# wgpu Roadmap

> **Unified Go WebGPU Package — Three Implementations, One API**
>
> Native Go (5 HAL backends), Rust FFI (wgpu-native), Browser (WASM). Zero CGO.

---

## Vision

**wgpu** is a complete WebGPU implementation in Pure Go. No CGO required — single binary deployment on all platforms.

### Core Principles

1. **Pure Go** — No CGO, FFI via goffi library
2. **Multi-Backend** — Vulkan, Metal, DX12, GLES, Software
3. **WebGPU Spec** — Follow W3C WebGPU specification
4. **Production-Ready** — Tested on Intel, NVIDIA, AMD, Apple

---

## Current State: v0.30.23

✅ **Triple-backend architecture (ADR-038)** — Native Go, Rust FFI, Browser WASM via build tags
✅ **All 5 Native HAL backends complete** (~127K LOC)
✅ **Three-layer Native stack** — wgpu API → wgpu/core → wgpu/hal
✅ **Complete public API** — consumers never import `wgpu/hal`
✅ **Core validation layer** — 15/17 Rust wgpu-core checks
✅ **Text rendering on all 3 GPU backends** — Vulkan, DX12, GLES
✅ **DX12 TDR fixed** — deferred resource destruction + DRED diagnostics
✅ **PendingWrites batching** — Rust wgpu-core pattern for WriteBuffer/WriteTexture
✅ **Enterprise fence architecture** — HAL owns fences, SubmissionIndex tracking
✅ **Deferred resource destruction** — ResourceRef (Go Arc) + DestroyQueue (Rust LifetimeTracker). Buffer onZero callback (Rust Drop parity) — refcount-driven HAL destruction, no stale submission index risk.
✅ **Per-command-buffer resource tracking** — Clone/Drop in encoders (Rust EncoderInFlight)
✅ **DX12 HLSL shader cache** — in-memory SHA-256 keyed, LRU eviction
✅ **DX12 DRED diagnostics** — auto-breadcrumbs + page fault tracking on TDR
✅ **DX12 DXIL direct compilation** — naga DXIL backend, SM 6.0+, zero external dependencies, first Pure Go DXIL generator. Full 2D rendering (text, SDF shapes, widgets) verified on Intel Iris Xe.
✅ **Blend constant draw-time validation** — Rust wgpu-core OptionalState pattern
✅ **Vulkan fence pool recycling** — matches Rust wgpu-hal maintain() before submit
✅ **WebGPU Buffer mapping API** — `Buffer.Map`/`MapAsync`/`MappedRange`/`Unmap`, `Device.Poll`
✅ **GLES Y-flip fix** — swapchain offscreen FBO + Present blit (Rust wgpu-hal parity)
✅ **Encoder pool** — per-Device shared pool (Rust CommandAllocator pattern)
✅ **Software DWM-safe presentation** — CreateDIBSection+BitBlt (SDL3/Qt6 pattern)
✅ **Software Linux X11 presentation** — XPutImage via goffi (Skia pattern)
✅ **Vulkan maxMemoryAllocationSize enforcement** — prevents SIGSEGV on large writes
✅ **Staging belt auto-chunking** — writes >64MB automatically split (Rust wgpu parity)
✅ **Late buffer binding validation** — draw/dispatch-time MinBindingSize=0 checks (Rust parity)
✅ **Compute dispatch barriers** — per-dispatch memory barriers on all GPU backends (Rust parity)
✅ **Dispatch workgroup validation** — count + size limits checked before dispatch
✅ **Validation Phase A — crash prevention** — 18 P0 checks (WriteBuffer bounds, BindGroup buffer validation, draw-time state with typed sentinel errors, PipelineLayout count, format type guards, Queue.Submit resource state)
✅ **Validation Phase B — correctness** — MinBindingSize early check, index buffer format mismatch, indirect buffer bounds overrun, depth/stencil aspect granularity, BindGroup destruction at Submit. Coverage 22% → ~45% of Rust wgpu-core.
✅ **DX12 buffer state tracking** — per-buffer D3D12_RESOURCE_STATES with automatic transition barriers (Rust BufferTracker pattern)
✅ **Pipeline constants passthrough** — Constants map flows from public API through HAL
✅ **Zero-init workgroup memory** — WebGPU spec default, plumbed through all layers
✅ **CopyTextureToTexture public API** — DMA hardware copy with sub-region support
✅ **Vulkan relay semaphores** — GPU-side submission ordering (Mesa ANV workaround)
✅ **WASM platform split** — root package _native.go/_browser.go, core/hal excluded from WASM build
✅ **Vulkan command buffer free list** — batch alloc 16 CBs, pool reset (Khronos/NVIDIA/ARM/Mesa/Rust parity)
✅ **Damage-aware surface presentation** — `PresentWithDamage()` with compositor dirty rects. First WebGPU implementation. Software + Vulkan + DX12 + GLES backends.
