nx_v128_share_probe.nx source

1// nx_v128_share_probe.nx -- F618 rung-1 HOT-LOOP RE-MEASURE (native): what share of the shipped-config 2// P-frame encode (416x320, emode 673, qp 22 -- the nx_vcodec_layout_gate recipe) is the 16x16 motion 3// search (vm_search_q, the SAD-dominated kernel the design names first)? Times (a) the full vv_enc_rct9 4// P encode and (b) an isolated whole-frame vm_search_q sweep with the encoder's own args, same frames. 5// The sweep is a LOWER BOUND on total SAD-family work (encode also does subpel refine, me8 partition 6// searches when armed, zero-SAD skip checks, intra SAD). license_tier: ORIGINAL 7import "nx_syscalls.nx" 8import "nx_video_codec_wasm.nx" 9const K_MAGIC_2026: i64 = 2026 10const K_MAGIC_524288: i64 = 524288 11const K_MAGIC_4096: i64 = 4096 12const K_MAGIC_1048576: i64 = 1048576 13const K_MAGIC_8192: i64 = 8192 14const K_MAGIC_1000000: i64 = 1000000 15 16const SPW: i64 = 416 17const SPH: i64 = 320 18const SPQP: i64 = 22 19const SPEMODE: i64 = 673 20const SPITER: i64 = 30 21 22func spw(s: *u8) -> i64 { var n: i64=0; while s[n]!=(0 as u8){n=n+1} sys_write(1,s,n); return 0 } 23func spn(v: i64) -> i64 { 24 let b: *u8=sys_mmap(28); var m: i64=v; if m<0{sys_write(1,"-" as *u8,1);m=0-m} 25 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} 26 var i: i64=0; while i<k{b[i]=t[k-1-i];i=i+1} sys_write(1,b,k); return 0 } 27 28func sp_fill(f0: *u8, f1: *u8, sz: i64) -> i64 { 29 let N: i64 = SPW*SPH 30 var i: i64 = 0 31 while i < N { 32 let x: i64 = i % SPW 33 let y: i64 = i / SPW 34 var v: i64 = (x + y) * 2 + ((x * y) % 17) 35 if x > (SPW/2) { v = v + 40 } 36 v = v % 256 37 f0[i] = v as u8 38 i = i + 1 } 39 i = N 40 while i < sz { f0[i] = 128 as u8; i = i + 1 } 41 var yy: i64 = 0 42 while yy < SPH { 43 var xx: i64 = 0 44 while xx < SPW { 45 var sx: i64 = xx - 2 46 if sx < 0 { sx = 0 } 47 f1[yy*SPW + xx] = f0[yy*SPW + sx] 48 xx = xx + 1 } 49 yy = yy + 1 } 50 i = N 51 while i < sz { f1[i] = f0[i]; i = i + 1 } 52 return 0 } 53 54func main() -> i64 { 55 let vmscr3: *i64 = sys_mmap(32) as *i64 // vm_search scratch (K_MAGIC_2026-07-29: hoisted ONE alloc; the per-call sys_mmap inside vm_search was the wasm 0-stub + native map-leak class) 56 spw("=== nx_v128_share_probe: motion-search share of shipped-config P encode (416x320 emode673 qp22) ===\n" as *u8) 57 let N: i64 = SPW*SPH 58 let C: i64 = (SPW/2)*(SPH/2) 59 let sz: i64 = N + 2*C 60 let f0: *u8 = sys_mmap(sz) 61 let f1: *u8 = sys_mmap(sz) 62 let zeros: *u8 = sys_mmap(sz) 63 let rEK: *u8 = sys_mmap(sz) 64 let rEP: *u8 = sys_mmap(sz) 65 let sK: *u8 = sys_mmap(K_MAGIC_524288) 66 let sP: *u8 = sys_mmap(K_MAGIC_524288) 67 let blk: *i64 = sys_mmap(K_MAGIC_4096) as *i64 68 let mv: *i64 = sys_mmap(128) as *i64 69 let rctx: *i64 = sys_mmap(K_MAGIC_4096) as *i64 70 let est: *i64 = sys_mmap(K_MAGIC_4096) as *i64 71 let probs: *i64 = sys_mmap(K_MAGIC_4096) as *i64 72 let rcbuf: *u8 = sys_mmap(K_MAGIC_1048576) 73 let t8c: *i64 = sys_mmap(K_MAGIC_8192) as *i64 74 let mvplane: *i64 = sys_mmap(K_MAGIC_8192) as *i64 75 vc_t8_init(t8c) 76 var i: i64 = 0 77 while i < sz { zeros[i] = 0 as u8; i = i + 1 } 78 sp_fill(f0, f1, sz) 79 // key(rct8) once -> rEK is the P-frame's prev (the shipped chain's reference) 80 rctx[0]=SPEMODE; rctx[1]=est as i64; rctx[2]=probs as i64; rctx[3]=rcbuf as i64; rctx[4]=t8c as i64 81 rctx[5]=0; rctx[6]=0; rctx[7]=0; rctx[8]=0; rctx[9]=mvplane as i64 82 let kb: i64 = vv_enc_rct8(f0, zeros, rEK, SPW, SPH, SPQP, 1, SPQP*188, sK, K_MAGIC_524288, blk, mv, rctx) 83 spw(" key bytes=" as *u8); spn(kb); spw("\n" as *u8) 84 // (a) FULL P encode timing (vv_enc_rct9, the vcv-10 P path) 85 let t0: i64 = sys_now_us() 86 var it: i64 = 0 87 while it < SPITER { 88 rctx[0]=SPEMODE; rctx[1]=est as i64; rctx[2]=probs as i64; rctx[3]=rcbuf as i64; rctx[4]=t8c as i64 89 rctx[5]=0; rctx[6]=0; rctx[7]=0; rctx[8]=0; rctx[9]=mvplane as i64 90 vv_enc_rct9(f1, rEK, rEP, SPW, SPH, SPQP, SPQP*94, sP, K_MAGIC_524288, blk, mv, rctx) 91 it = it + 1 } 92 let t1: i64 = sys_now_us() 93 let encus: i64 = (t1 - t0) / SPITER 94 // (b) isolated whole-frame 16x16 vm_search_q sweep (the encoder's own first-search args) 95 let BW: i64 = SPW / 16 96 let BH: i64 = SPH / 16 97 let t2: i64 = sys_now_us() 98 it = 0 99 var sadsum: i64 = 0 100 while it < SPITER { 101 var by: i64 = 0 102 while by < BH { 103 var bx: i64 = 0 104 while bx < BW { 105 let s: i64 = vm_search_q(f1, rEK, SPW, SPH, bx, by, 16, 16, mv, SPQP, vmscr3) 106 sadsum = sadsum + s 107 bx = bx + 1 } 108 by = by + 1 } 109 it = it + 1 } 110 let t3: i64 = sys_now_us() 111 let srchus: i64 = (t3 - t2) / SPITER 112 spw(" P-enc(vv_enc_rct9) = " as *u8); spn(encus); spw(" us/frame (" as *u8); spn(K_MAGIC_1000000/(encus+1)); spw(" fps)\n" as *u8) 113 spw(" 16x16 search sweep = " as *u8); spn(srchus); spw(" us/frame (sadsum=" as *u8); spn(sadsum); spw(")\n" as *u8) 114 spw(" search share (LOWER BOUND on SAD-family) = " as *u8); spn((srchus*1000)/(encus+1)); spw(" permil\n" as *u8) 115 return 0 }