Vaibhav Rathod's Notes

PathTracer Learning - Phase 3 - GPU and Vulkan

4 min read

  • Phase 3 — GPU and Vulkan Ray Tracing

    • Moving from CPU to GPU. Vulkan’s ray tracing extension (VK_KHR_ray_tracing_pipeline) exposes hardware RT units (NVIDIA RTX, AMD RDNA2+).
    • Parent: PathTracer Learning

  • 3.1 Vulkan Ray Tracing Overview

    • PathTracer Learning - Vulkan RT Pipeline
      • Full pipeline breakdown: ray gen, intersection, any-hit, closest-hit, miss shaders
      • Shader Binding Table (SBT) construction and indexing
    • PathTracer Learning - BLAS and TLAS
      • How geometry is organized for hardware traversal
    • Key extensions required
      • VK_KHR_ray_tracing_pipeline — the RT pipeline itself
      • VK_KHR_acceleration_structure — BLAS/TLAS management
      • VK_KHR_ray_query — inline ray queries from any shader stage
      • VK_KHR_buffer_device_address — GPU pointers (required for SBT)
      • VK_KHR_deferred_host_operations — async BLAS builds
      • VK_EXT_descriptor_indexing — bindless textures/buffers

  • 3.2 Acceleration Structures in Vulkan

    • PathTracer Learning - BLAS and TLAS
    • BLAS (Bottom-Level Acceleration Structure)
      • Contains actual geometry: triangles or AABBs
      • Built once per mesh (or rebuilt for skinned meshes)
      • VkAccelerationStructureGeometryKHR — describes the geometry
      • VkAccelerationStructureBuildGeometryInfoKHR — build parameters
      • Build flags: PREFER_FAST_TRACE vs PREFER_FAST_BUILD vs ALLOW_UPDATE
      • Compaction: reduces memory by 50-70% after build
    • TLAS (Top-Level Acceleration Structure)
      • Contains instances of BLASes with transform matrices
      • Rebuilt every frame (instances move, appear, disappear)
      • VkAccelerationStructureInstanceKHR:
    • PathTracer Learning - Concept - Device Address Bit
      • VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT — required for SBT and AS input buffers
      • vkGetBufferDeviceAddress() returns a VkDeviceAddress (uint64)
      • This is a raw GPU pointer — enables pointer arithmetic in shaders

  • 3.3 Shader Binding Table (SBT)

    • See PathTracer Learning - Vulkan RT Pipeline for full details
    • The SBT maps ray types to shader programs
    • Layout: [RayGen | Miss shaders | Hit groups | Callable shaders]
    • Each entry is a shader handle (32 bytes) + optional inline data
    • Alignment requirements
      • shaderGroupHandleSize — typically 32 bytes
      • shaderGroupBaseAlignment — typically 64 bytes
      • Each region must be aligned to shaderGroupBaseAlignment
    • Indexing formula
      • hitGroupIndex = instanceSBTOffset + geometryIndex * sbtStride + rayContributionToHitGroupIndex

  • 3.4 Ray Tracing Shaders (GLSL)

    • See PathTracer Learning - Vulkan RT Pipeline for complete shader examples
    • Ray generation shader (.rgen)
      • Entry point for each pixel
      • Calls traceRayEXT() to launch rays
      • For path tracing: implement the full bounce loop here (not recursively)
    • Closest-hit shader (.rchit)
      • Called for the nearest confirmed hit
      • Fetch vertex data via buffer device address (GL_EXT_buffer_reference)
      • Evaluate BRDF, trace shadow rays
    • Miss shader (.rmiss)
      • Called when ray hits nothing
      • Sample environment map
    • Any-hit shader (.rahit)
      • Called for every potential hit — use for alpha testing
      • ignoreIntersectionEXT() to reject transparent pixels

  • 3.5 Bindless Resources

    • Path tracers need access to many textures and buffers (one per mesh/material)
    • Bindless: bind a large array of descriptors, index at runtime
    • VK_EXT_descriptor_indexing — enables unbounded descriptor arrays
    • nonuniformEXT — required when index varies per invocation (non-uniform)
    • Alternative: push constants or inline data in SBT entries

  • 3.6 Iterative Path Tracing on GPU

    • Recursive approach: hit shader calls traceRayEXT → limited by maxPipelineRayRecursionDepth
    • Iterative approach (preferred): loop in ray generation shader
    • Payload carries all data needed between bounces

  • 3.7 Async Compute

  • 3.8 Ray Query (Inline RT)

    • Alternative to RT pipeline: query rays from any shader stage
    • VK_KHR_ray_query — no SBT, no separate shader stages
    • Useful for: shadow rays in compute shaders, AO, simple occlusion queries
    • Less flexible than RT pipeline but simpler to integrate

  • Phase 3 Checklist

    • Create BLAS from triangle mesh
    • Create TLAS with instances
    • Build and upload SBT
    • Write ray generation shader
    • Write closest-hit shader with BRDF evaluation
    • Write miss shader for sky/environment
    • Trace a primary ray and display result
    • Add shadow rays for direct lighting (NEE)
    • Implement iterative path tracing loop in rgen shader
    • Set up bindless texture array for materials
    • Implement BLAS compaction

Graph View

Backlinks