wgblas 2.2.1 → 2.3.0

package/dist/wgblas.browser.js

@@ -1,4 +1,4 @@
-var wgblas=(()=>{var ka=Object.create;var pe=Object.defineProperty;var Da=Object.getOwnPropertyDescriptor;var Na=Object.getOwnPropertyNames;var Pa=Object.getPrototypeOf,Ma=Object.prototype.hasOwnProperty;var ge=(r=>typeof require<"u"?require:typeof Proxy<"u"?new Proxy(r,{get:(t,e)=>(typeof require<"u"?require:t)[e]}):r)(function(r){if(typeof require<"u")return require.apply(this,arguments);throw Error('Dynamic require of "'+r+'" is not supported')});var O=(r,t,e)=>()=>{if(e)throw e[0];try{return r&&(t=r(r=0)),t}catch(o){throw e=[o],o}};var qe=(r,t)=>{for(var e in t)pe(r,e,{get:t[e],enumerable:!0})},Te=(r,t,e,o)=>{if(t&&typeof t=="object"||typeof t=="function")for(let a of Na(t))!Ma.call(r,a)&&a!==e&&pe(r,a,{get:()=>t[a],enumerable:!(o=Da(t,a))||o.enumerable});return r};var we=(r,t,e)=>(e=r!=null?ka(Pa(r)):{},Te(t||!r||!r.__esModule?pe(e,"default",{value:r,enumerable:!0}):e,r)),Ia=r=>Te(pe({},"__esModule",{value:!0}),r);var ke,$e=O(()=>{ke=`// sscal: x = alpha * x
+var wgblas=(()=>{var oi=Object.create;var pe=Object.defineProperty;var ai=Object.getOwnPropertyDescriptor;var ii=Object.getOwnPropertyNames;var si=Object.getPrototypeOf,ni=Object.prototype.hasOwnProperty;var we=(r=>typeof require<"u"?require:typeof Proxy<"u"?new Proxy(r,{get:(o,e)=>(typeof require<"u"?require:o)[e]}):r)(function(r){if(typeof require<"u")return require.apply(this,arguments);throw Error('Dynamic require of "'+r+'" is not supported')});var Z=(r,o,e)=>()=>{if(e)throw e[0];try{return r&&(o=r(r=0)),o}catch(a){throw e=[a],a}};var Fe=(r,o)=>{for(var e in o)pe(r,e,{get:o[e],enumerable:!0})},qe=(r,o,e,a)=>{if(o&&typeof o=="object"||typeof o=="function")for(let t of ii(o))!ni.call(r,t)&&t!==e&&pe(r,t,{get:()=>o[t],enumerable:!(a=ai(o,t))||a.enumerable});return r};var ge=(r,o,e)=>(e=r!=null?oi(si(r)):{},qe(o||!r||!r.__esModule?pe(e,"default",{value:r,enumerable:!0}):e,r)),li=r=>qe(pe({},"__esModule",{value:!0}),r);var ke,$e=Z(()=>{ke=`// sscal: x = alpha * x
 
 @group(0) @binding(0) var<storage, read_write> x: array<f32>;
 
@@ -21,7 +21,7 @@ fn main(
     x[id * params.x_inc] = params.alpha * x[id * params.x_inc];
   }
 }
-`});var Qe,Ze=O(()=>{Qe=`// cscal: x := alpha * x, complex. x is one interleaved f32 array
+`});var Qe,Je=Z(()=>{Qe=`// cscal: x := alpha * x, complex. x is one interleaved f32 array
 // (re0, im0, re1, im1, ...), matching Complex32Array/GpuVector's storage
 // (and cuBLAS's cuComplex / stdlib's Complex64Array) \u2014 no repacking needed
 // between JS and GPU.
@@ -54,7 +54,7 @@ fn main(
     x[base + 1u] = params.alphaRe * im + params.alphaIm * re;
   }
 }
-`});var rt,Je=O(()=>{rt=`// sswap: x <-> y
+`});var et,rt=Z(()=>{et=`// sswap: x <-> y
 
 @group(0) @binding(0) var<storage, read_write> x: array<f32>;
 @group(0) @binding(1) var<storage, read_write> y: array<f32>;
@@ -80,7 +80,7 @@ fn main(
     y[id * params.y_inc] = temp;
   }
 }
-`});var tt,et=O(()=>{tt=`// dswap: x <-> y, double-double (Dekker) f64 emulation of sswap. A swap is
+`});var ot,tt=Z(()=>{ot=`// dswap: x <-> y, double-double (Dekker) f64 emulation of sswap. A swap is
 // pure data movement \u2014 hi and lo are exchanged verbatim, with no arithmetic
 // at all \u2014 so (unlike dscal/daxpy/ddot) this needs no
 // ddMulProtected/ddAddProtected renormalizing barrier, and so no
@@ -118,7 +118,7 @@ fn main(
     yLo[iy] = tempLo;
   }
 }
-`});var at,ot=O(()=>{at=`// saxpy: y = alpha * x + y
+`});var it,at=Z(()=>{it=`// saxpy: y = alpha * x + y
 
 @group(0) @binding(0) var<storage, read>       x: array<f32>;
 @group(0) @binding(1) var<storage, read_write> y: array<f32>;
@@ -143,7 +143,7 @@ fn main(
     y[id * params.y_inc] = params.alpha * x[id * params.x_inc] + y[id * params.y_inc];
   }
 }
-`});var st,it=O(()=>{st=`// scopy: y = x
+`});var nt,st=Z(()=>{nt=`// scopy: y = x
 
 @group(0) @binding(0) var<storage, read>       x: array<f32>;
 @group(0) @binding(1) var<storage, read_write> y: array<f32>;
@@ -167,7 +167,7 @@ fn main(
     y[id * params.y_inc] = x[id * params.x_inc];
   }
 }
-`});var lt,nt=O(()=>{lt=`// dcopy: y = x, double-double (Dekker) f64 emulation of scopy. A copy is
+`});var ut,lt=Z(()=>{ut=`// dcopy: y = x, double-double (Dekker) f64 emulation of scopy. A copy is
 // pure data movement \u2014 hi and lo are transferred verbatim, with no
 // arithmetic at all \u2014 so (unlike dscal/daxpy/ddot) this needs no
 // ddMulProtected/ddAddProtected renormalizing barrier, and so no
@@ -201,7 +201,7 @@ fn main(
     yLo[iy] = xLo[ix];
   }
 }
-`});var ft,ut=O(()=>{ft=`// sdot: result = sum(x[i] * y[i])
+`});var ft,dt=Z(()=>{ft=`// sdot: result = sum(x[i] * y[i])
 // pass 1 dispatches exactly 2 * WGS workgroups; pass 2 uses reduction/sum.wgsl.
 
 @group(0) @binding(0) var<storage, read>       x:        array<f32>;
@@ -254,7 +254,7 @@ fn main(
 
   if (lid.x == 0u) { partials[wgid.x] = tile[0]; }
 }
-`});var De,mt=O(()=>{De=`// sum reduction: collapses 2*WGS partials into one scalar.
+`});var De,mt=Z(()=>{De=`// sum reduction: collapses 2*WGS partials into one scalar.
 // dispatch: 1 workgroup of WGS threads.
 // partials must have exactly 2*WGS entries.
 
@@ -280,7 +280,7 @@ fn reduce(
 
   if (i == 0u) { result[0] = tile[0]; }
 }
-`});var ct,dt=O(()=>{ct=`// sasum: result = sum(|x[i]|)
+`});var pt,ct=Z(()=>{pt=`// sasum: result = sum(|x[i]|)
 // pass 1 dispatches exactly 2 * WGS workgroups; pass 2 uses reduction/abssum.wgsl.
 
 @group(0) @binding(0) var<storage, read>       x:        array<f32>;
@@ -331,7 +331,7 @@ fn main(
 
   if (lid.x == 0u) { partials[wgid.x] = tile[0]; }
 }
-`});var gt,pt=O(()=>{gt=`// snrm2: result = sqrt(sum(x[i] * x[i])), computed via scaled accumulation
+`});var gt,wt=Z(()=>{gt=`// snrm2: result = sqrt(sum(x[i] * x[i])), computed via scaled accumulation
 // (Blue's algorithm / reference BLAS's SLASSQ) rather than naive squaring \u2014
 // naive \`sum += x_i * x_i\` overflows to inf for |x_i| \u2273 1.8e19 (f32's
 // squaring range is only sqrt(f32_max)) and loses precision on tiny
@@ -436,7 +436,7 @@ fn main(
     partialsSsq[wgid.x]   = tileSsq[0];
   }
 }
-`});var ht,wt=O(()=>{ht=`// scaledSum reduction: collapses 2*WGS (scale, ssq) partials from
+`});var bt,ht=Z(()=>{bt=`// scaledSum reduction: collapses 2*WGS (scale, ssq) partials from
 // snrm2.wgsl into the final norm \u2014 sqrt(scale\xB2 \xB7 ssq) == scale \xB7 sqrt(ssq).
 // Mirrors reduction/sum.wgsl's shape exactly, merging via ssqMerge (see
 // snrm2.wgsl for the derivation) instead of plain \`+\`, and taking the final
@@ -501,7 +501,7 @@ fn reduce_scaled(
     result[0] = tileScale[0] * sqrt(tileSsq[0]);
   }
 }
-`});var yt,bt=O(()=>{yt=`// isamax: returns index of element with largest absolute value
+`});var xt,yt=Z(()=>{xt=`// isamax: returns index of element with largest absolute value
 // pass 1 dispatches exactly 2 * WGS workgroups; pass 2 uses reduction/argmax.wgsl.
 
 @group(0) @binding(0) var<storage, read>       x:            array<f32>;
@@ -583,7 +583,7 @@ fn main(
     partials_idx[wgid.x] = tile_idx[0];
   }
 }
-`});var vt,xt=O(()=>{vt=`// amax reduction: collapses 2*WGS (value, index) pairs into one index.
+`});var _t,vt=Z(()=>{_t=`// amax reduction: collapses 2*WGS (value, index) pairs into one index.
 // dispatch: 1 workgroup of WGS threads.
 // partials_val and partials_idx must have exactly 2*WGS entries.
 
@@ -626,7 +626,7 @@ fn reduce(
 
   if (i == 0u) { result[0] = tile_idx[0]; }
 }
-`});var Kr,_t=O(()=>{Kr=`// Double-double arithmetic via Dekker's algorithm \u2014 an alternative to
+`});var Pr,Bt=Z(()=>{Pr=`// Double-double arithmetic via Dekker's algorithm \u2014 an alternative to
 // f64add.wgsl's bit-exact IEEE-754 emulation. Doesn't touch that path.
 //
 // A double-double number is a pair (hi, lo) of f32 with hi+lo approximating
@@ -644,7 +644,7 @@ struct DD {
   hi: f32,
   lo: f32,
 }
-`});var ye,Bt=O(()=>{ye=`// Requires f64/dekker.wgsl concatenated first for the DD struct.
+`});var ye,At=Z(()=>{ye=`// Requires f64/dekker.wgsl concatenated first for the DD struct.
 
 // |a| for a double-double pair. Negation is exact (no rounding), so this is
 // just a sign flip on both components \u2014 hi alone determines the pair's sign.
@@ -654,7 +654,7 @@ fn ddAbs(a: DD) -> DD {
   }
   return a;
 }
-`});var zr,St=O(()=>{zr=`// Requires f64/dekker.wgsl concatenated first for the DD struct.
+`});var Ir,St=Z(()=>{Ir=`// 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,At=O(()=>{Gt=`// dasum: sum(|x[i]|), double-double (Dekker). Same ILP=4 shape as sasum.wgsl;
+`});var Et,Gt=Z(()=>{Et=`// 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.
@@ -822,7 +822,7 @@ fn dasum_main(
     partialsLo[wgid.x] = tile[0].lo;
   }
 }
-`});var Ne,Et=O(()=>{Ne=`// sum reduction (f64, double-double): collapses 2*WGS partial (hi, lo) pairs
+`});var Le,kt=Z(()=>{Le=`// sum reduction (f64, double-double): collapses 2*WGS partial (hi, lo) pairs
 // into one, using ddAddProtected instead of plain f32 \`+\` (see
 // reduction/sum.wgsl for the f32 original this mirrors).
 // dispatch: 1 workgroup of WGS threads. partialsHi/partialsLo must have
@@ -862,7 +862,7 @@ fn reduce_f64(
     resultLo[0] = tile[0].lo;
   }
 }
-`});var $r,kt=O(()=>{$r=`// Requires f64/dekker.wgsl concatenated first for the DD struct, and
+`});var Cr,Dt=Z(()=>{Cr=`// Requires f64/dekker.wgsl concatenated first for the DD struct, and
 // f64/utils/add.wgsl for fsub/negf (bitcast-based subtraction/negation) and,
 // for ddMulProtected at the bottom, fastTwoSumProtected.
 //
@@ -961,7 +961,7 @@ fn ddMulProtected(a: DD, b: DD, threadSlot: u32) -> DD {
   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
+`});var Pt,Lt=Z(()=>{Pt=`// 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
 // a second input vector and a product where dasum takes an absolute value.
 //
@@ -1067,7 +1067,7 @@ fn ddot_main(
     partialsLo[wgid.x] = tile[0].lo;
   }
 }
-`});var Mt,Pt=O(()=>{Mt=`// dscal: x := alpha * x, double-double (Dekker) f64 emulation of sscal.
+`});var Mt,Nt=Z(()=>{Mt=`// dscal: x := alpha * x, double-double (Dekker) f64 emulation of sscal.
 // alpha and x are each an f32 (hi, lo) pair. See f64/utils/multiply.wgsl for
 // ddMulProtected and why plain ddMulRaw isn't safe without a renormalizing
 // barrier \u2014 that barrier needs a provably uniform loop trip count across
@@ -1127,7 +1127,7 @@ fn dscal_main(
     }
   }
 }
-`});var Lt,It=O(()=>{Lt=`// daxpy: y := alpha * x + y, double-double (Dekker) f64 emulation of saxpy.
+`});var Rt,It=Z(()=>{Rt=`// daxpy: y := alpha * x + y, double-double (Dekker) f64 emulation of saxpy.
 // Each element costs one ddMulProtected (alpha*x[i]) then one ddAddProtected
 // (+ y[i]) \u2014 the same two-protected-op shape ddot spends per term, applied
 // straight to the output instead of folded into a reduction. See dscal.wgsl
@@ -1193,7 +1193,7 @@ fn daxpy_main(
     }
   }
 }
-`});var Pe,Rt=O(()=>{Pe=`// Requires f64/dekker.wgsl concatenated first for the DD struct.
+`});var Pe,jt=Z(()=>{Pe=`// Requires f64/dekker.wgsl concatenated first for the DD struct.
 
 // a > b for double-double pairs. hi dominates (|lo| <= ulp(hi)/2 always), so
 // comparing hi alone is correct except on an exact hi tie, when lo breaks it.
@@ -1205,14 +1205,14 @@ fn ddGreater(a: DD, b: DD) -> bool {
   }
   return a.lo > b.lo;
 }
-`});var Tt,qt=O(()=>{Tt=`// Requires f64/dekker.wgsl concatenated first for the DD struct.
+`});var qt,Ft=Z(()=>{qt=`// Requires f64/dekker.wgsl concatenated first for the DD struct.
 
 // a == b for double-double pairs \u2014 exact field equality, no rounding
 // involved, so (like ddGreater) this needs no protection.
 fn ddEqual(a: DD, b: DD) -> bool {
   return a.hi == b.hi && a.lo == b.lo;
 }
-`});var Ft,Ct=O(()=>{Ft=`// idamax: returns index of element with largest absolute value (f64, double-double)
+`});var Tt,Ht=Z(()=>{Tt=`// idamax: returns index of element with largest absolute value (f64, double-double)
 // pass 1 dispatches exactly 2 * WGS workgroups; pass 2 uses reduction/argmaxF64.wgsl.
 // Concatenated after f64/dekker.wgsl (DD struct), f64/utils/abs.wgsl (ddAbs),
 // f64/utils/greater.wgsl (ddGreater), and f64/utils/equal.wgsl (ddEqual).
@@ -1308,7 +1308,7 @@ fn idamax_main(
     partialsIdx[wgid.x]   = tile_idx[0];
   }
 }
-`});var Wt,jt=O(()=>{Wt=`// amax reduction (f64, double-double): collapses 2*WGS (value, index) pairs
+`});var Wt,Ct=Z(()=>{Wt=`// amax reduction (f64, double-double): collapses 2*WGS (value, index) pairs
 // into one index, using ddGreater/ddEqual instead of plain f32 \`>\`/\`==\` (see
 // reduction/argmax.wgsl for the f32 original this mirrors).
 // dispatch: 1 workgroup of WGS threads. partialsValHi/partialsValLo/
@@ -1358,7 +1358,7 @@ fn reduce_f64(
 
   if (i == 0u) { result[0] = tile_idx[0]; }
 }
-`});var Ot,Ht=O(()=>{Ot=`// srot: x = c*x + s*y,  y = -s*x + c*y
+`});var Ot,Vt=Z(()=>{Ot=`// srot: x = c*x + s*y,  y = -s*x + c*y
 
 @group(0) @binding(0) var<storage, read_write> x: array<f32>;
 @group(0) @binding(1) var<storage, read_write> y: array<f32>;
@@ -1387,7 +1387,7 @@ fn main(
     y[id * params.y_inc] = -params.s * xi + params.c * yi;
   }
 }
-`});var Kt,Vt=O(()=>{Kt=`// drot: x = c*x + s*y, y = -s*x + c*y \u2014 double-double (Dekker) f64 emulation
+`});var Ut,Kt=Z(()=>{Ut=`// drot: x = c*x + s*y, y = -s*x + c*y \u2014 double-double (Dekker) f64 emulation
 // of srot. c, s, x, and y are each split into an f32 (hi, lo) pair. Each
 // element costs four ddMulProtected (c*x, s*y, -s*x, c*y) then two
 // ddAddProtected (the two sums) \u2014 negS is computed once outside the loop
@@ -1468,7 +1468,7 @@ fn drot_main(
     }
   }
 }
-`});var Ut,zt=O(()=>{Ut=`// srotm: applies modified Givens rotation H to vectors x and y.
+`});var Yt,zt=Z(()=>{Yt=`// srotm: applies modified Givens rotation H to vectors x and y.
 // param[0] = flag: -1 (full H), 0 (unit diagonal), 1 (unit off-diagonal)
 // param = [ flag, h11, h21, h12, h22 ]
 // flag == -2 (identity/no-op) is handled in JS before dispatch reaches here.
@@ -1518,7 +1518,7 @@ fn main(
     y[id * params.y_inc] = h21 * xi + h22 * yi;
   }
 }
-`});var Xt,Yt=O(()=>{Xt=`// drotm: applies a modified Givens rotation H to vectors x and y \u2014 double-
+`});var Xt,Zt=Z(()=>{Xt=`// drotm: applies a modified Givens rotation H to vectors x and y \u2014 double-
 // double (Dekker) f64 emulation of srotm. paramHi/paramLo[0] = flag: -1
 // (full H), 0 (unit diagonal), 1 (unit off-diagonal). param = [ flag, h11,
 // h21, h12, h22 ], each entry an f32 (hi, lo) pair.
@@ -1617,7 +1617,7 @@ fn drotm_main(
     }
   }
 }
-`});var Zt,$t=O(()=>{Zt=`// Requires f64/dekker.wgsl concatenated first for the DD struct, and
+`});var Ne,$t=Z(()=>{Ne=`// Requires f64/dekker.wgsl concatenated first for the DD struct, and
 // f64/utils/add.wgsl (ddSubProtected/ddAddProtected/negf) and
 // f64/utils/multiply.wgsl (ddMulProtected).
 //
@@ -1662,7 +1662,7 @@ fn ddDivProtected(a: DD, b: DD, threadSlot: u32) -> DD {
   let refined = ddAddProtected(DD(q1, 0.0), DD(q2, 0.0), threadSlot);
   return DD(select(refined.hi, q1, bIsZero), select(refined.lo, 0.0, bIsZero));
 }
-`});var Jt,Qt=O(()=>{Jt=`// Requires f64/dekker.wgsl concatenated first for the DD struct, and
+`});var Qt,Jt=Z(()=>{Qt=`// Requires f64/dekker.wgsl concatenated first for the DD struct, and
 // f64/utils/add.wgsl (ddSubProtected/ddAddProtected) and
 // f64/utils/multiply.wgsl (twoProdBit \u2014 squaring a plain f32 needs no
 // barrier, per multiply.wgsl's own note that twoProdBit is universally safe
@@ -1707,7 +1707,7 @@ fn ddSqrtProtected(a: DD, threadSlot: u32) -> DD {
   let result = ddAddProtected(DD(x0, 0.0), DD(correction, 0.0), threadSlot);
   return DD(select(result.hi, 0.0, isZero), select(result.lo, 0.0, isZero));
 }
-`});var eo,ro=O(()=>{eo=`// dnrm2: result = sqrt(sum(x[i] * x[i])), double-double (Dekker) f64
+`});var eo,ro=Z(()=>{eo=`// dnrm2: result = sqrt(sum(x[i] * x[i])), double-double (Dekker) f64
 // emulation of snrm2 \u2014 same scaled accumulation (Blue's algorithm), just
 // with \`scale\`/\`ssq\` as DD pairs (via ddDivProtected/ddMulProtected/
 // ddAddProtected/ddSqrtProtected) instead of plain f32. Squaring still
@@ -1874,7 +1874,7 @@ fn dnrm2_main(
     partialsSsqLo[wgid.x]   = tileSsqLo[0];
   }
 }
-`});var oo,to=O(()=>{oo=`// scaledSum reduction (f64, double-double): collapses 2*WGS (scale, ssq) DD
+`});var oo,to=Z(()=>{oo=`// scaledSum reduction (f64, double-double): collapses 2*WGS (scale, ssq) DD
 // partials from dnrm2.wgsl into the final norm \u2014 sqrt(scale\xB2 \xB7 ssq) ==
 // scale \xB7 sqrt(ssq), via ddMulProtected/ddSqrtProtected. Mirrors
 // reduction/scaledSum.wgsl's shape exactly; ssqMergeProtected is duplicated
@@ -1967,7 +1967,7 @@ fn reduce_scaled_f64(
     resultLo[0] = result.lo;
   }
 }
-`});var io,ao=O(()=>{io=`// sgemv_n: y = alpha * A * x + beta * y  (A is m\xD7n row-major, no-transpose)
+`});var io,ao=Z(()=>{io=`// sgemv_n: y = alpha * A * x + beta * y  (A is m\xD7n row-major, no-transpose)
 //
 // One workgroup per output row, with a grid-stride outer loop so the shader
 // still covers all rows when m exceeds maxComputeWorkgroupsPerDimension.
@@ -2044,7 +2044,7 @@ fn main(
     workgroupBarrier();
   }
 }
-`});var no,so=O(()=>{no=`// sgemv_t: y = alpha * A^T * x + beta * y  (A is m\xD7n row-major, transposed)
+`});var no,so=Z(()=>{no=`// sgemv_t: y = alpha * A^T * x + beta * y  (A is m\xD7n row-major, transposed)
 // each thread owns one column of A \u2192 one element of y (length n)
 // tiles over x (length m) using shared memory; four independent accumulators
 // let the GPU pipeline A reads across j within each tile (ILP=4)
@@ -2111,7 +2111,7 @@ fn main(
     y[yi] = select(acc, acc + params.beta * y[yi], params.beta != 0.0);
   }
 }
-`});var uo,lo=O(()=>{uo=`// ssymv: y = alpha * A * x + beta * y
+`});var uo,lo=Z(()=>{uo=`// ssymv: y = alpha * A * x + beta * y
 // A is n\xD7n symmetric, lower (uplo=0) or upper (uplo=1) triangle stored.
 // The logical matrix is fully dense (symmetric), so each row's dot product
 // sums over all n columns; entries on the unstored side of the diagonal are
@@ -2182,7 +2182,7 @@ fn main(
     }
   }
 }
-`});var mo,fo=O(()=>{mo=`// strmv: y = op(A) * x
+`});var mo,fo=Z(()=>{mo=`// strmv: y = op(A) * x
 // A is n\xD7n triangular, lower (uplo=0) or upper (uplo=1) triangle stored.
 // op(A) is A (trans=0) or A^T (trans=1).
 // diag=1 (unit) treats the diagonal as 1 without reading A's diagonal values.
@@ -2285,7 +2285,7 @@ fn main(
     }
   }
 }
-`});var Me,co=O(()=>{Me=`// strsv_invert_block: computes ONE column (workgroup_id.x) of ONE block's
+`});var Me,co=Z(()=>{Me=`// strsv_invert_block: computes ONE column (workgroup_id.x) of ONE block's
 // (workgroup_id.y) explicit inverse, via the same one-row-at-a-time
 // substitution as strsv_block.wgsl, but solving against a unit basis vector
 // e_col instead of the real right-hand side, and writing to a dense
@@ -2394,7 +2394,7 @@ fn strsv_invert_block_main(
     workgroupBarrier();
   }
 }
-`});var go,po=O(()=>{go=`// strsv_apply_inverse: given a precomputed block inverse (from
+`});var wo,po=Z(()=>{wo=`// strsv_apply_inverse: given a precomputed block inverse (from
 // strsv_invert_block.wgsl), computes this block's solution as a dense
 // matrix-vector multiply against the block's current remainder in x \u2014
 // replacing what the old strsv_block.wgsl did via a genuinely sequential,
@@ -2441,7 +2441,7 @@ fn strsv_apply_inverse_main(@builtin(local_invocation_id) lid: vec3u) {
   }
   x[(params.blockStart + lid.x) * params.incx] = acc;
 }
-`});var ho,wo=O(()=>{ho=`// strsv_update: subtracts a solved block's contribution from every
+`});var ho,go=Z(()=>{ho=`// strsv_update: subtracts a solved block's contribution from every
 // remaining row in parallel (one workgroup per row, like strmv.wgsl) \u2014
 // this is what turns strsv's O(n) sequential stages into O(n/blockSize).
 // No diag/masking needed: this region never touches the diagonal.
@@ -2516,7 +2516,7 @@ fn strsv_update_main(
     workgroupBarrier();
   }
 }
-`});var yo,bo=O(()=>{yo=`// sger: A := alpha * x * y^T + A  (rank-1 update, A is m\xD7n general/dense)
+`});var yo,bo=Z(()=>{yo=`// sger: A := alpha * x * y^T + A  (rank-1 update, A is m\xD7n general/dense)
 
