wgblas 2.2.0 → 2.2.1

package/dist/wgblas.browser.js

@@ -654,7 +654,7 @@ fn ddAbs(a: DD) -> DD {
   }
   return a;
 }
-`});var zr,At=O(()=>{zr=`// Requires f64/dekker.wgsl concatenated first for the DD struct.
+`});var zr,St=O(()=>{zr=`// Requires f64/dekker.wgsl concatenated first for the DD struct.
 
 // \u2500\u2500 A real compiler bug \u2014 read before touching anything below \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
 //
@@ -737,7 +737,7 @@ fn ddAddProtected(a: DD, b: DD, threadSlot: u32) -> DD {
 fn ddSubProtected(a: DD, b: DD, threadSlot: u32) -> DD {
   return ddAddProtected(a, DD(negf(b.hi), negf(b.lo)), threadSlot);
 }
-`});var Gt,St=O(()=>{Gt=`// dasum: sum(|x[i]|), double-double (Dekker). Same ILP=4 shape as sasum.wgsl;
+`});var Gt,At=O(()=>{Gt=`// dasum: sum(|x[i]|), double-double (Dekker). Same ILP=4 shape as sasum.wgsl;
 // see f64/utils/add.wgsl for ddAddProtected and why plain ddAdd isn't safe.
 // GpuVector input isn't pre-abs'd, so ddAbs() (f64/utils/abs.wgsl) applies
 // unconditionally below.
@@ -948,9 +948,18 @@ fn ddMulRaw(a: DD, b: DD, threadSlot: u32) -> DD {
   return DD(p.hi, crossAndLo);
 }
 
+// A third compiler bug (Intel Mesa ANV, via NIR dump): raw.hi never gets
+// materialized as one rounded value \u2014 the driver re-fuses a.hi*b.hi with
+// ffma at every use site instead, breaking the a+b == s+e identity
+// TwoSum-style algorithms depend on. Same fix as ddMulRaw's p.lo: force it
+// through workgroup memory + a barrier. (Verified: max forward-error factor
+// over 20000 trials dropped from >1e5 to ~3-4, Intel Mesa Iris Xe + NVIDIA GTX 1650.)
 fn ddMulProtected(a: DD, b: DD, threadSlot: u32) -> DD {
   let raw = ddMulRaw(a, b, threadSlot);
-  return fastTwoSumProtected(raw.hi, raw.lo, threadSlot);
+  dekkerScratch[threadSlot] = raw.hi;
+  workgroupBarrier();
+  let rawHi = dekkerScratch[threadSlot];
+  return fastTwoSumProtected(rawHi, raw.lo, threadSlot);
 }
 `});var Nt,Dt=O(()=>{Nt=`// ddot: sum(x[i] * y[i]), double-double (Dekker). Same ILP=4 shape as
 // dasum.wgsl, which this mirrors closely \u2014 the only structural difference is
@@ -2678,7 +2687,7 @@ fn main(
     }
   }
 }
-`});var ue,Ao=O(()=>{ue=`// sgemm_small: C = alpha * op(A) * op(B) + beta * C \u2014 small-tile half of
+`});var ue,So=O(()=>{ue=`// sgemm_small: C = alpha * op(A) * op(B) + beta * C \u2014 small-tile half of
 // the two-tier autotuned dispatch (see sgemm.mjs and sgemm_large.wgsl).
 // BM=BN=32, BK=8, TM=TN=2 \u2014 wins over the large tile below a 6x6=36
 // workgroup grid of 64-tiles, where the large tile doesn't have enough
@@ -2890,7 +2899,7 @@ fn main(
     }
   }
 }
-`});var fe,So=O(()=>{fe=`// sgemm_large: C = alpha * op(A) * op(B) + beta * C \u2014 large-tile half of
+`});var fe,Ao=O(()=>{fe=`// sgemm_large: C = alpha * op(A) * op(B) + beta * C \u2014 large-tile half of
 // the two-tier autotuned dispatch (see sgemm.mjs and sgemm_small.wgsl).
 // BM=BN=64, BK=8, TM=8, TN=4 (128 threads/workgroup) \u2014 the kernel 9
 // autotuning winner (temp/autotune_sweep.mjs, temp/gen_sweep_kernel.mjs,
@@ -3446,8 +3455,8 @@ fn main(@builtin(global_invocation_id) gid: vec3u) {
     strided[stridedIdx] = block[blockIdx];
   }
 }
-`});var Lo={};qe(Lo,{routineShaders:()=>or,shaderSources:()=>Ii});var or,Ii,Ro=O(()=>{$e();Ze();Je();et();ot();it();nt();ut();mt();dt();pt();wt();bt();xt();_t();Bt();At();St();Et();kt();Dt();Pt();It();Rt();qt();Ct();jt();Ht();Vt();zt();Yt();$t();Qt();ro();to();ao();so();lo();fo();co();po();wo();bo();xo();_o();Ao();So();Go();Eo();ko();No();Mo();or={};or.sscal={sscal:ke};or.cscal={cscal:Qe};or.sswap={sswap:rt};or.dswap={dswap:tt};or.saxpy={saxpy:at};or.scopy={scopy:st};or.dcopy={dcopy:lt};or.sdot={sdot:ft,"reduction/sum":De};or.sasum={sasum:ct,"reduction/sum":De};or.snrm2={snrm2:gt,"reduction/scaledSum":ht};or.isamax={isamax:yt,"reduction/argmax":vt};or.dasum={"f64/dekker":Kr,"f64/utils/abs":ye,"f64/utils/add":zr,dasum:Gt,"reduction/sumF64":Ne};or.ddot={"f64/dekker":Kr,"f64/utils/add":zr,"f64/utils/multiply":$r,ddot:Nt,"reduction/sumF64":Ne};or.dscal={"f64/dekker":Kr,"f64/utils/add":zr,"f64/utils/multiply":$r,dscal:Mt};or.daxpy={"f64/dekker":Kr,"f64/utils/add":zr,"f64/utils/multiply":$r,daxpy:Lt};or.idamax={"f64/dekker":Kr,"f64/utils/abs":ye,"f64/utils/greater":Pe,"f64/utils/equal":Tt,idamax:Ft,"reduction/argmaxF64":Wt};or.srot={srot:Ot};or.drot={"f64/dekker":Kr,"f64/utils/add":zr,"f64/utils/multiply":$r,drot:Kt};or.srotm={srotm:Ut};or.drotm={"f64/dekker":Kr,"f64/utils/add":zr,"f64/utils/multiply":$r,drotm:Xt};or.dnrm2={"f64/dekker":Kr,"f64/utils/abs":ye,"f64/utils/greater":Pe,"f64/utils/add":zr,"f64/utils/multiply":$r,"f64/utils/divide":Zt,"f64/utils/sqrt":Jt,dnrm2:eo,"reduction/scaledSumF64":oo};or.sgemv={sgemv_n:io,sgemv_t:no};or.ssymv={ssymv:uo};or.strmv={strmv:mo};or.strsv={strsv_invert_block:Me,strsv_apply_inverse:go,strsv_update:ho};or.sger={sger:yo};or.ssyr={ssyr:vo};or.ssyr2={ssyr2:Bo};or.sgemm={sgemm_small:ue,sgemm_large:fe};or.sgemmtr={sgemmtr_small:xe,sgemmtr_large:ve};or.ssyrk={sgemmtr_small:xe,sgemmtr_large:ve};or.ssyr2k={sgemmtr_small:xe,sgemmtr_large:ve};or.ssymm={sgemm_small:ue,sgemm_large:fe,symmetrize:Do};or.strmm={sgemm_small:ue,sgemm_large:fe,triangularize:Po};or.strsm={strsv_invert_block:Me,block_transfer:Io,sscal:ke,sgemm_small:ue,sgemm_large:fe};Ii=Object.assign({},...Object.values(or))});var Ti={};qe(Ti,{Complex32:()=>Wr,Complex32Array:()=>_r,Complex64:()=>jr,Complex64Array:()=>Gr,GpuMatrix:()=>X,GpuVector:()=>N,cleanup:()=>Oe,cscal:()=>To,dasum:()=>Uo,daxpy:()=>Ho,dcopy:()=>Vo,ddot:()=>Yo,dnrm2:()=>$o,drot:()=>ra,drotm:()=>ta,dscal:()=>Co,dswap:()=>jo,gpuName:()=>Ve,idamax:()=>Qo,init:()=>He,isamax:()=>Zo,randomFloat32Array:()=>Ue,randomFloat64Array:()=>Ye,randomTriangularFloat32Array:()=>Xe,sasum:()=>zo,saxpy:()=>Wo,scopy:()=>Oo,sdot:()=>Ko,sgemm:()=>da,sgemmtr:()=>ca,sgemv:()=>oa,sger:()=>ua,snrm2:()=>Xo,srot:()=>Jo,srotm:()=>ea,sscal:()=>qo,sswap:()=>Fo,ssymm:()=>wa,ssymv:()=>aa,ssyr:()=>fa,ssyr2:()=>ma,ssyr2k:()=>ga,ssyrk:()=>pa,strmm:()=>ha,strmv:()=>ia,strsm:()=>ba,strsv:()=>la});function Ce(r,t){return t?r.features.has("timestamp-query")?{requiredFeatures:["timestamp-query"]}:(console.warn("timestamp-query not supported on this device \u2014 benchmark mode disabled."),{}):{}}function Fe(r){if(!je(r))return{querySet:null,passDescriptor:void 0};let t=r.createQuerySet({type:"timestamp",count:2});return{querySet:t,passDescriptor:{timestampWrites:{querySet:t,beginningOfPassWriteIndex:0,endOfPassWriteIndex:1}}}}function Lr(r,t,e){if(!e)return null;let o=r.createBuffer({label:"timestamp-resolve",size:16,usage:GPUBufferUsage.QUERY_RESOLVE|GPUBufferUsage.COPY_SRC});t.resolveQuerySet(e,0,2,o,0);let a=r.createBuffer({label:"timestamp-readback",size:16,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});return t.copyBufferToBuffer(o,0,a,0,16),{tsReadBuffer:a,resolveBuffer:o,querySet:e}}async function P(r){if(!r)return;let{tsReadBuffer:t,resolveBuffer:e,querySet:o}=r;await t.mapAsync(GPUMapMode.READ);let a=new BigInt64Array(t.getMappedRange().slice());return t.unmap(),t.destroy(),e.destroy(),o.destroy(),Math.max(0,Number(a[1]-a[0]))/1e6}var Qr=null,Se=!1,Jr=new Map,le=new WeakMap,Vr=null,We=({powerPreference:r,benchmark:t})=>`${r}::${t}`;async function He({powerPreference:r="high-performance",benchmark:t=!1,dumpShaders:e=!1}={}){let o={powerPreference:r,benchmark:t,dumpShaders:e},a=We(o),i=Jr.get(a);if(i)return i;if(Qr)e!==Se&&typeof window>"u"&&console.warn(`dumpShaders: ${e} was requested, but the WebGPU instance was already created with dumpShaders: ${Se}. The first init() call fixes this for the process.`);else if(typeof window>"u"){let{create:m,globals:p}=await import("webgpu");Object.assign(globalThis,p),Qr=m(e?["enable-dawn-features=dump_shaders,disable_symbol_renaming"]:[]),Se=e}else e&&console.warn("dumpShaders has no effect in the browser \u2014 see init()'s docs."),Qr=navigator.gpu;if(!Qr)throw new Error("WebGPU not supported in this environment.");let s=await Qr.requestAdapter({powerPreference:r})??await Qr.requestAdapter();if(!s)throw new Error("No WebGPU adapter found.");let n=[...Ce(s,t).requiredFeatures??[]],f=await s.requestDevice({requiredFeatures:n});f.addEventListener("uncapturederror",m=>{console.error("Uncaptured GPU error:",m.error.message)});let l=n.includes("timestamp-query");return le.set(f,{adapter:s,benchmark:l,options:o}),Jr.set(a,f),Vr||(Vr=f),f}function Oe(r){if(r===void 0){for(let e of Jr.values())e.destroy();Jr.clear(),Vr=null;return}let t=le.get(r);t&&(Jr.delete(We(t.options)),le.delete(r),r.destroy(),Vr===r&&(Vr=Jr.values().next().value??null))}function Ve(r=Vr){let t=r&&le.get(r);if(!t)throw new Error("WebGPU adapter not initialized \u2014 call init() first.");let{device:e,description:o}=t.adapter.info;return{description:o||"unknown",device:e||"unknown"}}function je(r=Vr){return le.get(r)?.benchmark??!1}function re(){if(!Vr)throw new Error("WebGPU device not initialized \u2014 call init() first.");return Vr}function d(...r){r.flat().forEach(t=>t.destroy())}function Ge(r,t,e){let o=r.limits.maxStorageBufferBindingSize;if(t>o)throw new Error(`Buffer "${e}" needs ${t} bytes, exceeding this device's maxStorageBufferBindingSize (${o} bytes). The operands are too large for this device.`)}function x(r,t,e="blas-input",o=!1){let a=t.byteLength;Ge(r,a,e);let i=o?GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC:GPUBufferUsage.STORAGE,s=r.createBuffer({label:e,size:a,usage:i,mappedAtCreation:!0}),u=t.constructor;return new u(s.getMappedRange()).set(t),s.unmap(),s}function tr(r,t,e="blas-storage",o=0){return Ge(r,t,e),r.createBuffer({label:e,size:t,usage:GPUBufferUsage.STORAGE|o})}function Br(r,t,e="blas-result"){return Ge(r,t,e),r.createBuffer({label:e,size:t,usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC})}function G(r,t,e){let o=r.createBuffer({label:"blas-readback",size:e.size,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});return t.copyBufferToBuffer(e,0,o,0,e.size),o}var ee=16,Ke=new WeakMap;function La(r){let t=Ke.get(r);return t||(t=r.createBuffer({label:"blas-vec4-fallback",size:ee,usage:GPUBufferUsage.STORAGE}),Ke.set(r,t)),t}function Sr(r,t){let e=t instanceof GPUBuffer?t:t.buffer,o=t instanceof GPUBuffer?0:t.offset??0,a=t instanceof GPUBuffer?t.size:t.size??e.size-o,i=Math.floor(a/ee)*ee;return i<ee?{buffer:La(r),offset:0,size:ee}:{buffer:e,offset:o,size:i}}function Ee(r,t,e,o){if(t%4!==0)return!1;let a=r instanceof GPUBuffer?r:r.buffer,i=r instanceof GPUBuffer?0:r.offset??0,s=r instanceof GPUBuffer?a.size:r.size??a.size-i,u=Math.floor(s/ee)*4;if(u<=0)return!1;let n=(Math.max(e,1)-1)*t+(Math.max(o,1)-1);return Math.floor(n/4)*4+4<=u}function I(r,t,e="blas-params"){let o=t.length*4,a=Math.ceil(o/16)*16,i=new ArrayBuffer(a),s=new DataView(i);t.forEach(({value:n,type:f},l)=>{let m=l*4;if(f==="u32")s.setUint32(m,n,!0);else if(f==="i32")s.setInt32(m,n,!0);else if(f==="f32")s.setFloat32(m,n,!0);else throw new Error(`Unknown param type "${f}". Use "f32", "u32", or "i32".`)});let u=r.createBuffer({label:e,size:a,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});return r.queue.writeBuffer(u,0,i),u}async function S(r,t=Float32Array){try{await r.mapAsync(GPUMapMode.READ);let e=new t(r.getMappedRange().slice());return r.unmap(),e}finally{r.destroy()}}function rr(r){let t=r.length,e=new Float32Array(t),o=new Float32Array(t);for(let a=0;a<t;a++){let i=Math.fround(r[a]);e[a]=i,o[a]=Math.fround(r[a]-i)}return{hi:e,lo:o}}function dr(r,t){let e=r.length,o=new Float64Array(e);for(let a=0;a<e;a++)o[a]=r[a]+t[a];return o}var jr=class{constructor(t,e){this.re=t,this.im=e}},Gr=class extends Array{constructor(t){if(t===void 0){super();return}if(typeof t=="number"){super(t);for(let o=0;o<t;o++)this[o]=new jr(0,0);return}let e=Array.from(t);if(super(),e.length!==0){if(e[0]instanceof jr){for(let o of e){if(!(o instanceof jr))throw new Error("Complex64Array expects every element to be a Complex64.");this.push(o)}return}if(e.length%2!==0)throw new Error("Complex64Array expects an even number of interleaved [re, im, ...] values.");for(let o=0;o<e.length;o+=2){if(typeof e[o]!="number"||typeof e[o+1]!="number")throw new Error("Complex64Array expects interleaved [re, im, ...] values to be numbers.");this.push(new jr(e[o],e[o+1]))}}}};function te(r,t=r.length){let e=new Float32Array(t*2);for(let o=0;o<t;o++)e[o*2]=r[o].re,e[o*2+1]=r[o].im;return e}function he(r,t=r.length){let e=new Float64Array(t),o=new Float64Array(t);for(let l=0;l<t;l++)e[l]=r[l].re,o[l]=r[l].im;let{hi:a,lo:i}=rr(e),{hi:s,lo:u}=rr(o),n=new Float32Array(t*2),f=new Float32Array(t*2);for(let l=0;l<t;l++)n[l*2]=a[l],n[l*2+1]=s[l],f[l*2]=i[l],f[l*2+1]=u[l];return{hi:n,lo:f}}function be(r,t){let e=r.length/2,o=new Float32Array(e),a=new Float32Array(e),i=new Float32Array(e),s=new Float32Array(e);for(let l=0;l<e;l++)o[l]=r[l*2],i[l]=r[l*2+1],a[l]=t[l*2],s[l]=t[l*2+1];let u=dr(o,a),n=dr(i,s),f=new Gr(e);for(let l=0;l<e;l++)f[l]=new jr(u[l],n[l]);return f}var Wr=class{constructor(t,e){this.re=Math.fround(t),this.im=Math.fround(e)}},_r=class extends Array{constructor(t){if(t===void 0){super();return}if(typeof t=="number"){super(t);for(let o=0;o<t;o++)this[o]=new Wr(0,0);return}let e=Array.from(t);if(super(),e.length!==0){if(e[0]instanceof Wr){for(let o of e){if(!(o instanceof Wr))throw new Error("Complex32Array expects every element to be a Complex32.");this.push(o)}return}if(e.length%2!==0)throw new Error("Complex32Array expects an even number of interleaved [re, im, ...] values.");for(let o=0;o<e.length;o+=2){if(typeof e[o]!="number"||typeof e[o+1]!="number")throw new Error("Complex32Array expects interleaved [re, im, ...] values to be numbers.");this.push(new Wr(e[o],e[o+1]))}}}};var N=class r{constructor(t,e,o=Float32Array,a=null,i=null){this._buf=t,this._loBuf=a,this.length=e,this.dtype=o,this.device=i??re()}static from(t,e){let o=t instanceof GPUDevice,a=o?t:re(),i=o?e:t;if(i instanceof Float64Array){let{hi:u,lo:n}=rr(i),f=x(a,u,"gpu-vector-f64-hi",!0),l=x(a,n,"gpu-vector-f64-lo",!0);return new r(f,i.length,Float64Array,l,a)}if(i instanceof _r){let u=x(a,te(i),"gpu-vector-complex32",!0);return new r(u,i.length,_r,null,a)}if(i instanceof Gr){let{hi:u,lo:n}=he(i),f=x(a,u,"gpu-vector-complex64-hi",!0),l=x(a,n,"gpu-vector-complex64-lo",!0);return new r(f,i.length,Gr,l,a)}if(!(i instanceof Float32Array))throw new Error("GpuVector.from expects a Float32Array, Float64Array, Complex32Array, or Complex64Array.");let s=x(a,i,"gpu-vector",!0);return new r(s,i.length,i.constructor,null,a)}async read(){let t=this.device,e=t.createCommandEncoder(),o=G(t,e,this._buf);if(t.queue.submit([e.finish()]),this.dtype===_r)return new _r(await S(o,Float32Array));if(!this._loBuf)return S(o,this.dtype);let a=t.createCommandEncoder(),i=G(t,a,this._loBuf);t.queue.submit([a.finish()]);let[s,u]=await Promise.all([S(o,Float32Array),S(i,Float32Array)]);return this.dtype===Gr?be(s,u):dr(s,u)}destroy(){this._buf.destroy(),this._loBuf&&this._loBuf.destroy()}};var X=class r{constructor(t,e,o,a,i=null,s="row-major",u=null,n=Float32Array){this._buf=t,this._loBuf=i,this.rows=e,this.cols=o,this.lda=a,this.layout=s,this.dtype=n,this.device=u??re()}static from(t,...e){let o=t instanceof GPUDevice,a=o?t:re(),i=o?e.shift():t,[s,u,n,f="row-major"]=e;if(f!=="row-major"&&f!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");let l=f==="row-major";if(n===void 0&&(n=l?u:s),!(i instanceof Float32Array)&&!(i instanceof Float64Array)&&!(i instanceof _r)&&!(i instanceof Gr))throw new Error("GpuMatrix.from expects a Float32Array, Float64Array, Complex32Array, or Complex64Array.");if(!Number.isInteger(s)||s<=0)throw new Error("rows must be a positive integer.");if(!Number.isInteger(u)||u<=0)throw new Error("cols must be a positive integer.");let m=l?u:s;if(!Number.isInteger(n)||n<m)throw new Error(`lda must be an integer >= ${l?"cols":"rows"}.`);let p=l?s:u;if(i.length<p*n)throw new Error("data does not have enough elements for the given rows, cols, and lda.");if(i instanceof Float64Array){let g=p*n,{hi:w,lo:h}=rr(i.subarray(0,g)),b=x(a,w,"gpu-matrix-f64-hi",!0),y=x(a,h,"gpu-matrix-f64-lo",!0);return new r(b,s,u,n,y,f,a,Float64Array)}if(i instanceof _r){let g=x(a,te(i,p*n),"gpu-matrix-complex32",!0);return new r(g,s,u,n,null,f,a,_r)}if(i instanceof Gr){let{hi:g,lo:w}=he(i,p*n),h=x(a,g,"gpu-matrix-complex64-hi",!0),b=x(a,w,"gpu-matrix-complex64-lo",!0);return new r(h,s,u,n,b,f,a,Gr)}let c=x(a,i.subarray(0,p*n),"gpu-matrix",!0);return new r(c,s,u,n,null,f,a)}async read(){let t=this.device,e=t.createCommandEncoder(),o=G(t,e,this._buf);t.queue.submit([e.finish()]);let a=this.layout!=="column-major",i=a?this.rows:this.cols,s=a?this.cols:this.rows;if(this.dtype===_r){let f=new _r(await S(o,Float32Array));if(this.lda===s)return f;let l=new _r(i*s);for(let m=0;m<i;m++)for(let p=0;p<s;p++)l[m*s+p]=f[m*this.lda+p];return l}if(this._loBuf){let f=t.createCommandEncoder(),l=G(t,f,this._loBuf);t.queue.submit([f.finish()]);let[m,p]=await Promise.all([S(o,Float32Array),S(l,Float32Array)]);if(this.dtype===Gr){let w=be(m,p);if(this.lda===s)return w;let h=new Gr(i*s);for(let b=0;b<i;b++)for(let y=0;y<s;y++)h[b*s+y]=w[b*this.lda+y];return h}let c=dr(m,p);if(this.lda===s)return c;let g=new Float64Array(i*s);for(let w=0;w<i;w++)g.set(c.subarray(w*this.lda,w*this.lda+s),w*s);return g}let u=await S(o,Float32Array);if(this.lda===s)return u;let n=new Float32Array(i*s);for(let f=0;f<i;f++)n.set(u.subarray(f*this.lda,f*this.lda+s),f*s);return n}destroy(){this._buf.destroy(),this._loBuf&&this._loBuf.destroy()}};function ze(r){let t=r>>>0;return function(){t=t+1831565813|0;let e=Math.imul(t^t>>>15,1|t);return e=e+Math.imul(e^e>>>7,61|e)^e,((e^e>>>14)>>>0)/4294967296}}function Ue(r,t=-1,e=1,o){let a=new Float32Array(r),i=o===void 0?Math.random:ze(o);for(let s=0;s<r;s++)a[s]=t+i()*(e-t);return a}function Ye(r,t=-1,e=1,o){let a=new Float64Array(r),i=o===void 0?Math.random:ze(o);for(let s=0;s<r;s++)a[s]=t+i()*(e-t);return a}function Xe(r,t,e="lower",o=-1,a=1,i=5,s=15,u="row-major"){if(e!=="lower"&&e!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(u!=="row-major"&&u!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(t<r)throw new Error("lda must be >= n.");let n=u==="column-major",f=(m,p)=>n?p*t+m:m*t+p,l=new Float32Array(r*t);for(let m=0;m<r;m++){for(let p=0;p<r;p++){if(m===p)continue;(e==="lower"?p<m:p>m)&&(l[f(m,p)]=o+Math.random()*(a-o))}l[f(m,m)]=i+Math.random()*(s-i)}return l}function E(r,t,e,o=0){let a=e.map((i,s)=>({binding:o+s,resource:i instanceof GPUBuffer?{buffer:i}:i}));return r.createBindGroup({layout:t,entries:a})}function M(r,t){r.queue.submit([t.finish()])}function qr(r){let{querySet:t,passDescriptor:e}=Fe(r);return{commandEncoder:r.createCommandEncoder(),querySet:t,passDescriptor:e}}function gr(r,t,e,o,a){let i=r.beginComputePass(a);i.setPipeline(t),i.setBindGroup(0,e),typeof o=="number"?i.dispatchWorkgroups(o):i.dispatchWorkgroups(o.x,o.y,o.z??1),i.end()}function j(r,t,e,o){let{commandEncoder:a,querySet:i,passDescriptor:s}=qr(r);gr(a,t,e,o,s);let u=Lr(r,a,i);return{commandEncoder:a,ts:u}}var qi={},Ie=new WeakMap;async function D(r,t,e="main"){Ie.has(r)||Ie.set(r,new Map);let o=Ie.get(r),a=Array.isArray(t)?t:[t],i=`${a.join("+")}::${e}`;if(!o.has(i)){let s=Ri(r,a,e).catch(u=>{throw o.delete(i),u});o.set(i,s)}return o.get(i)}async function Li(r){if(typeof process>"u"||!process.versions?.node){let{shaderSources:t}=await Promise.resolve().then(()=>(Ro(),Lo)),e=t[r];if(!e)throw new Error(`Shader "${r}" not found in browser bundle.`);return e}else{let{readFileSync:t}=await import("fs"),{fileURLToPath:e}=await import("url"),{dirname:o,join:a}=await import("path"),i=o(e(qi.url));return t(a(i,`../shaders/${r}.wgsl`),"utf8")}}async function Ri(r,t,e="main"){let o=t.join("+"),a=await Promise.all(t.map(Li)),i=0,s=a.map((g,w)=>{let h=g.split(`
+`});var Lo={};qe(Lo,{routineShaders:()=>or,shaderSources:()=>Ii});var or,Ii,Ro=O(()=>{$e();Ze();Je();et();ot();it();nt();ut();mt();dt();pt();wt();bt();xt();_t();Bt();St();At();Et();kt();Dt();Pt();It();Rt();qt();Ct();jt();Ht();Vt();zt();Yt();$t();Qt();ro();to();ao();so();lo();fo();co();po();wo();bo();xo();_o();So();Ao();Go();Eo();ko();No();Mo();or={};or.sscal={sscal:ke};or.cscal={cscal:Qe};or.sswap={sswap:rt};or.dswap={dswap:tt};or.saxpy={saxpy:at};or.scopy={scopy:st};or.dcopy={dcopy:lt};or.sdot={sdot:ft,"reduction/sum":De};or.sasum={sasum:ct,"reduction/sum":De};or.snrm2={snrm2:gt,"reduction/scaledSum":ht};or.isamax={isamax:yt,"reduction/argmax":vt};or.dasum={"f64/dekker":Kr,"f64/utils/abs":ye,"f64/utils/add":zr,dasum:Gt,"reduction/sumF64":Ne};or.ddot={"f64/dekker":Kr,"f64/utils/add":zr,"f64/utils/multiply":$r,ddot:Nt,"reduction/sumF64":Ne};or.dscal={"f64/dekker":Kr,"f64/utils/add":zr,"f64/utils/multiply":$r,dscal:Mt};or.daxpy={"f64/dekker":Kr,"f64/utils/add":zr,"f64/utils/multiply":$r,daxpy:Lt};or.idamax={"f64/dekker":Kr,"f64/utils/abs":ye,"f64/utils/greater":Pe,"f64/utils/equal":Tt,idamax:Ft,"reduction/argmaxF64":Wt};or.srot={srot:Ot};or.drot={"f64/dekker":Kr,"f64/utils/add":zr,"f64/utils/multiply":$r,drot:Kt};or.srotm={srotm:Ut};or.drotm={"f64/dekker":Kr,"f64/utils/add":zr,"f64/utils/multiply":$r,drotm:Xt};or.dnrm2={"f64/dekker":Kr,"f64/utils/abs":ye,"f64/utils/greater":Pe,"f64/utils/add":zr,"f64/utils/multiply":$r,"f64/utils/divide":Zt,"f64/utils/sqrt":Jt,dnrm2:eo,"reduction/scaledSumF64":oo};or.sgemv={sgemv_n:io,sgemv_t:no};or.ssymv={ssymv:uo};or.strmv={strmv:mo};or.strsv={strsv_invert_block:Me,strsv_apply_inverse:go,strsv_update:ho};or.sger={sger:yo};or.ssyr={ssyr:vo};or.ssyr2={ssyr2:Bo};or.sgemm={sgemm_small:ue,sgemm_large:fe};or.sgemmtr={sgemmtr_small:xe,sgemmtr_large:ve};or.ssyrk={sgemmtr_small:xe,sgemmtr_large:ve};or.ssyr2k={sgemmtr_small:xe,sgemmtr_large:ve};or.ssymm={sgemm_small:ue,sgemm_large:fe,symmetrize:Do};or.strmm={sgemm_small:ue,sgemm_large:fe,triangularize:Po};or.strsm={strsv_invert_block:Me,block_transfer:Io,sscal:ke,sgemm_small:ue,sgemm_large:fe};Ii=Object.assign({},...Object.values(or))});var Ti={};qe(Ti,{Complex32:()=>Wr,Complex32Array:()=>_r,Complex64:()=>jr,Complex64Array:()=>Gr,GpuMatrix:()=>X,GpuVector:()=>N,cleanup:()=>Oe,cscal:()=>To,dasum:()=>Uo,daxpy:()=>Ho,dcopy:()=>Vo,ddot:()=>Yo,dnrm2:()=>$o,drot:()=>ra,drotm:()=>ta,dscal:()=>Co,dswap:()=>jo,gpuName:()=>Ve,idamax:()=>Qo,init:()=>He,isamax:()=>Zo,randomFloat32Array:()=>Ue,randomFloat64Array:()=>Ye,randomTriangularFloat32Array:()=>Xe,sasum:()=>zo,saxpy:()=>Wo,scopy:()=>Oo,sdot:()=>Ko,sgemm:()=>da,sgemmtr:()=>ca,sgemv:()=>oa,sger:()=>ua,snrm2:()=>Xo,srot:()=>Jo,srotm:()=>ea,sscal:()=>qo,sswap:()=>Fo,ssymm:()=>wa,ssymv:()=>aa,ssyr:()=>fa,ssyr2:()=>ma,ssyr2k:()=>ga,ssyrk:()=>pa,strmm:()=>ha,strmv:()=>ia,strsm:()=>ba,strsv:()=>la});function Ce(r,t){return t?r.features.has("timestamp-query")?{requiredFeatures:["timestamp-query"]}:(console.warn("timestamp-query not supported on this device \u2014 benchmark mode disabled."),{}):{}}function Fe(r){if(!je(r))return{querySet:null,passDescriptor:void 0};let t=r.createQuerySet({type:"timestamp",count:2});return{querySet:t,passDescriptor:{timestampWrites:{querySet:t,beginningOfPassWriteIndex:0,endOfPassWriteIndex:1}}}}function Lr(r,t,e){if(!e)return null;let o=r.createBuffer({label:"timestamp-resolve",size:16,usage:GPUBufferUsage.QUERY_RESOLVE|GPUBufferUsage.COPY_SRC});t.resolveQuerySet(e,0,2,o,0);let a=r.createBuffer({label:"timestamp-readback",size:16,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});return t.copyBufferToBuffer(o,0,a,0,16),{tsReadBuffer:a,resolveBuffer:o,querySet:e}}async function P(r){if(!r)return;let{tsReadBuffer:t,resolveBuffer:e,querySet:o}=r;await t.mapAsync(GPUMapMode.READ);let a=new BigInt64Array(t.getMappedRange().slice());return t.unmap(),t.destroy(),e.destroy(),o.destroy(),Math.max(0,Number(a[1]-a[0]))/1e6}var Qr=null,Ae=!1,Jr=new Map,le=new WeakMap,Vr=null,We=({powerPreference:r,benchmark:t})=>`${r}::${t}`;async function He({powerPreference:r="high-performance",benchmark:t=!1,dumpShaders:e=!1}={}){let o={powerPreference:r,benchmark:t,dumpShaders:e},a=We(o),i=Jr.get(a);if(i)return i;if(Qr)e!==Ae&&typeof window>"u"&&console.warn(`dumpShaders: ${e} was requested, but the WebGPU instance was already created with dumpShaders: ${Ae}. The first init() call fixes this for the process.`);else if(typeof window>"u"){let{create:m,globals:p}=await import("webgpu");Object.assign(globalThis,p),Qr=m(e?["enable-dawn-features=dump_shaders,disable_symbol_renaming"]:[]),Ae=e}else e&&console.warn("dumpShaders has no effect in the browser \u2014 see init()'s docs."),Qr=navigator.gpu;if(!Qr)throw new Error("WebGPU not supported in this environment.");let s=await Qr.requestAdapter({powerPreference:r})??await Qr.requestAdapter();if(!s)throw new Error("No WebGPU adapter found.");let n=[...Ce(s,t).requiredFeatures??