Native GPU tools

Performance

Producer target

Use the maximized, focused stress launcher for six profiles: baseline,128 bodies,256/320 triangles,512/1280,512/5120,128 repeat. Timed samples include hitches; focus, maximization or window-size changes restart that interval. Game FPS, fresh native fronts and submitted scene Hz are independent measurements.

WGSL diagnostics

Compile a bounded source without executing it. Native Dawn messages include one-based lines and UTF16 columns/spans. Compilation is explicit, not triggered on every keystroke. One request at a time; max16,384 UTF16 units. This is a trusted developer tool, not an untrusted-shader security sandbox.

Pinned native GPU runtime required

Native capability report

Inspect the actual adapter, advertised features, WGSL language extensions and limits. Optional features are not enabled just by discovery. This does not create a device, run a shader or benchmark. After checking WGSL, the report also includes that device's negotiated features. CBN safety budgets are listed separately from hardware limits.

Inspect on demand · no automatic work

Bounded GPU safety check

Explicitly run three tiny compute dispatches: an in-bounds read, an out-of-bounds read with an unbound allocation guard, then valid recovery. No barriers after invalid access, unbounded loops or deliberate driver faults. This is a diagnostic, not conformance sign-off or an FPS test.

Idle · never runs automatically

Asynchronous API surface

Check all eight Promise operations and the device.lost getter using invalid receivers. Reports rejection types and deferred reactions without creating an adapter, device, shader or GPU workload. Explicit developer diagnostic only; not conformance or FPS sign-off.

Idle · never runs automatically

Conversion microtask timing

Explicitly check 30 getter/coercion failures across adapter, device, compute pipeline, render pipeline and buffer mapping. Each rejection must preserve the original value and precede the next Promise microtask, without waiting for a native tick. Creates or reuses the workbench's single device and one temporary16-byte buffer; does not compile or execute a shader. Not a performance or conformance certification.

Idle · never runs automatically

Error and listener diagnostics

Explicitly check native Event flags, primitive capture options, cancellation, passive listeners, AbortSignal removal and one real native validation notice. Reuses the workbench device and destroys its temporary invalid buffer. Reports the still-missing native GPUDevice EventTarget brand and trusted GPU event flag honestly. No shader execution, automatic polling, stress workload or conformance claim.

Idle · never runs automatically

Shader coverage

WGSL compute, GPU-written indirect commands, depth-tested meshes, shadows, persistent buffers, specialization constants, explicit multi-group layouts and dynamic uniform/storage offsets. Real Dawn async compute/render compilation and shader compilation info with owned backing and owner-thread Promise delivery. Kinetic startup checks134 override/143 indirect/296 layout/144 async/153 diagnostic assertions. B checks textures and samplers; N checks stencil operations, masks, read-only attachments, comparisons and core occlusion queries. M negotiates the actual native timestamp-query feature and profiles compute/render passes. Full WebIDL, standard mapped ranges, GLSL conversion and complete CTS coverage remain unfinished.

Reusable native rendering

Kinetic Lab records 17 native render bundles once for its mesh levels, ping-pong states and sky/floor. Live compute and indirect buffers still update; repeated frames execute immutable commands instead of re-encoding each mesh in JavaScript. The backend also supports up to eight nullable color slots, per-target blends, independent resolves and depth-only pipelines. These are bounded capabilities, not a full WebGPU compliance claim or a new FPS result.

Developer keys

C = Kinetic Lab, V = registry models, B = texture lab, N = stencil/query lab, M = native GPU timings, F9 = compute HUD, F10 = workbench. Remap in Minecraft Controls. Text fields accept these letters as text. Opening another view waits for real retirement of the previous one. Escape closes only the overlay.

Single development opt-in

This GPU-capable jar stages its checksum-verified runtime in the background, discovers the Windows system shader compiler and warms the native producer before the first lab keypress. A hidden Kinetic page compiles its pipelines while animation and input are parked, then is reused when C opens it. No SDK installation or three path arguments are required. Normal launches still use the lightweight software backend. GPU support requires compatible Windows x64 hardware/drivers; failed preparation is logged and software tools remain available. Backend selection must precede the first CBN view; live DLL swapping is not supported.

Native offscreen canvas

Explicitly render one tiny native GPU frame using the engine's genuine OffscreenCanvas. Preview through a native ImageBitmap and export one lossless PNG Blob. Extended mode preserves a native linear-sRGB half-float front; this preview/PNG is intentionally SDR. Owner-thread only: worker WebGPU, synchronous bitmap transfer and HDR display remain unavailable. Snapshot readback can wait for GPU completion; this is not a per-frame streaming or FPS test.

Idle · never runs automatically

Canvas lifecycle inspector

On demand: six tiny native clears across RGBA8, BGRA8 and half-float backings at two sizes, real bitmap/PNG snapshots, empty-bitmap handling, safe transfer rejection and a final recovery frame. Uses genuine DOM or owner-thread OffscreenCanvas, not worker simulation. Readbacks are explicit diagnostics, never a per-frame background task. No HDR, FPS or conformance certification.

Idle · never runs automatically