Watch

feat(device): applications ask for the device features their shaders use

Device
- Device::Initialize(required, optional) takes DeviceFeatures, the five
  core feature structs vulkan10 to vulkan14. The library enables only what
  its own shaders and code use. Device::enabledFeatures says what was
  enabled, supportedFeatures what the GPU has, apiVersion the version in
  use.
- shaderInt16, shaderFloat16, storageBuffer16BitAccess, scalarBlockLayout
  and the uniform-buffer, storage-buffer and storage-image non-uniform
  indexing bits are no longer enabled for everyone. An application whose
  shaders take uint16_t spec constants passes
  {.vulkan10 = {.shaderInt16 = VK_TRUE}}. scalarBlockLayout stays on with
  the NVIDIA workaround, whose TLAS address may sit in a struct's tail
  padding.
- Every GPU is weighed before one is chosen: Vulkan 1.3, the swapchain and,
  below 1.4, maintenance5 extensions, the library's needs, the required
  features. One that can't run the application is skipped, saying why when
  it ranks higher. A CPU device runs it only when nothing else is listed,
  and CRAFTER_GRAPHICS_DEVICE never falls back. A GPU two drivers both list
  is tried through each.
- When no GPU qualifies, Initialize throws naming each GPU's gaps. It
  destroys the instance it made, before any Wayland connection, so it can
  be called again asking for less.
- maintenance5 is vulkan14's on a 1.4 GPU and VK_KHR_maintenance5's below.
  CRAFTER_GRAPHICS_NO_VULKAN_1_4 runs as a 1.3 application, to exercise
  that path on a 1.4 GPU.
- VK_KHR_spirv_1_4 and VK_KHR_shader_float_controls, core since 1.2, are no
  longer enabled, nor VK_EXT_shader_atomic_float, which was enabled without
  its feature struct. memoryDecompression's feature bit is queried before
  it is enabled.

Shaders
- VulkanShader checks a module's capabilities, declared directly or
  implicitly, against the enabled core features, the GPU's subgroup
  operations and the optional groups, and throws naming the field to pass,
  or saying the GPU lacks it. The tables are generated from vk.xml and
  SPIR-V's grammar.
- WorkaroundNvidiaAS::Patch sized a row-major matrix push member by its
  columns, so the TLAS address could land inside it.

Window
- The swapchain format is one whose own feature flags allow storage reads
  and writes without a format, which since Vulkan 1.3 each format says for
  itself; shaderStorageImageRead/WriteWithoutFormat are no longer required.

Tests
- DeviceFeatures: ResolveFeatures, ChooseDevice and CheckShaderCapabilities,
  without a GPU.
- FeatureRequests: required and optional features end to end, the 1.3
  path, and a retry after a refusal.
- PushConstantRewrite: a row-major mat3x4 push member.
- HeapASRead, HeapDynamicIndexHit, RTPassPushData and the ray-tracing
  examples ask for the features their shaders use.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

2026-10-31 06:40:02 +01:00

5.4 KiB

Raw Blame History

VulkanTriangle

The minimal ray-traced example. Renders a single static triangle through vkCmdTraceRaysKHR. No UI.

What it shows

Device::Initialize({.vulkan10 = {.shaderInt16 = VK_TRUE}}) + Window + swapchain bring-up. The raygen takes the heap's buffer start as a uint16_t spec constant, which needs shaderInt16; without asking for it, loading the shader throws, naming the feature.
A DescriptorHeapVulkan sized for one image + one buffer slot, with slot ranges allocated via the bump-allocator API (AllocateImageSlots, AllocateBufferSlots).
A PipelineRTVulkan built from raygen / miss / closesthit SPIR-V shaders compiled at build time.
Mesh::Build constructing a BLAS and RenderingElement3D::BuildTLAS the per-frame TLAS.
Direct descriptor writes via vkWriteResourceDescriptorsEXT for the swapchain views and TLAS device addresses.
RTPass{&pipeline} plugged into window.passes — the canonical way to add ray tracing to a window in this library.