 @group(0) @binding(0) var<storage, read>       x: array<f32>;
 @group(0) @binding(1) var<storage, read>       y: array<f32>;
@@ -2564,7 +2564,84 @@ fn main(
     }
   }
 }
-`});var vo,xo=O(()=>{vo=`// ssyr: A := alpha * x * x^T + A  (symmetric rank-1 update)
+`});var vo,xo=Z(()=>{vo=`// dger: A := alpha * x * y^T + A, double-double (Dekker) f64 emulation of
+// sger (rank-1 update). x, y, A, and alpha are each split into an f32
+// (hi, lo) pair; WGSL has no f64 type. One workgroup per row of A
+// (grid-stride over rows); within a row, ddMulProtected/ddAddProtected each
+// carry a workgroupBarrier(), so the column loop uses the same
+// uniform-main + ragged-tail split dscal.wgsl uses, rather than sger.wgsl's
+// 4-way ILP unroll \u2014 unrolling would pay the barrier round-trip four times
+// over per iteration for no benefit here.
+
+@group(0) @binding(0) var<storage, read>       xHi: array<f32>;
+@group(0) @binding(1) var<storage, read>       xLo: array<f32>;
+@group(0) @binding(2) var<storage, read>       yHi: array<f32>;
+@group(0) @binding(3) var<storage, read>       yLo: array<f32>;
+@group(0) @binding(4) var<storage, read_write> AHi: array<f32>;
+@group(0) @binding(5) var<storage, read_write> ALo: array<f32>;
+
+struct Params {
+  m:       u32,
+  n:       u32,
+  alphaHi: f32,
+  alphaLo: f32,
+  x_inc:   u32,
+  y_inc:   u32,
+  lda:     u32,
+}
+
+@group(0) @binding(6) var<uniform> params: Params;
+
+const WGS: u32 = 64u;
+
+@compute @workgroup_size(64)
+fn dger_main(
+  @builtin(workgroup_id)        wgid: vec3u,
+  @builtin(local_invocation_id) lid:  vec3u,
+  @builtin(num_workgroups)      nwg:  vec3u,
+) {
+  let alpha = DD(params.alphaHi, params.alphaLo);
+
+  // Column loop's trip count depends only on n and WGS, not on row \u2014 the
+  // same main/tail split applies uniformly to every row, so every thread
+  // reaches each protected call the same number of times overall.
+  let n_floor = (params.n / WGS) * WGS;
+  let mainIters = n_floor / WGS;
+  let hasTail = select(0u, 1u, n_floor < params.n);
+
+  for (var row = wgid.x; row < params.m; row += nwg.x) {
+    // Every thread in the workgroup redundantly computes the same alpha*x[row]
+    // \u2014 wasteful but harmless, and keeps the per-row protected-call count
+    // trivially identical across threads.
+    let ix = row * params.x_inc;
+    let xi = ddMulProtected(alpha, DD(xHi[ix], xLo[ix]), lid.x);
+    let row_base = row * params.lda;
+
+    for (var iter = 0u; iter < mainIters; iter++) {
+      let col = lid.x + iter * WGS;
+      let idx = row_base + col;
+      let iy = col * params.y_inc;
+      let prod = ddMulProtected(xi, DD(yHi[iy], yLo[iy]), lid.x);
+      let result = ddAddProtected(prod, DD(AHi[idx], ALo[idx]), lid.x);
+      AHi[idx] = result.hi;
+      ALo[idx] = result.lo;
+    }
+
+    for (var iter = 0u; iter < hasTail; iter++) {
+      let col = n_floor + lid.x;
+      let valid = col < params.n;
+      let idx = select(0u, row_base + col, valid);
+      let iy = select(0u, col * params.y_inc, valid);
+      let prod = ddMulProtected(xi, DD(yHi[iy], yLo[iy]), lid.x);
+      let result = ddAddProtected(prod, DD(AHi[idx], ALo[idx]), lid.x);
+      if (valid) {
+        AHi[idx] = result.hi;
+        ALo[idx] = result.lo;
+      }
+    }
+  }
+}
+`});var Bo,_o=Z(()=>{Bo=`// ssyr: A := alpha * x * x^T + A  (symmetric rank-1 update)
 // A is n\xD7n symmetric; only the triangle specified by uplo is referenced/updated,
 // the other triangle is implied by symmetry (not touched).
 
@@ -2624,7 +2701,91 @@ fn main(
     }
   }
 }
-`});var Bo,_o=O(()=>{Bo=`// ssyr2: A := alpha * x * y^T + alpha * y * x^T + A  (symmetric rank-2 update)
+`});var So,Ao=Z(()=>{So=`// dsyr: A := alpha * x * x^T + A, double-double (Dekker) f64 emulation of
+// ssyr (symmetric rank-1 update). x, A, and alpha are each split into an f32
+// (hi, lo) pair; WGSL has no f64 type. Only the triangle specified by uplo
+// is referenced/updated. One workgroup per row of A (grid-stride over
+// rows); within a row, ddMulProtected/ddAddProtected each carry a
+// workgroupBarrier(), so the column loop uses the same uniform-main +
+// ragged-tail split dger.wgsl uses over its *stored* range, rather than
+// ssyr.wgsl's 4-way ILP unroll.
+
+@group(0) @binding(0) var<storage, read>       xHi: array<f32>;
+@group(0) @binding(1) var<storage, read>       xLo: array<f32>;
+@group(0) @binding(2) var<storage, read_write> AHi: array<f32>;
+@group(0) @binding(3) var<storage, read_write> ALo: array<f32>;
+
+struct Params {
+  n:     u32,
+  alphaHi: f32,
+  alphaLo: f32,
+  incx:  u32,
+  lda:   u32,
+  uplo:  u32,  // 0 = lower, 1 = upper
+}
+
+@group(0) @binding(4) var<uniform> params: Params;
+
+const WGS: u32 = 64u;
+
+@compute @workgroup_size(64)
+fn dsyr_main(
+  @builtin(workgroup_id)        wgid: vec3u,
+  @builtin(local_invocation_id) lid:  vec3u,
+  @builtin(num_workgroups)      nwg:  vec3u,
+) {
+  let alpha = DD(params.alphaHi, params.alphaLo);
+
+  for (var row = wgid.x; row < params.n; row += nwg.x) {
+    let ix = row * params.incx;
+    let xi = ddMulProtected(alpha, DD(xHi[ix], xLo[ix]), lid.x);
+    let row_base = row * params.lda;
+
+    // Stored-triangle column range for this row: lower [0,row], upper [row,n).
+    var colStart: u32;
+    var colEnd: u32;
+    if params.uplo == 1u {
+      colStart = row;
+      colEnd = params.n;
+    } else {
+      colStart = 0u;
+      colEnd = row + 1u;
+    }
+
+    // Range length varies per row, but every thread in the workgroup runs
+    // the SAME row at the same time (the outer loop is shared), so the
+    // main/tail split computed from colStart/colEnd is already uniform
+    // across threads for this row \u2014 just not the same across different rows.
+    let rangeLen = colEnd - colStart;
+    let range_floor = colStart + (rangeLen / WGS) * WGS;
+    let mainIters = (range_floor - colStart) / WGS;
+    let hasTail = select(0u, 1u, range_floor < colEnd);
+
+    for (var iter = 0u; iter < mainIters; iter++) {
+      let col = colStart + lid.x + iter * WGS;
+      let idx = row_base + col;
+      let ic = col * params.incx;
+      let prod = ddMulProtected(xi, DD(xHi[ic], xLo[ic]), lid.x);
+      let result = ddAddProtected(prod, DD(AHi[idx], ALo[idx]), lid.x);
+      AHi[idx] = result.hi;
+      ALo[idx] = result.lo;
+    }
+
+    for (var iter = 0u; iter < hasTail; iter++) {
+      let col = range_floor + lid.x;
+      let valid = col < colEnd;
+      let idx = select(0u, row_base + col, valid);
+      let ic = select(0u, col * params.incx, valid);
+      let prod = ddMulProtected(xi, DD(xHi[ic], xLo[ic]), lid.x);
+      let result = ddAddProtected(prod, DD(AHi[idx], ALo[idx]), lid.x);
+      if (valid) {
+        AHi[idx] = result.hi;
+        ALo[idx] = result.lo;
+      }
+    }
+  }
+}
+`});var Eo,Go=Z(()=>{Eo=`// ssyr2: A := alpha * x * y^T + alpha * y * x^T + A  (symmetric rank-2 update)
 // A is n\xD7n symmetric; only the triangle specified by uplo is referenced/updated,
 // the other triangle is implied by symmetry (not touched).
 
