Historical task source. The Eve task board holds current task state.
Phase 169: Neith GPU Compute Shader Framework — Vulkan Compute / Metal Compute / DirectX 12 Compute / WebGPU Compute / CUDA / HIP / OpenCL / SYCL Unified Runtime#
Closes the compute-shader sovereignty gap. Phase 40.2 briefly mentions compute passes. Phase 95 covers GPU training clusters (ML side). Phase 157 + 161 rely heavily on GPU compute but assume the primitive exists. No phase delivers a first-class cross-backend compute-shader framework with wave intrinsics, shared memory, async compute, cooperative matrix ops, indirect dispatch, and language / reflection tooling that every modern engine, simulation, ML inference path, and procedural asset pipeline depends on.
Sovereign replacement for Vulkan / Metal / D3D12 / WebGPU compute runtimes, NVIDIA CUDA, AMD HIP, Khronos OpenCL, Intel oneAPI / SYCL, Apple Metal Performance Shaders, Microsoft DirectML (runtime portion), and shader-reflection tooling (SPIRV-Cross, DXC).
169.1 @neith/compute-abstraction — Backend Abstraction#
- 169.1.1 Single API (C / C++ / Rust bindings) over all backends
- 169.1.2 Vulkan compute backend — the provider contract, SPIR-V
dispatch protocol and validation are complete; the residual is the live
device capture, and the Linux server HAS a live Vulkan device (Mesa
lavapipe, software). Install
mesa-vulkan-drivers/vulkan-toolsand confirmvulkaninfo --summarylistsllvmpipe; park only with a named failed install attempt. Wire the real Vulkan provider inlibs/neith/compute-abstraction/crates/neith-compute-abstractionand capture a dispatch. Verify: a spec dispatches a known compute kernel (e.g. a prefix-sum) on lavapipe and the read-back result matches the CPU reference. - 169.1.3 Metal 3+ compute backend — the AIR provider contract is
complete; the residual is a live Metal device capture, and the MacBook Air
IS a Metal 3 device. Wire the real Metal provider in
libs/neith/compute-abstraction/crates/neith-compute-abstractionand capture a dispatch on the Mac (the earlier "requires a live provider/device" note is satisfied by the Mac). Verify: the same known kernel dispatches on the Mac's GPU and matches the CPU reference. - 169.1.4 D3D12 compute backend — DXIL provider contract complete; the
residual is a live D3D12 device capture. No Windows host in the fleet;
attempt it on a GitHub-hosted
windows-latestrunner using the WARP software D3D12 adapter (record the workflow/run URL), which can dispatch a DXIL compute kernel; park only with a named failed attempt. Verify: the known compute kernel dispatches through WARP D3D12 on the runner and the read-back result matches the CPU reference. - 169.1.5 WebGPU compute (Dawn / wgpu-class) backend — the WGSL provider
contract is complete; the residual is a live device capture, attemptable
with
wgpuon lavapipe (Linux) or Metal (Mac). Wire the real WebGPU provider inlibs/neith/compute-abstraction/crates/neith-compute-abstraction. Verify: the known kernel dispatches through wgpu on lavapipe and matches the CPU reference. - 169.1.6 CUDA backend (NVIDIA path) — PTX provider contract complete;
the residual is a live CUDA device capture. Needs an NVIDIA GPU; the fleet
has none (stated 2026-09-18).
blocked:hardware - 169.1.7 HIP backend (AMD path) — AMDGPU code-object provider contract
complete; the residual is a live ROCm/HIP device capture. Needs an AMD
ROCm GPU; the fleet has none (stated 2026-09-18).
blocked:hardware - 169.1.8 SYCL / oneAPI backend (Intel path) — SPIR-V/Level-Zero
provider contract complete; the residual is a live SYCL/oneAPI device
capture. Needs an Intel GPU with the Level Zero runtime; the fleet has
none (the Vulkan/WebGPU/Metal backends already give software/CPU and Mac
coverage) (stated 2026-09-18).
