conformtools 0.5.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +92 -0
- package/dist/arriraw-Y5VXHXNX.js +525 -0
- package/dist/browser/timeline_core_bg.wasm +0 -0
- package/dist/chunk-35U7WGES.js +3 -0
- package/dist/chunk-3XTUTEKJ.js +102 -0
- package/dist/chunk-7D7H6VFW.js +3 -0
- package/dist/chunk-AACY2V3L.js +3 -0
- package/dist/chunk-AHNDJFQR.js +3 -0
- package/dist/chunk-C2ESM27Z.js +3 -0
- package/dist/chunk-DKGGJM6Z.js +3 -0
- package/dist/chunk-DRHOPOC7.js +3 -0
- package/dist/chunk-HJXRLQOI.js +3 -0
- package/dist/chunk-IDDQR56R.js +7 -0
- package/dist/chunk-JEMLKBH2.js +3 -0
- package/dist/chunk-NYB4ISL3.js +3 -0
- package/dist/chunk-P7OVDUA5.js +3 -0
- package/dist/chunk-PSANHDZM.js +3 -0
- package/dist/chunk-QDWB7QU7.js +3 -0
- package/dist/chunk-RXHNBSER.js +3 -0
- package/dist/chunk-SP5RGDAM.js +3 -0
- package/dist/chunk-TK55FA56.js +3 -0
- package/dist/chunk-UCMNH6Q6.js +3 -0
- package/dist/chunk-UUMHVA7R.js +3 -0
- package/dist/chunk-WWGQFOB2.js +64 -0
- package/dist/chunk-YPRTUS5W.js +3 -0
- package/dist/cli.js +11 -0
- package/dist/clip-state-B6JOX3JE.js +3 -0
- package/dist/cr3-wasm-X6Q4YMUU.js +3 -0
- package/dist/dng-wasm-EAB24KDN.js +3 -0
- package/dist/dpx-wasm-OX2D24ZB.js +3 -0
- package/dist/esm-SFFBP7PS.js +3 -0
- package/dist/exr-wasm-Z6BBPJSN.js +3 -0
- package/dist/h264-JHQCJTWP.js +3 -0
- package/dist/h264-mxf-4M63NBQK.js +3 -0
- package/dist/http.js +4 -0
- package/dist/inspect-6OM4FTVW.js +46 -0
- package/dist/isobmff-XL33CVJR.js +3 -0
- package/dist/mcp.js +13 -0
- package/dist/media_core-VSTO6CS4.js +9 -0
- package/dist/media_core_bg-BKXQLNFA.js +3 -0
- package/dist/media_core_bg-XIM7Y4R3.js +3 -0
- package/dist/mxf-VIANACTI.js +3 -0
- package/dist/onnxruntime-web-stub-PKJ2NZBF.js +3 -0
- package/dist/ops-7HVBBXUQ.js +3 -0
- package/dist/prores-TQEZ3KFH.js +526 -0
- package/dist/r3d-2QCFX7UX.js +3 -0
- package/dist/r3d-AT6NNIY3.js +1025 -0
- package/dist/source-BDJAUQVR.js +3 -0
- package/dist/sscd-sweep-IXM4OD7O.js +3 -0
- package/dist/still-E4HLCCST.js +3 -0
- package/dist/tiff-wasm-XYTQNL2F.js +3 -0
- package/dist/timeline_core-VK2ZWPCO.js +3 -0
- package/dist/timeline_core_bg-QEOVAIWP.js +3 -0
- package/dist/udf_bg-UTQ6BVXX.js +3 -0
- package/dist/wasm/media_core_bg.wasm +0 -0
- package/dist/wasm/udf_bg.wasm +0 -0
- package/dist/wasm-threads/media_core_bg.wasm +0 -0
- package/dist/wav-DXF73Y56.js +3 -0
- package/dist/zod-IPSV75LP.js +3 -0
- package/package.json +41 -0
|
@@ -0,0 +1,1025 @@
|
|
|
1
|
+
#!/usr/bin/env node
|
|
2
|
+
import { createRequire as __conformRequire } from 'node:module'; const require = __conformRequire(import.meta.url);
|
|
3
|
+
import{a as e1}from"./chunk-7D7H6VFW.js";import{na as n1,ua as r1}from"./chunk-UUMHVA7R.js";import"./chunk-PSANHDZM.js";var t1=new Uint32Array([327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,327683,458759,458759,458759,458759,458759,458759,458759,458759,524296,524296,524296,524296,589839,589839,655376,655377,393227,393227,393227,393227,393227,393227,393227,393227,393227,393227,393227,393227,393227,393227,393227,393227,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262148,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,262146,458764,458764,458764,458764,458764,458764,458764,458764,458766,458766,458766,458766,458766,458766,458766,458766,393229,393229,393229,393229,393229,393229,393229,393229,393229,393229,393229,393229,393229,393229,393229,393229,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,327686,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,131081,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,196609,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262149,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154,262154]),a1=new Uint32Array([393223,393224,327686,327686,262149,262149,262149,262149,196609,196609,196609,196609,196609,196609,196609,196609,262146,262146,262146,262146,262147,262147,262147,262147,196612,196612,196612,196612,196612,196612,196612,196612,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536,65536]),i1=new Uint32Array([131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,131072,196610,196610,196610,196610,196610,196610,196610,196610,196611,196611,196611,196611,196611,196611,196611,196611,196612,196612,196612,196612,196612,196612,196612,196612,262149,262149,262149,262149,262150,262150,262150,262150,131073,131073,131073,131073,131073,131073,131073,131073,131073,131073,131073,131073,131073,131073,131073,131073]),u1=new Uint32Array([0,1,5,6,14,15,27,28,44,45,65,66,90,91,119,120,2,4,7,13,16,26,29,43,46,64,67,89,92,118,121,150,3,8,12,17,25,30,42,47,63,68,88,93,117,122,149,151,9,11,18,24,31,41,48,62,69,87,94,116,123,148,152,177,10,19,23,32,40,49,61,70,86,95,115,124,147,153,176,178,20,22,33,39,50,60,71,85,96,114,125,146,154,175,179,200,21,34,38,51,59,72,84,97,113,126,145,155,174,180,199,201,35,37,52,58,73,83,98,112,127,144,156,173,181,198,202,219,36,53,57,74,82,99,111,128,143,157,172,182,197,203,218,220,54,56,75,81,100,110,129,142,158,171,183,196,204,217,221,234,55,76,80,101,109,130,141,159,170,184,195,205,216,222,233,235,77,79,102,108,131,140,160,169,185,194,206,215,223,232,236,245,78,103,107,132,139,161,168,186,193,207,214,224,231,237,244,246,104,106,133,138,162,167,187,192,208,213,225,230,238,243,247,252,105,134,137,163,166,188,191,209,212,226,229,239,242,248,251,253,135,136,164,165,189,190,210,211,227,228,240,241,249,250,254,255]),z=new Int32Array([2048,2882,2841,2772,2676,2554,2408,2239,2048,1837,1609,1365,1108,841,565,284,2048,2772,2408,1837,1108,284,-565,-1365,-2048,-2554,-2841,-2882,-2676,-2239,-1609,-841,2048,2554,1609,284,-1108,-2239,-2841,-2772,-2048,-841,565,1837,2676,2882,2408,1365,2048,2239,565,-1365,-2676,-2772,-1609,284,2048,2882,2408,841,-1108,-2554,-2841,-1837,2048,1837,-565,-2554,-2676,-841,1609,2882,2048,-284,-2408,-2772,-1108,1365,2841,2239,2048,1365,-1609,-2882,-1108,1837,2841,841,-2048,-2772,-565,2239,2676,284,-2408,-2554,2048,841,-2408,-2239,1108,2882,565,-2554,-2048,1365,2841,284,-2676,-1837,1609,2772,2048,284,-2841,-841,2676,1365,-2408,-1837,2048,2239,-1609,-2554,1108,2772,-565,-2882,2048,-284,-2841,841,2676,-1365,-2408,1837,2048,-2239,-1609,2554,1108,-2772,-565,2882,2048,-841,-2408,2239,1108,-2882,565,2554,-2048,-1365,2841,-284,-2676,1837,1609,-2772,2048,-1365,-1609,2882,-1108,-1837,2841,-841,-2048,2772,-565,-2239,2676,-284,-2408,2554,2048,-1837,-565,2554,-2676,841,1609,-2882,2048,284,-2408,2772,-1108,-1365,2841,-2239,2048,-2239,565,1365,-2676,2772,-1609,-284,2048,-2882,2408,-841,-1108,2554,-2841,1837,2048,-2554,1609,-284,-1108,2239,-2841,2772,-2048,841,565,-1837,2676,-2882,2408,-1365,2048,-2772,2408,-1837,1108,-284,-565,1365,-2048,2554,-2841,2882,-2676,2239,-1609,841,2048,-2882,2841,-2772,2676,-2554,2408,-2239,2048,-1837,1609,-1365,1108,-841,565,-284]),s1={m16:931135616,lg_a:1098061119,lg_b:1125906863,lg_c:1046850094,lg_pa:1008981770,gc_k:1064554302,gc_p:1101529087,gc_t:1028443341,sd_eps:953267991,inv273:997199617,rr_c13:956301440,rr_c16:931135616,rr_c18:914358400,bil_sg:973078528,bil_srb:975874731,sobel_sc:931135616,bil_t7_0:1032826803,bil_t7_1:1040595071,bil_t7_2:1044597305,bil_t7_3:1046301409,bil_t7_4:1044597305,bil_t7_5:1040595071,bil_t7_6:1032826803,bil_t9_0:1031861895,bil_t9_1:1035879338,bil_t9_2:1039871874,bil_t9_3:1041522383,bil_t9_4:1042077871,bil_t9_5:1041522383,bil_t9_6:1039871874,bil_t9_7:1035879338,bil_t9_8:1031861895,gauss55_00:1065353216,gauss55_01:1082130432,gauss55_02:1088421888,gauss55_03:1082130432,gauss55_04:1065353216,gauss55_10:1082130432,gauss55_11:1098907648,gauss55_12:1104150528,gauss55_13:1098907648,gauss55_14:1082130432,gauss55_20:1088421888,gauss55_21:1104150528,gauss55_22:1109655552,gauss55_23:1104150528,gauss55_24:1088421888,gauss55_30:1082130432,gauss55_31:1098907648,gauss55_32:1104150528,gauss55_33:1098907648,gauss55_34:1082130432,gauss55_40:1065353216,gauss55_41:1082130432,gauss55_42:1088421888,gauss55_43:1082130432,gauss55_44:1065353216};function E(n,r){return(n[r]<<24|n[r+1]<<16|n[r+2]<<8|n[r+3])>>>0}function w(n,r){return n[r]<<8|n[r+1]}function q(n){let r=Math.min(n.length,65536),e=-1;for(let u=16;u+6<=r;u+=2){if(w(n,u)!==259)continue;let g=w(n,u+2),p=w(n,u+4);if(g<256||g>8192||p<128||p>8192)continue;let y=u-16;if(E(n,y+4)!==0||E(n,y+8)!==0)continue;let P=E(n,y);if(!(P<65536||y+P>n.length+8192)){e=y;break}}if(e<0)throw new Error("RED codestream container not found");let i=e,t=i+12,a=[];for(let u=0;u<8;u++)a.push(w(n,t+4+2*u));let o=i+28;if(o+512>n.length)throw new Error("R3D container table truncated");let b=a[1],m=a[2],f=i+540,v=v1(n,f,a[4],b);if(b%2!==0)throw new Error(`odd R3D canvas width ${b}`);return{canvasW:b,canvasH:m,rawFields:a,table512:n.subarray(o,o+512),streamOff:f,segments:v}}function v1(n,r,e,i){let t=Math.ceil(i/128),a=r;for(;(w(n,a)&4095)<256;)if(a+=2,a+2>n.length)throw new Error("R3D segment stream truncated");a+=1;let o=a,b=n[a-1],m=[];for(let f=0;f<e;f++){if(o+8>n.length)throw new Error("R3D segment header past essence end");let v=[w(n,o),w(n,o+2),w(n,o+4),w(n,o+6)],u=v.map(h=>h+7>>3),g=o+8,p=u[0]+u[1]+u[2]+u[3];if(g+p>n.length)throw new Error("R3D segment body past essence end");let y=g,P=v.map((h,x)=>{let C={off:y,bitLength:h,byteLength:u[x]};return y+=u[x],C});m.push({x:f%t*128,y:Math.floor(f/t)*16,quant:b,subbands:P}),b=n[g+p]??0,o=g+p+1}return m}function o1(n){let r=new Uint32Array(n.segments.length*4*8),e=0;for(let i of n.segments)for(let t=0;t<4;t++){let a=i.subbands[t];r[e]=a.off,r[e+1]=a.bitLength,r[e+2]=a.byteLength,r[e+3]=i.x,r[e+4]=i.y,r[e+5]=i.quant,r[e+6]=t,r[e+7]=0,e+=8}return r}function k(n){return x1[0]=n>>>0,h1[0]}var c1=new ArrayBuffer(4),x1=new Uint32Array(c1),h1=new Float32Array(c1),w1=1112556626;function B1(n){if(n.length<140)return null;let r=new DataView(n.buffer,n.byteOffset,n.byteLength);if(r.getUint32(0,!0)!==w1||r.getUint32(4,!0)!==1)return null;let e=r.getUint32(8,!0),i=12,t=[];for(let f=0;f<4;f++,i+=4)t.push(r.getUint32(i,!0));let a=[];for(let f=0;f<26;f++,i+=4)a.push(r.getUint32(i,!0));let o=r.getUint32(i,!0);i+=4;let b=r.getUint32(i,!0);if(i+=4,i+b>n.length)return null;let m=new TextDecoder().decode(n.subarray(i,i+b));return{wbBits:t,smatrixBits:a,exposureBits:o,fixFlashers:(e&2)!==0,model:m}}async function M(n){try{return await r1(),B1(n1(n??new Uint8Array(0)))}catch{return null}}var l1=`// R3D KOMODO entropy decode (kernel 1 of 2). One invocation per (segment, subband) job.