@@ -2687,7 +2848,871 @@ fn main(
     }
   }
 }
-`});var ue,So=O(()=>{ue=`// sgemm_small: C = alpha * op(A) * op(B) + beta * C \u2014 small-tile half of
+`});var Do,ko=Z(()=>{Do=`// dsyr2: A := alpha * x * y^T + alpha * y * x^T + A, double-double (Dekker)
+// f64 emulation of ssyr2 (symmetric rank-2 update). x, y, A, and alpha are
+// each split into an f32 (hi, lo) pair; WGSL has no f64 type. Only the
+// triangle specified by uplo is referenced/updated. One workgroup per row
+// of A (grid-stride over rows); within a row, ddMulProtected/ddAddProtected
+// each carry a workgroupBarrier(), so the column loop uses the same
+// uniform-main + ragged-tail split dsyr.wgsl uses over its *stored* range,
+// rather than ssyr2.wgsl's 4-way ILP unroll.
+
+@group(0) @binding(0) var<storage, read>       xHi: array<f32>;
+@group(0) @binding(1) var<storage, read>       xLo: array<f32>;
+@group(0) @binding(2) var<storage, read>       yHi: array<f32>;
+@group(0) @binding(3) var<storage, read>       yLo: array<f32>;
+@group(0) @binding(4) var<storage, read_write> AHi: array<f32>;
+@group(0) @binding(5) var<storage, read_write> ALo: array<f32>;
+
+struct Params {
+  n:       u32,
+  alphaHi: f32,
+  alphaLo: f32,
+  incx:    u32,
+  incy:    u32,
+  lda:     u32,
+  uplo:    u32,  // 0 = lower, 1 = upper
+}
+
+@group(0) @binding(6) var<uniform> params: Params;
+
+const WGS: u32 = 64u;
+
+@compute @workgroup_size(64)
+fn dsyr2_main(
+  @builtin(workgroup_id)        wgid: vec3u,
+  @builtin(local_invocation_id) lid:  vec3u,
+  @builtin(num_workgroups)      nwg:  vec3u,
+) {
+  let alpha = DD(params.alphaHi, params.alphaLo);
+
+  for (var row = wgid.x; row < params.n; row += nwg.x) {
+    let ix = row * params.incx;
+    let iy = row * params.incy;
+    let xi = ddMulProtected(alpha, DD(xHi[ix], xLo[ix]), lid.x);
+    let yi = ddMulProtected(alpha, DD(yHi[iy], yLo[iy]), lid.x);
+    let row_base = row * params.lda;
+
+    // Stored-triangle column range for this row: lower [0,row], upper [row,n).
+    var colStart: u32;
+    var colEnd: u32;
+    if params.uplo == 1u {
+      colStart = row;
+      colEnd = params.n;
+    } else {
+      colStart = 0u;
+      colEnd = row + 1u;
+    }
+
+    // Range length varies per row, but every thread in the workgroup runs
+    // the SAME row at the same time (the outer loop is shared), so the
+    // main/tail split computed from colStart/colEnd is already uniform
+    // across threads for this row \u2014 just not the same across different rows.
+    let rangeLen = colEnd - colStart;
+    let range_floor = colStart + (rangeLen / WGS) * WGS;
+    let mainIters = (range_floor - colStart) / WGS;
+    let hasTail = select(0u, 1u, range_floor < colEnd);
+
+    for (var iter = 0u; iter < mainIters; iter++) {
+      let col = colStart + lid.x + iter * WGS;
+      let idx = row_base + col;
+      let jc_x = col * params.incx;
+      let jc_y = col * params.incy;
+      let prod1 = ddMulProtected(xi, DD(yHi[jc_y], yLo[jc_y]), lid.x);
+      let prod2 = ddMulProtected(yi, DD(xHi[jc_x], xLo[jc_x]), lid.x);
+      let sum1 = ddAddProtected(prod1, prod2, lid.x);
+      let result = ddAddProtected(sum1, DD(AHi[idx], ALo[idx]), lid.x);
+      AHi[idx] = result.hi;
+      ALo[idx] = result.lo;
+    }
+
+    for (var iter = 0u; iter < hasTail; iter++) {
+      let col = range_floor + lid.x;
+      let valid = col < colEnd;
+      let idx = select(0u, row_base + col, valid);
+      let jc_x = select(0u, col * params.incx, valid);
+      let jc_y = select(0u, col * params.incy, valid);
+      let prod1 = ddMulProtected(xi, DD(yHi[jc_y], yLo[jc_y]), lid.x);
+      let prod2 = ddMulProtected(yi, DD(xHi[jc_x], xLo[jc_x]), lid.x);
+      let sum1 = ddAddProtected(prod1, prod2, lid.x);
+      let result = ddAddProtected(sum1, DD(AHi[idx], ALo[idx]), lid.x);
+      if (valid) {
+        AHi[idx] = result.hi;
+        ALo[idx] = result.lo;
+      }
+    }
+  }
+}
+`});var Po,Lo=Z(()=>{Po=`// dgemv_n: y := alpha * A * x + beta * y, double-double (Dekker) f64
+// emulation of sgemv_n (matrix-vector product, no transpose). A, x, y,
+// alpha, and beta are each split into an f32 (hi, lo) pair; WGSL has no f64
+// type. Same one-workgroup-per-output-row shape as sgemv_n.wgsl (grid-stride
+// over rows), but the per-row dot product is reduced via a full
+// within-workgroup DD tree reduction \u2014 the same reduction ddot.wgsl uses to
+// combine 64 lanes down to one \u2014 rather than sgemv_n.wgsl's 4-way ILP
+// unroll + shared-memory tree (ddMulProtected/ddAddProtected each carry a
+// workgroupBarrier(), so unrolling would pay that barrier four times over
+// per iteration for no benefit). Since each row is handled by exactly one
+// workgroup, no separate cross-workgroup reduction pass (ddot's reduce_f64)
+// is needed here \u2014 the tree reduction alone finishes the row.
+
+@group(0) @binding(0) var<storage, read>       AHi: array<f32>;
+@group(0) @binding(1) var<storage, read>       ALo: array<f32>;
+@group(0) @binding(2) var<storage, read>       xHi: array<f32>;
+@group(0) @binding(3) var<storage, read>       xLo: array<f32>;
+@group(0) @binding(4) var<storage, read_write> yHi: array<f32>;
+@group(0) @binding(5) var<storage, read_write> yLo: array<f32>;
+
+struct Params {
+  m:       u32,
+  n:       u32,
+  alphaHi: f32,
+  alphaLo: f32,
+  betaHi:  f32,
+  betaLo:  f32,
+  incx:    u32,
+  incy:    u32,
+  lda:     u32,
+}
+
+@group(0) @binding(6) var<uniform> params: Params;
+
+const WGS: u32 = 64u;
+var<workgroup> tile: array<DD, 64>;
+
+@compute @workgroup_size(64)
+fn dgemv_n_main(
+  @builtin(workgroup_id)        wgid: vec3u,
+  @builtin(local_invocation_id) lid:  vec3u,
+  @builtin(num_workgroups)      nwg:  vec3u,
+) {
+  let alpha = DD(params.alphaHi, params.alphaLo);
+  let beta  = DD(params.betaHi, params.betaLo);
+  let isBetaNonzero = params.betaHi != 0.0 || params.betaLo != 0.0;
+
+  // Column loop's trip count depends only on n and WGS, not on row \u2014 same
+  // main/tail split applies uniformly to every row (see dger.wgsl).
+  let n_floor   = (params.n / WGS) * WGS;
+  let mainIters = n_floor / WGS;
+  let hasTail   = select(0u, 1u, n_floor < params.n);
+
+  for (var row = wgid.x; row < params.m; row += nwg.x) {
+    let row_base = row * params.lda;
+    var acc = DD(0.0, 0.0);
+
+    for (var iter = 0u; iter < mainIters; iter++) {
+      let j  = lid.x + iter * WGS;
+      let ia = row_base + j;
+      let ix = j * params.incx;
+      let prod = ddMulProtected(DD(AHi[ia], ALo[ia]), DD(xHi[ix], xLo[ix]), lid.x);
+      acc = ddAddProtected(acc, prod, lid.x);
+    }
+
+    for (var iter = 0u; iter < hasTail; iter++) {
+      let j     = n_floor + lid.x;
+      let valid = j < params.n;
+      let ia    = select(0u, row_base + j, valid);
+      let ix    = select(0u, j * params.incx, valid);
+      let prod  = ddMulProtected(DD(AHi[ia], ALo[ia]), DD(xHi[ix], xLo[ix]), lid.x);
+      let contribution = DD(select(0.0, prod.hi, valid), select(0.0, prod.lo, valid));
+      acc = ddAddProtected(acc, contribution, lid.x);
+    }
+
+    tile[lid.x] = acc;
+    workgroupBarrier();
+
+    // Inactive threads combine against a throwaway partner and discard it
+    // (ddAddProtected must be called unconditionally by every thread).
+    for (var s = WGS / 2u; s > 0u; s >>= 1u) {
+      let partner = select(lid.x, lid.x + s, lid.x < s);
+      let combined = ddAddProtected(tile[lid.x], tile[partner], lid.x);
+      workgroupBarrier(); // all threads must read tile[] above before any write below
+      if (lid.x < s) { tile[lid.x] = combined; }
+      workgroupBarrier();
+    }
+
+    // tile[0] now holds the full row dot product, visible to every lane \u2014
+    // every thread redundantly finishes the O(1) alpha/beta combine so the
+    // protected ops below stay uniformly called (only the write is gated).
+    let dot = tile[0];
+    let scaled = ddMulProtected(alpha, dot, lid.x);
+    let iy = row * params.incy;
+    let yVal = DD(yHi[iy], yLo[iy]);
+    let betaTimesY = ddMulProtected(beta, yVal, lid.x);
+    let withBeta = ddAddProtected(scaled, betaTimesY, lid.x);
+    // BLAS beta==0 semantics: y is written, not accumulated \u2014 the result
+    // must not depend on y's prior value (though it is still read above).
+    let result = DD(
+      select(scaled.hi, withBeta.hi, isBetaNonzero),
+      select(scaled.lo, withBeta.lo, isBetaNonzero),
+    );
+
+    if (lid.x == 0u) {
+      yHi[iy] = result.hi;
+      yLo[iy] = result.lo;
+    }
+    // All 64 threads must agree before the next row reuses tile[].
+    workgroupBarrier();
+  }
+}
+`});var Mo,No=Z(()=>{Mo=`// dgemv_t: y := alpha * A^T * x + beta * y, double-double (Dekker) f64
+// emulation of sgemv_t (matrix-vector product, transposed). A, x, y, alpha,
+// and beta are each split into an f32 (hi, lo) pair; WGSL has no f64 type.
+// Same one-thread-per-output-column shape as sgemv_t.wgsl, tiling over x
+// (length m) via shared memory. Because ddMulProtected/ddAddProtected each
+// carry a workgroupBarrier(), every thread in the workgroup \u2014 including ones
+// whose column is out of range (n not a multiple of WGS) \u2014 must call them
+// the same number of times: unlike sgemv_t.wgsl's \`if (col < n)\` guard
+// around the whole accumulation, out-of-range threads here compute against
+// a clamped dummy column unconditionally and simply never write their
+// result (same technique dger.wgsl/ddot.wgsl use for their ragged tails).
+
+@group(0) @binding(0) var<storage, read>       AHi: array<f32>;
+@group(0) @binding(1) var<storage, read>       ALo: array<f32>;
+@group(0) @binding(2) var<storage, read>       xHi: array<f32>;
+@group(0) @binding(3) var<storage, read>       xLo: array<f32>;
+@group(0) @binding(4) var<storage, read_write> yHi: array<f32>;
+@group(0) @binding(5) var<storage, read_write> yLo: array<f32>;
+
+struct Params {
+  m:       u32,
+  n:       u32,
+  alphaHi: f32,
+  alphaLo: f32,
+  betaHi:  f32,
+  betaLo:  f32,
+  incx:    u32,
+  incy:    u32,
+  lda:     u32,
+}
+
+@group(0) @binding(6) var<uniform> params: Params;
+
+const WGS: u32 = 64u;
+var<workgroup> xTile: array<DD, 64>;
+
+@compute @workgroup_size(64)
+fn dgemv_t_main(
+  @builtin(global_invocation_id) gid: vec3u,
+  @builtin(local_invocation_id)  lid: vec3u,
+) {
+  let col      = gid.x;
+  let colValid = col < params.n;
+  // Clamp so out-of-range threads still index safely \u2014 their contribution
+  // is computed but never written back.
+  let safeCol  = select(0u, col, colValid);
+
+  var acc = DD(0.0, 0.0);
+
+  let m_floor   = (params.m / WGS) * WGS;
+  let mainIters = m_floor / WGS;
+
+  for (var iter = 0u; iter < mainIters; iter++) {
+    let base = iter * WGS;
+    // Cooperative load: all 64 threads fill xTile with x[base..base+WGS),
+    // independent of col \u2014 safe regardless of whether this thread's column
+    // is in range.
+    let ix = (base + lid.x) * params.incx;
+    xTile[lid.x] = DD(xHi[ix], xLo[ix]);
+    workgroupBarrier();
+
+    for (var j = 0u; j < WGS; j++) {
+      let ia = (base + j) * params.lda + safeCol;
+      let prod = ddMulProtected(DD(AHi[ia], ALo[ia]), xTile[j], lid.x);
+      acc = ddAddProtected(acc, prod, lid.x);
+    }
+    workgroupBarrier();
+  }
+
+  // Remainder rows (m not a multiple of WGS): the count is the same for
+  // every thread regardless of col, so this loop is already barrier-safe
+  // without needing a second tiling pass.
+  let remCount = params.m - m_floor;
+  for (var k = 0u; k < remCount; k++) {
+    let row = m_floor + k;
+    let ia  = row * params.lda + safeCol;
+    let ix  = row * params.incx;
+    let prod = ddMulProtected(DD(AHi[ia], ALo[ia]), DD(xHi[ix], xLo[ix]), lid.x);
+    acc = ddAddProtected(acc, prod, lid.x);
+  }
+
+  let alpha = DD(params.alphaHi, params.alphaLo);
+  let beta  = DD(params.betaHi, params.betaLo);
+  let scaled = ddMulProtected(alpha, acc, lid.x);
+
+  let iy = safeCol * params.incy;
+  let yVal = DD(yHi[iy], yLo[iy]);
+  let betaTimesY = ddMulProtected(beta, yVal, lid.x);
+  let withBeta = ddAddProtected(scaled, betaTimesY, lid.x);
+  let isBetaNonzero = params.betaHi != 0.0 || params.betaLo != 0.0;
+  // BLAS beta==0 semantics: y is written, not accumulated \u2014 the result
+  // must not depend on y's prior value (though it is still read above).
+  let result = DD(
+    select(scaled.hi, withBeta.hi, isBetaNonzero),
+    select(scaled.lo, withBeta.lo, isBetaNonzero),
+  );
+
+  if (colValid) {
+    yHi[iy] = result.hi;
+    yLo[iy] = result.lo;
+  }
+}
+`});var Ro,Io=Z(()=>{Ro=`// dsymv: y := alpha * A * x + beta * y, double-double (Dekker) f64 emulation
+// of ssymv (symmetric matrix-vector product). A, x, y, alpha, and beta are
+// each split into an f32 (hi, lo) pair; WGSL has no f64 type. A is n\xD7n
+// symmetric \u2014 only the triangle specified by uplo is physically stored, so
+// entries on the unstored side of the diagonal are fetched from their
+// mirror position (A[i,j] == A[j,i]), same as ssymv.wgsl. Same
+// one-workgroup-per-row + full within-workgroup DD tree reduction shape as
+// dgemv_n.wgsl (every row sums over all n columns here, unlike dsyr's
+// triangle-restricted range, since the logical matrix is fully dense).
+
+@group(0) @binding(0) var<storage, read>       AHi: array<f32>;
+@group(0) @binding(1) var<storage, read>       ALo: array<f32>;
+@group(0) @binding(2) var<storage, read>       xHi: array<f32>;
+@group(0) @binding(3) var<storage, read>       xLo: array<f32>;
+@group(0) @binding(4) var<storage, read_write> yHi: array<f32>;
+@group(0) @binding(5) var<storage, read_write> yLo: array<f32>;
+
+struct Params {
+  n:       u32,
+  alphaHi: f32,
+  alphaLo: f32,
+  betaHi:  f32,
+  betaLo:  f32,
+  incx:    u32,
+  incy:    u32,
+  lda:     u32,
+  uplo:    u32,  // 0 = lower, 1 = upper
+}
+
+@group(0) @binding(6) var<uniform> params: Params;
+
+const WGS: u32 = 64u;
+var<workgroup> tile: array<DD, 64>;
+
+// A[i,j] flat index for the symmetric matrix \u2014 stored position if (i,j) is
+// on the uplo side of the diagonal, mirrored from (j,i) otherwise. Pure
+// addressing (no protected op inside), so branching here is safe.
+fn symIdx(i: u32, j: u32) -> u32 {
+  var row: u32;
+  var col: u32;
+  if params.uplo == 0u {
+    // Lower: stored at (i,j) for j <= i, mirrored from (j,i) otherwise.
+    if j <= i { row = i; col = j; } else { row = j; col = i; }
+  } else {
+    // Upper: stored at (i,j) for j >= i, mirrored from (j,i) otherwise.
+    if j >= i { row = i; col = j; } else { row = j; col = i; }
+  }
+  return row * params.lda + col;
+}
+
+@compute @workgroup_size(64)
+fn dsymv_main(
+  @builtin(workgroup_id)        wgid: vec3u,
+  @builtin(local_invocation_id) lid:  vec3u,
+  @builtin(num_workgroups)      nwg:  vec3u,
+) {
+  let alpha = DD(params.alphaHi, params.alphaLo);
+  let beta  = DD(params.betaHi, params.betaLo);
+  let isBetaNonzero = params.betaHi != 0.0 || params.betaLo != 0.0;
+
+  // Column loop's trip count depends only on n and WGS, not on row \u2014 same
+  // main/tail split applies uniformly to every row (see dger.wgsl).
+  let n_floor   = (params.n / WGS) * WGS;
+  let mainIters = n_floor / WGS;
+  let hasTail   = select(0u, 1u, n_floor < params.n);
+
+  for (var i = wgid.x; i < params.n; i += nwg.x) {
+    var acc = DD(0.0, 0.0);
+
+    for (var iter = 0u; iter < mainIters; iter++) {
+      let j   = lid.x + iter * WGS;
+      let idx = symIdx(i, j);
+      let ix  = j * params.incx;
+      let prod = ddMulProtected(DD(AHi[idx], ALo[idx]), DD(xHi[ix], xLo[ix]), lid.x);
+      acc = ddAddProtected(acc, prod, lid.x);
+    }
+
+    for (var iter = 0u; iter < hasTail; iter++) {
+      let j     = n_floor + lid.x;
+      let valid = j < params.n;
+      let safeJ = select(0u, j, valid);
+      let idx   = symIdx(i, safeJ);
+      let ix    = safeJ * params.incx;
+      let prod  = ddMulProtected(DD(AHi[idx], ALo[idx]), DD(xHi[ix], xLo[ix]), lid.x);
+      let contribution = DD(select(0.0, prod.hi, valid), select(0.0, prod.lo, valid));
+      acc = ddAddProtected(acc, contribution, lid.x);
+    }
+
+    tile[lid.x] = acc;
+    workgroupBarrier();
+
+    // Inactive threads combine against a throwaway partner and discard it
+    // (ddAddProtected must be called unconditionally by every thread).
+    for (var s = WGS / 2u; s > 0u; s >>= 1u) {
+      let partner = select(lid.x, lid.x + s, lid.x < s);
+      let combined = ddAddProtected(tile[lid.x], tile[partner], lid.x);
+      workgroupBarrier(); // all threads must read tile[] above before any write below
+      if (lid.x < s) { tile[lid.x] = combined; }
+      workgroupBarrier();
+    }
+
+    // tile[0] now holds the full row dot product, visible to every lane \u2014
+    // every thread redundantly finishes the O(1) alpha/beta combine so the
+    // protected ops below stay uniformly called (only the write is gated).
+    let dot = tile[0];
+    let scaled = ddMulProtected(alpha, dot, lid.x);
+    let iy = i * params.incy;
+    let yVal = DD(yHi[iy], yLo[iy]);
+    let betaTimesY = ddMulProtected(beta, yVal, lid.x);
+    let withBeta = ddAddProtected(scaled, betaTimesY, lid.x);
+    // BLAS beta==0 semantics: y is written, not accumulated \u2014 the result
+    // must not depend on y's prior value (though it is still read above).
+    let result = DD(
+      select(scaled.hi, withBeta.hi, isBetaNonzero),
+      select(scaled.lo, withBeta.lo, isBetaNonzero),
+    );
+
+    if (lid.x == 0u) {
+      yHi[iy] = result.hi;
+      yLo[iy] = result.lo;
+    }
+    // All 64 threads must agree before the next row reuses tile[].
+    workgroupBarrier();
+  }
+}
+`});var Fo,jo=Z(()=>{Fo=`// dtrmv: y := op(A) * x, double-double (Dekker) f64 emulation of strmv
+// (triangular matrix-vector product). A, x, and y are each split into an f32
+// (hi, lo) pair; WGSL has no f64 type. A is n\xD7n triangular \u2014 only the
+// triangle specified by uplo is referenced. Like strmv.wgsl, wgblas's trmv
+// takes a separate output vector y rather than overwriting x in place (the
+// standard BLAS signature is in-place), so there is no aliasing/read-write-
+// ordering concern the way real in-place trmv would have \u2014 this is
+// structurally just dsymv.wgsl's per-row DD tree reduction with a triangular
+// (not full-n, not mirrored) column range per row, same shape as dsyr.wgsl's
+// per-row-varying range (main/tail split computed inside the row loop, since
+// range length depends on i).
+
+@group(0) @binding(0) var<storage, read>       AHi: array<f32>;
+@group(0) @binding(1) var<storage, read>       ALo: array<f32>;
+@group(0) @binding(2) var<storage, read>       xHi: array<f32>;
+@group(0) @binding(3) var<storage, read>       xLo: array<f32>;
+@group(0) @binding(4) var<storage, read_write> yHi: array<f32>;
+@group(0) @binding(5) var<storage, read_write> yLo: array<f32>;
+
+struct Params {
+  n:     u32,
+  incx:  u32,
+  incy:  u32,
