A deterministic GPU-kernel compiler, Naga AST validator, and CPU-oracle correctness certifier powered by etalon-wgsl-forge. It lowers typed KernelSpec + Schedule graphs into multi-backend shaders (WebGPU WGSL, Mobile WebGL2 GLSL ES 3.0, SPIR-V, HLSL, and MSL) without LLM heuristics.
Configure orthogonal tiling, workgroup geometry, vectorization width, and cooperative subgroup matrix acceleration below. Every parameter change synthesizes deterministic shader code (WGSL for WebGPU or #version 300 es GLSL for mobile WebGL2), evaluates the schedule verification contract via eval_program_with_receipt_wasm (forge.shader.*, gpu.compute.*, p64.runtime.*, nquin.wire.*), plots the operational intensity on the hardware Roofline model, and diffs a $16 \times 16$ output tile against the independent CPU reference oracle.
vibe:0.3/{forge,gpu,p64,nquin})
Evaluating schedule in vibe-wasm…
Waiting for receipt…
Tap/drag on the matrix grid to select active workgroup tile $(T_Y, T_X)$ · Right: Log-Log Roofline plot
Tap/hover cells to inspect $|\Delta| = |\text{GPU} - \text{CPU}|$ ($\max |\Delta| < 10^{-6}$) & live lowered shader
etalon-wgsl-forge Compiles & Certifies; the Engine Executes
Whether you are browsing high-level system design or auditing GPU execution safety, the core architectural separation is strict: QualiaDB owns domain semantics, policy admissibility (`ExactnessProfile` / `ADR 0003`), and weight containers (`P64`), while etalon-wgsl-forge owns deterministic shader lowering, Naga AST validation, and hardware-vs-CPU oracle certification:
etalon-wgsl-forge)Compiles GPU compute kernels from a typed mathematical specification (KernelSpec) and an orthogonal hardware execution schedule (Schedule), validates the generated shader AST via Naga, and certifies every kernel on real GPU silicon against an independent CPU reference oracle. Also transcodes GGUF/Safetensors weights into the 64-byte aligned p64 container.
qualia-core-db & etalon)The in-process runtime loads adapter-keyed certified shaders and p64 weight volumes (`{stem}.execution-profile.json`) to execute inference, graph sieves, and continuum physics. Sustained decode throughput is measured by the runtime engine (e.g., ~18.8 tok/s native Vulkan on RTX A2000 for SmolLM2-360M Q8; ~5.9 tok/s in-browser WebGPU).
KernelSpec + Schedule IR Across Target Backends
In etalon-wgsl-forge, a compute kernel is declared once as a pure mathematical DAG (KernelSpec). Tiling, workgroup dimensions, shared-memory staging, and vectorization width live in a separate Schedule descriptor so hardware tuning can never silently alter mathematical semantics. Every lowered shader is parsed and type-checked through Naga AST validation before emission to target backends:
Generated, Naga-validated, and oracle-certified kernel families include:
ETA-14) — Tiled GEMM, GEMV, transposed-B GEMM for native [out, in] weights, and double-single (df64) emulated 64-bit floating-point GEMM on GPUs lacking native FP64.etalon-signal and spatial HRTF audio.ETA-17) — Workgroup prefix scans, top-k selection, warp-transposed uvec4 GpuQuinSieve, and 8-pass LSD GpuRadixSorter in etalon-nquin.ETA-15) — Staggered Yee-grid 3D Maxwell curl updates (with CFL stability & discrete Poynting conservation checks) and 2D Shallow Water Equation (swe_solver) kernels in etalon-sim.ETA-14) — RMSNorm, rotary position embeddings (interleaved & NeoX RoPE), multi-head GQA attention with online softmax rescaling, and SwiGLU FFN.ETA-14, ETA-15, ETA-17)
Every generated kernel is dispatched on the target hardware adapter, read back with deterministic fencing, and compared element-by-element against an independent CPU reference oracle before its adapter-keyed receipt is sealed into QualiaDB:
A complete multi-head GQA transformer decode layer composed of etalon-wgsl-forge kernels is certified on real SmolLM2-360M weights (n_embd = 960, 15 Q-heads / 5 KV-heads, head_dim = 64, ffn = 2560, RoPE base 100k), matching the composed CPU reference oracle to a maximum relative error of 3.28×10⁻⁶ on RTX A2000 Vulkan.
3D FDTD electromagnetic and 2D SWE hydrodynamic kernels in etalon-sim enforce the Courant–Friedrichs–Lewy condition ($c\Delta t\sqrt{\Delta x^{-2}+\Delta y^{-2}+\Delta z^{-2}} \le 1$) prior to dispatch and verify discrete Poynting energy conservation against the 64-bit CPU reference integrator.
etalon-nquin certifies 48-byte NQuinWire GPU predicate sieves and 8-pass LSD radix sorts against the CPU reference implementation with ExactnessProfile::Exact (oracle_mismatches == 0) and 5-field XOR parity verification.
The Etalon silicon steward selects its GPU backend by capability intersection (f16, subgroup operations, timestamp queries) and records the active adapter in the execution profile; you can also pin a backend explicitly:
# Default is Vulkan (vendor-neutral across AMD / Intel / NVIDIA / ARM / Linux & Windows).
# Pin a backend explicitly:
QUALIA_WGPU_BACKEND=vulkan # default, full compute & graphics certification
QUALIA_WGPU_BACKEND=dx12 # Windows DirectX 12 hardware lane
QUALIA_WGPU_BACKEND=metal # Apple Silicon / macOS Metal lane
wasm32-unknown-unknown WGSL execution via etalon-wasm, and GLSL ES 3.0 mobile fallback.ADR 0003)
ExactnessProfile::Exact: Bitwise-deterministic integer, rational, symbolic, or 48-byte NQuin sieve/sort evaluation (oracle_mismatches == 0). Admissible for deontic, identity, and policy gates.ExactnessProfile::DeterministicApproximate { abs_bound, rel_bound }: Certified floating-point kernels with explicit finite error bounds verified against the CPU reference oracle.ExactnessProfile::InteractiveApproximate: Fast-math preview or stochastic rendering lanes — strictly rejected by admit_for_policy_decision() so visual approximations can never influence governance or rights decisions.P64 Weight & Profile Pairing: Converted models pair 64-byte aligned .p64 weight volumes (Q4_K_M / F16 / Q8_0 with per-tensor CRC-32C) with a sibling {stem}.execution-profile.json receipt.QualiaDB v0.0.41 • WGSL Forge • Implementation: etalon-wgsl-forge + qualia-core-db::wgsl_forge::etalon_bridge • CLI: qualia shader …