Full-DPP Wave64 Primitives for CDNA3: Eliminating LDS-Mediated Cross-Lane Communication on AMD Instinct MI300X

Research artifacts for the paper: Eliminating LDS-Mediated Cross-Lane Communication on CDNA3. Includes header-only HIP library (wave_reduce_dpp, wave_scan_dpp), Flash Attention kernels (fa_naive, fa_dpp, fa_dme, fa_mfma2), native HIP LLM runtime (cdna3::runtime::Session), all benchmark and validation source code, hardware benchmark results from MI300X VF (gfx942, ROCm 7.2), and the paper draft. Validated on AMD Instinct MI300X (gfx942): DPP reduction 1.79x over hipCUB WarpReduce, Flash Attention 18% speedup via DPP+DME, LLM runtime 90.7 tok/s (Qwen2.5-0.5B, 2.1x over naive), TTFT 47.5x at 1024 tokens.