+  lda:   u32,
+  trans: u32,  // 0 = no-transpose, 1 = transpose
+  uplo:  u32,  // 0 = lower, 1 = upper
+  diag:  u32,  // 0 = non-unit, 1 = unit
+}
+
+@group(0) @binding(6) var<uniform> params: Params;
+
+const WGS: u32 = 64u;
+var<workgroup> tile: array<DD, 64>;
+
+@compute @workgroup_size(64)
+fn dtrmv_main(
+  @builtin(workgroup_id)        wgid: vec3u,
+  @builtin(local_invocation_id) lid:  vec3u,
+  @builtin(num_workgroups)      nwg:  vec3u,
+) {
+  // Effective "upper" range for row i: op(A)[i,j] is stored at A[i,j] for
+  // no-transpose+upper or transpose+lower (range [i, n)); at A[j,i] for
+  // no-transpose+lower or transpose+upper (range [0, i]).
+  let isUpperRange = (params.trans == 0u) == (params.uplo == 1u);
+
+  for (var i = wgid.x; i < params.n; i += nwg.x) {
+    let rangeStart = select(0u, i, isUpperRange);
+    let rangeEnd   = select(i + 1u, params.n, isUpperRange);
+
+    // Range length varies per row, but every thread in the workgroup runs
+    // the SAME row at the same time, so the main/tail split computed from
+    // rangeStart/rangeEnd is already uniform across threads for this row \u2014
+    // just not the same across different rows (see dsyr.wgsl).
+    let rangeLen    = rangeEnd - rangeStart;
+    let range_floor = rangeStart + (rangeLen / WGS) * WGS;
+    let mainIters   = (range_floor - rangeStart) / WGS;
+    let hasTail     = select(0u, 1u, range_floor < rangeEnd);
+
+    var acc = DD(0.0, 0.0);
+
+    for (var iter = 0u; iter < mainIters; iter++) {
+      let j    = rangeStart + lid.x + iter * WGS;
+      let addr = select(j * params.lda + i, i * params.lda + j, params.trans == 0u);
+      let isUnitDiag = params.diag == 1u && j == i;
+      let aVal = DD(
+        select(AHi[addr], 1.0, isUnitDiag),
+        select(ALo[addr], 0.0, isUnitDiag),
+      );
+      let ix = j * params.incx;
+      let prod = ddMulProtected(aVal, DD(xHi[ix], xLo[ix]), lid.x);
+      acc = ddAddProtected(acc, prod, lid.x);
+    }
+
+    for (var iter = 0u; iter < hasTail; iter++) {
+      let j     = range_floor + lid.x;
+      let valid = j < rangeEnd;
+      let safeJ = select(rangeStart, j, valid); // rangeLen is always >= 1, so rangeStart is a safe in-range fallback
+      let addr  = select(safeJ * params.lda + i, i * params.lda + safeJ, params.trans == 0u);
+      let isUnitDiag = params.diag == 1u && safeJ == i;
+      let aVal = DD(
+        select(AHi[addr], 1.0, isUnitDiag),
+        select(ALo[addr], 0.0, isUnitDiag),
+      );
+      let ix = safeJ * params.incx;
+      let prod = ddMulProtected(aVal, DD(xHi[ix], xLo[ix]), lid.x);
+      let contribution = DD(select(0.0, prod.hi, valid), select(0.0, prod.lo, valid));
+      acc = ddAddProtected(acc, contribution, lid.x);
+    }
+
+    tile[lid.x] = acc;
+    workgroupBarrier();
+
+    // Inactive threads combine against a throwaway partner and discard it
+    // (ddAddProtected must be called unconditionally by every thread).
+    for (var s = WGS / 2u; s > 0u; s >>= 1u) {
+      let partner = select(lid.x, lid.x + s, lid.x < s);
+      let combined = ddAddProtected(tile[lid.x], tile[partner], lid.x);
+      workgroupBarrier(); // all threads must read tile[] above before any write below
+      if (lid.x < s) { tile[lid.x] = combined; }
+      workgroupBarrier();
+    }
+
+    if (lid.x == 0u) {
+      let iy = i * params.incy;
+      yHi[iy] = tile[0].hi;
+      yLo[iy] = tile[0].lo;
+    }
+    // All 64 threads must agree before the next row reuses tile[].
+    workgroupBarrier();
+  }
+}
+`});var Ho,qo=Z(()=>{Ho=`// dtrsv_invert_block: double-double (Dekker) f64 emulation of
+// strsv_invert_block.wgsl \u2014 computes ONE column (workgroup_id.x) of ONE
+// block's (workgroup_id.y) explicit inverse, via the same one-row-at-a-time
+// substitution, solving against a unit basis vector e_col. A, Ainv, and the
+// running accumulation are each split into an f32 (hi, lo) pair; WGSL has no
+// f64 type.
+//
+// This is the first routine in the f64 port order to use ddDivProtected (the
+// non-unit-diagonal division) \u2014 deliberately last per TODO.md, since
+// division is where dnrm2 found real double-double edge-case bugs during
+// the L1 port. The division only runs once per (column, step) \u2014 same
+// barrier-uniformity requirement as every other protected op here \u2014 so
+// every thread in the workgroup computes it redundantly from the
+// already-shared \`tile[0]\` reduction result (same pattern dgemv_n uses for
+// its O(1) alpha/beta combine after the per-row reduction), with only
+// \`lid.x==0\` writing the result. \`params.diag\` is uniform across the whole
+// dispatch (not per-thread), so branching on it to skip the division
+// entirely for a unit diagonal is safe \u2014 unlike dtrmv.wgsl's per-element
+// diag check, which varies per thread and needed select() instead.
+
+@group(0) @binding(0) var<storage, read>       AHi: array<f32>;
+@group(0) @binding(1) var<storage, read>       ALo: array<f32>;
+@group(0) @binding(2) var<storage, read_write> AinvHi: array<f32>;
+@group(0) @binding(3) var<storage, read_write> AinvLo: array<f32>;
+
+struct Params {
+  n:     u32,
+  lda:   u32,
+  trans: u32,  // 0 = no-transpose, 1 = transpose
+  uplo:  u32,  // 0 = lower, 1 = upper
+  diag:  u32,  // 0 = non-unit, 1 = unit
+}
+
+@group(0) @binding(4) var<uniform> params: Params;
+
+const WGS: u32 = 64u;
+const BLOCK_SIZE: u32 = 64u;
+var<workgroup> tile: array<DD, 64>;
+
+fn readA(i: u32, j: u32) -> DD {
+  let idx = select(j * params.lda + i, i * params.lda + j, params.trans == 0u);
+  return DD(AHi[idx], ALo[idx]);
+}
+
+@compute @workgroup_size(64)
+fn dtrsv_invert_block_main(
+  @builtin(workgroup_id)        wgid: vec3u,
+  @builtin(local_invocation_id) lid:  vec3u,
+) {
+  let col = wgid.x;
+  let blockIndex = wgid.y;
+  let blockStart = blockIndex * BLOCK_SIZE;
+  var blockEnd = blockStart + BLOCK_SIZE;
+  if (blockEnd > params.n) { blockEnd = params.n; }
+  let blockLen = blockEnd - blockStart;
+
+  if (col >= blockLen) { return; }
+
+  let ainvBase = blockIndex * BLOCK_SIZE * BLOCK_SIZE;
+  let forward = (params.trans == 0u) == (params.uplo == 0u);
+
+  if forward {
+    for (var r = lid.x; r < col; r += WGS) {
+      AinvHi[ainvBase + r * BLOCK_SIZE + col] = 0.0;
+      AinvLo[ainvBase + r * BLOCK_SIZE + col] = 0.0;
+    }
+  } else {
+    for (var r = col + 1u + lid.x; r < blockLen; r += WGS) {
+      AinvHi[ainvBase + r * BLOCK_SIZE + col] = 0.0;
+      AinvLo[ainvBase + r * BLOCK_SIZE + col] = 0.0;
+    }
+  }
+  storageBarrier();
+  workgroupBarrier();
+
+  let numSteps = select(col + 1u, blockLen - col, forward);
+  for (var step = 0u; step < numSteps; step++) {
+    let localRow = select(col - step, col + step, forward);
+    let i = blockStart + localRow;
+
+    // Accumulation range over lj: [col, localRow) forward, [localRow+1, col+1) backward.
+    // Range length varies per step, but every thread runs the SAME step at
+    // the same time, so the main/tail split is already uniform for this
+    // step \u2014 just not the same across different steps (see dsyr.wgsl).
+    let ljStart = select(localRow + 1u, col, forward);
+    let ljEnd   = select(col + 1u, localRow, forward);
+    let rangeLen    = ljEnd - ljStart;
+    let range_floor = ljStart + (rangeLen / WGS) * WGS;
+    let mainIters   = (range_floor - ljStart) / WGS;
+    let hasTail     = select(0u, 1u, range_floor < ljEnd);
+
+    var acc = DD(0.0, 0.0);
+
+    for (var iter = 0u; iter < mainIters; iter++) {
+      let lj = ljStart + lid.x + iter * WGS;
+      let aij = readA(i, blockStart + lj);
+      let aidx = ainvBase + lj * BLOCK_SIZE + col;
+      let ainv = DD(AinvHi[aidx], AinvLo[aidx]);
+      let prod = ddMulProtected(aij, ainv, lid.x);
+      acc = ddAddProtected(acc, prod, lid.x);
+    }
+
+    for (var iter = 0u; iter < hasTail; iter++) {
+      let lj    = range_floor + lid.x;
+      let valid = lj < ljEnd;
+      let safeLj = select(ljStart, lj, valid); // ljStart is always in-range when rangeLen > 0, the only case hasTail can be 1
+      let aij = readA(i, blockStart + safeLj);
+      let aidx = ainvBase + safeLj * BLOCK_SIZE + col;
+      let ainv = DD(AinvHi[aidx], AinvLo[aidx]);
+      let prod = ddMulProtected(aij, ainv, lid.x);
+      let contribution = DD(select(0.0, prod.hi, valid), select(0.0, prod.lo, valid));
+      acc = ddAddProtected(acc, contribution, lid.x);
+    }
+
+    tile[lid.x] = acc;
+    workgroupBarrier();
+
+    // Inactive threads combine against a throwaway partner and discard it
+    // (ddAddProtected must be called unconditionally by every thread).
+    for (var s = WGS / 2u; s > 0u; s >>= 1u) {
+      let partner = select(lid.x, lid.x + s, lid.x < s);
+      let combined = ddAddProtected(tile[lid.x], tile[partner], lid.x);
+      workgroupBarrier(); // all threads must read tile[] above before any write below
+      if (lid.x < s) { tile[lid.x] = combined; }
+      workgroupBarrier();
+    }
+
+    // tile[0] now holds the full accumulation, visible to every lane \u2014 every
+    // thread redundantly finishes the O(1) rhs/division so the protected
+    // ops below stay uniformly called (only the write is gated).
+    let e = DD(select(0.0, 1.0, localRow == col), 0.0);
+    let rhs = ddSubProtected(e, tile[0], lid.x);
+
+    var val: DD;
+    if params.diag == 1u {
+      val = rhs;
+    } else {
+      val = ddDivProtected(rhs, readA(i, i), lid.x);
+    }
+
+    if (lid.x == 0u) {
+      AinvHi[ainvBase + localRow * BLOCK_SIZE + col] = val.hi;
+      AinvLo[ainvBase + localRow * BLOCK_SIZE + col] = val.lo;
+    }
+    storageBarrier();
+    workgroupBarrier();
+  }
+}
+`});var Co,To=Z(()=>{Co=`// dtrsv_apply_inverse: double-double (Dekker) f64 emulation of
+// strsv_apply_inverse.wgsl \u2014 given a precomputed block inverse (from
+// dtrsv_invert_block.wgsl), computes this block's solution as a dense
+// matrix-vector multiply against the block's current remainder in x.
+// Ainv, x, and the per-row accumulation are each split into an f32 (hi, lo)
+// pair; WGSL has no f64 type.
+//
+// Unlike dsymv/dtrmv's per-row reduction across 64 threads, each thread here
+// owns one whole row's dot product independently (blockLen <= WGS, so one
+// thread per row already covers it, no tree reduction needed) \u2014 same shape
+// as dgemv_t.wgsl's per-thread accumulation. The f32 original early-returns
+// threads with \`lid.x >= blockLen\` (the last, possibly-short block); DD
+// can't do that, since every thread must call ddMulProtected/ddAddProtected
+// the same number of times. Instead every thread runs the identical
+// \`blockLen\`-iteration loop (reading past-blockLen rows of Ainv, which the
+// zero-initialized, never-written buffer backing them makes safe garbage \u2014
+// see dgemv_t.wgsl's clamped-dummy-index technique) and only the final
+// write is gated.
+
+@group(0) @binding(0) var<storage, read>       AinvHi: array<f32>;
+@group(0) @binding(1) var<storage, read>       AinvLo: array<f32>;
+@group(0) @binding(2) var<storage, read_write> xHi: array<f32>;
+@group(0) @binding(3) var<storage, read_write> xLo: array<f32>;
+
+struct Params {
+  incx:       u32,
+  blockIndex: u32,
+  blockStart: u32,
+  blockEnd:   u32,
+}
+
+@group(0) @binding(4) var<uniform> params: Params;
+
+const BLOCK_SIZE: u32 = 64u;
+var<workgroup> xLocal: array<DD, 64>;
+
+@compute @workgroup_size(64)
+fn dtrsv_apply_inverse_main(@builtin(local_invocation_id) lid: vec3u) {
+  let blockLen = params.blockEnd - params.blockStart;
+
+  if (lid.x < blockLen) {
+    let ix = (params.blockStart + lid.x) * params.incx;
+    xLocal[lid.x] = DD(xHi[ix], xLo[ix]);
+  }
+  workgroupBarrier();
+
+  let ainvBase = params.blockIndex * BLOCK_SIZE * BLOCK_SIZE;
+  var acc = DD(0.0, 0.0);
+  for (var j = 0u; j < blockLen; j++) {
+    let aidx = ainvBase + lid.x * BLOCK_SIZE + j;
+    let prod = ddMulProtected(DD(AinvHi[aidx], AinvLo[aidx]), xLocal[j], lid.x);
+    acc = ddAddProtected(acc, prod, lid.x);
+  }
+
+  if (lid.x < blockLen) {
+    let ix = (params.blockStart + lid.x) * params.incx;
+    xHi[ix] = acc.hi;
+    xLo[ix] = acc.lo;
+  }
+}
+`});var Vo,Wo=Z(()=>{Vo=`// dtrsv_update: double-double (Dekker) f64 emulation of strsv_update.wgsl \u2014
+// subtracts a solved block's contribution from every remaining row in
+// parallel (one workgroup per row, like dtrmv.wgsl). A, x, and the per-row
+// accumulation are each split into an f32 (hi, lo) pair; WGSL has no f64
+// type. No diag/masking needed: this region never touches the diagonal.
+//
+// The column range [blockStart, blockEnd) is the same fixed width for every
+// row in a dispatch (not per-row-varying the way dtrmv's triangular range
+// is), so the main/tail split is computed once, outside the row loop \u2014 same
+// shape as dger.wgsl's column loop.
+
+@group(0) @binding(0) var<storage, read>       AHi: array<f32>;
+@group(0) @binding(1) var<storage, read>       ALo: array<f32>;
+@group(0) @binding(2) var<storage, read_write> xHi: array<f32>;
+@group(0) @binding(3) var<storage, read_write> xLo: array<f32>;
+
+struct Params {
+  n:          u32,
+  incx:       u32,
+  lda:        u32,
+  trans:      u32,  // 0 = no-transpose, 1 = transpose
+  uplo:       u32,  // 0 = lower, 1 = upper
+  blockStart: u32,
+  blockEnd:   u32,  // exclusive
+}
+
+@group(0) @binding(4) var<uniform> params: Params;
+
+const WGS: u32 = 64u;
+var<workgroup> tile: array<DD, 64>;
+
+@compute @workgroup_size(64)
+fn dtrsv_update_main(
+  @builtin(workgroup_id)        wgid: vec3u,
+  @builtin(local_invocation_id) lid:  vec3u,
+  @builtin(num_workgroups)      nwg:  vec3u,
+) {
+  // forward: remaining rows are [blockEnd,n); backward: [0,blockStart).
+  // Uniform across the whole dispatch (derived from params only), so an
+  // early return here is safe \u2014 never a partial-workgroup divergence.
+  let forward = (params.trans == 0u) == (params.uplo == 0u);
+
+  var rangeStart: u32;
+  var rangeEnd: u32;
+  if forward {
+    rangeStart = params.blockEnd;
+    rangeEnd = params.n;
+  } else {
+    rangeStart = 0u;
+    rangeEnd = params.blockStart;
+  }
+
+  if (rangeStart >= rangeEnd) { return; }
+  let count = rangeEnd - rangeStart;
+
+  let colCount    = params.blockEnd - params.blockStart;
+  let col_floor   = (colCount / WGS) * WGS;
+  let mainIters   = col_floor / WGS;
+  let hasTail     = select(0u, 1u, col_floor < colCount);
+
+  for (var idx = wgid.x; idx < count; idx += nwg.x) {
+    let i = rangeStart + idx;
+    var acc = DD(0.0, 0.0);
+
+    for (var iter = 0u; iter < mainIters; iter++) {
+      let j = params.blockStart + lid.x + iter * WGS;
+      let aidx = select(j * params.lda + i, i * params.lda + j, params.trans == 0u);
+      let ix = j * params.incx;
+      let prod = ddMulProtected(DD(AHi[aidx], ALo[aidx]), DD(xHi[ix], xLo[ix]), lid.x);
+      acc = ddAddProtected(acc, prod, lid.x);
+    }
+
+    for (var iter = 0u; iter < hasTail; iter++) {
+      let j     = params.blockStart + col_floor + lid.x;
+      let valid = j < params.blockEnd;
+      let safeJ = select(params.blockStart, j, valid);
+      let aidx  = select(safeJ * params.lda + i, i * params.lda + safeJ, params.trans == 0u);
+      let ix    = safeJ * params.incx;
+      let prod  = ddMulProtected(DD(AHi[aidx], ALo[aidx]), DD(xHi[ix], xLo[ix]), lid.x);
+      let contribution = DD(select(0.0, prod.hi, valid), select(0.0, prod.lo, valid));
+      acc = ddAddProtected(acc, contribution, lid.x);
+    }
+
+    tile[lid.x] = acc;
+    workgroupBarrier();
+
+    // Inactive threads combine against a throwaway partner and discard it
+    // (ddAddProtected must be called unconditionally by every thread).
+    for (var s = WGS / 2u; s > 0u; s >>= 1u) {
+      let partner = select(lid.x, lid.x + s, lid.x < s);
+      let combined = ddAddProtected(tile[lid.x], tile[partner], lid.x);
+      workgroupBarrier(); // all threads must read tile[] above before any write below
+      if (lid.x < s) { tile[lid.x] = combined; }
+      workgroupBarrier();
+    }
+
+    let ix_i = i * params.incx;
+    let xVal = DD(xHi[ix_i], xLo[ix_i]);
+    // tile[0] now holds the full subtracted term, visible to every lane \u2014
+    // every thread redundantly finishes the O(1) subtraction so the
+    // protected op below stays uniformly called (only the write is gated).
+    let result = ddSubProtected(xVal, tile[0], lid.x);
+
+    if (lid.x == 0u) {
+      xHi[ix_i] = result.hi;
+      xLo[ix_i] = result.lo;
+    }
+    // All 64 threads must agree before the next row (grid-stride) reuses tile[].
+    workgroupBarrier();
+  }
+}
+`});var ue,Oo=Z(()=>{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
@@ -2899,7 +3924,7 @@ fn main(
     }
   }
 }
-`});var fe,Ao=O(()=>{fe=`// sgemm_large: C = alpha * op(A) * op(B) + beta * C \u2014 large-tile half of
+`});var de,Ko=Z(()=>{de=`// 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,
@@ -3105,7 +4130,7 @@ fn main(
     }
   }
 }
-`});var xe,Go=O(()=>{xe=`// sgemmtr_small: C := uplo(alpha * op(A) * op(B) + beta * C) \u2014 small-tile
+`});var xe,Uo=Z(()=>{xe=`// sgemmtr_small: C := uplo(alpha * op(A) * op(B) + beta * C) \u2014 small-tile
 // half of a two-tier dispatch, identical to sgemm_small.wgsl except the
 // final output write is gated to one triangle of C by \`uplo\` \u2014 see
 // sgemmtr_large.wgsl for the full rationale (shared by both tiers).
@@ -3218,7 +4243,7 @@ fn main(
     }
   }
 }
-`});var ve,Eo=O(()=>{ve=`// sgemmtr_large: C := uplo(alpha * op(A) * op(B) + beta * C) \u2014 large-tile
+`});var ve,zo=Z(()=>{ve=`// sgemmtr_large: C := uplo(alpha * op(A) * op(B) + beta * C) \u2014 large-tile
 // half of a two-tier dispatch, identical to sgemm_large.wgsl (see that file
 // for the BM/BN/BK/TM/TN autotuning rationale) except the final output write
 // is gated to one triangle of C by \`uplo\`, the same convention ssyr/ssyr2
@@ -3338,7 +4363,7 @@ fn main(
     }
   }
 }
-`});var Do,ko=O(()=>{Do=`// symmetrize: Adense := full dense expansion of a symmetric matrix stored
+`});var Zo,Yo=Z(()=>{Zo=`// symmetrize: Adense := full dense expansion of a symmetric matrix stored
 // with only its \`uplo\` triangle meaningful (the other triangle is implied
 // by symmetry: A[i,j] = A[j,i]). A plain element-wise pass, no tiling or
 // shared memory needed \u2014 used to materialize a dense operand for routines
@@ -3369,7 +4394,7 @@ fn main(@builtin(global_invocation_id) gid: vec3u) {
   let srcIdx = select(col * params.lda + row, row * params.lda + col, isStored);
   Adense[row * params.ldd + col] = A[srcIdx];
 }
-`});var Po,No=O(()=>{Po=`// triangularize: Adense := dense expansion of op(A) (A or A^T per \`trans\`),
+`});var $o,Xo=Z(()=>{$o=`// triangularize: Adense := dense expansion of op(A) (A or A^T per \`trans\`),
 // zero-filling the unstored triangle (exact for a matmul) so strmm can reuse
 // sgemm's kernel unchanged. \`diag=1\` substitutes 1.0 on the diagonal.
 
