Skip to content

v0.2.8

Choose a tag to compare

View all tags
@Phantasm0009 Phantasm0009 released this 03 Mar 18:27
· 12 commits to main since this release
v0.2.8
77832a6

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
Assets 2
Loading