nx_vcodec_satd16_gate.nx source

1import "nx_gate_base.nx" 2// nx_vcodec_satd16_gate.nx -- SATD quarter-pel refine (emode bit10=1024, ENCODER-ONLY) roundtrip on /tmp/seq.y4m. 3// Argmin-by-Hadamard-SATD picks the MV whose residual is cheapest to CODE (x264's classic refinement win) vs raw 4// SAD. Decoder-transparent (the wire carries the winning MV; bestsad is internal to the argmin) -> NO vcv gate 5// needed; this gate proves (a) enc/dec recon stays bit-exact with SATD on (separate chains + memcmp every frame) 6// and (b) the byte/MSE direction vs the SHIPPED config. Compares emode 681 (rc+sig+part+gentle-deblock = live 7// BUILD 815) vs 1705 (681|1024 = +SATD refine). Real BD-rate = the MSU oracle (nx_vcodec_msu_ours16). license: ORIGINAL 8import "nx_syscalls.nx" 9import "nx_video_codec_wasm.nx" 10 11const SW: i64 = 352 12const SH: i64 = 288 13const NF: i64 = 48 14 15func grow(name: *u8, ok: i64) -> i64 { if ok==1 { gw(" PASS " as *u8) } else { gw(" FAIL " as *u8) } gw(name); gw(" 16" as *u8); return ok } 17func gn(v: i64) -> i64 { 18 let b: *u8=sys_mmap(28); var m: i64=v; if m<0{sys_write(1,"-" as *u8,1);m=0-m} 19 let t: *u8=sys_mmap(28); var k: i64=0; if m==0{t[0]=48 as u8;k=1} while m>0{t[k]=(48+(m%10)) as u8;m=m/10;k=k+1} 20 var i: i64=0; while i<k{b[i]=t[k-1-i];i=i+1} sys_write(1,b,k); return 0 } 21func cpb(d: *u8, s: *u8, n: i64) -> i64 { var i: i64=0; while i<n { d[i]=s[i]; i=i+1 } return 0 } 22func eqb(a: *u8, b: *u8, n: i64) -> i64 { var i: i64=0; while i<n { if a[i]!=b[i] { return i } i=i+1 } return 0-1 } 23 24func frames(yuv: *u8, hi: i64, total: i64, fp: *i64, fdata: i64) -> i64 { 25 var nf: i64=0; var p: i64=hi 26 while nf < NF { if p>=total { break } 27 var q: i64=p; while q<total { if (yuv[q]&0xff)==10 { break } q=q+1 } q=q+1 28 if q+fdata>total { break } fp[nf]=(yuv as i64)+q; nf=nf+1; p=q+fdata } 29 return nf } 30 31func roundtrip(fp: *i64, qp: i64, emode: i64, sz: i64, ok: *i64) -> i64 { 32 let prevE: *u8=sys_mmap(sz+64); let reconE: *u8=sys_mmap(sz+64) 33 let prevD: *u8=sys_mmap(sz+64); let reconD: *u8=sys_mmap(sz+64) 34 let wire: *u8=sys_mmap(4194304); let blk: *i64=sys_mmap(512) as *i64; let mv: *i64=sys_mmap(128) as *i64 35 let estE: *i64=sys_mmap(64) as *i64; let probsE: *i64=sys_mmap(32*8) as *i64; let rcE: *u8=sys_mmap(4194304); let t8cE: *i64=sys_mmap(5120) as *i64 36 let estD: *i64=sys_mmap(64) as *i64; let probsD: *i64=sys_mmap(32*8) as *i64; let t8cD: *i64=sys_mmap(5120) as *i64 37 let rctxE: *i64=sys_mmap(64) as *i64; let rctxD: *i64=sys_mmap(64) as *i64 38 vc_t8_init(t8cE); vc_t8_init(t8cD) 39 var z: i64=0; while z<sz { prevE[z]=0 as u8; prevD[z]=0 as u8; z=z+1 } 40 rctxE[1]=estE as i64; rctxE[2]=probsE as i64; rctxE[3]=rcE as i64; rctxE[4]=t8cE as i64; rctxE[5]=0; rctxE[6]=0; rctxE[7]=0 