nx_qmatvec.nx

nx_qmatvec.nx -- the FIRST real kernel toward beating ggml/llama.cpp: a sovereign block- quantized DOT PRODUCT (the hot path of LLM decode), built bits-up. The roofline proved decode is MEMORY-bound, so the lever is MOVING LESS DATA -- this kernel does exactly that, and the ENGINEER verifies it 1:1 against the exact dot while the quality (error) is measured vs the Council's floor. Integer-only by design (no fp dequant) -- a real efficiency edge for the edge/energy north-star. Scheme (ggml Q4_0/Q8_0 structure, integer scales): weights in blocks of B=32; each block has ONE integer scale S = max|w| in the block; each weight becomes a signed code q in [-QMAX,QMAX] (4-bit: -7..7, or 8-bit: -127..127). Dequant w' = q*S/QMAX. The block scale FACTORS OUT of the inner loop (ggml's key trick): block_dot = (S/QMAX) * sum_i q_i*x_i, so the hot loop is pure integer multiply-accumulate of small codes, and S applies once per block. bytes moved (what DRAM must deliver, the memory-bound cost): Q4: 4 bits/weight + 2 bytes scale per 32 = 4.5 bits/weight Q8: 8 bits/weight + 2 bytes scale per 32 = 8.5 bits/weight vs int16 ref = 16 bits/weight license_tier: ORIGINAL Refs: ggml Q4_0/Q8_0 block quantization; the memory-bound roofline.