gogpu/wgpu

Summary Software backend's SPIR-V interpreter silently returns zero for ~36 unimplemented GLSL.std.450 opcodes. No warning — silent data corruption on any platform using the software backend (servers, embedded, CI, containers, any device without GPU). Current state hal/software/shader/glsl_ext.go implements 47 of ~83 GLSL.std.450 opcodes. The default case at line 345 returns ValUint(0) without any diagnostic. A shader using determinant() in CI silently gets 0.0. Short-term fix Add slog.Warn to the default return — stop hiding the problem: default: hal.Logger().Warn("unimplemented GLSL.std.450 opcode", "opcode", op) return ValUint(0) Architecture: Interpreter + Go Native SIMD are complementary The SPIR-V interpreter (~14K LOC) and the planned naga Go Native SIMD backend serve different roles: SPIR-V Interpreter naga Go Native SIMD Input Any SPIR-V binary naga IR (compile-time) When Runtime (dynamic shaders) Build-time (pre-compiled) Debugging Breakpoints, JSON tr

Summary Expand WebGPU validation coverage from ~45% to ~70% of Rust wgpu-core checks. Currently 55 validation functions across core/validate.go (26), internal/raytracing/validate.go (17), and core/indirect_validation.go (12). Key gaps: feature gates, texture usage conflicts, render pass compatibility. Priority areas Feature gate checks (highest impact) None of the 25 feature flags are validated at usage time: FeatureTimestampQuery — not checked on CreateQuerySet FeatureMultiDrawIndirect — not checked on multi-draw calls FeaturePushConstants — not checked on PipelineLayout with push constants FeatureShaderF16 — not checked on shader module with f16 FeatureRayQuery — checked in internal/raytracing/ (9 checks), but NOT in core/ Texture usage conflicts Complete conflict matrix (CopySrc+CopyDst mutual exclusion in same pass) Storage texture read-write conflict detection Multisampled texture usage restrictions Render pass compatibility Color attachment format mismatch with pip

Summary Windows GLES has a mutex-protected AdapterContext (167 LOC) with Lock()/Unlock() + runtime.LockOSThread() for thread-safe GL context switching. Linux GLES lacks this abstraction — context management is split between Instance and Surface with no formal locking. Current state Windows (hal/gles/adapter_context.go): Single WGL context owned by Instance Lock() / LockForDC() / Unlock() with mutex + runtime.LockOSThread() Surface borrows context via Lock — lightweight Linux (hal/gles/api_linux.go): X11/headless: Instance creates EGL context, shares with Surface (ownsContext: false) Wayland: Surface creates own EGL context (ownsContext: true) — GL objects not shared across surfaces No mutex, no Lock/Unlock, no runtime.LockOSThread() What's needed adapter_context_linux.go with EGL Lock/Unlock pattern matching Windows Mutex + runtime.LockOSThread() for thread safety on Linux Wayland context-per-surface as documented intentional divergence Refactor api_linux.go Instance to use

Summary Pipeline cache reduces pipeline creation time on second launch by caching driver-compiled GPU ISA to disk. Currently we pass VK_NULL_HANDLE / null on both Vulkan and DX12 — all pipelines recompile from scratch every launch. Why this matters For 15-30 pipelines (typical GUI app): Scenario Cold start Current (no cache) ~150-300ms With pipeline cache ~3-5ms Pipeline cache is independent of shader compiler — DXIL, FXC, or SPIR-V all benefit. The cache saves driver-compiled GPU ISA (machine code), not shader bytecode. Vulkan VkPipelineCache API bindings fully generated in vk/commands_gen.go (lines 1497-1541): CreatePipelineCache, DestroyPipelineCache, GetPipelineCacheData, MergePipelineCaches. All loaded but unused — pipeline.go:324 passes 0 as cache handle. Rust wgpu has full Vulkan pipeline cache implementation (wgpu-hal/src/vulkan/device.rs:2448-2466). Work: Add pipelineCache field to Vulkan Device Create cache at device init, pass to all vkCreateGraphicsPipeli

Summary Device.CreateQuerySet and Device.CreateRenderBundleEncoder are fully implemented on all 6 HAL backends but have no public API wrapper. Users must access HAL directly to use these features. What's ready (all 6 backends) CreateQuerySet: Vulkan: hal/vulkan/query.go DX12: hal/dx12/query.go Metal: hal/metal/device.go GLES: hal/gles/device.go Software: hal/software/device.go (returns ErrTimestampsNotSupported) Noop: hal/noop/device.go (returns ErrTimestampsNotSupported) CreateRenderBundleEncoder: implemented on all 6 backends. What's missing Two methods in root package (device_native.go, ~30 LOC each): func (d *Device) CreateQuerySet(desc *QuerySetDescriptor) (*QuerySet, error) func (d *Device) CreateRenderBundleEncoder(desc *RenderBundleEncoderDescriptor) (*RenderBundleEncoder, error) Plus corresponding public types if not yet in the root package. Impact Without CreateQuerySet, users cannot do timestamp or occlusion queries through the public API. This is a gap for profiling and p