized in Compat.java using Manifold preprocessor directives.

Core files under vulkan/src/main/java/com/braffolk/dhvulkan/:

  • VulkanRenderContext.java — Pipeline creation, UBO management, shader conversion, draw calls
  • VulkanRenderEngine.java — VulkanBackend impl, per-frame uniform fill, VBO cache, draw dispatch
  • DhVulkanFramebuffer.java — DH-owned Vulkan framebuffer with color+depth, auto-resize
  • DhCompositePipeline.java — Fullscreen triangle composite with depth bias
  • DhSsaoPipeline.java — 2-pass SSAO post-process
  • DhFogPipeline.java — 2-pass fog post-process (distance + height)
  • DhDepthReaderPipeline.java — Copies MC depth to sampleable R32F texture (required on NVIDIA Windows)
  • IVulkanRenderDelegate.java — Interface for the Vulkan delegate
  • DhVulkanMixinPlugin.java — IMixinConfigPlugin for resolving mixin conflicts between DH and VulkanMod
  • compat/Compat.java — Single source of truth for all version-specific API differences

Mixins under vulkan/src/main/java/com/braffolk/dhvulkan/mixin/:

  • MixinLodBufferContainer — Intercepts uploadBuffersDirect() to store raw vertex data directly
  • MixinLodRenderer — Redirects LodRenderer.renderLodPass() to the Vulkan delegate
  • MixinLevelRenderer — Triggers composite after MC terrain renders
  • MixinGLProxy — Creates dummy GLProxy, bypasses GL thread assertions
  • MixinGLBuffer — Cancels GL buffer operations (bind, create, upload, destroy)
  • MixinGLVertexBuffer — Duck interface for storing Vulkan buffer handles on VBOs
  • MixinGLState — Cancels GL state machine operations
  • MixinLightMapWrapper — Cancels raw GL calls in DH's LightMapWrapper (prevents crashes in Vulkan context)

Vulkan shaders under vulkan/src/main/resources/shaders/vulkan/:

  • dh_terrain.vert / dh_terrain.frag — Terrain rendering shaders (with gl_ClipDistance overdraw)
  • dh_apply.vert / dh_apply.frag — Composite shaders
  • dh_ssao.vert / dh_ssao.frag / dh_ssao_apply.frag — SSAO shaders
  • dh_fog.frag / dh_fog_apply.frag — Fog shaders
  • dh_depth_read.frag — MC depth reader (copies depth to R32F)

Phase 1: Fix Visual Rendering ✅ COMPLETE

Root causes found and fixed:

  1. White LODs + 180° culling: Vertex and fragment shaders had different UBO layouts at binding=0. Fragment shader read matrix data as uClipDistance → garbage values → wrong discard
  2. bool compile error: Changed bool → int in UBO but forgot !var → var == 0 conversion
  3. Dark colors: VTextureSelector.getImage(0) returns block atlas, not lightmap. Sampling atlas with lightmap UVs → garbage. Bypassed with full brightness constant
  4. Transparency: MC's blend state inherited by DH pipeline. Disabled PipelineState.blendInfo.enabled
  5. No depth writes: MC had depthMask=false. Overridden to true
  6. Lightmap always daytime: Three bugs — (a) texelFetch coordinates swapped (blockLight,skyLight) vs original (skyLight,blockLight), (b) unnecessary .bgr swizzle, (c) replaced CPU LightmapManager with MC's framebuffer lightmap via GlTexture.glId() → VkGlTexture → VulkanImage

Phase 2: Depth Integration ✅ COMPLETE

  • LODs render behind MC terrain via depth bias
  • Uses MC's projection matrix (not DH's) for compatible depth values
  • Lightmap: uses MC's framebuffer-rendered lightmap via VulkanMod's GL emulation layer

Phase 3: Transparency / Blending ✅ COMPLETE

  • Added setBlendState(boolean) to IVulkanRenderDelegate — toggles PipelineState.blendInfo
  • Sets blend functions: SRC_ALPHA/ONE_MINUS_SRC_ALPHA (RGB), ONE/ONE_MINUS_SRC_ALPHA (alpha)
  • Re-binds pipeline after blend state change (VulkanMod bakes blend into pipeline objects)

Phase 4: Buffer Management & Performance ✅ COMPLETE

  • VBO cache refactored with CachedBuffer (tracks Vulkan buffer + ByteBuffer identity)
  • Invalidation: detects when vulkanBufferHandle changes and frees old GPU buffer
  • Proper cleanup in cleanup() frees all cached Vulkan buffers
  • Vertex buffers use GPU_MEM (device-local VRAM) with automatic staging via VulkanMod