[]],f=await s.requestDevice({requiredFeatures:n});f.addEventListener("uncapturederror",m=>{console.error("Uncaptured GPU error:",m.error.message)});let l=n.includes("timestamp-query");return le.set(f,{adapter:s,benchmark:l,options:o}),Jr.set(a,f),Vr||(Vr=f),f}function Oe(r){if(r===void 0){for(let e of Jr.values())e.destroy();Jr.clear(),Vr=null;return}let t=le.get(r);t&&(Jr.delete(We(t.options)),le.delete(r),r.destroy(),Vr===r&&(Vr=Jr.values().next().value??null))}function Ve(r=Vr){let t=r&&le.get(r);if(!t)throw new Error("WebGPU adapter not initialized \u2014 call init() first.");let{device:e,description:o}=t.adapter.info;return{description:o||"unknown",device:e||"unknown"}}function je(r=Vr){return le.get(r)?.benchmark??!1}function re(){if(!Vr)throw new Error("WebGPU device not initialized \u2014 call init() first.");return Vr}function d(...r){r.flat().forEach(t=>t.destroy())}function Ge(r,t,e){let o=r.limits.maxStorageBufferBindingSize;if(t>o)throw new Error(`Buffer "${e}" needs ${t} bytes, exceeding this device's maxStorageBufferBindingSize (${o} bytes). The operands are too large for this device.`)}function x(r,t,e="blas-input",o=!1){let a=t.byteLength;Ge(r,a,e);let i=o?GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC:GPUBufferUsage.STORAGE,s=r.createBuffer({label:e,size:a,usage:i,mappedAtCreation:!0}),u=t.constructor;return new u(s.getMappedRange()).set(t),s.unmap(),s}function tr(r,t,e="blas-storage",o=0){return Ge(r,t,e),r.createBuffer({label:e,size:t,usage:GPUBufferUsage.STORAGE|o})}function Br(r,t,e="blas-result"){return Ge(r,t,e),r.createBuffer({label:e,size:t,usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC})}function G(r,t,e){let o=r.createBuffer({label:"blas-readback",size:e.size,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});return t.copyBufferToBuffer(e,0,o,0,e.size),o}var ee=16,Ke=new WeakMap;function La(r){let t=Ke.get(r);return t||(t=r.createBuffer({label:"blas-vec4-fallback",size:ee,usage:GPUBufferUsage.STORAGE}),Ke.set(r,t)),t}function Ar(r,t){let e=t instanceof GPUBuffer?t:t.buffer,o=t instanceof GPUBuffer?0:t.offset??0,a=t instanceof GPUBuffer?t.size:t.size??e.size-o,i=Math.floor(a/ee)*ee;return i<ee?{buffer:La(r),offset:0,size:ee}:{buffer:e,offset:o,size:i}}function Ee(r,t,e,o){if(t%4!==0)return!1;let a=r instanceof GPUBuffer?r:r.buffer,i=r instanceof GPUBuffer?0:r.offset??0,s=r instanceof GPUBuffer?a.size:r.size??a.size-i,u=Math.floor(s/ee)*4;if(u<=0)return!1;let n=(Math.max(e,1)-1)*t+(Math.max(o,1)-1);return Math.floor(n/4)*4+4<=u}function I(r,t,e="blas-params"){let o=t.length*4,a=Math.ceil(o/16)*16,i=new ArrayBuffer(a),s=new DataView(i);t.forEach(({value:n,type:f},l)=>{let m=l*4;if(f==="u32")s.setUint32(m,n,!0);else if(f==="i32")s.setInt32(m,n,!0);else if(f==="f32")s.setFloat32(m,n,!0);else throw new Error(`Unknown param type "${f}". Use "f32", "u32", or "i32".`)});let u=r.createBuffer({label:e,size:a,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});return r.queue.writeBuffer(u,0,i),u}async function A(r,t=Float32Array){try{await r.mapAsync(GPUMapMode.READ);let e=new t(r.getMappedRange().slice());return r.unmap(),e}finally{r.destroy()}}function rr(r){let t=r.length,e=new Float32Array(t),o=new Float32Array(t);for(let a=0;a<t;a++){let i=Math.fround(r[a]);e[a]=i,o[a]=Math.fround(r[a]-i)}return{hi:e,lo:o}}function dr(r,t){let e=r.length,o=new Float64Array(e);for(let a=0;a<e;a++)o[a]=r[a]+t[a];return o}var jr=class{constructor(t,e){this.re=t,this.im=e}},Gr=class extends Array{constructor(t){if(t===void 0){super();return}if(typeof t=="number"){super(t);for(let o=0;o<t;o++)this[o]=new jr(0,0);return}let e=Array.from(t);if(super(),e.length!==0){if(e[0]instanceof jr){for(let o of e){if(!(o instanceof jr))throw new Error("Complex64Array expects every element to be a Complex64.");this.push(o)}return}if(e.length%2!==0)throw new Error("Complex64Array expects an even number of interleaved [re, im, ...] values.");for(let o=0;o<e.length;o+=2){if(typeof e[o]!="number"||typeof e[o+1]!="number")throw new Error("Complex64Array expects interleaved [re, im, ...] values to be numbers.");this.push(new jr(e[o],e[o+1]))}}}};function te(r,t=r.length){let e=new Float32Array(t*2);for(let o=0;o<t;o++)e[o*2]=r[o].re,e[o*2+1]=r[o].im;return e}function he(r,t=r.length){let e=new Float64Array(t),o=new Float64Array(t);for(let l=0;l<t;l++)e[l]=r[l].re,o[l]=r[l].im;let{hi:a,lo:i}=rr(e),{hi:s,lo:u}=rr(o),n=new Float32Array(t*2),f=new Float32Array(t*2);for(let l=0;l<t;l++)n[l*2]=a[l],n[l*2+1]=s[l],f[l*2]=i[l],f[l*2+1]=u[l];return{hi:n,lo:f}}function be(r,t){let e=r.length/2,o=new Float32Array(e),a=new Float32Array(e),i=new Float32Array(e),s=new Float32Array(e);for(let l=0;l<e;l++)o[l]=r[l*2],i[l]=r[l*2+1],a[l]=t[l*2],s[l]=t[l*2+1];let u=dr(o,a),n=dr(i,s),f=new Gr(e);for(let l=0;l<e;l++)f[l]=new jr(u[l],n[l]);return f}var Wr=class{constructor(t,e){this.re=Math.fround(t),this.im=Math.fround(e)}},_r=class extends Array{constructor(t){if(t===void 0){super();return}if(typeof t=="number"){super(t);for(let o=0;o<t;o++)this[o]=new Wr(0,0);return}let e=Array.from(t);if(super(),e.length!==0){if(e[0]instanceof Wr){for(let o of e){if(!(o instanceof Wr))throw new Error("Complex32Array expects every element to be a Complex32.");this.push(o)}return}if(e.length%2!==0)throw new Error("Complex32Array expects an even number of interleaved [re, im, ...] values.");for(let o=0;o<e.length;o+=2){if(typeof e[o]!="number"||typeof e[o+1]!="number")throw new Error("Complex32Array expects interleaved [re, im, ...] values to be numbers.");this.push(new Wr(e[o],e[o+1]))}}}};var N=class r{constructor(t,e,o=Float32Array,a=null,i=null){this._buf=t,this._loBuf=a,this.length=e,this.dtype=o,this.device=i??re()}static from(t,e){let o=t instanceof GPUDevice,a=o?t:re(),i=o?e:t;if(i instanceof Float64Array){let{hi:u,lo:n}=rr(i),f=x(a,u,"gpu-vector-f64-hi",!0),l=x(a,n,"gpu-vector-f64-lo",!0);return new r(f,i.length,Float64Array,l,a)}if(i instanceof _r){let u=x(a,te(i),"gpu-vector-complex32",!0);return new r(u,i.length,_r,null,a)}if(i instanceof Gr){let{hi:u,lo:n}=he(i),f=x(a,u,"gpu-vector-complex64-hi",!0),l=x(a,n,"gpu-vector-complex64-lo",!0);return new r(f,i.length,Gr,l,a)}if(!(i instanceof Float32Array))throw new Error("GpuVector.from expects a Float32Array, Float64Array, Complex32Array, or Complex64Array.");let s=x(a,i,"gpu-vector",!0);return new r(s,i.length,i.constructor,null,a)}async read(){let t=this.device,e=t.createCommandEncoder(),o=G(t,e,this._buf);if(t.queue.submit([e.finish()]),this.dtype===_r)return new _r(await A(o,Float32Array));if(!this._loBuf)return A(o,this.dtype);let a=t.createCommandEncoder(),i=G(t,a,this._loBuf);t.queue.submit([a.finish()]);let[s,u]=await Promise.all([A(o,Float32Array),A(i,Float32Array)]);return this.dtype===Gr?be(s,u):dr(s,u)}destroy(){this._buf.destroy(),this._loBuf&&this._loBuf.destroy()}};var X=class r{constructor(t,e,o,a,i=null,s="row-major",u=null,n=Float32Array){this._buf=t,this._loBuf=i,this.rows=e,this.cols=o,this.lda=a,this.layout=s,this.dtype=n,this.device=u??re()}static from(t,...e){let o=t instanceof GPUDevice,a=o?t:re(),i=o?e.shift():t,[s,u,n,f="row-major"]=e;if(f!=="row-major"&&f!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");let l=f==="row-major";if(n===void 0&&(n=l?u:s),!(i instanceof Float32Array)&&!(i instanceof Float64Array)&&!(i instanceof _r)&&!(i instanceof Gr))throw new Error("GpuMatrix.from expects a Float32Array, Float64Array, Complex32Array, or Complex64Array.");if(!Number.isInteger(s)||s<=0)throw new Error("rows must be a positive integer.");if(!Number.isInteger(u)||u<=0)throw new Error("cols must be a positive integer.");let m=l?u:s;if(!Number.isInteger(n)||n<m)throw new Error(`lda must be an integer >= ${l?"cols":"rows"}.`);let p=l?s:u;if(i.length<p*n)throw new Error("data does not have enough elements for the given rows, cols, and lda.");if(i instanceof Float64Array){let g=p*n,{hi:w,lo:h}=rr(i.subarray(0,g)),b=x(a,w,"gpu-matrix-f64-hi",!0),y=x(a,h,"gpu-matrix-f64-lo",!0);return new r(b,s,u,n,y,f,a,Float64Array)}if(i instanceof _r){let g=x(a,te(i,p*n),"gpu-matrix-complex32",!0);return new r(g,s,u,n,null,f,a,_r)}if(i instanceof Gr){let{hi:g,lo:w}=he(i,p*n),h=x(a,g,"gpu-matrix-complex64-hi",!0),b=x(a,w,"gpu-matrix-complex64-lo",!0);return new r(h,s,u,n,b,f,a,Gr)}let c=x(a,i.subarray(0,p*n),"gpu-matrix",!0);return new r(c,s,u,n,null,f,a)}async read(){let t=this.device,e=t.createCommandEncoder(),o=G(t,e,this._buf);t.queue.submit([e.finish()]);let a=this.layout!=="column-major",i=a?this.rows:this.cols,s=a?this.cols:this.rows;if(this.dtype===_r){let f=new _r(await A(o,Float32Array));if(this.lda===s)return f;let l=new _r(i*s);for(let m=0;m<i;m++)for(let p=0;p<s;p++)l[m*s+p]=f[m*this.lda+p];return l}if(this._loBuf){let f=t.createCommandEncoder(),l=G(t,f,this._loBuf);t.queue.submit([f.finish()]);let[m,p]=await Promise.all([A(o,Float32Array),A(l,Float32Array)]);if(this.dtype===Gr){let w=be(m,p);if(this.lda===s)return w;let h=new Gr(i*s);for(let b=0;b<i;b++)for(let y=0;y<s;y++)h[b*s+y]=w[b*this.lda+y];return h}let c=dr(m,p);if(this.lda===s)return c;let g=new Float64Array(i*s);for(let w=0;w<i;w++)g.set(c.subarray(w*this.lda,w*this.lda+s),w*s);return g}let u=await A(o,Float32Array);if(this.lda===s)return u;let n=new Float32Array(i*s);for(let f=0;f<i;f++)n.set(u.subarray(f*this.lda,f*this.lda+s),f*s);return n}destroy(){this._buf.destroy(),this._loBuf&&this._loBuf.destroy()}};function ze(r){let t=r>>>0;return function(){t=t+1831565813|0;let e=Math.imul(t^t>>>15,1|t);return e=e+Math.imul(e^e>>>7,61|e)^e,((e^e>>>14)>>>0)/4294967296}}function Ue(r,t=-1,e=1,o){let a=new Float32Array(r),i=o===void 0?Math.random:ze(o);for(let s=0;s<r;s++)a[s]=t+i()*(e-t);return a}function Ye(r,t=-1,e=1,o){let a=new Float64Array(r),i=o===void 0?Math.random:ze(o);for(let s=0;s<r;s++)a[s]=t+i()*(e-t);return a}function Xe(r,t,e="lower",o=-1,a=1,i=5,s=15,u="row-major"){if(e!=="lower"&&e!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(u!=="row-major"&&u!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(t<r)throw new Error("lda must be >= n.");let n=u==="column-major",f=(m,p)=>n?p*t+m:m*t+p,l=new Float32Array(r*t);for(let m=0;m<r;m++){for(let p=0;p<r;p++){if(m===p)continue;(e==="lower"?p<m:p>m)&&(l[f(m,p)]=o+Math.random()*(a-o))}l[f(m,m)]=i+Math.random()*(s-i)}return l}function E(r,t,e,o=0){let a=e.map((i,s)=>({binding:o+s,resource:i instanceof GPUBuffer?{buffer:i}:i}));return r.createBindGroup({layout:t,entries:a})}function M(r,t){r.queue.submit([t.finish()])}function qr(r){let{querySet:t,passDescriptor:e}=Fe(r);return{commandEncoder:r.createCommandEncoder(),querySet:t,passDescriptor:e}}function gr(r,t,e,o,a){let i=r.beginComputePass(a);i.setPipeline(t),i.setBindGroup(0,e),typeof o=="number"?i.dispatchWorkgroups(o):i.dispatchWorkgroups(o.x,o.y,o.z??1),i.end()}function j(r,t,e,o){let{commandEncoder:a,querySet:i,passDescriptor:s}=qr(r);gr(a,t,e,o,s);let u=Lr(r,a,i);return{commandEncoder:a,ts:u}}var qi={},Ie=new WeakMap;async function D(r,t,e="main"){Ie.has(r)||Ie.set(r,new Map);let o=Ie.get(r),a=Array.isArray(t)?t:[t],i=`${a.join("+")}::${e}`;if(!o.has(i)){let s=Ri(r,a,e).catch(u=>{throw o.delete(i),u});o.set(i,s)}return o.get(i)}async function Li(r){if(typeof process>"u"||!process.versions?.node){let{shaderSources:t}=await Promise.resolve().then(()=>(Ro(),Lo)),e=t[r];if(!e)throw new Error(`Shader "${r}" not found in browser bundle.`);return e}else{let{readFileSync:t}=await import("fs"),{fileURLToPath:e}=await import("url"),{dirname:o,join:a}=await import("path"),i=o(e(qi.url));return t(a(i,`../shaders/${r}.wgsl`),"utf8")}}async function Ri(r,t,e="main"){let o=t.join("+"),a=await Promise.all(t.map(Li)),i=0,s=a.map((g,w)=>{let h=g.split(`
 `).length,b={name:t[w],startLine:i+1,endLine:i+h};return i+=h,b}),u=g=>{let w=g&&s.find(h=>g>=h.startLine&&g<=h.endLine);return w?`${w.name}.wgsl:${g-w.startLine+1}`:`line ${g}`},n=a.join(`
 `),f=r.createShaderModule({label:o,code:n}),m=(await f.getCompilationInfo()).messages.filter(g=>g.type==="error");if(m.length>0)throw new Error(`Shader "${o}" compilation failed:
 ${m.map(g=>`  ${u(g.lineNum)}: ${g.message}`).join(`
-`)}`);let p=e==="main"?{module:f}:{module:f,entryPoint:e},c=r.createComputePipeline({label:o,layout:"auto",compute:p});return c._shaderModule=f,c}function cr(r,t,e){let o=r.limits.maxComputeWorkgroupsPerDimension;return e===void 0?Math.min(Math.ceil(t/64),o):{x:Math.min(Math.ceil(e/8),o),y:Math.min(Math.ceil(t/8),o)}}function U(r,t,e,o="x"){let a=r.limits.maxComputeWorkgroupsPerDimension;if(t>a)throw new Error(`${e}: this problem needs ${t} workgroups in ${o}, but the device allows ${a} (maxComputeWorkgroupsPerDimension). The operands are too large for this device \u2014 split the operation into smaller blocks.`);return t}function Zr(r,t,e,o){return o===void 0?U(r,Math.ceil(e/64),t):{x:U(r,Math.ceil(o/8),t,"x"),y:U(r,Math.ceil(e/8),t,"y")}}function q(r){if(!(r instanceof GPUDevice))throw new Error("device must be a GPUDevice.")}function T(r,t,e){for(let[o,a]of Object.entries(e))if(!(!(a instanceof N)&&!(a instanceof X))&&a.device!==r)throw new Error(`${t}: ${o} belongs to a different GPUDevice than the one passed in. GPU buffers cannot be shared across devices \u2014 recreate the operand on this device, or call the routine with the device that owns it.`)}async function qo(r,t,e,o,a){let i=o instanceof N;if(q(r),T(r,"sscal",{x:o}),!Number.isInteger(t)||!Number.isInteger(a))throw new Error("n and incx must be integers.");if(typeof e!="number")throw new Error("alpha must be a number.");if(Number.isNaN(e))throw new Error("alpha must not be NaN.");if(!Number.isFinite(e))throw new Error("alpha must be finite.");if(a<=0)throw new Error("incx must be positive.");if(!(o instanceof Float32Array)&&!(o instanceof N))throw new Error("x must be a Float32Array or GpuVector.");if(t<=0)return i?{}:{x:o};if(o.length<(t-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");let s=await D(r,"sscal"),u=null,n=null,f=null;try{u=i?o._buf:x(r,o,"sscal-x",!0),n=I(r,[{value:t,type:"u32"},{value:e,type:"f32"},{value:a,type:"u32"}],"sscal-params");let l=E(r,s.getBindGroupLayout(0),[u,n]),{commandEncoder:m,ts:p}=j(r,s,l,cr(r,t));f=i?null:G(r,m,u),M(r,m);let c=await P(p);if(i)return c!==void 0?{gpuTimeMs:c}:{};let g=await S(f,Float32Array);return f=null,c!==void 0?{x:g,gpuTimeMs:c}:{x:g}}finally{!i&&u&&d(u),n&&d(n),f&&d(f)}}async function To(r,t,e,o,a){let i=o instanceof N;if(q(r),T(r,"cscal",{x:o}),!Number.isInteger(t)||!Number.isInteger(a))throw new Error("n and incx must be integers.");if(!(e instanceof Wr))throw new Error("alpha must be a Complex32.");if(Number.isNaN(e.re)||Number.isNaN(e.im))throw new Error("alpha must not be NaN.");if(!Number.isFinite(e.re)||!Number.isFinite(e.im))throw new Error("alpha must be finite.");if(a<=0)throw new Error("incx must be positive.");if(!(o instanceof _r)&&!i)throw new Error("x must be a Complex32Array or GpuVector.");if(i&&o.dtype!==_r)throw new Error("x must be a Complex32Array-backed GpuVector.");if(t<=0)return i?{}:{x:o};if(o.length<(t-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");let s=await D(r,"cscal"),u=null,n=null,f=null;try{u=i?o._buf:x(r,te(o),"cscal-x",!0),n=I(r,[{value:t,type:"u32"},{value:e.re,type:"f32"},{value:e.im,type:"f32"},{value:a,type:"u32"}],"cscal-params");let l=E(r,s.getBindGroupLayout(0),[u,n]),{commandEncoder:m,ts:p}=j(r,s,l,cr(r,t));f=i?null:G(r,m,u),M(r,m);let c=await P(p);if(i)return c!==void 0?{gpuTimeMs:c}:{};let g=await S(f,Float32Array);f=null;let w=new _r(g);return c!==void 0?{x:w,gpuTimeMs:c}:{x:w}}finally{!i&&u&&d(u),n&&d(n),f&&d(f)}}async function Co(r,t,e,o,a){let i=o instanceof N;if(q(r),!Number.isInteger(t)||!Number.isInteger(a))throw new Error("n and incx must be integers.");if(typeof e!="number")throw new Error("alpha must be a number.");if(Number.isNaN(e))throw new Error("alpha must not be NaN.");if(!Number.isFinite(e))throw new Error("alpha must be finite.");if(!(o instanceof Float64Array)&&!i)throw new Error("x must be a Float64Array or GpuVector.");if(i&&o.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(a<=0)throw new Error("incx must be positive.");if(T(r,"dscal",{x:o}),t<=0)return i?{}:{x:o};if(o.length<(t-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");let u=await D(r,[...["f64/dekker","f64/utils/add","f64/utils/multiply"],"dscal"]),{hi:n,lo:f}=rr(new Float64Array([e])),l=null,m=null,p=null,c=null,g=null;try{if(i)l=o._buf,m=o._loBuf;else{let{hi:k,lo:B}=rr(o);l=x(r,k,"dscal-xHi",!0),m=x(r,B,"dscal-xLo",!0)}p=I(r,[{value:t,type:"u32"},{value:n[0],type:"f32"},{value:f[0],type:"f32"},{value:a,type:"u32"}],"dscal-params");let w=E(r,u.getBindGroupLayout(0),[l,m,p]),{commandEncoder:h,ts:b}=j(r,u,w,cr(r,t));c=i?null:G(r,h,l),g=i?null:G(r,h,m),M(r,h);let y=await P(b);if(i)return y!==void 0?{gpuTimeMs:y}:{};let v=await S(c,Float32Array);c=null;let _=await S(g,Float32Array);g=null;let A=dr(v,_);return y!==void 0?{x:A,gpuTimeMs:y}:{x:A}}finally{!i&&l&&d(l),!i&&m&&d(m),p&&d(p),c&&d(c),g&&d(g)}}async function Fo(r,t,e,o,a,i){let s=e instanceof N,u=a instanceof N;if(q(r),T(r,"sswap",{x:e,y:a}),!Number.isInteger(t)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!(e instanceof Float32Array)&&!(e instanceof N))throw new Error("x must be a Float32Array or GpuVector.");if(!(a instanceof Float32Array)&&!(a instanceof N))throw new Error("y must be a Float32Array or GpuVector.");if(e.constructor!==a.constructor)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(t<=0)return s?{}:{x:e,y:a};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(a.length<(t-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=await D(r,"sswap"),f=null,l=null,m=null,p=null,c=null;try{f=s?e._buf:x(r,e,"sswap-x",!0),l=u?a._buf:x(r,a,"sswap-y",!0),m=I(r,[{value:t,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"sswap-params");let g=E(r,n.getBindGroupLayout(0),[f,l,m]),{commandEncoder:w,ts:h}=j(r,n,g,cr(r,t));p=s?null:G(r,w,f),c=u?null:G(r,w,l),M(r,w);let b=await P(h);if(s)return b!==void 0?{gpuTimeMs:b}:{};let y=await S(p,Float32Array);p=null;let v=await S(c,Float32Array);return c=null,b!==void 0?{x:y,y:v,gpuTimeMs:b}:{x:y,y:v}}finally{!s&&f&&d(f),!u&&l&&d(l),m&&d(m),p&&d(p),c&&d(c)}}async function jo(r,t,e,o,a,i){let s=e instanceof N,u=a instanceof N;if(q(r),T(r,"dswap",{x:e,y:a}),!Number.isInteger(t)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!(e instanceof Float64Array)&&!s)throw new Error("x must be a Float64Array or GpuVector.");if(!(a instanceof Float64Array)&&!u)throw new Error("y must be a Float64Array or GpuVector.");if(s&&e.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(u&&a.dtype!==Float64Array)throw new Error("y must be a Float64Array-backed GpuVector.");if(s!==u)throw new Error("x and y must be the same type (both Float64Array or both GpuVector).");if(t<=0)return s?{}:{x:e,y:a};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(a.length<(t-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=await D(r,"dswap"),f=null,l=null,m=null,p=null,c=null,g=null,w=null,h=null,b=null;try{if(s)f=e._buf,l=e._loBuf,m=a._buf,p=a._loBuf;else{let W=rr(e),V=rr(a);f=x(r,W.hi,"dswap-xHi",!0),l=x(r,W.lo,"dswap-xLo",!0),m=x(r,V.hi,"dswap-yHi",!0),p=x(r,V.lo,"dswap-yLo",!0)}c=I(r,[{value:t,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"dswap-params");let y=E(r,n.getBindGroupLayout(0),[f,l,m,p,c]),{commandEncoder:v,ts:_}=j(r,n,y,cr(r,t));g=s?null:G(r,v,f),w=s?null:G(r,v,l),h=u?null:G(r,v,m),b=u?null:G(r,v,p),M(r,v);let A=await P(_);if(s)return A!==void 0?{gpuTimeMs:A}:{};let k=await S(g,Float32Array);g=null;let B=await S(w,Float32Array);w=null;let L=await S(h,Float32Array);h=null;let C=await S(b,Float32Array);b=null;let R=dr(k,B),F=dr(L,C);return A!==void 0?{x:R,y:F,gpuTimeMs:A}:{x:R,y:F}}finally{!s&&f&&d(f),!s&&l&&d(l),!u&&m&&d(m),!u&&p&&d(p),c&&d(c),g&&d(g),w&&d(w),h&&d(h),b&&d(b)}}async function Wo(r,t,e,o,a,i,s){let u=o instanceof N,n=i instanceof N;if(q(r),T(r,"saxpy",{x:o,y:i}),!Number.isInteger(t)||!Number.isInteger(a)||!Number.isInteger(s))throw new Error("n, incx, and incy must be integers.");if(typeof e!="number")throw new Error("alpha must be a number.");if(Number.isNaN(e))throw new Error("alpha must not be NaN.");if(!Number.isFinite(e))throw new Error("alpha must be finite.");if(a<=0||s<=0)throw new Error("incx and incy must be positive.");if(!u&&!(o instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!n&&!(i instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(u!==n)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(t<=0)return n?{}:{y:i};if(o.length<(t-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");if(i.length<(t-1)*s+1)throw new Error("y does not have enough elements for the given n and incy.");let f=await D(r,"saxpy"),l=null,m=null,p=null,c=null;try{l=u?o._buf:x(r,o,"saxpy-x",!1),m=n?i._buf:x(r,i,"saxpy-y",!0),p=I(r,[{value:t,type:"u32"},{value:e,type:"f32"},{value:a,type:"u32"},{value:s,type:"u32"}],"saxpy-params");let g=E(r,f.getBindGroupLayout(0),[l,m,p]),{commandEncoder:w,ts:h}=j(r,f,g,cr(r,t));c=n?null:G(r,w,m),M(r,w);let b=await P(h);if(n)return b!==void 0?{gpuTimeMs:b}:{};let y=await S(c,Float32Array);return c=null,b!==void 0?{y,gpuTimeMs:b}:{y}}finally{!u&&l&&d(l),!n&&m&&d(m),p&&d(p),c&&d(c)}}async function Ho(r,t,e,o,a,i,s){let u=o instanceof N,n=i instanceof N;if(q(r),!Number.isInteger(t)||!Number.isInteger(a)||!Number.isInteger(s))throw new Error("n, incx, and incy must be integers.");if(typeof e!="number")throw new Error("alpha must be a number.");if(Number.isNaN(e))throw new Error("alpha must not be NaN.");if(!Number.isFinite(e))throw new Error("alpha must be finite.");if(!(o instanceof Float64Array)&&!u)throw new Error("x must be a Float64Array or GpuVector.");if(!(i instanceof Float64Array)&&!n)throw new Error("y must be a Float64Array or GpuVector.");if(u&&o.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(n&&i.dtype!==Float64Array)throw new Error("y must be a Float64Array-backed GpuVector.");if(u!==n)throw new Error("x and y must be the same type (both Float64Array or both GpuVector).");if(a<=0||s<=0)throw new Error("incx and incy must be positive.");if(T(r,"daxpy",{x:o,y:i}),t<=0)return n?{}:{y:i};if(o.length<(t-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");if(i.length<(t-1)*s+1)throw new Error("y does not have enough elements for the given n and incy.");let l=await D(r,[...["f64/dekker","f64/utils/add","f64/utils/multiply"],"daxpy"]),{hi:m,lo:p}=rr(new Float64Array([e])),c=null,g=null,w=null,h=null,b=null,y=null,v=null;try{if(u)c=o._buf,g=o._loBuf,w=i._buf,h=i._loBuf;else{let F=rr(o),W=rr(i);c=x(r,F.hi,"daxpy-xHi",!1),g=x(r,F.lo,"daxpy-xLo",!1),w=x(r,W.hi,"daxpy-yHi",!0),h=x(r,W.lo,"daxpy-yLo",!0)}b=I(r,[{value:t,type:"u32"},{value:m[0],type:"f32"},{value:p[0],type:"f32"},{value:a,type:"u32"},{value:s,type:"u32"}],"daxpy-params");let _=E(r,l.getBindGroupLayout(0),[c,g,w,h,b]),{commandEncoder:A,ts:k}=j(r,l,_,cr(r,t));y=n?null:G(r,A,w),v=n?null:G(r,A,h),M(r,A);let B=await P(k);if(n)return B!==void 0?{gpuTimeMs:B}:{};let L=await S(y,Float32Array);y=null;let C=await S(v,Float32Array);v=null;let R=dr(L,C);return B!==void 0?{y:R,gpuTimeMs:B}:{y:R}}finally{!u&&c&&d(c),!u&&g&&d(g),!n&&w&&d(w),!n&&h&&d(h),b&&d(b),y&&d(y),v&&d(v)}}async function Oo(r,t,e,o,a,i){let s=e instanceof N,u=a instanceof N;if(q(r),T(r,"scopy",{x:e,y:a}),!Number.isInteger(t)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!s&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!u&&!(a instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(s!==u)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(t<=0)return u?{}:{y:a};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(a.length<(t-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=await D(r,"scopy"),f=null,l=null,m=null,p=null;try{f=s?e._buf:x(r,e,"scopy-x",!1),l=u?a._buf:x(r,a,"scopy-y",!0),m=I(r,[{value:t,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"scopy-params");let c=E(r,n.getBindGroupLayout(0),[f,l,m]),{commandEncoder:g,ts:w}=j(r,n,c,cr(r,t));p=u?null:G(r,g,l),M(r,g);let h=await P(w);if(u)return h!==void 0?{gpuTimeMs:h}:{};let b=await S(p,Float32Array);return p=null,h!==void 0?{y:b,gpuTimeMs:h}:{y:b}}finally{!s&&f&&d(f),!u&&l&&d(l),m&&d(m),p&&d(p)}}async function Vo(r,t,e,o,a,i){let s=e instanceof N,u=a instanceof N;if(q(r),T(r,"dcopy",{x:e,y:a}),!Number.isInteger(t)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!(e instanceof Float64Array)&&!s)throw new Error("x must be a Float64Array or GpuVector.");if(!(a instanceof Float64Array)&&!u)throw new Error("y must be a Float64Array or GpuVector.");if(s&&e.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(u&&a.dtype!==Float64Array)throw new Error("y must be a Float64Array-backed GpuVector.");if(s!==u)throw new Error("x and y must be the same type (both Float64Array or both GpuVector).");if(t<=0)return u?{}:{y:a};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(a.length<(t-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=await D(r,"dcopy"),f=null,l=null,m=null,p=null,c=null,g=null,w=null;try{if(s)f=e._buf,l=e._loBuf,m=a._buf,p=a._loBuf;else{let B=rr(e),L=rr(a);f=x(r,B.hi,"dcopy-xHi",!1),l=x(r,B.lo,"dcopy-xLo",!1),m=x(r,L.hi,"dcopy-yHi",!0),p=x(r,L.lo,"dcopy-yLo",!0)}c=I(r,[{value:t,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"dcopy-params");let h=E(r,n.getBindGroupLayout(0),[f,l,m,p,c]),{commandEncoder:b,ts:y}=j(r,n,h,cr(r,t));g=u?null:G(r,b,m),w=u?null:G(r,b,p),M(r,b);let v=await P(y);if(u)return v!==void 0?{gpuTimeMs:v}:{};let _=await S(g,Float32Array);g=null;let A=await S(w,Float32Array);w=null;let k=dr(_,A);return v!==void 0?