@@ -3413,7 +4438,7 @@ fn main(@builtin(global_invocation_id) gid: vec3u) {
 
   Adense[row * params.ldd + col] = select(0.0, A[srcRow * params.lda + srcCol], isMeaningful);
 }
-`});var Io,Mo=O(()=>{Io=`// block_transfer: gather/scatter/scatter-subtract between a tight (blockLen
+`});var Qo,Jo=Z(()=>{Qo=`// block_transfer: gather/scatter/scatter-subtract between a tight (blockLen
 // x otherLen) block and a sub-range of a strided (any ld, row/col-major)
 // buffer \u2014 needed since block offsets aren't 256-byte-aligned and block
 // rows/cols aren't always one contiguous range for copyBufferToBuffer.
@@ -3455,8 +4480,8 @@ 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(()=>{$e();Ze();Je();et();ot();it();nt();ut();mt();dt();pt();wt();bt();xt();_t();Bt();St();At();Et();kt();Dt();Pt();It();Rt();qt();Ct();jt();Ht();Vt();zt();Yt();$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 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($),!g&&K&&d(K),Y&&d(Y),d(lr),d(ir)}}return Ia(Ti);})();
+`});var ra={};Fe(ra,{routineShaders:()=>ir,shaderSources:()=>bs});var ir,bs,ea=Z(()=>{$e();Je();rt();tt();at();st();lt();dt();mt();ct();wt();ht();yt();vt();Bt();At();St();Gt();kt();Dt();Lt();Nt();It();jt();Ft();Ht();Ct();Vt();Kt();zt();Zt();$t();Jt();ro();to();ao();so();lo();fo();co();po();go();bo();xo();_o();Ao();Go();ko();Lo();No();Io();jo();qo();To();Wo();Oo();Ko();Uo();zo();Yo();Xo();Jo();ir={};ir.sscal={sscal:ke};ir.cscal={cscal:Qe};ir.sswap={sswap:et};ir.dswap={dswap:ot};ir.saxpy={saxpy:it};ir.scopy={scopy:nt};ir.dcopy={dcopy:ut};ir.sdot={sdot:ft,"reduction/sum":De};ir.sasum={sasum:pt,"reduction/sum":De};ir.snrm2={snrm2:gt,"reduction/scaledSum":bt};ir.isamax={isamax:xt,"reduction/argmax":_t};ir.dasum={"f64/dekker":Pr,"f64/utils/abs":ye,"f64/utils/add":Ir,dasum:Et,"reduction/sumF64":Le};ir.ddot={"f64/dekker":Pr,"f64/utils/add":Ir,"f64/utils/multiply":Cr,ddot:Pt,"reduction/sumF64":Le};ir.dscal={"f64/dekker":Pr,"f64/utils/add":Ir,"f64/utils/multiply":Cr,dscal:Mt};ir.daxpy={"f64/dekker":Pr,"f64/utils/add":Ir,"f64/utils/multiply":Cr,daxpy:Rt};ir.idamax={"f64/dekker":Pr,"f64/utils/abs":ye,"f64/utils/greater":Pe,"f64/utils/equal":qt,idamax:Tt,"reduction/argmaxF64":Wt};ir.srot={srot:Ot};ir.drot={"f64/dekker":Pr,"f64/utils/add":Ir,"f64/utils/multiply":Cr,drot:Ut};ir.srotm={srotm:Yt};ir.drotm={"f64/dekker":Pr,"f64/utils/add":Ir,"f64/utils/multiply":Cr,drotm:Xt};ir.dnrm2={"f64/dekker":Pr,"f64/utils/abs":ye,"f64/utils/greater":Pe,"f64/utils/add":Ir,"f64/utils/multiply":Cr,"f64/utils/divide":Ne,"f64/utils/sqrt":Qt,dnrm2:eo,"reduction/scaledSumF64":oo};ir.sgemv={sgemv_n:io,sgemv_t:no};ir.ssymv={ssymv:uo};ir.strmv={strmv:mo};ir.strsv={strsv_invert_block:Me,strsv_apply_inverse:wo,strsv_update:ho};ir.sger={sger:yo};ir.dger={"f64/dekker":Pr,"f64/utils/add":Ir,"f64/utils/multiply":Cr,dger:vo};ir.ssyr={ssyr:Bo};ir.dsyr={"f64/dekker":Pr,"f64/utils/add":Ir,"f64/utils/multiply":Cr,dsyr:So};ir.ssyr2={ssyr2:Eo};ir.dsyr2={"f64/dekker":Pr,"f64/utils/add":Ir,"f64/utils/multiply":Cr,dsyr2:Do};ir.dgemv={"f64/dekker":Pr,"f64/utils/add":Ir,"f64/utils/multiply":Cr,dgemv_n:Po,dgemv_t:Mo};ir.dsymv={"f64/dekker":Pr,"f64/utils/add":Ir,"f64/utils/multiply":Cr,dsymv:Ro};ir.dtrmv={"f64/dekker":Pr,"f64/utils/add":Ir,"f64/utils/multiply":Cr,dtrmv:Fo};ir.dtrsv={"f64/dekker":Pr,"f64/utils/add":Ir,"f64/utils/multiply":Cr,"f64/utils/divide":Ne,dtrsv_invert_block:Ho,dtrsv_apply_inverse:Co,dtrsv_update:Vo};ir.sgemm={sgemm_small:ue,sgemm_large:de};ir.sgemmtr={sgemmtr_small:xe,sgemmtr_large:ve};ir.ssyrk={sgemmtr_small:xe,sgemmtr_large:ve};ir.ssyr2k={sgemmtr_small:xe,sgemmtr_large:ve};ir.ssymm={sgemm_small:ue,sgemm_large:de,symmetrize:Zo};ir.strmm={sgemm_small:ue,sgemm_large:de,triangularize:$o};ir.strsm={strsv_invert_block:Me,block_transfer:Qo,sscal:ke,sgemm_small:ue,sgemm_large:de};bs=Object.assign({},...Object.values(ir))});var _s={};Fe(_s,{Complex32:()=>Or,Complex32Array:()=>Br,Complex64:()=>Vr,Complex64Array:()=>Lr,GpuMatrix:()=>X,GpuVector:()=>M,cleanup:()=>Oe,cscal:()=>oa,dasum:()=>ca,daxpy:()=>la,dcopy:()=>da,ddot:()=>pa,dgemv:()=>Ra,dger:()=>La,dnrm2:()=>ga,drot:()=>xa,drotm:()=>_a,dscal:()=>aa,dswap:()=>sa,dsymv:()=>ja,dsyr:()=>Na,dsyr2:()=>Ia,dtrmv:()=>Fa,dtrsv:()=>Ta,gpuName:()=>Ke,idamax:()=>ba,init:()=>Ve,isamax:()=>ha,randomFloat32Array:()=>Ye,randomFloat64Array:()=>Ze,randomTriangularFloat32Array:()=>Xe,sasum:()=>ma,saxpy:()=>na,scopy:()=>ua,sdot:()=>fa,sgemm:()=>Ca,sgemmtr:()=>Wa,sgemv:()=>Ba,sger:()=>Da,snrm2:()=>wa,srot:()=>ya,srotm:()=>va,sscal:()=>ta,sswap:()=>ia,ssymm:()=>Ka,ssymv:()=>Aa,ssyr:()=>Pa,ssyr2:()=>Ma,ssyr2k:()=>Oa,ssyrk:()=>Va,strmm:()=>Ua,strmv:()=>Sa,strsm:()=>za,strsv:()=>ka});function He(r,o){return o?r.features.has("timestamp-query")?{requiredFeatures:["timestamp-query"]}:(console.warn("timestamp-query not supported on this device \u2014 benchmark mode disabled."),{}):{}}function Te(r){if(!Ce(r))return{querySet:null,passDescriptor:void 0};let o=r.createQuerySet({type:"timestamp",count:2});return{querySet:o,passDescriptor:{timestampWrites:{querySet:o,beginningOfPassWriteIndex:0,endOfPassWriteIndex:1}}}}function kr(r,o,e){if(!e)return null;let a=r.createBuffer({label:"timestamp-resolve",size:16,usage:GPUBufferUsage.QUERY_RESOLVE|GPUBufferUsage.COPY_SRC});o.resolveQuerySet(e,0,2,a,0);let t=r.createBuffer({label:"timestamp-readback",size:16,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});return o.copyBufferToBuffer(a,0,t,0,16),{tsReadBuffer:t,resolveBuffer:a,querySet:e}}async function j(r){if(!r)return;let{tsReadBuffer:o,resolveBuffer:e,querySet:a}=r;await o.mapAsync(GPUMapMode.READ);let t=new BigInt64Array(o.getMappedRange().slice());return o.unmap(),o.destroy(),e.destroy(),a.destroy(),Math.max(0,Number(t[1]-t[0]))/1e6}var Jr=null,Se=!1,Qr=new Map,le=new WeakMap,zr=null,We=({powerPreference:r,benchmark:o})=>`${r}::${o}`;async function Ve({powerPreference:r="high-performance",benchmark:o=!1,dumpShaders:e=!1}={}){let a={powerPreference:r,benchmark:o,dumpShaders:e},t=We(a),i=Qr.get(t);if(i)return i;if(Jr)e!==Se&&typeof window>"u"&&console.warn(`dumpShaders: ${e} was requested, but the WebGPU instance was already created with dumpShaders: ${Se}. The first init() call fixes this for the process.`);else if(typeof window>"u"){let{create:m,globals:w}=await import("webgpu");Object.assign(globalThis,w),Jr=m(e?["enable-dawn-features=dump_shaders,disable_symbol_renaming"]:[]),Se=e}else e&&console.warn("dumpShaders has no effect in the browser \u2014 see init()'s docs."),Jr=navigator.gpu;if(!Jr)throw new Error("WebGPU not supported in this environment.");let s=await Jr.requestAdapter({powerPreference:r})??await Jr.requestAdapter();if(!s)throw new Error("No WebGPU adapter found.");let n=[...He(s,o).requiredFeatures??[]],d=await s.requestDevice({requiredFeatures:n});d.addEventListener("uncapturederror",m=>{console.error("Uncaptured GPU error:",m.error.message)});let l=n.includes("timestamp-query");return le.set(d,{adapter:s,benchmark:l,options:a}),Qr.set(t,d),zr||(zr=d),d}function Oe(r){if(r===void 0){for(let e of Qr.values())e.destroy();Qr.clear(),zr=null;return}let o=le.get(r);o&&(Qr.delete(We(o.options)),le.delete(r),r.destroy(),zr===r&&(zr=Qr.values().next().value??null))}function Ke(r=zr){let o=r&&le.get(r);if(!o)throw new Error("WebGPU adapter not initialized \u2014 call init() first.");let{device:e,description:a}=o.adapter.info;return{description:a||"unknown",device:e||"unknown"}}function Ce(r=zr){return le.get(r)?.benchmark??!1}function re(){if(!zr)throw new Error("WebGPU device not initialized \u2014 call init() first.");return zr}function f(...r){r.flat().forEach(o=>o.destroy())}function Ge(r,o,e){let a=r.limits.maxStorageBufferBindingSize;if(o>a)throw new Error(`Buffer "${e}" needs ${o} bytes, exceeding this device's maxStorageBufferBindingSize (${a} bytes). The operands are too large for this device.`)}function h(r,o,e="blas-input",a=!1){let t=o.byteLength;Ge(r,t,e);let i=a?GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC:GPUBufferUsage.STORAGE,s=r.createBuffer({label:e,size:t,usage:i,mappedAtCreation:!0}),u=o.constructor;return new u(s.getMappedRange()).set(o),s.unmap(),s}function fr(r,o,e="blas-storage",a=0){return Ge(r,o,e),r.createBuffer({label:e,size:o,usage:GPUBufferUsage.STORAGE|a})}function Ar(r,o,e="blas-result"){return Ge(r,o,e),r.createBuffer({label:e,size:o,usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC})}function E(r,o,e){let a=r.createBuffer({label:"blas-readback",size:e.size,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});return o.copyBufferToBuffer(e,0,a,0,e.size),a}var ee=16,Ue=new WeakMap;function ui(r){let o=Ue.get(r);return o||(o=r.createBuffer({label:"blas-vec4-fallback",size:ee,usage:GPUBufferUsage.STORAGE}),Ue.set(r,o)),o}function Dr(r,o){let e=o instanceof GPUBuffer?o:o.buffer,a=o instanceof GPUBuffer?0:o.offset??0,t=o instanceof GPUBuffer?o.size:o.size??e.size-a,i=Math.floor(t/ee)*ee;return i<ee?{buffer:ui(r),offset:0,size:ee}:{buffer:e,offset:a,size:i}}function Ee(r,o,e,a){if(o%4!==0)return!1;let t=r instanceof GPUBuffer?r:r.buffer,i=r instanceof GPUBuffer?0:r.offset??0,s=r instanceof GPUBuffer?t.size:r.size??t.size-i,u=Math.floor(s/ee)*4;if(u<=0)return!1;let n=(Math.max(e,1)-1)*o+(Math.max(a,1)-1);return Math.floor(n/4)*4+4<=u}function q(r,o,e="blas-params"){let a=o.length*4,t=Math.ceil(a/16)*16,i=new ArrayBuffer(t),s=new DataView(i);o.forEach(({value:n,type:d},l)=>{let m=l*4;if(d==="u32")s.setUint32(m,n,!0);else if(d==="i32")s.setInt32(m,n,!0);else if(d==="f32")s.setFloat32(m,n,!0);else throw new Error(`Unknown param type "${d}". Use "f32", "u32", or "i32".`)});let u=r.createBuffer({label:e,size:t,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});return r.queue.writeBuffer(u,0,i),u}async function G(r,o=Float32Array){try{await r.mapAsync(GPUMapMode.READ);let e=new o(r.getMappedRange().slice());return r.unmap(),e}finally{r.destroy()}}function U(r){let o=r.length,e=new Float32Array(o),a=new Float32Array(o);for(let t=0;t<o;t++){let i=Math.fround(r[t]);e[t]=i,a[t]=Math.fround(r[t]-i)}return{hi:e,lo:a}}function dr(r,o){let e=r.length,a=new Float64Array(e);for(let t=0;t<e;t++)a[t]=r[t]+o[t];return a}var Vr=class{constructor(o,e){this.re=o,this.im=e}},Lr=class extends Array{constructor(o){if(o===void 0){super();return}if(typeof o=="number"){super(o);for(let a=0;a<o;a++)this[a]=new Vr(0,0);return}let e=Array.from(o);if(super(),e.length!==0){if(e[0]instanceof Vr){for(let a of e){if(!(a instanceof Vr))throw new Error("Complex64Array expects every element to be a Complex64.");this.push(a)}return}if(e.length%2!==0)throw new Error("Complex64Array expects an even number of interleaved [re, im, ...] values.");for(let a=0;a<e.length;a+=2){if(typeof e[a]!="number"||typeof e[a+1]!="number")throw new Error("Complex64Array expects interleaved [re, im, ...] values to be numbers.");this.push(new Vr(e[a],e[a+1]))}}}};function te(r,o=r.length){let e=new Float32Array(o*2);for(let a=0;a<o;a++)e[a*2]=r[a].re,e[a*2+1]=r[a].im;return e}function he(r,o=r.length){let e=new Float64Array(o),a=new Float64Array(o);for(let l=0;l<o;l++)e[l]=r[l].re,a[l]=r[l].im;let{hi:t,lo:i}=U(e),{hi:s,lo:u}=U(a),n=new Float32Array(o*2),d=new Float32Array(o*2);for(let l=0;l<o;l++)n[l*2]=t[l],n[l*2+1]=s[l],d[l*2]=i[l],d[l*2+1]=u[l];return{hi:n,lo:d}}function be(r,o){let e=r.length/2,a=new Float32Array(e),t=new Float32Array(e),i=new Float32Array(e),s=new Float32Array(e);for(let l=0;l<e;l++)a[l]=r[l*2],i[l]=r[l*2+1],t[l]=o[l*2],s[l]=o[l*2+1];let u=dr(a,t),n=dr(i,s),d=new Lr(e);for(let l=0;l<e;l++)d[l]=new Vr(u[l],n[l]);return d}var Or=class{constructor(o,e){this.re=Math.fround(o),this.im=Math.fround(e)}},Br=class extends Array{constructor(o){if(o===void 0){super();return}if(typeof o=="number"){super(o);for(let a=0;a<o;a++)this[a]=new Or(0,0);return}let e=Array.from(o);if(super(),e.length!==0){if(e[0]instanceof Or){for(let a of e){if(!(a instanceof Or))throw new Error("Complex32Array expects every element to be a Complex32.");this.push(a)}return}if(e.length%2!==0)throw new Error("Complex32Array expects an even number of interleaved [re, im, ...] values.");for(let a=0;a<e.length;a+=2){if(typeof e[a]!="number"||typeof e[a+1]!="number")throw new Error("Complex32Array expects interleaved [re, im, ...] values to be numbers.");this.push(new Or(e[a],e[a+1]))}}}};var M=class r{constructor(o,e,a=Float32Array,t=null,i=null){this._buf=o,this._loBuf=t,this.length=e,this.dtype=a,this.device=i??re()}static from(o,e){let a=o instanceof GPUDevice,t=a?o:re(),i=a?e:o;if(i instanceof Float64Array){let{hi:u,lo:n}=U(i),d=h(t,u,"gpu-vector-f64-hi",!0),l=h(t,n,"gpu-vector-f64-lo",!0);return new r(d,i.length,Float64Array,l,t)}if(i instanceof Br){let u=h(t,te(i),"gpu-vector-complex32",!0);return new r(u,i.length,Br,null,t)}if(i instanceof Lr){let{hi:u,lo:n}=he(i),d=h(t,u,"gpu-vector-complex64-hi",!0),l=h(t,n,"gpu-vector-complex64-lo",!0);return new r(d,i.length,Lr,l,t)}if(!(i instanceof Float32Array))throw new Error("GpuVector.from expects a Float32Array, Float64Array, Complex32Array, or Complex64Array.");let s=h(t,i,"gpu-vector",!0);return new r(s,i.length,i.constructor,null,t)}async read(){let o=this.device,e=o.createCommandEncoder(),a=E(o,e,this._buf);if(o.queue.submit([e.finish()]),this.dtype===Br)return new Br(await G(a,Float32Array));if(!this._loBuf)return G(a,this.dtype);let t=o.createCommandEncoder(),i=E(o,t,this._loBuf);o.queue.submit([t.finish()]);let[s,u]=await Promise.all([G(a,Float32Array),G(i,Float32Array)]);return this.dtype===Lr?be(s,u):dr(s,u)}destroy(){this._buf.destroy(),this._loBuf&&this._loBuf.destroy()}};var X=class r{constructor(o,e,a,t,i=null,s="row-major",u=null,n=Float32Array){this._buf=o,this._loBuf=i,this.rows=e,this.cols=a,this.lda=t,this.layout=s,this.dtype=n,this.device=u??re()}static from(o,...e){let a=o instanceof GPUDevice,t=a?o:re(),i=a?e.shift():o,[s,u,n,d="row-major"]=e;if(d!=="row-major"&&d!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");let l=d==="row-major";if(n===void 0&&(n=l?u:s),!(i instanceof Float32Array)&&!(i instanceof Float64Array)&&!(i instanceof Br)&&!(i instanceof Lr))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 w=l?s:u;if(i.length<w*n)throw new Error("data does not have enough elements for the given rows, cols, and lda.");if(i instanceof Float64Array){let p=w*n,{hi:g,lo:y}=U(i.subarray(0,p)),b=h(t,g,"gpu-matrix-f64-hi",!0),x=h(t,y,"gpu-matrix-f64-lo",!0);return new r(b,s,u,n,x,d,t,Float64Array)}if(i instanceof Br){let p=h(t,te(i,w*n),"gpu-matrix-complex32",!0);return new r(p,s,u,n,null,d,t,Br)}if(i instanceof Lr){let{hi:p,lo:g}=he(i,w*n),y=h(t,p,"gpu-matrix-complex64-hi",!0),b=h(t,g,"gpu-matrix-complex64-lo",!0);return new r(y,s,u,n,b,d,t,Lr)}let c=h(t,i.subarray(0,w*n),"gpu-matrix",!0);return new r(c,s,u,n,null,d,t)}async read(){let o=this.device,e=o.createCommandEncoder(),a=E(o,e,this._buf);o.queue.submit([e.finish()]);let t=this.layout!=="column-major",i=t?this.rows:this.cols,s=t?this.cols:this.rows;if(this.dtype===Br){let d=new Br(await G(a,Float32Array));if(this.lda===s)return d;let l=new Br(i*s);for(let m=0;m<i;m++)for(let w=0;w<s;w++)l[m*s+w]=d[m*this.lda+w];return l}if(this._loBuf){let d=o.createCommandEncoder(),l=E(o,d,this._loBuf);o.queue.submit([d.finish()]);let[m,w]=await Promise.all([G(a,Float32Array),G(l,Float32Array)]);if(this.dtype===Lr){let g=be(m,w);if(this.lda===s)return g;let y=new Lr(i*s);for(let b=0;b<i;b++)for(let x=0;x<s;x++)y[b*s+x]=g[b*this.lda+x];return y}let c=dr(m,w);if(this.lda===s)return c;let p=new Float64Array(i*s);for(let g=0;g<i;g++)p.set(c.subarray(g*this.lda,g*this.lda+s),g*s);return p}let u=await G(a,Float32Array);if(this.lda===s)return u;let n=new Float32Array(i*s);for(let d=0;d<i;d++)n.set(u.subarray(d*this.lda,d*this.lda+s),d*s);return n}destroy(){this._buf.destroy(),this._loBuf&&this._loBuf.destroy()}};function ze(r){let o=r>>>0;return function(){o=o+1831565813|0;let e=Math.imul(o^o>>>15,1|o);return e=e+Math.imul(e^e>>>7,61|e)^e,((e^e>>>14)>>>0)/4294967296}}function Ye(r,o=-1,e=1,a){let t=new Float32Array(r),i=a===void 0?Math.random:ze(a);for(let s=0;s<r;s++)t[s]=o+i()*(e-o);return t}function Ze(r,o=-1,e=1,a){let t=new Float64Array(r),i=a===void 0?Math.random:ze(a);for(let s=0;s<r;s++)t[s]=o+i()*(e-o);return t}function Xe(r,o,e="lower",a=-1,t=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(o<r)throw new Error("lda must be >= n.");let n=u==="column-major",d=(m,w)=>n?w*o+m:m*o+w,l=new Float32Array(r*o);for(let m=0;m<r;m++){for(let w=0;w<r;w++){if(m===w)continue;(e==="lower"?w<m:w>m)&&(l[d(m,w)]=a+Math.random()*(t-a))}l[d(m,m)]=i+Math.random()*(s-i)}return l}function D(r,o,e,a=0){let t=e.map((i,s)=>({binding:a+s,resource:i instanceof GPUBuffer?{buffer:i}:i}));return r.createBindGroup({layout:o,entries:t})}function F(r,o){r.queue.submit([o.finish()])}function Mr(r){let{querySet:o,passDescriptor:e}=Te(r);return{commandEncoder:r.createCommandEncoder(),querySet:o,passDescriptor:e}}function hr(r,o,e,a,t){let i=r.beginComputePass(t);i.setPipeline(o),i.setBindGroup(0,e),typeof a=="number"?i.dispatchWorkgroups(a):i.dispatchWorkgroups(a.x,a.y,a.z??1),i.end()}function O(r,o,e,a){let{commandEncoder:t,querySet:i,passDescriptor:s}=Mr(r);hr(t,o,e,a,s);let u=kr(r,t,i);return{commandEncoder:t,ts:u}}var vs={},Ie=new WeakMap;async function P(r,o,e="main"){Ie.has(r)||Ie.set(r,new Map);let a=Ie.get(r),t=Array.isArray(o)?o:[o],i=`${t.join("+")}::${e}`;if(!a.has(i)){let s=xs(r,t,e).catch(u=>{throw a.delete(i),u});a.set(i,s)}return a.get(i)}async function ys(r){if(typeof process>"u"||!process.versions?.node){let{shaderSources:o}=await Promise.resolve().then(()=>(ea(),ra)),e=o[r];if(!e)throw new Error(`Shader "${r}" not found in browser bundle.`);return e}else{let{readFileSync:o}=await import("fs"),{fileURLToPath:e}=await import("url"),{dirname:a,join:t}=await import("path"),i=a(e(vs.url));return o(t(i,`../shaders/${r}.wgsl`),"utf8")}}async function xs(r,o,e="main"){let a=o.join("+"),t=await Promise.all(o.map(ys)),i=0,s=t.map((p,g)=>{let y=p.split(`
+`).length,b={name:o[g],startLine:i+1,endLine:i+y};return i+=y,b}),u=p=>{let g=p&&s.find(y=>p>=y.startLine&&p<=y.endLine);return g?`${g.name}.wgsl:${p-g.startLine+1}`:`line ${p}`},n=t.join(`
+`),d=r.createShaderModule({label:a,code:n}),m=(await d.getCompilationInfo()).messages.filter(p=>p.type==="error");if(m.length>0)throw new Error(`Shader "${a}" compilation failed:
+${m.map(p=>`  ${u(p.lineNum)}: ${p.message}`).join(`
+`)}`);let w=e==="main"?{module:d}:{module:d,entryPoint:e},c=r.createComputePipeline({label:a,layout:"auto",compute:w});return c._shaderModule=d,c}function br(r,o,e){let a=r.limits.maxComputeWorkgroupsPerDimension;return e===void 0?Math.min(Math.ceil(o/64),a):{x:Math.min(Math.ceil(e/8),a),y:Math.min(Math.ceil(o/8),a)}}function or(r,o,e,a="x"){let t=r.limits.maxComputeWorkgroupsPerDimension;if(o>t)throw new Error(`${e}: this problem needs ${o} workgroups in ${a}, but the device allows ${t} (maxComputeWorkgroupsPerDimension). The operands are too large for this device \u2014 split the operation into smaller blocks.`);return o}function Yr(r,o,e,a){return a===void 0?or(r,Math.ceil(e/64),o):{x:or(r,Math.ceil(a/8),o,"x"),y:or(r,Math.ceil(e/8),o,"y")}}function H(r){if(!(r instanceof GPUDevice))throw new Error("device must be a GPUDevice.")}function T(r,o,e){for(let[a,t]of Object.entries(e))if(!(!(t instanceof M)&&!(t instanceof X))&&t.device!==r)throw new Error(`${o}: ${a} 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 ta(r,o,e,a,t){let i=a instanceof M;if(H(r),T(r,"sscal",{x:a}),!Number.isInteger(o)||!Number.isInteger(t))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(t<=0)throw new Error("incx must be positive.");if(!(a instanceof Float32Array)&&!(a instanceof M))throw new Error("x must be a Float32Array or GpuVector.");if(o<=0)return i?{}:{x:a};if(a.length<(o-1)*t+1)throw new Error("x does not have enough elements for the given n and incx.");let s=await P(r,"sscal"),u=null,n=null,d=null;try{u=i?a._buf:h(r,a,"sscal-x",!0),n=q(r,[{value:o,type:"u32"},{value:e,type:"f32"},{value:t,type:"u32"}],"sscal-params");let l=D(r,s.getBindGroupLayout(0),[u,n]),{commandEncoder:m,ts:w}=O(r,s,l,br(r,o));d=i?null:E(r,m,u),F(r,m);let c=await j(w);if(i)return c!==void 0?{gpuTimeMs:c}:{};let p=await G(d,Float32Array);return d=null,c!==void 0?{x:p,gpuTimeMs:c}:{x:p}}finally{!i&&u&&f(u),n&&f(n),d&&f(d)}}async function oa(r,o,e,a,t){let i=a instanceof M;if(H(r),T(r,"cscal",{x:a}),!Number.isInteger(o)||!Number.isInteger(t))throw new Error("n and incx must be integers.");if(!(e instanceof Or))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(t<=0)throw new Error("incx must be positive.");if(!(a instanceof Br)&&!i)throw new Error("x must be a Complex32Array or GpuVector.");if(i&&a.dtype!==Br)throw new Error("x must be a Complex32Array-backed GpuVector.");if(o<=0)return i?{}:{x:a};if(a.length<(o-1)*t+1)throw new Error("x does not have enough elements for the given n and incx.");let s=await P(r,"cscal"),u=null,n=null,d=null;try{u=i?a._buf:h(r,te(a),"cscal-x",!0),n=q(r,[{value:o,type:"u32"},{value:e.re,type:"f32"},{value:e.im,type:"f32"},{value:t,type:"u32"}],"cscal-params");let l=D(r,s.getBindGroupLayout(0),[u,n]),{commandEncoder:m,ts:w}=O(r,s,l,br(r,o));d=i?null:E(r,m,u),F(r,m);let c=await j(w);if(i)return c!==void 0?{gpuTimeMs:c}:{};let p=await G(d,Float32Array);d=null;let g=new Br(p);return c!==void 0?{x:g,gpuTimeMs:c}:{x:g}}finally{!i&&u&&f(u),n&&f(n),d&&f(d)}}async function aa(r,o,e,a,t){let i=a instanceof M;if(H(r),!Number.isInteger(o)||!Number.isInteger(t))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 instanceof Float64Array)&&!i)throw new Error("x must be a Float64Array or GpuVector.");if(i&&a.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(t<=0)throw new Error("incx must be positive.");if(T(r,"dscal",{x:a}),o<=0)return i?