|
|
4
|
+
// Faithful port of decode_subband_raw in rust/media-core/src/r3d/entropy.rs.
|
|
5
|
+
// Writes RAW coefficients (scan order) straight into coeff[job*2048 ..] \u2014 the buffer must be
|
|
6
|
+
// zero-initialized; only nonzero decoded positions are written (deterministic across reruns).
|
|
7
|
+
// tab layout (u32 words): T410[1024] @0, T210[64] @1024, T310[64] @1088, SCAN[256] @1152,
|
|
8
|
+
// BASIS[256] @1408 (i32 bit patterns), BITPLANE[256] @1664.
|
|
9
|
+
|
|
10
|
+
struct Params { jobCount: u32, w: u32, h: u32, planeSize: u32 };
|
|
11
|
+
@group(0) @binding(0) var<uniform> params: Params;
|
|
12
|
+
@group(0) @binding(1) var<storage, read> essence: array<u32>;
|
|
13
|
+
@group(0) @binding(2) var<storage, read> jobs: array<u32>; // 8 u32 per job
|
|
14
|
+
@group(0) @binding(3) var<storage, read> tab: array<u32>;
|
|
15
|
+
@group(0) @binding(4) var<storage, read_write> coeff: array<i32>;
|
|
16
|
+
|
|
17
|
+
const T410_OFF: u32 = 0u;
|
|
18
|
+
const T210_OFF: u32 = 1024u;
|
|
19
|
+
const T310_OFF: u32 = 1088u;
|
|
20
|
+
const SCAN_OFF: u32 = 1152u;
|
|
21
|
+
const BASIS_OFF: u32 = 1408u;
|
|
22
|
+
const BP_OFF: u32 = 1664u;
|
|
23
|
+
|
|
24
|
+
var<private> DC_BASE: array<i32, 6> = array<i32, 6>(0, 16, 32, 64, 128, 256);
|
|
25
|
+
var<private> DC_BITS: array<i32, 6> = array<i32, 6>(2, 2, 3, 4, 5, 6);
|
|
26
|
+
|
|
27
|
+
|
|
28
|
+
// BitReader state (per invocation)
|
|
29
|
+
var<private> rc: u32;
|
|
30
|
+
var<private> rbc: i32;
|
|
31
|
+
var<private> rbyc: i32;
|
|
32
|
+
var<private> rptr: u32; // next byte index, relative to body start
|
|
33
|
+
var<private> bodyOff: u32; // absolute byte offset of body in essence
|
|
34
|
+
var<private> bodyLen: u32; // body byte length; reads past end return 0
|
|
35
|
+
var<private> jobBase: u32; // job*2048 into coeff
|
|
36
|
+
|
|
37
|
+
fn ebyte(i: u32) -> u32 { return (essence[i >> 2u] >> ((i & 3u) * 8u)) & 0xffu; }
|
|
38
|
+
fn bodyByte(i: u32) -> u32 {
|
|
39
|
+
if (i < bodyLen) { return ebyte(bodyOff + i); }
|
|
40
|
+
return 0u;
|
|
41
|
+
}
|
|
42
|
+
fn nextByte() -> u32 {
|
|
43
|
+
let b = bodyByte(rptr);
|
|
44
|
+
rptr = rptr + 1u;
|
|
45
|
+
return b;
|
|
46
|
+
}
|
|
47
|
+
fn neg(x: u32) -> bool { return (x & 0x80000000u) != 0u; }
|
|
48
|
+
|
|
49
|
+
fn refillA() {
|
|
50
|
+
if (rbc > 7) {
|
|
51
|
+
var iv: i32 = rbc + 8;
|
|
52
|
+
var g: u32 = 0u;
|
|
53
|
+
loop {
|
|
54
|
+
g = g + 1u; if (g > 8u) { break; }
|
|
55
|
+
rbyc = rbyc + 1;
|
|
56
|
+
rbc = iv - 16;
|
|
57
|
+
let b = nextByte();
|
|
58
|
+
rc = rc + (b << (u32(iv - 16) & 31u));
|
|
59
|
+
iv = iv - 8;
|
|
60
|
+
if (iv <= 15) { break; }
|
|
61
|
+
}
|
|
62
|
+
}
|
|
63
|
+
}
|
|
64
|
+
fn consume(shift: i32) {
|
|
65
|
+
rc = rc << (u32(shift) & 31u);
|
|
66
|
+
rbc = rbc + shift;
|
|
67
|
+
refillA();
|
|
68
|
+
}
|
|
69
|
+
fn fc14(u: u32) -> i32 {
|
|
70
|
+
if (((u - 0x10u) >> 16u) == 0u) {
|
|
71
|
+
return i32((u + 0xfff0u) & 0xffffu) + 0x41;
|
|
72
|
+
}
|
|
73
|
+
return 0x10030;
|
|
74
|
+
}
|
|
75
|
+
fn fa78(symIn: i32) -> i32 {
|
|
76
|
+
let sym = symIn % 9;
|
|
77
|
+
if (sym != 8) {
|
|
78
|
+
if (sym < 4) { return sym + 1; }
|
|
79
|
+
let base = (1 << (u32(sym) - 2u)) + 1; // FA_BASE[sym-4]: 5,9,17,33
|
|
80
|
+
var iVar5: i32 = -1;
|
|
81
|
+
var preC: u32 = rc;
|
|
82
|
+
var g: u32 = 0u;
|
|
83
|
+
loop {
|
|
84
|
+
g = g + 1u; if (g > 64u) { break; }
|
|
85
|
+
let iVar6 = iVar5;
|
|
86
|
+
var sh = preC << 1u;
|
|
87
|
+
rc = sh;
|
|
88
|
+
if (rbc == 7) {
|
|
89
|
+
rbc = 0;
|
|
90
|
+
sh = sh + nextByte();
|
|
91
|
+
rc = sh;
|
|
92
|
+
rbyc = rbyc + 1;
|
|
93
|
+
let ss = neg(preC);
|
|
94
|
+
iVar5 = iVar6 + 1;
|
|
95
|
+
preC = sh;
|
|
96
|
+
if (ss) { break; }
|
|
97
|
+
} else {
|
|
98
|
+
rbc = rbc + 1;
|
|
99
|
+
let cont = !neg(preC);
|
|
100
|
+
iVar5 = iVar6 + 1;
|
|
101
|
+
preC = sh;
|
|
102
|
+
if (!cont) { break; }
|
|
103
|
+
}
|
|
104
|
+
}
|
|
105
|
+
let uVar7 = rc;
|
|
106
|
+
let sufbits = sym - 4;
|
|
107
|
+
var uVar8: u32 = 0u;
|
|
108
|
+
if (sym >= 5) {
|
|
109
|
+
uVar8 = uVar7 >> (u32(0x24 - sym) & 31u);
|
|
110
|
+
consume(sufbits);
|
|
111
|
+
}
|
|
112
|
+
return base + i32((uVar8 + (u32(iVar5) << (u32(sufbits) & 31u))) & 0xffffu);
|
|
113
|
+
}
|
|
114
|
+
// sym == 8
|
|
115
|
+
var uVar2 = rc;
|
|
116
|
+
var iVar4 = rbc;
|
|
117
|
+
var uVar7v: i32 = 5;
|
|
118
|
+
if (!neg(uVar2)) {
|
|
119
|
+
var iVar3: i32 = 5;
|
|
120
|
+
var iVar5: i32 = 0;
|
|
121
|
+
var g: u32 = 0u;
|
|
122
|
+
var brk = false;
|
|
123
|
+
loop {
|
|
124
|
+
g = g + 1u; if (g > 64u) { break; }
|
|
125
|
+
var inner = false;
|
|
126
|
+
var g2: u32 = 0u;
|
|
127
|
+
loop {
|
|
128
|
+
g2 = g2 + 1u; if (g2 > 64u) { break; }
|
|
129
|
+
iVar5 = iVar3;
|
|
130
|
+
uVar2 = uVar2 << 1u;
|
|
131
|
+
rc = uVar2;
|
|
132
|
+
if (iVar4 != 7) { break; }
|
|
133
|
+
iVar4 = 0;
|
|
134
|
+
rbc = 0;
|
|
135
|
+
uVar2 = uVar2 + nextByte();
|
|
136
|
+
rc = uVar2;
|
|
137
|
+
rbyc = rbyc + 1;
|
|
138
|
+
iVar3 = iVar5 + 1;
|
|
139
|
+
if (neg(uVar2)) { inner = true; break; }
|
|
140
|
+
}
|
|
141
|
+
if (inner) { break; }
|
|
142
|
+
iVar4 = iVar4 + 1;
|
|
143
|
+
rbc = iVar4;
|
|
144
|
+
iVar3 = iVar5 + 1;
|
|
145
|
+
if (neg(uVar2)) { break; }
|
|
146
|
+
}
|
|
147
|
+
let v = iVar5 + 1;
|
|
148
|
+
if (v < 1) { return fc14(0u); }
|
|
149
|
+
uVar7v = v;
|
|
150
|
+
}
|
|
151
|
+
let uVar8 = uVar2 >> (bitcast<u32>(-uVar7v) & 31u);
|
|
152
|
+
rc = uVar2 << (u32(uVar7v) & 31u);
|
|
153
|
+
rbc = iVar4 + uVar7v;
|
|
154
|
+
refillA();
|
|
155
|
+
return fc14(uVar8);
|
|
156
|
+
}
|
|
157
|
+
// f8d4: one symbol -> value = zero-run skip, run = signed coefficient.