Phase 5: State Management ✅ COMPLETE

  • Full save/restore in beginFrame()/endFrame(): cull, depthMask, depthFun, topology, polygonMode, blendInfo (all 6 fields)
  • Explicit topology = TRIANGLE_LIST, polygonMode = FILL set per frame

Phase 6: Framebuffer / Render Pass Integration ✅ COMPLETE

What was implemented

  • DH-owned Vulkan framebuffer (DhVulkanFramebuffer.java) — color (RGBA8) + depth attachments with SAMPLED_BIT
  • Render pass with LOAD_OP_CLEAR/STORE_OP_STORE on both attachments, finalLayout = SHADER_READ_ONLY_OPTIMAL
  • Render pass switching — flow: End MC pass → Begin DH pass → Render LODs → End DH pass → Rebind MC pass → Composite
  • Composite pipeline (DhCompositePipeline.java) — fullscreen triangle (via gl_VertexIndex) with depth writes via gl_FragDepth
  • Composite shaders (dh_apply.vert, dh_apply.frag) — Vulkan GLSL 450, Y-flip for framebuffer coords
  • Depth bias (+0.0001 in composite shader) — LODs always slightly behind MC terrain, no z-fighting
  • uClipDistance — overdraw prevention via gl_ClipDistance[0] (hardware vertex-level clipping) + fragment shader dithered fade
  • Window resize — auto-recreate framebuffer via Renderer.addOnResizeCallback()
  • Back-face culling enabled — ~50% fragment reduction, +50fps on M1 Max (170fps vs 120fps)
  • Explicit depth test — depthTest=true, depthFun=GL_LEQUAL ensures Early-Z hardware is active
  • Removed polygonOffset — no longer needed with own framebuffer + depth bias

Caveats & future considerations

  • Transparent block overlap: Per-pixel MC depth comparison is now implemented (see MC Depth Comparison below). The deferredComposite path copies MC's depth to an R32F texture for sampling.
  • Vulkan constants inlined: DhVulkanFramebuffer.java and DhCompositePipeline.java inline Vulkan VK10 constants as static final int because org.lwjgl.vulkan is not on the fabric module's compile classpath.
  • VRenderSystem uses GL constants: All depthFun and similar values must use GL constants (e.g., GL_LEQUAL=515, GL_ALWAYS=519), not raw Vulkan values. VulkanMod converts them internally via PipelineState.DepthState.glToVulkan().
  • Dummy UBO at binding 0: VulkanMod's pipeline system requires at least one UBO. The composite pipeline uses a 4-byte dummy.

Phase 7: Shader Features

  • SSAO — Vulkan-native 2-pass post-process, see details below
  • Noise texture — integrated into terrain fragment shader (flat_shaded.frag): procedural per-block dithering via uNoiseEnabled, uNoiseSteps, uNoiseIntensity, uNoiseDropoff uniforms populated from config
  • Fog — Vulkan-native 2-pass post-process via DhFogPipeline.java, see details below
  • Earth curvature — vertex shader curves terrain based on uEarthRadius
  • Wireframe debug — needs VK_POLYGON_MODE_LINE pipeline variant
  • Cloud rendering — custom Vulkan cloud renderer with VBO mesh geometry, correct depth compositing against LOD terrain. Supports fast/fancy modes. VM 0.6+ only.

SSAO Implementation Details

Files:

  • DhSsaoPipeline.java — orchestrates both passes, manages resources
  • dh_ssao.vert — shared fullscreen triangle vertex shader (gl_VertexIndex, Y-flip)
  • dh_ssao.frag — pass 1: spiral depth-sampling occlusion computation
  • dh_ssao_apply.frag — pass 2: optional bilateral blur + multiplicative apply

Architecture:

  • Pass 1 renders into an intermediate R16F Framebuffer (raw occlusion values 0.0–1.0)
  • Pass 2 reads the R16F texture + DH depth, applies blur (currently disabled for perf), and multiplicatively blends onto DH's color buffer using LOAD_OP_LOAD
  • Inserted in VulkanRenderDelegate.endFrame() between ending DH's render pass and the composite pass
  • Gated on Config.Client.Advanced.Graphics.Ssao.enableSsao

Current Parameters (tuned for Vulkan path):