{}:{x:a};if(a.length<(o-1)*t+1)throw new Error("x does not have enough elements for the given n and incx.");let u=await P(r,[...["f64/dekker","f64/utils/add","f64/utils/multiply"],"dscal"]),{hi:n,lo:d}=U(new Float64Array([e])),l=null,m=null,w=null,c=null,p=null;try{if(i)l=a._buf,m=a._loBuf;else{let{hi:S,lo:_}=U(a);l=h(r,S,"dscal-xHi",!0),m=h(r,_,"dscal-xLo",!0)}w=q(r,[{value:o,type:"u32"},{value:n[0],type:"f32"},{value:d[0],type:"f32"},{value:t,type:"u32"}],"dscal-params");let g=D(r,u.getBindGroupLayout(0),[l,m,w]),{commandEncoder:y,ts:b}=O(r,u,g,br(r,o));c=i?null:E(r,y,l),p=i?null:E(r,y,m),F(r,y);let x=await j(b);if(i)return x!==void 0?{gpuTimeMs:x}:{};let v=await G(c,Float32Array);c=null;let B=await G(p,Float32Array);p=null;let A=dr(v,B);return x!==void 0?{x:A,gpuTimeMs:x}:{x:A}}finally{!i&&l&&f(l),!i&&m&&f(m),w&&f(w),c&&f(c),p&&f(p)}}async function ia(r,o,e,a,t,i){let s=e instanceof M,u=t instanceof M;if(H(r),T(r,"sswap",{x:e,y:t}),!Number.isInteger(o)||!Number.isInteger(a)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(a<=0||i<=0)throw new Error("incx and incy must be positive.");if(!(e instanceof Float32Array)&&!(e instanceof M))throw new Error("x must be a Float32Array or GpuVector.");if(!(t instanceof Float32Array)&&!(t instanceof M))throw new Error("y must be a Float32Array or GpuVector.");if(e.constructor!==t.constructor)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(o<=0)return s?{}:{x:e,y:t};if(e.length<(o-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(o-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=await P(r,"sswap"),d=null,l=null,m=null,w=null,c=null;try{d=s?e._buf:h(r,e,"sswap-x",!0),l=u?t._buf:h(r,t,"sswap-y",!0),m=q(r,[{value:o,type:"u32"},{value:a,type:"u32"},{value:i,type:"u32"}],"sswap-params");let p=D(r,n.getBindGroupLayout(0),[d,l,m]),{commandEncoder:g,ts:y}=O(r,n,p,br(r,o));w=s?null:E(r,g,d),c=u?null:E(r,g,l),F(r,g);let b=await j(y);if(s)return b!==void 0?{gpuTimeMs:b}:{};let x=await G(w,Float32Array);w=null;let v=await G(c,Float32Array);return c=null,b!==void 0?{x,y:v,gpuTimeMs:b}:{x,y:v}}finally{!s&&d&&f(d),!u&&l&&f(l),m&&f(m),w&&f(w),c&&f(c)}}async function sa(r,o,e,a,t,i){let s=e instanceof M,u=t instanceof M;if(H(r),T(r,"dswap",{x:e,y:t}),!Number.isInteger(o)||!Number.isInteger(a)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(a<=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(!(t 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&&t.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(o<=0)return s?{}:{x:e,y:t};if(e.length<(o-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(o-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=await P(r,"dswap"),d=null,l=null,m=null,w=null,c=null,p=null,g=null,y=null,b=null;try{if(s)d=e._buf,l=e._loBuf,m=t._buf,w=t._loBuf;else{let R=U(e),W=U(t);d=h(r,R.hi,"dswap-xHi",!0),l=h(r,R.lo,"dswap-xLo",!0),m=h(r,W.hi,"dswap-yHi",!0),w=h(r,W.lo,"dswap-yLo",!0)}c=q(r,[{value:o,type:"u32"},{value:a,type:"u32"},{value:i,type:"u32"}],"dswap-params");let x=D(r,n.getBindGroupLayout(0),[d,l,m,w,c]),{commandEncoder:v,ts:B}=O(r,n,x,br(r,o));p=s?null:E(r,v,d),g=s?null:E(r,v,l),y=u?null:E(r,v,m),b=u?null:E(r,v,w),F(r,v);let A=await j(B);if(s)return A!==void 0?{gpuTimeMs:A}:{};let S=await G(p,Float32Array);p=null;let _=await G(g,Float32Array);g=null;let N=await G(y,Float32Array);y=null;let I=await G(b,Float32Array);b=null;let k=dr(S,_),L=dr(N,I);return A!==void 0?{x:k,y:L,gpuTimeMs:A}:{x:k,y:L}}finally{!s&&d&&f(d),!s&&l&&f(l),!u&&m&&f(m),!u&&w&&f(w),c&&f(c),p&&f(p),g&&f(g),y&&f(y),b&&f(b)}}async function na(r,o,e,a,t,i,s){let u=a instanceof M,n=i instanceof M;if(H(r),T(r,"saxpy",{x:a,y:i}),!Number.isInteger(o)||!Number.isInteger(t)||!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(t<=0||s<=0)throw new Error("incx and incy must be positive.");if(!u&&!(a 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(o<=0)return n?{}:{y:i};if(a.length<(o-1)*t+1)throw new Error("x does not have enough elements for the given n and incx.");if(i.length<(o-1)*s+1)throw new Error("y does not have enough elements for the given n and incy.");let d=await P(r,"saxpy"),l=null,m=null,w=null,c=null;try{l=u?a._buf:h(r,a,"saxpy-x",!1),m=n?i._buf:h(r,i,"saxpy-y",!0),w=q(r,[{value:o,type:"u32"},{value:e,type:"f32"},{value:t,type:"u32"},{value:s,type:"u32"}],"saxpy-params");let p=D(r,d.getBindGroupLayout(0),[l,m,w]),{commandEncoder:g,ts:y}=O(r,d,p,br(r,o));c=n?null:E(r,g,m),F(r,g);let b=await j(y);if(n)return b!==void 0?{gpuTimeMs:b}:{};let x=await G(c,Float32Array);return c=null,b!==void 0?{y:x,gpuTimeMs:b}:{y:x}}finally{!u&&l&&f(l),!n&&m&&f(m),w&&f(w),c&&f(c)}}async function la(r,o,e,a,t,i,s){let u=a instanceof M,n=i instanceof M;if(H(r),!Number.isInteger(o)||!Number.isInteger(t)||!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 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&&a.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(t<=0||s<=0)throw new Error("incx and incy must be positive.");if(T(r,"daxpy",{x:a,y:i}),o<=0)return n?{}:{y:i};if(a.length<(o-1)*t+1)throw new Error("x does not have enough elements for the given n and incx.");if(i.length<(o-1)*s+1)throw new Error("y does not have enough elements for the given n and incy.");let l=await P(r,[...["f64/dekker","f64/utils/add","f64/utils/multiply"],"daxpy"]),{hi:m,lo:w}=U(new Float64Array([e])),c=null,p=null,g=null,y=null,b=null,x=null,v=null;try{if(u)c=a._buf,p=a._loBuf,g=i._buf,y=i._loBuf;else{let L=U(a),R=U(i);c=h(r,L.hi,"daxpy-xHi",!1),p=h(r,L.lo,"daxpy-xLo",!1),g=h(r,R.hi,"daxpy-yHi",!0),y=h(r,R.lo,"daxpy-yLo",!0)}b=q(r,[{value:o,type:"u32"},{value:m[0],type:"f32"},{value:w[0],type:"f32"},{value:t,type:"u32"},{value:s,type:"u32"}],"daxpy-params");let B=D(r,l.getBindGroupLayout(0),[c,p,g,y,b]),{commandEncoder:A,ts:S}=O(r,l,B,br(r,o));x=n?null:E(r,A,g),v=n?null:E(r,A,y),F(r,A);let _=await j(S);if(n)return _!==void 0?{gpuTimeMs:_}:{};let N=await G(x,Float32Array);x=null;let I=await G(v,Float32Array);v=null;let k=dr(N,I);return _!==void 0?{y:k,gpuTimeMs:_}:{y:k}}finally{!u&&c&&f(c),!u&&p&&f(p),!n&&g&&f(g),!n&&y&&f(y),b&&f(b),x&&f(x),v&&f(v)}}async function ua(r,o,e,a,t,i){let s=e instanceof M,u=t instanceof M;if(H(r),T(r,"scopy",{x:e,y:t}),!Number.isInteger(o)||!Number.isInteger(a)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(a<=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&&!(t 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(o<=0)return u?{}:{y:t};if(e.length<(o-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(o-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=await P(r,"scopy"),d=null,l=null,m=null,w=null;try{d=s?e._buf:h(r,e,"scopy-x",!1),l=u?t._buf:h(r,t,"scopy-y",!0),m=q(r,[{value:o,type:"u32"},{value:a,type:"u32"},{value:i,type:"u32"}],"scopy-params");let c=D(r,n.getBindGroupLayout(0),[d,l,m]),{commandEncoder:p,ts:g}=O(r,n,c,br(r,o));w=u?null:E(r,p,l),F(r,p);let y=await j(g);if(u)return y!==void 0?{gpuTimeMs:y}:{};let b=await G(w,Float32Array);return w=null,y!==void 0?{y:b,gpuTimeMs:y}:{y:b}}finally{!s&&d&&f(d),!u&&l&&f(l),m&&f(m),w&&f(w)}}async function da(r,o,e,a,t,i){let s=e instanceof M,u=t instanceof M;if(H(r),T(r,"dcopy",{x:e,y:t}),!Number.isInteger(o)||!Number.isInteger(a)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(a<=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(!(t 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&&t.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(o<=0)return u?{}:{y:t};if(e.length<(o-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(o-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=await P(r,"dcopy"),d=null,l=null,m=null,w=null,c=null,p=null,g=null;try{if(s)d=e._buf,l=e._loBuf,m=t._buf,w=t._loBuf;else{let _=U(e),N=U(t);d=h(r,_.hi,"dcopy-xHi",!1),l=h(r,_.lo,"dcopy-xLo",!1),m=h(r,N.hi,"dcopy-yHi",!0),w=h(r,N.lo,"dcopy-yLo",!0)}c=q(r,[{value:o,type:"u32"},{value:a,type:"u32"},{value:i,type:"u32"}],"dcopy-params");let y=D(r,n.getBindGroupLayout(0),[d,l,m,w,c]),{commandEncoder:b,ts:x}=O(r,n,y,br(r,o));p=u?null:E(r,b,m),g=u?null:E(r,b,w),F(r,b);let v=await j(x);if(u)return v!==void 0?{gpuTimeMs:v}:{};let B=await G(p,Float32Array);p=null;let A=await G(g,Float32Array);g=null;let S=dr(B,A);return v!==void 0?{y:S,gpuTimeMs:v}:{y:S}}finally{!s&&d&&f(d),!s&&l&&f(l),!u&&m&&f(m),!u&&w&&f(w),c&&f(c),p&&f(p),g&&f(g)}}async function fa(r,o,e,a,t,i){let s=e instanceof M,u=t instanceof M;if(H(r),T(r,"sdot",{x:e,y:t}),!Number.isInteger(o)||!Number.isInteger(a)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(a<=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&&!(t 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(o<=0)return{dot:0};if(e.length<(o-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(o-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=await P(r,"sdot"),d=await P(r,"reduction/sum"),l=null,m=null,w=null,c=null,p=null,g=null;try{l=s?e._buf:h(r,e,"sdot-x",!1),m=u?t._buf:h(r,t,"sdot-y",!1),w=fr(r,512,"sdot-partials"),c=Ar(r,4,"sdot-result"),p=q(r,[{value:o,type:"u32"},{value:a,type:"u32"},{value:i,type:"u32"}],"sdot-params");let y=D(r,n.getBindGroupLayout(0),[l,m,w,p]),{commandEncoder:b,ts:x}=O(r,n,y,128);F(r,b);let v=D(r,d.getBindGroupLayout(0),[w,c]),{commandEncoder:B,ts:A}=O(r,d,v,1);g=E(r,B,c),F(r,B);let S=G(g,Float32Array);g=null;let[_,N,I]=await Promise.all([j(x),j(A),S]);return _!==void 0&&N!==void 0?{dot:I[0],gpuTimeMs:_+N}:{dot:I[0]}}finally{!s&&l&&f(l),!u&&m&&f(m),w&&f(w),c&&f(c),p&&f(p),g&&f(g)}}async function ma(r,o,e,a){let t=e instanceof M;if(H(r),T(r,"sasum",{x:e}),!Number.isInteger(o)||!Number.isInteger(a))throw new Error("n and incx must be integers.");if(a<=0)throw new Error("incx must be positive.");if(!t&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(o<=0)return{asum:0};if(e.length<(o-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");let i=await P(r,"sasum"),s=await P(r,"reduction/sum"),u=null,n=null,d=null,l=null,m=null;try{u=t?e._buf:h(r,e,"sasum-x",!1),n=fr(r,512,"sasum-partials"),d=Ar(r,4,"sasum-result"),l=q(r,[{value:o,type:"u32"},{value:a,type:"u32"}],"sasum-params");let w=D(r,i.getBindGroupLayout(0),[u,n,l]),{commandEncoder:c,ts:p}=O(r,i,w,128);F(r,c);let g=D(r,s.getBindGroupLayout(0),[n,d]),{commandEncoder:y,ts:b}=O(r,s,g,1);m=E(r,y,d),F(r,y);let x=G(m,Float32Array);m=null;let[v,B,A]=await Promise.all([j(p),j(b),x]);return v!==void 0&&B!==void 0?{asum:A[0],gpuTimeMs:v+B}:{asum:A[0]}}finally{!t&&u&&f(u),n&&f(n),d&&f(d),l&&f(l),m&&f(m)}}async function ca(r,o,e,a){let t=e instanceof M;if(H(r),T(r,"dasum",{x:e}),!Number.isInteger(o)||!Number.isInteger(a))throw new Error("n and incx must be integers.");if(a<=0)throw new Error("incx must be positive.");if(!t&&!(e instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(t&&e.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(o<=0)return{asum:0};if(e.length<(o-1)*a+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 P(r,[...i,"dasum"]),u=await P(r,[...i,"reduction/sumF64"]),n=null,d=null,l=null,m=null,w=null,c=null,p=null,g=null,y=null;try{if(t)n=e._buf,d=e._loBuf;else{let{hi:V,lo:C}=U(e.map(Math.abs));n=h(r,V,"dasum-xHi",!1),d=h(r,C,"dasum-xLo",!1)}l=fr(r,512,"dasum-partialsHi"),m=fr(r,512,"dasum-partialsLo"),w=Ar(r,4,"dasum-result-hi"),c=Ar(r,4,"dasum-result-lo"),p=q(r,[{value:o,type:"u32"},{value:a,type:"u32"}],"dasum-params");let b=D(r,s.getBindGroupLayout(0),[n,d,l,m,p]),{commandEncoder:x,ts:v}=O(r,s,b,128);F(r,x);let B=D(r,u.getBindGroupLayout(0),[l,m,w,c]),{commandEncoder:A,ts:S}=O(r,u,B,1);g=E(r,A,w),y=E(r,A,c),F(r,A);let _=G(g,Float32Array),N=G(y,Float32Array);g=null,y=null;let[I,k,L,R]=await Promise.all([j(v),j(S),_,N]),W=dr(L,R)[0];return I!==void 0&&k!==void 0?{asum:W,gpuTimeMs:I+k}:{asum:W}}finally{!t&&n&&f(n),!t&&d&&f(d),l&&f(l),m&&f(m),w&&f(w),c&&f(c),p&&f(p),g&&f(g),y&&f(y)}}async function pa(r,o,e,a,t,i){let s=e instanceof M,u=t instanceof M;if(H(r),T(r,"ddot",{x:e,y:t}),!Number.isInteger(o)||!Number.isInteger(a)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(a<=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&&!(t 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&&t.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(o<=0)return{dot:0};if(e.length<(o-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(o-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"],d=await P(r,[...n,"f64/utils/multiply","ddot"]),l=await P(r,[...n,"reduction/sumF64"]),m=null,w=null,c=null,p=null,g=null,y=null,b=null,x=null,v=null,B=null,A=null;try{if(s)m=e._buf,w=e._loBuf,c=t._buf,p=t._loBuf;else{let $=U(e),J=U(t);m=h(r,$.hi,"ddot-xHi",!1),w=h(r,$.lo,"ddot-xLo",!1),c=h(r,J.hi,"ddot-yHi",!1),p=h(r,J.lo,"ddot-yLo",!1)}g=fr(r,512,"ddot-partialsHi"),y=fr(r,512,"ddot-partialsLo"),b=Ar(r,4,"ddot-result-hi"),x=Ar(r,4,"ddot-result-lo"),v=q(r,[{value:o,type:"u32"},{value:a,type:"u32"},{value:i,type:"u32"}],"ddot-params");let S=D(r,d.getBindGroupLayout(0),[m,w,c,p,g,y,v]),{commandEncoder:_,ts:N}=O(r,d,S,128);F(r,_);let I=D(r,l.getBindGroupLayout(0),[g,y,b,x]),{commandEncoder:k,ts:L}=O(r,l,I,1);B=E(r,k,b),A=E(r,k,x),F(r,k);let R=G(B,Float32Array),W=G(A,Float32Array);B=null,A=null;let[V,C,z,K]=await Promise.all([j(N),j(L),R,W]),Y=dr(z,K)[0];return V!==void 0&&C!==void 0?{dot:Y,gpuTimeMs:V+C}:{dot:Y}}finally{!s&&m&&f(m),!s&&w&&f(w),!u&&c&&f(c),!u&&p&&f(p),g&&f(g),y&&f(y),b&&f(b),x&&f(x),v&&f(v),B&&f(B),A&&f(A)}}async function wa(r,o,e,a){let t=e instanceof M;if(H(r),T(r,"snrm2",{x:e}),!Number.isInteger(o)||!Number.isInteger(a))throw new Error("n and incx must be integers.");if(a<=0)throw new Error("incx must be positive.");if(!t&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(o<=0)return{nrm2:0};if(e.length<(o-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");let i=await P(r,"snrm2"),s=await P(r,"reduction/scaledSum"),u=null,n=null,d=null,l=null,m=null,w=null;try{u=t?e._buf:h(r,e,"snrm2-x",!1),n=fr(r,512,"snrm2-partials-scale"),d=fr(r,512,"snrm2-partials-ssq"),l=Ar(r,4,"snrm2-result"),m=q(r,[{value:o,type:"u32"},{value:a,type:"u32"}],"snrm2-params");let c=D(r,i.getBindGroupLayout(0),[u,n,d,m]),{commandEncoder:p,ts:g}=O(r,i,c,128);F(r,p);let y=D(r,s.getBindGroupLayout(0),[n,d,l]),{commandEncoder:b,ts:x}=O(r,s,y,1);w=E(r,b,l),F(r,b);let v=G(w,Float32Array);w=null;let[B,A,S]=await Promise.all([j(g),j(x),v]),_=S[0];return B!==void 0&&A!==void 0?{nrm2:_,gpuTimeMs:B+A}:{nrm2:_}}finally{!t&&u&&f(u),n&&f(n),d&&f(d),l&&f(l),m&&f(m),w&&f(w)}}async function ga(r,o,e,a){let t=e instanceof M;if(H(r),T(r,"dnrm2",{x:e}),!Number.isInteger(o)||!Number.isInteger(a))throw new Error("n and incx must be integers.");if(a<=0)throw new Error("incx must be positive.");if(!t&&!(e instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(t&&e.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(o<=0)return{nrm2:0};if(e.length<(o-1)*a+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 P(r,[...i,"dnrm2"]),u=await P(r,[...i,"reduction/scaledSumF64"]),n=null,d=null,l=null,m=null,w=null,c=null,p=null,g=null,y=null,b=null,x=null;try{if(t)n=e._buf,d=e._loBuf;else{let{hi:z,lo:K}=U(e);n=h(r,z,"dnrm2-xHi",!1),d=h(r,K,"dnrm2-xLo",!1)}l=fr(r,512,"dnrm2-partials-scaleHi"),m=fr(r,512,"dnrm2-partials-scaleLo"),w=fr(r,512,"dnrm2-partials-ssqHi"),c=fr(r,512,"dnrm2-partials-ssqLo"),p=Ar(r,4,"dnrm2-result-hi"),g=Ar(r,4,"dnrm2-result-lo"),y=q(r,[{value:o,type:"u32"},{value:a,type:"u32"}],"dnrm2-params");let v=D(r,s.getBindGroupLayout(0),[n,d,l,m,w,c,y]),{commandEncoder:B,ts:A}=O(r,s,v,128);F(r,B);let S=D(r,u.getBindGroupLayout(0),[l,m,w,c,p,g]),{commandEncoder:_,ts:N}=O(r,u,S,1);b=E(r,_,p),x=E(r,_,g),F(r,_);let I=G(b,Float32Array),k=G(x,Float32Array);b=null,x=null;let[L,R,W,V]=await Promise.all([j(A),j(N),I,k]),C=dr(W,V)[0];return L!==void 0&&R!==void 0?{nrm2:C,gpuTimeMs:L+R}:{nrm2:C}}finally{!t&&n&&f(n),!t&&d&&f(d),l&&f(l),m&&f(m),w&&f(w),c&&f(c),p&&f(p),g&&f(g),y&&f(y),b&&f(b),x&&f(x)}}async function ha(r,o,e,a){let t=e instanceof M;if(H(r),T(r,"isamax",{x:e}),!Number.isInteger(o)||!Number.isInteger(a))throw new Error("n and incx must be integers.");if(a<=0)throw new Error("incx must be positive.");if(!t&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(o<=0)return{index:0};if(e.length<(o-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");let i=await P(r,"isamax"),s=await P(r,"reduction/argmax"),u=null,n=null,d=null,l=null,m=null,w=null;try{u=t?e._buf:h(r,e,"isamax-x",!1),n=fr(r,512,"isamax-partials-val"),d=fr(r,512,"isamax-partials-idx"),l=Ar(r,4,"isamax-result"),m=q(r,[{value:o,type:"u32"},{value:a,type:"u32"}],"isamax-params");let c=D(r,i.getBindGroupLayout(0),[u,n,d,m]),{commandEncoder:p,ts:g}=O(r,i,c,128);F(r,p);let y=D(r,s.getBindGroupLayout(0),[n,d,l]),{commandEncoder:b,ts:x}=O(r,s,y,1);w=E(r,b,l),F(r,b);let v=G(w,Uint32Array);w=null;let[B,A,S]=await Promise.all([j(g),j(x),v]),_=S[0];return B!==void 0&&A!==void 0?{index:_,gpuTimeMs:B+A}:{index:_}}finally{!t&&u&&f(u),n&&f(n),d&&f(d),l&&f(l),m&&f(m),w&&f(w)}}async function ba(r,o,e,a){let t=e instanceof M;if(H(r),T(r,"idamax",{x:e}),!Number.isInteger(o)||!Number.isInteger(a))throw new Error("n and incx must be integers.");if(a<=0)throw new Error("incx must be positive.");if(!t&&!(e instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(t&&e.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(o<=0)return{index:0};if(e.length<(o-1)*a+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 P(r,[...i,"idamax"],"idamax_main"),u=await P(r,[...i,"reduction/argmaxF64"],"reduce_f64"),n=null,d=null,l=null,m=null,w=null,c=null,p=null,g=null;try{if(t)n=e._buf,d=e._loBuf;else{let{hi:L,lo:R}=U(e);n=h(r,L,"idamax-xHi",!1),d=h(r,R,"idamax-xLo",!1)}l=fr(r,512,"idamax-partials-val-hi"),m=fr(r,512,"idamax-partials-val-lo"),w=fr(r,512,"idamax-partials-idx"),c=Ar(r,4,"idamax-result"),p=q(r,[{value:o,type:"u32"},{value:a,type:"u32"}],"idamax-params");let y=D(r,s.getBindGroupLayout(0),[n,d,l,m,w,p]),{commandEncoder:b,ts:x}=O(r,s,y,128);F(r,b);let v=D(r,u.getBindGroupLayout(0),[l,m,w,c]),{commandEncoder:B,ts:A}=O(r,u,v,1);g=E(r,B,c),F(r,B);let S=G(g,Uint32Array);g=null;let[_,N,I]=await Promise.all([j(x),j(A),S]),k=I[0];return _!==void 0&&N!==void 0?{index:k,gpuTimeMs:_+N}:{index:k}}finally{!t&&n&&f(n),!t&&d&&f(d),l&&f(l),m&&f(m),w&&f(w),c&&f(c),p&&f(p),g&&f(g)}}async function ya(r,o,e,a,t,i,s,u){let n=e instanceof M,d=t instanceof M;if(H(r),T(r,"srot",{x:e,y:t}),!Number.isInteger(o)||!Number.isInteger(a)||!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(a<=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(!d&&!(t instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(n!==d)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(o<=0)return n?{}:{x:e,y:t};if(e.length<(o-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(o-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let l=await P(r,"srot"),m=null,w=null,c=null,p=null,g=null;try{m=n?e._buf:h(r,e,"srot-x",!0),w=d?t._buf:h(r,t,"srot-y",!0),c=q(r,[{value:o,type:"u32"},{value:s,type:"f32"},{value:u,type:"f32"},{value:a,type:"u32"},{value:i,type:"u32"}],"srot-params");let y=D(r,l.getBindGroupLayout(0),[m,w,c]),{commandEncoder:b,ts:x}=O(r,l,y,br(r,o));p=n?null:E(r,b,m),g=d?null:E(r,b,w),F(r,b);let v=await j(x);if(n)return v!==void 0?{gpuTimeMs:v}:{};let B=G(p,Float32Array),A=G(g,Float32Array);p=null,g=null;let[S,_]=await Promise.all([B,A]);return v!==void 0?{x:S,y:_,gpuTimeMs:v}:{x:S,y:_}}finally{!n&&m&&f(m),!d&&w&&f(w),c&&f(c),p&&f(p),g&&f(g)}}async function xa(r,o,e,a,t,i,s,u){let n=e instanceof M,d=t instanceof M;if(H(r),T(r,"drot",{x:e,y:t}),!Number.isInteger(o)||!Number.isInteger(a)||!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(a<=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(!(t instanceof Float64Array)&&!d)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(d&&t.dtype!==Float64Array)throw new Error("y must be a Float64Array-backed GpuVector.");if(n!==d)throw new Error("x and y must be the same type (both Float64Array or both GpuVector).");if(o<=0)return n?{}:{x:e,y:t};if(e.length<(o-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(o-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let m=await P(r,[...["f64/dekker","f64/utils/add","f64/utils/multiply"],"drot"]),{hi:w,lo:c}=U(new Float64Array([s])),{hi:p,lo:g}=U(new Float64Array([u])),y=null,b=null,x=null,v=null,B=null,A=null,S=null,_=null,N=null;try{if(n)y=e._buf,b=e._loBuf,x=t._buf,v=t._loBuf;else{let $=U(e),J=U(t);y=h(r,$.hi,"drot-xHi",!0),b=h(r,$.lo,"drot-xLo",!0),x=h(r,J.hi,"drot-yHi",!0),v=h(r,J.lo,"drot-yLo",!0)}B=q(r,[{value:o,type:"u32"},{value:w[0],type:"f32"},{value:c[0],type:"f32"},{value:p[0],type:"f32"},{value:g[0],type:"f32"},{value:a,type:"u32"},{value:i,type:"u32"}],"drot-params");let I=D(r,m.getBindGroupLayout(0),[y,b,x,v,B]),{commandEncoder:k,ts:L}=O(r,m,I,br(r,o));A=n?null:E(r,k,y),S=n?null:E(r,k,b),_=d?null:E(r,k,x),N=d?null:E(r,k,v),F(r,k);let R=await j(L);if(n)return R!==void 0?{gpuTimeMs:R}:{};let W=await G(A,Float32Array);A=null;let V=await G(S,Float32Array);S=null;let C=await G(_,Float32Array);_=null;let z=await G(N,Float32Array);N=null;let K=dr(W,V),Y=dr(C,z);return R!==void 0?{x:K,y:Y,gpuTimeMs:R}:{x:K,y:Y}}finally{!n&&y&&f(y),!n&&b&&f(b),!d&&x&&f(x),!d&&v&&f(v),B&&f(B),A&&f(A),S&&f(S),_&&f(_),N&&f(N)}}async function va(r,o,e,a,t,i,s){let u=e instanceof M,n=t instanceof M;if(H(r),T(r,"srotm",{x:e,y:t}),!Number.isInteger(o)||!Number.isInteger(a)||!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(a<=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&&!(t 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(o<=0||s[0]===-2)return u?