|
|
158
|
+
struct VR { value: i32, run: i32 };
|
|
159
|
+
fn f8d4() -> VR {
|
|
160
|
+
let e = tab[T410_OFF + ((rc >> 22u) & 0x3ffu)];
|
|
161
|
+
let sh = i32(e >> 16u);
|
|
162
|
+
let sym = i32(e & 0xffffu);
|
|
163
|
+
consume(sh);
|
|
164
|
+
var run = fa78(sym & 0xffff);
|
|
165
|
+
let iVar3 = rc;
|
|
166
|
+
rc = iVar3 << 1u;
|
|
167
|
+
if (rbc == 7) {
|
|
168
|
+
rc = rc + nextByte();
|
|
169
|
+
rbc = 0;
|
|
170
|
+
rbyc = rbyc + 1;
|
|
171
|
+
} else {
|
|
172
|
+
rbc = rbc + 1;
|
|
173
|
+
}
|
|
174
|
+
if (neg(iVar3)) { run = -run; }
|
|
175
|
+
var value: i32 = 0;
|
|
176
|
+
if (sym >= 9) {
|
|
177
|
+
let e2 = tab[T210_OFF + ((rc >> 26u) & 0x3fu)];
|
|
178
|
+
let sh2 = i32(e2 >> 16u);
|
|
179
|
+
let sym2 = i32(e2 & 0xffffu);
|
|
180
|
+
consume(sh2);
|
|
181
|
+
value = fa78(sym2 & 0xffff);
|
|
182
|
+
}
|
|
183
|
+
return VR(value, run);
|
|
184
|
+
}
|
|
185
|
+
fn dcSign(mag: i32) -> i32 {
|
|
186
|
+
let iVar5 = rc;
|
|
187
|
+
rc = iVar5 << 1u;
|
|
188
|
+
if (rbc == 7) {
|
|
189
|
+
rc = rc + nextByte();
|
|
190
|
+
rbc = 0;
|
|
191
|
+
rbyc = rbyc + 1;
|
|
192
|
+
} else {
|
|
193
|
+
rbc = rbc + 1;
|
|
194
|
+
}
|
|
195
|
+
if (neg(iVar5)) { return -mag; }
|
|
196
|
+
return mag;
|
|
197
|
+
}
|
|
198
|
+
fn f5c4() {
|
|
199
|
+
var iVar9: i32 = 0;
|
|
200
|
+
for (var sub: u32 = 0u; sub < 8u; sub = sub + 1u) {
|
|
201
|
+
let idx = ((rc >> 24u) & 0xfcu) >> 2u;
|
|
202
|
+
let ent = tab[T310_OFF + idx];
|
|
203
|
+
let sh = i32(ent >> 16u);
|
|
204
|
+
let symfull = i32(ent & 0xffffu);
|
|
205
|
+
rc = rc << (u32(sh) & 31u);
|
|
206
|
+
rbc = rbc + sh;
|
|
207
|
+
refillA();
|
|
208
|
+
if (symfull == 6) {
|
|
209
|
+
var uVar2: i32 = 8;
|
|
210
|
+
var uVar12 = rc;
|
|
211
|
+
var uVar14 = rbc;
|
|
212
|
+
var do70c = true;
|
|
213
|
+
var uVar13: u32 = 0u;
|
|
214
|
+
if (!neg(uVar12)) {
|
|
215
|
+
var uVar13s: i32 = 0;
|
|
216
|
+
var g: u32 = 0u;
|
|
217
|
+
loop {
|
|
218
|
+
g = g + 1u; if (g > 64u) { break; }
|
|
219
|
+
var inner = false;
|
|
220
|
+
var g2: u32 = 0u;
|
|
221
|
+
loop {
|
|
222
|
+
g2 = g2 + 1u; if (g2 > 64u) { break; }
|
|
223
|
+
uVar13s = uVar2;
|
|
224
|
+
uVar12 = uVar12 << 1u;
|
|
225
|
+
rc = uVar12;
|
|
226
|
+
if (uVar14 != 7) { break; }
|
|
227
|
+
uVar14 = 0;
|
|
228
|
+
rbc = 0;
|
|
229
|
+
uVar12 = uVar12 + nextByte();
|
|
230
|
+
rc = uVar12;
|
|
231
|
+
rbyc = rbyc + 1;
|
|
232
|
+
uVar2 = uVar13s + 1;
|
|
233
|
+
if (neg(uVar12)) { inner = true; break; }
|
|
234
|
+
}
|
|
235
|
+
if (inner) { break; }
|
|
236
|
+
uVar14 = uVar14 + 1;
|
|
237
|
+
rbc = uVar14;
|
|
238
|
+
uVar2 = uVar13s + 1;
|
|
239
|
+
if (neg(uVar12)) { break; }
|
|
240
|
+
}
|
|
241
|
+
uVar2 = uVar13s + 1;
|
|
242
|
+
if (uVar2 <= 0) {
|
|
243
|
+
uVar13 = 0u;
|
|
244
|
+
do70c = false;
|
|
245
|
+
}
|
|
246
|
+
}
|
|
247
|
+
if (do70c) {
|
|
248
|
+
uVar13 = uVar12 >> (bitcast<u32>(-uVar2) & 31u);
|
|
249
|
+
rc = uVar12 << (u32(uVar2) & 31u);
|
|
250
|
+
rbc = uVar14 + uVar2;
|
|
251
|
+
refillA();
|
|
252
|
+
}
|
|
253
|
+
var iVar5: i32 = 0x1017f;
|
|
254
|
+
if (((uVar13 - 0x80u) >> 16u) == 0u) {
|
|
255
|
+
iVar5 = i32((uVar13 + 0xff80u) & 0xffffu) + 0x200;
|
|
256
|
+
}
|
|
257
|
+
iVar9 = iVar9 + dcSign(iVar5);
|
|
258
|
+
coeff[jobBase + sub * 256u] = iVar9;
|
|
259
|
+
continue;
|
|
260
|
+
}
|
|
261
|
+
let sym = symfull;
|
|
262
|
+
let base = DC_BASE[sym];
|
|
263
|
+
let nbits = DC_BITS[sym];
|
|
264
|
+
var uVar12 = rc;
|
|
265
|
+
var uVar14 = rbc;
|
|
266
|
+
var iVar11: i32 = -1;
|
|
267
|
+
var uVar13: u32 = 0u;
|
|
268
|
+
var g: u32 = 0u;
|
|
269
|
+
loop {
|
|
270
|
+
g = g + 1u; if (g > 64u) { break; }
|
|
271
|
+
let iVar15 = iVar11;
|
|
272
|
+
uVar13 = uVar12 << 1u;
|
|
273
|
+
rc = uVar13;
|
|
274
|
+
if (uVar14 != 7) {
|
|
275
|
+
uVar14 = uVar14 + 1;
|
|
276
|
+
rbc = uVar14;
|
|
277
|
+
let b = !neg(uVar12);
|
|
278
|
+
uVar12 = uVar13;
|
|
279
|
+
iVar11 = iVar15 + 1;
|
|
280
|
+
if (!b) { break; }
|
|
281
|
+
} else {
|
|
282
|
+
uVar14 = 0;
|
|
283
|
+
rbc = 0;
|
|
284
|
+
uVar13 = uVar13 + nextByte();
|
|
285
|
+
rc = uVar13;
|
|
286
|
+
let b = neg(uVar12);
|
|
287
|
+
uVar12 = uVar13;
|
|
288
|
+
iVar11 = iVar15 + 1;
|
|
289
|
+
rbyc = rbyc + 1;
|
|
290
|
+
if (b) { break; }
|
|
291
|
+
}
|
|
292
|
+
}
|
|
293
|
+
rc = uVar13 << (u32(nbits) & 31u);
|
|
294
|
+
rbc = uVar14 + nbits;
|
|
295
|
+
refillA();
|
|
296
|
+
let suffix = uVar13 >> (bitcast<u32>(-nbits) & 31u);
|
|
297
|
+
let iVar5 = base + i32((suffix + (u32(iVar11) << (u32(nbits) & 31u))) & 0xffffu);
|
|
298
|
+
if (iVar5 != 0) { iVar9 = iVar9 + dcSign(iVar5); }
|
|
299
|
+
coeff[jobBase + sub * 256u] = iVar9;
|
|
300
|
+
}
|
|
301
|
+
}
|
|
302
|
+
|
|
303
|
+
@compute @workgroup_size(64)
|
|
304
|
+
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
305
|
+
let job = gid.x;
|
|
306
|
+
if (job >= params.jobCount) { return; }
|
|
307
|
+
let jb = job * 8u;
|
|
308
|
+
bodyOff = jobs[jb];
|
|
309
|
+
let bitLen = i32(jobs[jb + 1u]);
|
|
310
|
+
bodyLen = jobs[jb + 2u];
|
|
311
|
+
jobBase = job * 2048u;
|
|
312
|
+
|
|
313
|
+
// BitReader::new
|
|
314
|
+
rc = (bodyByte(0u) << 24u) | (bodyByte(1u) << 16u) | (bodyByte(2u) << 8u) | bodyByte(3u);
|
|
315
|
+
rbc = 0;
|
|
316
|
+
rbyc = 0;
|
|
317
|
+
rptr = 4u;
|
|
318
|
+
|
|
319
|
+
// DCs (8 subblocks), then AC loop (decode_subband_raw)
|
|
320
|
+
f5c4();
|
|
321
|
+
var u6: i32 = 8;
|
|
322
|
+
var g: u32 = 0u;
|
|
323
|
+
loop {
|
|
324
|
+
g = g + 1u; if (g > 4096u) { break; }
|
|
325
|
+
if (!(rbc + rbyc * 8 < bitLen)) { break; }
|
|
326
|
+
let vr = f8d4();
|
|
327
|
+
if (vr.value < 0) { break; }
|
|
328
|
+
let i1 = vr.value + u6;
|
|
329
|
+
if (i1 >= 2048) { break; }
|
|
330
|
+
let ip = u32(i1);
|
|
331
|
+
coeff[jobBase + (ip % 8u) * 256u + (ip >> 3u)] = vr.run; // pos_map
|
|
332
|
+
u6 = i1 + 1;
|
|
333
|
+
if (!(rbc + rbyc * 8 < bitLen)) { break; }
|
|
334
|
+
}
|
|
335
|
+
}
|
|
336
|
+
`;var f1=`// R3D KOMODO descan + dequant + 16x16 integer basis transform + scatter (kernel 2 of 2).
|
|
337
|
+
// Cooperative: ONE WORKGROUP (256 threads) per (job, subblock); deq/tmp live in shared memory,
|
|
338
|
+
// each thread computes one element of each pass. Mirrors decode_subband + scatter_subband.
|
|
339
|
+
// Dispatch is 2D; block index = wid.x + wid.y * num_workgroups.x.
|
|
340
|
+
// tab layout matches r3d-vld.wgsl.
|
|
341
|
+
|
|
342
|
+
struct Params { blockCount: u32, w: u32, h: u32, planeSize: u32 };
|
|
343
|
+
@group(0) @binding(0) var<uniform> params: Params;
|
|
344
|
+
@group(0) @binding(1) var<storage, read> jobs: array<u32>; // 8 u32 per job (same buffer as kernel 1)
|
|
345
|
+
@group(0) @binding(2) var<storage, read> tab: array<u32>;
|
|
346
|
+
@group(0) @binding(3) var<storage, read> coeff: array<i32>; // raw coeffs, job*2048 + sb*256 + k
|
|
347
|
+
@group(0) @binding(4) var<storage, read_write> planes: array<i32>;
|
|
348
|
+
|
|
349
|
+
const SCAN_OFF: u32 = 1152u;
|
|
350
|
+
const BASIS_OFF: u32 = 1408u;
|
|
351
|
+
const BP_OFF: u32 = 1664u;
|
|
352
|
+
|
|
353
|
+
var<workgroup> sBASIS: array<i32, 256>;
|
|
354
|
+
var<workgroup> deq: array<i32, 256>;
|
|
355
|
+
var<workgroup> tmp: array<i32, 256>;
|
|
356
|
+
|
|
357
|
+
@compute @workgroup_size(256)
|
|
358
|
+
fn main(@builtin(workgroup_id) wid: vec3<u32>,
|
|
359
|
+
@builtin(num_workgroups) nw: vec3<u32>,
|
|
360
|
+
@builtin(local_invocation_id) lid: vec3<u32>) {
|
|
361
|
+
let b = wid.x + wid.y * nw.x;
|
|
362
|
+
let t = lid.x;
|
|
363
|
+
sBASIS[t] = bitcast<i32>(tab[BASIS_OFF + t]);
|
|
364
|
+
if (b >= params.blockCount) { return; }
|
|
365
|
+
let job = b >> 3u;
|
|
366
|
+
let sb = b & 7u;
|
|
367
|
+
let jb = job * 8u;
|
|
368
|
+
let segX = jobs[jb + 3u];
|
|
369
|
+
let segY = jobs[jb + 4u];
|
|
370
|
+
let quant = i32(jobs[jb + 5u]);
|
|
371
|
+
let plane = jobs[jb + 6u];
|
|
372
|
+
let base = job * 2048u + sb * 256u;
|
|
373
|
+
|
|
374
|
+
// descan + dequant: one coefficient per thread
|
|
375
|
+
{
|
|
376
|
+
let src = coeff[base + tab[SCAN_OFF + t]];
|
|
377
|
+
let bp = i32(tab[BP_OFF + t]);
|
|
378
|
+
if (bp != 0) { deq[t] = src * bp * quant; } else { deq[t] = src * 2; }
|
|
379
|
+
}
|
|
380
|
+
workgroupBarrier();
|
|
381
|
+
// pass 1: tmp[r*16+c] over deq rows
|
|
382
|
+
{
|
|
383
|
+
let r = t >> 4u;
|
|
384
|
+
let c = t & 15u;
|
|
385
|
+
var acc: i32 = 0;
|
|
386
|
+
let bb = c * 16u;
|
|
387
|
+
let rr0 = r * 16u;
|
|
388
|
+
for (var k: u32 = 0u; k < 16u; k = k + 1u) {
|
|
389
|
+
acc = acc + sBASIS[bb + k] * deq[rr0 + k];
|
|
390
|
+
}
|
|
391
|
+
tmp[r * 16u + c] = (acc + 0x1000) >> 13u;
|
|
392
|
+
}
|
|
393
|
+
workgroupBarrier();
|
|
394
|
+
// pass 2 + scatter: res[i*16+j] -> plane[(segY+i), segX+sb*16+j]
|
|
395
|
+
{
|
|
396
|
+
let i = t >> 4u;
|
|
397
|
+
let j = t & 15u;
|
|
398
|
+
var acc: i32 = 0;
|
|
399
|
+
let bb = i * 16u;
|
|
400
|
+
for (var k: u32 = 0u; k < 16u; k = k + 1u) {
|
|
401
|
+
acc = acc + sBASIS[bb + k] * tmp[k * 16u + j];
|
|
402
|
+
}
|
|
403
|
+
let py = segY + i;
|
|
404
|
+
let px = segX + sb * 16u + j;
|
|
405
|
+
if (py < params.h && px < params.w) {
|
|
406
|
+
planes[plane * params.planeSize + py * params.w + px] = (acc + 0x1000) >> 13u;
|
|
407
|
+
}
|
|
408
|
+
}
|
|
409
|
+
}
|
|
410
|
+
`;var d1=`// R3D IPP2 develop chain (stages 1-15 of rust/media-core/src/r3d/develop.rs) on WebGPU.
