Complete parity with wgpu#594
Draft
inner-daemons wants to merge 71 commits into
Draft
Conversation
- RenderBundleEncoder::new: add None device arg (new optional param) - VertexBufferLayout: wrap in Some (now Option<VertexBufferLayout>) - RequestAdapterOptions: add apply_limit_buckets: false field - enumerate_adapters: add apply_limit_buckets bool arg - render_bundle_encoder_finish: pass Box directly, not *encoder - wgpu_render_bundle_set_vertex_buffer: buffer_id → Some(buffer_id) - render_pass_set_vertex_buffer: buffer_id → Some(buffer_id) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Exposes per-format GPU texture capabilities (allowed usages + feature flags) that wgpu-core already queries via the HAL but previously had no C API surface. Adds WGPUNativeTextureFormatCapabilities struct and WGPUNativeTextureFormatFeatureFlags bitmask to wgpu.h. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
6 tasks
- Add 14 WGPUNativeTextureFormat_Astc*Sfloat codes (0x0003000A..0x00030017) starting after R64Uint (0x00030009) to avoid value collision - Add WGPUNativeFeature_VulkanExternalMemoryFd (0x00030041) - Extend WGPUNativeLimits with mesh shader and ray tracing limit fields - to_native_texture_format: expand ASTC HDR channel to 14 native codes - write_limits_struct: fill all new extended limit fields - map_required_limits: read all new extended limit fields Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds WGPUNativeFeature_VulkanExternalMemoryDmaBuf (0x00030042) to expose the Linux DMA-BUF external memory Vulkan feature through the C API. Maps to wgpu::Features::VULKAN_EXTERNAL_MEMORY_DMA_BUF in both features_to_native and feature_from_native. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
These formats exist in the standard WebGPU C header (WGPUTextureFormat_*) but were missing from wgpu-native's Rust conv.rs, causing a panic when wgpuAdapterGetTextureFormatCapabilities was called with these formats. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…rativeMatrixProperties
…ctureArray binding - ffi/wgpu.h: Add WGPUWgslLanguageFeatures bitflag typedef with 3 constants (ReadOnlyAndReadWriteStorageTextures, Packed4x8IntegerDotProduct, PointerCompositeAccess) - ffi/wgpu.h: Declare wgpuGetWgslLanguageFeatures() and wgpuSurfaceDiscardTexture() - ffi/wgpu.h: Add tlases/tlasCount array fields to WGPUBindGroupEntryExtras - src/lib.rs: Implement wgpuGetWgslLanguageFeatures via naga ImplementedLanguageExtension - src/lib.rs: Implement wgpuSurfaceDiscardTexture with double-discard guard - src/conv.rs: Handle AccelerationStructureArray in map_bind_group_entry Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- ffi/wgpu.h: Add WGPUDownlevelFlags bitmask, WGPUShaderModel enum, WGPUDownlevelCapabilities struct, and wgpuAdapterGetDownlevelCapabilities() - ffi/wgpu.h: Add WGPUAllocationReport, WGPUMemoryBlockReport, WGPUAllocatorReport structs with wgpuDeviceGetAllocatorReport() and wgpuAllocatorReportFreeMembers() - src/lib.rs: Implement wgpuAdapterGetDownlevelCapabilities via wgpu-core - src/lib.rs: Implement wgpuDeviceGetAllocatorReport and wgpuAllocatorReportFreeMembers; allocations array and name strings are heap-allocated and freed by the caller - src/conv.rs: Add map_downlevel_capabilities helper Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds NON_POWER_OF_TWO_MIPMAPPED_TEXTURES, INDEPENDENT_BLEND, DEPTH_TEXTURE_AND_BUFFER_COPIES, WEBGPU_TEXTURE_FORMAT_SUPPORT, BUFFER_BINDINGS_NOT_16_BYTE_ALIGNED, FULL_DRAW_INDEX_UINT32, VIEW_FORMATS, SURFACE_VIEW_FORMATS, NONBLOCKING_QUERY_RESOLVE, SHADER_F16_IN_F32, and MSL2_1 to WGPUDownlevelFlags and updates map_downlevel_capabilities to include them. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
map_bind_group_layout_entry previously only read the binding array count from the legacy WGPUBindGroupLayoutEntryExtras.count extension. The WebGPU spec now puts this value on WGPUBindGroupLayoutEntry.bindingArraySize directly. Fall through to bindingArraySize if the extension count is 0, so both old and new callers work correctly. Without this fix, callers using the standard field got count=None in the wgt::BindGroupLayoutEntry, causing Metal's bindless path to be skipped and naga MSL translation to fail with "mapping of ResourceBinding is missing". Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Previously, wgpuBlasPrepareCompactAsync just logged errors on failure (e.g. CompactionUnsupported) without pushing them to the device's validation error scope. Callers using push_error_scope/pop_error_scope to detect validation errors would see no error, causing tests to fail. Fix by adding error_sink to WGPUBlasImpl (propagated from the device at creation time) and calling handle_error() in wgpuBlasPrepareCompactAsync and wgpuQueueCompactBlas so errors reach the active error scope. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…ceCreateBlas - Handle WGPUBlasGeometryKind_AABBs by mapping aabbDescriptors to wgt::BlasAABBGeometrySizeDescriptor; previously panicked unconditionally on any non-triangle geometry kind. - Preserve the None/Some distinction for index_count that wgpu-core uses to detect mismatched index descriptors: - indexFormat=Undefined + indexCount>0 → (None, Some(count)) - indexFormat!=Undefined + indexCount=0 → (Some(format), None) Both cases hit wgpu-core's MissingIndexData validation error. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
1. Align BlasTriangleGeometrySizeDescriptor index mapping in
wgpuCommandEncoderBuildAccelerationStructures with wgpuDeviceCreateBlas.
