Fixed

• Memory leak: validation maps pinned every resource per frame — SetBindGroup copied every bound buffer/texture into encoder maps per draw call. Maps never cleared after Submit/Release, pinning BindGroup objects indefinitely. Now tracks bind groups only; validation walks boundBuffers/boundTextures transitively. Contributor: @samyfodil (#291)

• Validation error precedence — buffer/texture errors now deterministically beat bind group errors regardless of map iteration order. Matches Rust wgpu ordering.

Hotfix for v0.30.28 deadlock: Submit→Triage→onZero→LastSubmissionIndex→Queue.mu. Changed lastSubmissionIndex from mutex-protected uint64 to atomic.Uint64.

Fixed

• Resource lifecycle: BindGroup/Pipeline Release() bypasses ref-counting —
  Release() called dq.Defer(lastSubmissionIndex) directly, ignoring the
  ResourceRef ref-counting system. On shared encoder path, this caused
  use-after-free: HAL resource destroyed while GPU still processing commands.
  Now Release() calls ref.Drop() — HAL destruction deferred until ALL
  refs (user + GPU) are dropped. Matches Rust wgpu Arc<BindGroup> pattern.
  Applied to: BindGroup, RenderPipeline, ComputePipeline. (#287)

• DestroyQueue deadlock: Triage → onZero → Defer re-entry —
  Triage() held mutex while executing callbacks. When onZero fired and
  called Defer(), it tried to acquire the same mutex → deadlock. Now
  callbacks execute outside the lock. Same fix applied to FlushAll().

Tests

• 6 new lifecycle tests: refcount transitions, Release-before-Submit,
  Release-without-Clone, mixed-resource TrackedSubmission, buffer binding scenario

Fixed

• GLES: MappedAtCreation buffer data silently discarded on Unmap —
  UnmapBuffer only flushed shadow data to GL when BufferUsageMapWrite was set.
  Per WebGPU spec, MappedAtCreation does NOT require MapWrite usage. Buffers
  with Uniform|CopyDst + MappedAtCreation (the standard g3d pattern) had their
  data thrown away — zero MVP matrices, zero vertices, zero indices. Root cause of
  invisible 3D geometry on GLES. (#284)

• GLES: depth/stencil attachment point chosen by format —
  GL_DEPTH_ATTACHMENT for depth-only formats, GL_DEPTH_STENCIL_ATTACHMENT for
  combined. Wrong attachment point caused FBO incomplete (0x506). Rust wgpu-hal
  parity (command.rs:577-580). (#284)

• GLES: depth clear masked by stale pipeline state — ClearDepthCommand now
  calls glDepthMask(true) + glClearDepth(value) before glClear. Prevents
  stale DepthMask(false) from prior pipeline silently masking depth clear. (#284)

• GLES: viewport depth range ignored — SetViewportCommand now calls
  glDepthRange(minDepth, maxDepth). Was ignoring depth range fields entirely.
  Rust wgpu-hal parity (queue.rs:1295-1296). (#284)

Added

• gl.Context.ClearDepth() and gl.Context.DepthRange() wrapper methods for
  both Windows (syscall, double) and Linux (goffi, float32 for GLES).

Changed

• deps: gpucontext v0.22.0 → v0.23.0

Fixed

• GLES: depth/stencil not attached to swapchain FBO on surface render pass —
  setupSurfaceTarget() had an early return that skipped depth/stencil attachment.
  3D content with depth testing was invisible on GLES while Vulkan and DX12 worked
  correctly. Now attaches depth/stencil to the swapchain FBO via
  AttachDepthStencilToFBOCommand. Matches Rust wgpu-hal GLES begin_render_pass
  which uses the same draw_fbo path for both surface and offscreen targets. (#284)

Changed

• deps: gpucontext v0.22.0 → v0.23.0

[0.30.25] - 2026-07-29

Changed

• deps: gpucontext v0.21.1 → v0.22.0 — FontSmoothing type + coordinate space docs (gogpu#396, gogpu#398)

Fixed

• Browser/Rust Surface missing methods — PresentPixels, WritePixels,
  SetPrepareFrame, and SetPresentsWithTransaction now exist on all backends.
  Missing methods return clear errors or no-op, matching ADR-047 (unified public
  API). Fixes WASM build failure introduced by gogpu PR #370. (#281)

• Metal checkptr abort under -race — ObjC block callback trampolines used
  unsafe.Pointer(blockPtr + 32) which violates Go's pointer provenance rules.
  Replaced with unsafe.Add(unsafe.Pointer(blockPtr), 32) (Go 1.17+) at all 4
  block callback locations. go test -race now works on Metal compute. (#280)

Added

• Compile-time Surface API contract — anonymous interface assertion
  (var _ interface{...} = (*Surface)(nil)) in each backend file enforces that
  all public methods exist at compile time (ADR-047 enforcement).

• CI cross-compile job — GOOS=js GOARCH=wasm (browser) and -tags=rust
  (Rust FFI) builds added to CI pipeline. Missing Surface methods now fail the
  PR, not the downstream consumer.