✅ **Automatic resource lifecycle** — `runtime.AddCleanup` for Buffer/BindGroup (ADR-018, Rust Arc+Drop pattern). GC safety net prevents per-frame resource leaks.
✅ **Zero-allocation WriteBuffer batching** — pre-allocated BufferCopy + stack barrier arrays. All PendingWrites hot paths 0 allocs/op.
✅ **Full SPIR-V interpreter** — 7 phases (~10K LOC): vertex/fragment/compute on CPU, texture sampling, GLSL.std.450 intrinsics, control flow, atomics, workgroup shared memory. Shader debugger with breakpoints and JSON trace. For debugging/CI, not production.
✅ **DX12 timestamp queries** — CreateQuerySet, EndQuery, ResolveQueryData (Rust wgpu-hal parity)
✅ **Queue thread safety** — Submit/WriteBuffer/WriteTexture serialized via sync.Mutex (Rust wgpu-core parity)
✅ **GLES compute memory barriers** — glMemoryBarrier for storage→draw/dispatch transitions (Rust parity)
✅ **Software render pass instrumentation** — slog debug events + RenderPassStats for CI e2e assertions
✅ **Browser WebGPU backend** — complete `syscall/js` → `navigator.gpu` implementation (~6500 LOC). Instance, Adapter, Device, Resources, Pipelines, Command Recording, Queue Submit, Surface/Canvas, Buffer Mapping. Bypasses core/hal (Rust wgpu pattern). 97 TextureFormats, 31 VertexFormats, 29+ tests. Zero external dependencies.
✅ **GLES hidden window context (Windows)** — GL context owned by Instance on hidden 1×1 HWND, shared via mutex-protected `AdapterContext`. Adapter/Device/Queue survive Surface destruction. Follows Rust wgpu-hal `wgl.rs` `AdapterContext::lock()`/`lock_with_dc()` pattern. Surface lightweight — no context ownership.
✅ **Metal stencil state translation** — MTLDepthStencilState front/back face states, stencilAttachmentPixelFormat, ObjC descriptor leak fix (v0.30.4, @samyfodil)
✅ **Metal IOSurface leak fix** — retain/release balance, presentsWithTransaction, frameSemaphore, UMA StorageModeShared (v0.30.10, @lkmavi)
✅ **Software DrawIndexed** — index buffer resolution (uint16/uint32, baseVertex), vertex fetch integration (v0.30.5, @samyfodil)
✅ **Software R8/format-aware textures** — `formatBytesPerPixel` → `gputypes.BlockCopySize()` canonical (87 formats, Rust parity). Format-aware copy commands, Clear, readTexel (v0.30.5-v0.30.7)
✅ **Software MSAA honest reporting** — removed false Multisample/MultisampleResolve capabilities, CreateTexture rejects SampleCount>1 (v0.30.8)
✅ **Software BGRA swizzle on alpha blending** — framebuffer readback R↔B swap for Windows BGRA8 surfaces (v0.30.8)
✅ **Software buffer binding offsets** — CreateBindGroup stores Offset/Size, SetBindGroup accepts dynamicOffsets (v0.30.8)
✅ **GLES LockOSThread in integration tests** — EGL context thread affinity fix, eliminates CI flaky (v0.30.7)
✅ **TextureView.Texture() accessor** — Rust wgpu parity, all 3 build targets (v0.30.11)
✅ **SurfaceTexture.AsTexture()** — spec-compliant Queue.WriteTexture to surface textures (v0.30.11)
✅ **Software blit state validation** — Skia-inspired blitStateValid() checks blend/depth/mask/MSAA (v0.30.11)
✅ **Software PresentPixels extension** — atomic CPU pixel write + present. 3-layer (HAL→core→public), hal.PixelPresenter/PixelWriter interfaces. 3 copies → 1 copy. SDL3/Qt6/rio pattern. Non-standard WebGPU extension (v0.30.14)
✅ **GLES instance-level EGL context (Linux)** — surfaceless/pbuffer context at CreateInstance (Rust wgpu-hal `egl.rs:846` parity). Tiered config selection WINDOW+PBUFFER → PBUFFER-only (Rust `egl.rs:218-293`). Shared context in CreateSurface on X11/headless, own context on Wayland.
✅ **GLES FFI pointer convention fix (Linux)** — 30+ GL calls in `context_linux.go` fixed: pointer-type args corrected for goffi `CallFunction` (PR #210, @lkmavi). ADR-044 documents convention. CI test verifies GenBuffers/GenTextures through goffi.
✅ **GLES fence sync ordering** — `glFenceSync` before `glFlush` (was reversed). PollCompleted uses non-blocking `glGetSynciv`. Confirmed by ANGLE bug 6464, virglrenderer commit 21bbc9ea, Mesa Gallium. `Fence.Signal` returns error (Rust parity).
✅ **GLES GLSL version propagation (Rust parity)** — detected `GL_SHADING_LANGUAGE_VERSION` flows from adapter → device → naga GLSL writer. No more hardcoded `#version 430`. Runtime binding fallback (`glGetUniformBlockIndex` + `glUniformBlockBinding` + `glUniform1i`) for GL < 4.2 where `layout(binding=N)` unavailable. `shaderBindingLayout` capability flag (Rust `SHADER_BINDING_LAYOUT` parity). Triangle verified on WSL2 Mesa d3d12 GL 4.1.
