nx_vq_train_gate.nx source

1// nx_vq_train_gate.nx -- proves Lloyd's training (nx_vq_train) lowers VQ distortion at the SAME bitrate (better codebook 2// coverage), the bridge from raw VQ toward the neural codebook. Trains a K-entry codebook on T period-varied LPC 3// reflection vectors and measures total distortion before vs after training. 4import "nx_syscalls_x86_64.nx" 5import "nx_lpc_autocorr.nx" 6import "nx_lpc_levinson.nx" 7import "nx_vq.nx" 8import "nx_vq_train.nx" 9 10func g_puts(s: *u8) -> i64 { var n: i64=0; while s[n]!=(0 as u8){n=n+1} sys_write(1,s,n); return 0 } 11func g_pn(v: i64) -> i64 { 12 let b: *u8 = sys_mmap(28); var x: i64 = v 13 if x < 0 { b[0]=45; sys_write(1,b,1); x = 0 - x } 14 if x == 0 { b[0]=48; sys_write(1,b,1); return 0 } 15 var d: i64=0; var y: i64=x 16 while y>0 { d=d+1; y=y/10 } 17 var i: i64=d-1; y=x 18 while i>=0 { b[i]=(48+(y%10)) as u8; y=y/10; i=i-1 } 19 sys_write(1,b,d); return 0 20} 21func g_check(name: *u8, cond: i64) -> i64 { 22 if cond==1 { g_puts(" PASS " as *u8) } else { g_puts(" FAIL " as *u8) } 23 g_puts(name); g_puts("\n" as *u8); return cond 24} 25func st16(buf: *u8, idx: i64, v: i64) -> i64 { var x: i64 = v; if x < 0 { x = x + 65536 } buf[idx*2] = x & 0xff; buf[idx*2+1] = (x >> 8) & 0xff; return 0 } 26func ld64(buf: *u8, byteoff: i64) -> i64 { var v: i64 = 0; var b: i64 = 0; while b < 8 { v = v | (buf[byteoff + b] << (b*8)); b = b + 1 } return v } 27func fill_parabola(s: *u8, n: i64, period: i64) -> i64 { var i: i64=0; while i<n { let p: i64 = i % period; st16(s, i, 40 + p*(period-p)*2); i=i+1 } return 0 } 28 29func main() -> i64 { 30 g_puts("nx_vq_train gate (Lloyd k-means lowers VQ distortion at same bitrate, MEASURED)\n" as *u8) 31 var pass: i64 = 0; var total: i64 = 0 32 let n: i64 = 64; let order: i64 = 10; let D: i64 = order 33 let T: i64 = 16; let K: i64 = 8 34 35 let s_pcm: *u8 = sys_mmap(n*2) 36 let R: *u8 = sys_mmap((order+1)*8) 37 let kb: *u8 = sys_mmap((order+1)*8) 38 let ab: *u8 = sys_mmap((order+1)*8) 39 let eb: *u8 = sys_mmap((order+1)*8) 40 let train: *i64 = sys_mmap(T*D*8) as *i64 41 let cb: *i64 = sys_mmap(K*D*8) as *i64 42 let assign: *i64 = sys_mmap(T*8) as *i64 43 let sumb: *i64 = sys_mmap(K*D*8) as *i64 44 let cnt: *i64 = sys_mmap(K*8) as *i64 45 46 // training set: T period-varied reflection vectors, in Q15 (>>15 from levinson Q30) to keep the squared distances bounded 47 var t: i64 = 0 48 while t < T { 49 fill_parabola(s_pcm, n, 8 + t) 50 nx_lpc_autocorr(s_pcm, n, order, R) 51 nx_lpc_levinson(R, order, kb, ab, eb) 52 var i: i64 = 0 53 while i < D { train[t*D + i] = ld64(kb, i*8) >> 15; i = i + 1 } 54 t = t + 1 55 } 56 // initialise the codebook with K evenly-spaced training vectors (a coarse start) 57 var c: i64 = 0 58 while c < K { var i: i64 = 0; while i < D { cb[c*D + i] = train[(c*2)*D + i]; i = i + 1 } c = c + 1 } 59 60 let dist_before: i64 = vqt_distortion(train, T, cb, K, D) 61 var it: i64 = 0 62 while it < 6 { vqt_iterate(train, T, cb, K, D, assign, sumb, cnt); it = it + 1 } 63 let dist_after: i64 = vqt_distortion(train, T, cb, K, D) 64 65 g_puts(" [measure] codebook K=" as *u8); g_pn(K); g_puts(" over T=" as *u8); g_pn(T); g_puts(" vectors: distortion before=" as *u8); g_pn(dist_before); g_puts(" after 6 Lloyd iters=" as *u8); g_pn(dist_after); g_puts("\n" as *u8) 66 if dist_before > 0 { g_puts(" [measure] distortion cut by training = " as *u8); g_pn((dist_before - dist_after) * 100 / dist_before); g_puts("% (same 3-bit index, better coverage)\n" as *u8) } 67 68 pass = pass + g_check("training lowers VQ distortion (Lloyd is non-increasing)" as *u8, dist_after <= dist_before); total=total+1 69 pass = pass + g_check("training measurably improves coverage (>= 10% distortion cut)" as *u8, dist_after * 10 <= dist_before * 9); total=total+1 70 71 g_puts("---- vq train gate: passed " as *u8); g_pn(pass); g_puts(" / " as *u8); g_pn(total); g_puts(" ----\n" as *u8) 72 if pass == total { g_puts("verdict=GREEN\n" as *u8); sys_exit(0); return 0 } 73 g_puts("verdict=RED\n" as *u8); sys_exit(1); return 1 74}