|
|
411
|
+
// Production port of experiments/prores-wgsl-spike/r3d-develop/r3d-develop.wgsl with two
|
|
412
|
+
// changes: _d0 compands the entropy-lane channel planes directly (the spike read a
|
|
413
|
+
// pre-built u16 raw_pre buffer that shipping code no longer constructs \u2014 see
|
|
414
|
+
// frame.rs komodo_stage2), and _d16 packs the final u16 Log3G10 output to the RGBA8 the
|
|
415
|
+
// viewer's Full tier hands out (frame.rs log3g10_rgba8: high byte, alpha 255).
|
|
416
|
+
// One entry point per stage; ping-pong storage buffers bound per-dispatch by index.ts.
|
|
417
|
+
// Determinism recipe (arriraw/NOTES.md): every f32 multiply that feeds an add is
|
|
418
|
+
// mulr(a,b)=fma(a,b,zeroUniform); every ordered add/sub in a chain is
|
|
419
|
+
// addr/subr (fma with the runtime one) so the backend can neither contract nor
|
|
420
|
+
// reassociate; every division is div_cr (correctly-rounded Markstein). The Log3G10
|
|
421
|
+
// log10 and the gamut-compress powf are evaluated in double-f32 (dd) arithmetic,
|
|
422
|
+
// accurate to ~2^-45 \u2014 enough that the f32 rounding matches libm's f64-based results
|
|
423
|
+
// (the JS reference proved libm == round(f64) for this frame, bit-exact vs Rust).
|
|
424
|
+
//
|
|
425
|
+
// Buffers (harness): inA/inB i32 planes, inF f32 planes, outI/outF, rawU packed u16,
|
|
426
|
+
// outU packed u16 pairs. Plane-major layout, n = w*h per plane.
|
|
427
|
+
|
|
428
|
+
struct U {
|
|
429
|
+
w: u32, h: u32, n: u32, flags: u32,
|
|
430
|
+
zero: f32, one: f32, exposure: f32, pad0: f32,
|
|
431
|
+
wb: vec4<f32>,
|
|
432
|
+
nd: vec4<f32>,
|
|
433
|
+
smr: vec4<f32>, // sm[0..4]
|
|
434
|
+
smg: vec4<f32>, // sm[4..8]
|
|
435
|
+
smb: vec4<f32>, // sm[8..12]
|
|
436
|
+
};
|
|
437
|
+
@group(0) @binding(0) var<uniform> P: U;
|
|
438
|
+
@group(0) @binding(1) var<storage, read> inA: array<i32>;
|
|
439
|
+
@group(0) @binding(2) var<storage, read> inB: array<i32>;
|
|
440
|
+
@group(0) @binding(3) var<storage, read> inF: array<f32>;
|
|
441
|
+
@group(0) @binding(4) var<storage, read_write> outI: array<i32>;
|
|
442
|
+
@group(0) @binding(5) var<storage, read_write> outF: array<f32>;
|
|
443
|
+
@group(0) @binding(7) var<storage, read_write> outU: array<u32>;
|
|
444
|
+
@group(0) @binding(8) var<storage, read_write> outR: array<u32>;
|
|
445
|
+
|
|
446
|
+
//__CONSTS__
|
|
447
|
+
|
|
448
|
+
// ---- determinism helpers ----
|
|
449
|
+
fn mulr(a: f32, b: f32) -> f32 { return fma(a, b, P.zero); }
|
|
450
|
+
fn addr(a: f32, b: f32) -> f32 { return fma(b, P.one, a); }
|
|
451
|
+
fn subr(a: f32, b: f32) -> f32 { return fma(-b, P.one, a); }
|
|
452
|
+
fn recip_cr(s: f32) -> f32 {
|
|
453
|
+
let y0 = 1.0 / s;
|
|
454
|
+
let e0 = fma(-s, y0, 1.0);
|
|
455
|
+
let y1 = fma(e0, y0, y0);
|
|
456
|
+
let e1 = fma(-s, y1, 1.0);
|
|
457
|
+
return fma(e1, y1, y1);
|
|
458
|
+
}
|
|
459
|
+
fn div_cr(a: f32, b: f32) -> f32 { // correctly-rounded a/b (Markstein)
|
|
460
|
+
let y = recip_cr(b);
|
|
461
|
+
let q0 = mulr(a, y);
|
|
462
|
+
let r = fma(-b, q0, a);
|
|
463
|
+
return fma(r, y, q0);
|
|
464
|
+
}
|
|
465
|
+
|
|
466
|
+
// ---- double-f32 (dd) arithmetic; every internal add/sub opaque via addr/subr ----
|
|
467
|
+
fn two_sum(a: f32, b: f32) -> vec2<f32> {
|
|
468
|
+
let s = addr(a, b);
|
|
469
|
+
let bb = subr(s, a);
|
|
470
|
+
let err = addr(subr(a, subr(s, bb)), subr(b, bb));
|
|
471
|
+
return vec2<f32>(s, err);
|
|
472
|
+
}
|
|
473
|
+
fn qts(a: f32, b: f32) -> vec2<f32> { // quick_two_sum, |a| >= |b|
|
|
474
|
+
let s = addr(a, b);
|
|
475
|
+
return vec2<f32>(s, subr(b, subr(s, a)));
|
|
476
|
+
}
|
|
477
|
+
fn two_prod(a: f32, b: f32) -> vec2<f32> {
|
|
478
|
+
let p = mulr(a, b);
|
|
479
|
+
return vec2<f32>(p, fma(a, b, -p));
|
|
480
|
+
}
|
|
481
|
+
fn dd_add(x: vec2<f32>, y: vec2<f32>) -> vec2<f32> {
|
|
482
|
+
let s = two_sum(x.x, y.x);
|
|
483
|
+
let t = two_sum(x.y, y.y);
|
|
484
|
+
var lo = addr(s.y, t.x);
|
|
485
|
+
let r = qts(s.x, lo);
|
|
486
|
+
lo = addr(r.y, t.y);
|
|
487
|
+
return qts(r.x, lo);
|
|
488
|
+
}
|
|
489
|
+
fn dd_add_f(x: vec2<f32>, b: f32) -> vec2<f32> {
|
|
490
|
+
let s = two_sum(x.x, b);
|
|
491
|
+
return qts(s.x, addr(s.y, x.y));
|
|
492
|
+
}
|
|
493
|
+
fn dd_mul(x: vec2<f32>, y: vec2<f32>) -> vec2<f32> {
|
|
494
|
+
let p = two_prod(x.x, y.x);
|
|
495
|
+
let lo = addr(p.y, addr(mulr(x.x, y.y), mulr(x.y, y.x)));
|
|
496
|
+
return qts(p.x, lo);
|
|
497
|
+
}
|
|
498
|
+
fn dd_mul_f(x: vec2<f32>, b: f32) -> vec2<f32> {
|
|
499
|
+
let p = two_prod(x.x, b);
|
|
500
|
+
return qts(p.x, addr(p.y, mulr(x.y, b)));
|
|
501
|
+
}
|
|
502
|
+
fn dd_div(x: vec2<f32>, y: vec2<f32>) -> vec2<f32> {
|
|
503
|
+
let q1 = div_cr(x.x, y.x);
|
|
504
|
+
let r1 = dd_add(x, dd_mul_f(y, -q1));
|
|
505
|
+
let q2 = div_cr(r1.x, y.x);
|
|
506
|
+
let r2 = dd_add(r1, dd_mul_f(y, -q2));
|
|
507
|
+
let q3 = div_cr(r2.x, y.x);
|
|
508
|
+
return dd_add_f(qts(q1, q2), q3);
|
|
509
|
+
}
|
|
510
|
+
|
|
511
|
+
// ---- dd log mantissa core: x = 2^e * m, ln(m) via 2*atanh((m-1)/(m+1)) ----
|
|
512
|
+
struct LogM { lm: vec2<f32>, e: f32 };
|
|
513
|
+
fn logm(x: f32) -> LogM {
|
|
514
|
+
let bits = bitcast<u32>(x);
|
|
515
|
+
var e = i32(bits >> 23u) - 127;
|
|
516
|
+
var m = bitcast<f32>((bits & 0x007FFFFFu) | 0x3F800000u);
|
|
517
|
+
if (m > 1.4142135381698608) { m = m * 0.5; e = e + 1; }
|
|
518
|
+
let z = subr(m, 1.0); // exact (Sterbenz)
|
|
519
|
+
let den = two_sum(2.0, z);
|
|
520
|
+
let s = dd_div(vec2<f32>(z, 0.0), den);
|
|
521
|
+
let t = dd_mul(s, s);
|
|
522
|
+
let tf = t.x;
|
|
523
|
+
// f32 tail: 1/9 + t/11 + ... + t^6/21 (contributes at t^4 ~ 7.7e-7 -> 2^-48)
|
|
524
|
+
var R = fma(tf, C_1_21, C_1_19);
|
|
525
|
+
R = fma(tf, R, C_1_17);
|
|
526
|
+
R = fma(tf, R, C_1_15);
|
|
527
|
+
R = fma(tf, R, C_1_13);
|
|
528
|
+
R = fma(tf, R, C_1_11);
|
|
529
|
+
R = fma(tf, R, C_1_9);
|
|
530
|
+
var acc = dd_add(DD_1_7, dd_mul_f(t, R));
|
|
531
|
+
acc = dd_add(DD_1_5, dd_mul(t, acc));
|
|
532
|
+
acc = dd_add(DD_1_3, dd_mul(t, acc));
|
|
533
|
+
let q = dd_add_f(dd_mul(t, acc), 1.0);
|
|
534
|
+
let lm = dd_mul(dd_mul_f(s, 2.0), q);
|
|
535
|
+
return LogM(lm, f32(e));
|
|
536
|
+
}
|
|
537
|
+
fn log10_cr(x: f32) -> f32 {
|
|
538
|
+
let l = logm(x);
|
|
539
|
+
let r = dd_add(dd_mul_f(DD_LOG10_2, l.e), dd_mul(l.lm, DD_INVLN10));
|
|
540
|
+
return addr(r.x, r.y);
|
|
541
|
+
}
|
|
542
|
+
fn pow_gp(x: f32) -> f32 { // x^GC_P, x in (0, ~1.0001]; caller handles x <= 0
|
|
543
|
+
if (x < 1.1754944e-38) { return 0.0; }
|
|
544
|
+
let l = logm(x);
|
|
545
|
+
let l2 = dd_add(dd_mul(l.lm, DD_INVLN2), vec2<f32>(l.e, 0.0));
|
|
546
|
+
let y = dd_mul_f(l2, GC_P);
|
|
547
|
+
if (y.x < -150.0) { return 0.0; }
|
|
548
|
+
let kf = round(mulr(y.x, 16.0));
|
|
549
|
+
let ki = i32(kf);
|
|
550
|
+
let f = dd_add_f(y, -(kf * 0.0625)); // kf/16 exact
|
|
551
|
+
let u = dd_mul(f, DD_LN2);
|
|
552
|
+
let uf = u.x;
|
|
553
|
+
let u2 = uf * uf;
|
|
554
|
+
let tail = u2 * u2 * fma(uf, fma(uf, C_1_720, C_1_120), C_1_24);
|
|
555
|
+
var acc = dd_mul(u, DD_1_6);
|
|
556
|
+
acc = dd_add_f(acc, 0.5);
|
|
557
|
+
acc = dd_mul(u, acc);
|
|
558
|
+
acc = dd_add_f(acc, 1.0);
|
|
559
|
+
acc = dd_mul(u, acc);
|
|
560
|
+
acc = dd_add_f(acc, 1.0);
|
|
561
|
+
acc = dd_add_f(acc, tail);
|
|
562
|
+
let r = dd_mul(acc, EXP2TAB[ki & 15]);
|
|
563
|
+
var v = addr(r.x, r.y);
|
|
564
|
+
var e2 = ki >> 4u;
|
|
565
|
+
if (e2 < -126) {
|
|
566
|
+
if (e2 < -149) { return 0.0; }
|
|
567
|
+
v = v * 1.1754943508222875e-38; // 2^-126
|
|
568
|
+
e2 = e2 + 126;
|
|
569
|
+
}
|
|
570
|
+
return v * bitcast<f32>(u32(e2 + 127) << 23u);
|
|
571
|
+
}
|
|
572
|
+
|
|
573
|
+
// ---- shared sampling helpers ----
|
|
574
|
+
fn cx(v: i32, m: u32) -> u32 { return u32(clamp(v, 0, i32(m) - 1)); }
|
|
575
|
+
fn idxc(x: i32, y: i32) -> u32 { return cx(y, P.h) * P.w + cx(x, P.w); }
|
|
576
|
+
fn clampi(v: i32) -> i32 { return clamp(v, 0, 65535); }
|
|
577
|
+
fn fcl(v: f32) -> i32 { return i32(floor(clamp(v, 0.0, 65535.0))); }
|
|
578
|
+
|
|
579
|
+
const WGX = 16u;
|
|
580
|
+
const WGY = 16u;
|
|
581
|
+
|
|
582
|
+
// ==== stage 1: compand + preDeGAS2 + plane extract ====
|
|
583
|
+
// Input inA = the entropy lane's 4 channel planes [LL, HH, HL, LH]; RED's raw_pre
|
|
584
|
+
// order is [HL, LL, LH, HH] = entropy planes [2, 0, 3, 1]. The compand is
|
|
585
|
+
// clip(16\xB7coeff + 32768) into 0..=65535 (komodo_stage2) \u2014 pre-clamping the
|
|
586
|
+
// coefficient to \xB14096 makes 16\xB7v+32768 overflow-free while leaving the clipped
|
|
587
|
+
// result identical to the Rust i64 formula for every input.