✅ **Wayland SHM presentation (enterprise quality)** — display wrapper pattern (Qt6 parity), `wl_display_roundtrip_queue`, proper proxy cleanup, triple-buffer freeze fix (bufferBusyMap pointer + roundtrip_queue dispatch). Verified on WSL2.
✅ **Codecov OIDC** — replaced token + GPG verification with GitHub OIDC. No more intermittent CI failures from GPG keyserver.
✅ **PresentPixels check-before-mutate** — validate `hal.PixelPresenter` before discarding acquired texture. Rust wgpu validate-then-mutate pattern (v0.30.20)
✅ **DX12 UMA GPU classification** — `CheckFeatureSupport(D3D12_FEATURE_ARCHITECTURE)` replaces VRAM heuristic. Exact Rust wgpu parity. CacheCoherentUMA stored for future memory pool optimization (v0.30.20, @Zeroes1)
✅ **goffi v0.6.0 — ADR-049 three-tier FFI error handling** — 764 call sites migrated. Tier 1 (creation): check error. Tier 2 (void/hot path): infallible by GPU spec. Tier 3 (Wayland syscalls): errno diagnostics. Enterprise research: Rust wgpu-hal, Mesa, ash (v0.30.17)
✅ **vk-gen enterprise improvements** — C array params decay to pointers, deterministic output (sorted map keys), auto-gofmt via go/format, correct build tag (v0.30.18)
✅ **Software CopyTextureToBuffer row stride** — row-by-row copy respecting BytesPerRow (v0.30.21)
✅ **Software blend state wiring** — extract blend from Fragment.Targets into raster pipeline (v0.30.21)
✅ **Software BGRA readTexel** — R/B channel swap for BGRA8Unorm/Srgb textures (v0.30.21)
✅ **Metal MSAA storage mode fix** — `MTLGPUFamilyApple1` detection replaces `hasUnifiedMemory` for texture storage mode. Intel Mac SIGABRT on MSAA fixed. Apple Silicon keeps Shared optimization (v0.30.22, @AnyCPU #271)
✅ **Device teardown GPU drain** — private `waitIdle()` bypasses released guard, `queue.release()` after drain. Rust `Queue::Drop` parity (@besmpl #264)
✅ **Vulkan swapchain fail-closed** — transactional reconfiguration, capability snapshot validation, broken-state tracking, semaphore/fence error propagation (@besmpl #265)
✅ **Explicit mock adapter** — `NewInstance` no longer fabricates mock adapter. Rust wgpu parity (@besmpl #266)
✅ **Surface-qualified adapter selection** — `RequestAdapterWithSurface` validates via `vkGetPhysicalDeviceSurfaceSupportKHR`. Iterates all queue families — ahead of Rust wgpu (@besmpl #267)
✅ **Surface lifetime ownership** — centralized acquisition/teardown with lease system. Instance owns release ordering. Vulkan device-level swapchain tracking (@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 between pool create/drain. Go 1.14+ async preemption fix (@besmpl #260)
✅ **Metal texture array copy shape** — `setArrayLength` for 1D/2D arrays instead of `setDepth`. Rust wgpu `device.rs:436-453` parity. metalCopyPlan decomposition (@besmpl #261)
✅ **Fixed-count indirect multi-draw** — `drawCount` on DrawIndirect/DrawIndexedIndirect. 2 improvements over Rust: feature check for both draw kinds + offset advance in fallback loop (gfx-rs/wgpu#9870). `internal/indirect` overflow-safe arithmetic (@besmpl #253)
✅ **Metal ICB optimization** — indirect command buffers for large indexed multi-draws (1024-52428 commands). Original optimization beyond Rust wgpu (Rust uses simple loop). Deferred render encoder + compute dispatch (@besmpl #263)
✅ **DX12 submission-ordered state reconciliation** — command-local state tracker, preamble barriers at submit time, per-plane depth/stencil tracking, preamble pool, COM lifecycle safety (@besmpl #262)
✅ **Rust v29 typed surface targets** — `SurfaceTarget`/`SurfaceTargetUnsafe` with opaque constructors for Win32, Xlib, Wayland, Android NDK, Metal, Web. `hal.SurfaceTarget.RequireKind()` discriminator. Same contract across native/rust/browser (@besmpl #273)
✅ **Android arm64 Vulkan (guarded preview)** — API 29+, Bionic loader via goffi v0.6.1, Rust v29 WSI parity. Callbacks rejected pending physical-device proof (@besmpl #268)
✅ **Headless surface readback** — `HeadlessSurfaceTarget` + `Surface.ReadPixels()` for golden image testing. `hal.PixelReader` optional capability. Closes #256 (@besmpl #276)