{}:{x:e,y:t};if(e.length<(o-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(o-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let d=await P(r,"srotm"),l=null,m=null,w=null,c=null,p=null,g=null;try{l=u?e._buf:h(r,e,"srotm-x",!0),m=n?t._buf:h(r,t,"srotm-y",!0),w=h(r,s,"srotm-param",!1),c=q(r,[{value:o,type:"u32"},{value:a,type:"u32"},{value:i,type:"u32"}],"srotm-params");let y=D(r,d.getBindGroupLayout(0),[l,m,w,c]),{commandEncoder:b,ts:x}=O(r,d,y,br(r,o));p=u?null:E(r,b,l),g=n?null:E(r,b,m),F(r,b);let v=await j(x);if(u)return v!==void 0?{gpuTimeMs:v}:{};let B=G(p,Float32Array),A=G(g,Float32Array);p=null,g=null;let[S,_]=await Promise.all([B,A]);return v!==void 0?{x:S,y:_,gpuTimeMs:v}:{x:S,y:_}}finally{!u&&l&&f(l),!n&&m&&f(m),w&&f(w),c&&f(c),p&&f(p),g&&f(g)}}async function _a(r,o,e,a,t,i,s){let u=e instanceof M,n=t instanceof M;if(H(r),T(r,"drotm",{x:e,y:t}),!Number.isInteger(o)||!Number.isInteger(a)||!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(a<=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(!(t 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&&t.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(o<=0||s[0]===-2)return u?{}:{x:e,y:t};if(e.length<(o-1)*a+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(o-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let l=await P(r,[...["f64/dekker","f64/utils/add","f64/utils/multiply"],"drotm"]),{hi:m,lo:w}=U(s),c=null,p=null,g=null,y=null,b=null,x=null,v=null,B=null,A=null,S=null,_=null;try{if(u)c=e._buf,p=e._loBuf,g=t._buf,y=t._loBuf;else{let Y=U(e),$=U(t);c=h(r,Y.hi,"drotm-xHi",!0),p=h(r,Y.lo,"drotm-xLo",!0),g=h(r,$.hi,"drotm-yHi",!0),y=h(r,$.lo,"drotm-yLo",!0)}b=h(r,m,"drotm-paramHi",!1),x=h(r,w,"drotm-paramLo",!1),v=q(r,[{value:o,type:"u32"},{value:a,type:"u32"},{value:i,type:"u32"}],"drotm-params");let N=D(r,l.getBindGroupLayout(0),[c,p,g,y,b,x,v]),{commandEncoder:I,ts:k}=O(r,l,N,br(r,o));B=u?null:E(r,I,c),A=u?null:E(r,I,p),S=n?null:E(r,I,g),_=n?null:E(r,I,y),F(r,I);let L=await j(k);if(u)return L!==void 0?{gpuTimeMs:L}:{};let R=await G(B,Float32Array);B=null;let W=await G(A,Float32Array);A=null;let V=await G(S,Float32Array);S=null;let C=await G(_,Float32Array);_=null;let z=dr(R,W),K=dr(V,C);return L!==void 0?{x:z,y:K,gpuTimeMs:L}:{x:z,y:K}}finally{!u&&c&&f(c),!u&&p&&f(p),!n&&g&&f(g),!n&&y&&f(y),b&&f(b),x&&f(x),v&&f(v),B&&f(B),A&&f(A),S&&f(S),_&&f(_)}}async function Ba(r,o,e,a,t,i,s,u,n,d,l,m,w="row-major"){let c=i instanceof X,p=u instanceof M,g=l instanceof M;if(H(r),T(r,"sgemv",{A:i,x:u,y:l}),o!=="no-transpose"&&o!=="transpose")throw new Error("trans 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 t!="number")throw new Error("alpha must be a number.");if(Number.isNaN(t))throw new Error("alpha must not be NaN.");if(!Number.isFinite(t))throw new Error("alpha must be finite.");if(typeof d!="number")throw new Error("beta must be a number.");if(Number.isNaN(d))throw new Error("beta must not be NaN.");if(!Number.isFinite(d))throw new Error("beta must be finite.");if(!Number.isInteger(e)||!Number.isInteger(a)||!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(!p&&!(u instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!g&&!(l instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(p!==g)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(p&&!c)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(c&&!p)throw new Error("x and y must be GpuVectors when A is a GpuMatrix.");if(p&&u._buf===l._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(c&&g&&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<a))throw new Error("A is too small for the given m and n.");if(e<0||a<0)throw new Error("m and n must be non-negative.");if(e===0||a===0)return g?{}:{y:l};(c?i.layout:w)==="column-major"&&([e,a]=[a,e],o=o==="no-transpose"?"transpose":"no-transpose");let b=o==="no-transpose",x=b?a:e,v=b?e:a;if(s<a)throw new Error("lda must be >= n.");if(!c&&i.length<(e-1)*s+a)throw new Error("A does not have enough elements for the given m, n, and lda.");if(u.length<(x-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 P(r,b?"sgemv_n":"sgemv_t"),S=null,_=null,N=null,I=null;try{S=c?i._buf:h(r,i,"sgemv-A",!1),_=p?u._buf:h(r,u,"sgemv-x",!1),N=g?l._buf:h(r,l,"sgemv-y",!0),I=q(r,[{value:e,type:"u32"},{value:a,type:"u32"},{value:t,type:"f32"},{value:d,type:"f32"},{value:n,type:"u32"},{value:m,type:"u32"},{value:s,type:"u32"}],"sgemv-params");let k=D(r,A.getBindGroupLayout(0),[S,_,N,I]),L=b?Math.min(e,r.limits.maxComputeWorkgroupsPerDimension):Yr(r,"sgemv",v),{commandEncoder:R,ts:W}=O(r,A,k,L),V=g?null:E(r,R,N);F(r,R);let C=await j(W);if(g)return C!==void 0?{gpuTimeMs:C}:{};let z=await G(V,Float32Array);return C!==void 0?{y:z,gpuTimeMs:C}:{y:z}}finally{!c&&S&&f(S),!p&&_&&f(_),!g&&N&&f(N),I&&f(I)}}async function Aa(r,o,e,a,t,i,s,u,n,d,l,m="row-major"){let w=s instanceof M,c=d instanceof M,p=t instanceof X;if(H(r),T(r,"ssymv",{A:t,x:s,y:d}),o!=="lower"&&o!=="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 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 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(!p&&!(t instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!w&&!(s instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!c&&!(d instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(w!==c)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(w&&!p)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(p&&!w)throw new Error("x and y must be GpuVectors when A is a GpuMatrix.");if(w&&s._buf===d._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(p&&i!==t.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(p&&(t.rows<e||t.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:d};if(!p&&t.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(d.length<(e-1)*l+1)throw new Error("y does not have enough elements for the given n and incy.");let y=(p?t.layout:m)==="column-major"?o==="upper":o==="lower",b=await P(r,"ssymv"),x=null,v=null,B=null,A=null;try{x=p?t._buf:h(r,t,"ssymv-A",!1),v=w?s._buf:h(r,s,"ssymv-x",!1),B=c?d._buf:h(r,d,"ssymv-y",!0),A=q(r,[{value:e,type:"u32"},{value:a,type:"f32"},{value:n,type:"f32"},{value:u,type:"u32"},{value:l,type:"u32"},{value:i,type:"u32"},{value:y?0:1,type:"u32"}],"ssymv-params");let S=D(r,b.getBindGroupLayout(0),[x,v,B,A]),_=Math.min(e,r.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:N,ts:I}=O(r,b,S,_),k=c?null:E(r,N,B);F(r,N);let L=await j(I);if(c)return L!==void 0?{gpuTimeMs:L}:{};let R=await G(k,Float32Array);return L!==void 0?{y:R,gpuTimeMs:L}:{y:R}}finally{!p&&x&&f(x),!w&&v&&f(v),!c&&B&&f(B),A&&f(A)}}async function Sa(r,o,e,a,t,i,s,u,n,d,l,m="row-major"){let w=u instanceof M,c=d instanceof M,p=i instanceof X,g=a==="unit";if(H(r),T(r,"strmv",{A:i,x:u,y:d}),o!=="lower"&&o!=="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(!g&&a!=="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(t)||!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<t)throw new Error("lda must be >= n.");if(!p&&!(i instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!w&&!(u instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!c&&!(d instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(w!==c)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(w&&u._buf===d._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(w&&!p)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(p&&!w)throw new Error("x and y must be GpuVectors when A is a GpuMatrix.");if(p&&c&&i._buf===d._buf)throw new Error("A and y must not reference the same GPU buffer.");if(p&&s!==i.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(p&&(i.rows<t||i.cols<t))throw new Error("A is too small for the given n.");if(t<0)throw new Error("n must be non-negative.");if(t===0)return c?{}:{y:d};if(!p&&i.length<(t-1)*s+t)throw new Error("A does not have enough elements for the given n and lda.");if(u.length<(t-1)*n+1)throw new Error("x does not have enough elements for the given n and incx.");if(d.length<(t-1)*l+1)throw new Error("y does not have enough elements for the given n and incy.");let b=(p?i.layout:m)==="column-major",x=b?o==="upper":o==="lower",v=b?e==="transpose":e==="no-transpose",B=await P(r,"strmv"),A=null,S=null,_=null,N=null;try{A=p?i._buf:h(r,i,"strmv-A",!1),S=w?u._buf:h(r,u,"strmv-x",!1),_=c?d._buf:h(r,d,"strmv-y",!0),N=q(r,[{value:t,type:"u32"},{value:n,type:"u32"},{value:l,type:"u32"},{value:s,type:"u32"},{value:v?0:1,type:"u32"},{value:x?0:1,type:"u32"},{value:g?1:0,type:"u32"}],"strmv-params");let I=D(r,B.getBindGroupLayout(0),[A,S,_,N]),k=Math.min(t,r.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:L,ts:R}=O(r,B,I,k),W=c?null:E(r,L,_);F(r,L);let V=await j(R);if(c)return V!==void 0?{gpuTimeMs:V}:{};let C=await G(W,Float32Array);return V!==void 0?{y:C,gpuTimeMs:V}:{y:C}}finally{!p&&A&&f(A),!w&&S&&f(S),!c&&_&&f(_),N&&f(N)}}function Ga(r,o,e){let a=new ArrayBuffer(r*o),t=new DataView(a);for(let i=0;i<r;i++){let s=e(i),u=i*o;s.forEach((n,d)=>t.setUint32(u+d*4,n,!0))}return a}function Ea(r,o,e){let a=r.createBuffer({label:e,size:o.byteLength,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});return r.queue.writeBuffer(a,0,o),a}async function ka(r,o,e,a,t,i,s,u,n,d="row-major"){let l=u instanceof M,m=i instanceof X,w=a==="unit";if(H(r),T(r,"strsv",{A:i,x:u}),o!=="lower"&&o!=="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&&a!=="non-unit")throw new Error("diag must be 'unit' or 'non-unit'.");if(d!=="row-major"&&d!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(!Number.isInteger(t)||!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<t)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<t||i.cols<t))throw new Error("A is too small for the given n.");if(t<0)throw new Error("n must be non-negative.");if(t===0)return l?{}:{x:u};if(!m&&i.length<(t-1)*s+t)throw new Error("A does not have enough elements for the given n and lda.");if(u.length<(t-1)*n+1)throw new Error("x does not have enough elements for the given n and incx.");let p=(m?i.layout:d)==="column-major",g=p?o==="upper":o==="lower",y=p?e==="transpose":e==="no-transpose",b=await P(r,"strsv_invert_block"),x=await P(r,"strsv_apply_inverse"),v=await P(r,"strsv_update"),B=y===g,A=[];for(let C=0;C<t;C+=64)A.push(C);B||A.reverse();let S=A.length,_=r.limits.maxComputeWorkgroupsPerDimension,N=r.limits.minUniformBufferOffsetAlignment,I=null,k=null,L=null,R=null,W=null,V=null;try{I=m?i._buf:h(r,i,"strsv-A",!1),k=l?u._buf:h(r,u,"strsv-x",!0),L=fr(r,S*64*64*4,"strsv-Ainv");let C=Ga(S,N,ar=>{let sr=ar*64,lr=Math.min(sr+64,t);return[n,ar,sr,lr]});R=Ea(r,C,"strsv-apply-params");let z=Ga(S,N,ar=>{let sr=ar*64,lr=Math.min(sr+64,t);return[t,n,s,y?0:1,g?0:1,sr,lr]});W=Ea(r,z,"strsv-update-params");let{commandEncoder:K,querySet:Y}=Mr(r);V=q(r,[{value:t,type:"u32"},{value:s,type:"u32"},{value:y?0:1,type:"u32"},{value:g?0:1,type:"u32"},{value:w?1:0,type:"u32"}],"strsv-invert-params");let $=D(r,b.getBindGroupLayout(0),[I,L,V]);hr(K,b,$,{x:64,y:S},Y?{timestampWrites:{querySet:Y,beginningOfPassWriteIndex:0}}:void 0);for(let ar=0;ar<A.length;ar++){let sr=A[ar],lr=Math.min(sr+64,t),ur=sr/64,mr=ar===A.length-1,yr=ur*N,wr=D(r,x.getBindGroupLayout(0),[L,k,{buffer:R,offset:yr,size:16}]);hr(K,x,wr,1,mr&&Y?{timestampWrites:{querySet:Y,endOfPassWriteIndex:1}}:void 0);let gr=B?t-lr:sr;if(gr===0)continue;let Gr=D(r,v.getBindGroupLayout(0),[I,k,{buffer:W,offset:yr,size:32}]),Nr=Math.min(gr,_);hr(K,v,Gr,Nr)}let nr=kr(r,K,Y),er=l?null:E(r,K,k);F(r,K);let Q=await j(nr);if(l)return Q!==void 0?{gpuTimeMs:Q}:{};let rr=await G(er,Float32Array);return Q!==void 0?{x:rr,gpuTimeMs:Q}:{x:rr}}finally{!m&&I&&f(I),!l&&k&&f(k),L&&f(L),R&&f(R),W&&f(W),V&&f(V)}}async function Da(r,o,e,a,t,i,s,u,n,d,l="row-major"){let m=n instanceof X;if(H(r),T(r,"sger",{A:n,x:t,y:s}),l!=="row-major"&&l!=="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(!Number.isInteger(o)||!Number.isInteger(e)||!Number.isInteger(i)||!Number.isInteger(u)||!Number.isInteger(d))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&&d!==n.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(m&&(n.rows<o||n.cols<e))throw new Error("A is too small for the given m and n.");(m?n.layout:l)==="column-major"&&([o,e]=[e,o],[t,s]=[s,t],[i,u]=[u,i]);let c=t instanceof M,p=s instanceof M;if(d<e)throw new Error("lda must be >= n.");if(!c&&!(t 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(c!==p)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===t._buf)throw new Error("A and x must not reference the same GPU buffer.");if(m&&p&&n._buf===s._buf)throw new Error("A and y must not reference the same GPU buffer.");if(o<0||e<0)throw new Error("m and n must be non-negative.");if(o===0||e===0)return m?{}:{A:n};if(!m&&n.length<(o-1)*d+e)throw new Error("A does not have enough elements for the given m, n, and lda.");if(t.length<(o-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 g=await P(r,"sger"),y=null,b=null,x=null,v=null;try{y=c?t._buf:h(r,t,"sger-x",!1),b=p?s._buf:h(r,s,"sger-y",!1),x=m?n._buf:h(r,n,"sger-A",!0),v=q(r,[{value:o,type:"u32"},{value:e,type:"u32"},{value:a,type:"f32"},{value:i,type:"u32"},{value:u,type:"u32"},{value:d,type:"u32"}],"sger-params");let B=D(r,g.getBindGroupLayout(0),[y,b,x,v]),A=Math.min(o,r.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:S,ts:_}=O(r,g,B,A),N=m?null:E(r,S,x);F(r,S);let I=await j(_);if(m)return I!==void 0?{gpuTimeMs:I}:{};let k=await G(N,Float32Array);return I!==void 0?{A:k,gpuTimeMs:I}:{A:k}}finally{!c&&y&&f(y),!p&&b&&f(b),!m&&x&&f(x),v&&f(v)}}async function La(r,o,e,a,t,i,s,u,n,d,l="row-major"){let m=n instanceof X;if(H(r),T(r,"dger",{A:n,x:t,y:s}),l!=="row-major"&&l!=="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(!Number.isInteger(o)||!Number.isInteger(e)||!Number.isInteger(i)||!Number.isInteger(u)||!Number.isInteger(d))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 Float64Array))throw new Error("A must be a Float64Array or GpuMatrix.");if(m&&n.dtype!==Float64Array)throw new Error("A must be a Float64Array-backed GpuMatrix.");if(m&&d!==n.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(m&&(n.rows<o||n.cols<e))throw new Error("A is too small for the given m and n.");(m?n.layout:l)==="column-major"&&([o,e]=[e,o],[t,s]=[s,t],[i,u]=[u,i]);let c=t instanceof M,p=s instanceof M;if(d<e)throw new Error("lda must be >= n.");if(!c&&!(t instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(!p&&!(s instanceof Float64Array))throw new Error("y must be a Float64Array or GpuVector.");if(c&&t.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(p&&s.dtype!==Float64Array)throw new Error("y must be a Float64Array-backed GpuVector.");if(c!==p)throw new Error("x and y must be the same type (both Float64Array 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===t._buf)throw new Error("A and x must not reference the same GPU buffer.");if(m&&p&&n._buf===s._buf)throw new Error("A and y must not reference the same GPU buffer.");if(o<0||e<0)throw new Error("m and n must be non-negative.");if(o===0||e===0)return m?{}:{A:n};if(!m&&n.length<(o-1)*d+e)throw new Error("A does not have enough elements for the given m, n, and lda.");if(t.length<(o-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 y=await P(r,[...["f64/dekker","f64/utils/add","f64/utils/multiply"],"dger"],"dger_main"),{hi:b,lo:x}=U(new Float64Array([a])),v=null,B=null,A=null,S=null,_=null,N=null,I=null,k=null,L=null;try{if(c)v=t._buf,B=t._loBuf,A=s._buf,S=s._loBuf,_=n._buf,N=n._loBuf;else{let J=U(t),nr=U(s),er=U(n);v=h(r,J.hi,"dger-xHi",!1),B=h(r,J.lo,"dger-xLo",!1),A=h(r,nr.hi,"dger-yHi",!1),S=h(r,nr.lo,"dger-yLo",!1),_=h(r,er.hi,"dger-AHi",!0),N=h(r,er.lo,"dger-ALo",!0)}I=q(r,[{value:o,type:"u32"},{value:e,type:"u32"},{value:b[0],type:"f32"},{value:x[0],type:"f32"},{value:i,type:"u32"},{value:u,type:"u32"},{value:d,type:"u32"}],"dger-params");let R=D(r,y.getBindGroupLayout(0),[v,B,A,S,_,N,I]),W=Math.min(o,r.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:V,ts:C}=O(r,y,R,W);k=m?null:E(r,V,_),L=m?null:E(r,V,N),F(r,V);let z=await j(C);if(m)return z!==void 0?{gpuTimeMs:z}:{};let K=await G(k,Float32Array);k=null;let Y=await G(L,Float32Array);L=null;let $=dr(K,Y);return z!==void 0?{A:$,gpuTimeMs:z}:{A:$}}finally{!c&&v&&f(v),!c&&B&&f(B),!p&&A&&f(A),!p&&S&&f(S),!m&&_&&f(_),!m&&N&&f(N),I&&f(I),k&&f(k),L&&f(L)}}async function Pa(r,o,e,a,t,i,s,u,n="row-major"){let d=t instanceof M,l=s instanceof X;if(H(r),T(r,"ssyr",{A:s,x:t}),o!=="lower"&&o!=="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 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(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(!d&&!(t instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(d&&!l)throw new Error("A must be a GpuMatrix when x is a GpuVector.");if(l&&!d)throw new Error("x must be a GpuVector when A is a GpuMatrix.");if(l&&d&&s._buf===t._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(t.length<(e-1)*i+1)throw new Error("x does not have enough elements for the given n and incx.");let w=(l?s.layout:n)==="column-major"?o==="upper":o==="lower",c=await P(r,"ssyr"),p=null,g=null,y=null;try{p=d?t._buf:h(r,t,"ssyr-x",!1),g=l?s._buf:h(r,s,"ssyr-A",!0),y=q(r,[{value:e,type:"u32"},{value:a,type:"f32"},{value:i,type:"u32"},{value:u,type:"u32"},{value:w?0:1,type:"u32"}],"ssyr-params");let b=D(r,c.getBindGroupLayout(0),[p,g,y]),x=Math.min(e,r.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:v,ts:B}=O(r,c,b,x),A=l?null:E(r,v,g);F(r,v);let S=await j(B);if(l)return S!==void 0?{gpuTimeMs:S}:{};let _=await G(A,Float32Array);return S!==void 0?{A:_,gpuTimeMs:S}:{A:_}}finally{!d&&p&&f(p),!l&&g&&f(g),y&&f(y)}}async function Na(r,o,e,a,t,i,s,u,n="row-major"){let d=t instanceof M,l=s instanceof X;if(H(r),T(r,"dsyr",{A:s,x:t}),o!=="lower"&&o!=="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 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(i<=0)throw new Error("incx must be positive.");if(u<e)throw new Error("lda must be >= n.");if(!l&&!(s instanceof Float64Array))throw new Error("A must be a Float64Array or GpuMatrix.");if(l&&s.dtype!==Float64Array)throw new Error("A must be a Float64Array-backed GpuMatrix.");if(!d&&!(t instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(d&&t.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(d&&!l)throw new Error("A must be a GpuMatrix when x is a GpuVector.");if(l&&!d)throw new Error("x must be a GpuVector when A is a GpuMatrix.");if(l&&d&&s._buf===t._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(t.length<(e-1)*i+1)throw new Error("x does not have enough elements for the given n and incx.");let w=(l?s.layout:n)==="column-major"?o==="upper":o==="lower",p=await P(r,[...["f64/dekker","f64/utils/add","f64/utils/multiply"],"dsyr"],"dsyr_main"),{hi:g,lo:y}=U(new Float64Array([a])),b=null,x=null,v=null,B=null,A=null,S=null,_=null;try{if(d)b=t._buf,x=t._loBuf,v=s._buf,B=s._loBuf;else{let z=U(t),K=U(s);b=h(r,z.hi,"dsyr-xHi",!1),x=h(r,z.lo,"dsyr-xLo",!1),v=h(r,K.hi,"dsyr-AHi",!0),B=h(r,K.lo,"dsyr-ALo",!0)}A=q(r,[{value:e,type:"u32"},{value:g[0],type:"f32"},{value:y[0],type:"f32"},{value:i,type:"u32"},{value:u,type:"u32"},{value:w?0:1,type:"u32"}],"dsyr-params");let N=D(r,p.getBindGroupLayout(0),[b,x,v,B,A]),I=Math.min(e,r.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:k,ts:L}=O(r,p,N,I);S=l?null:E(r,k,v),_=l?null:E(r,k,B),F(r,k);let R=await j(L);if(l)return R!==void 0?{gpuTimeMs:R}:{};let W=await G(S,Float32Array);S=null;let V=await G(_,Float32Array);_=null;let C=dr(W,V);return R!==void 0?{A:C,gpuTimeMs:R}:{A:C}}finally{!d&&b&&f(b),!d&&x&&f(x),!l&&v&&f(v),!l&&B&&f(B),A&&f(A),S&&f(S),_&&f(_)}}async function Ma(r,o,e,a,t,i,s,u,n,d,l="row-major"){let m=t instanceof M,w=s instanceof M,c=n instanceof X;if(H(r),T(r,"ssyr2",{A:n,x:t,y:s}),o!=="lower"&&o!=="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(d))throw new Error("n, incx, incy, and lda must be integers.");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(i<=0||u<=0)throw new Error("incx and incy must be positive.");if(d<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&&!(t instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!w&&!(s instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(m!==w)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===t._buf)throw new Error("A and x must not reference the same GPU buffer.");if(c&&w&&n._buf===s._buf)throw new Error("A and y must not reference the same GPU buffer.");if(m&&t._buf===s._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(c&&d!==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)*d+e)throw new Error("A does not have enough elements for the given n and lda.");if(t.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 g=(c?n.layout:l)==="column-major"?o==="upper":o==="lower",y=await P(r,"ssyr2"),b=null,x=null,v=null,B=null;try{b=m?t._buf:h(r,t,"ssyr2-x",!1),x=w?s._buf:h(r,s,"ssyr2-y",!1),v=c?n._buf:h(r,n,"ssyr2-A",!0),B=q(r,[{value:e,type:"u32"},{value:a,type:"f32"},{value:i,type:"u32"},{value:u,type:"u32"},{value:d,type:"u32"},{value:g?0:1,type:"u32"}],"ssyr2-params");let A=D(r,y.getBindGroupLayout(0),[b,x,v,B]),S=Math.min(e,r.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:_,ts:N}=O(r,y,A,S),I=c?null:E(r,_,v);F(r,_);let k=await j(N);if(c)return k!==void 0?{gpuTimeMs:k}:{};let L=await G(I,Float32Array);return k!==void 0?{A:L,gpuTimeMs:k}:{A:L}}finally{!m&&b&&f(b),!w&&x&&f(x),!c&&v&&f(v),B&&f(B)}}async function Ia(r,o,e,a,t,i,s,u,n,d,l="row-major"){let m=t instanceof M,w=s instanceof M,c=n instanceof X;if(H(r),T(r,"dsyr2",{A:n,x:t,y:s}),o!=="lower"&&o!=="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(d))throw new Error("n, incx, incy, and lda must be integers.");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(i<=0||u<=0)throw new Error("incx and incy must be positive.");if(d<e)throw new Error("lda must be >= n.");if(!c&&!(n instanceof Float64Array))throw new Error("A must be a Float64Array or GpuMatrix.");if(c&&n.dtype!==Float64Array)throw new Error("A must be a Float64Array-backed GpuMatrix.");if(!m&&!