{y:k,gpuTimeMs:v}:{y:k}}finally{!s&&f&&d(f),!s&&l&&d(l),!u&&m&&d(m),!u&&p&&d(p),c&&d(c),g&&d(g),w&&d(w)}}async function Ko(r,t,e,o,a,i){let s=e instanceof N,u=a instanceof N;if(q(r),T(r,"sdot",{x:e,y:a}),!Number.isInteger(t)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!s&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!u&&!(a instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(s!==u)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(t<=0)return{dot:0};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(a.length<(t-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=await D(r,"sdot"),f=await D(r,"reduction/sum"),l=null,m=null,p=null,c=null,g=null,w=null;try{l=s?e._buf:x(r,e,"sdot-x",!1),m=u?a._buf:x(r,a,"sdot-y",!1),p=tr(r,512,"sdot-partials"),c=Br(r,4,"sdot-result"),g=I(r,[{value:t,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"sdot-params");let h=E(r,n.getBindGroupLayout(0),[l,m,p,g]),{commandEncoder:b,ts:y}=j(r,n,h,128);M(r,b);let v=E(r,f.getBindGroupLayout(0),[p,c]),{commandEncoder:_,ts:A}=j(r,f,v,1);w=G(r,_,c),M(r,_);let k=S(w,Float32Array);w=null;let[B,L,C]=await Promise.all([P(y),P(A),k]);return B!==void 0&&L!==void 0?{dot:C[0],gpuTimeMs:B+L}:{dot:C[0]}}finally{!s&&l&&d(l),!u&&m&&d(m),p&&d(p),c&&d(c),g&&d(g),w&&d(w)}}async function zo(r,t,e,o){let a=e instanceof N;if(q(r),T(r,"sasum",{x:e}),!Number.isInteger(t)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!a&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(t<=0)return{asum:0};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=await D(r,"sasum"),s=await D(r,"reduction/sum"),u=null,n=null,f=null,l=null,m=null;try{u=a?e._buf:x(r,e,"sasum-x",!1),n=tr(r,512,"sasum-partials"),f=Br(r,4,"sasum-result"),l=I(r,[{value:t,type:"u32"},{value:o,type:"u32"}],"sasum-params");let p=E(r,i.getBindGroupLayout(0),[u,n,l]),{commandEncoder:c,ts:g}=j(r,i,p,128);M(r,c);let w=E(r,s.getBindGroupLayout(0),[n,f]),{commandEncoder:h,ts:b}=j(r,s,w,1);m=G(r,h,f),M(r,h);let y=S(m,Float32Array);m=null;let[v,_,A]=await Promise.all([P(g),P(b),y]);return v!==void 0&&_!==void 0?{asum:A[0],gpuTimeMs:v+_}:{asum:A[0]}}finally{!a&&u&&d(u),n&&d(n),f&&d(f),l&&d(l),m&&d(m)}}async function Uo(r,t,e,o){let a=e instanceof N;if(q(r),T(r,"dasum",{x:e}),!Number.isInteger(t)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!a&&!(e instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(a&&e.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(t<=0)return{asum:0};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=["f64/dekker","f64/utils/abs","f64/utils/add"],s=await D(r,[...i,"dasum"]),u=await D(r,[...i,"reduction/sumF64"]),n=null,f=null,l=null,m=null,p=null,c=null,g=null,w=null,h=null;try{if(a)n=e._buf,f=e._loBuf;else{let{hi:z,lo:H}=rr(e.map(Math.abs));n=x(r,z,"dasum-xHi",!1),f=x(r,H,"dasum-xLo",!1)}l=tr(r,512,"dasum-partialsHi"),m=tr(r,512,"dasum-partialsLo"),p=Br(r,4,"dasum-result-hi"),c=Br(r,4,"dasum-result-lo"),g=I(r,[{value:t,type:"u32"},{value:o,type:"u32"}],"dasum-params");let b=E(r,s.getBindGroupLayout(0),[n,f,l,m,g]),{commandEncoder:y,ts:v}=j(r,s,b,128);M(r,y);let _=E(r,u.getBindGroupLayout(0),[l,m,p,c]),{commandEncoder:A,ts:k}=j(r,u,_,1);w=G(r,A,p),h=G(r,A,c),M(r,A);let B=S(w,Float32Array),L=S(h,Float32Array);w=null,h=null;let[C,R,F,W]=await Promise.all([P(v),P(k),B,L]),V=dr(F,W)[0];return C!==void 0&&R!==void 0?{asum:V,gpuTimeMs:C+R}:{asum:V}}finally{!a&&n&&d(n),!a&&f&&d(f),l&&d(l),m&&d(m),p&&d(p),c&&d(c),g&&d(g),w&&d(w),h&&d(h)}}async function Yo(r,t,e,o,a,i){let s=e instanceof N,u=a instanceof N;if(q(r),T(r,"ddot",{x:e,y:a}),!Number.isInteger(t)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!s&&!(e instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(!u&&!(a instanceof Float64Array))throw new Error("y must be a Float64Array or GpuVector.");if(s&&e.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(u&&a.dtype!==Float64Array)throw new Error("y must be a Float64Array-backed GpuVector.");if(s!==u)throw new Error("x and y must be the same type (both Float64Array or both GpuVector).");if(t<=0)return{dot:0};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(a.length<(t-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=["f64/dekker","f64/utils/add"],f=await D(r,[...n,"f64/utils/multiply","ddot"]),l=await D(r,[...n,"reduction/sumF64"]),m=null,p=null,c=null,g=null,w=null,h=null,b=null,y=null,v=null,_=null,A=null;try{if(s)m=e._buf,p=e._loBuf,c=a._buf,g=a._loBuf;else{let ir=rr(e),lr=rr(a);m=x(r,ir.hi,"ddot-xHi",!1),p=x(r,ir.lo,"ddot-xLo",!1),c=x(r,lr.hi,"ddot-yHi",!1),g=x(r,lr.lo,"ddot-yLo",!1)}w=tr(r,512,"ddot-partialsHi"),h=tr(r,512,"ddot-partialsLo"),b=Br(r,4,"ddot-result-hi"),y=Br(r,4,"ddot-result-lo"),v=I(r,[{value:t,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"ddot-params");let k=E(r,f.getBindGroupLayout(0),[m,p,c,g,w,h,v]),{commandEncoder:B,ts:L}=j(r,f,k,128);M(r,B);let C=E(r,l.getBindGroupLayout(0),[w,h,b,y]),{commandEncoder:R,ts:F}=j(r,l,C,1);_=G(r,R,b),A=G(r,R,y),M(r,R);let W=S(_,Float32Array),V=S(A,Float32Array);_=null,A=null;let[z,H,$,K]=await Promise.all([P(L),P(F),W,V]),Y=dr($,K)[0];return z!==void 0&&H!==void 0?{dot:Y,gpuTimeMs:z+H}:{dot:Y}}finally{!s&&m&&d(m),!s&&p&&d(p),!u&&c&&d(c),!u&&g&&d(g),w&&d(w),h&&d(h),b&&d(b),y&&d(y),v&&d(v),_&&d(_),A&&d(A)}}async function Xo(r,t,e,o){let a=e instanceof N;if(q(r),T(r,"snrm2",{x:e}),!Number.isInteger(t)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!a&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(t<=0)return{nrm2:0};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=await D(r,"snrm2"),s=await D(r,"reduction/scaledSum"),u=null,n=null,f=null,l=null,m=null,p=null;try{u=a?e._buf:x(r,e,"snrm2-x",!1),n=tr(r,512,"snrm2-partials-scale"),f=tr(r,512,"snrm2-partials-ssq"),l=Br(r,4,"snrm2-result"),m=I(r,[{value:t,type:"u32"},{value:o,type:"u32"}],"snrm2-params");let c=E(r,i.getBindGroupLayout(0),[u,n,f,m]),{commandEncoder:g,ts:w}=j(r,i,c,128);M(r,g);let h=E(r,s.getBindGroupLayout(0),[n,f,l]),{commandEncoder:b,ts:y}=j(r,s,h,1);p=G(r,b,l),M(r,b);let v=S(p,Float32Array);p=null;let[_,A,k]=await Promise.all([P(w),P(y),v]),B=k[0];return _!==void 0&&A!==void 0?{nrm2:B,gpuTimeMs:_+A}:{nrm2:B}}finally{!a&&u&&d(u),n&&d(n),f&&d(f),l&&d(l),m&&d(m),p&&d(p)}}async function $o(r,t,e,o){let a=e instanceof N;if(q(r),T(r,"dnrm2",{x:e}),!Number.isInteger(t)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!a&&!(e instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(a&&e.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(t<=0)return{nrm2:0};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=["f64/dekker","f64/utils/abs","f64/utils/greater","f64/utils/add","f64/utils/multiply","f64/utils/divide","f64/utils/sqrt"],s=await D(r,[...i,"dnrm2"]),u=await D(r,[...i,"reduction/scaledSumF64"]),n=null,f=null,l=null,m=null,p=null,c=null,g=null,w=null,h=null,b=null,y=null;try{if(a)n=e._buf,f=e._loBuf;else{let{hi:$,lo:K}=rr(e);n=x(r,$,"dnrm2-xHi",!1),f=x(r,K,"dnrm2-xLo",!1)}l=tr(r,512,"dnrm2-partials-scaleHi"),m=tr(r,512,"dnrm2-partials-scaleLo"),p=tr(r,512,"dnrm2-partials-ssqHi"),c=tr(r,512,"dnrm2-partials-ssqLo"),g=Br(r,4,"dnrm2-result-hi"),w=Br(r,4,"dnrm2-result-lo"),h=I(r,[{value:t,type:"u32"},{value:o,type:"u32"}],"dnrm2-params");let v=E(r,s.getBindGroupLayout(0),[n,f,l,m,p,c,h]),{commandEncoder:_,ts:A}=j(r,s,v,128);M(r,_);let k=E(r,u.getBindGroupLayout(0),[l,m,p,c,g,w]),{commandEncoder:B,ts:L}=j(r,u,k,1);b=G(r,B,g),y=G(r,B,w),M(r,B);let C=S(b,Float32Array),R=S(y,Float32Array);b=null,y=null;let[F,W,V,z]=await Promise.all([P(A),P(L),C,R]),H=dr(V,z)[0];return F!==void 0&&W!==void 0?{nrm2:H,gpuTimeMs:F+W}:{nrm2:H}}finally{!a&&n&&d(n),!a&&f&&d(f),l&&d(l),m&&d(m),p&&d(p),c&&d(c),g&&d(g),w&&d(w),h&&d(h),b&&d(b),y&&d(y)}}async function Zo(r,t,e,o){let a=e instanceof N;if(q(r),T(r,"isamax",{x:e}),!Number.isInteger(t)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!a&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(t<=0)return{index:0};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=await D(r,"isamax"),s=await D(r,"reduction/argmax"),u=null,n=null,f=null,l=null,m=null,p=null;try{u=a?e._buf:x(r,e,"isamax-x",!1),n=tr(r,512,"isamax-partials-val"),f=tr(r,512,"isamax-partials-idx"),l=Br(r,4,"isamax-result"),m=I(r,[{value:t,type:"u32"},{value:o,type:"u32"}],"isamax-params");let c=E(r,i.getBindGroupLayout(0),[u,n,f,m]),{commandEncoder:g,ts:w}=j(r,i,c,128);M(r,g);let h=E(r,s.getBindGroupLayout(0),[n,f,l]),{commandEncoder:b,ts:y}=j(r,s,h,1);p=G(r,b,l),M(r,b);let v=S(p,Uint32Array);p=null;let[_,A,k]=await Promise.all([P(w),P(y),v]),B=k[0];return _!==void 0&&A!==void 0?{index:B,gpuTimeMs:_+A}:{index:B}}finally{!a&&u&&d(u),n&&d(n),f&&d(f),l&&d(l),m&&d(m),p&&d(p)}}async function Qo(r,t,e,o){let a=e instanceof N;if(q(r),T(r,"idamax",{x:e}),!Number.isInteger(t)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!a&&!(e instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(a&&e.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(t<=0)return{index:0};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=["f64/dekker","f64/utils/abs","f64/utils/greater","f64/utils/equal"],s=await D(r,[...i,"idamax"],"idamax_main"),u=await D(r,[...i,"reduction/argmaxF64"],"reduce_f64"),n=null,f=null,l=null,m=null,p=null,c=null,g=null,w=null;try{if(a)n=e._buf,f=e._loBuf;else{let{hi:F,lo:W}=rr(e);n=x(r,F,"idamax-xHi",!1),f=x(r,W,"idamax-xLo",!1)}l=tr(r,512,"idamax-partials-val-hi"),m=tr(r,512,"idamax-partials-val-lo"),p=tr(r,512,"idamax-partials-idx"),c=Br(r,4,"idamax-result"),g=I(r,[{value:t,type:"u32"},{value:o,type:"u32"}],"idamax-params");let h=E(r,s.getBindGroupLayout(0),[n,f,l,m,p,g]),{commandEncoder:b,ts:y}=j(r,s,h,128);M(r,b);let v=E(r,u.getBindGroupLayout(0),[l,m,p,c]),{commandEncoder:_,ts:A}=j(r,u,v,1);w=G(r,_,c),M(r,_);let k=S(w,Uint32Array);w=null;let[B,L,C]=await Promise.all([P(y),P(A),k]),R=C[0];return B!==void 0&&L!==void 0?{index:R,gpuTimeMs:B+L}:{index:R}}finally{!a&&n&&d(n),!a&&f&&d(f),l&&d(l),m&&d(m),p&&d(p),c&&d(c),g&&d(g),w&&d(w)}}async function Jo(r,t,e,o,a,i,s,u){let n=e instanceof N,f=a instanceof N;if(q(r),T(r,"srot",{x:e,y:a}),!Number.isInteger(t)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(typeof s!="number")throw new Error("c must be a number.");if(typeof u!="number")throw new Error("s must be a number.");if(Number.isNaN(s)||Number.isNaN(u))throw new Error("c and s must not be NaN.");if(!Number.isFinite(s))throw new Error("c must be finite.");if(!Number.isFinite(u))throw new Error("s must be finite.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!n&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!f&&!(a instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(n!==f)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(t<=0)return n?{}:{x:e,y:a};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(a.length<(t-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let l=await D(r,"srot"),m=null,p=null,c=null,g=null,w=null;try{m=n?e._buf:x(r,e,"srot-x",!0),p=f?a._buf:x(r,a,"srot-y",!0),c=I(r,[{value:t,type:"u32"},{value:s,type:"f32"},{value:u,type:"f32"},{value:o,type:"u32"},{value:i,type:"u32"}],"srot-params");let h=E(r,l.getBindGroupLayout(0),[m,p,c]),{commandEncoder:b,ts:y}=j(r,l,h,cr(r,t));g=n?null:G(r,b,m),w=f?null:G(r,b,p),M(r,b);let v=await P(y);if(n)return v!==void 0?{gpuTimeMs:v}:{};let _=S(g,Float32Array),A=S(w,Float32Array);g=null,w=null;let[k,B]=await Promise.all([_,A]);return v!==void 0?{x:k,y:B,gpuTimeMs:v}:{x:k,y:B}}finally{!n&&m&&d(m),!f&&p&&d(p),c&&d(c),g&&d(g),w&&d(w)}}async function ra(r,t,e,o,a,i,s,u){let n=e instanceof N,f=a instanceof N;if(q(r),T(r,"drot",{x:e,y:a}),!Number.isInteger(t)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(typeof s!="number")throw new Error("c must be a number.");if(typeof u!="number")throw new Error("s must be a number.");if(Number.isNaN(s)||Number.isNaN(u))throw new Error("c and s must not be NaN.");if(!Number.isFinite(s))throw new Error("c must be finite.");if(!Number.isFinite(u))throw new Error("s must be finite.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!(e instanceof Float64Array)&&!n)throw new Error("x must be a Float64Array or GpuVector.");if(!(a instanceof Float64Array)&&!f)throw new Error("y must be a Float64Array or GpuVector.");if(n&&e.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(f&&a.dtype!==Float64Array)throw new Error("y must be a Float64Array-backed GpuVector.");if(n!==f)throw new Error("x and y must be the same type (both Float64Array or both GpuVector).");if(t<=0)return n?{}:{x:e,y:a};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(a.length<(t-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let m=await D(r,[...["f64/dekker","f64/utils/add","f64/utils/multiply"],"drot"]),{hi:p,lo:c}=rr(new Float64Array([s])),{hi:g,lo:w}=rr(new Float64Array([u])),h=null,b=null,y=null,v=null,_=null,A=null,k=null,B=null,L=null;try{if(n)h=e._buf,b=e._loBuf,y=a._buf,v=a._loBuf;else{let ir=rr(e),lr=rr(a);h=x(r,ir.hi,"drot-xHi",!0),b=x(r,ir.lo,"drot-xLo",!0),y=x(r,lr.hi,"drot-yHi",!0),v=x(r,lr.lo,"drot-yLo",!0)}_=I(r,[{value:t,type:"u32"},{value:p[0],type:"f32"},{value:c[0],type:"f32"},{value:g[0],type:"f32"},{value:w[0],type:"f32"},{value:o,type:"u32"},{value:i,type:"u32"}],"drot-params");let C=E(r,m.getBindGroupLayout(0),[h,b,y,v,_]),{commandEncoder:R,ts:F}=j(r,m,C,cr(r,t));A=n?null:G(r,R,h),k=n?null:G(r,R,b),B=f?null:G(r,R,y),L=f?null:G(r,R,v),M(r,R);let W=await P(F);if(n)return W!==void 0?{gpuTimeMs:W}:{};let V=await S(A,Float32Array);A=null;let z=await S(k,Float32Array);k=null;let H=await S(B,Float32Array);B=null;let $=await S(L,Float32Array);L=null;let K=dr(V,z),Y=dr(H,$);return W!==void 0?{x:K,y:Y,gpuTimeMs:W}:{x:K,y:Y}}finally{!n&&h&&d(h),!n&&b&&d(b),!f&&y&&d(y),!f&&v&&d(v),_&&d(_),A&&d(A),k&&d(k),B&&d(B),L&&d(L)}}async function ea(r,t,e,o,a,i,s){let u=e instanceof N,n=a instanceof N;if(q(r),T(r,"srotm",{x:e,y:a}),!Number.isInteger(t)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(!(s instanceof Float32Array)||s.length!==5)throw new Error("param must be a Float32Array of length 5.");if(s[0]!==-2&&s[0]!==-1&&s[0]!==0&&s[0]!==1)throw new Error("param[0] (flag) must be one of -2, -1, 0, or 1.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!u&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!n&&!(a instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(u!==n)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(t<=0||s[0]===-2)return u?{}:{x:e,y:a};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(a.length<(t-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let f=await D(r,"srotm"),l=null,m=null,p=null,c=null,g=null,w=null;try{l=u?e._buf:x(r,e,"srotm-x",!0),m=n?a._buf:x(r,a,"srotm-y",!0),p=x(r,s,"srotm-param",!1),c=I(r,[{value:t,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"srotm-params");let h=E(r,f.getBindGroupLayout(0),[l,m,p,c]),{commandEncoder:b,ts:y}=j(r,f,h,cr(r,t));g=u?null:G(r,b,l),w=n?null:G(r,b,m),M(r,b);let v=await P(y);if(u)return v!==void 0?{gpuTimeMs:v}:{};let _=S(g,Float32Array),A=S(w,Float32Array);g=null,w=null;let[k,B]=await Promise.all([_,A]);return v!==void 0?{x:k,y:B,gpuTimeMs:v}:{x:k,y:B}}finally{!u&&l&&d(l),!n&&m&&d(m),p&&d(p),c&&d(c),g&&d(g),w&&d(w)}}async function ta(r,t,e,o,a,i,s){let u=e instanceof N,n=a instanceof N;if(q(r),T(r,"drotm",{x:e,y:a}),!Number.isInteger(t)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(!(s instanceof Float64Array)||s.length!==5)throw new Error("param must be a Float64Array of length 5.");if(s[0]!==-2&&s[0]!==-1&&s[0]!==0&&s[0]!==1)throw new Error("param[0] (flag) must be one of -2, -1, 0, or 1.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!(e instanceof Float64Array)&&!u)throw new Error("x must be a Float64Array or GpuVector.");if(!(a instanceof Float64Array)&&!n)throw new Error("y must be a Float64Array or GpuVector.");if(u&&e.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(n&&a.dtype!==Float64Array)throw new Error("y must be a Float64Array-backed GpuVector.");if(u!==n)throw new Error("x and y must be the same type (both Float64Array or both GpuVector).");if(t<=0||s[0]===-2)return u?{}:{x:e,y:a};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(a.length<(t-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let l=await D(r,[...["f64/dekker","f64/utils/add","f64/utils/multiply"],"drotm"]),{hi:m,lo:p}=rr(s),c=null,g=null,w=null,h=null,b=null,y=null,v=null,_=null,A=null,k=null,B=null;try{if(u)c=e._buf,g=e._loBuf,w=a._buf,h=a._loBuf;else{let Y=rr(e),ir=rr(a);c=x(r,Y.hi,"drotm-xHi",!0),g=x(r,Y.lo,"drotm-xLo",!0),w=x(r,ir.hi,"drotm-yHi",!0),h=x(r,ir.lo,"drotm-yLo",!0)}b=x(r,m,"drotm-paramHi",!1),y=x(r,p,"drotm-paramLo",!1),v=I(r,[{value:t,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"drotm-params");let L=E(r,l.getBindGroupLayout(0),[c,g,w,h,b,y,v]),{commandEncoder:C,ts:R}=j(r,l,L,cr(r,t));_=u?null:G(r,C,c),A=u?null:G(r,C,g),k=n?null:G(r,C,w),B=n?null:G(r,C,h),M(r,C);let F=await P(R);if(u)return F!==void 0?{gpuTimeMs:F}:{};let W=await S(_,Float32Array);_=null;let V=await S(A,Float32Array);A=null;let z=await S(k,Float32Array);k=null;let H=await S(B,Float32Array);B=null;let $=dr(W,V),K=dr(z,H);return F!==void 0?{x:$,y:K,gpuTimeMs:F}:{x:$,y:K}}finally{!u&&c&&d(c),!u&&g&&d(g),!n&&w&&d(w),!n&&h&&d(h),b&&d(b),y&&d(y),v&&d(v),_&&d(_),A&&d(A),k&&d(k),B&&d(B)}}async function oa(r,t,e,o,a,i,s,u,n,f,l,m,p="row-major"){let c=i instanceof X,g=u instanceof N,w=l instanceof N;if(q(r),T(r,"sgemv",{A:i,x:u,y:l}),t!=="no-transpose"&&t!=="transpose")throw new Error("trans must be 'no-transpose' or 'transpose'.");if(p!=="row-major"&&p!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(typeof a!="number")throw new Error("alpha must be a number.");if(Number.isNaN(a))throw new Error("alpha must not be NaN.");if(!Number.isFinite(a))throw new Error("alpha must be finite.");if(typeof f!="number")throw new Error("beta must be a number.");if(Number.isNaN(f))throw new Error("beta must not be NaN.");if(!Number.isFinite(f))throw new Error("beta must be finite.");if(!Number.isInteger(e)||!Number.isInteger(o)||!Number.isInteger(n)||!Number.isInteger(m)||!Number.isInteger(s))throw new Error("m, n, incx, incy, and lda must be integers.");if(n<=0||m<=0)throw new Error("incx and incy must be positive.");if(!c&&!(i instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!g&&!(u instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!w&&!(l instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(g!==w)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(g&&!c)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(c&&!g)throw new Error("x and y must be GpuVectors when A is a GpuMatrix.");if(g&&u._buf===l._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(c&&w&&i._buf===l._buf)throw new Error("A and y must not reference the same GPU buffer.");if(c&&s!==i.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(c&&(i.rows<e||i.cols<o))throw new Error("A is too small for the given m and n.");if(e<0||o<0)throw new Error("m and n must be non-negative.");if(e===0||o===0)return w?{}:{y:l};(c?i.layout:p)==="column-major"&&([e,o]=[o,e],t=t==="no-transpose"?"transpose":"no-transpose");let b=t==="no-transpose",y=b?o:e,v=b?e:o;if(s<o)throw new Error("lda must be >= n.");if(!c&&i.length<(e-1)*s+o)throw new Error("A does not have enough elements for the given m, n, and lda.");if(u.length<(y-1)*n+1)throw new Error("x does not have enough elements for the given dimensions and incx.");if(l.length<(v-1)*m+1)throw new Error("y does not have enough elements for the given dimensions and incy.");let A=await D(r,b?"sgemv_n":"sgemv_t"),k=null,B=null,L=null,C=null;try{k=c?i._buf:x(r,i,"sgemv-A",!1),B=g?u._buf:x(r,u,"sgemv-x",!1),L=w?l._buf:x(r,l,"sgemv-y",!0),C=I(r,[{value:e,type:"u32"},{value:o,type:"u32"},{value:a,type:"f32"},{value:f,type:"f32"},{value:n,type:"u32"},{value:m,type:"u32"},{value:s,type:"u32"}],"sgemv-params");let R=E(r,A.getBindGroupLayout(0),[k,B,L,C]),F=b?Math.min(e,r.limits.maxComputeWorkgroupsPerDimension):Zr(r,"sgemv",v),{commandEncoder:W,ts:V}=j(r,A,R,F),z=w?null:G(r,W,L);M(r,W);let H=await P(V);if(w)return H!==void 0?{gpuTimeMs:H}:{};let $=await S(z,Float32Array);return H!==void 0?{y:$,gpuTimeMs:H}:{y:$}}finally{!c&&k&&d(k),!g&&B&&d(B),!w&&L&&d(L),C&&d(C)}}async function aa(r,t,e,o,a,i,s,u,n,f,l,m="row-major"){let p=s instanceof N,c=f instanceof N,g=a instanceof X;if(q(r),T(r,"ssymv",{A:a,x:s,y:f}),t!=="lower"&&t!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(m!=="row-major"&&m!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(!Number.isInteger(e)||!Number.isInteger(u)||!Number.isInteger(l)||!Number.isInteger(i))throw new Error("n, incx, incy, and lda must be integers.");if(typeof o!="number")throw new Error("alpha must be a number.");if(Number.isNaN(o))throw new Error("alpha must not be NaN.");if(!Number.isFinite(o))throw new Error("alpha must be finite.");if(typeof n!="number")throw new Error("beta must be a number.");if(Number.isNaN(n))throw new Error("beta must not be NaN.");if(!Number.isFinite(n))throw new Error("beta must be finite.");if(u<=0||l<=0)throw new Error("incx and incy must be positive.");if(i<e)throw new Error("lda must be >= n.");if(!g&&!(a instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!p&&!(s instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!c&&!(f instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(p!==c)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(p&&!g)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(g&&!p)throw new Error("x and y must be GpuVectors when A is a GpuMatrix.");if(p&&s._buf===f._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(g&&i!==a.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(g&&(a.rows<e||a.cols<e))throw new Error("A is too small for the given n.");if(e<0)throw new Error("n must be non-negative.");if(e===0)return c?{}:{y:f};if(!g&&a.length<(e-1)*i+e)throw new Error("A does not have enough elements for the given n and lda.");if(s.length<(e-1)*u+1)throw new Error("x does not have enough elements for the given n and incx.");if(f.length<(e-1)*l+1)throw new Error("y does not have enough elements for the given n and incy.");let h=(g?a.layout:m)==="column-major"?t==="upper":t==="lower",b=await D(r,"ssymv"),y=null,v=null,_=null,A=null;try{y=g?a._buf:x(r,a,"ssymv-A",!1),v=p?s._buf:x(r,s,"ssymv-x",!1),_=c?f._buf:x(r,f,"ssymv-y",!0),A=I(r,[{value:e,type:"u32"},{value:o,type:"f32"},{value:n,type:"f32"},{value:u,type:"u32"},{value:l,type:"u32"},{value:i,type:"u32"},{value:h?0:1,type:"u32"}],"ssymv-params");let k=E(r,b.getBindGroupLayout(0),[y,v,_,A]),B=Math.min(e,r.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:L,ts:C}=j(r,b,k,B),R=c?null:G(r,L,_);M(r,L);let F=await P(C);if(c)return F!==void 0?{gpuTimeMs:F}:{};let W=await S(R,Float32Array);return F!==void 0?{y:W,gpuTimeMs:F}:{y:W}}finally{!g&&y&&d(y),!p&&v&&d(v),!c&&_&&d(_),A&&d(A)}}async function ia(r,t,e,o,a,i,s,u,n,f,l,m="row-major"){let p=u instanceof N,c=f instanceof N,g=i instanceof X,w=o==="unit";if(q(r),T(r,"strmv",{A:i,x:u,y:f}),t!=="lower"&&t!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(e!=="no-transpose"&&e!=="transpose")throw new Error("trans must be 'no-transpose' or 'transpose'.");if(!w&&o!=="non-unit")throw new Error("diag must be 'unit' or 'non-unit'.");if(m!=="row-major"&&m!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(!Number.isInteger(a)||!Number.isInteger(n)||!Number.isInteger(l)||!Number.isInteger(s))throw new Error("n, incx, incy, and lda must be integers.");if(n<=0||l<=0)throw new Error("incx and incy must be positive.");if(s<a)throw new Error("lda must be >= n.");if(!g&&!(i instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!p&&!(u instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!c&&!(f instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(p!==c)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(p&&u._buf===f._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(p&&!g)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(g&&!p)throw new Error("x and y must be GpuVectors when A is a GpuMatrix.");if(g&&c&&i._buf===f._buf)throw new Error("A and y must not reference the same GPU buffer.");if(g&&s!==i.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(g&&(i.rows<a||i.cols<a))throw new Error("A is too small for the given n.");if(a<0)throw new Error("n must be non-negative.");if(a===0)return c?{}:{y:f};if(!g&&i.length<(a-1)*s+a)throw new Error("A does not have enough elements for the given n and lda.");if(u.length<(a-1)*n+1)throw new Error("x does not have enough elements for the given n and incx.");if(f.length<(a-1)*l+1)throw new Error("y does not have enough elements for the given n and incy.");let b=(g?i.layout:m)==="column-major",y=b?t==="upper":t==="lower",v=b?e==="transpose":e==="no-transpose",_=await D(r,"strmv"),A=null,k=null,B=null,L=null;try{A=g?i._buf:x(r,i,"strmv-A",!1),k=p?u._buf:x(r,u,"strmv-x",!1),B=c?f._buf:x(r,f,"strmv-y",!0),L=I(r,[{value:a,type:"u32"},{value:n,type:"u32"},{value:l,type:"u32"},{value:s,type:"u32"},{value:v?0:1,type:"u32"},{value:y?0:1,type:"u32"},{value:w?1:0,type:"u32"}],"strmv-params");let C=E(r,_.getBindGroupLayout(0),[A,k,B,L]),R=Math.min(a,r.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:F,ts:W}=j(r,_,C,R),V=c?null:G(r,F,B);M(r,F);let z=await P(W);if(c)return z!==void 0?{gpuTimeMs:z}:{};let H=await S(V,Float32Array);return z!