|
|
588
|
+
fn cmpd(base: u32, i: u32) -> i32 { // base = entropy-plane offset (plane * n)
|
|
589
|
+
let v = clamp(inA[base + i], -4096, 4096);
|
|
590
|
+
return clamp(16 * v + 32768, 0, 65535);
|
|
591
|
+
}
|
|
592
|
+
@compute @workgroup_size(16, 16)
|
|
593
|
+
fn _d0(@builtin(global_invocation_id) g: vec3<u32>) {
|
|
594
|
+
let x = g.x; let y = g.y;
|
|
595
|
+
if (x >= P.w || y >= P.h) { return; }
|
|
596
|
+
let n = P.n; let i = y * P.w + x;
|
|
597
|
+
let a = cmpd(0u, i); // raw_pre p1 (LL = entropy plane 0) center
|
|
598
|
+
outI[i] = cmpd(2u * n, i); // raw_pre p0 (HL = entropy plane 2)
|
|
599
|
+
outI[n + i] = a;
|
|
600
|
+
outI[2u * n + i] = cmpd(3u * n, i); // raw_pre p2 (LH = entropy plane 3)
|
|
601
|
+
var avg: i32;
|
|
602
|
+
if (x < P.w - 1u && y < P.h - 1u) {
|
|
603
|
+
avg = (a + cmpd(0u, i + 1u) + cmpd(0u, i + P.w) + cmpd(0u, i + P.w + 1u)) >> 2u;
|
|
604
|
+
} else if (x == P.w - 1u && y < P.h - 1u) {
|
|
605
|
+
avg = (a + cmpd(0u, i + P.w)) >> 1u;
|
|
606
|
+
} else if (y == P.h - 1u && x < P.w - 1u) {
|
|
607
|
+
avg = (a + cmpd(0u, i + 1u)) >> 1u;
|
|
608
|
+
} else {
|
|
609
|
+
avg = a;
|
|
610
|
+
}
|
|
611
|
+
let v = avg + 2 * cmpd(n, i) - 65536; // raw_pre p3 (HH = entropy plane 1)
|
|
612
|
+
outI[3u * n + i] = clampi(v);
|
|
613
|
+
}
|
|
614
|
+
|
|
615
|
+
// ==== stage 2: degas16_to_image ====
|
|
616
|
+
@compute @workgroup_size(16, 16)
|
|
617
|
+
fn _d1(@builtin(global_invocation_id) g: vec3<u32>) {
|
|
618
|
+
let x = i32(g.x); let y = i32(g.y);
|
|
619
|
+
if (g.x >= P.w || g.y >= P.h) { return; }
|
|
620
|
+
let n = P.n; let i = g.y * P.w + g.x;
|
|
621
|
+
let p1c = inA[n + i];
|
|
622
|
+
let p3c = inA[3u * n + i];
|
|
623
|
+
let p1r = inA[n + idxc(x + 1, y)];
|
|
624
|
+
let p3u = inA[3u * n + idxc(x, y - 1)];
|
|
625
|
+
let p1d = inA[n + idxc(x, y + 1)];
|
|
626
|
+
let p3l = inA[3u * n + idxc(x - 1, y)];
|
|
627
|
+
var est2 = (p1c + p3c + p1r + p3u) >> 2u;
|
|
628
|
+
var est1 = (p1c + p3c + p1d + p3l) >> 2u;
|
|
629
|
+
if (g.x == 0u && g.y > 0u && g.y < P.h - 1u) { est1 = (p1c + 2 * p3c + p1d) >> 2u; }
|
|
630
|
+
if (g.x == P.w - 1u && g.y > 0u && g.y < P.h - 1u) { est2 = (2 * p1c + p3c + p3u) >> 2u; }
|
|
631
|
+
if (g.y == 0u) { est2 = (2 * p1c + 2 * p3c) >> 2u; est1 = (p1c + 2 * p3c + p1d) >> 2u; }
|
|
632
|
+
if (g.y == P.h - 1u) { est2 = (2 * p1c + p3c + p3u) >> 2u; est1 = (2 * p1c + 2 * p3c) >> 2u; }
|
|
633
|
+
outI[i] = clampi(2 * inA[i] - 65536 + est2);
|
|
634
|
+
outI[n + i] = p1c;
|
|
635
|
+
outI[2u * n + i] = clampi(2 * inA[2u * n + i] - 65536 + est1);
|
|
636
|
+
outI[3u * n + i] = p3c;
|
|
637
|
+
}
|
|
638
|
+
|
|
639
|
+
// ==== stage 3: fix_flashers ====
|
|
640
|
+
const FOFF = array<vec2<i32>, 8>( // (dy, dx) in develop.rs order
|
|
641
|
+
vec2<i32>(0, -1), vec2<i32>(0, 1), vec2<i32>(-1, 0), vec2<i32>(1, 0),
|
|
642
|
+
vec2<i32>(1, -1), vec2<i32>(1, 1), vec2<i32>(-1, -1), vec2<i32>(-1, 1));
|
|
643
|
+
@compute @workgroup_size(16, 16)
|
|
644
|
+
fn _d2(@builtin(global_invocation_id) g: vec3<u32>) {
|
|
645
|
+
let x = i32(g.x); let y = i32(g.y);
|
|
646
|
+
if (g.x >= P.w || g.y >= P.h) { return; }
|
|
647
|
+
let n = P.n; let i = g.y * P.w + g.x;
|
|
648
|
+
for (var c = 0u; c < 4u; c++) {
|
|
649
|
+
let base = c * n;
|
|
650
|
+
let ctr = inA[base + i];
|
|
651
|
+
var sum = 0;
|
|
652
|
+
var maxdiff = 0;
|
|
653
|
+
var localmax = true;
|
|
654
|
+
var nbr: array<i32, 8>;
|
|
655
|
+
for (var k = 0u; k < 8u; k++) {
|
|
656
|
+
let v = inA[base + idxc(x + FOFF[k].y, y + FOFF[k].x)];
|
|
657
|
+
nbr[k] = v;
|
|
658
|
+
sum += v;
|
|
659
|
+
}
|
|
660
|
+
let avg8 = sum / 8;
|
|
661
|
+
for (var k = 0u; k < 8u; k++) {
|
|
662
|
+
maxdiff = max(maxdiff, abs(avg8 - nbr[k]));
|
|
663
|
+
if (nbr[k] > ctr) { localmax = false; }
|
|
664
|
+
}
|
|
665
|
+
let ratio = div_cr(f32(abs(ctr - avg8)), f32(max(maxdiff, 1)));
|
|
666
|
+
var o = ctr;
|
|
667
|
+
if (localmax && maxdiff > 0 && ratio > 11.0) { o = avg8; }
|
|
668
|
+
outI[base + i] = o;
|
|
669
|
+
}
|
|
670
|
+
}
|
|
671
|
+
|
|
672
|
+
// ==== stage 4: sobel4B on greens (ch1, ch3 -> 2 planes) ====
|
|
673
|
+
fn sobel_one(base: u32, x: i32, y: i32) -> i32 {
|
|
674
|
+
let ul = inA[base + idxc(x - 1, y - 1)];
|
|
675
|
+
let u = inA[base + idxc(x, y - 1)];
|
|
676
|
+
let ur = inA[base + idxc(x + 1, y - 1)];
|
|
677
|
+
let l = inA[base + idxc(x - 1, y)];
|
|
678
|
+
let c = inA[base + idxc(x, y)];
|
|
679
|
+
let r = inA[base + idxc(x + 1, y)];
|
|
680
|
+
let dl = inA[base + idxc(x - 1, y + 1)];
|
|
681
|
+
let d = inA[base + idxc(x, y + 1)];
|
|
682
|
+
let dr = inA[base + idxc(x + 1, y + 1)];
|
|
683
|
+
// integer sums are exact in f32 (< 2^24) \u2014 matches Rust's f32 adds bit-for-bit
|
|
684
|
+
let g1 = mulr(f32((ul + 2 * u + ur) - (l + 2 * c + r)), SOBEL_SC);
|
|
685
|
+
let g2 = mulr(f32((ul + 2 * l + dl) - (u + 2 * c + d)), SOBEL_SC);
|
|
686
|
+
let g3 = mulr(f32((dl + 2 * d + dr) - (l + 2 * c + r)), SOBEL_SC);
|
|
687
|
+
let g4 = mulr(f32((ur + 2 * r + dr) - (u + 2 * c + d)), SOBEL_SC);
|
|
688
|
+
let sumsq = addr(addr(addr(mulr(g2, g2), mulr(g1, g1)), mulr(g4, g4)), mulr(g3, g3));
|
|
689
|
+
return fcl(16776960.0 * sumsq); // 256*65535 exact
|
|
690
|
+
}
|
|
691
|
+
@compute @workgroup_size(16, 16)
|
|
692
|
+
fn _d3(@builtin(global_invocation_id) g: vec3<u32>) {
|
|
693
|
+
if (g.x >= P.w || g.y >= P.h) { return; }
|
|
694
|
+
let i = g.y * P.w + g.x;
|
|
695
|
+
outI[i] = sobel_one(P.n, i32(g.x), i32(g.y));
|
|
696
|
+
outI[P.n + i] = sobel_one(3u * P.n, i32(g.x), i32(g.y));
|
|
697
|
+
}
|
|
698
|
+
|
|
699
|
+
// ==== stage 5: spatialNR ====
|
|
700
|
+
@compute @workgroup_size(16, 16)
|
|
701
|
+
fn _d4(@builtin(global_invocation_id) g: vec3<u32>) {
|
|
702
|
+
if (g.x >= P.w || g.y >= P.h) { return; }
|
|
703
|
+
let n = P.n; let i = g.y * P.w + g.x;
|
|
704
|
+
outI[i] = inA[i];
|
|
705
|
+
outI[2u * n + i] = inA[2u * n + i];
|
|
706
|
+
let c1i = inA[n + i];
|
|
707
|
+
let c3i = inA[3u * n + i];
|
|
708
|
+
if (g.x == 0u || g.y == 0u || g.x == P.w - 1u || g.y == P.h - 1u) {
|
|
709
|
+
outI[n + i] = c1i;
|
|
710
|
+
outI[3u * n + i] = c3i;
|
|
711
|
+
return;
|
|
712
|
+
}
|
|
713
|
+
let edge = max(mulr(f32(inB[i]), SOBEL_SC), mulr(f32(inB[n + i]), SOBEL_SC));
|
|
714
|
+
let c1 = f32(c1i);
|
|
715
|
+
let c3 = f32(c3i);
|
|
716
|
+
let gavg = f32(c1i + c3i) * 0.5; // exact
|
|
717
|
+
let base = mulr(gavg, subr(1.0, edge));
|
|
718
|
+
outI[n + i] = fcl(addr(base, mulr(c1, edge)));
|
|
719
|
+
outI[3u * n + i] = fcl(addr(base, mulr(c3, edge)));
|
|
720
|
+
}
|
|
721
|
+
|
|
722
|
+
// ==== stages 6-7: separable bilateral (mirror boundary) ====
|
|
723
|
+
fn mir(idx: i32, nn: i32) -> u32 {
|
|
724
|
+
var i = abs(idx);
|
|
725
|
+
if (i >= nn) { i = 2 * nn - 2 - i; }
|
|
726
|
+
return u32(i);
|
|
727
|
+
}
|
|
728
|
+
fn bilat(gx: u32, gy: u32, axis: u32) {
|
|
729
|
+
let n = P.n; let i = gy * P.w + gx;
|
|
730
|
+
var acc = vec4<f32>(0.0);
|
|
731
|
+
var ws = vec4<f32>(0.0);
|
|
732
|
+
let cv = vec4<i32>(inA[i], inA[n + i], inA[2u * n + i], inA[3u * n + i]);
|
|
733
|
+
for (var k = 0u; k < 9u; k++) {
|
|
734
|
+
let off = i32(k) - 4;
|
|
735
|
+
var si: u32;
|
|
736
|
+
if (axis == 0u) { si = mir(i32(gy) + off, i32(P.h)) * P.w + gx; }
|
|
737
|
+
else { si = gy * P.w + mir(i32(gx) + off, i32(P.w)); }
|
|
738
|
+
// channels 0,2: 9-tap, range srb
|
|
739
|
+
for (var ci = 0u; ci < 4u; ci += 2u) {
|
|
740
|
+
let s = inA[ci * n + si];
|
|
741
|
+
let sf = f32(s);
|
|
742
|
+
let rw = mulr(subr(1.0, clamp(mulr(f32(abs(s - cv[ci])), BIL_SRB), 0.0, 1.0)), T9[k]);
|
|
743
|
+
acc[ci] = addr(acc[ci], mulr(rw, sf));
|
|
744
|
+
ws[ci] = addr(ws[ci], rw);
|
|
745
|
+
}
|
|
746
|
+
// channels 1,3: 7-tap (k in 1..=7), range sg
|
|
747
|
+
if (k >= 1u && k <= 7u) {
|
|
748
|
+
for (var ci = 1u; ci < 4u; ci += 2u) {
|
|
749
|
+
let s = inA[ci * n + si];
|
|
750
|
+
let sf = f32(s);
|
|
751
|
+
let rw = mulr(subr(1.0, clamp(mulr(f32(abs(s - cv[ci])), BIL_SG), 0.0, 1.0)), T7[k - 1u]);
|
|
752
|
+
