code wiki / _hdl_build / nx_qmatvec_test.nx
nx_qmatvec_test.nx
buildroot/runtime/_hdl_build/nx_qmatvec_test.nx
about
nx_qmatvec_test.nx -- the ENGINEER verifies the sovereign quantized dot kernel and measures the
quality/data frontier. Exit 0 only if all hold:
(1) Q4 block-quantized dot matches the EXACT dot within the quality floor (real, bounded error).
(2) Q8 is much TIGHTER (more bits -> better quality) -- the frontier is real and monotone.
(3) data moved: Q4 < Q8 < int16 ref, and Q4 is >=3x less than the ref -> the memory-bound
speedup (throughput ~ 1/bytes) is >=3x where ggml SIMD micro-opt gave +0.8%.
Deterministic correlated vectors (no rng): X = W + bounded noise, so the exact dot is large and
the relative error is stable. license_tier: ORIGINAL
dependencies 2 imports · 0 importers
imports: nx_qmatvec.nxnx_syscalls.nx
imported by: nobody (leaf or entry point)
call flow from main pre-order; caps 40 nodes / depth 6 declared; ↻ = already shown
structs
| none |
consts
| none |
functions
| 13 | func qt_puts(s: *u8) -> i64 { var n: i64 = 0; while s[n] != (0 as u8) { n = n + 1 } sys_write(1, s, n); return 0 } |
| 14 | func qt_num(v: i64) -> i64 { let bb: *u8 = sys_mmap(28); var m: i64=v; if m<0 {m=0-m; sys_write(1,"-" as *u8,1)}; let t: *u8 = sys_mmap(28); var k: i64=0; if m==0 {t[0]=48;k=1}; while m>0 {t[k]=48+(m%10); m=m/10; k=k+1}; var i: i64=0; while i<k {bb[i]=t[k-1-i]; i=i+1}; sys_write(1, bb, k); return 0 } |
| 16 | func main() -> i64 |