==void 0?{y:H,gpuTimeMs:z}:{y:H}}finally{!g&&A&&d(A),!p&&k&&d(k),!c&&B&&d(B),L&&d(L)}}function sa(r,t,e){let o=new ArrayBuffer(r*t),a=new DataView(o);for(let i=0;i<r;i++){let s=e(i),u=i*t;s.forEach((n,f)=>a.setUint32(u+f*4,n,!0))}return o}function na(r,t,e){let o=r.createBuffer({label:e,size:t.byteLength,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});return r.queue.writeBuffer(o,0,t),o}async function la(r,t,e,o,a,i,s,u,n,f="row-major"){let l=u instanceof N,m=i instanceof X,p=o==="unit";if(q(r),T(r,"strsv",{A:i,x:u}),t!=="lower"&&t!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(e!=="no-transpose"&&e!=="transpose")throw new Error("trans must be 'no-transpose' or 'transpose'.");if(!p&&o!=="non-unit")throw new Error("diag must be 'unit' or 'non-unit'.");if(f!=="row-major"&&f!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(!Number.isInteger(a)||!Number.isInteger(n)||!Number.isInteger(s))throw new Error("n, incx, and lda must be integers.");if(n<=0)throw new Error("incx must be positive.");if(s<a)throw new Error("lda must be >= n.");if(!m&&!(i instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!l&&!(u instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(l&&!m)throw new Error("A must be a GpuMatrix when x is a GpuVector.");if(m&&!l)throw new Error("x must be a GpuVector when A is a GpuMatrix.");if(m&&l&&i._buf===u._buf)throw new Error("A and x must not reference the same GPU buffer.");if(m&&s!==i.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(m&&(i.rows<a||i.cols<a))throw new Error("A is too small for the given n.");if(a<0)throw new Error("n must be non-negative.");if(a===0)return l?{}:{x:u};if(!m&&i.length<(a-1)*s+a)throw new Error("A does not have enough elements for the given n and lda.");if(u.length<(a-1)*n+1)throw new Error("x does not have enough elements for the given n and incx.");let g=(m?i.layout:f)==="column-major",w=g?t==="upper":t==="lower",h=g?e==="transpose":e==="no-transpose",b=await D(r,"strsv_invert_block"),y=await D(r,"strsv_apply_inverse"),v=await D(r,"strsv_update"),_=h===w,A=[];for(let H=0;H<a;H+=64)A.push(H);_||A.reverse();let k=A.length,B=r.limits.maxComputeWorkgroupsPerDimension,L=r.limits.minUniformBufferOffsetAlignment,C=null,R=null,F=null,W=null,V=null,z=null;try{C=m?i._buf:x(r,i,"strsv-A",!1),R=l?u._buf:x(r,u,"strsv-x",!0),F=tr(r,k*64*64*4,"strsv-Ainv");let H=sa(k,L,er=>{let Z=er*64,J=Math.min(Z+64,a);return[n,er,Z,J]});W=na(r,H,"strsv-apply-params");let $=sa(k,L,er=>{let Z=er*64,J=Math.min(Z+64,a);return[a,n,s,h?0:1,w?0:1,Z,J]});V=na(r,$,"strsv-update-params");let{commandEncoder:K,querySet:Y}=qr(r);z=I(r,[{value:a,type:"u32"},{value:s,type:"u32"},{value:h?0:1,type:"u32"},{value:w?0:1,type:"u32"},{value:p?1:0,type:"u32"}],"strsv-invert-params");let ir=E(r,b.getBindGroupLayout(0),[C,F,z]);gr(K,b,ir,{x:64,y:k},Y?{timestampWrites:{querySet:Y,beginningOfPassWriteIndex:0}}:void 0);for(let er=0;er<A.length;er++){let Z=A[er],J=Math.min(Z+64,a),sr=Z/64,pr=er===A.length-1,br=sr*L,fr=E(r,y.getBindGroupLayout(0),[F,R,{buffer:W,offset:br,size:16}]);gr(K,y,fr,1,pr&&Y?{timestampWrites:{querySet:Y,endOfPassWriteIndex:1}}:void 0);let yr=_?a-J:Z;if(yr===0)continue;let Cr=E(r,v.getBindGroupLayout(0),[C,R,{buffer:V,offset:br,size:32}]),Rr=Math.min(yr,B);gr(K,v,Cr,Rr)}let hr=Lr(r,K,Y),nr=l?null:G(r,K,R);M(r,K);let mr=await P(hr);if(l)return mr!==void 0?{gpuTimeMs:mr}:{};let Q=await S(nr,Float32Array);return mr!==void 0?{x:Q,gpuTimeMs:mr}:{x:Q}}finally{!m&&C&&d(C),!l&&R&&d(R),F&&d(F),W&&d(W),V&&d(V),z&&d(z)}}async function ua(r,t,e,o,a,i,s,u,n,f,l="row-major"){let m=n instanceof X;if(q(r),T(r,"sger",{A:n,x:a,y:s}),l!=="row-major"&&l!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(typeof o!="number")throw new Error("alpha must be a number.");if(Number.isNaN(o))throw new Error("alpha must not be NaN.");if(!Number.isFinite(o))throw new Error("alpha must be finite.");if(!Number.isInteger(t)||!Number.isInteger(e)||!Number.isInteger(i)||!Number.isInteger(u)||!Number.isInteger(f))throw new Error("m, n, incx, incy, and lda must be integers.");if(i<=0||u<=0)throw new Error("incx and incy must be positive.");if(!m&&!(n instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(m&&f!==n.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(m&&(n.rows<t||n.cols<e))throw new Error("A is too small for the given m and n.");(m?n.layout:l)==="column-major"&&([t,e]=[e,t],[a,s]=[s,a],[i,u]=[u,i]);let c=a instanceof N,g=s instanceof N;if(f<e)throw new Error("lda must be >= n.");if(!c&&!(a instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!g&&!(s instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(c!==g)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(c&&!m)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(m&&!c)throw new Error("x and y must be GpuVectors when A is a GpuMatrix.");if(m&&c&&n._buf===a._buf)throw new Error("A and x must not reference the same GPU buffer.");if(m&&g&&n._buf===s._buf)throw new Error("A and y must not reference the same GPU buffer.");if(t<0||e<0)throw new Error("m and n must be non-negative.");if(t===0||e===0)return m?{}:{A:n};if(!m&&n.length<(t-1)*f+e)throw new Error("A does not have enough elements for the given m, n, and lda.");if(a.length<(t-1)*i+1)throw new Error("x does not have enough elements for the given m and incx.");if(s.length<(e-1)*u+1)throw new Error("y does not have enough elements for the given n and incy.");let w=await D(r,"sger"),h=null,b=null,y=null,v=null;try{h=c?a._buf:x(r,a,"sger-x",!1),b=g?s._buf:x(r,s,"sger-y",!1),y=m?n._buf:x(r,n,"sger-A",!0),v=I(r,[{value:t,type:"u32"},{value:e,type:"u32"},{value:o,type:"f32"},{value:i,type:"u32"},{value:u,type:"u32"},{value:f,type:"u32"}],"sger-params");let _=E(r,w.getBindGroupLayout(0),[h,b,y,v]),A=Math.min(t,r.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:k,ts:B}=j(r,w,_,A),L=m?null:G(r,k,y);M(r,k);let C=await P(B);if(m)return C!==void 0?{gpuTimeMs:C}:{};let R=await S(L,Float32Array);return C!==void 0?{A:R,gpuTimeMs:C}:{A:R}}finally{!c&&h&&d(h),!g&&b&&d(b),!m&&y&&d(y),v&&d(v)}}async function fa(r,t,e,o,a,i,s,u,n="row-major"){let f=a instanceof N,l=s instanceof X;if(q(r),T(r,"ssyr",{A:s,x:a}),t!=="lower"&&t!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(n!=="row-major"&&n!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(!Number.isInteger(e)||!Number.isInteger(i)||!Number.isInteger(u))throw new Error("n, incx, and lda must be integers.");if(typeof o!="number")throw new Error("alpha must be a number.");if(Number.isNaN(o))throw new Error("alpha must not be NaN.");if(!Number.isFinite(o))throw new Error("alpha must be finite.");if(i<=0)throw new Error("incx must be positive.");if(u<e)throw new Error("lda must be >= n.");if(!l&&!(s instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!f&&!(a instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(f&&!l)throw new Error("A must be a GpuMatrix when x is a GpuVector.");if(l&&!f)throw new Error("x must be a GpuVector when A is a GpuMatrix.");if(l&&f&&s._buf===a._buf)throw new Error("A and x must not reference the same GPU buffer.");if(l&&u!==s.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(l&&(s.rows<e||s.cols<e))throw new Error("A is too small for the given n.");if(e<0)throw new Error("n must be non-negative.");if(e===0)return l?{}:{A:s};if(!l&&s.length<(e-1)*u+e)throw new Error("A does not have enough elements for the given n and lda.");if(a.length<(e-1)*i+1)throw new Error("x does not have enough elements for the given n and incx.");let p=(l?s.layout:n)==="column-major"?t==="upper":t==="lower",c=await D(r,"ssyr"),g=null,w=null,h=null;try{g=f?a._buf:x(r,a,"ssyr-x",!1),w=l?s._buf:x(r,s,"ssyr-A",!0),h=I(r,[{value:e,type:"u32"},{value:o,type:"f32"},{value:i,type:"u32"},{value:u,type:"u32"},{value:p?0:1,type:"u32"}],"ssyr-params");let b=E(r,c.getBindGroupLayout(0),[g,w,h]),y=Math.min(e,r.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:v,ts:_}=j(r,c,b,y),A=l?null:G(r,v,w);M(r,v);let k=await P(_);if(l)return k!==void 0?{gpuTimeMs:k}:{};let B=await S(A,Float32Array);return k!==void 0?{A:B,gpuTimeMs:k}:{A:B}}finally{!f&&g&&d(g),!l&&w&&d(w),h&&d(h)}}async function ma(r,t,e,o,a,i,s,u,n,f,l="row-major"){let m=a instanceof N,p=s instanceof N,c=n instanceof X;if(q(r),T(r,"ssyr2",{A:n,x:a,y:s}),t!=="lower"&&t!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(l!=="row-major"&&l!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(!Number.isInteger(e)||!Number.isInteger(i)||!Number.isInteger(u)||!Number.isInteger(f))throw new Error("n, incx, incy, and lda must be integers.");if(typeof o!="number")throw new Error("alpha must be a number.");if(Number.isNaN(o))throw new Error("alpha must not be NaN.");if(!Number.isFinite(o))throw new Error("alpha must be finite.");if(i<=0||u<=0)throw new Error("incx and incy must be positive.");if(f<e)throw new Error("lda must be >= n.");if(!c&&!(n instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!m&&!(a instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!p&&!(s instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(m!==p)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(m&&!c)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(c&&!m)throw new Error("x and y must be GpuVectors when A is a GpuMatrix.");if(c&&m&&n._buf===a._buf)throw new Error("A and x must not reference the same GPU buffer.");if(c&&p&&n._buf===s._buf)throw new Error("A and y must not reference the same GPU buffer.");if(m&&a._buf===s._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(c&&f!==n.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(c&&(n.rows<e||n.cols<e))throw new Error("A is too small for the given n.");if(e<0)throw new Error("n must be non-negative.");if(e===0)return c?{}:{A:n};if(!c&&n.length<(e-1)*f+e)throw new Error("A does not have enough elements for the given n and lda.");if(a.length<(e-1)*i+1)throw new Error("x does not have enough elements for the given n and incx.");if(s.length<(e-1)*u+1)throw new Error("y does not have enough elements for the given n and incy.");let w=(c?n.layout:l)==="column-major"?t==="upper":t==="lower",h=await D(r,"ssyr2"),b=null,y=null,v=null,_=null;try{b=m?a._buf:x(r,a,"ssyr2-x",!1),y=p?s._buf:x(r,s,"ssyr2-y",!1),v=c?n._buf:x(r,n,"ssyr2-A",!0),_=I(r,[{value:e,type:"u32"},{value:o,type:"f32"},{value:i,type:"u32"},{value:u,type:"u32"},{value:f,type:"u32"},{value:w?0:1,type:"u32"}],"ssyr2-params");let A=E(r,h.getBindGroupLayout(0),[b,y,v,_]),k=Math.min(e,r.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:B,ts:L}=j(r,h,A,k),C=c?null:G(r,B,v);M(r,B);let R=await P(L);if(c)return R!==void 0?{gpuTimeMs:R}:{};let F=await S(C,Float32Array);return R!==void 0?{A:F,gpuTimeMs:R}:{A:F}}finally{!m&&b&&d(b),!p&&y&&d(y),!c&&v&&d(v),_&&d(_)}}async function da(r,t,e,o,a,i,s,u,n,f,l,m,p,c,g="row-major"){let w=u instanceof X,h=f instanceof X,b=p instanceof X;if(q(r),T(r,"sgemm",{A:u,B:f,C:p}),t!=="no-transpose"&&t!=="transpose")throw new Error("transA must be 'no-transpose' or 'transpose'.");if(e!=="no-transpose"&&e!=="transpose")throw new Error("transB must be 'no-transpose' or 'transpose'.");if(g!=="row-major"&&g!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(typeof s!="number")throw new Error("alpha must be a number.");if(Number.isNaN(s))throw new Error("alpha must not be NaN.");if(!Number.isFinite(s))throw new Error("alpha must be finite.");if(typeof m!="number")throw new Error("beta must be a number.");if(Number.isNaN(m))throw new Error("beta must not be NaN.");if(!Number.isFinite(m))throw new Error("beta must be finite.");if(!Number.isInteger(o)||!Number.isInteger(a)||!Number.isInteger(i)||!Number.isInteger(n)||!Number.isInteger(l)||!Number.isInteger(c))throw new Error("m, n, k, lda, ldb, and ldc must be integers.");if(!w&&!(u instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!h&&!(f instanceof Float32Array))throw new Error("B must be a Float32Array or GpuMatrix.");if(!b&&!(p instanceof Float32Array))throw new Error("C must be a Float32Array or GpuMatrix.");if((w||h)&&!b)throw new Error("C must be a GpuMatrix when A or B is a GpuMatrix.");if(b&&(!w||!h))throw new Error("A and B must be GpuMatrix when C is a GpuMatrix.");if(o<0||a<0||i<0)throw new Error("m, n, and k must be non-negative.");if(n<=0||l<=0||c<=0)throw new Error("lda, ldb, and ldc must be positive.");if(o===0||a===0)return b?{}:{C:p};let y=w?u.layout:g,v=h?f.layout:g,_=b?p.layout:g,A=y==="column-major"?i:o,k=y==="column-major"?o:i,B=t==="no-transpose"?A:k,L=t==="no-transpose"?k:A;if(n<L)throw new Error(`lda must be >= ${y==="column-major"?"rows":"cols"} of A as stored.`);if(w){if(n!==u.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");let[J,sr]=t==="no-transpose"?[o,i]:[i,o];if(u.rows<J||u.cols<sr)throw new Error("A is too small for the given m, k, and transA.")}else if(u.length<(B-1)*n+L)throw new Error("A does not have enough elements for the given dimensions and lda.");let C=v==="column-major"?a:i,R=v==="column-major"?i:a,F=e==="no-transpose"?C:R,W=e==="no-transpose"?R:C;if(l<W)throw new Error(`ldb must be >= ${v==="column-major"?"rows":"cols"} of B as stored.`);if(h){if(l!==f.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");let[J,sr]=e==="no-transpose"?[i,a]:[a,i];if(f.rows<J||f.cols<sr)throw new Error("B is too small for the given n, k, and transB.")}else if(f.length<(F-1)*l+W)throw new Error("B does not have enough elements for the given dimensions and ldb.");let V=_==="column-major"?a:o,z=_==="column-major"?o:a;if(c<z)throw new Error(`ldc must be >= ${_==="column-major"?"rows":"cols"} of C as stored.`);if(b){if(c!==p.lda)throw new Error("ldc must match C.lda when C is a GpuMatrix.");if(p.rows<o||p.cols<a)throw new Error("C is too small for the given m and n.")}else if(p.length<(V-1)*c+z)throw new Error("C does not have enough elements for the given dimensions and ldc.");y==="column-major"&&(t=t==="no-transpose"?"transpose":"no-transpose"),v==="column-major"&&(e=e==="no-transpose"?"transpose":"no-transpose"),_==="column-major"&&([u,f]=[f,u],[w,h]=[h,w],[n,l]=[l,n],[t,e]=[e==="no-transpose"?"transpose":"no-transpose",t==="no-transpose"?"transpose":"no-transpose"],[o,a]=[a,o]);let H=Math.ceil(a/64),$=Math.ceil(o/64),K=H*$>=36,Y=await D(r,K?"sgemm_large":"sgemm_small"),ir=w?u._buf:x(r,u,"sgemm-A",!1),lr=h?f._buf:x(r,f,"sgemm-B",!1),hr=b?p._buf:x(r,p,"sgemm-C",!0),nr=t==="no-transpose",mr=e==="no-transpose",Q=nr&&Ee(ir,n,o,i),er=Ee(lr,l,mr?i:a,mr?a:i),Z=I(r,[{value:o,type:"u32"},{value:a,type:"u32"},{value:i,type:"u32"},{value:s,type:"f32"},{value:m,type:"f32"},{value:n,type:"u32"},{value:l,type:"u32"},{value:c,type:"u32"},{value:t==="transpose"?1:0,type:"u32"},{value:e==="transpose"?1:0,type:"u32"},{value:Q?1:0,type:"u32"},{value:er?1:0,type:"u32"}],"sgemm-params");try{let J=E(r,Y.getBindGroupLayout(0),[ir,Sr(r,ir),lr,Sr(r,lr),hr,Z]),sr=K?{x:U(r,H,"sgemm","x"),y:U(r,$,"sgemm","y")}:{x:U(r,Math.ceil(a/32),"sgemm","x"),y:U(r,Math.ceil(o/32),"sgemm","y")},{commandEncoder:pr,ts:br}=j(r,Y,J,sr),fr=b?null:G(r,pr,hr);M(r,pr);let ur=await P(br);if(b)return ur!==void 0?{gpuTimeMs:ur}:{};let yr=await S(fr,Float32Array);return ur!==void 0?{C:yr,gpuTimeMs:ur}:{C:yr}}finally{w||d(ir),h||d(lr),b||d(hr),d(Z)}}async function ca(r,t,e,o,a,i,s,u,n,f,l,m,p,c,g,w="row-major"){let h=n instanceof X,b=l instanceof X,y=c instanceof X;if(q(r),T(r,"sgemmtr",{A:n,B:l,C:c}),t!=="lower"&&t!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(e!=="no-transpose"&&e!=="transpose")throw new Error("transA must be 'no-transpose' or 'transpose'.");if(o!=="no-transpose"&&o!=="transpose")throw new Error("transB must be 'no-transpose' or 'transpose'.");if(w!=="row-major"&&w!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(typeof u!="number")throw new Error("alpha must be a number.");if(Number.isNaN(u))throw new Error("alpha must not be NaN.");if(!Number.isFinite(u))throw new Error("alpha must be finite.");if(typeof p!="number")throw new Error("beta must be a number.");if(Number.isNaN(p))throw new Error("beta must not be NaN.");if(!Number.isFinite(p))throw new Error("beta must be finite.");if(!Number.isInteger(a)||!Number.isInteger(i)||!Number.isInteger(s)||!Number.isInteger(f)||!Number.isInteger(m)||!Number.isInteger(g))throw new Error("m, n, k, lda, ldb, and ldc must be integers.");if(!h&&!(n instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!b&&!(l instanceof Float32Array))throw new Error("B must be a Float32Array or GpuMatrix.");if(!y&&!(c instanceof Float32Array))throw new Error("C must be a Float32Array or GpuMatrix.");if((h||b)&&!y)throw new Error("C must be a GpuMatrix when A or B is a GpuMatrix.");if(y&&(!h||!b))throw new Error("A and B must be GpuMatrix when C is a GpuMatrix.");if(a<0||i<0||s<0)throw new Error("m, n, and k must be non-negative.");if(f<=0||m<=0||g<=0)throw new Error("lda, ldb, and ldc must be positive.");if(a===0||i===0)return y?{}:{C:c};let v=h?n.layout:w,_=b?l.layout:w,A=y?c.layout:w,k=v==="column-major"?s:a,B=v==="column-major"?a:s,L=e==="no-transpose"?k:B,C=e==="no-transpose"?B:k;if(f<C)throw new Error(`lda must be >= ${v==="column-major"?"rows":"cols"} of A as stored.`);if(h){if(f!==n.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");let[Q,er]=e==="no-transpose"?[a,s]:[s,a];if(n.rows<Q||n.cols<er)throw new Error("A is too small for the given m, k, and transA.")}else if(n.length<(L-1)*f+C)throw new Error("A does not have enough elements for the given dimensions and lda.");let R=_==="column-major"?i:s,F=_==="column-major"?s:i,W=o==="no-transpose"?R:F,V=o==="no-transpose"?F:R;if(m<V)throw new Error(`ldb must be >= ${_==="column-major"?"rows":"cols"} of B as stored.`);if(b){if(m!==l.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");let[Q,er]=o==="no-transpose"?[s,i]:[i,s];if(l.rows<Q||l.cols<er)throw new Error("B is too small for the given n, k, and transB.")}else if(l.length<(W-1)*m+V)throw new Error("B does not have enough elements for the given dimensions and ldb.");let z=A==="column-major"?i:a,H=A==="column-major"?a:i;if(g<H)throw new Error(`ldc must be >= ${A==="column-major"?"rows":"cols"} of C as stored.`);if(y){if(g!==c.lda)throw new Error("ldc must match C.lda when C is a GpuMatrix.");if(c.rows<a||c.cols<i)throw new Error("C is too small for the given m and n.")}else if(c.length<(z-1)*g+H)throw new Error("C does not have enough elements for the given dimensions and ldc.");v==="column-major"&&(e=e==="no-transpose"?"transpose":"no-transpose"),_==="column-major"&&(o=o==="no-transpose"?"transpose":"no-transpose"),A==="column-major"&&([n,l]=[l,n],[h,b]=[b,h],[f,m]=[m,f],[e,o]=[o==="no-transpose"?"transpose":"no-transpose",e==="no-transpose"?"transpose":"no-transpose"],[a,i]=[i,a],t=t==="lower"?"upper":"lower");let $=Math.ceil(i/64),K=Math.ceil(a/64),Y=$*K>=36,ir=await D(r,Y?"sgemmtr_large":"sgemmtr_small"),lr=h?n._buf:x(r,n,"sgemmtr-A",!1),hr=b?l._buf:x(r,l,"sgemmtr-B",!1),nr=y?c._buf:x(r,c,"sgemmtr-C",!0),mr=I(r,[{value:a,type:"u32"},{value:i,type:"u32"},{value:s,type:"u32"},{value:u,type:"f32"},{value:p,type:"f32"},{value:f,type:"u32"},{value:m,type:"u32"},{value:g,type:"u32"},{value:e==="transpose"?1:0,type:"u32"},{value:o==="transpose"?1:0,type:"u32"},{value:t==="upper"?1:0,type:"u32"}],"sgemmtr-params");try{let Q=E(r,ir.getBindGroupLayout(0),[lr,hr,nr,mr]),er=Y?{x:U(r,$,"sgemmtr","x"),y:U(r,K,"sgemmtr","y")}:{x:U(r,Math.ceil(i/32),"sgemmtr","x"),y:U(r,Math.ceil(a/32),"sgemmtr","y")},{commandEncoder:Z,ts:J}=j(r,ir,Q,er),sr=y?null:G(r,Z,nr);M(r,Z);let pr=await P(J);if(y)return pr!==void 0?{gpuTimeMs:pr}:{};let br=await S(sr,Float32Array);return pr!==void 0?{C:br,gpuTimeMs:pr}:{C:br}}finally{h||d(lr),b||d(hr),y||d(nr),d(mr)}}async function pa(r,t,e,o,a,i,s,u,n,f,l,m="row-major"){let p=s instanceof X,c=f instanceof X;if(q(r),T(r,"ssyrk",{A:s,C:f}),t!=="lower"&&t!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(e!=="no-transpose"&&e!=="transpose")throw new Error("trans must be 'no-transpose' or 'transpose'.");if(m!=="row-major"&&m!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(typeof i!="number")throw new Error("alpha must be a number.");if(Number.isNaN(i))throw new Error("alpha must not be NaN.");if(!Number.isFinite(i))throw new Error("alpha must be finite.");if(typeof n!="number")throw new Error("beta must be a number.");if(Number.isNaN(n))throw new Error("beta must not be NaN.");if(!Number.isFinite(n))throw new Error("beta must be finite.");if(!Number.isInteger(o)||!Number.isInteger(a)||!Number.isInteger(u)||!Number.isInteger(l))throw new Error("n, k, lda, and ldc must be integers.");if(!p&&!(s instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!c&&!(f instanceof Float32Array))throw new Error("C must be a Float32Array or GpuMatrix.");if(p&&!c)throw new Error("C must be a GpuMatrix when A is a GpuMatrix.");if(c&&!p)throw new Error("A must be a GpuMatrix when C is a GpuMatrix.");if(o<0||a<0)throw new Error("n and k must be non-negative.");if(u<=0||l<=0)throw new Error("lda and ldc must be positive.");if(o===0)return c?{}:{C:f};let g=p?s.layout:m,w=c?f.layout:m,h=g==="column-major"?a:o,b=g==="column-major"?o:a,y=e==="no-transpose"?h:b,v=e==="no-transpose"?b:h;if(u<v)throw new Error(`lda must be >= ${g==="column-major"?"rows":"cols"} of A as stored.`);if(p){if(u!==s.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");let[H,$]=e==="no-transpose"?[o,a]:[a,o];if(s.rows<H||s.cols<$)throw new Error("A is too small for the given n, k, and trans.")}else if(s.length<(y-1)*u+v)throw new Error("A does not have enough elements for the given dimensions and lda.");if(l<o)throw new Error("ldc must be >= n.");if(c){if(l!==f.lda)throw new Error("ldc must match C.lda when C is a GpuMatrix.");if(f.rows<o||f.cols<o)throw new Error("C is too small for the given n.")}else if(f.length<(o-1)*l+o)throw new Error("C does not have enough elements for the given dimensions and ldc.");let _=e;g==="column-major"&&(_=_==="no-transpose"?"transpose":"no-transpose");let A=_==="no-transpose"?"transpose":"no-transpose",k=t;w==="column-major"&&([_,A]=[A==="no-transpose"?"transpose":"no-transpose",_==="no-transpose"?"transpose":"no-transpose"],k=k==="lower"?"upper":"lower");let B=Math.ceil(o/64),L=Math.ceil(o/64),C=B*L>=36,R=await D(r,C?"sgemmtr_large":"sgemmtr_small"),F=p?s._buf:x(r,s,"ssyrk-A",!1),W=c?f._buf:x(r,f,"ssyrk-C",!0),V=p?tr(r,F.size,"ssyrk-B",GPUBufferUsage.COPY_DST):x(r,s,"ssyrk-B",!1),z=I(r,[{value:o,type:"u32"},{value:o,type:"u32"},{value:a,type:"u32"},{value:i,type:"f32"},{value:n,type:"f32"},{value:u,type:"u32"},{value:u,type:"u32"},{value:l,type:"u32"},{value:_==="transpose"?1:0,type:"u32"},{value:A==="transpose"?1:0,type:"u32"},{value:k==="upper"?1:0,type:"u32"}],"ssyrk-params");try{let H=E(r,R.getBindGroupLayout(0),[F,V,W,z]),$=C?{x:U(r,B,"ssyrk","x"),y:U(r,L,"ssyrk","y")}:{x:U(r,Math.ceil(o/32),"ssyrk","x"),y:U(r,Math.ceil(o/32),"ssyrk","y")},{commandEncoder:K,querySet:Y,passDescriptor:ir}=qr(r);p&&K.copyBufferToBuffer(F,0,V,0,F.size),gr(K,R,H,$,ir);let lr=Lr(r,K,Y),hr=c?null:G(r,K,W);M(r,K);let nr=await P(lr);if(c)return nr!==void 0?{gpuTimeMs:nr}:{};let mr=await S(hr,Float32Array);return nr!==void 0?{C:mr,gpuTimeMs:nr}:{C:mr}}finally{p||d(F),d(V),c||d(W),d(z)}}async function ga(r,t,e,o,a,i,s,u,n,f,l,m,p,c="row-major"){let g=s instanceof X,w=n instanceof X,h=m instanceof X;if(q(r),T(r,"ssyr2k",{A:s,B:n,C:m}),t!=="lower"&&t!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(e!=="no-transpose"&&e!=="transpose")throw new Error("trans must be 'no-transpose' or 'transpose'.");if(c!=="row-major"&&c!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(typeof i!="number")throw new Error("alpha must be a number.");if(Number.isNaN(i))throw new Error("alpha must not be NaN.");if(!Number.isFinite(i))throw new Error("alpha must be finite.");if(typeof l!="number")throw new Error("beta must be a number.");if(Number.isNaN(l))throw new Error("beta must not be NaN.");if(!Number.isFinite(l))throw new Error("beta must be finite.");if(!Number.isInteger(o)||!Number.isInteger(a)||!Number.isInteger(u)||!Number.isInteger(f)||!Number.isInteger(p))throw new Error("n, k, lda, ldb, and ldc must be integers.");if(!g&&!(s instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!w&&!(n instanceof Float32Array))throw new Error("B must be a Float32Array or GpuMatrix.");if(!h&&!(m instanceof Float32Array))throw new Error("C must be a Float32Array or GpuMatrix.");if((g||w)&&!h)throw new Error("C must be a GpuMatrix when A or B is a GpuMatrix.");if(h&&(!g||!w))throw new Error("A and B must be GpuMatrix when C is a GpuMatrix.");if(o<0||a<0)throw new Error("n and k must be non-negative.");if(u<=0||f<=0||p<=0)throw new Error("lda, ldb, and ldc must be positive.");if(o===0)return h?{}:{C:m};let b=g?s.layout:c,y=w?n.layout:c,v=h?m.layout:c,_=b==="column-major"?a:o,A=b==="column-major"?o:a,k=e==="no-transpose"?_:A,B=e==="no-transpose"?A:_;if(u<B)throw new Error(`lda must be >= ${b==="column-major"?"rows":"cols"} of A as stored.`);if(g){if(u!==s.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");let[J,sr]=e==="no-transpose"?[o,a]:[a,o];if(s.rows<J||s.cols<sr)throw new Error("A is too small for the given n, k, and trans.")