  • uSampleCount=4 (GL uses 6, reduced for perf)
  • uRadius=4.0, uStrength=0.18, uMinLight=0.30, uBias=0.025
  • gBlurRadius=0 (blur disabled for performance, bilinear filtering on R16F suffices)
  • Pass 2 render pass is cached (created once, reused every frame)

Important

Critical: Do NOT apply Y-flip to depth reconstruction in dh_ssao.frag. The Vulkan vertex shader flips TexCoord.y for framebuffer sampling, and the natural instinct is to negate Y when converting TexCoord→NDC for the inverse projection matrix (since MC's projection uses GL convention with Y=+1 at top). However, testing showed this produces overly aggressive, unnatural occlusion with harsh dark halos. The non-flipped version produces visually correct, MC-like ambient occlusion. This is likely because VulkanMod's viewport flip already handles the Y convention, making the depth values in DH's framebuffer consistent with the TexCoord mapping.

Warning

Performance: SSAO at retina resolution is expensive. On M1 Max at retina fullscreen (~5M pixels), SSAO costs ~35% FPS (155→99). Half-resolution rendering was attempted but caused horizontal banding artifacts — likely needs better bilinear upscale handling or a dedicated upscale pass. Current mitigations: sample count reduced to 4, blur disabled. Future options: compute shader single-pass, temporal accumulation, or quarter-res with smart upscale.

Note

Projection matrix: The Vulkan SSAO uses mcProjectionMatrix (not dhProjectionMatrix) because DH's LODs are rendered with MC's projection in the Vulkan path (for depth compatibility with MC terrain compositing). This differs from the GL path which uses dhProjectionMatrix. The parameters were tuned accordingly.

Fog Implementation Details

Files:

  • DhFogPipeline.java — orchestrates both passes, manages resources (~25 config-driven uniforms)
  • dh_fog.frag — pass 1: depth → world pos reconstruction, far fog + height fog (3 falloff types, 10 mixing modes)
  • dh_fog_apply.frag — pass 2: depth-gated fog texture passthrough with SRC_ALPHA blend
  • Reuses dh_ssao.vert for fullscreen triangle vertex shader

Architecture:

  • Pass 1 renders into intermediate RGBA16F Framebuffer (fog color + alpha 0.0–1.0)
  • Pass 2 reads fog texture + DH depth, applies depth-gated fog with SRC_ALPHA / ONE_MINUS_SRC_ALPHA blend onto DH's color buffer
  • Inserted in VulkanRenderDelegate.endFrame() after SSAO, before composite
  • Gated on Config.Client.Advanced.Graphics.Fog.enableDhFog
  • Pass 2 render pass is cached (same pattern as SSAO)

Important

Projection matrix for fog: Like SSAO, the fog pipeline uses mcProjectionMatrix * dhModelViewMatrix for world position reconstruction. Initially dhProjectionMatrix was used (matching the GL path), but this caused all LODs to render as 100% fog because the depth values were produced with mcProjectionMatrix. The inverse matrix must always match the matrix that produced the depth buffer.

MC Depth Comparison ✅ COMPLETE

Goal: Per-pixel depth comparison between MC terrain and DH LODs, so LODs are hidden wherever MC has rendered.

Implementation

  • DhDepthReaderPipeline.java — Renders MC's swapchain depth into a sampleable R32F color texture via a separate render pass. Required because Vulkan can't sample a depth attachment while it's bound to the active render pass.
  • dh_depth_read.frag — Simple passthrough shader that samples the depth texture and writes to gl_FragDepth / color output.
  • deferredComposite() in VulkanRenderDelegate — Called from MixinLevelRenderer after MC terrain renders. Reads MC depth via the depth reader, then composites DH's framebuffer with per-pixel depth comparison.
  • Works on all platforms: macOS (MoltenVK), Windows (NVIDIA, AMD), Linux.

Fade Mode Support

  • Respects DH's vanillaFadeMode setting:
    • NONE: Composites in endFrame() immediately (no deferred pass). MC depth available for debug mode 6 only.
    • SINGLE/DOUBLE: endFrame() composites without MC depth, then deferredComposite() re-composites with MC depth from the depth reader.
  • Unified rendering path — no version-specific branching.