blocked:hardware - 169.1.9 CPU reference backend
- 169.1.10 Capability flags (FP16 / FP64 / wave ops / subgroup / coop-matrix)
- 169.1.11 Device enumeration + best-device heuristic
- 169.1.12 Multi-GPU context
- 169.1.13 Feature negotiation with graceful fall-back
- 169.1.14 Driver-version allowlist
- 169.1.15 Diagnostic logging of chosen path
169.2 @neith/compute-lang — Shader Language & Compiler#
- 169.2.1 Sovereign shader IR (stable format)
- 169.2.2 WGSL front-end
- 169.2.3 HLSL / SM 6.x front-end
- 169.2.4 GLSL 4.6 / Vulkan-GLSL front-end
- 169.2.5 MSL / Slang transcompile
- 169.2.6 Rust-GPU compute kernels
- 169.2.7 SPIR-V target
- 169.2.8 DXIL target
- 169.2.9 AIR (Apple) target
- 169.2.10 WGSL target
- 169.2.11 PTX / SASS target (CUDA)
- 169.2.12 GCN / RDNA / GCN2 ISA diagnostic
- 169.2.13 Optimisation passes (SSA, DCE, loop unroll, vectorise)
- 169.2.14 Reflection (bindings / resources / constants)
- 169.2.15 Shader-hash caching + PSO pre-build
169.3 @neith/compute-resources — Buffers, Textures & Bindings#
- 169.3.1 Typed / untyped storage buffer
- 169.3.2 Structured buffer with stride
- 169.3.3 Append / counter buffer
- 169.3.4 Storage texture 1D / 2D / 3D / array / cube
- 169.3.5 Sampler (linear / nearest / filtering)
- 169.3.6 Descriptor heap / bindless resources
- 169.3.7 Push constants / root constants
- 169.3.8 Uniform / constant buffer
- 169.3.9 Indirect dispatch buffer
- 169.3.10 Raytracing acceleration structures
- 169.3.11 Shared memory (groupshared / threadgroup)
- 169.3.12 Cooperative matrix (WMMA / MFMA) types
- 169.3.13 Half / double / int8 / int16 types
- 169.3.14 Persistent mapping
- 169.3.15 Sparse / tiled resource support
169.4 @neith/compute-dispatch — Dispatch Model#
- 169.4.1 Synchronous
dispatch(x, y, z) - 169.4.2 Indirect dispatch from buffer
- 169.4.3 Mesh-shader compute (when supported)
- 169.4.4 Ray-tracing dispatch
- 169.4.5 Async compute queue
- 169.4.6 Copy queue separation
- 169.4.7 Pipeline barriers with fine-grained stage masks
- 169.4.8 Fence / event cross-queue sync
- 169.4.9 Multi-queue topology (compute + graphics + copy)
- 169.4.10 Timeline semaphores
- 169.4.11 Query pool (timestamp / occlusion)
- 169.4.12 Transient resource aliasing
- 169.4.13 Workload-priority hints
- 169.4.14 Per-dispatch profiling capture
- 169.4.15 Automatic batching of small dispatches
169.5 @neith/compute-wave — Wave / Subgroup / Cooperative Operations#
- 169.5.1 Wave vote (all / any / ballot)
- 169.5.2 Wave broadcast
- 169.5.3 Wave reduction (sum / min / max / and / or / xor)
- 169.5.4 Wave prefix-scan
- 169.5.5 Shuffle / swizzle / rotate
- 169.5.6 Cooperative matrix multiply-accumulate
- 169.5.7 FP8 / FP16 / BF16 cooperative-matrix support
- 169.5.8 Tile / work-group cooperative intrinsics
- 169.5.9 Barrier (memory / control)
- 169.5.10 Fence semantics (acquire / release)
- 169.5.11 Atomic ops (add / cmpxchg / float atomics)
- 169.5.12 Ballot → compaction helper
- 169.5.13 Warp / wave helper macros per backend
- 169.5.14 Wait / yield instructions
- 169.5.15 Feature-detect portable fallback
169.6 @neith/compute-memory — Memory Model & Allocators#
- 169.6.1 Device-local / host-visible / uncached writeable regions
- 169.6.2 Staging buffer ring
- 169.6.3 Paged virtual memory for large buffers
- 169.6.4 Bump allocator for transient
- 169.6.5 Pool allocator per-size class
- 169.6.6 VMA-class integration (AMD-VMA / D3D12MA)
- 169.6.7 Cross-queue resource aliasing
- 169.6.8 Pinned / pageable host memory
- 169.6.9 Unified memory (UMA) path
- 169.6.10 Memory leak detector
- 169.6.11 Budget alarm
- 169.6.12 Defragmentation pass
- 169.6.13 Memory priority hints (GPU)
- 169.6.14 Lazy allocation (on-first-use)
- 169.6.15 Sparse-residency management
169.7 @neith/compute-ml — ML Inference & Tensor Ops#
- 169.7.1 Tensor primitive (rank / dtype / layout)
- 169.7.2 Gemm / Conv2D / DepthwiseConv / Attention kernels
- 169.7.3 Fused activations (ReLU / GELU / SiLU / tanh)
- 169.7.4 LayerNorm / RMSNorm / BatchNorm
- 169.7.5 Softmax / LogSumExp
- 169.7.6 Flash-attention-class kernel
- 169.7.7 Quantisation (INT8 / INT4 / FP8) kernels
- 169.7.8 ONNX / GGUF runtime importer (Phase 93 reuse)
- 169.7.9 Graph optimiser (fusion / folding)
- 169.7.10 Profiling (op-level)
- 169.7.11 Auto-schedule / auto-tune
- 169.7.12 Determinism toggle
- 169.7.13 Training gradient kernels (for on-device)