}else if(s.length<(k-1)*u+B)throw new Error("A does not have enough elements for the given dimensions and lda.");let L=y==="column-major"?a:o,C=y==="column-major"?o:a,R=e==="no-transpose"?L:C,F=e==="no-transpose"?C:L;if(f<F)throw new Error(`ldb must be >= ${y==="column-major"?"rows":"cols"} of B as stored.`);if(w){if(f!==n.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");let[J,sr]=e==="no-transpose"?[o,a]:[a,o];if(n.rows<J||n.cols<sr)throw new Error("B is too small for the given n, k, and trans.")}else if(n.length<(R-1)*f+F)throw new Error("B does not have enough elements for the given dimensions and ldb.");if(p<o)throw new Error("ldc must be >= n.");if(h){if(p!==m.lda)throw new Error("ldc must match C.lda when C is a GpuMatrix.");if(m.rows<o||m.cols<o)throw new Error("C is too small for the given n.")}else if(m.length<(o-1)*p+o)throw new Error("C does not have enough elements for the given dimensions and ldc.");let W=e;b==="column-major"&&(W=W==="no-transpose"?"transpose":"no-transpose");let V=e;y==="column-major"&&(V=V==="no-transpose"?"transpose":"no-transpose");let z=v==="column-major"?t==="lower"?"upper":"lower":t,H=J=>J==="no-transpose"?"transpose":"no-transpose";function $(J,sr,pr,br,fr,ur){let yr=J,Cr=H(br);return v!=="column-major"?{transX:yr,X:sr,ldX:pr,transY:Cr,Y:fr,ldY:ur}:{transX:H(Cr),X:fr,ldX:ur,transY:H(yr),Y:sr,ldY:pr}}let K=Math.ceil(o/64),Y=Math.ceil(o/64),ir=K*Y>=36,lr=await D(r,ir?"sgemmtr_large":"sgemmtr_small"),hr=ir?{x:U(r,K,"ssyr2k","x"),y:U(r,Y,"ssyr2k","y")}:{x:U(r,Math.ceil(o/32),"ssyr2k","x"),y:U(r,Math.ceil(o/32),"ssyr2k","y")},nr=g?s._buf:x(r,s,"ssyr2k-A",!1),mr=w?n._buf:x(r,n,"ssyr2k-B",!1),Q=h?m._buf:x(r,m,"ssyr2k-C",!0),er=null,Z=null;try{let J=$(W,nr,u,V,mr,f),sr=$(V,mr,f,W,nr,u),pr=(Ir,Ar)=>I(r,[{value:o,type:"u32"},{value:o,type:"u32"},{value:a,type:"u32"},{value:i,type:"f32"},{value:Ar,type:"f32"},{value:Ir.ldX,type:"u32"},{value:Ir.ldY,type:"u32"},{value:p,type:"u32"},{value:Ir.transX==="transpose"?1:0,type:"u32"},{value:Ir.transY==="transpose"?1:0,type:"u32"},{value:z==="upper"?1:0,type:"u32"}],"ssyr2k-params");er=pr(J,l),Z=pr(sr,1);let br=E(r,lr.getBindGroupLayout(0),[J.X,J.Y,Q,er]),fr=E(r,lr.getBindGroupLayout(0),[sr.X,sr.Y,Q,Z]),{commandEncoder:ur,querySet:yr}=qr(r),Cr=yr?{timestampWrites:{querySet:yr,beginningOfPassWriteIndex:0}}:void 0,Rr=yr?{timestampWrites:{querySet:yr,endOfPassWriteIndex:1}}:void 0;gr(ur,lr,br,hr,Cr),gr(ur,lr,fr,hr,Rr);let Tr=Lr(r,ur,yr),Er=h?null:G(r,ur,Q);M(r,ur);let vr=await P(Tr);if(h)return vr!==void 0?{gpuTimeMs:vr}:{};let xr=await S(Er,Float32Array);return vr!==void 0?{C:xr,gpuTimeMs:vr}:{C:xr}}finally{g||d(nr),w||d(mr),h||d(Q),er&&d(er),Z&&d(Z)}}async function wa(r,t,e,o,a,i,s,u,n,f,l,m,p,c="row-major"){let g=s instanceof X,w=n instanceof X,h=m instanceof X;if(q(r),T(r,"ssymm",{A:s,B:n,C:m}),t!=="left"&&t!=="right")throw new Error("side must be 'left' or 'right'.");if(e!=="lower"&&e!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(c!=="row-major"&&c!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(typeof i!="number")throw new Error("alpha must be a number.");if(Number.isNaN(i))throw new Error("alpha must not be NaN.");if(!Number.isFinite(i))throw new Error("alpha must be finite.");if(typeof l!="number")throw new Error("beta must be a number.");if(Number.isNaN(l))throw new Error("beta must not be NaN.");if(!Number.isFinite(l))throw new Error("beta must be finite.");if(!Number.isInteger(o)||!Number.isInteger(a)||!Number.isInteger(u)||!Number.isInteger(f)||!Number.isInteger(p))throw new Error("m, n, lda, ldb, and ldc must be integers.");if(!g&&!(s instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!w&&!(n instanceof Float32Array))throw new Error("B must be a Float32Array or GpuMatrix.");if(!h&&!(m instanceof Float32Array))throw new Error("C must be a Float32Array or GpuMatrix.");if((g||w)&&!h)throw new Error("C must be a GpuMatrix when A or B is a GpuMatrix.");if(h&&(!g||!w))throw new Error("A and B must be GpuMatrix when C is a GpuMatrix.");if(o<0||a<0)throw new Error("m and n must be non-negative.");if(o===0||a===0)return h?{}:{C:m};let b=g?s.layout:c,y=w?n.layout:c,v=h?m.layout:c,_=t==="left"?o:a;if(u<_)throw new Error("lda must be >= "+(t==="left"?"m":"n")+".");if(g){if(u!==s.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(s.rows<_||s.cols<_)throw new Error("A is too small for the given m/n and side.")}else if(s.length<(_-1)*u+_)throw new Error("A does not have enough elements for the given dimensions and lda.");let A=y==="column-major"?a:o,k=y==="column-major"?o:a;if(f<k)throw new Error(`ldb must be >= ${y==="column-major"?"rows":"cols"} of B as stored.`);if(w){if(f!==n.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");if(n.rows<o||n.cols<a)throw new Error("B is too small for the given m and n.")}else if(n.length<(A-1)*f+k)throw new Error("B does not have enough elements for the given dimensions and ldb.");let B=v==="column-major"?a:o,L=v==="column-major"?o:a;if(p<L)throw new Error(`ldc must be >= ${v==="column-major"?"rows":"cols"} of C as stored.`);if(h){if(p!==m.lda)throw new Error("ldc must match C.lda when C is a GpuMatrix.");if(m.rows<o||m.cols<a)throw new Error("C is too small for the given m and n.")}else if(m.length<(B-1)*p+L)throw new Error("C does not have enough elements for the given dimensions and ldc.");let C=b==="column-major"?e==="lower"?"upper":"lower":e,R=y==="column-major"?"transpose":"no-transpose",F="no-transpose",W=o,V=a,z=_,H=t==="left"?F:R,$=t==="left"?R:F,K=ur=>ur==="no-transpose"?"transpose":"no-transpose",Y=t==="right";v==="column-major"&&([H,$]=[K($),K(H)],Y=!Y,[W,V]=[V,W]);let ir=_,lr=Math.ceil(V/64),hr=Math.ceil(W/64),nr=lr*hr>=36,mr=await D(r,nr?"sgemm_large":"sgemm_small"),Q=await D(r,"symmetrize"),er=nr?{x:U(r,lr,"ssymm","x"),y:U(r,hr,"ssymm","y")}:{x:U(r,Math.ceil(V/32),"ssymm","x"),y:U(r,Math.ceil(W/32),"ssymm","y")},Z=g?s._buf:x(r,s,"ssymm-A",!1),J=w?n._buf:x(r,n,"ssymm-B",!1),sr=h?m._buf:x(r,m,"ssymm-C",!0),pr=tr(r,_*ir*4,"ssymm-Adense"),br=null,fr=null;try{br=I(r,[{value:_,type:"u32"},{value:u,type:"u32"},{value:ir,type:"u32"},{value:C==="upper"?1:0,type:"u32"}],"ssymm-sym-params");let ur=E(r,Q.getBindGroupLayout(0),[Z,pr,br]),yr=Y?J:pr,Cr=Y?f:ir,Rr=Y?pr:J;fr=I(r,[{value:W,type:"u32"},{value:V,type:"u32"},{value:z,type:"u32"},{value:i,type:"f32"},{value:l,type:"f32"},{value:Cr,type:"u32"},{value:Y?ir:f,type:"u32"},{value:p,type:"u32"},{value:H==="transpose"?1:0,type:"u32"},{value:$==="transpose"?1:0,type:"u32"}],"ssymm-gemm-params");let Er=E(r,mr.getBindGroupLayout(0),[yr,Sr(r,yr),Rr,Sr(r,Rr),sr,fr]),{commandEncoder:vr,querySet:xr}=qr(r),Ir=xr?{timestampWrites:{querySet:xr,beginningOfPassWriteIndex:0}}:void 0,Ar=xr?{timestampWrites:{querySet:xr,endOfPassWriteIndex:1}}:void 0;gr(vr,Q,ur,{x:Math.ceil(_/8),y:Math.ceil(_/8)},Ir),gr(vr,mr,Er,er,Ar);let Fr=Lr(r,vr,xr),Or=h?null:G(r,vr,sr);M(r,vr);let Ur=await P(Fr);if(h)return Ur!==void 0?{gpuTimeMs:Ur}:{};let me=await S(Or,Float32Array);return Ur!==void 0?{C:me,gpuTimeMs:Ur}:{C:me}}finally{g||d(Z),w||d(J),h||d(sr),d(pr),br&&d(br),fr&&d(fr)}}async function ha(r,t,e,o,a,i,s,u,n,f,l,m,p="row-major"){let c=n instanceof X,g=l instanceof X,w=a==="unit";if(q(r),T(r,"strmm",{A:n,B:l}),t!=="left"&&t!=="right")throw new Error("side must be 'left' or 'right'.");if(e!=="lower"&&e!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(o!=="no-transpose"&&o!=="transpose")throw new Error("transA must be 'no-transpose' or 'transpose'.");if(!w&&a!=="non-unit")throw new Error("diag must be 'unit' or 'non-unit'.");if(p!=="row-major"&&p!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(typeof u!="number")throw new Error("alpha must be a number.");if(Number.isNaN(u))throw new Error("alpha must not be NaN.");if(!Number.isFinite(u))throw new Error("alpha must be finite.");if(!Number.isInteger(i)||!Number.isInteger(s)||!Number.isInteger(f)||!Number.isInteger(m))throw new Error("m, n, lda, and ldb must be integers.");if(!c&&!(n instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!g&&!(l instanceof Float32Array))throw new Error("B must be a Float32Array or GpuMatrix.");if(c!==g)throw new Error("A and B must both be GpuMatrix or both be Float32Array.");if(i<0||s<0)throw new Error("m and n must be non-negative.");if(i===0||s===0)return g?{}:{B:l};let h=c?n.layout:p,b=g?l.layout:p,y=t==="left"?i:s;if(f<y)throw new Error("lda must be >= "+(t==="left"?"m":"n")+".");if(c){if(f!==n.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(n.rows<y||n.cols<y)throw new Error("A is too small for the given m/n and side.")}else if(n.length<(y-1)*f+y)throw new Error("A does not have enough elements for the given dimensions and lda.");let v=b==="column-major"?s:i,_=b==="column-major"?i:s;if(m<_)throw new Error(`ldb must be >= ${b==="column-major"?"rows":"cols"} of B as stored.`);if(g){if(m!==l.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");if(l.rows<i||l.cols<s)throw new Error("B is too small for the given m and n.")}else if(l.length<(v-1)*m+_)throw new Error("B does not have enough elements for the given dimensions and ldb.");let A=h==="column-major"?e==="lower"?"upper":"lower":e,k=h==="column-major"?o==="no-transpose"?"transpose":"no-transpose":o,B=b==="column-major"?"transpose":"no-transpose",L="no-transpose",C=i,R=s,F=y,W=t==="left"?L:B,V=t==="left"?B:L,z=br=>br==="no-transpose"?"transpose":"no-transpose",H=t==="right";b==="column-major"&&([W,V]=[z(V),z(W)],H=!H,[C,R]=[R,C]);let $=y,K=Math.ceil(R/64),Y=Math.ceil(C/64),ir=K*Y>=36,lr=await D(r,ir?"sgemm_large":"sgemm_small"),hr=await D(r,"triangularize"),nr=ir?{x:U(r,K,"strmm","x"),y:U(r,Y,"strmm","y")}:{x:U(r,Math.ceil(R/32),"strmm","x"),y:U(r,Math.ceil(C/32),"strmm","y")},mr=null,Q=null,er=null,Z=null,J=null,sr=null,pr=!1;try{mr=c?n._buf:x(r,n,"strmm-A",!1),Q=g?l._buf:x(r,l,"strmm-B",!0),er=tr(r,y*$*4,"strmm-Adense"),Z=tr(r,v*m*4,"strmm-out",GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST),J=I(r,[{value:y,type:"u32"},{value:f,type:"u32"},{value:$,type:"u32"},{value:A==="upper"?1:0,type:"u32"},{value:k==="transpose"?1:0,type:"u32"},{value:w?1:0,type:"u32"}],"strmm-tri-params");let br=E(r,hr.getBindGroupLayout(0),[mr,er,J]),fr=H?Q:er,ur=H?m:$,yr=H?er:Q;sr=I(r,[{value:C,type:"u32"},{value:R,type:"u32"},{value:F,type:"u32"},{value:u,type:"f32"},{value:0,type:"f32"},{value:ur,type:"u32"},{value:H?$:m,type:"u32"},{value:m,type:"u32"},{value:W==="transpose"?1:0,type:"u32"},{value:V==="transpose"?1:0,type:"u32"}],"strmm-gemm-params");let Rr=E(r,lr.getBindGroupLayout(0),[fr,Sr(r,fr),yr,Sr(r,yr),Z,sr]),{commandEncoder:Tr,querySet:Er}=qr(r);Tr.copyBufferToBuffer(Q,0,Z,0,Math.min(Q.size,Z.size));let vr=Er?{timestampWrites:{querySet:Er,beginningOfPassWriteIndex:0}}:void 0,xr=Er?{timestampWrites:{querySet:Er,endOfPassWriteIndex:1}}:void 0;gr(Tr,hr,br,{x:Math.ceil(y/8),y:Math.ceil(y/8)},vr),gr(Tr,lr,Rr,nr,xr);let Ir=Lr(r,Tr,Er),Ar=g?null:G(r,Tr,Z);M(r,Tr);let Fr=await P(Ir);if(g)return d(l._buf),l._buf=Z,pr=!0,Fr!==void 0?{gpuTimeMs:Fr}:{};let Or=await S(Ar,Float32Array);return Fr!==void 0?{B:Or,gpuTimeMs:Fr}:{B:Or}}finally{!c&&mr&&d(mr),!g&&Q&&d(Q),er&&d(er),Z&&!pr&&d(Z),J&&d(J),sr&&d(sr)}}async function ba(r,t,e,o,a,i,s,u,n,f,l,m,p="row-major"){let c=n instanceof X,g=l instanceof X,w=a==="unit";if(q(r),T(r,"strsm",{A:n,B:l}),t!=="left"&&t!=="right")throw new Error("side must be 'left' or 'right'.");if(e!=="lower"&&e!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(o!=="no-transpose"&&o!=="transpose")throw new Error("transA must be 'no-transpose' or 'transpose'.");if(!w&&a!=="non-unit")throw new Error("diag must be 'unit' or 'non-unit'.");if(p!=="row-major"&&p!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(typeof u!="number")throw new Error("alpha must be a number.");if(Number.isNaN(u))throw new Error("alpha must not be NaN.");if(!Number.isFinite(u))throw new Error("alpha must be finite.");if(!Number.isInteger(i)||!Number.isInteger(s)||!Number.isInteger(f)||!Number.isInteger(m))throw new Error("m, n, lda, and ldb must be integers.");if(!c&&!(n instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!g&&!(l instanceof Float32Array))throw new Error("B must be a Float32Array or GpuMatrix.");if(c!==g)throw new Error("A and B must both be GpuMatrix or both be Float32Array.");if(i<0||s<0)throw new Error("m and n must be non-negative.");if(i===0||s===0)return g?{}:{B:l};let h=c?n.layout:p,b=g?l.layout:p,y=t==="left"?i:s;if(f<y)throw new Error("lda must be >= "+(t==="left"?"m":"n")+".");if(c){if(f!==n.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(n.rows<y||n.cols<y)throw new Error("A is too small for the given m/n and side.")}else if(n.length<(y-1)*f+y)throw new Error("A does not have enough elements for the given dimensions and lda.");let v=b==="column-major"?s:i,_=b==="column-major"?i:s;if(m<_)throw new Error(`ldb must be >= ${b==="column-major"?"rows":"cols"} of B as stored.`);if(g){if(m!==l.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");if(l.rows<i||l.cols<s)throw new Error("B is too small for the given m and n.")}else if(l.length<(v-1)*m+_)throw new Error("B does not have enough elements for the given dimensions and ldb.");let A=h==="column-major"?e==="lower"?"upper":"lower":e,k=h==="column-major"?o==="no-transpose"?"transpose":"no-transpose":o,B=t==="left"?s:i,L=t==="left",C=k==="no-transpose"==(A==="lower"),R=t==="left"?C:!C,F=[];for(let Q=0;Q<y;Q+=64)F.push(Q);R||F.reverse();let W=F.length,V=await D(r,"strsv_invert_block"),z=await D(r,"block_transfer"),H=await D(r,"sscal"),$=null,K=null,Y=null,ir=[],lr=[];function hr(Q,er){let Z=tr(r,Q,er);return lr.push(Z),Z}function nr(Q,er){let Z=I(r,Q,er);return ir.push(Z),Z}let mr=(v-1)*m+_;try{$=c?n._buf:x(r,n,"strsm-A",!1),K=g?l._buf:x(r,l,"strsm-B",!0),Y=tr(r,W*64*64*4,"strsm-Ainv");let Q=null;if(u!==1&&u!==0){let Er=nr([{value:mr,type:"u32"},{value:u,type:"f32"},{value:1,type:"u32"}],"strsm-scale-params");Q=E(r,H.getBindGroupLayout(0),[K,Er])}let er=nr([{value:y,type:"u32"},{value:f,type:"u32"},{value:k==="transpose"?1:0,type:"u32"},{value:A==="upper"?1:0,type:"u32"},{value:w?1:0,type:"u32"}],"strsm-invert-params"),Z=E(r,V.getBindGroupLayout(0),[$,Y,er]),J=hr(64*B*4,"strsm-Bblock"),sr=hr(64*B*4,"strsm-Xblock"),pr=hr(y*64*4,"strsm-Aoff"),br=hr(y*B*4,"strsm-delta"),{commandEncoder:fr,querySet:ur}=qr(r);if(u===0){let Er=Math.ceil(_/64),vr=Math.ceil(v/64),xr=Er*vr>=36,Ir=await D(r,xr?"sgemm_large":"sgemm_small"),Ar=nr([{value:v,type:"u32"},{value:_,type:"u32"},{value:0,type:"u32"},{value:0,type:"f32"},{value:0,type:"f32"},{value:1,type:"u32"},{value:1,type:"u32"},{value:m,type:"u32"},{value:0,type:"u32"},{value:0,type:"u32"}],"strsm-zero-params"),Fr=E(r,Ir.getBindGroupLayout(0),[Y,Sr(r,Y),Y,Sr(r,Y),K,Ar]),Or=xr?{x:U(r,Er,"strsm","x"),y:U(r,vr,"strsm","y")}:{x:U(r,Math.ceil(_/32),"strsm","x"),y:U(r,Math.ceil(v/32),"strsm","y")};gr(fr,Ir,Fr,Or,ur?{timestampWrites:{querySet:ur,beginningOfPassWriteIndex:0,endOfPassWriteIndex:1}}:void 0)}else{Q&&gr(fr,H,Q,cr(r,mr)),gr(fr,V,Z,{x:64,y:W},ur?{timestampWrites:{querySet:ur,beginningOfPassWriteIndex:0}}:void 0);for(let vr=0;vr<F.length;vr++){let xr=F[vr],Ir=Math.min(xr+64,y),Ar=Ir-xr,Fr=xr/64,Or=vr===F.length-1,Ur=nr([{value:xr,type:"u32"},{value:Ar,type:"u32"},{value:0,type:"u32"},{value:B,type:"u32"},{value:m,type:"u32"},{value:b==="column-major"?1:0,type:"u32"},{value:L?1:0,type:"u32"},{value:2,type:"u32"}],"strsm-gather-B-params"),me=E(r,z.getBindGroupLayout(0),[J,K,Ur]);gr(fr,z,me,Zr(r,"strsm",Ar,B));{let Yr=Ar,Xr=B,_e=Ar,ae=Math.ceil(Xr/64),ie=Math.ceil(Yr/64),se=ae*ie>=36,ne=await D(r,se?"sgemm_large":"sgemm_small"),Be=nr([{value:Yr,type:"u32"},{value:Xr,type:"u32"},{value:_e,type:"u32"},{value:1,type:"f32"},{value:0,type:"f32"},{value:64,type:"u32"},{value:B,type:"u32"},{value:B,type:"u32"},{value:t==="right"?1:0,type:"u32"},{value:0,type:"u32"}],"strsm-apply-params"),ce={buffer:Y,offset:Fr*64*64*4,size:4096*4},Ae=E(r,ne.getBindGroupLayout(0),[ce,Sr(r,ce),J,Sr(r,J),sr,Be]),Ea=se?{x:U(r,ae,"strsm","x"),y:U(r,ie,"strsm","y")}:{x:U(r,Math.ceil(Xr/32),"strsm","x"),y:U(r,Math.ceil(Yr/32),"strsm","y")};gr(fr,ne,Ae,Ea)}let de=R?Ir:0,Le=R?y:xr,Re=de<Le,ya=nr([{value:xr,type:"u32"},{value:Ar,type:"u32"},{value:0,type:"u32"},{value:B,type:"u32"},{value:m,type:"u32"},{value:b==="column-major"?1:0,type:"u32"},{value:L?1:0,type:"u32"},{value:0,type:"u32"}],"strsm-scatter-params"),xa=E(r,z.getBindGroupLayout(0),[sr,K,ya]),va=Or&&!Re&&ur?{timestampWrites:{querySet:ur,endOfPassWriteIndex:1}}:void 0;if(gr(fr,z,xa,Zr(r,"strsm",Ar,B),va),!Re)continue;let oe=Le-de,_a=nr([{value:de,type:"u32"},{value:oe,type:"u32"},{value:xr,type:"u32"},{value:Ar,type:"u32"},{value:f,type:"u32"},{value:k==="transpose"?1:0,type:"u32"},{value:L?1:0,type:"u32"},{value:2,type:"u32"}],"strsm-gather-A-params"),Ba=E(r,z.getBindGroupLayout(0),[pr,$,_a]);gr(fr,z,Ba,Zr(r,"strsm",oe,Ar));{let Yr=oe,Xr=B,_e=Ar,ae=Math.ceil(Xr/64),ie=Math.ceil(Yr/64),se=ae*ie>=36,ne=await D(r,se?"sgemm_large":"sgemm_small"),Be=nr([{value:Yr,type:"u32"},{value:Xr,type:"u32"},{value:_e,type:"u32"},{value:1,type:"f32"},{value:0,type:"f32"},{value:Ar,type:"u32"},{value:B,type:"u32"},{value:B,type:"u32"},{value:0,type:"u32"},{value:0,type:"u32"}],"strsm-update-params"),ce=E(r,ne.getBindGroupLayout(0),[pr,Sr(r,pr),sr,Sr(r,sr),br,Be]),Ae=se?{x:U(r,ae,"strsm","x"),y:U(r,ie,"strsm","y")}:{x:U(r,Math.ceil(Xr/32),"strsm","x"),y:U(r,Math.ceil(Yr/32),"strsm","y")};gr(fr,ne,ce,Ae)}let Aa=nr([{value:de,type:"u32"},{value:oe,type:"u32"},{value:0,type:"u32"},{value:B,type:"u32"},{value:m,type:"u32"},{value:b==="column-major"?1:0,type:"u32"},{value:L?1:0,type:"u32"},{value:1,type:"u32"}],"strsm-scatter-sub-params"),Sa=E(r,z.getBindGroupLayout(0),[br,K,Aa]),Ga=Or&&ur?{timestampWrites:{querySet:ur,endOfPassWriteIndex:1}}:void 0;gr(fr,z,Sa,Zr(r,"strsm",oe,B),Ga)}}let yr=Lr(r,fr,ur),Cr=g?null:G(r,fr,K);M(r,fr);let Rr=await P(yr);if(g)return Rr!==void 0?{gpuTimeMs:Rr}:{};let Tr=await S(Cr,Float32Array);return Rr!==void 0?{B:Tr,gpuTimeMs:Rr}:{B:Tr}}finally{!c&&$&&d($),!g&&K&&d(K),Y&&d(Y),d(lr),d(ir)}}return Ia(Ti);})();
+`)}`);let p=e==="main"?{module:f}:{module:f,entryPoint:e},c=r.createComputePipeline({label:o,layout:"auto",compute:p});return c._shaderModule=f,c}function cr(r,t,e){let o=r.limits.maxComputeWorkgroupsPerDimension;return e===void 0?Math.min(Math.ceil(t/64),o):{x:Math.min(Math.ceil(e/8),o),y:Math.min(Math.ceil(t/8),o)}}function U(r,t,e,o="x"){let a=r.limits.maxComputeWorkgroupsPerDimension;if(t>a)throw new Error(`${e}: this problem needs ${t} workgroups in ${o}, but the device allows ${a} (maxComputeWorkgroupsPerDimension). The operands are too large for this device \u2014 split the operation into smaller blocks.`);return t}function Zr(r,t,e,o){return o===void 0?U(r,Math.ceil(e/64),t):{x:U(r,Math.ceil(o/8),t,"x"),y:U(r,Math.ceil(e/8),t,"y")}}function q(r){if(!(r instanceof GPUDevice))throw new Error("device must be a GPUDevice.")}function T(r,t,e){for(let[o,a]of Object.entries(e))if(!(!(a instanceof N)&&!(a instanceof X))&&a.device!==r)throw new Error(`${t}: ${o} belongs to a different GPUDevice than the one passed in. GPU buffers cannot be shared across devices \u2014 recreate the operand on this device, or call the routine with the device that owns it.`)}async function qo(r,t,e,o,a){let i=o instanceof N;if(q(r),T(r,"sscal",{x:o}),!Number.isInteger(t)||!Number.isInteger(a))throw new Error("n and incx must be integers.");if(typeof e!="number")throw new Error("alpha must be a number.");if(Number.isNaN(e))throw new Error("alpha must not be NaN.");if(!Number.isFinite(e))throw new Error("alpha must be finite.");if(a<=0)throw new Error("incx must be positive.");if(!(o instanceof Float32Array)&&!(o instanceof N))throw new Error("x must be a Float32Array or GpuVector.");if(t<=0)return i?{}:{x:o};if(o.length<(t-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");let s=await D(r,"sscal"),u=null,n=null,f=null;try{u=i?o._buf:x(r,o,"sscal-x",!0),n=I(r,[{value:t,type:"u32"},{value:e,type:"f32"},{value:a,type:"u32"}],"sscal-params");let l=E(r,s.getBindGroupLayout(0),[u,n]),{commandEncoder:m,ts:p}=j(r,s,l,cr(r,t));f=i?null:G(r,m,u),M(r,m);let c=await P(p);if(i)return c!==void 0?{gpuTimeMs:c}:{};let g=await A(f,Float32Array);return f=null,c!==void 0?{x:g,gpuTimeMs:c}:{x:g}}finally{!i&&u&&d(u),n&&d(n),f&&d(f)}}async function To(r,t,e,o,a){let i=o instanceof N;if(q(r),T(r,"cscal",{x:o}),!Number.isInteger(t)||!Number.isInteger(a))throw new Error("n and incx must be integers.");if(!(e instanceof Wr))throw new Error("alpha must be a Complex32.");if(Number.isNaN(e.re)||Number.isNaN(e.im))throw new Error("alpha must not be NaN.");if(!Number.isFinite(e.re)||!Number.isFinite(e.im))throw new Error("alpha must be finite.");if(a<=0)throw new Error("incx must be positive.");if(!(o instanceof _r)&&!i)throw new Error("x must be a Complex32Array or GpuVector.");if(i&&o.dtype!==_r)throw new Error("x must be a Complex32Array-backed GpuVector.");if(t<=0)return i?{}:{x:o};if(o.length<(t-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");let s=await D(r,"cscal"),u=null,n=null,f=null;try{u=i?o._buf:x(r,te(o),"cscal-x",!0),n=I(r,[{value:t,type:"u32"},{value:e.re,type:"f32"},{value:e.im,type:"f32"},{value:a,type:"u32"}],"cscal-params");let l=E(r,s.getBindGroupLayout(0),[u,n]),{commandEncoder:m,ts:p}=j(r,s,l,cr(r,t));f=i?null:G(r,m,u),M(r,m);let c=await P(p);if(i)return c!==void 0?{gpuTimeMs:c}:{};let g=await A(f,Float32Array);f=null;let w=new _r(g);return c!==void 0?{x:w,gpuTimeMs:c}:{x:w}}finally{!i&&u&&d(u),n&&d(n),f&&d(f)}}async function Co(r,t,e,o,a){let i=o instanceof N;if(q(r),!Number.isInteger(t)||!Number.isInteger(a))throw new Error("n and incx must be integers.");if(typeof e!="number")throw new Error("alpha must be a number.");if(Number.isNaN(e))throw new Error("alpha must not be NaN.");if(!Number.isFinite(e))throw new Error("alpha must be finite.");if(!(o instanceof Float64Array)&&!i)throw new Error("x must be a Float64Array or GpuVector.");if(i&&o.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(a<=0)throw new Error("incx must be positive.");if(T(r,"dscal",{x:o}),t<=0)return i?{}:{x:o};if(o.length<(t-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");let u=await D(r,[...["f64/dekker","f64/utils/add","f64/utils/multiply"],"dscal"]),{hi:n,lo:f}=rr(new Float64Array([e])),l=null,m=null,p=null,c=null,g=null;try{if(i)l=o._buf,m=o._loBuf;else{let{hi:k,lo:B}=rr(o);l=x(r,k,"dscal-xHi",!0),m=x(r,B,"dscal-xLo",!0)}p=I(r,[{value:t,type:"u32"},{value:n[0],type:"f32"},{value:f[0],type:"f32"},{value:a,type:"u32"}],"dscal-params");let w=E(r,u.getBindGroupLayout(0),[l,m,p]),{commandEncoder:h,ts:b}=j(r,u,w,cr(r,t));c=i?null:G(r,h,l),g=i?null:G(r,h,m),M(r,h);let y=await P(b);if(i)return y!==void 0?{gpuTimeMs:y}:{};let v=await A(c,Float32Array);c=null;let _=await A(g,Float32Array);g=null;let S=dr(v,_);return y!==void 0?{x:S,gpuTimeMs:y}:{x:S}}finally{!i&&l&&d(l),!i&&m&&d(m),p&&d(p),c&&d(c),g&&d(g)}}async function Fo(r,t,e,o,a,i){let s=e instanceof N,u=a instanceof N;if(q(r),T(r,"sswap",{x:e,y:a}),!Number.isInteger(t)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!(e instanceof Float32Array)&&!(e instanceof N))throw new Error("x must be a Float32Array or GpuVector.");if(!(a instanceof Float32Array)&&!(a instanceof N))throw new Error("y must be a Float32Array or GpuVector.");if(e.constructor!==a.constructor)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(t<=0)return s?{}:{x:e,y:a};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(a.length<(t-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=await D(r,"sswap"),f=null,l=null,m=null,p=null,c=null;try{f=s?e._buf:x(r,e,"sswap-x",!0),l=u?a._buf:x(r,a,"sswap-y",!0),m=I(r,[{value:t,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"sswap-params");let g=E(r,n.getBindGroupLayout(0),[f,l,m]),{commandEncoder:w,ts:h}=j(r,n,g,cr(r,t));p=s?null:G(r,w,f),c=u?null:G(r,w,l),M(r,w);let b=await P(h);if(s)return b!==void 0?{gpuTimeMs:b}:{};let y=await A(p,Float32Array);p=null;let v=await A(c,Float32Array);return c=null,b!==void 0?{x:y,y:v,gpuTimeMs:b}:{x:y,y:v}}finally{!s&&f&&d(f),!u&&l&&d(l),m&&d(m),p&&d(p),c&&d(c)}}async function jo(r,t,e,o,a,i){let s=e instanceof N,u=a instanceof N;if(q(r),T(r,"dswap",{x:e,y:a}),!Number.isInteger(t)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!(e instanceof Float64Array)&&!s)throw new Error("x must be a Float64Array or GpuVector.");if(!(a instanceof Float64Array)&&!u)throw new Error("y must be a Float64Array or GpuVector.");if(s&&e.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(u&&a.dtype!==Float64Array)throw new Error("y must be a Float64Array-backed GpuVector.");if(s!==u)throw new Error("x and y must be the same type (both Float64Array or both GpuVector).");if(t<=0)return s?{}:{x:e,y:a};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(a.length<(t-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=await D(r,"dswap"),f=null,l=null,m=null,p=null,c=null,g=null,w=null,h=null,b=null;try{if(s)f=e._buf,l=e._loBuf,m=a._buf,p=a._loBuf;else{let W=rr(e),V=rr(a);f=x(r,W.hi,"dswap-xHi",!0),l=x(r,W.lo,"dswap-xLo",!0),m=x(r,V.hi,"dswap-yHi",!0),p=x(r,V.lo,"dswap-yLo",!0)}c=I(r,[{value:t,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"dswap-params");let y=E(r,n.getBindGroupLayout(0),[f,l,m,p,c]),{commandEncoder:v,ts:_}=j(r,n,y,cr(r,t));g=s?null:G(r,v,f),w=s?null:G(r,v,l),h=u?null:G(r,v,m),b=u?null:G(r,v,p),M(r,v);let S=await P(_);if(s)return S!