v1.0.2

accel-gpu v1.0.2

Fixed

• WebKit browser backend consistency failure for WebGL matmul
• Hardened shader indexing with integer coordinate paths for stable cross-browser output

Result

• Backend consistency tests now pass reliably across Chromium, Firefox, and WebKit

v1.0.1

accel-gpu v1.0.1

Fixed

• WebGL matmul output indexing bug (output texture coordinates now mapped correctly)
• Restored strict cross-backend matmul consistency in browser tests

Added

• Backend consistency tests for matmul, inv, ft, conv2d
• Memory stress tests for gpu.tidy(...) and gpu.scoped(...)
• Canonical error-message validation tests
• Bundle-size smoke checks with CI enforcement
• Backend tolerance and debugging guide

Changed

• Standardized validation errors across linear, matrix, conv, and FFT paths
• Updated roadmap/changelog/docs for post-1.0 hardening completion

v1.0.0

accel-gpu 1.0.0

Production-ready major release with hardened memory management, improved type safety, and fully unified docs/navigation experience.

Highlights

• Memory management hardened:
  • gpu.tidy(...) now preserves returned arrays and disposes intermediates correctly
  • Nested tidy scope retention supported
  • FinalizationRegistry fallback cleanup remains in place
• TypeScript shape safety (opt-in):
  • Generic GPUArray<Shape> patterns
  • Shape-aware overloads for matmul, dot, and transpose
• Unified docs/site UX:
  • VitePress docs as the main GitHub Pages landing experience
  • Consistent themed navigation across landing, guides, demos, playground, and benchmarks
  • Added dedicated memory management guide
• Packaging and tree-shaking:
  • sideEffects: false
  • Subpath exports (math, linalg, ml, signal, data)

Validation

• Build + DTS generation passing
• Unit tests passing
• Docs build and site artifact generation passing

Notes

• Tag: v1.0.0
• Source of release notes: latest changelog + production hardening updates

v0.2.11

Highlights

• Added opt-in TypeScript shape safety using generic GPUArray<Shape> patterns.
• Improved compile-time ergonomics with shape-aware overloads for matmul, dot, and transpose.
• Fixed GitHub Pages docs routing by ensuring VitePress output is included in deploy artifacts.
• Refreshed docs presentation for a cleaner, more structured experience.

Added

• TypeScript Shape Safety
  • Generic shape aliases and typed context array creation paths
  • Shape-aware overloads for key linear algebra ops
• Docs Styling Enhancements
  • Custom VitePress theme styles
  • Cleaner home, quick start, and API page structure

Changed

• Pages Build Pipeline
  • build:site now runs docs build and packages docs routes into deploy/
  • /guide/quickstart now resolves correctly on GitHub Pages
• Docs UX
  • Improved nav/sidebar grouping, footer, and local search configuration

Validation

• Build and type declaration generation passing
• Unit tests passing
• Package published as accel-gpu@0.2.11

Upgrade Notes

• No breaking runtime API changes.
• Shape typing is opt-in and backward compatible with existing usage.

v0.2.10

Highlights

• Added raw binary ingestion with fromBuffer(...) and gpu.fromBuffer(...) for ArrayBuffer and SharedArrayBuffer inputs.
• Introduced a dedicated VitePress documentation scaffold under docs/.
• Migrated long-form README API content to focused docs pages and kept README concise.

Added

• Data Interop
  • fromBuffer(...) helper
  • gpu.fromBuffer(...) context API
  • Validation for byte alignment, offsets, and length bounds
• Documentation
  • VitePress config and starter pages (Overview, Quick Start, API)
  • Embedded playground/example iframes in docs pages
• Testing & CI
  • Cross-browser Playwright coverage (Chromium, Firefox, WebKit)
  • Browser-test workflow in GitHub Actions

Changed

• Improved package modularity and tree-shaking ergonomics:
  • sideEffects: false
  • Subpath exports: accel-gpu/math, accel-gpu/linalg, accel-gpu/ml, accel-gpu/signal, accel-gpu/data
• Added/validated memory-safety primitives:
  • gpu.tidy(fn) scoped disposal alias
  • FinalizationRegistry best-effort cleanup fallback
• README refocused to link out to the docs site for full API details.

Validation

• Unit tests passing
• Cross-browser Playwright suite passing
• Docs build passing

Upgrade Notes

• No breaking API changes.
• 0.2.10 is a patch release focused on DX, packaging clarity, docs structure, and ingestion performance paths.

v0.2.9

Patch release after published v0.2.8 lock. Includes latest docs/demo/package updates.

v0.2.8

Changelog

All notable changes to accel-gpu will be documented in this file.