### Remaining validation (planned)
- **Phase C** (P2): Spec compliance edge cases, feature gates
- Validation Phase C — spec compliance edge cases, feature gates

| Backend | Platform | Status |
|---------|----------|--------|
| Vulkan | Windows, Linux, macOS | ✅ Stable — text, compute, MSAA |
| Metal | macOS | ✅ Stable — naga MSL 91/91 |
| DX12 | Windows | ✅ Stable — TDR fixed, PendingWrites, deferred destruction |
| GLES | Windows, Linux | ✅ Stable — hidden window (Win), instance-level EGL (Linux), FFI pointer fix (@lkmavi), GLSL version propagation (GL 4.1+ verified), runtime binding fallback (GL <4.2), Wayland EGL surfaceless/pbuffer, tiered config (WindowBit-only for Wayland), fence sync, compute barriers |
| Software | Windows, Linux, macOS | ✅ Stable — windowed presentation (GDI/X11/Wayland SHM/CG+Metal), SPIR-V interpreter. Wayland: display wrapper pattern, roundtrip_queue, triple-buffer. |

→ **See [CHANGELOG.md](CHANGELOG.md) for detailed per-version notes**

---

## Roadmap

### Near-term (Q3 2026)

**Release v0.31.0** (after naga #82 merge):
- [ ] Bump deps: goffi v0.6.2, webgpu v0.5.4, naga v0.17.16
- [ ] CODEOWNERS: @besmpl for Android paths
- [ ] CHANGELOG for 17 merged PRs
- [ ] Integrate `gogpu/galloc` into Vulkan memory pools (replace BuddyAllocator)