==void 0?{gpuTimeMs:S}:{};let k=await A(g,Float32Array);g=null;let B=await A(w,Float32Array);w=null;let L=await A(h,Float32Array);h=null;let C=await A(b,Float32Array);b=null;let R=dr(k,B),F=dr(L,C);return S!==void 0?{x:R,y:F,gpuTimeMs:S}:{x:R,y:F}}finally{!s&&f&&d(f),!s&&l&&d(l),!u&&m&&d(m),!u&&p&&d(p),c&&d(c),g&&d(g),w&&d(w),h&&d(h),b&&d(b)}}async function Wo(r,t,e,o,a,i,s){let u=o instanceof N,n=i instanceof N;if(q(r),T(r,"saxpy",{x:o,y:i}),!Number.isInteger(t)||!Number.isInteger(a)||!Number.isInteger(s))throw new Error("n, incx, and incy must be integers.");if(typeof e!="number")throw new Error("alpha must be a number.");if(Number.isNaN(e))throw new Error("alpha must not be NaN.");if(!Number.isFinite(e))throw new Error("alpha must be finite.");if(a<=0||s<=0)throw new Error("incx and incy must be positive.");if(!u&&!(o instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!n&&!(i instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(u!==n)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(t<=0)return n?{}:{y:i};if(o.length<(t-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");if(i.length<(t-1)*s+1)throw new Error("y does not have enough elements for the given n and incy.");let f=await D(r,"saxpy"),l=null,m=null,p=null,c=null;try{l=u?o._buf:x(r,o,"saxpy-x",!1),m=n?i._buf:x(r,i,"saxpy-y",!0),p=I(r,[{value:t,type:"u32"},{value:e,type:"f32"},{value:a,type:"u32"},{value:s,type:"u32"}],"saxpy-params");let g=E(r,f.getBindGroupLayout(0),[l,m,p]),{commandEncoder:w,ts:h}=j(r,f,g,cr(r,t));c=n?null:G(r,w,m),M(r,w);let b=await P(h);if(n)return b!==void 0?{gpuTimeMs:b}:{};let y=await A(c,Float32Array);return c=null,b!==void 0?{y,gpuTimeMs:b}:{y}}finally{!u&&l&&d(l),!n&&m&&d(m),p&&d(p),c&&d(c)}}async function Ho(r,t,e,o,a,i,s){let u=o instanceof N,n=i instanceof N;if(q(r),!Number.isInteger(t)||!Number.isInteger(a)||!Number.isInteger(s))throw new Error("n, incx, and incy must be integers.");if(typeof e!="number")throw new Error("alpha must be a number.");if(Number.isNaN(e))throw new Error("alpha must not be NaN.");if(!Number.isFinite(e))throw new Error("alpha must be finite.");if(!(o instanceof Float64Array)&&!u)throw new Error("x must be a Float64Array or GpuVector.");if(!(i instanceof Float64Array)&&!n)throw new Error("y must be a Float64Array or GpuVector.");if(u&&o.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(n&&i.dtype!==Float64Array)throw new Error("y must be a Float64Array-backed GpuVector.");if(u!==n)throw new Error("x and y must be the same type (both Float64Array or both GpuVector).");if(a<=0||s<=0)throw new Error("incx and incy must be positive.");if(T(r,"daxpy",{x:o,y:i}),t<=0)return n?{}:{y:i};if(o.length<(t-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");if(i.length<(t-1)*s+1)throw new Error("y does not have enough elements for the given n and incy.");let l=await D(r,[...["f64/dekker","f64/utils/add","f64/utils/multiply"],"daxpy"]),{hi:m,lo:p}=rr(new Float64Array([e])),c=null,g=null,w=null,h=null,b=null,y=null,v=null;try{if(u)c=o._buf,g=o._loBuf,w=i._buf,h=i._loBuf;else{let F=rr(o),W=rr(i);c=x(r,F.hi,"daxpy-xHi",!1),g=x(r,F.lo,"daxpy-xLo",!1),w=x(r,W.hi,"daxpy-yHi",!0),h=x(r,W.lo,"daxpy-yLo",!0)}b=I(r,[{value:t,type:"u32"},{value:m[0],type:"f32"},{value:p[0],type:"f32"},{value:a,type:"u32"},{value:s,type:"u32"}],"daxpy-params");let _=E(r,l.getBindGroupLayout(0),[c,g,w,h,b]),{commandEncoder:S,ts:k}=j(r,l,_,cr(r,t));y=n?null:G(r,S,w),v=n?null:G(r,S,h),M(r,S);let B=await P(k);if(n)return B!==void 0?{gpuTimeMs:B}:{};let L=await A(y,Float32Array);y=null;let C=await A(v,Float32Array);v=null;let R=dr(L,C);return B!==void 0?{y:R,gpuTimeMs:B}:{y:R}}finally{!u&&c&&d(c),!u&&g&&d(g),!n&&w&&d(w),!n&&h&&d(h),b&&d(b),y&&d(y),v&&d(v)}}async function Oo(r,t,e,o,a,i){let s=e instanceof N,u=a instanceof N;if(q(r),T(r,"scopy",{x:e,y:a}),!Number.isInteger(t)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!s&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!u&&!(a instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(s!==u)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(t<=0)return u?{}:{y:a};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(a.length<(t-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=await D(r,"scopy"),f=null,l=null,m=null,p=null;try{f=s?e._buf:x(r,e,"scopy-x",!1),l=u?a._buf:x(r,a,"scopy-y",!0),m=I(r,[{value:t,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"scopy-params");let c=E(r,n.getBindGroupLayout(0),[f,l,m]),{commandEncoder:g,ts:w}=j(r,n,c,cr(r,t));p=u?null:G(r,g,l),M(r,g);let h=await P(w);if(u)return h!==void 0?{gpuTimeMs:h}:{};let b=await A(p,Float32Array);return p=null,h!==void 0?{y:b,gpuTimeMs:h}:{y:b}}finally{!s&&f&&d(f),!u&&l&&d(l),m&&d(m),p&&d(p)}}async function Vo(r,t,e,o,a,i){let s=e instanceof N,u=a instanceof N;if(q(r),T(r,"dcopy",{x:e,y:a}),!Number.isInteger(t)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!(e instanceof Float64Array)&&!s)throw new Error("x must be a Float64Array or GpuVector.");if(!(a instanceof Float64Array)&&!u)throw new Error("y must be a Float64Array or GpuVector.");if(s&&e.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(u&&a.dtype!==Float64Array)throw new Error("y must be a Float64Array-backed GpuVector.");if(s!==u)throw new Error("x and y must be the same type (both Float64Array or both GpuVector).");if(t<=0)return u?{}:{y:a};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(a.length<(t-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=await D(r,"dcopy"),f=null,l=null,m=null,p=null,c=null,g=null,w=null;try{if(s)f=e._buf,l=e._loBuf,m=a._buf,p=a._loBuf;else{let B=rr(e),L=rr(a);f=x(r,B.hi,"dcopy-xHi",!1),l=x(r,B.lo,"dcopy-xLo",!1),m=x(r,L.hi,"dcopy-yHi",!0),p=x(r,L.lo,"dcopy-yLo",!0)}c=I(r,[{value:t,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"dcopy-params");let h=E(r,n.getBindGroupLayout(0),[f,l,m,p,c]),{commandEncoder:b,ts:y}=j(r,n,h,cr(r,t));g=u?null:G(r,b,m),w=u?null:G(r,b,p),M(r,b);let v=await P(y);if(u)return v!==void 0?{gpuTimeMs:v}:{};let _=await A(g,Float32Array);g=null;let S=await A(w,Float32Array);w=null;let k=dr(_,S);return v!==void 0?{y:k,gpuTimeMs:v}:{y:k}}finally{!s&&f&&d(f),!s&&l&&d(l),!u&&m&&d(m),!u&&p&&d(p),c&&d(c),g&&d(g),w&&d(w)}}async function Ko(r,t,e,o,a,i){let s=e instanceof N,u=a instanceof N;if(q(r),T(r,"sdot",{x:e,y:a}),!Number.isInteger(t)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!s&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!u&&!(a instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(s!==u)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(t<=0)return{dot:0};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(a.length<(t-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=await D(r,"sdot"),f=await D(r,"reduction/sum"),l=null,m=null,p=null,c=null,g=null,w=null;try{l=s?e._buf:x(r,e,"sdot-x",!1),m=u?a._buf:x(r,a,"sdot-y",!1),p=tr(r,512,"sdot-partials"),c=Br(r,4,"sdot-result"),g=I(r,[{value:t,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"sdot-params");let h=E(r,n.getBindGroupLayout(0),[l,m,p,g]),{commandEncoder:b,ts:y}=j(r,n,h,128);M(r,b);let v=E(r,f.getBindGroupLayout(0),[p,c]),{commandEncoder:_,ts:S}=j(r,f,v,1);w=G(r,_,c),M(r,_);let k=A(w,Float32Array);w=null;let[B,L,C]=await Promise.all([P(y),P(S),k]);return B!==void 0&&L!==void 0?{dot:C[0],gpuTimeMs:B+L}:{dot:C[0]}}finally{!s&&l&&d(l),!u&&m&&d(m),p&&d(p),c&&d(c),g&&d(g),w&&d(w)}}async function zo(r,t,e,o){let a=e instanceof N;if(q(r),T(r,"sasum",{x:e}),!Number.isInteger(t)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!a&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(t<=0)return{asum:0};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=await D(r,"sasum"),s=await D(r,"reduction/sum"),u=null,n=null,f=null,l=null,m=null;try{u=a?e._buf:x(r,e,"sasum-x",!1),n=tr(r,512,"sasum-partials"),f=Br(r,4,"sasum-result"),l=I(r,[{value:t,type:"u32"},{value:o,type:"u32"}],"sasum-params");let p=E(r,i.getBindGroupLayout(0),[u,n,l]),{commandEncoder:c,ts:g}=j(r,i,p,128);M(r,c);let w=E(r,s.getBindGroupLayout(0),[n,f]),{commandEncoder:h,ts:b}=j(r,s,w,1);m=G(r,h,f),M(r,h);let y=A(m,Float32Array);m=null;let[v,_,S]=await Promise.all([P(g),P(b),y]);return v!==void 0&&_!==void 0?{asum:S[0],gpuTimeMs:v+_}:{asum:S[0]}}finally{!a&&u&&d(u),n&&d(n),f&&d(f),l&&d(l),m&&d(m)}}async function Uo(r,t,e,o){let a=e instanceof N;if(q(r),T(r,"dasum",{x:e}),!Number.isInteger(t)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!a&&!(e instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(a&&e.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(t<=0)return{asum:0};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=["f64/dekker","f64/utils/abs","f64/utils/add"],s=await D(r,[...i,"dasum"]),u=await D(r,[...i,"reduction/sumF64"]),n=null,f=null,l=null,m=null,p=null,c=null,g=null,w=null,h=null;try{if(a)n=e._buf,f=e._loBuf;else{let{hi:z,lo:H}=rr(e.map(Math.abs));n=x(r,z,"dasum-xHi",!1),f=x(r,H,"dasum-xLo",!1)}l=tr(r,512,"dasum-partialsHi"),m=tr(r,512,"dasum-partialsLo"),p=Br(r,4,"dasum-result-hi"),c=Br(r,4,"dasum-result-lo"),g=I(r,[{value:t,type:"u32"},{value:o,type:"u32"}],"dasum-params");let b=E(r,s.getBindGroupLayout(0),[n,f,l,m,g]),{commandEncoder:y,ts:v}=j(r,s,b,128);M(r,y);let _=E(r,u.getBindGroupLayout(0),[l,m,p,c]),{commandEncoder:S,ts:k}=j(r,u,_,1);w=G(r,S,p),h=G(r,S,c),M(r,S);let B=A(w,Float32Array),L=A(h,Float32Array);w=null,h=null;let[C,R,F,W]=await Promise.all([P(v),P(k),B,L]),V=dr(F,W)[0];return C!==void 0&&R!==void 0?{asum:V,gpuTimeMs:C+R}:{asum:V}}finally{!a&&n&&d(n),!a&&f&&d(f),l&&d(l),m&&d(m),p&&d(p),c&&d(c),g&&d(g),w&&d(w),h&&d(h)}}async function Yo(r,t,e,o,a,i){let s=e instanceof N,u=a instanceof N;if(q(r),T(r,"ddot",{x:e,y:a}),!Number.isInteger(t)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!s&&!(e instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(!u&&!(a instanceof Float64Array))throw new Error("y must be a Float64Array or GpuVector.");if(s&&e.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(u&&a.dtype!==Float64Array)throw new Error("y must be a Float64Array-backed GpuVector.");if(s!==u)throw new Error("x and y must be the same type (both Float64Array or both GpuVector).");if(t<=0)return{dot:0};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(a.length<(t-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=["f64/dekker","f64/utils/add"],f=await D(r,[...n,"f64/utils/multiply","ddot"]),l=await D(r,[...n,"reduction/sumF64"]),m=null,p=null,c=null,g=null,w=null,h=null,b=null,y=null,v=null,_=null,S=null;try{if(s)m=e._buf,p=e._loBuf,c=a._buf,g=a._loBuf;else{let ir=rr(e),lr=rr(a);m=x(r,ir.hi,"ddot-xHi",!1),p=x(r,ir.lo,"ddot-xLo",!1),c=x(r,lr.hi,"ddot-yHi",!1),g=x(r,lr.lo,"ddot-yLo",!1)}w=tr(r,512,"ddot-partialsHi"),h=tr(r,512,"ddot-partialsLo"),b=Br(r,4,"ddot-result-hi"),y=Br(r,4,"ddot-result-lo"),v=I(r,[{value:t,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"ddot-params");let k=E(r,f.getBindGroupLayout(0),[m,p,c,g,w,h,v]),{commandEncoder:B,ts:L}=j(r,f,k,128);M(r,B);let C=E(r,l.getBindGroupLayout(0),[w,h,b,y]),{commandEncoder:R,ts:F}=j(r,l,C,1);_=G(r,R,b),S=G(r,R,y),M(r,R);let W=A(_,Float32Array),V=A(S,Float32Array);_=null,S=null;let[z,H,$,K]=await Promise.all([P(L),P(F),W,V]),Y=dr($,K)[0];return z!==void 0&&H!==void 0?{dot:Y,gpuTimeMs:z+H}:{dot:Y}}finally{!s&&m&&d(m),!s&&p&&d(p),!u&&c&&d(c),!u&&g&&d(g),w&&d(w),h&&d(h),b&&d(b),y&&d(y),v&&d(v),_&&d(_),S&&d(S)}}async function Xo(r,t,e,o){let a=e instanceof N;if(q(r),T(r,"snrm2",{x:e}),!Number.isInteger(t)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!a&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(t<=0)return{nrm2:0};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=await D(r,"snrm2"),s=await D(r,"reduction/scaledSum"),u=null,n=null,f=null,l=null,m=null,p=null;try{u=a?e._buf:x(r,e,"snrm2-x",!1),n=tr(r,512,"snrm2-partials-scale"),f=tr(r,512,"snrm2-partials-ssq"),l=Br(r,4,"snrm2-result"),m=I(r,[{value:t,type:"u32"},{value:o,type:"u32"}],"snrm2-params");let c=E(r,i.getBindGroupLayout(0),[u,n,f,m]),{commandEncoder:g,ts:w}=j(r,i,c,128);M(r,g);let h=E(r,s.getBindGroupLayout(0),[n,f,l]),{commandEncoder:b,ts:y}=j(r,s,h,1);p=G(r,b,l),M(r,b);let v=A(p,Float32Array);p=null;let[_,S,k]=await Promise.all([P(w),P(y),v]),B=k[0];return _!==void 0&&S!==void 0?{nrm2:B,gpuTimeMs:_+S}:{nrm2:B}}finally{!a&&u&&d(u),n&&d(n),f&&d(f),l&&d(l),m&&d(m),p&&d(p)}}async function $o(r,t,e,o){let a=e instanceof N;if(q(r),T(r,"dnrm2",{x:e}),!Number.isInteger(t)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!a&&!(e instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(a&&e.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(t<=0)return{nrm2:0};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=["f64/dekker","f64/utils/abs","f64/utils/greater","f64/utils/add","f64/utils/multiply","f64/utils/divide","f64/utils/sqrt"],s=await D(r,[...i,"dnrm2"]),u=await D(r,[...i,"reduction/scaledSumF64"]),n=null,f=null,l=null,m=null,p=null,c=null,g=null,w=null,h=null,b=null,y=null;try{if(a)n=e._buf,f=e._loBuf;else{let{hi:$,lo:K}=rr(e);n=x(r,$,"dnrm2-xHi",!1),f=x(r,K,"dnrm2-xLo",!1)}l=tr(r,512,"dnrm2-partials-scaleHi"),m=tr(r,512,"dnrm2-partials-scaleLo"),p=tr(r,512,"dnrm2-partials-ssqHi"),c=tr(r,512,"dnrm2-partials-ssqLo"),g=Br(r,4,"dnrm2-result-hi"),w=Br(r,4,"dnrm2-result-lo"),h=I(r,[{value:t,type:"u32"},{value:o,type:"u32"}],"dnrm2-params");let v=E(r,s.getBindGroupLayout(0),[n,f,l,m,p,c,h]),{commandEncoder:_,ts:S}=j(r,s,v,128);M(r,_);let k=E(r,u.getBindGroupLayout(0),[l,m,p,c,g,w]),{commandEncoder:B,ts:L}=j(r,u,k,1);b=G(r,B,g),y=G(r,B,w),M(r,B);let C=A(b,Float32Array),R=A(y,Float32Array);b=null,y=null;let[F,W,V,z]=await Promise.all([P(S),P(L),C,R]),H=dr(V,z)[0];return F!==void 0&&W!==void 0?{nrm2:H,gpuTimeMs:F+W}:{nrm2:H}}finally{!a&&n&&d(n),!a&&f&&d(f),l&&d(l),m&&d(m),p&&d(p),c&&d(c),g&&d(g),w&&d(w),h&&d(h),b&&d(b),y&&d(y)}}async function Zo(r,t,e,o){let a=e instanceof N;if(q(r),T(r,"isamax",{x:e}),!Number.isInteger(t)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!a&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(t<=0)return{index:0};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=await D(r,"isamax"),s=await D(r,"reduction/argmax"),u=null,n=null,f=null,l=null,m=null,p=null;try{u=a?e._buf:x(r,e,"isamax-x",!1),n=tr(r,512,"isamax-partials-val"),f=tr(r,512,"isamax-partials-idx"),l=Br(r,4,"isamax-result"),m=I(r,[{value:t,type:"u32"},{value:o,type:"u32"}],"isamax-params");let c=E(r,i.getBindGroupLayout(0),[u,n,f,m]),{commandEncoder:g,ts:w}=j(r,i,c,128);M(r,g);let h=E(r,s.getBindGroupLayout(0),[n,f,l]),{commandEncoder:b,ts:y}=j(r,s,h,1);p=G(r,b,l),M(r,b);let v=A(p,Uint32Array);p=null;let[_,S,k]=await Promise.all([P(w),P(y),v]),B=k[0];return _!==void 0&&S!==void 0?{index:B,gpuTimeMs:_+S}:{index:B}}finally{!a&&u&&d(u),n&&d(n),f&&d(f),l&&d(l),m&&d(m),p&&d(p)}}async function Qo(r,t,e,o){let a=e instanceof N;if(q(r),T(r,"idamax",{x:e}),!Number.isInteger(t)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!a&&!(e instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(a&&e.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(t<=0)return{index:0};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=["f64/dekker","f64/utils/abs","f64/utils/greater","f64/utils/equal"],s=await D(r,[...i,"idamax"],"idamax_main"),u=await D(r,[...i,"reduction/argmaxF64"],"reduce_f64"),n=null,f=null,l=null,m=null,p=null,c=null,g=null,w=null;try{if(a)n=e._buf,f=e._loBuf;else{let{hi:F,lo:W}=rr(e);n=x(r,F,"idamax-xHi",!1),f=x(r,W,"idamax-xLo",!1)}l=tr(r,512,"idamax-partials-val-hi"),m=tr(r,512,"idamax-partials-val-lo"),p=tr(r,512,"idamax-partials-idx"),c=Br(r,4,"idamax-result"),g=I(r,[{value:t,type:"u32"},{value:o,type:"u32"}],"idamax-params");let h=E(r,s.getBindGroupLayout(0),[n,f,l,m,p,g]),{commandEncoder:b,ts:y}=j(r,s,h,128);M(r,b);let v=E(r,u.getBindGroupLayout(0),[l,m,p,c]),{commandEncoder:_,ts:S}=j(r,u,v,1);w=G(r,_,c),M(r,_);let k=A(w,Uint32Array);w=null;let[B,L,C]=await Promise.all([P(y),P(S),k]),R=C[0];return B!==void 0&&L!==void 0?{index:R,gpuTimeMs:B+L}:{index:R}}finally{!a&&n&&d(n),!a&&f&&d(f),l&&d(l),m&&d(m),p&&d(p),c&&d(c),g&&d(g),w&&d(w)}}async function Jo(r,t,e,o,a,i,s,u){let n=e instanceof N,f=a instanceof N;if(q(r),T(r,"srot",{x:e,y:a}),!Number.isInteger(t)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(typeof s!="number")throw new Error("c must be a number.");if(typeof u!="number")throw new Error("s must be a number.");if(Number.isNaN(s)||Number.isNaN(u))throw new Error("c and s must not be NaN.");if(!Number.isFinite(s))throw new Error("c must be finite.");if(!Number.isFinite(u))throw new Error("s must be finite.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!n&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!f&&!(a instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(n!==f)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(t<=0)return n?{}:{x:e,y:a};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(a.length<(t-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let l=await D(r,"srot"),m=null,p=null,c=null,g=null,w=null;try{m=n?e._buf:x(r,e,"srot-x",!0),p=f?a._buf:x(r,a,"srot-y",!0),c=I(r,[{value:t,type:"u32"},{value:s,type:"f32"},{value:u,type:"f32"},{value:o,type:"u32"},{value:i,type:"u32"}],"srot-params");let h=E(r,l.getBindGroupLayout(0),[m,p,c]),{commandEncoder:b,ts:y}=j(r,l,h,cr(r,t));g=n?null:G(r,b,m),w=f?null:G(r,b,p),M(r,b);let v=await P(y);if(n)return v!==void 0?{gpuTimeMs:v}:{};let _=A(g,Float32Array),S=A(w,Float32Array);g=null,w=null;let[k,B]=await Promise.all([_,S]);return v!==void 0?{x:k,y:B,gpuTimeMs:v}:{x:k,y:B}}finally{!n&&m&&d(m),!f&&p&&d(p),c&&d(c),g&&d(g),w&&d(w)}}async function ra(r,t,e,o,a,i,s,u){let n=e instanceof N,f=a instanceof N;if(q(r),T(r,"drot",{x:e,y:a}),!Number.isInteger(t)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(typeof s!="number")throw new Error("c must be a number.");if(typeof u!="number")throw new Error("s must be a number.");if(Number.isNaN(s)||Number.isNaN(u))throw new Error("c and s must not be NaN.");if(!Number.isFinite(s))throw new Error("c must be finite.");if(!Number.isFinite(u))throw new Error("s must be finite.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!(e instanceof Float64Array)&&!n)throw new Error("x must be a Float64Array or GpuVector.");if(!(a instanceof Float64Array)&&!f)throw new Error("y must be a Float64Array or GpuVector.");if(n&&e.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(f&&a.dtype!==Float64Array)throw new Error("y must be a Float64Array-backed GpuVector.");if(n!==f)throw new Error("x and y must be the same type (both Float64Array or both GpuVector).");if(t<=0)return n?{}:{x:e,y:a};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(a.length<(t-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let m=await D(r,[...["f64/dekker","f64/utils/add","f64/utils/multiply"],"drot"]),{hi:p,lo:c}=rr(new Float64Array([s])),{hi:g,lo:w}=rr(new Float64Array([u])),h=null,b=null,y=null,v=null,_=null,S=null,k=null,B=null,L=null;try{if(n)h=e._buf,b=e._loBuf,y=a._buf,v=a._loBuf;else{let ir=rr(e),lr=rr(a);h=x(r,ir.hi,"drot-xHi",!0),b=x(r,ir.lo,"drot-xLo",!0),y=x(r,lr.hi,"drot-yHi",!0),v=x(r,lr.lo,"drot-yLo",!0)}_=I(r,[{value:t,type:"u32"},{value:p[0],type:"f32"},{value:c[0],type:"f32"},{value:g[0],type:"f32"},{value:w[0],type:"f32"},{value:o,type:"u32"},{value:i,type:"u32"}],"drot-params");let C=E(r,m.getBindGroupLayout(0),[h,b,y,v,_]),{commandEncoder:R,ts:F}=j(r,m,C,cr(r,t));S=n?null:G(r,R,h),k=n?null:G(r,R,b),B=f?null:G(r,R,y),L=f?null:G(r,R,v),M(r,R);let W=await P(F);if(n)return W!==void 0?{gpuTimeMs:W}:{};let V=await A(S,Float32Array);S=null;let z=await A(k,Float32Array);k=null;let H=await A(B,Float32Array);B=null;let $=await A(L,Float32Array);L=null;let K=dr(V,z),Y=dr(H,$);return W!==void 0?{x:K,y:Y,gpuTimeMs:W}:{x:K,y:Y}}finally{!n&&h&&d(h),!n&&b&&d(b),!f&&y&&d(y),!f&&v&&d(v),_&&d(_),S&&d(S),k&&d(k),B&&d(B),L&&d(L)}}async function ea(r,t,e,o,a,i,s){let u=e instanceof N,n=a instanceof N;if(q(r),T(r,"srotm",{x:e,y:a}),!Number.isInteger(t)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(!(s instanceof Float32Array)||s.length!==5)throw new Error("param must be a Float32Array of length 5.");if(s[0]!==-2&&s[0]!==-1&&s[0]!==0&&s[0]!==1)throw new Error("param[0] (flag) must be one of -2, -1, 0, or 1.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!u&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!n&&!(a instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(u!==n)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(t<=0||s[0]===-2)return u?{}:{x:e,y:a};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(a.length<(t-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let f=await D(r,"srotm"),l=null,m=null,p=null,c=null,g=null,w=null;try{l=u?e._buf:x(r,e,"srotm-x",!0),m=n?a._buf:x(r,a,"srotm-y",!0),p=x(r,s,"srotm-param",!1),c=I(r,[{value:t,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"srotm-params");let h=E(r,f.getBindGroupLayout(0),[l,m,p,c]),{commandEncoder:b,ts:y}=j(r,f,h,cr(r,t));g=u?null:G(r,b,l),w=n?null:G(r,b,m),M(r,b);let v=await P(y);if(u)return v!==void 0?{gpuTimeMs:v}:{};let _=A(g,Float32Array),S=A(w,Float32Array);g=null,w=null;let[k,B]=await Promise.all([_,S]);return v!==void 0?{x:k,y:B,gpuTimeMs:v}:{x:k,y:B}}finally{!u&&l&&d(l),!n&&m&&d(m),p&&d(p),c&&d(c),g&&d(g),w&&d(w)}}async function ta(r,t,e,o,a,i,s){let u=e instanceof N,n=a instanceof N;if(q(r),T(r,"drotm",{x:e,y:a}),!Number.isInteger(t)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(!(s instanceof Float64Array)||s.length!==5)throw new Error("param must be a Float64Array of length 5.");if(s[0]!==-2&&s[0]!==-1&&s[0]!==0&&s[0]!==1)throw new Error("param[0] (flag) must be one of -2, -1, 0, or 1.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!(e instanceof Float64Array)&&!u)throw new Error("x must be a Float64Array or GpuVector.");if(!(a instanceof Float64Array)&&!n)throw new Error("y must be a Float64Array or GpuVector.");if(u&&e.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(n&&a.dtype!==Float64Array)throw new Error("y must be a Float64Array-backed GpuVector.");if(u!==n)throw new Error("x and y must be the same type (both Float64Array or both GpuVector).");if(t<=0||s[0]===-2)return u?{}:{x:e,y:a};if(e.length<(t-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(a.length<(t-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let l=await D(r,[...["f64/dekker","f64/utils/add","f64/utils/multiply"],"drotm"]),{hi:m,lo:p}=rr(s),c=null,g=null,w=null,h=null,b=null,y=null,v=null,_=null,S=null,k=null,B=null;try{if(u)c=e._buf,g=e._loBuf,w=a._buf,h=a._loBuf;else{let Y=rr(e),ir=rr(a);c=x(r,Y.hi,"drotm-xHi",!0),g=x(r,Y.lo,"drotm-xLo",!0),w=x(r,ir.hi,"drotm-yHi",!0),h=x(r,ir.lo,"drotm-yLo",!0)}b=x(r,m,"drotm-paramHi",!1),y=x(r,p,"drotm-paramLo",!1),v=I(r,[{value:t,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"drotm-params");let L=E(r,l.getBindGroupLayout(0),[c,g,w,h,b,y,v]),{commandEncoder:C,ts:R}=j(r,l,L,cr(r,t));_=u?null:G(r,C,c),S=u?null:G(r,C,g),k=n?null:G(r,C,w),B=n?null:G(r,C,h),M(r,C);let F=await P(R);if(u)return F!==void 0?{gpuTimeMs:F}:{};let W=await A(_,Float32Array);_=null;let V=await A(S,Float32Array);S=null;let z=await A(k,Float32Array);k=null;let H=await A(B,Float32Array);B=null;let $=dr(W,V),K=dr(z,H);return F!==void 0?{x:$,y:K,gpuTimeMs:F}:{x:$,y:K}}finally{!u&&c&&d(c),!u&&g&&d(g),!n&&w&&d(w),!n&&h&&d(h),b&&d(b),y&&d(y),v&&d(v),_&&d(_),S&&d(S),k&&d(k),B&&d(B)}}async function oa(r,t,e,o,a,i,s,u,n,f,l,m,p="row-major"){let c=i instanceof X,g=u instanceof N,w=l instanceof N;if(q(r),T(r,"sgemv",{A:i,x:u,y:l}),t!=="no-transpose"&&t!=="transpose")throw new Error("trans must be 'no-transpose' or 'transpose'.");if(p!=="row-major"&&p!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(typeof a!="number")throw new Error("alpha must be a number.");if(Number.isNaN(a))throw new Error("alpha must not be NaN.");if(!Number.isFinite(a))throw new Error("alpha must be finite.");if(typeof f!="number")throw new Error("beta must be a number.");if(Number.isNaN(f))throw new Error("beta must not be NaN.");if(!Number.isFinite(f))throw new Error("beta must be finite.");if(!Number.isInteger(e)||!Number.isInteger(o)||!Number.isInteger(n)||!Number.isInteger(m)||!Number.isInteger(s))throw new Error("m, n, incx, incy, and lda must be integers.");if(n<=0||m<=0)throw new Error("incx and incy must be positive.");if(!c&&!(i instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!g&&!(u instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!w&&!(l instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(g!==w)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(g&&!c)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(c&&!g)throw new Error("x and y must be GpuVectors when A is a GpuMatrix.");if(g&&u._buf===l._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(c&&w&&i._buf===l._buf)throw new Error("A and y must not reference the same GPU buffer.");if(c&&s!==i.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(c&&(i.rows<e||i.cols<o))throw new Error("A is too small for the given m and n.");if(e<0||o<0)throw new Error("m and n must be non-negative.");if(e===0||o===0)return w?