nx_dequant_bench.nx

nx_dequant_bench.nx -- pure dequant throughput measurement. Earns its own numbers, claims nothing about other frameworks. Per the no-strawman-perf-comparisons feedback (2026-05-19): any "Nx faster" or "ratio" claim requires running BOTH sides on the same workload + hardware. This bench measures OUR Q4_K / Q8_0 / F32 dequant kernels. Comparing to llama.cpp / ggml requires running llama.cpp on the same hardware (deferred to Task #8). What this bench measures: * ns per dequantized value (avg over K runs) * MB/s input throughput (bytes-of-quantized-data / second) * Sanity check: output values not stuck at NX_GL_Q10_INF_SAT (catches catastrophic overflow / NaN propagation) Hardware caveat: this runs on qemu-riscv64-static (emulated RV64 on an x86_64 host). Numbers will be 5-20x slower than native because qemu user-mode dynamic translation costs overhead. Native RV64 silicon would be faster; native x86_64 (the actual deployment target) faster still. Report the qemu number AS the qemu number, not as "Nishi performance". genealogy_id: standard_microbench_harness + ggml_q4k_dequant_ref lineage_id: substrate_dequant_bench_v1