Repository navigation
Tags
Tags
release: v0.31.6 (gogpu#320) * chore: prepare release v0.31.6 - CHANGELOG: add v0.31.6 entry (PR gogpu#312 copy usage tracking, docs updates) - AGENTS.md: update version v0.30.22 → v0.31.6 and all dep versions - README.md: goffi v0.6.1 → v0.6.3 - ARCHITECTURE.md: update all dep versions to current - .gitignore: add .backup-investigation/ * fix: exclude nolintlint for browser convert files (CI lint blind spot) CI lint runs on ubuntu (GOOS=linux) — browser files with //go:build js,wasm are invisible. nolintlint flags //nolint:goconst as unused because goconst never runs on these files in CI. Same root cause as Metal lint blind spot (v0.31.3). Widened exclusion to cover all convert_*.go browser files. * fix: complete cross-platform lint exclusions (browser + rust backends) Extend golangci-lint exclusions for all platform-specific files invisible to CI lint runner (ubuntu, GOOS=linux). Known upstream bug (golangci/golangci-lint#3833). Browser (internal/browser/): revive var-naming (ref_ avoids Ref() collision), gocritic, nestif (JS interop nesting). Rust (*_rust.go): nolintlint for invisible //nolint directives. Browser mapped_range: nolintlint for //nolint:gosec. Verified: 0 issues on all 4 targets (Windows, Linux, macOS, WASM).
fix(browser): nil descriptor guards for all Device.Create* methods (g… …ogpu#316) * fix(browser): nil descriptor guards for all Device.Create* methods (gogpu#315) Browser WASM shim panicked on nil descriptors in CreateTextureView and other Create* methods. Native and Rust backends already handle nil correctly. This aligns browser behavior: - CreateTextureView/CreateSampler: nil desc → default (WebGPU spec) - CreateBuffer/Texture/ShaderModule/BindGroupLayout/PipelineLayout/ BindGroup/RenderPipeline/ComputePipeline: nil desc → error Reported by @darkliquid — blocks ironwail-go WASM HUD rendering. * docs: CHANGELOG v0.31.3 * fix(metal): resolve 10 macOS-only lint issues invisible to Linux CI CI lint runs on ubuntu-latest only — darwin build-tagged files were never checked. Fixed all 10 issues found by macOS cross-platform lint: - misspell: honour → honor (US locale per .golangci.yml) - nolintlint: remove 3 stale //nolint:nestif (nestif excluded for hal/ in config) - revive: rename max → maxCount (shadows Go 1.21 builtin) - unconvert: remove unnecessary uintptr(unsafe.Sizeof()) conversion - unparam: add //nolint:unparam for writeTextureShared (error return for API parity) - unused: remove dead test helper createMetalTextureCopyTestArray - whitespace: remove leading newline in applyBindGroup
feat(core): resource tracker + inline present barrier (ADR-060, gogpu… …#308, gogpu#307) (gogpu#310) * feat(core): TextureTracker + DeviceTracker foundation (ADR-060, gogpu#308) TextureTracker mirrors existing BufferTracker pattern (core/track/buffer.go): - TextureUses flags (10 states, Rust wgpu-types parity) - SkipBarrier: skip when state unchanged AND all ordered - TextureUsageScope: per-command-buffer usage collection with conflict detection - TextureTracker: device-level state tracking with Merge() barrier generation - DeviceTracker: wrapper with MergeTextureScope and TrackPresentTexture Removed pendingBufferBarriers/pendingTextureBarriers stubs from core/command.go (unused for 8 months). Added textureScope to CommandBufferMutable, DeviceTracker to Device. 1,273 LOC new code (386 tracker + 606 tests + 116 device + 165 device tests). 22 test functions, 174 test cases, 3 benchmarks. Zero lint issues. Submit-time wiring (queue_native.go) follows in next commit. * fix(vulkan): barrier submit signals present semaphore (ADR-060, gogpu#308) ensurePresentLayout now allocates a present semaphore via allocPresentSemaphore() and signals it from the barrier submit. present() reordered: ensurePresentLayout runs BEFORE presentWaitSemaphores so the barrier semaphore is included in vkQueuePresentKHR's wait list. This fixes a spec-correctness issue where the barrier relied solely on implicit queue ordering. The present engine now explicitly waits for the barrier to complete. TODOs added for full submit-time inline barrier injection via DeviceTracker — requires core TextureView → TrackerIndex wiring. * feat(vulkan): inline present barrier in EndEncoding — zero extra vkQueueSubmit (ADR-060, gogpu#308) Barrier (COLOR_ATTACHMENT_OPTIMAL → PRESENT_SRC_KHR) injected inside the user's command buffer before vkEndCommandBuffer. ensurePresentLayout() returns early (layout already PRESENT_SRC). Zero separate submit. Multi-submit handled: reverse barrier (PRESENT_SRC → COLOR_ATTACHMENT) in BeginRenderPass when prior submit transitioned to PRESENT_SRC but current pass needs COLOR_ATTACHMENT for LoadOp::Load. Single-submit: 1 vkQueueSubmit (was 2). Fixes FPS regression on Intel. Multi-submit: forward+reverse barriers inline, ensurePresentLayout as fallback for edge cases (blit-only paths). Also: moved maintidx exclusion to .golangci.yml HAL path, removed 10 redundant //nolint:maintidx directives across HAL backends. * feat(core): multi-CB encoder + textureScope population (gogpu#10, gogpu#11, ADR-060, gogpu#308) Multi-CB encoder (Rust InnerCommandEncoder parity): - cbList []hal.CommandBuffer on CoreCommandEncoder - OpenPass/CloseCB/CloseAndSwap/CloseAndPushFront/CloseIfOpen - Finish() returns ALL CBs, Submit() flattens via halBufferList() - 18 tests covering all CB ordering patterns textureScope population in BeginRenderPass: - Texture.trackingData wired to real TrackerIndexAllocator - TextureView.Parent references parent Texture for TrackerIndex - populateTextureScope() records COLOR_TARGET, DEPTH_STENCIL usage - Texture.Destroy() frees TrackerIndex - 12 tests covering allocation, freeing, reuse, scope population HAL interface unchanged. All backends unchanged. * feat: submit-time texture barrier injection via DeviceTracker (gogpu#12, ADR-060, gogpu#308) Submit() now merges each CB's textureScope into DeviceTracker, generates PendingTransitions, and prepends barrier CB to allBuffers. Data flow: BeginRenderPass → populateTextureScope → Finish → Submit → MergeTextureScope → BarrierCBFromTransitions → prepend. Wiring: public Texture holds coreTexture with TrackerIndex. SurfaceTexture propagates coreTexture. convertRenderPassDesc wires core.TextureView.Parent for TrackerIndex access. 18 new tests (6 core + 12 public API). Zero lint issues. * feat(core): wire BufferTracker + usage conflict validation (gogpu#13, gogpu#14, ADR-060, gogpu#308) BufferTracker (dormant 8 months) wired to Submit path: - Buffer.trackingData uses real TrackerIndexAllocator - bufferScope on CommandBufferMutable, populated for SetVertexBuffer, SetIndexBuffer, CopyBufferToBuffer/Texture - Submit merges buffer+texture scopes, generates combined barrier CB Usage conflict validation (WebGPU spec): - populateTextureScope returns error on conflict (COLOR_TARGET + RESOURCE) - RecordBufferUsage detects write+read conflicts (STORAGE_WRITE + UNIFORM) 15 new tests, 1 updated. Zero lint issues. * chore: cleanup stale comments and misleading docs (gogpu#15, ADR-060, gogpu#308) - core/resource.go: remove 'placeholder types' and 'CORE-006 stub' comments (types are fully defined with HAL integration, tracking is real) - interpreter.go: fix wrong 'control flow NOT implemented' comment (loops, conditionals, phi, switch all implemented since v0.27.0) - swapchain.go: update ensurePresentLayout doc to describe it as FALLBACK (inline barrier in EndEncoding handles common case) - pending_writes.go: update comment to reference ADR-060 tracker status * fix: cache swapchain core.Texture per-surface, not per-frame (gogpu#307, ADR-060, gogpu#308) GetCurrentTexture() was creating a new core.Texture with a new TrackerIndex on every frame. Nothing ever released them — leaked ~120K objects/min at 2000 FPS, causing 8GB+ memory growth (gogpu#307) and FPS drop from growing tracker scope slices. Fix: cache one core.Texture on Surface (swapchainTexture field). Created lazily on first GetCurrentTexture, reused across frames. Destroyed on Unconfigure/Configure/Release/retireDevice. Rust wgpu creates per-frame too but Arc<Texture> Drop auto-frees. Our cache is even better — zero per-frame allocation. * fix: barrier encoder stale submissionIndex + TrackerIndex leak + DiscardEncoding layout (ADR-060) Three bugs found by enterprise validation (3 parallel agents): 1. recordAndTrackAllBarriers used lastSubmissionIndex.Load() (previous submit's index) for barrier encoder Defer. Moved to postSubmit with actual subIdx — encoder recycled at correct time. 2. Texture.Release() did not free TrackerIndex from coreTexture — leaked for every user-created texture until device destroy. Now releases TrackingData on Release. 3. DiscardEncoding did not reset swapchain image layout tracking — stale COLOR_ATTACHMENT_OPTIMAL remained after discarded render pass while actual GPU layout is UNDEFINED. Now resets to UNDEFINED. * docs: CHANGELOG for ADR-060 core resource tracker (gogpu#308) * chore: bump gpucontext v0.26.0 + finalize CHANGELOG v0.31.0
fix(vulkan): align swapchain FinalLayout with Rust wgpu/Dawn (ADR-059) ( gogpu#304) Render passes for swapchain images now use COLOR_ATTACHMENT_OPTIMAL instead of PRESENT_SRC_KHR as FinalLayout. Explicit barrier to PRESENT_SRC_KHR in ensurePresentLayout() before present. Eliminated synchronous vkWaitForFences GPU stall — barrier submit uses no fence, command pool reset deferred to acquireNextImage() where acquire fence guarantees completion. Matches Rust wgpu (device.rs:119-120) and Dawn (RenderPassCache.cpp:173-174).
fix(vulkan): accumulated present semaphores for multi-submit sync (AD… …R-058) (gogpu#302) Per-image semaphore pool replaces single present semaphore. Each queue.Submit() signals a new semaphore from the pool, vkQueuePresentKHR waits on all accumulated semaphores. Fixes g3d#22 race on TBDR GPUs (Asahi Linux) where only the first submit was synchronized with present. Rust wgpu SwapchainPresentSemaphores parity (native.rs:743-843). Pools grow on demand, recycle after present — zero overhead for the common single-submit-per-frame case.
PreviousNext