{}:{y:l};(c?i.layout:p)==="column-major"&&([e,o]=[o,e],t=t==="no-transpose"?"transpose":"no-transpose");let b=t==="no-transpose",y=b?o:e,v=b?e:o;if(s<o)throw new Error("lda must be >= n.");if(!c&&i.length<(e-1)*s+o)throw new Error("A does not have enough elements for the given m, n, and lda.");if(u.length<(y-1)*n+1)throw new Error("x does not have enough elements for the given dimensions and incx.");if(l.length<(v-1)*m+1)throw new Error("y does not have enough elements for the given dimensions and incy.");let S=await D(r,b?"sgemv_n":"sgemv_t"),k=null,B=null,L=null,C=null;try{k=c?i._buf:x(r,i,"sgemv-A",!1),B=g?u._buf:x(r,u,"sgemv-x",!1),L=w?l._buf:x(r,l,"sgemv-y",!0),C=I(r,[{value:e,type:"u32"},{value:o,type:"u32"},{value:a,type:"f32"},{value:f,type:"f32"},{value:n,type:"u32"},{value:m,type:"u32"},{value:s,type:"u32"}],"sgemv-params");let R=E(r,S.getBindGroupLayout(0),[k,B,L,C]),F=b?Math.min(e,r.limits.maxComputeWorkgroupsPerDimension):Zr(r,"sgemv",v),{commandEncoder:W,ts:V}=j(r,S,R,F),z=w?null:G(r,W,L);M(r,W);let H=await P(V);if(w)return H!==void 0?{gpuTimeMs:H}:{};let $=await A(z,Float32Array);return H!==void 0?{y:$,gpuTimeMs:H}:{y:$}}finally{!c&&k&&d(k),!g&&B&&d(B),!w&&L&&d(L),C&&d(C)}}async function aa(r,t,e,o,a,i,s,u,n,f,l,m="row-major"){let p=s instanceof N,c=f instanceof N,g=a instanceof X;if(q(r),T(r,"ssymv",{A:a,x:s,y:f}),t!=="lower"&&t!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(m!=="row-major"&&m!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(!Number.isInteger(e)||!Number.isInteger(u)||!Number.isInteger(l)||!Number.isInteger(i))throw new Error("n, incx, incy, and lda must be integers.");if(typeof o!="number")throw new Error("alpha must be a number.");if(Number.isNaN(o))throw new Error("alpha must not be NaN.");if(!Number.isFinite(o))throw new Error("alpha must be finite.");if(typeof n!="number")throw new Error("beta must be a number.");if(Number.isNaN(n))throw new Error("beta must not be NaN.");if(!Number.isFinite(n))throw new Error("beta must be finite.");if(u<=0||l<=0)throw new Error("incx and incy must be positive.");if(i<e)throw new Error("lda must be >= n.");if(!g&&!(a instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!p&&!(s instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!c&&!(f instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(p!==c)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(p&&!g)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(g&&!p)throw new Error("x and y must be GpuVectors when A is a GpuMatrix.");if(p&&s._buf===f._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(g&&i!==a.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(g&&(a.rows<e||a.cols<e))throw new Error("A is too small for the given n.");if(e<0)throw new Error("n must be non-negative.");if(e===0)return c?{}:{y:f};if(!g&&a.length<(e-1)*i+e)throw new Error("A does not have enough elements for the given n and lda.");if(s.length<(e-1)*u+1)throw new Error("x does not have enough elements for the given n and incx.");if(f.length<(e-1)*l+1)throw new Error("y does not have enough elements for the given n and incy.");let h=(g?a.layout:m)==="column-major"?t==="upper":t==="lower",b=await D(r,"ssymv"),y=null,v=null,_=null,S=null;try{y=g?a._buf:x(r,a,"ssymv-A",!1),v=p?s._buf:x(r,s,"ssymv-x",!1),_=c?f._buf:x(r,f,"ssymv-y",!0),S=I(r,[{value:e,type:"u32"},{value:o,type:"f32"},{value:n,type:"f32"},{value:u,type:"u32"},{value:l,type:"u32"},{value:i,type:"u32"},{value:h?0:1,type:"u32"}],"ssymv-params");let k=E(r,b.getBindGroupLayout(0),[y,v,_,S]),B=Math.min(e,r.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:L,ts:C}=j(r,b,k,B),R=c?null:G(r,L,_);M(r,L);let F=await P(C);if(c)return F!==void 0?{gpuTimeMs:F}:{};let W=await A(R,Float32Array);return F!==void 0?{y:W,gpuTimeMs:F}:{y:W}}finally{!g&&y&&d(y),!p&&v&&d(v),!c&&_&&d(_),S&&d(S)}}async function ia(r,t,e,o,a,i,s,u,n,f,l,m="row-major"){let p=u instanceof N,c=f instanceof N,g=i instanceof X,w=o==="unit";if(q(r),T(r,"strmv",{A:i,x:u,y:f}),t!=="lower"&&t!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(e!=="no-transpose"&&e!=="transpose")throw new Error("trans must be 'no-transpose' or 'transpose'.");if(!w&&o!=="non-unit")throw new Error("diag must be 'unit' or 'non-unit'.");if(m!=="row-major"&&m!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(!Number.isInteger(a)||!Number.isInteger(n)||!Number.isInteger(l)||!Number.isInteger(s))throw new Error("n, incx, incy, and lda must be integers.");if(n<=0||l<=0)throw new Error("incx and incy must be positive.");if(s<a)throw new Error("lda must be >= n.");if(!g&&!(i instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!p&&!(u instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!c&&!(f instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(p!==c)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(p&&u._buf===f._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(p&&!g)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(g&&!p)throw new Error("x and y must be GpuVectors when A is a GpuMatrix.");if(g&&c&&i._buf===f._buf)throw new Error("A and y must not reference the same GPU buffer.");if(g&&s!==i.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(g&&(i.rows<a||i.cols<a))throw new Error("A is too small for the given n.");if(a<0)throw new Error("n must be non-negative.");if(a===0)return c?{}:{y:f};if(!g&&i.length<(a-1)*s+a)throw new Error("A does not have enough elements for the given n and lda.");if(u.length<(a-1)*n+1)throw new Error("x does not have enough elements for the given n and incx.");if(f.length<(a-1)*l+1)throw new Error("y does not have enough elements for the given n and incy.");let b=(g?i.layout:m)==="column-major",y=b?t==="upper":t==="lower",v=b?e==="transpose":e==="no-transpose",_=await D(r,"strmv"),S=null,k=null,B=null,L=null;try{S=g?i._buf:x(r,i,"strmv-A",!1),k=p?u._buf:x(r,u,"strmv-x",!1),B=c?f._buf:x(r,f,"strmv-y",!0),L=I(r,[{value:a,type:"u32"},{value:n,type:"u32"},{value:l,type:"u32"},{value:s,type:"u32"},{value:v?0:1,type:"u32"},{value:y?0:1,type:"u32"},{value:w?1:0,type:"u32"}],"strmv-params");let C=E(r,_.getBindGroupLayout(0),[S,k,B,L]),R=Math.min(a,r.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:F,ts:W}=j(r,_,C,R),V=c?null:G(r,F,B);M(r,F);let z=await P(W);if(c)return z!==void 0?{gpuTimeMs:z}:{};let H=await A(V,Float32Array);return z!==void 0?{y:H,gpuTimeMs:z}:{y:H}}finally{!g&&S&&d(S),!p&&k&&d(k),!c&&B&&d(B),L&&d(L)}}function sa(r,t,e){let o=new ArrayBuffer(r*t),a=new DataView(o);for(let i=0;i<r;i++){let s=e(i),u=i*t;s.forEach((n,f)=>a.setUint32(u+f*4,n,!0))}return o}function na(r,t,e){let o=r.createBuffer({label:e,size:t.byteLength,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});return r.queue.writeBuffer(o,0,t),o}async function la(r,t,e,o,a,i,s,u,n,f="row-major"){let l=u instanceof N,m=i instanceof X,p=o==="unit";if(q(r),T(r,"strsv",{A:i,x:u}),t!=="lower"&&t!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(e!=="no-transpose"&&e!=="transpose")throw new Error("trans must be 'no-transpose' or 'transpose'.");if(!p&&o!=="non-unit")throw new Error("diag must be 'unit' or 'non-unit'.");if(f!=="row-major"&&f!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(!Number.isInteger(a)||!Number.isInteger(n)||!Number.isInteger(s))throw new Error("n, incx, and lda must be integers.");if(n<=0)throw new Error("incx must be positive.");if(s<a)throw new Error("lda must be >= n.");if(!m&&!(i instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!l&&!(u instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(l&&!m)throw new Error("A must be a GpuMatrix when x is a GpuVector.");if(m&&!l)throw new Error("x must be a GpuVector when A is a GpuMatrix.");if(m&&l&&i._buf===u._buf)throw new Error("A and x must not reference the same GPU buffer.");if(m&&s!==i.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(m&&(i.rows<a||i.cols<a))throw new Error("A is too small for the given n.");if(a<0)throw new Error("n must be non-negative.");if(a===0)return l?{}:{x:u};if(!m&&i.length<(a-1)*s+a)throw new Error("A does not have enough elements for the given n and lda.");if(u.length<(a-1)*n+1)throw new Error("x does not have enough elements for the given n and incx.");let g=(m?i.layout:f)==="column-major",w=g?t==="upper":t==="lower",h=g?e==="transpose":e==="no-transpose",b=await D(r,"strsv_invert_block"),y=await D(r,"strsv_apply_inverse"),v=await D(r,"strsv_update"),_=h===w,S=[];for(let H=0;H<a;H+=64)S.push(H);_||S.reverse();let k=S.length,B=r.limits.maxComputeWorkgroupsPerDimension,L=r.limits.minUniformBufferOffsetAlignment,C=null,R=null,F=null,W=null,V=null,z=null;try{C=m?i._buf:x(r,i,"strsv-A",!1),R=l?u._buf:x(r,u,"strsv-x",!0),F=tr(r,k*64*64*4,"strsv-Ainv");let H=sa(k,L,er=>{let Z=er*64,J=Math.min(Z+64,a);return[n,er,Z,J]});W=na(r,H,"strsv-apply-params");let $=sa(k,L,er=>{let Z=er*64,J=Math.min(Z+64,a);return[a,n,s,h?0:1,w?0:1,Z,J]});V=na(r,$,"strsv-update-params");let{commandEncoder:K,querySet:Y}=qr(r);z=I(r,[{value:a,type:"u32"},{value:s,type:"u32"},{value:h?0:1,type:"u32"},{value:w?0:1,type:"u32"},{value:p?1:0,type:"u32"}],"strsv-invert-params");let ir=E(r,b.getBindGroupLayout(0),[C,F,z]);gr(K,b,ir,{x:64,y:k},Y?{timestampWrites:{querySet:Y,beginningOfPassWriteIndex:0}}:void 0);for(let er=0;er<S.length;er++){let Z=S[er],J=Math.min(Z+64,a),sr=Z/64,pr=er===S.length-1,br=sr*L,fr=E(r,y.getBindGroupLayout(0),[F,R,{buffer:W,offset:br,size:16}]);gr(K,y,fr,1,pr&&Y?{timestampWrites:{querySet:Y,endOfPassWriteIndex:1}}:void 0);let yr=_?a-J:Z;if(yr===0)continue;let Cr=E(r,v.getBindGroupLayout(0),[C,R,{buffer:V,offset:br,size:32}]),Rr=Math.min(yr,B);gr(K,v,Cr,Rr)}let hr=Lr(r,K,Y),nr=l?null:G(r,K,R);M(r,K);let mr=await P(hr);if(l)return mr!==void 0?{gpuTimeMs:mr}:{};let Q=await A(nr,Float32Array);return mr!==void 0?{x:Q,gpuTimeMs:mr}:{x:Q}}finally{!m&&C&&d(C),!l&&R&&d(R),F&&d(F),W&&d(W),V&&d(V),z&&d(z)}}async function ua(r,t,e,o,a,i,s,u,n,f,l="row-major"){let m=n instanceof X;if(q(r),T(r,"sger",{A:n,x:a,y:s}),l!=="row-major"&&l!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(typeof o!="number")throw new Error("alpha must be a number.");if(Number.isNaN(o))throw new Error("alpha must not be NaN.");if(!Number.isFinite(o))throw new Error("alpha must be finite.");if(!Number.isInteger(t)||!Number.isInteger(e)||!Number.isInteger(i)||!Number.isInteger(u)||!Number.isInteger(f))throw new Error("m, n, incx, incy, and lda must be integers.");if(i<=0||u<=0)throw new Error("incx and incy must be positive.");if(!m&&!(n instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(m&&f!==n.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(m&&(n.rows<t||n.cols<e))throw new Error("A is too small for the given m and n.");(m?n.layout:l)==="column-major"&&([t,e]=[e,t],[a,s]=[s,a],[i,u]=[u,i]);let c=a instanceof N,g=s instanceof N;if(f<e)throw new Error("lda must be >= n.");if(!c&&!(a instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!g&&!(s instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(c!==g)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(c&&!m)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(m&&!c)throw new Error("x and y must be GpuVectors when A is a GpuMatrix.");if(m&&c&&n._buf===a._buf)throw new Error("A and x must not reference the same GPU buffer.");if(m&&g&&n._buf===s._buf)throw new Error("A and y must not reference the same GPU buffer.");if(t<0||e<0)throw new Error("m and n must be non-negative.");if(t===0||e===0)return m?{}:{A:n};if(!m&&n.length<(t-1)*f+e)throw new Error("A does not have enough elements for the given m, n, and lda.");if(a.length<(t-1)*i+1)throw new Error("x does not have enough elements for the given m and incx.");if(s.length<(e-1)*u+1)throw new Error("y does not have enough elements for the given n and incy.");let w=await D(r,"sger"),h=null,b=null,y=null,v=null;try{h=c?a._buf:x(r,a,"sger-x",!1),b=g?s._buf:x(r,s,"sger-y",!1),y=m?n._buf:x(r,n,"sger-A",!0),v=I(r,[{value:t,type:"u32"},{value:e,type:"u32"},{value:o,type:"f32"},{value:i,type:"u32"},{value:u,type:"u32"},{value:f,type:"u32"}],"sger-params");let _=E(r,w.getBindGroupLayout(0),[h,b,y,v]),S=Math.min(t,r.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:k,ts:B}=j(r,w,_,S),L=m?null:G(r,k,y);M(r,k);let C=await P(B);if(m)return C!==void 0?{gpuTimeMs:C}:{};let R=await A(L,Float32Array);return C!==void 0?{A:R,gpuTimeMs:C}:{A:R}}finally{!c&&h&&d(h),!g&&b&&d(b),!m&&y&&d(y),v&&d(v)}}async function fa(r,t,e,o,a,i,s,u,n="row-major"){let f=a instanceof N,l=s instanceof X;if(q(r),T(r,"ssyr",{A:s,x:a}),t!=="lower"&&t!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(n!=="row-major"&&n!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(!Number.isInteger(e)||!Number.isInteger(i)||!Number.isInteger(u))throw new Error("n, incx, and lda must be integers.");if(typeof o!="number")throw new Error("alpha must be a number.");if(Number.isNaN(o))throw new Error("alpha must not be NaN.");if(!Number.isFinite(o))throw new Error("alpha must be finite.");if(i<=0)throw new Error("incx must be positive.");if(u<e)throw new Error("lda must be >= n.");if(!l&&!(s instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!f&&!(a instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(f&&!l)throw new Error("A must be a GpuMatrix when x is a GpuVector.");if(l&&!f)throw new Error("x must be a GpuVector when A is a GpuMatrix.");if(l&&f&&s._buf===a._buf)throw new Error("A and x must not reference the same GPU buffer.");if(l&&u!==s.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(l&&(s.rows<e||s.cols<e))throw new Error("A is too small for the given n.");if(e<0)throw new Error("n must be non-negative.");if(e===0)return l?{}:{A:s};if(!l&&s.length<(e-1)*u+e)throw new Error("A does not have enough elements for the given n and lda.");if(a.length<(e-1)*i+1)throw new Error("x does not have enough elements for the given n and incx.");let p=(l?s.layout:n)==="column-major"?t==="upper":t==="lower",c=await D(r,"ssyr"),g=null,w=null,h=null;try{g=f?a._buf:x(r,a,"ssyr-x",!1),w=l?s._buf:x(r,s,"ssyr-A",!0),h=I(r,[{value:e,type:"u32"},{value:o,type:"f32"},{value:i,type:"u32"},{value:u,type:"u32"},{value:p?0:1,type:"u32"}],"ssyr-params");let b=E(r,c.getBindGroupLayout(0),[g,w,h]),y=Math.min(e,r.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:v,ts:_}=j(r,c,b,y),S=l?null:G(r,v,w);M(r,v);let k=await P(_);if(l)return k!==void 0?{gpuTimeMs:k}:{};let B=await A(S,Float32Array);return k!==void 0?{A:B,gpuTimeMs:k}:{A:B}}finally{!f&&g&&d(g),!l&&w&&d(w),h&&d(h)}}async function ma(r,t,e,o,a,i,s,u,n,f,l="row-major"){let m=a instanceof N,p=s instanceof N,c=n instanceof X;if(q(r),T(r,"ssyr2",{A:n,x:a,y:s}),t!=="lower"&&t!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(l!=="row-major"&&l!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(!Number.isInteger(e)||!Number.isInteger(i)||!Number.isInteger(u)||!Number.isInteger(f))throw new Error("n, incx, incy, and lda must be integers.");if(typeof o!="number")throw new Error("alpha must be a number.");if(Number.isNaN(o))throw new Error("alpha must not be NaN.");if(!Number.isFinite(o))throw new Error("alpha must be finite.");if(i<=0||u<=0)throw new Error("incx and incy must be positive.");if(f<e)throw new Error("lda must be >= n.");if(!c&&!(n instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!m&&!(a instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!p&&!(s instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(m!==p)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(m&&!c)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(c&&!m)throw new Error("x and y must be GpuVectors when A is a GpuMatrix.");if(c&&m&&n._buf===a._buf)throw new Error("A and x must not reference the same GPU buffer.");if(c&&p&&n._buf===s._buf)throw new Error("A and y must not reference the same GPU buffer.");if(m&&a._buf===s._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(c&&f!==n.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(c&&(n.rows<e||n.cols<e))throw new Error("A is too small for the given n.");if(e<0)throw new Error("n must be non-negative.");if(e===0)return c?{}:{A:n};if(!c&&n.length<(e-1)*f+e)throw new Error("A does not have enough elements for the given n and lda.");if(a.length<(e-1)*i+1)throw new Error("x does not have enough elements for the given n and incx.");if(s.length<(e-1)*u+1)throw new Error("y does not have enough elements for the given n and incy.");let w=(c?n.layout:l)==="column-major"?t==="upper":t==="lower",h=await D(r,"ssyr2"),b=null,y=null,v=null,_=null;try{b=m?a._buf:x(r,a,"ssyr2-x",!1),y=p?s._buf:x(r,s,"ssyr2-y",!1),v=c?n._buf:x(r,n,"ssyr2-A",!0),_=I(r,[{value:e,type:"u32"},{value:o,type:"f32"},{value:i,type:"u32"},{value:u,type:"u32"},{value:f,type:"u32"},{value:w?0:1,type:"u32"}],"ssyr2-params");let S=E(r,h.getBindGroupLayout(0),[b,y,v,_]),k=Math.min(e,r.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:B,ts:L}=j(r,h,S,k),C=c?null:G(r,B,v);M(r,B);let R=await P(L);if(c)return R!==void 0?{gpuTimeMs:R}:{};let F=await A(C,Float32Array);return R!==void 0?{A:F,gpuTimeMs:R}:{A:F}}finally{!m&&b&&d(b),!p&&y&&d(y),!c&&v&&d(v),_&&d(_)}}async function da(r,t,e,o,a,i,s,u,n,f,l,m,p,c,g="row-major"){let w=u instanceof X,h=f instanceof X,b=p instanceof X;if(q(r),T(r,"sgemm",{A:u,B:f,C:p}),t!=="no-transpose"&&t!=="transpose")throw new Error("transA must be 'no-transpose' or 'transpose'.");if(e!=="no-transpose"&&e!=="transpose")throw new Error("transB must be 'no-transpose' or 'transpose'.");if(g!=="row-major"&&g!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(typeof s!="number")throw new Error("alpha must be a number.");if(Number.isNaN(s))throw new Error("alpha must not be NaN.");if(!Number.isFinite(s))throw new Error("alpha must be finite.");if(typeof m!="number")throw new Error("beta must be a number.");if(Number.isNaN(m))throw new Error("beta must not be NaN.");if(!Number.isFinite(m))throw new Error("beta must be finite.");if(!Number.isInteger(o)||!Number.isInteger(a)||!Number.isInteger(i)||!Number.isInteger(n)||!Number.isInteger(l)||!Number.isInteger(c))throw new Error("m, n, k, lda, ldb, and ldc must be integers.");if(!w&&!(u instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!h&&!(f instanceof Float32Array))throw new Error("B must be a Float32Array or GpuMatrix.");if(!b&&!(p instanceof Float32Array))throw new Error("C must be a Float32Array or GpuMatrix.");if((w||h)&&!b)throw new Error("C must be a GpuMatrix when A or B is a GpuMatrix.");if(b&&(!w||!h))throw new Error("A and B must be GpuMatrix when C is a GpuMatrix.");if(o<0||a<0||i<0)throw new Error("m, n, and k must be non-negative.");if(n<=0||l<=0||c<=0)throw new Error("lda, ldb, and ldc must be positive.");if(o===0||a===0)return b?{}:{C:p};let y=w?u.layout:g,v=h?f.layout:g,_=b?p.layout:g,S=y==="column-major"?i:o,k=y==="column-major"?o:i,B=t==="no-transpose"?S:k,L=t==="no-transpose"?k:S;if(n<L)throw new Error(`lda must be >= ${y==="column-major"?"rows":"cols"} of A as stored.`);if(w){if(n!==u.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");let[J,sr]=t==="no-transpose"?[o,i]:[i,o];if(u.rows<J||u.cols<sr)throw new Error("A is too small for the given m, k, and transA.")}else if(u.length<(B-1)*n+L)throw new Error("A does not have enough elements for the given dimensions and lda.");let C=v==="column-major"?a:i,R=v==="column-major"?i:a,F=e==="no-transpose"?C:R,W=e==="no-transpose"?R:C;if(l<W)throw new Error(`ldb must be >= ${v==="column-major"?"rows":"cols"} of B as stored.`);if(h){if(l!==f.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");let[J,sr]=e==="no-transpose"?[i,a]:[a,i];if(f.rows<J||f.cols<sr)throw new Error("B is too small for the given n, k, and transB.")}else if(f.length<(F-1)*l+W)throw new Error("B does not have enough elements for the given dimensions and ldb.");let V=_==="column-major"?a:o,z=_==="column-major"?o:a;if(c<z)throw new Error(`ldc must be >= ${_==="column-major"?"rows":"cols"} of C as stored.`);if(b){if(c!==p.lda)throw new Error("ldc must match C.lda when C is a GpuMatrix.");if(p.rows<o||p.cols<a)throw new Error("C is too small for the given m and n.")}else if(p.length<(V-1)*c+z)throw new Error("C does not have enough elements for the given dimensions and ldc.");y==="column-major"&&(t=t==="no-transpose"?"transpose":"no-transpose"),v==="column-major"&&(e=e==="no-transpose"?"transpose":"no-transpose"),_==="column-major"&&([u,f]=[f,u],[w,h]=[h,w],[n,l]=[l,n],[t,e]=[e==="no-transpose"?"transpose":"no-transpose",t==="no-transpose"?"transpose":"no-transpose"],[o,a]=[a,o]);let H=Math.ceil(a/64),$=Math.ceil(o/64),K=H*$>=36,Y=await D(r,K?"sgemm_large":"sgemm_small"),ir=w?u._buf:x(r,u,"sgemm-A",!1),lr=h?f._buf:x(r,f,"sgemm-B",!1),hr=b?p._buf:x(r,p,"sgemm-C",!0),nr=t==="no-transpose",mr=e==="no-transpose",Q=nr&&Ee(ir,n,o,i),er=Ee(lr,l,mr?i:a,mr?a:i),Z=I(r,[{value:o,type:"u32"},{value:a,type:"u32"},{value:i,type:"u32"},{value:s,type:"f32"},{value:m,type:"f32"},{value:n,type:"u32"},{value:l,type:"u32"},{value:c,type:"u32"},{value:t==="transpose"?1:0,type:"u32"},{value:e==="transpose"?1:0,type:"u32"},{value:Q?1:0,type:"u32"},{value:er?1:0,type:"u32"}],"sgemm-params");try{let J=E(r,Y.getBindGroupLayout(0),[ir,Ar(r,ir),lr,Ar(r,lr),hr,Z]),sr=K?{x:U(r,H,"sgemm","x"),y:U(r,$,"sgemm","y")}:{x:U(r,Math.ceil(a/32),"sgemm","x"),y:U(r,Math.ceil(o/32),"sgemm","y")},{commandEncoder:pr,ts:br}=j(r,Y,J,sr),fr=b?null:G(r,pr,hr);M(r,pr);let ur=await P(br);if(b)return ur!==void 0?{gpuTimeMs:ur}:{};let yr=await A(fr,Float32Array);return ur!==void 0?{C:yr,gpuTimeMs:ur}:{C:yr}}finally{w||d(ir),h||d(lr),b||d(hr),d(Z)}}async function ca(r,t,e,o,a,i,s,u,n,f,l,m,p,c,g,w="row-major"){let h=n instanceof X,b=l instanceof X,y=c instanceof X;if(q(r),T(r,"sgemmtr",{A:n,B:l,C:c}),t!=="lower"&&t!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(e!=="no-transpose"&&e!=="transpose")throw new Error("transA must be 'no-transpose' or 'transpose'.");if(o!=="no-transpose"&&o!=="transpose")throw new Error("transB must be 'no-transpose' or 'transpose'.");if(w!=="row-major"&&w!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(typeof u!="number")throw new Error("alpha must be a number.");if(Number.isNaN(u))throw new Error("alpha must not be NaN.");if(!Number.isFinite(u))throw new Error("alpha must be finite.");if(typeof p!="number")throw new Error("beta must be a number.");if(Number.isNaN(p))throw new Error("beta must not be NaN.");if(!Number.isFinite(p))throw new Error("beta must be finite.");if(!Number.isInteger(a)||!Number.isInteger(i)||!Number.isInteger(s)||!Number.isInteger(f)||!Number.isInteger(m)||!Number.isInteger(g))throw new Error("m, n, k, lda, ldb, and ldc must be integers.");if(!h&&!(n instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!b&&!(l instanceof Float32Array))throw new Error("B must be a Float32Array or GpuMatrix.");if(!y&&!(c instanceof Float32Array))throw new Error("C must be a Float32Array or GpuMatrix.");if((h||b)&&!y)throw new Error("C must be a GpuMatrix when A or B is a GpuMatrix.");if(y&&(!h||!b))throw new Error("A and B must be GpuMatrix when C is a GpuMatrix.");if(a<0||i<0||s<0)throw new Error("m, n, and k must be non-negative.");if(f<=0||m<=0||g<=0)throw new Error("lda, ldb, and ldc must be positive.");if(a===0||i===0)return y?{}:{C:c};let v=h?n.layout:w,_=b?l.layout:w,S=y?c.layout:w,k=v==="column-major"?s:a,B=v==="column-major"?a:s,L=e==="no-transpose"?k:B,C=e==="no-transpose"?B:k;if(f<C)throw new Error(`lda must be >= ${v==="column-major"?"rows":"cols"} of A as stored.`);if(h){if(f!==n.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");let[Q,er]=e==="no-transpose"?[a,s]:[s,a];if(n.rows<Q||n.cols<er)throw new Error("A is too small for the given m, k, and transA.")}else if(n.length<(L-1)*f+C)throw new Error("A does not have enough elements for the given dimensions and lda.");let R=_==="column-major"?i:s,F=_==="column-major"?s:i,W=o==="no-transpose"?R:F,V=o==="no-transpose"?F:R;if(m<V)throw new Error(`ldb must be >= ${_==="column-major"?"rows":"cols"} of B as stored.`);if(b){if(m!==l.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");let[Q,er]=o==="no-transpose"?[s,i]:[i,s];if(l.rows<Q||l.cols<er)throw new Error("B is too small for the given n, k, and transB.")}else if(l.length<(W-1)*m+V)throw new Error("B does not have enough elements for the given dimensions and ldb.");let z=S==="column-major"?i:a,H=S==="column-major"?a:i;if(g<H)throw new Error(`ldc must be >= ${S==="column-major"?"rows":"cols"} of C as stored.`);if(y){if(g!==c.lda)throw new Error("ldc must match C.lda when C is a GpuMatrix.");if(c.rows<a||c.cols<i)throw new Error("C is too small for the given m and n.")}else if(c.length<(z-1)*g+H)throw new Error("C does not have enough elements for the given dimensions and ldc.");v==="column-major"&&(e=e==="no-transpose"?"transpose":"no-transpose"),_==="column-major"&&(o=o==="no-transpose"?"transpose":"no-transpose"),S==="column-major"&&([n,l]=[l,n],[h,b]=[b,h],[f,m]=[m,f],[e,o]=[o==="no-transpose"?"transpose":"no-transpose",e==="no-transpose"?"transpose":"no-transpose"],[a,i]=[i,a],t=t==="lower"?"upper":"lower");let $=Math.ceil(i/64),K=Math.ceil(a/64),Y=$*K>=36,ir=await D(r,Y?"sgemmtr_large":"sgemmtr_small"),lr=h?n._buf:x(r,n,"sgemmtr-A",!1),hr=b?l._buf:x(r,l,"sgemmtr-B",!1),nr=y?c._buf:x(r,c,"sgemmtr-C",!0),mr=I(r,[{value:a,type:"u32"},{value:i,type:"u32"},{value:s,type:"u32"},{value:u,type:"f32"},{value:p,type:"f32"},{value:f,type:"u32"},{value:m,type:"u32"},{value:g,type:"u32"},{value:e==="transpose"?1:0,type:"u32"},{value:o==="transpose"?1:0,type:"u32"},{value:t==="upper"?1:0,type:"u32"}],"sgemmtr-params");try{let Q=E(r,ir.getBindGroupLayout(0),[lr,hr,nr,mr]),er=Y?