acc[ci] = addr(acc[ci], mulr(rw, sf));
|
|
753
|
+
ws[ci] = addr(ws[ci], rw);
|
|
754
|
+
}
|
|
755
|
+
}
|
|
756
|
+
}
|
|
757
|
+
for (var ci = 0u; ci < 4u; ci++) {
|
|
758
|
+
outI[ci * n + i] = fcl(div_cr(acc[ci], ws[ci]));
|
|
759
|
+
}
|
|
760
|
+
}
|
|
761
|
+
@compute @workgroup_size(16, 16)
|
|
762
|
+
fn _d5(@builtin(global_invocation_id) g: vec3<u32>) {
|
|
763
|
+
if (g.x >= P.w || g.y >= P.h) { return; }
|
|
764
|
+
bilat(g.x, g.y, 1u);
|
|
765
|
+
}
|
|
766
|
+
@compute @workgroup_size(16, 16)
|
|
767
|
+
fn _d6(@builtin(global_invocation_id) g: vec3<u32>) {
|
|
768
|
+
if (g.x >= P.w || g.y >= P.h) { return; }
|
|
769
|
+
bilat(g.x, g.y, 0u);
|
|
770
|
+
}
|
|
771
|
+
|
|
772
|
+
// ==== stage 8: ratio_creation (denoised image -> 2 f32 planes) ====
|
|
773
|
+
@compute @workgroup_size(16, 16)
|
|
774
|
+
fn _d7(@builtin(global_invocation_id) g: vec3<u32>) {
|
|
775
|
+
if (g.x >= P.w || g.y >= P.h) { return; }
|
|
776
|
+
let x = i32(g.x); let y = i32(g.y);
|
|
777
|
+
let n = P.n; let i = g.y * P.w + g.x;
|
|
778
|
+
let c0 = f32(inA[i]);
|
|
779
|
+
let c2 = f32(inA[2u * n + i]);
|
|
780
|
+
let g13 = inA[n + i] + inA[3u * n + i];
|
|
781
|
+
let sr = f32(g13 + inA[n + idxc(x + 1, y)] + inA[3u * n + idxc(x, y - 1)]); // exact
|
|
782
|
+
let gr = addr(sr * 0.25, RC_P); // *0.25 exact
|
|
783
|
+
var rr = 1.0;
|
|
784
|
+
if (gr > 0.0) { rr = div_cr(c0, gr); }
|
|
785
|
+
outF[i] = rr;
|
|
786
|
+
let sb = f32(g13 + inA[n + idxc(x, y + 1)] + inA[3u * n + idxc(x - 1, y)]);
|
|
787
|
+
let gb = addr(sb * 0.25, RC_P);
|
|
788
|
+
var rb = 1.0;
|
|
789
|
+
if (gb > 0.0) { rb = div_cr(c2, gb); }
|
|
790
|
+
outF[n + i] = rb;
|
|
791
|
+
}
|
|
792
|
+
|
|
793
|
+
// ==== stage 9: ratio_reconstruct (sharp=inA, sob=inB, ratio=inF -> outI) ====
|
|
794
|
+
fn rz(base: u32, x: i32, y: i32) -> f32 { // read_texture semantics: zero out of bounds
|
|
795
|
+
if (x < 0 || y < 0 || x >= i32(P.w) || y >= i32(P.h)) { return 0.0; }
|
|
796
|
+
return inF[base + u32(y) * P.w + u32(x)];
|
|
797
|
+
}
|
|
798
|
+
fn edge_at(base: u32, xi: i32, yi: i32) -> f32 {
|
|
799
|
+
let a = rz(base, xi - 1, yi - 1); // r(-1,-1) args (dy,dx) -> rz(x+dx, y+dy)
|
|
800
|
+
let b = rz(base, xi - 1, yi); // r(0,-1)
|
|
801
|
+
let c = rz(base, xi, yi - 1); // r(-1,0)
|
|
802
|
+
let d = rz(base, xi, yi + 1); // r(1,0)
|
|
803
|
+
let e = rz(base, xi + 1, yi); // r(0,1)
|
|
804
|
+
let f = rz(base, xi + 1, yi + 1); // r(1,1)
|
|
805
|
+
let g3a = rz(base, xi - 1, yi + 1); // r(1,-1)
|
|
806
|
+
let g3b = rz(base, xi + 1, yi - 1); // r(-1,1)
|
|
807
|
+
let g1 = mulr(addr(addr(addr(subr(subr(-2.0 * a, b), c), d), e), 2.0 * f), RR_C16);
|
|
808
|
+
let g2 = mulr(addr(addr(addr(subr(subr(-2.0 * b, a), g3a), g3b), 2.0 * e), f), RR_C16);
|
|
809
|
+
let g3 = mulr(addr(subr(addr(subr(a, mulr(3.0, g3a)), 2.0 * c), d), g3b), RR_C16);
|
|
810
|
+
let g4 = mulr(subr(addr(addr(subr(subr(c, b), 2.0 * d), 2.0 * g3b), e), f), RR_C16);
|
|
811
|
+
let m = max(max(abs(g1), abs(g2)), max(abs(g3), abs(g4)));
|
|
812
|
+
return clamp(mulr(m, 8000.0), 0.0, 1.0);
|
|
813
|
+
}
|
|
814
|
+
fn recon_path(ci: f32, rt: f32, edge: f32, gs: f32, ss: f32) -> i32 {
|
|
815
|
+
let recon0 = addr(mulr(mulr(mulr(rt, 0.25), subr(1.0, edge)), gs), mulr(edge, ci));
|
|
816
|
+
let a2 = clamp(mulr(recon0, RR_C13), 0.0, 65535.0);
|
|
817
|
+
let sc = mulr(ss, RR_C18);
|
|
818
|
+
let t = sc * sc;
|
|
819
|
+
let wgt = clamp(t * t, 0.0, 1.0);
|
|
820
|
+
let recon1 = addr(mulr(subr(1.0, wgt), recon0), mulr(wgt, ci));
|
|
821
|
+
return fcl(addr(addr(mulr(recon0, subr(1.0, a2)), 0.5), mulr(recon1, a2)));
|
|
822
|
+
}
|
|
823
|
+
@compute @workgroup_size(16, 16)
|
|
824
|
+
fn _d8(@builtin(global_invocation_id) g: vec3<u32>) {
|
|
825
|
+
if (g.x >= P.w || g.y >= P.h) { return; }
|
|
826
|
+
let x = i32(g.x); let y = i32(g.y);
|
|
827
|
+
let n = P.n; let i = g.y * P.w + g.x;
|
|
828
|
+
let f13 = inA[n + i] + inA[3u * n + i];
|
|
829
|
+
let s01 = inB[i] + inB[n + i];
|
|
830
|
+
// R path
|
|
831
|
+
let gs_r = f32(f13 + inA[n + idxc(x + 1, y)] + inA[3u * n + idxc(x, y - 1)]);
|
|
832
|
+
let ss_r = f32(s01 + inB[idxc(x + 1, y)] + inB[n + idxc(x, y - 1)]);
|
|
833
|
+
outI[i] = recon_path(f32(inA[i]), inF[i], edge_at(0u, x, y), gs_r, ss_r);
|
|
834
|
+
outI[n + i] = inA[n + i];
|
|
835
|
+
// B path
|
|
836
|
+
let gs_b = f32(f13 + inA[3u * n + idxc(x - 1, y)] + inA[n + idxc(x, y + 1)]);
|
|
837
|
+
let ss_b = f32(inB[idxc(x, y + 1)] + s01 + inB[n + idxc(x - 1, y)]);
|
|
838
|
+
outI[2u * n + i] = recon_path(f32(inA[2u * n + i]), inF[n + i], edge_at(n, x, y), gs_b, ss_b);
|
|
839
|
+
outI[3u * n + i] = inA[3u * n + i];
|
|
840
|
+
}
|
|
841
|
+
|
|
842
|
+
// ==== stage 10: white_balance + delinearize ====
|
|
843
|
+
@compute @workgroup_size(16, 16)
|
|
844
|
+
fn _d9(@builtin(global_invocation_id) g: vec3<u32>) {
|
|
845
|
+
if (g.x >= P.w || g.y >= P.h) { return; }
|
|
846
|
+
let n = P.n; let i = g.y * P.w + g.x;
|
|
847
|
+
for (var c = 0u; c < 4u; c++) {
|
|
848
|
+
let q = f32(inA[c * n + i]) * 0.0000152587890625; // /65536 exact
|
|
849
|
+
let val = (q * q) * 65536.0; // exact scale of the rounded square
|
|
850
|
+
let v = addr(64.5, mulr(P.wb[c], subr(val, 64.0)));
|
|
851
|
+
outI[c * n + i] = fcl(v);
|
|
852
|
+
}
|
|
853
|
+
}
|
|
854
|
+
|
|
855
|
+
// ==== stage 11: half_premium_balanced -> planar RGB ====
|
|
856
|
+
@compute @workgroup_size(16, 16)
|
|
857
|
+
fn _d10(@builtin(global_invocation_id) g: vec3<u32>) {
|
|
858
|
+
if (g.x >= P.w || g.y >= P.h) { return; }
|
|
859
|
+
let x = i32(g.x); let y = i32(g.y);
|
|
860
|
+
let n = P.n; let i = g.y * P.w + g.x;
|
|
861
|
+
let rc = inA[i];
|
|
862
|
+
let bc = inA[2u * n + i];
|
|
863
|
+
let avg_r = (rc + inA[idxc(x - 1, y)] + inA[idxc(x, y + 1)] + inA[idxc(x - 1, y + 1)]) >> 2u;
|
|
864
|
+
let avg_b = (bc + inA[2u * n + idxc(x + 1, y)] + inA[2u * n + idxc(x, y - 1)] + inA[2u * n + idxc(x + 1, y - 1)]) >> 2u;
|
|
865
|
+
outI[i] = fcl(addr(mulr(0.75, f32(rc)), 0.25 * f32(avg_r)));
|
|
866
|
+
outI[2u * n + i] = fcl(addr(mulr(0.75, f32(bc)), 0.25 * f32(avg_b)));
|
|
867
|
+
outI[n + i] = fcl(0.5 * f32(inA[n + i] + inA[3u * n + i]));
|
|
868
|
+
}
|
|
869
|
+
|
|
870
|
+
// ==== shadow perceptual measure ====
|
|
871
|
+
fn shadow_pm(r: f32, gg: f32, b: f32) -> f32 {
|
|
872
|
+
let rp = addr(max(subr(r, 64.0), 0.0), 64.0);
|
|
873
|
+
let gp2 = addr(max(subr(gg, 64.0), 0.0), 64.0);
|
|
874
|
+
let bp = addr(max(subr(b, 64.0), 0.0), 64.0);
|
|
875
|
+
let sumsq = addr(addr(mulr(rp, rp), mulr(gp2, gp2)), mulr(bp, bp));
|
|
876
|
+
let ssum = addr(addr(addr(rp, SD_EPS), gp2), bp);
|
|
877
|
+
return div_cr(sumsq, ssum);
|
|
878
|
+
}
|
|
879
|
+
|
|
880
|
+
// ==== stage 12a: measure ====
|
|
881
|
+
@compute @workgroup_size(16, 16)
|
|
882
|
+
fn _d11(@builtin(global_invocation_id) g: vec3<u32>) {
|
|
883
|
+
if (g.x >= P.w || g.y >= P.h) { return; }
|
|
884
|
+
let n = P.n; let i = g.y * P.w + g.x;
|
|
885
|
+
outF[i] = shadow_pm(f32(inA[i]), f32(inA[n + i]), f32(inA[2u * n + i]));
|
|
886
|
+
}
|
|
887
|
+
|
|
888
|
+
// ==== stage 12b: 5x5 gauss of the measure (reflect boundary, PINNED accumulate order) ====
|
|
889
|
+
fn refl(idx: i32, nn: i32) -> u32 {
|
|
890
|
+
var i = abs(idx);
|
|
891
|
+
if (i >= nn) { i = 2 * nn - 1 - i; }
|
|
892
|
+
return u32(i);
|
|
893
|
+
}
|
|
894
|
+
@compute @workgroup_size(16, 16)
|
|
895
|
+
fn _d12(@builtin(global_invocation_id) g: vec3<u32>) {
|
|
896
|
+
if (g.x >= P.w || g.y >= P.h) { return; }
|
|
897
|
+
var acc = 0.0;
|
|
898
|
+
for (var dy = -2; dy <= 2; dy++) {
|
|
899
|
+
let rb = refl(i32(g.y) + dy, i32(P.h)) * P.w;
|
|
900
|
+
for (var dx = -2; dx <= 2; dx++) {
|
|
901
|
+
acc = addr(acc, mulr(GAUSS[u32(dy + 2) * 5u + u32(dx + 2)], inF[rb + refl(i32(g.x) + dx, i32(P.w))]));
|
|
902
|
+
}
|
|
903
|
+
}
|
|
904
|
+
outF[g.y * P.w + g.x] = mulr(acc, INV273);
|
|
905
|
+
}
|
|
906
|
+
|
|
907
|
+
// ==== stage 13: shadow_desat ====
|
|
908
|
+
@compute @workgroup_size(16, 16)
|
|
909
|
+
fn _d13(@builtin(global_invocation_id) g: vec3<u32>) {
