v0.4.0

Note: OpenCL support is not possible on static Linux builds.

Release Highlights

• Overhaul GF16 processing backend
  • Remove GF-Complete components, rewrite framework and improve general region handling
  • Fully separate ISA compilation units to support dynamic dispatch (also enables static Linux builds)
  • Add dot-product, region interleaving, chunk packing and prefetching optimisations
  • Add new calculation kernels: CLMul for NEON, Affine AVX variant for x86 (for Alder Lake and later CPUs) and experimental Shuffle2x/Affine2x variants
  • Add ARM SVE and SVE2 support
  • More optimisations during initialisation for various kernels, coefficient computation, and tweaked loop-tiling parameters
  • Improve transposition performance for Xor-Jit kernel, plus add single-use JIT optimisations
  • Rework multi-threading and remove OpenMP dependency; threading now manually managed via libuv
  • Add experimental OpenCL backend for GPGPU acceleration
    • Disabled by default - must be manually enabled
    • Have noticed it generate incorrect output, particularly on non-Windows hosts - use with caution!
• Add internal checksumming support to help detect memory errors during GF16 computation
• Improve concurrency when transferring to/from GF backend and hashing
• Redo MD5/CRC32 implementation for better optimisation
  • Input hashing now uses a stitched 2xMD5+CRC32 implementation
  • Add ASM MD5 implementation for x64/ARMv6/AArch64 (unsupported in MSVC)
  • Add ARM NEON and SVE2 MD5 implementations
  • Full SIMD width multi-buffer implementations
  • Remove node-yencode dependency
• Add support for concurrently processing multiple files to work around bottlenecks with single threaded input hashing
• Support concurrent I/O requests with chunked reading
• Support for compiling under MSVC/Clang-CL for Windows ARM/64 targets
• Separate GUI frontend available
• Improve progress display accuracy
• Various bug fixes