{x:U(r,$,"sgemmtr","x"),y:U(r,K,"sgemmtr","y")}:{x:U(r,Math.ceil(i/32),"sgemmtr","x"),y:U(r,Math.ceil(a/32),"sgemmtr","y")},{commandEncoder:Z,ts:J}=j(r,ir,Q,er),sr=y?null:G(r,Z,nr);M(r,Z);let pr=await P(J);if(y)return pr!==void 0?{gpuTimeMs:pr}:{};let br=await A(sr,Float32Array);return pr!==void 0?{C:br,gpuTimeMs:pr}:{C:br}}finally{h||d(lr),b||d(hr),y||d(nr),d(mr)}}async function pa(r,t,e,o,a,i,s,u,n,f,l,m="row-major"){let p=s instanceof X,c=f instanceof X;if(q(r),T(r,"ssyrk",{A:s,C:f}),t!=="lower"&&t!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(e!=="no-transpose"&&e!=="transpose")throw new Error("trans must be 'no-transpose' or 'transpose'.");if(m!=="row-major"&&m!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(typeof i!="number")throw new Error("alpha must be a number.");if(Number.isNaN(i))throw new Error("alpha must not be NaN.");if(!Number.isFinite(i))throw new Error("alpha must be finite.");if(typeof n!="number")throw new Error("beta must be a number.");if(Number.isNaN(n))throw new Error("beta must not be NaN.");if(!Number.isFinite(n))throw new Error("beta must be finite.");if(!Number.isInteger(o)||!Number.isInteger(a)||!Number.isInteger(u)||!Number.isInteger(l))throw new Error("n, k, lda, and ldc must be integers.");if(!p&&!(s instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!c&&!(f instanceof Float32Array))throw new Error("C must be a Float32Array or GpuMatrix.");if(p&&!c)throw new Error("C must be a GpuMatrix when A is a GpuMatrix.");if(c&&!p)throw new Error("A must be a GpuMatrix when C is a GpuMatrix.");if(o<0||a<0)throw new Error("n and k must be non-negative.");if(u<=0||l<=0)throw new Error("lda and ldc must be positive.");if(o===0)return c?{}:{C:f};let g=p?s.layout:m,w=c?f.layout:m,h=g==="column-major"?a:o,b=g==="column-major"?o:a,y=e==="no-transpose"?h:b,v=e==="no-transpose"?b:h;if(u<v)throw new Error(`lda must be >= ${g==="column-major"?"rows":"cols"} of A as stored.`);if(p){if(u!==s.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");let[H,$]=e==="no-transpose"?[o,a]:[a,o];if(s.rows<H||s.cols<$)throw new Error("A is too small for the given n, k, and trans.")}else if(s.length<(y-1)*u+v)throw new Error("A does not have enough elements for the given dimensions and lda.");if(l<o)throw new Error("ldc must be >= n.");if(c){if(l!==f.lda)throw new Error("ldc must match C.lda when C is a GpuMatrix.");if(f.rows<o||f.cols<o)throw new Error("C is too small for the given n.")}else if(f.length<(o-1)*l+o)throw new Error("C does not have enough elements for the given dimensions and ldc.");let _=e;g==="column-major"&&(_=_==="no-transpose"?"transpose":"no-transpose");let S=_==="no-transpose"?"transpose":"no-transpose",k=t;w==="column-major"&&([_,S]=[S==="no-transpose"?"transpose":"no-transpose",_==="no-transpose"?"transpose":"no-transpose"],k=k==="lower"?"upper":"lower");let B=Math.ceil(o/64),L=Math.ceil(o/64),C=B*L>=36,R=await D(r,C?"sgemmtr_large":"sgemmtr_small"),F=p?s._buf:x(r,s,"ssyrk-A",!1),W=c?f._buf:x(r,f,"ssyrk-C",!0),V=p?tr(r,F.size,"ssyrk-B",GPUBufferUsage.COPY_DST):x(r,s,"ssyrk-B",!1),z=I(r,[{value:o,type:"u32"},{value:o,type:"u32"},{value:a,type:"u32"},{value:i,type:"f32"},{value:n,type:"f32"},{value:u,type:"u32"},{value:u,type:"u32"},{value:l,type:"u32"},{value:_==="transpose"?1:0,type:"u32"},{value:S==="transpose"?1:0,type:"u32"},{value:k==="upper"?1:0,type:"u32"}],"ssyrk-params");try{let H=E(r,R.getBindGroupLayout(0),[F,V,W,z]),$=C?{x:U(r,B,"ssyrk","x"),y:U(r,L,"ssyrk","y")}:{x:U(r,Math.ceil(o/32),"ssyrk","x"),y:U(r,Math.ceil(o/32),"ssyrk","y")},{commandEncoder:K,querySet:Y,passDescriptor:ir}=qr(r);p&&K.copyBufferToBuffer(F,0,V,0,F.size),gr(K,R,H,$,ir);let lr=Lr(r,K,Y),hr=c?null:G(r,K,W);M(r,K);let nr=await P(lr);if(c)return nr!==void 0?{gpuTimeMs:nr}:{};let mr=await A(hr,Float32Array);return nr!==void 0?{C:mr,gpuTimeMs:nr}:{C:mr}}finally{p||d(F),d(V),c||d(W),d(z)}}async function ga(r,t,e,o,a,i,s,u,n,f,l,m,p,c="row-major"){let g=s instanceof X,w=n instanceof X,h=m instanceof X;if(q(r),T(r,"ssyr2k",{A:s,B:n,C:m}),t!=="lower"&&t!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(e!=="no-transpose"&&e!=="transpose")throw new Error("trans must be 'no-transpose' or 'transpose'.");if(c!=="row-major"&&c!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(typeof i!="number")throw new Error("alpha must be a number.");if(Number.isNaN(i))throw new Error("alpha must not be NaN.");if(!Number.isFinite(i))throw new Error("alpha must be finite.");if(typeof l!="number")throw new Error("beta must be a number.");if(Number.isNaN(l))throw new Error("beta must not be NaN.");if(!Number.isFinite(l))throw new Error("beta must be finite.");if(!Number.isInteger(o)||!Number.isInteger(a)||!Number.isInteger(u)||!Number.isInteger(f)||!Number.isInteger(p))throw new Error("n, k, lda, ldb, and ldc must be integers.");if(!g&&!(s instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!w&&!(n instanceof Float32Array))throw new Error("B must be a Float32Array or GpuMatrix.");if(!h&&!(m instanceof Float32Array))throw new Error("C must be a Float32Array or GpuMatrix.");if((g||w)&&!h)throw new Error("C must be a GpuMatrix when A or B is a GpuMatrix.");if(h&&(!g||!w))throw new Error("A and B must be GpuMatrix when C is a GpuMatrix.");if(o<0||a<0)throw new Error("n and k must be non-negative.");if(u<=0||f<=0||p<=0)throw new Error("lda, ldb, and ldc must be positive.");if(o===0)return h?{}:{C:m};let b=g?s.layout:c,y=w?n.layout:c,v=h?m.layout:c,_=b==="column-major"?a:o,S=b==="column-major"?o:a,k=e==="no-transpose"?_:S,B=e==="no-transpose"?S:_;if(u<B)throw new Error(`lda must be >= ${b==="column-major"?"rows":"cols"} of A as stored.`);if(g){if(u!==s.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");let[J,sr]=e==="no-transpose"?[o,a]:[a,o];if(s.rows<J||s.cols<sr)throw new Error("A is too small for the given n, k, and trans.")}else if(s.length<(k-1)*u+B)throw new Error("A does not have enough elements for the given dimensions and lda.");let L=y==="column-major"?a:o,C=y==="column-major"?o:a,R=e==="no-transpose"?L:C,F=e==="no-transpose"?C:L;if(f<F)throw new Error(`ldb must be >= ${y==="column-major"?"rows":"cols"} of B as stored.`);if(w){if(f!==n.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");let[J,sr]=e==="no-transpose"?[o,a]:[a,o];if(n.rows<J||n.cols<sr)throw new Error("B is too small for the given n, k, and trans.")}else if(n.length<(R-1)*f+F)throw new Error("B does not have enough elements for the given dimensions and ldb.");if(p<o)throw new Error("ldc must be >= n.");if(h){if(p!==m.lda)throw new Error("ldc must match C.lda when C is a GpuMatrix.");if(m.rows<o||m.cols<o)throw new Error("C is too small for the given n.")}else if(m.length<(o-1)*p+o)throw new Error("C does not have enough elements for the given dimensions and ldc.");let W=e;b==="column-major"&&(W=W==="no-transpose"?"transpose":"no-transpose");let V=e;y==="column-major"&&(V=V==="no-transpose"?"transpose":"no-transpose");let z=v==="column-major"?t==="lower"?"upper":"lower":t,H=J=>J==="no-transpose"?"transpose":"no-transpose";function $(J,sr,pr,br,fr,ur){let yr=J,Cr=H(br);return v!=="column-major"?{transX:yr,X:sr,ldX:pr,transY:Cr,Y:fr,ldY:ur}:{transX:H(Cr),X:fr,ldX:ur,transY:H(yr),Y:sr,ldY:pr}}let K=Math.ceil(o/64),Y=Math.ceil(o/64),ir=K*Y>=36,lr=await D(r,ir?"sgemmtr_large":"sgemmtr_small"),hr=ir?{x:U(r,K,"ssyr2k","x"),y:U(r,Y,"ssyr2k","y")}:{x:U(r,Math.ceil(o/32),"ssyr2k","x"),y:U(r,Math.ceil(o/32),"ssyr2k","y")},nr=g?s._buf:x(r,s,"ssyr2k-A",!1),mr=w?n._buf:x(r,n,"ssyr2k-B",!1),Q=h?m._buf:x(r,m,"ssyr2k-C",!0),er=null,Z=null;try{let J=$(W,nr,u,V,mr,f),sr=$(V,mr,f,W,nr,u),pr=(Ir,Sr)=>I(r,[{value:o,type:"u32"},{value:o,type:"u32"},{value:a,type:"u32"},{value:i,type:"f32"},{value:Sr,type:"f32"},{value:Ir.ldX,type:"u32"},{value:Ir.ldY,type:"u32"},{value:p,type:"u32"},{value:Ir.transX==="transpose"?1:0,type:"u32"},{value:Ir.transY==="transpose"?1:0,type:"u32"},{value:z==="upper"?1:0,type:"u32"}],"ssyr2k-params");er=pr(J,l),Z=pr(sr,1);let br=E(r,lr.getBindGroupLayout(0),[J.X,J.Y,Q,er]),fr=E(r,lr.getBindGroupLayout(0),[sr.X,sr.Y,Q,Z]),{commandEncoder:ur,querySet:yr}=qr(r),Cr=yr?{timestampWrites:{querySet:yr,beginningOfPassWriteIndex:0}}:void 0,Rr=yr?{timestampWrites:{querySet:yr,endOfPassWriteIndex:1}}:void 0;gr(ur,lr,br,hr,Cr),gr(ur,lr,fr,hr,Rr);let Tr=Lr(r,ur,yr),Er=h?null:G(r,ur,Q);M(r,ur);let vr=await P(Tr);if(h)return vr!==void 0?{gpuTimeMs:vr}:{};let xr=await A(Er,Float32Array);return vr!==void 0?{C:xr,gpuTimeMs:vr}:{C:xr}}finally{g||d(nr),w||d(mr),h||d(Q),er&&d(er),Z&&d(Z)}}async function wa(r,t,e,o,a,i,s,u,n,f,l,m,p,c="row-major"){let g=s instanceof X,w=n instanceof X,h=m instanceof X;if(q(r),T(r,"ssymm",{A:s,B:n,C:m}),t!=="left"&&t!=="right")throw new Error("side must be 'left' or 'right'.");if(e!=="lower"&&e!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(c!=="row-major"&&c!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(typeof i!="number")throw new Error("alpha must be a number.");if(Number.isNaN(i))throw new Error("alpha must not be NaN.");if(!Number.isFinite(i))throw new Error("alpha must be finite.");if(typeof l!="number")throw new Error("beta must be a number.");if(Number.isNaN(l))throw new Error("beta must not be NaN.");if(!Number.isFinite(l))throw new Error("beta must be finite.");if(!Number.isInteger(o)||!Number.isInteger(a)||!Number.isInteger(u)||!Number.isInteger(f)||!Number.isInteger(p))throw new Error("m, n, lda, ldb, and ldc must be integers.");if(!g&&!(s instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!w&&!(n instanceof Float32Array))throw new Error("B must be a Float32Array or GpuMatrix.");if(!h&&!(m instanceof Float32Array))throw new Error("C must be a Float32Array or GpuMatrix.");if((g||w)&&!h)throw new Error("C must be a GpuMatrix when A or B is a GpuMatrix.");if(h&&(!g||!w))throw new Error("A and B must be GpuMatrix when C is a GpuMatrix.");if(o<0||a<0)throw new Error("m and n must be non-negative.");if(o===0||a===0)return h?{}:{C:m};let b=g?s.layout:c,y=w?n.layout:c,v=h?m.layout:c,_=t==="left"?o:a;if(u<_)throw new Error("lda must be >= "+(t==="left"?"m":"n")+".");if(g){if(u!==s.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(s.rows<_||s.cols<_)throw new Error("A is too small for the given m/n and side.")}else if(s.length<(_-1)*u+_)throw new Error("A does not have enough elements for the given dimensions and lda.");let S=y==="column-major"?a:o,k=y==="column-major"?o:a;if(f<k)throw new Error(`ldb must be >= ${y==="column-major"?"rows":"cols"} of B as stored.`);if(w){if(f!==n.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");if(n.rows<o||n.cols<a)throw new Error("B is too small for the given m and n.")}else if(n.length<(S-1)*f+k)throw new Error("B does not have enough elements for the given dimensions and ldb.");let B=v==="column-major"?a:o,L=v==="column-major"?o:a;if(p<L)throw new Error(`ldc must be >= ${v==="column-major"?"rows":"cols"} of C as stored.`);if(h){if(p!==m.lda)throw new Error("ldc must match C.lda when C is a GpuMatrix.");if(m.rows<o||m.cols<a)throw new Error("C is too small for the given m and n.")}else if(m.length<(B-1)*p+L)throw new Error("C does not have enough elements for the given dimensions and ldc.");let C=b==="column-major"?e==="lower"?"upper":"lower":e,R=y==="column-major"?"transpose":"no-transpose",F="no-transpose",W=o,V=a,z=_,H=t==="left"?F:R,$=t==="left"?R:F,K=ur=>ur==="no-transpose"?"transpose":"no-transpose",Y=t==="right";v==="column-major"&&([H,$]=[K($),K(H)],Y=!Y,[W,V]=[V,W]);let ir=_,lr=Math.ceil(V/64),hr=Math.ceil(W/64),nr=lr*hr>=36,mr=await D(r,nr?"sgemm_large":"sgemm_small"),Q=await D(r,"symmetrize"),er=nr?{x:U(r,lr,"ssymm","x"),y:U(r,hr,"ssymm","y")}:{x:U(r,Math.ceil(V/32),"ssymm","x"),y:U(r,Math.ceil(W/32),"ssymm","y")},Z=g?s._buf:x(r,s,"ssymm-A",!1),J=w?n._buf:x(r,n,"ssymm-B",!1),sr=h?m._buf:x(r,m,"ssymm-C",!0),pr=tr(r,_*ir*4,"ssymm-Adense"),br=null,fr=null;try{br=I(r,[{value:_,type:"u32"},{value:u,type:"u32"},{value:ir,type:"u32"},{value:C==="upper"?1:0,type:"u32"}],"ssymm-sym-params");let ur=E(r,Q.getBindGroupLayout(0),[Z,pr,br]),yr=Y?J:pr,Cr=Y?f:ir,Rr=Y?pr:J;fr=I(r,[{value:W,type:"u32"},{value:V,type:"u32"},{value:z,type:"u32"},{value:i,type:"f32"},{value:l,type:"f32"},{value:Cr,type:"u32"},{value:Y?ir:f,type:"u32"},{value:p,type:"u32"},{value:H==="transpose"?1:0,type:"u32"},{value:$==="transpose"?1:0,type:"u32"}],"ssymm-gemm-params");let Er=E(r,mr.getBindGroupLayout(0),[yr,Ar(r,yr),Rr,Ar(r,Rr),sr,fr]),{commandEncoder:vr,querySet:xr}=qr(r),Ir=xr?{timestampWrites:{querySet:xr,beginningOfPassWriteIndex:0}}:void 0,Sr=xr?{timestampWrites:{querySet:xr,endOfPassWriteIndex:1}}:void 0;gr(vr,Q,ur,{x:Math.ceil(_/8),y:Math.ceil(_/8)},Ir),gr(vr,mr,Er,er,Sr);let Fr=Lr(r,vr,xr),Or=h?null:G(r,vr,sr);M(r,vr);let Ur=await P(Fr);if(h)return Ur!==void 0?{gpuTimeMs:Ur}:{};let me=await A(Or,Float32Array);return Ur!==void 0?{C:me,gpuTimeMs:Ur}:{C:me}}finally{g||d(Z),w||d(J),h||d(sr),d(pr),br&&d(br),fr&&d(fr)}}async function ha(r,t,e,o,a,i,s,u,n,f,l,m,p="row-major"){let c=n instanceof X,g=l instanceof X,w=a==="unit";if(q(r),T(r,"strmm",{A:n,B:l}),t!=="left"&&t!=="right")throw new Error("side must be 'left' or 'right'.");if(e!=="lower"&&e!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(o!=="no-transpose"&&o!=="transpose")throw new Error("transA must be 'no-transpose' or 'transpose'.");if(!w&&a!=="non-unit")throw new Error("diag must be 'unit' or 'non-unit'.");if(p!=="row-major"&&p!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(typeof u!="number")throw new Error("alpha must be a number.");if(Number.isNaN(u))throw new Error("alpha must not be NaN.");if(!Number.isFinite(u))throw new Error("alpha must be finite.");if(!Number.isInteger(i)||!Number.isInteger(s)||!Number.isInteger(f)||!Number.isInteger(m))throw new Error("m, n, lda, and ldb must be integers.");if(!c&&!(n instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!g&&!(l instanceof Float32Array))throw new Error("B must be a Float32Array or GpuMatrix.");if(c!==g)throw new Error("A and B must both be GpuMatrix or both be Float32Array.");if(i<0||s<0)throw new Error("m and n must be non-negative.");if(i===0||s===0)return g?{}:{B:l};let h=c?n.layout:p,b=g?l.layout:p,y=t==="left"?i:s;if(f<y)throw new Error("lda must be >= "+(t==="left"?"m":"n")+".");if(c){if(f!==n.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(n.rows<y||n.cols<y)throw new Error("A is too small for the given m/n and side.")}else if(n.length<(y-1)*f+y)throw new Error("A does not have enough elements for the given dimensions and lda.");let v=b==="column-major"?s:i,_=b==="column-major"?i:s;if(m<_)throw new Error(`ldb must be >= ${b==="column-major"?"rows":"cols"} of B as stored.`);if(g){if(m!==l.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");if(l.rows<i||l.cols<s)throw new Error("B is too small for the given m and n.")}else if(l.length<(v-1)*m+_)throw new Error("B does not have enough elements for the given dimensions and ldb.");let S=h==="column-major"?e==="lower"?"upper":"lower":e,k=h==="column-major"?o==="no-transpose"?"transpose":"no-transpose":o,B=b==="column-major"?"transpose":"no-transpose",L="no-transpose",C=i,R=s,F=y,W=t==="left"?L:B,V=t==="left"?B:L,z=br=>br==="no-transpose"?"transpose":"no-transpose",H=t==="right";b==="column-major"&&([W,V]=[z(V),z(W)],H=!H,[C,R]=[R,C]);let $=y,K=Math.ceil(R/64),Y=Math.ceil(C/64),ir=K*Y>=36,lr=await D(r,ir?"sgemm_large":"sgemm_small"),hr=await D(r,"triangularize"),nr=ir?{x:U(r,K,"strmm","x"),y:U(r,Y,"strmm","y")}:{x:U(r,Math.ceil(R/32),"strmm","x"),y:U(r,Math.ceil(C/32),"strmm","y")},mr=null,Q=null,er=null,Z=null,J=null,sr=null,pr=!1;try{mr=c?n._buf:x(r,n,"strmm-A",!1),Q=g?l._buf:x(r,l,"strmm-B",!0),er=tr(r,y*$*4,"strmm-Adense"),Z=tr(r,v*m*4,"strmm-out",GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST),J=I(r,[{value:y,type:"u32"},{value:f,type:"u32"},{value:$,type:"u32"},{value:S==="upper"?1:0,type:"u32"},{value:k==="transpose"?1:0,type:"u32"},{value:w?1:0,type:"u32"}],"strmm-tri-params");let br=E(r,hr.getBindGroupLayout(0),[mr,er,J]),fr=H?Q:er,ur=H?m:$,yr=H?er:Q;sr=I(r,[{value:C,type:"u32"},{value:R,type:"u32"},{value:F,type:"u32"},{value:u,type:"f32"},{value:0,type:"f32"},{value:ur,type:"u32"},{value:H?$:m,type:"u32"},{value:m,type:"u32"},{value:W==="transpose"?1:0,type:"u32"},{value:V==="transpose"?1:0,type:"u32"}],"strmm-gemm-params");let Rr=E(r,lr.getBindGroupLayout(0),[fr,Ar(r,fr),yr,Ar(r,yr),Z,sr]),{commandEncoder:Tr,querySet:Er}=qr(r);Tr.copyBufferToBuffer(Q,0,Z,0,Math.min(Q.size,Z.size));let vr=Er?{timestampWrites:{querySet:Er,beginningOfPassWriteIndex:0}}:void 0,xr=Er?{timestampWrites:{querySet:Er,endOfPassWriteIndex:1}}:void 0;gr(Tr,hr,br,{x:Math.ceil(y/8),y:Math.ceil(y/8)},vr),gr(Tr,lr,Rr,nr,xr);let Ir=Lr(r,Tr,Er),Sr=g?null:G(r,Tr,Z);M(r,Tr);let Fr=await P(Ir);if(g)return d(l._buf),l._buf=Z,pr=!0,Fr!==void 0?{gpuTimeMs:Fr}:{};let Or=await A(Sr,Float32Array);return Fr!==void 0?{B:Or,gpuTimeMs:Fr}:{B:Or}}finally{!c&&mr&&d(mr),!g&&Q&&d(Q),er&&d(er),Z&&!pr&&d(Z),J&&d(J),sr&&d(sr)}}async function ba(r,t,e,o,a,i,s,u,n,f,l,m,p="row-major"){let c=n instanceof X,g=l instanceof X,w=a==="unit";if(q(r),T(r,"strsm",{A:n,B:l}),t!=="left"&&t!=="right")throw new Error("side must be 'left' or 'right'.");if(e!=="lower"&&e!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(o!=="no-transpose"&&o!=="transpose")throw new Error("transA must be 'no-transpose' or 'transpose'.");if(!w&&a!=="non-unit")throw new Error("diag must be 'unit' or 'non-unit'.");if(p!=="row-major"&&p!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(typeof u!="number")throw new Error("alpha must be a number.");if(Number.isNaN(u))throw new Error("alpha must not be NaN.");if(!Number.isFinite(u))throw new Error("alpha must be finite.");if(!Number.isInteger(i)||!Number.isInteger(s)||!Number.isInteger(f)||!Number.isInteger(m))throw new Error("m, n, lda, and ldb must be integers.");if(!c&&!(n instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!g&&!(l instanceof Float32Array))throw new Error("B must be a Float32Array or GpuMatrix.");if(c!==g)throw new Error("A and B must both be GpuMatrix or both be Float32Array.");if(i<0||s<0)throw new Error("m and n must be non-negative.");if(i===0||s===0)return g?{}:{B:l};let h=c?n.layout:p,b=g?l.layout:p,y=t==="left"?i:s;if(f<y)throw new Error("lda must be >= "+(t==="left"?"m":"n")+".");if(c){if(f!==n.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(n.rows<y||n.cols<y)throw new Error("A is too small for the given m/n and side.")}else if(n.length<(y-1)*f+y)throw new Error("A does not have enough elements for the given dimensions and lda.");let v=b==="column-major"?s:i,_=b==="column-major"?i:s;if(m<_)throw new Error(`ldb must be >= ${b==="column-major"?"rows":"cols"} of B as stored.`);if(g){if(m!==l.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");if(l.rows<i||l.cols<s)throw new Error("B is too small for the given m and n.")}else if(l.length<(v-1)*m+_)throw new Error("B does not have enough elements for the given dimensions and ldb.");let S=h==="column-major"?e==="lower"?"upper":"lower":e,k=h==="column-major"?o==="no-transpose"?"transpose":"no-transpose":o,B=t==="left"?s:i,L=t==="left",C=k==="no-transpose"==(S==="lower"),R=t==="left"?C:!C,F=[];for(let Q=0;Q<y;Q+=64)F.push(Q);R||F.reverse();let W=F.length,V=await D(r,"strsv_invert_block"),z=await D(r,"block_transfer"),H=await D(r,"sscal"),$=null,K=null,Y=null,ir=[],lr=[];function hr(Q,er){let Z=tr(r,Q,er);return lr.push(Z),Z}function nr(Q,er){let Z=I(r,Q,er);return ir.push(Z),Z}let mr=(v-1)*m+_;try{$=c?n._buf:x(r,n,"strsm-A",!1),K=g?l._buf:x(r,l,"strsm-B",!0),Y=tr(r,W*64*64*4,"strsm-Ainv");let Q=null;if(u!==1&&u!==0){let Er=nr([{value:mr,type:"u32"},{value:u,type:"f32"},{value:1,type:"u32"}],"strsm-scale-params");Q=E(r,H.getBindGroupLayout(0),[K,Er])}let er=nr([{value:y,type:"u32"},{value:f,type:"u32"},{value:k==="transpose"?1:0,type:"u32"},{value:S==="upper"?1:0,type:"u32"},{value:w?1:0,type:"u32"}],"strsm-invert-params"),Z=E(r,V.getBindGroupLayout(0),[$,Y,er]),J=hr(64*B*4,"strsm-Bblock"),sr=hr(64*B*4,"strsm-Xblock"),pr=hr(y*64*4,"strsm-Aoff"),br=hr(y*B*4,"strsm-delta"),{commandEncoder:fr,querySet:ur}=qr(r);if(u===0){let Er=Math.ceil(_/64),vr=Math.ceil(v/64),xr=Er*vr>=36,Ir=await D(r,xr?"sgemm_large":"sgemm_small"),Sr=nr([{value:v,type:"u32"},{value:_,type:"u32"},{value:0,type:"u32"},{value:0,type:"f32"},{value:0,type:"f32"},{value:1,type:"u32"},{value:1,type:"u32"},{value:m,type:"u32"},{value:0,type:"u32"},{value:0,type:"u32"}],"strsm-zero-params"),Fr=E(r,Ir.getBindGroupLayout(0),[Y,Ar(r,Y),Y,Ar(r,Y),K,Sr]),Or=xr?{x:U(r,Er,"strsm","x"),y:U(r,vr,"strsm","y")}:{x:U(r,Math.ceil(_/32),"strsm","x"),y:U(r,Math.ceil(v/32),"strsm","y")};gr(fr,Ir,Fr,Or,ur?{timestampWrites:{querySet:ur,beginningOfPassWriteIndex:0,endOfPassWriteIndex:1}}:void 0)}else{Q&&gr(fr,H,Q,cr(r,mr)),gr(fr,V,Z,{x:64,y:W},ur?{timestampWrites:{querySet:ur,beginningOfPassWriteIndex:0}}:void 0);for(let vr=0;vr<F.length;vr++){let xr=F[vr],Ir=Math.min(xr+64,y),Sr=Ir-xr,Fr=xr/64,Or=vr===F.length-1,Ur=nr([{value:xr,type:"u32"},{value:Sr,type:"u32"},{value:0,type:"u32"},{value:B,type:"u32"},{value:m,type:"u32"},{value:b==="column-major"?1:0,type:"u32"},{value:L?1:0,type:"u32"},{value:2,type:"u32"}],"strsm-gather-B-params"),me=E(r,z.getBindGroupLayout(0),[J,K,Ur]);gr(fr,z,me,Zr(r,"strsm",Sr,B));{let Yr=Sr,Xr=B,_e=Sr,ae=Math.ceil(Xr/64),ie=Math.ceil(Yr/64),se=ae*ie>=36,ne=await D(r,se?"sgemm_large":"sgemm_small"),Be=nr([{value:Yr,type:"u32"},{value:Xr,type:"u32"},{value:_e,type:"u32"},{value:1,type:"f32"},{value:0,type:"f32"},{value:64,type:"u32"},{value:B,type:"u32"},{value:B,type:"u32"},{value:t==="right"?1:0,type:"u32"},{value:0,type:"u32"}],"strsm-apply-params"),ce={buffer:Y,offset:Fr*64*64*4,size:4096*4},Se=E(r,ne.getBindGroupLayout(0),[ce,Ar(r,ce),J,Ar(r,J),sr,Be]),Ea=se?{x:U(r,ae,"strsm","x"),y:U(r,ie,"strsm","y")}:{x:U(r,Math.ceil(Xr/32),"strsm","x"),y:U(r,Math.ceil(Yr/32),"strsm","y")};gr(fr,ne,Se,Ea)}let de=R?Ir:0,Le=R?y:xr,Re=de<Le,ya=nr([{value:xr,type:"u32"},{value:Sr,type:"u32"},{value:0,type:"u32"},{value:B,type:"u32"},{value:m,type:"u32"},{value:b==="column-major"?1:0,type:"u32"},{value:L?1:0,type:"u32"},{value:0,type:"u32"}],"strsm-scatter-params"),xa=E(r,z.getBindGroupLayout(0),[sr,K,ya]),va=Or&&!Re&&ur?{timestampWrites:{querySet:ur,endOfPassWriteIndex:1}}:void 0;if(gr(fr,z,xa,Zr(r,"strsm",Sr,B),va),!Re)continue;let oe=Le-de,_a=nr([{value:de,type:"u32"},{value:oe,type:"u32"},{value:xr,type:"u32"},{value:Sr,type:"u32"},{value:f,type:"u32"},{value:k==="transpose"?1:0,type:"u32"},{value:L?1:0,type:"u32"},{value:2,type:"u32"}],"strsm-gather-A-params"),Ba=E(r,z.getBindGroupLayout(0),[pr,$,_a]);gr(fr,z,Ba,Zr(r,"strsm",oe,Sr));{let Yr=oe,Xr=B,_e=Sr,ae=Math.ceil(Xr/64),ie=Math.ceil(Yr/64),se=ae*ie>=36,ne=await D(r,se?"sgemm_large":"sgemm_small"),Be=nr([{value:Yr,type:"u32"},{value:Xr,type:"u32"},{value:_e,type:"u32"},{value:1,type:"f32"},{value:0,type:"f32"},{value:Sr,type:"u32"},{value:B,type:"u32"},{value:B,type:"u32"},{value:0,type:"u32"},{value:0,type:"u32"}],"strsm-update-params"),ce=E(r,ne.getBindGroupLayout(0),[pr,Ar(r,pr),sr,Ar(r,sr),br,Be]),Se=se?{x:U(r,ae,"strsm","x"),y:U(r,ie,"strsm","y")}:{x:U(r,Math.ceil(Xr/32),"strsm","x"),y:U(r,Math.ceil(Yr/32),"strsm","y")};gr(fr,ne,ce,Se)}let Sa=nr([{value:de,type:"u32"},{value:oe,type:"u32"},{value:0,type:"u32"},{value:B,type:"u32"},{value:m,type:"u32"},{value:b==="column-major"?1:0,type:"u32"},{value:L?1:0,type:"u32"},{value:1,type:"u32"}],"strsm-scatter-sub-params"),Aa=E(r,z.getBindGroupLayout(0),[br,K,Sa]),Ga=Or&&ur?{timestampWrites:{querySet:ur,endOfPassWriteIndex:1}}:void 0;gr(fr,z,Aa,Zr(r,"strsm",oe,B),Ga)}}let yr=Lr(r,fr,ur),Cr=g?null:G(r,fr,K);M(r,fr);let Rr=await P(yr);if(g)return Rr!==void 0?{gpuTimeMs:Rr}:{};let Tr=await A(Cr,Float32Array);return Rr!==void 0?{B:Tr,gpuTimeMs:Rr}:{B:Tr}}finally{!c&&$&&d($),!g&&K&&d(K),Y&&d(Y),d(lr),d(ir)}}return Ia(Ti);})();

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "wgblas",
-  "version": "2.2.0",
+  "version": "2.2.1",
   "description": "BLAS on WebGPU",
   "type": "module",
   "main": "index.mjs",

package/src/shaders/f64/utils/multiply.wgsl

@@ -84,7 +84,16 @@ fn ddMulRaw(a: DD, b: DD, threadSlot: u32) -> DD {
   return DD(p.hi, crossAndLo);
 }
 
+// A third compiler bug (Intel Mesa ANV, via NIR dump): raw.hi never gets
+// materialized as one rounded value — the driver re-fuses a.hi*b.hi with
+// ffma at every use site instead, breaking the a+b == s+e identity
+// TwoSum-style algorithms depend on. Same fix as ddMulRaw's p.lo: force it
+// through workgroup memory + a barrier. (Verified: max forward-error factor
+// over 20000 trials dropped from >1e5 to ~3-4, Intel Mesa Iris Xe + NVIDIA GTX 1650.)
 fn ddMulProtected(a: DD, b: DD, threadSlot: u32) -> DD {
   let raw = ddMulRaw(a, b, threadSlot);
-  return fastTwoSumProtected(raw.hi, raw.lo, threadSlot);
+  dekkerScratch[threadSlot] = raw.hi;
+  workgroupBarrier();
+  let rawHi = dekkerScratch[threadSlot];
+  return fastTwoSumProtected(rawHi, raw.lo, threadSlot);
 }