CI

• Android SDK setup: use pre-installed SDK instead of third-party action (v3/v4 both crash on Preview license)
• goffi version check updated v0.6.1 → v0.6.2 in Android preview script
• DX12 command.go formatting fixed

What's Changed

Fixed

• Device teardown GPU drain ordering — Device.Release() drains GPU work before destroying staging buffers. Matches Rust wgpu Queue::Drop. (@besmpl #264)
• Vulkan swapchain fail-closed lifecycle — transactional reconfiguration, capability validation, broken-state tracking. (@besmpl #265)
• Explicit mock adapter construction — core.NewInstance no longer fabricates mock adapter. Rust wgpu parity. (@besmpl #266)
• Surface lifetime ownership — centralized acquisition/teardown with lease system, deterministic release ordering. (@besmpl #269)
• Vulkan nil pipeline layout guard — HAL rejects nil layouts before Vulkan call. (@besmpl #257)
• Metal autorelease pool thread pinning — LockOSThread prevents goroutine migration. Go 1.14+ preemption fix. (@besmpl #260)
• Metal texture array copy shape — setArrayLength for 1D/2D arrays instead of setDepth. Rust parity. (@besmpl #261)
• DX12 submission-ordered state reconciliation — command-local tracking, preamble barriers, depth/stencil planes. (@besmpl #262)
• nil WritePixels receiver — Surface.WritePixels on nil surface returns ErrReleased. (@besmpl #277)
• Metal storage-mode test fixture — aligned with isAppleGPU policy from v0.30.22. (@besmpl #275)

Added

• Surface-qualified adapter selection — RequestAdapterWithSurface via vkGetPhysicalDeviceSurfaceSupportKHR. Ahead of Rust wgpu. (@besmpl #267)
• Typed surface targets (Rust v29 parity) — CreateSurfaceFromTarget, CreateSurfaceUnsafe with opaque constructors. (@besmpl #273)
• Android arm64 Vulkan (guarded preview) — API 29+, Bionic loader via goffi v0.6.1. (@besmpl #268)
• Fixed-count indirect multi-draw — MultiDrawIndirect/MultiDrawIndexedIndirect with overflow-safe arithmetic. 2 improvements over Rust wgpu. (@besmpl #253)
• Metal ICB optimization — indirect command buffers for large indexed multi-draws. Original optimization beyond Rust. (@besmpl #263)
• Headless surface readback — HeadlessSurfaceTarget + Surface.ReadPixels() for golden image testing. (@besmpl #276)

Changed

• Bump naga v0.17.15 → v0.17.16 — MSL 64-bit atomics rejection (@besmpl)
• Bump goffi v0.6.1 → v0.6.2 — Windows AMD64 float returns fix (@besmpl)
• Bump webgpu v0.5.3 → v0.5.4 — Android surface, timestamp period, callbacks (@besmpl)
• DX12 lint cleanup — builtin shadow fix, struct returns, dead code removal
• CODEOWNERS — @besmpl added for Android paths
• ROADMAP + ARCHITECTURE — updated for 18 merged PRs
• CONTRIBUTING.md — Smart Coding framework, AI-assisted policy

Full Changelog: v0.30.22...v0.30.23

What's Changed

Fixed

• Metal MSAA texture storage mode crash on Intel Mac — use MTLGPUFamilyApple1 detection instead of hasUnifiedMemory for texture storage mode selection. Non-Apple GPU family devices (Intel/AMD) now always use MTLStorageModePrivate for textures. Apple Silicon keeps MTLStorageModeShared optimization for single-sample textures. Fixes #271.

Details

Intel integrated GPUs (e.g. Iris Plus 655) report hasUnifiedMemory=true but are NOT Apple GPU family. Metal spec requires MTLStorageModePrivate for multisample textures on non-Apple-family devices. Our previous code used hasUnifiedMemory to select MTLStorageModeShared for all textures, causing SIGABRT on the first frame when creating MSAA render targets.

The fix adds isAppleGPU detection via the existing DeviceSupportsFamily(device, MTLGPUFamilyApple1) infrastructure, which is the correct semantic check for storage mode capability. This preserves our MTLStorageModeShared optimization for single-sample textures on Apple Silicon (better than Rust wgpu-hal, which uses Private for all textures unconditionally).

Thanks to @AnyCPU for the detailed bug report with root cause analysis, patch, and regression test.

Full Changelog: v0.30.21...v0.30.22

What's Changed

Fixed

• Software CopyTextureToBuffer row stride — row-by-row copy respecting BytesPerRow from BufferLayout. Was flat copy(dst, src) ignoring row alignment, causing progressive row shift.

• Software configureRasterPipeline blend state — extract blend state from Fragment.Targets[0].Blend into the raster pipeline. Was always BlendDisabled, preventing premultiplied alpha compositing. Added convertBlendState, convertBlendFactor, convertBlendOp.

• Software readTexel BGRA format — swap R/B channels for BGRA8Unorm / BGRA8UnormSrgb textures. Replaced hardcoded format constants with gputypes imports.

Full Changelog: v0.30.20...v0.30.21