**GLES Linux Enterprise Parity:**
- [ ] Shared AdapterContext for Linux — mutex + LockOSThread (FEAT-GLES-003). Windows parity.
- [ ] Systematic GL error checking layer — `checkGL()` after every call (ADR-046)

**Public API Quality (#218):**
- [ ] Remove HAL leaks from public API (`Surface.HAL()`, `Surface.SetPrepareFrame()`)
- [ ] Expose `Device.CreateQuerySet` — HAL ready on all 6 backends, needs public wrapper
- [ ] Document Surface single-thread requirement

**DX12:**
- [x] ~~DeviceTextureTracker~~ → submission-ordered state reconciliation (#262, done)
- [ ] DXIL as default shader path (currently opt-in via `GOGPU_DX12_DXIL=1`)
- [ ] Integrate galloc into DX12 descriptor heap management

### Mid-term (Q4 2026)

**API Stabilization:**
- [ ] Remove 31 type aliases + 46 const re-exports from `types.go` — users import `gputypes` directly
- [ ] Move `StencilOperation` from hal alias to gputypes definition
- [ ] `Device.CreateRenderBundleEncoder` — render bundle optimization
- [ ] `Device.GetTextureFormatFeatures` — per-format capability query
- [ ] GetSurfaceCapabilities on all backends (currently Vulkan-only)
- [ ] Validation Phase C — spec compliance edge cases, feature gates

**Platform Expansion:**
- [x] ~~Android~~ — Vulkan surface via typed `SurfaceTarget` (#268/#273, guarded preview)
- [ ] Android physical-device evidence (API 29 + API 36 arm64)
- [ ] **iOS** — Metal backend ready (naga MSL 91/91), needs platform integration

### v1.0.0 — Production Release (November 2027, Go 18th birthday)

Target: stable, documented, conformant WebGPU implementation in Pure Go.

- [ ] Full WebGPU specification compliance
- [ ] API stability guarantee — no breaking changes after v1.0
- [ ] Conformance test suite (WebGPU CTS subset)
- [ ] Comprehensive documentation (pkg.go.dev, examples, migration guide)
- [ ] Full render/compute pass validation (resource transitions)
- [ ] Clean public API — no HAL/core leaks, no type aliases, typed errors

### Future

**Platform:**
- [ ] **WebGL2 fallback** — GLES backend `_js.go` for browsers without WebGPU
- [ ] **Embedded Linux** — headless compute (Mesa surfaceless, Raspberry Pi)

**Advanced GPU Features:**
- [ ] Ray tracing extensions (VK_KHR_ray_tracing_pipeline)
- [ ] Bindless resources (descriptor indexing)
- [ ] Mesh shaders (VK_EXT_mesh_shader)
- [ ] Video decode/encode (VK_KHR_video_queue)

---

## Architecture

```
                    WebGPU API (core/)
                          │
          ┌───────────────┼───────────────┐
          │               │               │
          ▼               ▼               ▼
      Instance        Device           Queue
          │               │               │
          └───────────────┼───────────────┘
                          │
                   HAL Interface
                          │
     ┌──────┬──────┬──────┼──────┬──────┐
     ▼      ▼      ▼      ▼      ▼      ▼
  Vulkan  Metal   DX12   GLES  Software Noop
```