The format is based on Keep a Changelog.

[0.2.8] - 2026-03-03

Added

• WebGPU kernels — conv2d, maxPool2d, avgPool2d now dispatch on WebGPU backend
• WebGPU FFT kernels — fft(), ifft(), and fftMagnitude() now use GPU compute on WebGPU backend
• GPU matrix ops — inv(), qr(), and svd() now use iterative WebGPU paths on WebGPU backend
• CPU worker path (experimental) — init({ forceCPU: true, worker: true })
• CPU WASM path (experimental) — init({ forceCPU: true, preferWasmCPU: true, wasmModule })
• Runtime backend flags — gpu.workerEnabled, gpu.cpuEngine
• Scoped lifecycle API — gpu.scoped(fn) for deterministic disposal of temporary arrays
• Training helpers — gradients(...) (numerical gradients) and sgdStep(...)
• Arrow interop — Apache Arrow-like vector/column import via fromArrow(...) and gpu.fromArrow(...)
• Killer app demos — example/audio (local audio spectrogram) and example/vector-search (browser-native vector search)
• Interactive docs playground — embedded playground on landing page

Changed

• conv2d, pooling, and FFT ops keep existing CPU/WebGL behavior as fallback while using WebGPU when available
• Updated docs and site examples for worker/WASM init options and accelerated op coverage
• Added troubleshooting guidance: /.well-known/appspecific/com.chrome.devtools.json 404 and 304 asset responses in local dev server logs are expected and harmless
• Added automatic scalar-chain fusion optimization for add/sub/mul/div before buffer materialization
• Added FinalizationRegistry best-effort cleanup for leaked arrays

[0.2.7] - 2025-03-02

Changed

• Re-publish with docs and site updates

[0.2.6] - 2025-03-01

Added

• Reductions — variance(), std(), argmax(), argmin()
• Axis-specific reductions — sum(axis?), mean(axis?), max(axis?)
• Activations — gelu(), leakyRelu(alpha)
• Comparison — equal(), greater(), less(), clamp(min, max)
• Slicing — slice(), get(), set(), concat(), split()
• Shape — flatten(), squeeze(), unsqueeze(), broadcast()
• Memory — dispose(), isDisposed, toArraySync() (CPU only)
• Matrix ops — inv(), det(), solve(), qr(), svd() (CPU)
• ML — maxPool2d, avgPool2d, conv2d, batchNorm, normalize()
• FFT & signal — fft(), ifft(), fftMagnitude(), spectrogram()
• Other — norm(), outer(), mse(), crossEntropy()
• Profiling — enableProfiling(), recordOp(), getProfilingResults(), init({ profiling: true })

[0.2.5] - 2025-03-01

Added

• JSDoc — Full documentation for public API (init, InitOptions, AccelContext, GPUArray, and all ops)
• ESLint + Prettier — Lint and format scripts; consistent code style
• Vitest — Test suite replacing scripts/test.mjs; npm test runs Vitest
• Dependabot — Weekly dependency updates for npm and GitHub Actions
• Bundlephobia badge — Package size badge in README
• Package icon — icon.png for docs and README

Changed

• ESLint flat config (no --ext); Prettier formats all src/**/*.ts
• README includes icon and Bundlephobia badge; favicon in docs

[0.2.0] - 2025-03-01

Added

• WebGL2 fallback — Full WebGL2 backend when WebGPU unavailable (Safari, Firefox, older Chrome)
• Shape inference — matmul(gpu, A, B) infers M, N, K from array shapes
• Method chaining — a.add(b).mul(2) returns this for chaining (await each step)
• reshape() — Reshape arrays with shape metadata
• CPU fallback — Automatic fallback when WebGPU/WebGL unavailable (Node, headless)
• Buffer pooling — Reuse GPUBuffers for better performance
• fromImageData() / toCanvas() — Canvas integration for image processing
• layerNorm — Layer normalization kernel for transformers
• attentionScores — Q @ K^T / sqrt(dim) for attention
• Clear error messages — Descriptive errors with shape info
• Benchmark page — Compare WebGPU vs WebGL vs CPU performance
• Playground — Interactive code editor
• forceCPU init option — Force CPU backend for testing
• forceWebGL init option — Force WebGL2 backend for testing

Changed

• Package renamed from @accel/gpu to accel-gpu
• init() now uses WebGPU → WebGL2 → CPU fallback chain
• matmul, softmax, transpose support shape inference from array metadata

[0.1.0] - Initial Release

• Core API: init(), gpu.array(), toArray()
• Math ops: add, mul, sum, max
• Linear algebra: matmul, dot, transpose
• ML: softmax
• WebGPU backend