|
|
910
|
+
if (g.x >= P.w || g.y >= P.h) { return; }
|
|
911
|
+
let n = P.n; let i = g.y * P.w + g.x;
|
|
912
|
+
let ri = inA[i]; let gi = inA[n + i]; let bi = inA[2u * n + i];
|
|
913
|
+
outI[i] = ri; outI[n + i] = gi; outI[2u * n + i] = bi;
|
|
914
|
+
let r = f32(ri); let gg = f32(gi); let b = f32(bi);
|
|
915
|
+
let delta = subr(inF[i], 64.0);
|
|
916
|
+
var factor = SD_STRENGTH;
|
|
917
|
+
if (delta >= 0.0) {
|
|
918
|
+
factor = addr(mulr(clamp(div_cr(abs(delta), 6.0), 0.0, 1.0), SD_OMS), SD_STRENGTH);
|
|
919
|
+
}
|
|
920
|
+
let pm = shadow_pm(r, gg, b);
|
|
921
|
+
let m76 = mulr(subr(1.0, factor), pm);
|
|
922
|
+
let r_is_max = r >= b && r > gg;
|
|
923
|
+
let b_is_max = !r_is_max && b > gg;
|
|
924
|
+
if (r_is_max) { outI[i] = fcl(addr(addr(mulr(factor, r), m76), 0.5)); }
|
|
925
|
+
if (b_is_max) { outI[2u * n + i] = fcl(addr(addr(mulr(factor, b), m76), 0.5)); }
|
|
926
|
+
}
|
|
927
|
+
|
|
928
|
+
// ==== stage 14: matrix + gamut compress + Log3G10 ====
|
|
929
|
+
@compute @workgroup_size(16, 16)
|
|
930
|
+
fn _d14(@builtin(global_invocation_id) g: vec3<u32>) {
|
|
931
|
+
if (g.x >= P.w || g.y >= P.h) { return; }
|
|
932
|
+
let n = P.n; let i = g.y * P.w + g.x;
|
|
933
|
+
var t: vec3<f32>;
|
|
934
|
+
var tc: vec3<f32>;
|
|
935
|
+
var xr: vec3<f32>;
|
|
936
|
+
var m: vec3<f32>;
|
|
937
|
+
for (var k = 0u; k < 3u; k++) {
|
|
938
|
+
let v = mulr(f32(inA[k * n + i]), M16);
|
|
939
|
+
t[k] = div_cr(v, P.nd[k]);
|
|
940
|
+
tc[k] = clamp(t[k], 0.0, 65535.0);
|
|
941
|
+
}
|
|
942
|
+
let mxc = max(tc[0], max(tc[1], tc[2]));
|
|
943
|
+
let mxu = max(t[0], max(t[1], t[2]));
|
|
944
|
+
for (var k = 0u; k < 3u; k++) {
|
|
945
|
+
var renorm = t[k];
|
|
946
|
+
if (mxc > 0.0) { renorm = mulr(mxu, div_cr(tc[k], mxc)); }
|
|
947
|
+
xr[k] = mulr(renorm, P.nd[k]);
|
|
948
|
+
}
|
|
949
|
+
let r1 = addr(addr(addr(mulr(P.smr.x, xr[0]), mulr(P.smr.y, xr[1])), mulr(P.smr.z, xr[2])), P.smr.w);
|
|
950
|
+
let g1 = addr(addr(addr(mulr(P.smg.x, xr[0]), mulr(P.smg.y, xr[1])), mulr(P.smg.z, xr[2])), P.smg.w);
|
|
951
|
+
let b1 = addr(addr(addr(mulr(P.smb.x, xr[0]), mulr(P.smb.y, xr[1])), mulr(P.smb.z, xr[2])), P.smb.w);
|
|
952
|
+
m[0] = subr(mulr(r1, 65535.0), 64.0);
|
|
953
|
+
m[1] = subr(mulr(g1, 65535.0), 64.0);
|
|
954
|
+
m[2] = subr(mulr(b1, 65535.0), 64.0);
|
|
955
|
+
let mx2 = max(m[0], max(m[1], m[2]));
|
|
956
|
+
if (mx2 > 0.0) {
|
|
957
|
+
for (var k = 0u; k < 3u; k++) {
|
|
958
|
+
let md = div_cr(m[k], mx2);
|
|
959
|
+
var comp: f32;
|
|
960
|
+
if (md < -1.0) { comp = 0.0; }
|
|
961
|
+
else if (md < GC_T) { comp = mulr(pow_gp(max(mulr(addr(md, 1.0), GC_K), 0.0)), GC_T); }
|
|
962
|
+
else { comp = md; }
|
|
963
|
+
m[k] = mulr(comp, mx2);
|
|
964
|
+
}
|
|
965
|
+
}
|
|
966
|
+
for (var k = 0u; k < 3u; k++) {
|
|
967
|
+
let lin = addr(mulr(mulr(m[k], M16), P.exposure), LG_PA);
|
|
968
|
+
var enc: f32;
|
|
969
|
+
if (lin < 0.0) { enc = mulr(lin, LG_A); }
|
|
970
|
+
else { enc = mulr(LG_C, log10_cr(max(addr(mulr(lin, LG_B), 1.0), 1.0e-30))); }
|
|
971
|
+
outI[k * n + i] = fcl(addr(mulr(enc, 65535.0), 0.5));
|
|
972
|
+
}
|
|
973
|
+
}
|
|
974
|
+
|
|
975
|
+
// ==== stage 15: ipp2 identity LUTs -> interleaved u16 (2 pixels/thread for u32 packing) ====
|
|
976
|
+
fn ipp2_one(p: f32) -> u32 {
|
|
977
|
+
let c = clamp(mulr(p, M16), 0.0, 1.0);
|
|
978
|
+
let c2 = clamp(mulr(c, 65535.0), 0.0, 65535.0);
|
|
979
|
+
let idx = trunc(c2);
|
|
980
|
+
let wv = clamp(mulr(idx, M16), 0.0, 1.0);
|
|
981
|
+
return u32(floor(mulr(wv, 65535.0)));
|
|
982
|
+
}
|
|
983
|
+
@compute @workgroup_size(16, 16)
|
|
984
|
+
fn _d15(@builtin(global_invocation_id) g: vec3<u32>) {
|
|
985
|
+
let tid = g.y * (P.w / 2u) + g.x; // over n/2 (w even)
|
|
986
|
+
if (g.x >= P.w / 2u || g.y >= P.h) { return; }
|
|
987
|
+
let n = P.n;
|
|
988
|
+
let i0 = 2u * tid;
|
|
989
|
+
var v: array<u32, 6>;
|
|
990
|
+
for (var px = 0u; px < 2u; px++) {
|
|
991
|
+
for (var k = 0u; k < 3u; k++) {
|
|
992
|
+
v[px * 3u + k] = ipp2_one(f32(inA[k * n + i0 + px]));
|
|
993
|
+
}
|
|
994
|
+
}
|
|
995
|
+
outU[3u * tid] = v[0] | (v[1] << 16u);
|
|
996
|
+
outU[3u * tid + 1u] = v[2] | (v[3] << 16u);
|
|
997
|
+
outU[3u * tid + 2u] = v[4] | (v[5] << 16u);
|
|
998
|
+
}
|
|
999
|
+
|
|
1000
|
+
// ==== viewer pack: interleaved u16 Log3G10 -> RGBA8 (high byte, alpha 255) ====
|
|
1001
|
+
// Mirrors frame.rs log3g10_rgba8 \u2014 the Full-tier RGBA the CPU lane returns.
|
|
1002
|
+
fn out16(j: u32) -> u32 { return (outU[j >> 1u] >> ((j & 1u) * 16u)) & 0xFFFFu; }
|
|
1003
|
+
@compute @workgroup_size(16, 16)
|
|
1004
|
+
fn _d16(@builtin(global_invocation_id) g: vec3<u32>) {
|
|
1005
|
+
if (g.x >= P.w || g.y >= P.h) { return; }
|
|
1006
|
+
let i = g.y * P.w + g.x;
|
|
1007
|
+
let r = out16(3u * i) >> 8u;
|
|
1008
|
+
let gg = out16(3u * i + 1u) >> 8u;
|
|
1009
|
+
let b = out16(3u * i + 2u) >> 8u;
|
|
1010
|
+
outR[i] = r | (gg << 8u) | (b << 16u) | 0xFF000000u;
|
|
1011
|
+
}
|
|
1012
|
+
`;var A1=1,P1=4,B=8,N=64,S=128,k1=1,b1=!1;function O(n,r){return b1||(b1=!0,console.warn(`[media/gpu/r3d] GPU lane disabled after failure at ${n}; falling back to CPU:`,r)),null}var _=Math.fround;function V(n){let r=Number(n).toString();return/[.e]/.test(r)||(r+=".0"),r}function A(n){let r=_(n),e=_(n-r);return`vec2<f32>(${V(r)}, ${V(e)})`}function V1(){let n={};for(let[i,t]of Object.entries(s1))n[i]=k(t);let r="",e=(i,t)=>{r+=`const ${i}: f32 = ${V(t)};
|
|
1013
|
+
`};e("M16",n.m16),e("LG_A",n.lg_a),e("LG_B",n.lg_b),e("LG_C",n.lg_c),e("LG_PA",n.lg_pa),e("GC_K",n.gc_k),e("GC_P",n.gc_p),e("GC_T",n.gc_t),e("SD_EPS",n.sd_eps),e("INV273",n.inv273),e("RR_C13",n.rr_c13),e("RR_C16",n.rr_c16),e("RR_C18",n.rr_c18),e("BIL_SG",n.bil_sg),e("BIL_SRB",n.bil_srb),e("SOBEL_SC",n.sobel_sc),e("RC_P",_(.1)),e("SD_STRENGTH",_(.1)),e("SD_OMS",_(1-_(.1))),e("C_1_9",_(1/9)),e("C_1_11",_(1/11)),e("C_1_13",_(1/13)),e("C_1_15",_(1/15)),e("C_1_17",_(1/17)),e("C_1_19",_(1/19)),e("C_1_21",_(1/21)),e("C_1_24",_(1/24)),e("C_1_120",_(1/120)),e("C_1_720",_(1/720)),r+=`const T7 = array<f32,7>(${Array.from({length:7},(i,t)=>V(n[`bil_t7_${t}`])).join(", ")});
|
|
1014
|
+
`,r+=`const T9 = array<f32,9>(${Array.from({length:9},(i,t)=>V(n[`bil_t9_${t}`])).join(", ")});
|
|
1015
|
+
`;{let i=[];for(let t=0;t<5;t++)for(let a=0;a<5;a++)i.push(V(n[`gauss55_${t}${a}`]));r+=`const GAUSS = array<f32,25>(${i.join(", ")});
|
|
1016
|
+
`}return r+=`const DD_1_3 = ${A(1/3)};
|
|
1017
|
+
const DD_1_5 = ${A(1/5)};
|
|
1018
|
+
const DD_1_6 = ${A(1/6)};
|
|
1019
|
+
const DD_1_7 = ${A(1/7)};
|
|
1020
|
+
`,r+=`const DD_LOG10_2 = ${A(Math.log10(2))};
|
|
1021
|
+
const DD_INVLN10 = ${A(1/Math.log(10))};
|
|
1022
|
+
`,r+=`const DD_INVLN2 = ${A(1/Math.log(2))};
|
|
1023
|
+
const DD_LN2 = ${A(Math.log(2))};
|
|
1024
|
+
`,r+=`const EXP2TAB = array<vec2<f32>,16>(${Array.from({length:16},(i,t)=>A(Math.pow(2,t/16))).join(", ")});
|
|
1025
|
+