The build path was using a simpler two-branch check that differed from
creation in two edge cases:
- indexFormat=Undefined, indexCount>0: build silently dropped the count
(gave (None,None)) while creation passed (None,Some(N)) to wgpu-core.
- indexFormat=defined, indexCount=0: build gave (Some(fmt),Some(0)) while
creation gave (Some(fmt),None), causing a descriptor mismatch in
wgpu-core validation at build time.
2. Fix aliased &T / &mut T UB in all three AdapterInfoExtras nextInChain
blocks (wgpuAdapterGetInfo, wgpuAdapterInfoFreeMembers,
wgpuDeviceGetAdapterInfo). The previous pattern called
`ptr.as_ref()` to check sType (creating a live shared reference) and
then immediately created `&mut *(ptr as *mut WGPUAdapterInfoExtras)` to
the same allocation — aliased &/&mut is UB under Rust's memory model.
Fixed by reading sType directly via raw pointer dereference
(`(*ptr).sType`) so no shared reference is live when the &mut is made.
3. Document that wgpuShaderModuleGetCompilationInfo always invokes its
callback synchronously regardless of callback_info.mode, consistent with
the same known limitation in wgpuBufferMapAsync. Changed the inline
`WGPUFuture { id: 0 }` literal to NULL_FUTURE for consistency.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This was referenced May 28, 2026
- Remove stale FIXME on wgpuDevicePoll (function was already reworked) - Fix misleading comment in create_shader_module_passthrough: WGSL routes through the validated path because wgpu-native has no unvalidated WGSL; binary formats use WGPUShaderModuleDescriptorPassthrough - Move map_cooperative_scalar_type and map_state_to_u32 from lib.rs into conv.rs where conversion helpers belong - Add comment above SetLabel no-ops explaining they are permanent because wgpu-core does not expose a set-label API for these resource types Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
device.rs - Replace 17× four-line label boilerplate (let label + let label_sv) with single conv::opt_str_to_string_view(desc.label) call (~51 lines saved) - Extract FragmentStateStorage struct and build_fragment_state() helper, eliminating ~80 lines of identical fragment state build code duplicated between create_render_pipeline and create_mesh_pipeline - Remove #[allow(unused_assignments)] from create_render_pipeline (no longer needed); keep it on create_mesh_pipeline for task stage with accurate comment - Trim Drop for CDevice comment: 5 lines → 2 - Merge two separate AccelerationStructure chain-pointer comments into one - Collapse "Wire chain pointers" comment blocks: 3 lines → 1 each - Trim set_device_lost_callback log message and preamble comment - Trim destroy() comment block: 9 lines → 2 (points to adapter.rs for detail) - Trim CQueue field doc comments: 3-line prose → 1-line each adapter.rs - Remove "Build the feature list" and "Build the limits structs" section comments (variable names are self-documenting) - Trim device_lost_cb block comment: 11 lines → 4 (remove CURRENT STATE / FORWARD COMPAT / KNOWN LIMITATION sub-headings, keep the core fact) - Merge two near-identical SAFETY comments on device_lost_ptr / handler_ptr into one comment before the struct literal - Remove "Capture adapter info before creating device" comment command.rs - Remove six obvious section comments: "Color attachments.", "Hole in color attachments array.", "Depth stencil attachment.", "Timestamp writes.", "Occlusion query set.", "Convert optional timestamp writes." — all restate the immediately following variable names - Remove "None means copy to end" comment on unwrap_or(u64::MAX) - Trim TLAS build comment: 2 lines → 1 - Replace 3× label boilerplate with conv::opt_str_to_string_view pass.rs - Add CBuffer to the use crate::resource import - Replace 15× crate::resource::CBuffer with CBuffer (CBuffer was imported everywhere else in the crate but missing from this file's use statement) lib.rs (wgpu-c-backend) - Replace println! with log::debug! (log is already a dependency) - Trim instance_factory comment: 4 lines → 1 - Trim panic propagation module comment: remove last 2 lines about spontaneous callbacks being "inherently racy" (design-doc content) - Extract not_found() local function to deduplicate the 13-line RequestAdapterError::NotFound struct literal constructed twice in request_adapter (once in cb callback, once as fallback) resource.rs - Trim CBuffer.is_mapped field comment: 4 lines → 1 - Trim map_async spin-poll comment: 6 lines → 2 - Trim get_mapped_range guard comment: 3 lines → 1 - Trim CBufferMappedRange.read_only