---

## Released Versions

| Version | Date | Highlights |
|---------|------|------------|
| **v0.30.22** | 2026-07 | Metal: MSAA storage mode crash on Intel Mac — `MTLGPUFamilyApple1` detection (#271) |
| **v0.30.21** | 2026-07 | Software: CopyTextureToBuffer row stride, blend state wiring, BGRA readTexel |
| **v0.30.20** | 2026-07 | PresentPixels check-before-mutate + DX12 UMA classification (@Zeroes1 #254) |
| **v0.30.17-19** | 2026-07 | goffi v0.6.0 ADR-049 (764 FFI call sites), vk-gen array params + deterministic output, webgpu v0.5.3 |
| **v0.30.15** | 2026-07 | PresentPixels/WritePixels 3-layer API, software blit debug cleanup |
| **v0.30.0** | 2026-06 | **BREAKING: Unified public API (ADR-047).** StencilOperation→gputypes, Device.released→atomic.Bool, sentinel methods, MinBindGroups, CopyBufferToTexture+ClearBuffer native, browser fence stubs, browser-compute example. |
| **v0.29.16** | 2026-06 | HAL wrapper stubs for Rust/Browser builds — public API compiles on all build targets. README updated. |
| **v0.29.15** | 2026-06 | naga v0.17.15 (HLSL sampler fix, @georgebuilds), x/sys v0.46.0. |
| **v0.29.14** | 2026-06 | GLES: missing vertexBuffers in Linux pipeline — one-line fix enables full gg rendering (SDF, text, widgets) on GLES Linux. @lkmavi. |
| **v0.29.13** | 2026-06 | **GLES enterprise parity**: GLSL version propagation, runtime binding fallback (GL <4.2), MSAA validation, GLES function names, lazy VAO, compute VERTEX_ATTRIB barrier, Wayland EGL fixes. First triangle on GLES WSL2 Wayland. Credits: @lkmavi (PR #210, #215). |
| **v0.29.12** | 2026-06 | GLES: Wayland EGL nativeDisplay=0 surfaceless fallback. CreateSurface Path A guard. |
| **v0.29.11** | 2026-06 | GLES: shared context + tiered config + CI GL test. Completes ADR-045 enterprise parity. |
| **v0.29.10** | 2026-06 | GLES: instance-level EGL context (Linux), surfaceless fallback, FFI pointer fix (PR #210, @lkmavi). |
| **v0.29.9** | 2026-06 | Software: Wayland SHM triple-buffer freeze fix (bufferBusyMap + roundtrip_queue). |
| **v0.29.8** | 2026-06 | Software: Wayland display wrapper pattern — queue mismatch fix. |
| **v0.29.7** | 2026-06 | Software: eager Wayland wl_shm init in Configure (SIGSEGV fix, gogpu#292). |
| **v0.29.6** | 2026-06 | GLES: Fence.Signal returns error on glFenceSync failure (Rust parity). |
| **v0.29.5** | 2026-06 | GLES: fence sync ordering — glFenceSync before glFlush. Confirmed by ANGLE/virglrenderer/Mesa. |
| **v0.29.4** | 2026-06 | Software: dedicated wl_event_queue for SHM buffer release dispatch. |
| **v0.29.3** | 2026-06 | Software: Wayland SHM present (enterprise). GLES: compute/instancing/topology fixes. |
| **v0.29.2** | 2026-06 | Vulkan: VK_ERROR_SURFACE_LOST_KHR (Rust parity). goffi v0.5.3. |
| **v0.29.1** | 2026-05 | go-webgpu/webgpu v0.5.2. |
| **v0.29.0** | 2026-05 | ADR-038: Rust backend at public API level (browser pattern). |
| **v0.28.11** | 2026-05 | Core: direct-write trackedRefs (pass → encoder, zero intermediate copies). |
| **v0.28.10** | 2026-05 | Core: pre-allocate trackedRefs in pass encoders (ML OOM fix). |
| **v0.28.9** | 2026-05 | Core: refcount-driven Buffer destruction via onZero (Rust Drop parity). Eliminates Phase 1 stale index risk. |
| **v0.28.8** | 2026-05 | Core: Clone buffer ResourceRefs in SetBindGroup — prevents use-after-free (Rust wgpu parity). |
| **v0.28.7** | 2026-05 | Core: deferred GLES enumeration in RequestAdapter (adapter name fix). |