Overdraw Prevention (Hybrid Approach)

  • Vertex shader: gl_ClipDistance[0] = length(vertexWorldPos.xz) - uClipDistance — hardware clips LOD geometry within MC's render distance. Zero fragment shader cost for clipped geometry.
  • Fragment shader: Dithered fade via smoothstep(uClipDistance, uClipDistance * 1.5, viewDist) + Bayer matrix noise — smooth visual transition when dithering is enabled.
  • uClipDistance computed from renderDistChunks * 16 * overdraw, matching original DH's distribution.

Important

Do NOT modify the projection matrix for overdraw clipping. An earlier attempt used Mat4f.setClipPlanes() to push the near clip plane to the overdraw distance. This broke depth ordering — LODs rendered in front of MC terrain because the different near/far clip planes produced incompatible depth buffer values. The gl_ClipDistance approach clips without affecting depth values.

Warning

VM 0.4.2 descriptor caching issue. On VulkanMod 0.4.2, texture descriptors bound to VTextureSelector slots persist across pipeline binds. If endFrame() binds a white fallback texture to slot 4 (for null MC depth), that fallback descriptor persists into deferredComposite()'s composite call, overriding the real MC depth texture. Fix: endFrame() skips compositing entirely for debug mode 6 when deferredComposite will run.

Note

Depth reader timing is critical. readDepth() must be called AFTER Renderer.endRenderPass() (MC depth is no longer an active attachment) and BEFORE Compat.rebindMainTarget() (MC's render pass restarts). The prepareMcDepthForSampling() / restoreMcDepthView() split ensures depth-only image views persist through the descriptor write.

Known Quirks & Platform Differences

Caution

NVIDIA: VulkanMod's GPU_MEM vertex buffers produce incorrect data in fullscreen quad geometry. When using a 4-vertex quad with index buffer (drawIndexed(vbo, ibo, 6)), NVIDIA only rasterized one of the two triangles. The vertex buffer data was read incorrectly, likely caused by VulkanMod's VMA sub-allocation offsets or async staging buffer transfers.

Fix: All fullscreen passes (composite, SSAO, fog) generate positions from gl_VertexIndex in the vertex shader, bypassing vertex buffer data entirely. A single oversized triangle at NDC (-1,-1), (3,-1), (-1,3) covers the entire viewport — the GPU clips the excess. This is also a Vulkan best practice (fewer vertices, no shared diagonal edge, no helper lane overhead).

The vPosition input is declared but unused (required by VulkanMod's pipeline vertex format binding). A dummy 24-byte vertex buffer is still bound.

Warning

Composite depth test must use GL_ALWAYS (519). The composite runs before MC terrain, so MC's depth buffer contains DONT_CARE garbage. GL_LEQUAL fails against random values. MoltenVK implicitly clears to 1.0, hiding the bug on macOS.

Phase 8: Iris/Shader Pack Compatibility — N/A

  • VulkanMod does not support shader packs — this phase is blocked until it does
  • DH's Iris integration (IRIS_ACCESSOR) is skipped for Vulkan path
  • Custom shader overrides via IDhApiShaderProgram remain GL-only

Phase 9: Edge Cases & Robustness

  • Handle renderer reset — cleanup() now resets initFailed flag, allowing re-initialization after settings changes. Previously, a transient init failure permanently disabled rendering with no errors.
  • Config hot-reload — DhVulkanConfig.reload() throttled to once per second (was reading JSON from disk every frame)
  • Per-frame allocation optimization — pre-allocated Mat4f, float[16], Vec3f fields avoid heap allocations on the hot path
  • Handle VulkanMod not being loaded (graceful fallback to GL)
  • Thread safety: DH runs LOD generation on worker threads — Vulkan buffer uploads must happen on the render thread

Phase 11: Multi-Version Support ✅ COMPLETE

Goal: Support MC 1.20.6 (VulkanMod 0.4.2) alongside MC 1.21.11 (VulkanMod 0.6.1) from one codebase.

Key Differences Between VM 0.4.2 and VM 0.6.1

Area VM 0.6.1 (MC 1.21.11) VM 0.4.2 (MC 1.20.6)
Buffer imports net.vulkanmod.vulkan.memory.buffer.* net.vulkanmod.vulkan.memory.*
IndexBuffer Constructor takes IndexType.UINT32 No IndexType param
Drawer.drawIndexed Supports UINT32 Hardcodes UINT16
Buffer.scheduleFree scheduleFree() freeBuffer()
Uniform suppliers Info.setBufferSupplier() at construction Only resolves built-in MC names; custom uniforms need manual wiring
VertexFormatElement new VFE(id, index, Type, Usage, count) new VFE(index, Type, Usage, count)
VertexFormat VertexFormat.builder().add().build() new VertexFormat(ImmutableMap)
SwapChain access Renderer.getInstance().getSwapChain() Vulkan.getSwapChain()
beginRenderPass Renderer.getInstance().beginRenderPass() Requires reflection to access currentCmdBuffer
GlTexture lightmap GlTexture → VkGlTexture bridge exists No GL texture bridge (lightmap unavailable)
Noise intensity config Pre-scaled 0–1 float Unscaled integer (needs * 0.01)
Mixin conflicts None DH's MixinTextureUtil conflicts with VM's MTextureUtil
GL context Partial GL emulation No GL context — raw GL calls crash

