Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

For the complete changelog, see GitHub CHANGELOG.md.

---

Unreleased

Added

• Comprehensive bilingual documentation (English & Chinese)
• New documentation site structure under docs/
• API reference documentation
• Architecture design documentation
• Performance optimization guide
• Development contribution guide

---

2.0.1 - 2026-04-16

Security

• Fixed 5 moderate npm vulnerabilities (esbuild, vite, vitest)
  • CVE-2024-XXXX: Regular expression denial of service in esbuild
  • CVE-2024-XXXX: Path traversal vulnerability in vite
• Updated vitest from v1.6.1 to v4.1.4
• Updated @vitest/coverage-v8 from v1.6.1 to v4.1.4

Changed

• Updated @webgpu/types from 0.1.68 to 0.1.69
• Updated @types/node to latest v20.x

---

2.0.0 - 2026-03-09

Added

• Zero-copy Tensor.reshape() via tensor views
  • Significant memory reduction for reshape-heavy models
  • Tensor._createView() static method for buffer sharing
  • Tensor._isView flag to track view tensors
• GPUContext.deferDestroy() for safe async buffer cleanup
  • Prevents use-after-free race conditions
  • Automatic cleanup of temporary GPU resources
• GPUContext.sync() and GPUContext.waitForSubmittedWork() methods
  • Better control over GPU synchronization
• FlattenOperator for CNN-to-FC layer transitions
• DenseOperator for fully connected layers
• Conv2dBiasReLUOperator - fused kernel optimization
  • 3x memory traffic reduction vs sequential execution

Fixed

• [Critical] Tensor.reshape() unnecessary GPU buffer allocation
  • Now uses zero-copy views, reducing memory pressure
• [Critical] InferenceEngine.infer() GPU memory leak
  • Intermediate tensors now properly destroyed after use
• [Critical] Race conditions in operator temp buffer cleanup
  • Now uses deferDestroy() pattern consistently

Changed

• Tensor.destroy() is now a no-op for view tensors
• All operators use deferred buffer destruction pattern

Performance

Operation	Before	After	Improvement
Flatten	Full buffer copy	Zero overhead	~100%
Fused Conv2d+Bias+ReLU	6 mem ops	2 mem ops	3×
Memory leaks	Persistent	Properly freed	Stable

---

1.0.0 - 2026-03-09

Added

• Standardized CI workflow with permissions and concurrency configuration
• Node.js 20 validation pipeline (lint, test, build)
• Automated security scanning for dependencies

Changed

• lint script now runs TypeScript type checking
• Updated CI triggers for better resource efficiency

---

0.9.0 - 2026-03-10

Added

• GitHub Pages deployment workflow
• Bilingual documentation support
  • Chinese documentation (README.zh-CN.md)
• Architecture diagram in index.md
• SEO metadata in _config.yml

Changed

• Refined Pages workflow path triggers for documentation files only
• Optimized Jekyll build exclusions

---

0.8.0 - 2026-01-08

Added

• tests/setup.ts - WebGPU globals for Node.js test environment
• tests/tsconfig.json - separate TypeScript config for tests
• @vitest/coverage-v8 for V8 coverage reporting

Fixed

• Vitest failing in Node.js due to missing WebGPU globals
• Operator test mocks not writing computed results to output buffers
• GPUContext.test.ts failing when navigator is read-only

Changed

• All operator tests have properly mocked WebGPU buffer lifecycle
• 54 tests passing (expanded from 47)

---

0.1.0 - 2025-02-13

Added

• Initial project infrastructure

  • MIT LICENSE
  • .editorconfig for consistent code formatting
  • TypeScript 5.x with strict mode
  • Vitest for testing with coverage

• Core Components

  • GPUContext - WebGPU device management with error handling
  • Tensor - GPU buffer management with layout support (NCHW/NHWC)
  • Custom error classes for all failure modes

• Neural Network Operators

  • ReLUOperator - Element-wise activation
  • SoftmaxOperator - Numerically stable probability distribution
  • MaxPoolOperator - 2D max pooling
  • Conv2dOperator - Direct convolution
  • DenseOperator - Matrix multiplication
  • FlattenOperator - Tensor reshaping

• Inference Engine

  • InferenceEngine - Complete inference pipeline
  • ModelLoader - JSON model loading

• Testing Framework

  • Property-based testing with fast-check
  • 100+ iterations per property test
  • CPU reference implementations for validation

• Examples

  • MNIST digit classification demo
  • Performance benchmark demo

• Performance Features

  • Hand-written WGSL compute shaders
  • Kernel fusion support architecture
  • Im2Col utilities for experimentation

Stats

• ~2,000 lines of source code
• ~1,100 lines of tests
• 7 neural network operators
• 13 property-based tests
• Zero runtime dependencies

---