| **v0.28.6** | 2026-05 | **GLES hidden window context** (Rust parity). Instance-owned GL context, AdapterContext mutex, Surface lightweight. Fixes context death on window close. |
| **v0.28.5** | 2026-05 | Metal: defer pool.Drain, drawable count. Core: indirect validation nil guard (#189). |
| **v0.28.4** | 2026-05 | macOS blit (@k-chimi), GLES Rust parity (fence+copies+timestamps+adapter), GPU dispatch indirect validation. |
| **v0.28.0** | 2026-05 | **Browser WebGPU backend** (WASM-001). Complete `syscall/js` → `navigator.gpu`. 6500 LOC, 5 phases, zero deps. First Pure Go WebGPU in the browser. |
| **v0.27.5** | 2026-05 | Defensive NULL handle guard in TransitionTextures/Buffers (Vulkan, DX12, public API). Prevents crash on destroyed resource barriers. |
| **v0.27.4** | 2026-05 | goffi v0.5.1 (struct ABI, XMM return, CGO_ENABLED=1), x/sys v0.44.0, flaky TestThread_CallAsync fix |
| **v0.27.3** | 2026-05 | Software render pass instrumentation (slog + RenderPassStats), Metal MsgSend docs |
| **v0.27.2** | 2026-05 | DX12 timestamp queries, Queue mutex, GLES compute barriers, Vulkan timestampPeriod fix |
| **v0.27.1** | 2026-05 | MSAA resolve LoadOp=CLEAR, Vulkan offscreen ImageLayoutGeneral, persistent stencil, naga v0.17.13 |
| **v0.27.0** | 2026-05 | **Full SPIR-V interpreter** (7 phases, ~10K LOC), shader debugger, compute HAL, particles rendering, tagged union optimization, naga v0.17.11, flaky test fix |
| **v0.26.12** | 2026-05 | **Test coverage** (core 85.5%, root 78.4%), Metal entry point fix (#168 by @k-chimi), naga v0.17.10 |
| **v0.26.11** | 2026-04 | **DX12 indirect dispatch/draw** — ExecuteIndirect + CommandSignature (was last GPU backend with stubs) |
| **v0.26.10** | 2026-04 | **Validation Phase B** — 5 P1 correctness checks (MinBindingSize, index format mismatch, indirect bounds, depth/stencil aspects, BindGroup Submit). Coverage 37% → 45%. |
| **v0.26.9** | 2026-04 | **Validation Phase A** — 18 P0 crash prevention checks (WriteBuffer bounds, BindGroup buffer, draw-time state, PipelineLayout, format guards, Submit resource state). Coverage 22% → 37%. |
| **v0.25.6** | 2026-04 | Vulkan command buffer free list (VK-CMD-001): batch alloc, pool reset, enterprise parity |
| **v0.25.5** | 2026-04 | WASM platform split (Phase 0): _native.go/_browser.go file split, core/hal excluded from WASM |
| **v0.25.4** | 2026-04 | Late buffer binding validation (VAL-006) + Vulkan relay semaphores (VK-SYNC-001) |
| **v0.25.3** | 2026-04 | Fix SIGSEGV on large WriteBuffer (#142): maxMemoryAllocationSize enforcement, staging belt auto-chunking, MappedSize bounds checking |
| **v0.25.2** | 2026-04 | gputypes v0.5.0 (PrimitiveState zero defaults) |
| **v0.25.1** | 2026-04 | Linux X11 software presentation via XPutImage (Skia pattern) |
| **v0.25.0** | 2026-04 | **WebGPU Buffer mapping API**, **DXIL full rendering** (naga v0.17.4), GLES Y-flip fix, sampler heap plumbing, pipeline error logging. Breaking: `Queue.ReadBuffer` removed. |
| **v0.24.7** | 2026-04 | DWM-safe software presentation (CreateDIBSection+BitBlt), rendering optimizations |
| **v0.23.8** | 2026-04 | Metal vertex buffer fix, GLES per-type binding counters, StagingBelt alignment |
| **v0.23.7** | 2026-04 | naga v0.16.4 (HLSL 72/72 parity, 330× faster FXC array init) |
| **v0.23.6** | 2026-04 | Deferred resource destruction, DX12 shader cache, DRED diagnostics |