Implementation

  • All version differences handled in Compat.java via #if MC_VER >= MC_1_21_1 / #else Manifold directives
  • DhVulkanMixinPlugin (IMixinConfigPlugin) strips conflicting DH mixins on MC 1.20.6
  • MixinLightMapWrapper cancels raw GL calls that would crash in VM 0.4.2's Vulkan-only context
  • Compat.drawIndexed() bypasses VM 0.4.2's Drawer to issue raw Vulkan commands with VK_INDEX_TYPE_UINT32
  • Compat.setUniformSuppliers() manually wires DH's MappedBuffer suppliers into VM 0.4.2's Uniform subclasses

Phase 10: Extension Mod Port ✅ COMPLETE

Goal: Port the Vulkan backend from a DH source fork to a standalone Fabric extension mod that works alongside unmodified DH releases.

Key Design Decision: Clean Override vs Monkeypatch

The fork modified LodBufferContainer.uploadBuffersDirect() directly to add a Vulkan code path that bypasses vbo.uploadBuffer() and stores raw ByteBuffer data on the VBO. The initial extension mod approach used a low-level mixin on GLBuffer.uploadBuffer() to intercept data — this caused missing east/west faces because the interception was too deep in the call stack and version-dependent.

The fix: MixinLodBufferContainer intercepts uploadBuffersDirect() at HEAD, replicating the fork's approach. This grabs raw ByteBuffers from the quad builder before any GL operations, stores them via a duck interface, and cancels the entire GL upload path.

Files Created/Modified

  • [NEW] MixinLodBufferContainer.java — Clean buffer upload interception
  • [NEW] MixinLodRenderer.java — Redirects rendering to VulkanRenderDelegate
  • [NEW] MixinLevelRenderer.java — Composite trigger after MC terrain
  • [NEW] MixinGLProxy.java — Dummy GLProxy, thread assertion bypass
  • [NEW] MixinGLBuffer.java — GL call cancellation
  • [NEW] MixinGLVertexBuffer.java — Duck interface (IVulkanVertexBuffer)
  • [NEW] DhVulkanModEntrypoint.java — Fabric entrypoint
  • [NEW] DhVulkanConfig.java — Mod config

Lessons Learned

  • Intercept data at the highest stable level (LodBufferContainer), not deep in the GL stack
  • Duck interfaces (IVulkanVertexBuffer) are the clean way to add fields to DH classes via mixin
  • DH's GLProxy.queueRunningOnRenderThread() works correctly even with a dummy GLProxy
  • Terrain shaders (standard.vert, flat_shaded.frag) load from DH's jar at runtime — no need to bundle

Phase 8: DH 3.0 Rendering API Integration -- IN PROGRESS

Phase 8a (core refactor) complete. VulkanBackend interface extracted, pipelines in core/pipeline/, RenderUniforms/VkVertexData created, DH 2.4 mixins organized under mixin/shared/. Both DH 2.4.x and 3.0.x supported from a single codebase via DhConfigHelper abstraction.

DH 3.0 introduces AbstractDhRenderApiDefinition, a pluggable rendering backend with injectable singletons for all render passes (IDhMetaRenderer, IDhTerrainRenderer, IDhSsaoRenderer, IDhFogRenderer, IDhFarFadeRenderer) and factory methods (IVertexBufferWrapper, ILodContainerUniformBufferWrapper). This eliminates 6 of our 8 current mixins.

Instead of intercepting DH's GL pipeline via mixins, we register a VkDhRenderApiDefinition that provides Vulkan implementations of each renderer interface. DH calls our code directly.

Still needed: MixinLightMapWrapper (DH's MixinLightTexture calls GL-dependent uploadLightmap() every frame from outside the rendering API), MixinLevelRenderer (post-MC-terrain composite timing), and a small MixinDependencySetup to register our API definition.