- 169.7.14 Distributed (ring-allreduce / NCCL / sovereign RCCL)
- 169.7.15 Bridge to Phase 47 Nous runtime
169.8 @neith/compute-sim — Simulation Kernels Library#
- 169.8.1 N-body (direct + tree-code)
- 169.8.2 SPH / FLIP / APIC solver building blocks
- 169.8.3 PCG / multigrid pressure solver
- 169.8.4 XPBD constraint solver
- 169.8.5 FFT (real / complex, 1D / 2D / 3D)
- 169.8.6 Radix / merge sort
- 169.8.7 Prefix sum / scan primitives
- 169.8.8 Hash grid + spatial partition
- 169.8.9 BVH build + refit
- 169.8.10 Raymarch SDF
- 169.8.11 Marching cubes / dual contour
- 169.8.12 Voxelisation
- 169.8.13 Mesh boolean CSG
- 169.8.14 Procedural noise (Perlin / Simplex / curl / worley)
- 169.8.15 Cross-backend performance conformance
169.9 @neith/compute-render — Render Helpers#
- 169.9.1 GPU-driven culling (frustum + occlusion)
- 169.9.2 Hierarchical z-buffer
- 169.9.3 Mesh / vertex fetch compute path
- 169.9.4 GPU light list / cluster build
- 169.9.5 GPU bone palette skinning
- 169.9.6 GPU morph target blend
- 169.9.7 GPU particles (Phase 161) runtime
- 169.9.8 Post-process kernels (bloom / DOF / TAA / CAS)
- 169.9.9 Compute-based rasterisation (software)
- 169.9.10 GPU virtual texture
169.10 @neith/compute-debug — Debug & Validation#
- 169.10.1 Resource-validation layer
- 169.10.2 Race-condition detector
- 169.10.3 Out-of-bounds access check
- 169.10.4 NaN / Inf hunter
- 169.10.5 Shader assert() runtime
- 169.10.6 Printf from shader to debug console
- 169.10.7 Step-debug in Phase 160 profiler
- 169.10.8 Memory sanitizer
- 169.10.9 Determinism test harness
- 169.10.10 Bindless resource leak tracker
169.11 @neith/compute-sdk — Developer SDK#
- 169.11.1 C / C++ / Rust / Swift bindings
- 169.11.2 Python bindings for offline scripts
- 169.11.3 JavaScript / TypeScript WebGPU wrapper
- 169.11.4 Kernel-authoring DSL macros
- 169.11.5 Examples (matmul / raymarch / fluid / FFT)
- 169.11.6 Benchmark harness
- 169.11.7 Stress-test suite
- 169.11.8 Version compat tests
- 169.11.9 Profiler integration (Phase 160)
- 169.11.10 Documentation + tutorials
169.12 Testing & Validation#
- 169.12.1 Cross-backend correctness (bit-reproducible for FP16)
- 169.12.2 Performance vs a CUDA reference ≥ 85% — the paired-median
benchmark grader is complete; certification needs a live NVIDIA/CUDA
capture. Needs an NVIDIA GPU to produce the CUDA reference; the fleet has
none (stated 2026-09-18).
blocked:hardware - 169.12.3 Performance vs a HIP reference ≥ 85% — the grader is
complete; certification needs a live AMD/HIP capture. Needs an AMD GPU to
produce the HIP reference; the fleet has none (stated 2026-09-18).
blocked:hardware - 169.12.4 WebGPU browser conformance — the physical-adapter no-skip evidence grader is complete; the residual is a live browser capture, which headless Chromium provides (its Dawn WebGPU backend runs in the repo's Playwright harness; enable WebGPU with the appropriate flags). Verify: the conformance suite runs the compute kernels in headless Chromium and the grader records a physical-adapter, no-skip pass.
- 169.12.5 24-hour soak (no driver crash) — the monotonic-duration/dispatch/crash/device-loss grader is complete; the residual is a live 24-hour run, attemptable on lavapipe (Linux) or the Mac. Verify: a 24-hour continuous-dispatch soak on lavapipe completes with zero device-loss or validation errors and the grader accepts the captured duration.
- 169.12.6 Multi-GPU scaling ≥ 80% efficiency — the scaling grader is
complete; certification needs two or more live peer GPUs. Needs two or
more peer GPUs on one host; the fleet has a single-GPU Mac and a GPU-less
Linux server (stated 2026-09-18).
blocked:hardware - 169.12.7 Memory-leak burn-in
- 169.12.8 Determinism mode regression
- 169.12.9 Dogfood: run the compute kernels of the Phase 42 engine, the Phase 157 sim and the Phase 47 inference path through this framework on the software/CPU devices (lavapipe on Linux, Metal on the Mac). Verify: a spec drives one representative kernel from each of the three subsystems through the abstraction and asserts each result matches its CPU reference, with the durations recorded.
- 169.12.10 Console platform (Phase 137) validation — the devkit grader
is complete; certification needs a live console devkit capture. Needs a
console devkit under NDA; the fleet has none (stated 2026-09-18).
blocked:hardware