SRCNN‑cpp
• SRCNN (Super Resolution Convolution Neural Network) implementation with pure C++.
• Transform convolution calculation to matrix multiplication for re‑using techniques of matrix multiplication optimization.
• Optimize matrix multiplication from scratch with 10x speedup.
sse2neon
• A translator from Intel SSE intrinsics to Arm/Aarch64 NEON implementation.
• Support obtaining hardware timestamp on Armv7‑A.
• Improve precision of reciprocal of square‑root conversion to at least 24 bits, giving more accurate result for ray tracing programs such
as Blender and Embree.
• Fix incorrect result of reciprocal of square‑root conversion when dealing with positive/negative infinity.
• Introduce fast and deterministic RNG for testing on various platform/environment.
simdjson
• Parsing gigabytes of JSON per second.
• Implement JSON indented output string functions with CRTP for best performance.