`,d1.replace("//__CONSTS__",r)}var F=[{entry:"_d0",binds:{1:"PL",4:"B0"},checkpoint:null},{entry:"_d1",binds:{1:"B0",4:"B1"},checkpoint:{name:"s0",buf:"B1",words:n=>4*n}},{entry:"_d2",binds:{1:"B1",4:"B0"},checkpoint:{name:"s1",buf:"B0",words:n=>4*n}},{entry:"_d3",binds:{1:"B0",4:"SOB"},checkpoint:{name:"s2",buf:"SOB",words:n=>2*n}},{entry:"_d4",binds:{1:"B0",2:"SOB",4:"B1"},checkpoint:{name:"s3",buf:"B1",words:n=>4*n}},{entry:"_d5",binds:{1:"B1",4:"B2"},checkpoint:{name:"s4",buf:"B2",words:n=>4*n}},{entry:"_d6",binds:{1:"B2",4:"B0"},checkpoint:{name:"s5",buf:"B0",words:n=>4*n}},{entry:"_d7",binds:{1:"B0",5:"B2"},checkpoint:{name:"s6",buf:"B2",words:n=>2*n}},{entry:"_d8",binds:{1:"B1",2:"SOB",3:"B2",4:"B0"},checkpoint:{name:"s7",buf:"B0",words:n=>4*n}},{entry:"_d9",binds:{1:"B0",4:"B1"},checkpoint:{name:"s8",buf:"B1",words:n=>4*n}},{entry:"_d10",binds:{1:"B1",4:"B0"},checkpoint:{name:"s9",buf:"B0",words:n=>3*n}},{entry:"_d11",binds:{1:"B0",5:"B2"},checkpoint:null},{entry:"_d12",binds:{3:"B2",5:"SOB"},checkpoint:{name:"s10",buf:"SOB",words:n=>n}},{entry:"_d13",binds:{1:"B0",3:"SOB",4:"B1"},checkpoint:null},{entry:"_d14",binds:{1:"B1",4:"B0"},checkpoint:null},{entry:"_d15",binds:{1:"B0",7:"OUT"},checkpoint:{name:"s11",buf:"OUT",words:n=>3*n>>1},halfW:!0},{entry:"_d16",binds:{7:"OUT",8:"RGBA"},checkpoint:{name:"s12",buf:"RGBA",words:n=>n}}],$=class n{constructor(r,e,i,t,a){this.essenceBuf=null;this.essenceCap=0;this.jobsBuf=null;this.jobsCap=0;this.coeffBuf=null;this.coeffCap=0;this.frames=new Map;this.queueTail=Promise.resolve();this.device=r,this.vldPipe=e,this.xfPipe=i,this.developPipes=t,this.tabBuf=a}static async create(r){let e=async(g,p)=>{let y=r.createShaderModule({code:g}),h=(await y.getCompilationInfo()).messages.filter(x=>x.type==="error");if(h.length)throw new Error(`${p}: ${h.map(x=>`${x.lineNum}:${x.linePos} ${x.message}`).join(" | ")}`);return y},i=await e(l1,"r3d-vld"),t=await e(f1,"r3d-xform"),a=await e(V1(),"r3d-develop"),o=(g,p)=>r.createComputePipelineAsync({layout:"auto",compute:{module:g,entryPoint:p}}),[b,m,...f]=await Promise.all([o(i,"main"),o(t,"main"),...F.map(g=>o(a,g.entry))]),v=r.createBuffer({size:1920*4,usage:S|B}),u=new Uint32Array(1920);return u.set(t1,0),u.set(a1,1024),u.set(i1,1088),u.set(u1,1152),u.set(new Uint32Array(z.buffer,z.byteOffset,256),1408),r.queue.writeBuffer(v,0,u),new n(r,b,m,f,v)}run(r,e){let i=this.queueTail.then(()=>this.decodeInner(r,e),()=>this.decodeInner(r,e));return this.queueTail=i.catch(()=>{}),i}frameBuffers(r){let e=`${r.canvasW}x${r.canvasH}`,i=this.frames.get(e);if(i)return i;let t=this.device,a=r.canvasW*r.canvasH,o=(m,f=0)=>t.createBuffer({size:m,usage:S|P1|B|f}),b={w:r.canvasW,h:r.canvasH,n:a,PL:o(4*a*4),B0:o(4*a*4),B1:o(4*a*4),B2:o(4*a*4),SOB:o(2*a*4),OUT:o(3*a*2),RGBA:o(a*4),uni:t.createBuffer({size:112,usage:N|B}),params1:t.createBuffer({size:16,usage:N|B}),params2:t.createBuffer({size:16,usage:N|B}),developBinds:[],flashCopy:{src:null,dst:null,bytes:4*a*4}};return b.flashCopy.src=b.B1,b.flashCopy.dst=b.B0,b.developBinds=F.map((m,f)=>{let v=[{binding:0,resource:{buffer:b.uni}}];for(let[u,g]of Object.entries(m.binds))v.push({binding:Number(u),resource:{buffer:b[g]}});return t.createBindGroup({layout:this.developPipes[f].getBindGroupLayout(0),entries:v})}),this.frames.set(e,b),b}growEssence(r){return(!this.essenceBuf||this.essenceCap<r)&&(this.essenceBuf?.destroy?.(),this.essenceCap=Math.max(r,this.essenceCap*2),this.essenceBuf=this.device.createBuffer({size:this.essenceCap,usage:S|B})),this.essenceBuf}growJobs(r){return(!this.jobsBuf||this.jobsCap<r)&&(this.jobsBuf?.destroy?.(),this.jobsCap=Math.max(r,this.jobsCap*2),this.jobsBuf=this.device.createBuffer({size:this.jobsCap,usage:S|B})),this.jobsBuf}growCoeff(r){return(!this.coeffBuf||this.coeffCap<r)&&(this.coeffBuf?.destroy?.(),this.coeffCap=Math.max(r,this.coeffCap*2),this.coeffBuf=this.device.createBuffer({size:this.coeffCap,usage:S|B})),this.coeffBuf}async decodeInner(r,e){let i=performance.now(),t=this.device,a=q(r),o=a.canvasW*a.canvasH,m=(t.limits??{}).maxStorageBufferBindingSize??128*1024*1024,f=a.segments.length*4,v=f*2048*4;if(4*o*4>m||v>m||r.length+16>m)throw new Error(`frame exceeds storage binding limit (${a.canvasW}x${a.canvasH}, ${f} jobs)`);let u=this.frameBuffers(a),g=new Uint32Array(256);for(let s=0;s<256;s++)g[s]=a.table512[s];t.queue.writeBuffer(this.tabBuf,1664*4,g);let p=new Uint8Array(r.length+8+3&-4);p.set(r);let y=this.growEssence(p.length);t.queue.writeBuffer(y,0,p.buffer,0,p.length),p.length<this.essenceCap&&t.queue.writeBuffer(y,p.length,new Uint8Array(Math.min(this.essenceCap-p.length,16)));let P=o1(a),h=this.growJobs(P.byteLength);t.queue.writeBuffer(h,0,P);let x=this.growCoeff(v),C=f*8;t.queue.writeBuffer(u.params1,0,new Uint32Array([f,a.canvasW,a.canvasH,o])),t.queue.writeBuffer(u.params2,0,new Uint32Array([C,a.canvasW,a.canvasH,o]));let R=e.paramsBits;{let s=new ArrayBuffer(112),c=new Uint32Array(s),l=new Float32Array(s);c[0]=a.canvasW,c[1]=a.canvasH,c[2]=o,c[3]=0,l[4]=0,l[5]=1,l[6]=k(R.exposureBits),l[7]=0;for(let d=0;d<4;d++)l[8+d]=k(R.wbBits[d]);for(let d=0;d<3;d++)l[12+d]=k(R.smatrixBits[12+d]);l[15]=0;for(let d=0;d<12;d++)l[16+d]=k(R.smatrixBits[d]);t.queue.writeBuffer(u.uni,0,s)}let g1=t.createBindGroup({layout:this.vldPipe.getBindGroupLayout(0),entries:[{binding:0,resource:{buffer:u.params1}},{binding:1,resource:{buffer:y}},{binding:2,resource:{buffer:h}},{binding:3,resource:{buffer:this.tabBuf}},{binding:4,resource:{buffer:x}}]}),p1=t.createBindGroup({layout:this.xfPipe.getBindGroupLayout(0),entries:[{binding:0,resource:{buffer:u.params2}},{binding:1,resource:{buffer:h}},{binding:2,resource:{buffer:this.tabBuf}},{binding:3,resource:{buffer:x}},{binding:4,resource:{buffer:u.PL}}]});t.pushErrorScope("out-of-memory"),t.pushErrorScope("validation");{let s=t.createCommandEncoder();s.clearBuffer(x,0,v);let c=s.beginComputePass();c.setPipeline(this.vldPipe),c.setBindGroup(0,g1),c.dispatchWorkgroups(Math.ceil(f/64)),c.end();let l=s.beginComputePass();l.setPipeline(this.xfPipe),l.setBindGroup(0,p1),l.dispatchWorkgroups(Math.ceil(C/16),16),l.end(),t.queue.submit([s.finish()])}let m1=Math.ceil(a.canvasW/16),_1=Math.ceil(a.canvasH/16),y1=Math.ceil(a.canvasW/2/16),L={},U=[],G=(s,c,l,d)=>{let Z=t.createBuffer({size:l,usage:B|A1});d.copyBufferToBuffer(c,0,Z,0,l),U.push({name:s,buf:c,bytes:l,stage:Z})},K=e.captureOnly?new Set(e.captureOnly):null,I=s=>e.captureStages===!0&&(K===null||K.has(s));{let s=t.createCommandEncoder();I("planes")&&G("planes",u.PL,4*o*4,s);for(let c=0;c<F.length;c++){let l=F[c];if(!(l.entry==="_d16"&&!e.wantRgba&&!I("s12"))){if(l.entry==="_d2"&&!R.fixFlashers)s.copyBufferToBuffer(u.flashCopy.src,0,u.flashCopy.dst,0,u.flashCopy.bytes);else{let d=s.beginComputePass();d.setPipeline(this.developPipes[c]),d.setBindGroup(0,u.developBinds[c]),d.dispatchWorkgroups(l.halfW?y1:m1,_1),d.end()}l.checkpoint&&I(l.checkpoint.name)&&G(l.checkpoint.name,u[l.checkpoint.buf],l.checkpoint.words(o)*4,s)}}e.wantLog16&&G("out_log16",u.OUT,3*o*2,s),e.wantRgba&&G("out_rgba",u.RGBA,o*4,s),t.queue.submit([s.finish()])}let J=await t.popErrorScope(),X=await t.popErrorScope();if(J||X){for(let s of U)s.stage.destroy?.();throw new Error(`GPU error: ${(J??X).message}`)}let j=null,T=null,Y=null;for(let s of U){await s.stage.mapAsync(k1);let c=s.stage.getMappedRange().slice(0);s.stage.unmap(),s.stage.destroy?.(),s.name==="out_log16"?j=new Uint16Array(c):s.name==="out_rgba"?T=new Uint8Array(c):s.name==="planes"?(Y=new Int32Array(c),L.planes=c):L[s.name]=c}let Q=s=>{if(!s)return!1;let c=new Uint32Array(s),l=Math.max(1,Math.floor(c.length/4096));for(let d=0;d<c.length;d+=l)if(c[d]!==0)return!1;return!0};if(Q(j?.buffer??null)||Q(T?.buffer??null))throw new Error("GPU output came back unwritten (dispatch killed?)");return{width:a.canvasW,height:a.canvasH,log16:j,rgba:T,stages:L,planes:Y,wallMs:performance.now()-i}}},W=null;function D(){return W||(W=(async()=>{try{let n=await e1();return n?await $.create(n):null}catch(n){return O("engine init",n),null}})()),W}async function M1(){return await D()!==null}function H(n){return M(n)}async function N1(n,r){let e=await D();if(!e)return null;let i=await H(r);if(!i)return null;try{let t=await e.run(n,{paramsBits:i,wantLog16:!1,wantRgba:!0});return t?.rgba?{width:t.width,height:t.height,rgba:t.rgba}:null}catch(t){return O("rgba decode",t)}}async function W1(n,r){let e=await D();if(!e)return null;let i=await H(r);if(!i)return null;try{let t=await e.run(n,{paramsBits:i,wantLog16:!0,wantRgba:!1});return t?.log16?{width:t.width,height:t.height,rgb16:t.log16.subarray(0,t.width*t.height*3)}:null}catch(t){return O("rgb16 decode",t)}}async function $1(n,r){let e=await D();if(!e)return null;let i=r?.captureOnly,t=r?.paramsBits??await H(void 0);if(!t)return null;try{return await e.run(n,{paramsBits:t,wantLog16:!i||i.includes("out_log16"),wantRgba:!i||i.includes("out_rgba"),captureStages:r?.captureStages??!1,captureOnly:i})}catch(a){return O("debug decode",a)}}export{$1 as decodeR3dGpuDebug,W1 as decodeR3dRgb16Gpu,N1 as decodeR3dRgbaGpu,M as developParamsBits,q as parseR3dFrame,M1 as r3dGpuAvailable};
|