41 rctxD[1]=estD as i64; rctxD[2]=probsD as i64; rctxD[3]=0; rctxD[4]=t8cD as i64; rctxD[5]=0; rctxD[6]=0; rctxD[7]=0 42 ok[0]=0-1; ok[1]=0 43 var pB: i64=0 44 var f: i64=0 45 while f < NF { 46 let cur: *u8=fp[f] as *u8 47 var key: i64=0; if f==0 { key=1 } 48 rctxE[0]=emode 49 let nb: i64 = vv_enc_rct8(cur, prevE, reconE, SW, SH, qp, key, qp*30, wire, 4194304, blk, mv, rctxE) 50 if nb<=0 { gw(" enc FAIL frame " as *u8); gn(f); gw("\n" as *u8); ok[0]=f; return pB } 51 rctxD[0]=emode 52 let dr: i64 = vv_dec_rct8(prevD, reconD, SW, SH, qp, wire, nb, blk, mv, rctxD) 53 if dr<0 { gw(" dec FAIL frame " as *u8); gn(f); gw("\n" as *u8); ok[0]=f; return pB } 54 let mm: i64 = eqb(reconE, reconD, sz) 55 if mm >= 0 { if ok[0] < 0 { ok[0]=f; gw(" DESYNC frame " as *u8); gn(f); gw(" at byte " as *u8); gn(mm); gw("\n" as *u8) } } 56 var pxi: i64=0; let NLp: i64=SW*SH 57 while pxi<NLp { let d: i64=(reconE[pxi]&0xff)-(cur[pxi]&0xff); ok[1]=ok[1]+d*d; pxi=pxi+1 } 58 if key==1 { } else { pB=pB+nb } 59 cpb(prevE, reconE, sz); cpb(prevD, reconD, sz) 60 f=f+1 61 } 62 return pB } 63 64func onerun(fp: *i64, sz: i64, qp: i64, emode: i64, label: *u8) -> i64 { 65 let ok: *i64=sys_mmap(16) as *i64 66 let pB: i64=roundtrip(fp, qp, emode, sz, ok) 67 let NL: i64=NF*SW*SH; let mmse: i64=ok[1]*1000/NL 68 gw(" " as *u8); gw(label); gw(" emode=" as *u8); gn(emode); gw(" P-total=" as *u8); gn(pB); gw("B lumaMSEx1000=" as *u8); gn(mmse); gw(" bitexact=" as *u8) 69 if ok[0] < 0 { gw("GREEN" as *u8) } else { gw("RED(frame " as *u8); gn(ok[0]); gw(")" as *u8) } 70 gw("\n" as *u8) 71 return pB } 72func sweepqp(fp: *i64, sz: i64, qp: i64) -> i64 { 73 gw(" --- qp=" as *u8); gn(qp); gw(" ---\n" as *u8) 74 let base: i64=onerun(fp, sz, qp, 681, "SHIP 815 (gentle, SAD refine) " as *u8) 75 let feat: i64=onerun(fp, sz, qp, 681+1024, "+SATD refine (bit10) " as *u8) 76 if base>0 { gw(" --> SATD byte delta = " as *u8); gn((base-feat)*1000/base); gw(" per-mille (positive = SATD SMALLER)\n" as *u8) } 77 return 0 } 78 79func main() -> i64 { 80 gw("=== nx_vcodec_satd16_gate: SATD quarter-pel refine (bit10) rct8 roundtrip on /tmp/seq.y4m ===\n" as *u8) 81 let box: *i64 = sys_mmap(16) as *i64 82 let yuv: *u8 = sys_read_file("/tmp/seq.y4m" as *u8, box) 83 if (yuv as i64)==0 { gw("cannot read /tmp/seq.y4m -> RED\n" as *u8); return 1 } 84 let total: i64=box[0] 85 var hi: i64=0; while hi<total { if (yuv[hi]&0xff)==10 { break } hi=hi+1 } hi=hi+1 86 let C2: i64=(SW/2)*(SH/2); let fdata: i64=SW*SH+2*C2; let sz: i64=fdata 87 let fp: *i64=sys_mmap(128*8) as *i64 88 let nf: i64=frames(yuv, hi, total, fp, fdata) 89 if nf < NF { gw("too few frames -> RED\n" as *u8); return 1 } 90 sweepqp(fp, sz, 8) 91 sweepqp(fp, sz, 20) 92 sweepqp(fp, sz, 32) 93 gw("SATD16: DONE\n" as *u8) 94 return 0 }