| **v0.23.5** | 2026-04 | GLES coordinate space, Vulkan fence recycling, blend constant validation |
| **v0.23.4** | 2026-04 | GLES text fix, DX12 TDR (descriptor UAF), StagingBelt |
| **v0.23.3** | 2026-04 | GLES blur fix, enterprise logging system |
| **v0.23.2** | 2026-04 | DX12 sampler heap (Rust pattern), GLES BindingMap |
| **v0.23.1** | 2026-04 | Text/texture rendering on all non-Vulkan backends |
| **v0.23.0** | 2026-03 | Enterprise fence architecture, naga v0.15.0 |
| **v0.22.2** | 2026-03 | Metal per-type slots, GLES scissor, goffi v0.5.0 |
| **v0.22.1** | 2026-03 | Vulkan/GLES/DX12 patches |
| **v0.21.3** | 2026-03 | Validation layer + GLES/DX12/Software fixes |
| **v0.21.0** | 2026-03 | wgpu public API migration |
| **v0.18.0** | 2026-02 | Public API root package (20 types, WebGPU-aligned) |
| v0.17.1 | 2026-02 | Metal MSAA texture view crash fix |
| v0.17.0 | 2026-02 | Wayland Vulkan surface creation |
| **v0.16.16** | 2026-02 | Vulkan X11/macOS surface pointer fix (gogpu#106) |
| v0.16.15 | 2026-02 | Software backend always compiled, no build tags (gogpu#106) |
| v0.16.14 | 2026-02 | Vulkan null surface handle guard (gogpu#106), naga v0.14.3 |
| v0.16.13 | 2026-02 | Vulkan: debug_utils via GetInstanceProcAddr (gogpu#98) |
| v0.16.12 | 2026-02 | Vulkan debug object naming (VK-VAL-002, gogpu#98) |
| v0.16.11 | 2026-02 | Vulkan zero-extent swapchain fix (VK-VAL-001, gogpu#98) |
| v0.16.10 | 2026-02 | Vulkan pre-acquire semaphore wait (VK-IMPL-004) |
| v0.16.6 | 2026-02 | Metal debug logging (23 log points), goffi v0.3.9 |
| v0.16.5 | 2026-02 | Vulkan per-encoder command pools |
| v0.16.4 | 2026-02 | Timeline semaphore, FencePool, batch alloc, hot-path benchmarks |
| v0.16.3 | 2026-02 | Per-frame fence tracking, GLES VSync, WaitIdle interface |
| v0.16.2 | 2026-02 | Metal autorelease pool LIFO fix (macOS Tahoe crash) |
| v0.16.1 | 2026-02 | Vulkan framebuffer cache invalidation fix |
| v0.16.0 | 2026-02 | Full GLES pipeline, structured logging, MSAA, Metal/DX12 features |
| v0.15.1 | 2026-02 | DX12 WriteBuffer/WriteTexture fix, shader pipeline fix |
| v0.15.0 | 2026-02 | ReadBuffer for compute shader readback |
| v0.14.0 | 2026-02 | Leak detection, error scopes, thread safety |
| v0.13.x | 2026-02 | Format capabilities, render bundles, naga v0.11.1 |
| v0.12.0 | 2026-01 | BufferRowLength fix, NativeHandle, WriteBuffer |
| v0.11.x | 2026-01 | gputypes migration, webgpu.h compliance |
| v0.10.x | 2026-01 | HAL integration, multi-thread architecture |
| v0.9.x | 2026-01 | Vulkan fixes (Intel, features, limits) |
| v0.8.x | 2025-12 | DX12 backend, 5 HAL backends complete |
| v0.7.x | 2025-12 | Metal shader pipeline (WGSL→MSL) |
| v0.6.0 | 2025-12 | Metal backend |
| v0.5.0 | 2025-12 | Software rasterization |
| v0.4.0 | 2025-12 | Vulkan + GLES backends |
| v0.1-3 | 2025-10 | Core types, validation, HAL interface |

→ **See [CHANGELOG.md](CHANGELOG.md) for detailed release notes**

---

## Contributing

We welcome contributions! Priority areas:

1. **Compute Shaders** — Full compute pipeline support
2. **WebAssembly** — Browser WebGPU bindings
3. **Mobile** — Android and iOS support
4. **Performance** — Optimization and benchmarks

See [CONTRIBUTING.md](CONTRIBUTING.md) for guidelines.

---

## Non-Goals

- **Game engine** — See gogpu/gogpu
- **2D graphics** — See gogpu/gg
- **GUI toolkit** — See gogpu/ui (planned)

---

## License

MIT License — see [LICENSE](LICENSE) for details.