field comment: 4 lines → 1 - Trim CQueueWriteBuffer module comment: 3 lines → 1 - Replace label boilerplate in CTexture::create_view with opt_str_to_string_view src/lib.rs - Fix misplaced doc comments on wgpuDevicePoll: /// lines must precede all attributes, not appear between #[no_mangle] and pub fn - Simplify wgpuDevicePoll wait branch: fold two PollType::Wait arms (differing only in submission_index: Some/None) into one using .copied(); replace if/else timeout block with (timeout_ns != 0).then(|| ...) (15 lines → 3) - Trim wgpuShaderModuleGetCompilationInfo TODO comment: 4 lines → 2 src/conv.rs - Rewrite WGPU_NATIVE_STORAGE_TEXTURE_ACCESS_ATOMIC and WGPU_NATIVE_TEXTURE_FORMAT_R64_UINT comments to lead with the non-obvious numeric value/range position rather than restating the constant name CHANGELOG.md - Tighten feature-parity paragraph: remove vague "and more. Additionally..." trailing sentence; replace with "Now tested against wgpu's test suite." wgpu-c-backend/readme.md + explanation.txt - Merge explanation.txt content into readme.md under a "Directory structure" section; delete explanation.txt (.txt file for doc content is non-standard) - Trim redundant opening sentence from readme.md description run-wgpu-tests.py - Collapse 3 print statements in copy_c_backend (Removing / Copying) into one "Syncing" line printed before the operation Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Address 21 doc/comment issues found in audit of PR gfx-rs#594: Wrong/misleading: - adapter.rs: rewrite SAFETY comment on callback userdata pointers; the previous text claimed the Boxes were "stored in CDevice" at the point they are registered, but they are still local stack allocations at that point — moved into CDevice only on the success path - device.rs: s/wgpu-core/wgpu-native/ in create_buffer validation comment; the validation happens at the wgpu-native layer, not wgpu-core directly - resource.rs: mark the null-return behaviour of wgpuBufferGetMappedRange for Read buffers as "observed, not a specified API contract" Redundant/inconsistent: - device.rs: remove 9 section-label comments (// Vertex state., // Primitive state. ×2, // Depth stencil. ×2, // Multisample. ×2, // Task stage (optional)., // Mesh stage (required).) that restate variable names - device.rs: unify the duplicate "Wire chain pointers" comment so both occurrences carry the Box stability explanation - pass.rs: align transition_resources no-op comment with the more informative command.rs version ("Metal, Vulkan, etc.") Missing documentation: - adapter.rs: expand Box-vs-Arc comment to explain that handler_ptr is the userdata1 raw pointer for the C callback - device.rs: add /// doc comment on set_device_lost_callback explaining that GPU-initiated loss never fires it (only Device::destroy() does) - device.rs: add comment to create_staging_buffer explaining the CPU staging model and cross-referencing CQueueWriteBuffer - device.rs: condense CQueue field docs for shared fields to cross- reference CDevice rather than duplicating the explanation - pass.rs: add End-in-Drop explanation to CComputePass and CRenderPass Drop impls (wgpu ends a pass by dropping the pass object) - resource.rs: expand spin-poll comment in map_async to explain that it semi-blocks the caller and what the 50 ms cap means - surface.rs: document the silent Fifo fallback in configure() when present_mode_to_native returns None Organisation: - src/lib.rs: move wgpuExternalTextureSetLabel into the grouped SetLabel no-op block; update the block comment from "most resource types" to "these resource types" Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Author
|
@Vipitis Do you have a discord or something? I'd like to be able to communicate with you about this since I think you will be one of the main users. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Checklist
cargo clippyreports no issuescargo docreports no issuescargo denyissues have been fixed or added todeny.tomlAdded new functionality. @githubnameDescription
This is this sister PR to gfx-rs/wgpu#9570. It brings wgpu-native into full parity with wgpu such that all features are supported, and all tests pass.
I cannot guarantee that everything is 100% right but I plan to look over every change and claude has also looked over every change, and like I mentioned it passes the full test suite of wgpu.
The only real breaking change is the removal of
wgpuDeviceCreateShaderModuleSpirVwhich has been replaced with wgpu's new passthrough API. It is easily replaced.Related Issues