It's the smallest sensible test of the bindless VK_EXT_descriptor_heap

ray-tracing path.

Run

cd examples/VulkanTriangle
crafter-build -r

You should see a 1280×720 window with an RGB-barycentric triangle filling roughly the centre. On the NVIDIA driver this works through an engine-side workaround for a driver fault — see below.

Native status — NVIDIA driver fault, worked around

On NVIDIA driver 610.43.02 (Vulkan 1.4) reading the acceleration structure through VK_EXT_descriptor_heap aborts the device with VK_ERROR_DEVICE_LOST on the first frame. This is a driver-side fault in the brand-new descriptor-heap acceleration-structure path, not an engine bug (full investigation in #7, summarised below).

The engine works around it transparently (issue #15). On the NVIDIA proprietary driver only, VulkanShader rewrites the compiled SPIR-V at module-load time so that every OpLoad of an accelerationStructureEXT out of the heap becomes a load of the TLAS device address (from a push-constant block) followed by OpConvertUToAccelerationStructureKHR — which reads no descriptor and so never touches the faulting path. RTPass feeds the active frame's TLAS address in as push data. SPIR-V allows only one push-constant block per entry point, so when a shader already declares one the TLAS address is appended to that block (rather than adding a second, which would fail validation — issue #18); shaders without a push constant get a freshly synthesized single-member block. raygen.glsl and the example code are unchanged; acceleration structures still bind into the heap normally. On every other driver the workaround is inert. It's gated on Device::workaroundDescriptorHeapAS and confined to one fenced block in interfaces/Crafter.Graphics-ShaderVulkan.cppm so it can be deleted wholesale once a fixed NVIDIA driver ships.

Status on later drivers

On 615.78.08 the fault is still there. With the rewrite off (CRAFTER_GRAPHICS_NO_AS_WORKAROUND set), traceRayEXT still loses the device, and an inline ray query that reads the TLAS from the heap no longer faults but renders nothing. Mesa's llvmpipe now advertises VK_EXT_descriptor_heap too and renders the plain heap read correctly, which is a second implementation backing the conclusion below. tests/HeapASRead renders this triangle through both paths and checks the pixels; HeapASRead --no-workaround is the quick check for a new driver, reporting each path on its own.

The underlying fault (#7)

The fault was traced to the acceleration-structure read through VK_EXT_descriptor_heap, not to the engine's RT setup:

The BLAS/TLAS build is correct and finishes before rendering (Window::FinishInit does vkQueueWaitIdle). The built TLAS instance has an identity transform, mask = 0xFF, and the correct BLAS device address.
The AS descriptor is written correctly — vkWriteResourceDescriptorsEXT stores the TLAS device address at the expected heap byte offset (verified by dumping the raw heap bytes after the write).
The Khronos validation layers (1.4.350, current) report zero errors for the whole frame, including the SBT regions handed to vkCmdTraceRaysKHR.
Storage images and buffers bound through the same descriptor heap work — with traceRayEXT removed, the raygen shader's imageStore renders correctly, so the heap binding / image path is sound.
Both the ray-tracing pipeline (traceRayEXT) and inline ray query (rayQueryEXT, which uses no shader binding table) fault identically when they read the acceleration structure from the heap. That isolates the fault to the AS-via-heap read, not the SBT or the RT pipeline.
The fault reproduces even with the AS descriptor written at heap byte 0 and read at shader index 0 (no descriptor offset/stride ambiguity), and is unaffected by the pAddressRange size.
VK_EXT_descriptor_heap is brand new; at the time NVIDIA was the only implementation on this machine that advertised it (llvmpipe did not), so there was no second conformant implementation to cross-check against. Mesa's llvmpipe has it now; see "Status on later drivers" above.

Conclusion: this is a driver-side fault in NVIDIA's VK_EXT_descriptor_heap acceleration-structure path, not an engine bug, and it should be reported to NVIDIA. Until a fixed driver ships, the SPIR-V rewrite above keeps the native RT path working; once it does, remove the workaround and the heap AS read becomes the direct path again.

5.4 KiB Raw Blame History Unescape Escape