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.

2.0.0 - 2026-03-09

Fixed

• MemoryPool lifecycle bug: clear() was erasing tracking for in-use blocks; deallocate() left stale entries
• atomicMin/atomicMax for negative floats: compute_quant_params_kernel gave incorrect results for negative values; replaced with CAS-based atomic float min/max

Added

• core/warp_utils.hpp: Shared warp-level reduction primitives (warp_reduce_max/sum/min, warp_broadcast, block_reduce_sum/max)
• detail::fill_kernel: GPU-side fill kernel for Tensor::fill on non-byte types

Changed

• FlashAttention kernel rewrite: Moved output accumulator from per-thread registers to shared memory, reducing register pressure from 256 bytes/thread; cooperative tile loading; reduced default block sizes
• normalization.hpp no longer depends on softmax.hpp (uses warp_utils.hpp directly)
• Tensor::fill now uses a GPU kernel instead of host-memory roundtrip for non-byte types

Unreleased

1.0.0 - 2024-01-01

Added

• Initial release with core kernel implementations
• GEMM Kernels
  • Naive GEMM implementation
  • Tiled GEMM with shared memory optimization
  • Double-buffered GEMM for latency hiding
  • Tensor Core GEMM using WMMA API
• Attention Kernels
  • FlashAttention-style fused attention kernel
  • Memory-efficient attention computation
• Normalization Kernels
  • LayerNorm implementation
  • RMSNorm implementation
  • BatchNorm implementation
• Convolution Kernels
  • Naive 2D convolution
  • Im2Col-based convolution
  • Depthwise separable convolution
• Sparse Operations
  • CSR and CSC sparse matrix formats
  • Sparse Matrix-Vector multiplication (SpMV)
  • Sparse Matrix-Matrix multiplication (SpMM)
• Elementwise Operations
  • Fused elementwise kernel support
  • Common activation functions (ReLU, GELU, SiLU)
• Memory Management
  • Memory pool for efficient GPU memory allocation
  • Aligned vector for CPU-side data
  • Tensor abstraction with automatic memory management
• Python Bindings
  • pybind11-based Python interface
  • NumPy array interoperability
• Documentation
  • API reference documentation
  • Architecture overview
  • Optimization guide
  • Modern C++/CUDA best practices guide
• Benchmarks
  • GEMM performance benchmarks
  • Attention kernel benchmarks
  • Convolution benchmarks
• Testing
  • Unit tests for all kernel implementations
  • Correctness validation against reference implementations

Dependencies

• CUDA Toolkit 11.0+ (12.x recommended)
• CMake 3.18+
• C++17 compatible compiler
• pybind11 (for Python bindings)