Architecture: Three-Layer Design

dhvulkan/
├── core/             # Vulkan engine, DH-agnostic, zero DH imports
│   ├── VulkanBackend       -- central rendering interface
│   ├── VulkanRenderEngine  -- refactored from VulkanRenderDelegate
│   ├── pipeline/           -- SSAO, fog, composite, depth reader
│   └── data/               -- RenderUniforms, VkVertexData
├── dh24/             # DH 2.4.x integration, all 8 current mixins + adapter
│   ├── mixin/              -- MixinLodRenderer, MixinGLProxy, etc.
│   ├── duck/               -- IVulkanVertexBuffer, etc.
│   └── Dh24RenderDelegate  -- translates DH 2.4 types to core types
├── api/              # DH 3.0 integration, API impls + 2-3 mixins
│   ├── VkDhRenderApiDefinition
│   ├── renderer/           -- VkMetaRenderer, VkTerrainRenderer, etc.
│   ├── VkVertexBufferWrapper
│   └── mixin/              -- MixinDependencySetup, MixinLightMapWrapper, MixinLevelRenderer
├── bridge/           # Version detection + integration interface
│   ├── DhIntegration       -- interface both paths implement
│   └── DhVersionDetector   -- runtime class presence check
└── compat/           # MC version compat (existing)

Nothing in core/, api/, bridge/, or compat/ imports from dh24/. Dropping DH 2.4 support = delete dh24/ + edit 2 files (dh-vulkanmod.mixins.json, DhVulkanModEntrypoint.java).

DhVersionDetector checks for AbstractDhRenderApiDefinition class presence at runtime. DhVulkanMixinPlugin.shouldApplyMixin() conditionally loads only the appropriate mixin set.

Migration

  1. Phase 8a: Refactor core. Extract VulkanBackend interface from VulkanRenderDelegate, move pipelines to core/pipeline/, create RenderUniforms/VkVertexData. No behavior change. Move current mixins to dh24/. Verify existing DH 2.4 functionality.
  2. Phase 8b: Add DH 3.0 path. Implement VkDhRenderApiDefinition + renderer impls + VkVertexBufferWrapper. Add MixinDependencySetup. Wire conditional loading. Test with DH 3.0.

Key Files Reference

File Purpose
compat/Compat.java Single source of truth for all version-specific API differences
VulkanRenderContext.java Pipeline, UBOs, shader conversion, draw API
VulkanRenderEngine.java Per-frame uniforms, VBO cache, draw dispatch, render pass switching
VulkanCloudRenderer.java Custom cloud rendering with VBO mesh and depth compositing
DhFrameProfiler.java Lightweight per-frame timing for all rendering phases
DhVulkanFramebuffer.java DH-owned Vulkan framebuffer (color + depth)
DhCompositePipeline.java Fullscreen triangle composite pipeline
DhSsaoPipeline.java 2-pass SSAO post-process
DhFogPipeline.java 2-pass fog post-process
DhDepthReaderPipeline.java Copies MC depth to sampleable R32F texture
DhVulkanMixinPlugin.java IMixinConfigPlugin for mixin conflict resolution
DhVulkanConfig.java JSON config with throttled hot-reload
MixinLodBufferContainer.java Intercepts buffer upload, stores raw vertex data
MixinLodRenderer.java Redirects LodRenderer to Vulkan delegate
MixinLevelRenderer.java Triggers deferredComposite after MC terrain
MixinLightMapWrapper.java Cancels raw GL calls in DH's lightmap code
dh_terrain.vert / dh_terrain.frag Terrain shaders (gl_ClipDistance overdraw)
dh_apply.vert / dh_apply.frag Vulkan composite shaders
dh_depth_read.frag MC depth reader shader
VertexFormats.java (DH) DH's custom vertex format definition (16 bytes)
PipelineState.java (VulkanMod) Render state (cull, blend, depth)

Vertex Format (16 bytes per vertex)

Offset  Size  GL Attr  Vulkan Format              Shader Input
0       8B    attr 0   R16G16B16A16_SINT          ivec4 vPosition (x,y,z + meta)
8       4B    attr 1   R8G8B8A8_UNORM             vec4 color (RGBA)
12      4B    attr 2   R32_SINT                   (material/normal/padding, unused by shader currently)

DH packs position (3 unsigned shorts) + light/micro-offset metadata into vPosition.a. The shader casts ivec4 → uvec4 for bitwise operations on the metadata.