(t instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(!w&&!(s instanceof Float64Array))throw new Error("y must be a Float64Array or GpuVector.");if(m&&t.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(w&&s.dtype!==Float64Array)throw new Error("y must be a Float64Array-backed GpuVector.");if(m!==w)throw new Error("x and y must be the same type (both Float64Array 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===t._buf)throw new Error("A and x must not reference the same GPU buffer.");if(c&&w&&n._buf===s._buf)throw new Error("A and y must not reference the same GPU buffer.");if(m&&t._buf===s._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(c&&d!==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)*d+e)throw new Error("A does not have enough elements for the given n and lda.");if(t.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 g=(c?n.layout:l)==="column-major"?o==="upper":o==="lower",b=await P(r,[...["f64/dekker","f64/utils/add","f64/utils/multiply"],"dsyr2"],"dsyr2_main"),{hi:x,lo:v}=U(new Float64Array([a])),B=null,A=null,S=null,_=null,N=null,I=null,k=null,L=null,R=null;try{if(m)B=t._buf,A=t._loBuf,S=s._buf,_=s._loBuf,N=n._buf,I=n._loBuf;else{let nr=U(t),er=U(s),Q=U(n);B=h(r,nr.hi,"dsyr2-xHi",!1),A=h(r,nr.lo,"dsyr2-xLo",!1),S=h(r,er.hi,"dsyr2-yHi",!1),_=h(r,er.lo,"dsyr2-yLo",!1),N=h(r,Q.hi,"dsyr2-AHi",!0),I=h(r,Q.lo,"dsyr2-ALo",!0)}k=q(r,[{value:e,type:"u32"},{value:x[0],type:"f32"},{value:v[0],type:"f32"},{value:i,type:"u32"},{value:u,type:"u32"},{value:d,type:"u32"},{value:g?0:1,type:"u32"}],"dsyr2-params");let W=D(r,b.getBindGroupLayout(0),[B,A,S,_,N,I,k]),V=Math.min(e,r.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:C,ts:z}=O(r,b,W,V);L=c?null:E(r,C,N),R=c?null:E(r,C,I),F(r,C);let K=await j(z);if(c)return K!==void 0?{gpuTimeMs:K}:{};let Y=await G(L,Float32Array);L=null;let $=await G(R,Float32Array);R=null;let J=dr(Y,$);return K!==void 0?{A:J,gpuTimeMs:K}:{A:J}}finally{!m&&B&&f(B),!m&&A&&f(A),!w&&S&&f(S),!w&&_&&f(_),!c&&N&&f(N),!c&&I&&f(I),k&&f(k),L&&f(L),R&&f(R)}}async function Ra(r,o,e,a,t,i,s,u,n,d,l,m,w="row-major"){let c=i instanceof X,p=u instanceof M,g=l instanceof M;if(H(r),T(r,"dgemv",{A:i,x:u,y:l}),o!=="no-transpose"&&o!=="transpose")throw new Error("trans 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 t!="number")throw new Error("alpha must be a number.");if(Number.isNaN(t))throw new Error("alpha must not be NaN.");if(!Number.isFinite(t))throw new Error("alpha must be finite.");if(typeof d!="number")throw new Error("beta must be a number.");if(Number.isNaN(d))throw new Error("beta must not be NaN.");if(!Number.isFinite(d))throw new Error("beta must be finite.");if(!Number.isInteger(e)||!Number.isInteger(a)||!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 Float64Array))throw new Error("A must be a Float64Array or GpuMatrix.");if(c&&i.dtype!==Float64Array)throw new Error("A must be a Float64Array-backed GpuMatrix.");if(!p&&!(u instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(!g&&!(l instanceof Float64Array))throw new Error("y must be a Float64Array or GpuVector.");if(p&&u.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(g&&l.dtype!==Float64Array)throw new Error("y must be a Float64Array-backed GpuVector.");if(p!==g)throw new Error("x and y must be the same type (both Float64Array or both GpuVector).");if(p&&!c)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(c&&!p)throw new Error("x and y must be GpuVectors when A is a GpuMatrix.");if(p&&u._buf===l._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(c&&g&&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<a))throw new Error("A is too small for the given m and n.");if(e<0||a<0)throw new Error("m and n must be non-negative.");if(e===0||a===0)return g?{}:{y:l};(c?i.layout:w)==="column-major"&&([e,a]=[a,e],o=o==="no-transpose"?"transpose":"no-transpose");let b=o==="no-transpose",x=b?a:e,v=b?e:a;if(s<a)throw new Error("lda must be >= n.");if(!c&&i.length<(e-1)*s+a)throw new Error("A does not have enough elements for the given m, n, and lda.");if(u.length<(x-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 B=["f64/dekker","f64/utils/add","f64/utils/multiply"],A=b?"dgemv_n":"dgemv_t",S=b?"dgemv_n_main":"dgemv_t_main",_=await P(r,[...B,A],S),{hi:N,lo:I}=U(new Float64Array([t])),{hi:k,lo:L}=U(new Float64Array([d])),R=null,W=null,V=null,C=null,z=null,K=null,Y=null,$=null,J=null;try{if(c)R=i._buf,W=i._loBuf;else{let mr=U(i);R=h(r,mr.hi,"dgemv-AHi",!1),W=h(r,mr.lo,"dgemv-ALo",!1)}if(p)V=u._buf,C=u._loBuf;else{let mr=U(u);V=h(r,mr.hi,"dgemv-xHi",!1),C=h(r,mr.lo,"dgemv-xLo",!1)}if(g)z=l._buf,K=l._loBuf;else{let mr=U(l);z=h(r,mr.hi,"dgemv-yHi",!0),K=h(r,mr.lo,"dgemv-yLo",!0)}Y=q(r,[{value:e,type:"u32"},{value:a,type:"u32"},{value:N[0],type:"f32"},{value:I[0],type:"f32"},{value:k[0],type:"f32"},{value:L[0],type:"f32"},{value:n,type:"u32"},{value:m,type:"u32"},{value:s,type:"u32"}],"dgemv-params");let nr=D(r,_.getBindGroupLayout(0),[R,W,V,C,z,K,Y]),er=b?Math.min(e,r.limits.maxComputeWorkgroupsPerDimension):Yr(r,"dgemv",v),{commandEncoder:Q,ts:rr}=O(r,_,nr,er);$=g?null:E(r,Q,z),J=g?null:E(r,Q,K),F(r,Q);let ar=await j(rr);if(g)return ar!==void 0?{gpuTimeMs:ar}:{};let sr=await G($,Float32Array);$=null;let lr=await G(J,Float32Array);J=null;let ur=dr(sr,lr);return ar!==void 0?{y:ur,gpuTimeMs:ar}:{y:ur}}finally{!c&&R&&f(R),!c&&W&&f(W),!p&&V&&f(V),!p&&C&&f(C),!g&&z&&f(z),!g&&K&&f(K),Y&&f(Y),$&&f($),J&&f(J)}}async function ja(r,o,e,a,t,i,s,u,n,d,l,m="row-major"){let w=s instanceof M,c=d instanceof M,p=t instanceof X;if(H(r),T(r,"dsymv",{A:t,x:s,y:d}),o!=="lower"&&o!=="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 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 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(!p&&!(t instanceof Float64Array))throw new Error("A must be a Float64Array or GpuMatrix.");if(p&&t.dtype!==Float64Array)throw new Error("A must be a Float64Array-backed GpuMatrix.");if(!w&&!(s instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(!c&&!(d instanceof Float64Array))throw new Error("y must be a Float64Array or GpuVector.");if(w&&s.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(c&&d.dtype!==Float64Array)throw new Error("y must be a Float64Array-backed GpuVector.");if(w!==c)throw new Error("x and y must be the same type (both Float64Array or both GpuVector).");if(w&&!p)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(p&&!w)throw new Error("x and y must be GpuVectors when A is a GpuMatrix.");if(w&&s._buf===d._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(p&&c&&t._buf===d._buf)throw new Error("A and y must not reference the same GPU buffer.");if(p&&i!==t.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(p&&(t.rows<e||t.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:d};if(!p&&t.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(d.length<(e-1)*l+1)throw new Error("y does not have enough elements for the given n and incy.");let y=(p?t.layout:m)==="column-major"?o==="upper":o==="lower",x=await P(r,[...["f64/dekker","f64/utils/add","f64/utils/multiply"],"dsymv"],"dsymv_main"),{hi:v,lo:B}=U(new Float64Array([a])),{hi:A,lo:S}=U(new Float64Array([n])),_=null,N=null,I=null,k=null,L=null,R=null,W=null,V=null,C=null;try{if(p)_=t._buf,N=t._loBuf;else{let rr=U(t);_=h(r,rr.hi,"dsymv-AHi",!1),N=h(r,rr.lo,"dsymv-ALo",!1)}if(w)I=s._buf,k=s._loBuf;else{let rr=U(s);I=h(r,rr.hi,"dsymv-xHi",!1),k=h(r,rr.lo,"dsymv-xLo",!1)}if(c)L=d._buf,R=d._loBuf;else{let rr=U(d);L=h(r,rr.hi,"dsymv-yHi",!0),R=h(r,rr.lo,"dsymv-yLo",!0)}W=q(r,[{value:e,type:"u32"},{value:v[0],type:"f32"},{value:B[0],type:"f32"},{value:A[0],type:"f32"},{value:S[0],type:"f32"},{value:u,type:"u32"},{value:l,type:"u32"},{value:i,type:"u32"},{value:y?0:1,type:"u32"}],"dsymv-params");let z=D(r,x.getBindGroupLayout(0),[_,N,I,k,L,R,W]),K=Math.min(e,r.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:Y,ts:$}=O(r,x,z,K);V=c?null:E(r,Y,L),C=c?null:E(r,Y,R),F(r,Y);let J=await j($);if(c)return J!==void 0?{gpuTimeMs:J}:{};let nr=await G(V,Float32Array);V=null;let er=await G(C,Float32Array);C=null;let Q=dr(nr,er);return J!==void 0?{y:Q,gpuTimeMs:J}:{y:Q}}finally{!p&&_&&f(_),!p&&N&&f(N),!w&&I&&f(I),!w&&k&&f(k),!c&&L&&f(L),!c&&R&&f(R),W&&f(W),V&&f(V),C&&f(C)}}async function Fa(r,o,e,a,t,i,s,u,n,d,l,m="row-major"){let w=u instanceof M,c=d instanceof M,p=i instanceof X,g=a==="unit";if(H(r),T(r,"dtrmv",{A:i,x:u,y:d}),o!=="lower"&&o!=="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(!g&&a!=="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(t)||!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<t)throw new Error("lda must be >= n.");if(!p&&!(i instanceof Float64Array))throw new Error("A must be a Float64Array or GpuMatrix.");if(p&&i.dtype!==Float64Array)throw new Error("A must be a Float64Array-backed GpuMatrix.");if(!w&&!(u instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(!c&&!(d instanceof Float64Array))throw new Error("y must be a Float64Array or GpuVector.");if(w&&u.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(c&&d.dtype!==Float64Array)throw new Error("y must be a Float64Array-backed GpuVector.");if(w!==c)throw new Error("x and y must be the same type (both Float64Array or both GpuVector).");if(w&&u._buf===d._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(w&&!p)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(p&&!w)throw new Error("x and y must be GpuVectors when A is a GpuMatrix.");if(p&&c&&i._buf===d._buf)throw new Error("A and y must not reference the same GPU buffer.");if(p&&s!==i.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(p&&(i.rows<t||i.cols<t))throw new Error("A is too small for the given n.");if(t<0)throw new Error("n must be non-negative.");if(t===0)return c?{}:{y:d};if(!p&&i.length<(t-1)*s+t)throw new Error("A does not have enough elements for the given n and lda.");if(u.length<(t-1)*n+1)throw new Error("x does not have enough elements for the given n and incx.");if(d.length<(t-1)*l+1)throw new Error("y does not have enough elements for the given n and incy.");let b=(p?i.layout:m)==="column-major",x=b?o==="upper":o==="lower",v=b?e==="transpose":e==="no-transpose",A=await P(r,[...["f64/dekker","f64/utils/add","f64/utils/multiply"],"dtrmv"],"dtrmv_main"),S=null,_=null,N=null,I=null,k=null,L=null,R=null,W=null,V=null;try{if(p)S=i._buf,_=i._loBuf;else{let Q=U(i);S=h(r,Q.hi,"dtrmv-AHi",!1),_=h(r,Q.lo,"dtrmv-ALo",!1)}if(w)N=u._buf,I=u._loBuf;else{let Q=U(u);N=h(r,Q.hi,"dtrmv-xHi",!1),I=h(r,Q.lo,"dtrmv-xLo",!1)}if(c)k=d._buf,L=d._loBuf;else{let Q=U(d);k=h(r,Q.hi,"dtrmv-yHi",!0),L=h(r,Q.lo,"dtrmv-yLo",!0)}R=q(r,[{value:t,type:"u32"},{value:n,type:"u32"},{value:l,type:"u32"},{value:s,type:"u32"},{value:v?0:1,type:"u32"},{value:x?0:1,type:"u32"},{value:g?1:0,type:"u32"}],"dtrmv-params");let C=D(r,A.getBindGroupLayout(0),[S,_,N,I,k,L,R]),z=Math.min(t,r.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:K,ts:Y}=O(r,A,C,z);W=c?null:E(r,K,k),V=c?null:E(r,K,L),F(r,K);let $=await j(Y);if(c)return $!==void 0?{gpuTimeMs:$}:{};let J=await G(W,Float32Array);W=null;let nr=await G(V,Float32Array);V=null;let er=dr(J,nr);return $!==void 0?{y:er,gpuTimeMs:$}:{y:er}}finally{!p&&S&&f(S),!p&&_&&f(_),!w&&N&&f(N),!w&&I&&f(I),!c&&k&&f(k),!c&&L&&f(L),R&&f(R),W&&f(W),V&&f(V)}}function qa(r,o,e){let a=new ArrayBuffer(r*o),t=new DataView(a);for(let i=0;i<r;i++){let s=e(i),u=i*o;s.forEach((n,d)=>t.setUint32(u+d*4,n,!0))}return a}function Ha(r,o,e){let a=r.createBuffer({label:e,size:o.byteLength,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});return r.queue.writeBuffer(a,0,o),a}async function Ta(r,o,e,a,t,i,s,u,n,d="row-major"){let l=u instanceof M,m=i instanceof X,w=a==="unit";if(H(r),T(r,"dtrsv",{A:i,x:u}),o!=="lower"&&o!=="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&&a!=="non-unit")throw new Error("diag must be 'unit' or 'non-unit'.");if(d!=="row-major"&&d!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(!Number.isInteger(t)||!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<t)throw new Error("lda must be >= n.");if(!m&&!(i instanceof Float64Array))throw new Error("A must be a Float64Array or GpuMatrix.");if(m&&i.dtype!==Float64Array)throw new Error("A must be a Float64Array-backed GpuMatrix.");if(!l&&!(u instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(l&&u.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed 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<t||i.cols<t))throw new Error("A is too small for the given n.");if(t<0)throw new Error("n must be non-negative.");if(t===0)return l?{}:{x:u};if(!m&&i.length<(t-1)*s+t)throw new Error("A does not have enough elements for the given n and lda.");if(u.length<(t-1)*n+1)throw new Error("x does not have enough elements for the given n and incx.");let p=(m?i.layout:d)==="column-major",g=p?o==="upper":o==="lower",y=p?e==="transpose":e==="no-transpose",b=["f64/dekker","f64/utils/add","f64/utils/multiply","f64/utils/divide"],x=await P(r,[...b,"dtrsv_invert_block"],"dtrsv_invert_block_main"),v=await P(r,[...b,"dtrsv_apply_inverse"],"dtrsv_apply_inverse_main"),B=await P(r,[...b,"dtrsv_update"],"dtrsv_update_main"),A=y===g,S=[];for(let nr=0;nr<t;nr+=64)S.push(nr);A||S.reverse();let _=S.length,N=r.limits.maxComputeWorkgroupsPerDimension,I=r.limits.minUniformBufferOffsetAlignment,k=null,L=null,R=null,W=null,V=null,C=null,z=null,K=null,Y=null,$=null,J=null;try{if(m)k=i._buf,L=i._loBuf;else{let tr=U(i);k=h(r,tr.hi,"dtrsv-AHi",!1),L=h(r,tr.lo,"dtrsv-ALo",!1)}if(l)R=u._buf,W=u._loBuf;else{let tr=U(u);R=h(r,tr.hi,"dtrsv-xHi",!0),W=h(r,tr.lo,"dtrsv-xLo",!0)}V=fr(r,_*64*64*4,"dtrsv-AinvHi"),C=fr(r,_*64*64*4,"dtrsv-AinvLo");let nr=qa(_,I,tr=>{let gr=tr*64,Gr=Math.min(gr+64,t);return[n,tr,gr,Gr]});z=Ha(r,nr,"dtrsv-apply-params");let er=qa(_,I,tr=>{let gr=tr*64,Gr=Math.min(gr+64,t);return[t,n,s,y?0:1,g?0:1,gr,Gr]});K=Ha(r,er,"dtrsv-update-params");let{commandEncoder:Q,querySet:rr}=Mr(r);Y=q(r,[{value:t,type:"u32"},{value:s,type:"u32"},{value:y?0:1,type:"u32"},{value:g?0:1,type:"u32"},{value:w?1:0,type:"u32"}],"dtrsv-invert-params");let ar=D(r,x.getBindGroupLayout(0),[k,L,V,C,Y]);hr(Q,x,ar,{x:64,y:_},rr?{timestampWrites:{querySet:rr,beginningOfPassWriteIndex:0}}:void 0);for(let tr=0;tr<S.length;tr++){let gr=S[tr],Gr=Math.min(gr+64,t),Nr=gr/64,Tr=tr===S.length-1,Sr=Nr*I,xr=D(r,v.getBindGroupLayout(0),[V,C,R,W,{buffer:z,offset:Sr,size:16}]);hr(Q,v,xr,1,Tr&&rr?{timestampWrites:{querySet:rr,endOfPassWriteIndex:1}}:void 0);let Er=A?t-Gr:gr;if(Er===0)continue;let _r=D(r,B.getBindGroupLayout(0),[k,L,R,W,{buffer:K,offset:Sr,size:32}]),Wr=Math.min(Er,N);hr(Q,B,_r,Wr)}let lr=kr(r,Q,rr);$=l?null:E(r,Q,R),J=l?null:E(r,Q,W),F(r,Q);let ur=await j(lr);if(l)return ur!==void 0?{gpuTimeMs:ur}:{};let mr=await G($,Float32Array);$=null;let yr=await G(J,Float32Array);J=null;let wr=dr(mr,yr);return ur!==void 0?{x:wr,gpuTimeMs:ur}:{x:wr}}finally{!m&&k&&f(k),!m&&L&&f(L),!l&&R&&f(R),!l&&W&&f(W),V&&f(V),C&&f(C),z&&f(z),K&&f(K),Y&&f(Y),$&&f($),J&&f(J)}}async function Ca(r,o,e,a,t,i,s,u,n,d,l,m,w,c,p="row-major"){let g=u instanceof X,y=d instanceof X,b=w instanceof X;if(H(r),T(r,"sgemm",{A:u,B:d,C:w}),o!=="no-transpose"&&o!=="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(p!=="row-major"&&p!=="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(a)||!Number.isInteger(t)||!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(!g&&!(u instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!y&&!(d instanceof Float32Array))throw new Error("B must be a Float32Array or GpuMatrix.");if(!b&&!(w instanceof Float32Array))throw new Error("C must be a Float32Array or GpuMatrix.");if((g||y)&&!b)throw new Error("C must be a GpuMatrix when A or B is a GpuMatrix.");if(b&&(!g||!y))throw new Error("A and B must be GpuMatrix when C is a GpuMatrix.");if(a<0||t<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(a===0||t===0)return b?{}:{C:w};let x=g?u.layout:p,v=y?d.layout:p,B=b?w.layout:p,A=x==="column-major"?i:a,S=x==="column-major"?a:i,_=o==="no-transpose"?A:S,N=o==="no-transpose"?S:A;if(n<N)throw new Error(`lda must be >= ${x==="column-major"?"rows":"cols"} of A as stored.`);if(g){if(n!==u.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");let[lr,ur]=o==="no-transpose"?[a,i]:[i,a];if(u.rows<lr||u.cols<ur)throw new Error("A is too small for the given m, k, and transA.")}else if(u.length<(_-1)*n+N)throw new Error("A does not have enough elements for the given dimensions and lda.");let I=v==="column-major"?t:i,k=v==="column-major"?i:t,L=e==="no-transpose"?I:k,R=e==="no-transpose"?k:I;if(l<R)throw new Error(`ldb must be >= ${v==="column-major"?"rows":"cols"} of B as stored.`);if(y){if(l!==d.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");let[lr,ur]=e==="no-transpose"?[i,t]:[t,i];if(d.rows<lr||d.cols<ur)throw new Error("B is too small for the given n, k, and transB.")}else if(d.length<(L-1)*l+R)throw new Error("B does not have enough elements for the given dimensions and ldb.");let W=B==="column-major"?t:a,V=B==="column-major"?a:t;if(c<V)throw new Error(`ldc must be >= ${B==="column-major"?"rows":"cols"} of C as stored.`);if(b){if(c!==w.lda)throw new Error("ldc must match C.lda when C is a GpuMatrix.");if(w.rows<a||w.cols<t)throw new Error("C is too small for the given m and n.")}else if(w.length<(W-1)*c+V)throw new Error("C does not have enough elements for the given dimensions and ldc.");x==="column-major"&&(o=o==="no-transpose"?"transpose":"no-transpose"),v==="column-major"&&(e=e==="no-transpose"?"transpose":"no-transpose"),B==="column-major"&&([u,d]=[d,u],[g,y]=[y,g],[n,l]=[l,n],[o,e]=[e==="no-transpose"?"transpose":"no-transpose",o==="no-transpose"?"transpose":"no-transpose"],[a,t]=[t,a]);let C=Math.ceil(t/64),z=Math.ceil(a/64),K=C*z>=36,Y=await P(r,K?"sgemm_large":"sgemm_small"),$=g?u._buf:h(r,u,"sgemm-A",!1),J=y?d._buf:h(r,d,"sgemm-B",!1),nr=b?w._buf:h(r,w,"sgemm-C",!0),er=o==="no-transpose",Q=e==="no-transpose",rr=er&&Ee($,n,a,i),ar=Ee(J,l,Q?i:t,Q?t:i),sr=q(r,[{value:a,type:"u32"},{value:t,type:"u32"},{value:i,type:"u32"},{value:s,type:"f32"},{value:m,type:"f32"},{value:n,type:"u32"},{value:l,type:"u32"},{value:c,type:"u32"},{value:o==="transpose"?1:0,type:"u32"},{value:e==="transpose"?1:0,type:"u32"},{value:rr?1:0,type:"u32"},{value:ar?1:0,type:"u32"}],"sgemm-params");try{let lr=D(r,Y.getBindGroupLayout(0),[$,Dr(r,$),J,Dr(r,J),nr,sr]),ur=K?{x:or(r,C,"sgemm","x"),y:or(r,z,"sgemm","y")}:{x:or(r,Math.ceil(t/32),"sgemm","x"),y:or(r,Math.ceil(a/32),"sgemm","y")},{commandEncoder:mr,ts:yr}=O(r,Y,lr,ur),wr=b?null:E(r,mr,nr);F(r,mr);let tr=await j(yr);if(b)return tr!==void 0?{gpuTimeMs:tr}:{};let gr=await G(wr,Float32Array);return tr!==void 0?{C:gr,gpuTimeMs:tr}:{C:gr}}finally{g||f($),y||f(J),b||f(nr),f(sr)}}async function Wa(r,o,e,a,t,i,s,u,n,d,l,m,w,c,p,g="row-major"){let y=n instanceof X,b=l instanceof X,x=c instanceof X;if(H(r),T(r,"sgemmtr",{A:n,B:l,C:c}),o!=="lower"&&o!=="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(a!=="no-transpose"&&a!=="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 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 w!="number")throw new Error("beta must be a number.");if(Number.isNaN(w))throw new Error("beta must not be NaN.");if(!Number.isFinite(w))throw new Error("beta must be finite.");if(!Number.isInteger(t)||!Number.isInteger(i)||!Number.isInteger(s)||!Number.isInteger(d)||!Number.isInteger(m)||!Number.isInteger(p))throw new Error("m, n, k, lda, ldb, and ldc must be integers.");if(!y&&!(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(!x&&!(c instanceof Float32Array))throw new Error("C must be a Float32Array or GpuMatrix.");if((y||b)&&!x)throw new Error("C must be a GpuMatrix when A or B is a GpuMatrix.");if(x&&(!y||!b))throw new Error("A and B must be GpuMatrix when C is a GpuMatrix.");if(t<0||i<0||s<0)throw new Error("m, n, and k must be non-negative.");if(d<=0||m<=0||p<=0)throw new Error("lda, ldb, and ldc must be positive.");if(t===0||i===0)return x?{}:{C:c};let v=y?n.layout:g,B=b?l.layout:g,A=x?c.layout:g,S=v==="column-major"?s:t,_=v==="column-major"?t:s,N=e==="no-transpose"?S:_,I=e==="no-transpose"?_:S;if(d<I)throw new Error(`lda must be >= ${v==="column-major"?"rows":"cols"} of A as stored.`);if(y){if(d!==n.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");let[rr,ar]=e==="no-transpose"?[t,s]:[s,t];if(n.rows<rr||n.cols<ar)throw new Error("A is too small for the given m, k, and transA.")}else if(n.length<(N-1)*d+I)throw new Error("A does not have enough elements for the given dimensions and lda.");let k=B==="column-major"?i:s,L=B==="column-major"?s:i,R=a==="no-transpose"?k:L,W=a==="no-transpose"?L:k;if(m<W)throw new Error(`ldb must be >= ${B==="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[rr,ar]=a==="no-transpose"?[s,i]:[i,s];if(l.rows<rr||l.cols<ar)throw new Error("B is too small for the given n, k, and transB.")}else if(l.length<(R-1)*m+W)throw new Error("B does not have enough elements for the given dimensions and ldb.");let V=A==="column-major"?i:t,C=A==="column-major"?t:i;if(p<C)throw new Error(`ldc must be >= ${A==="column-major"?"rows":"cols"} of C as stored.`);if(x){if(p!==c.lda)throw new Error("ldc must match C.lda when C is a GpuMatrix.");if(c.rows<t||c.cols<i)throw new Error("C is too small for the given m and n.")}else if(c.length<(V-1)*p+C)throw new Error("C does not have enough elements for the given dimensions and ldc.");v==="column-major"&&(e=e==="no-transpose"?"transpose":"no-transpose"),B==="column-major"&&(a=a==="no-transpose"?"transpose":"no-transpose"),A==="column-major"&&([n,l]=[l,n],[y,b]=[b,y],[d,m]=[m,d],[e,a]=[a==="no-transpose"?"transpose":"no-transpose",e==="no-transpose"?"transpose":"no-transpose"],[t,i]=[i,t],o=o==="lower"?"upper":"lower");let z=Math.ceil(i/64),K=Math.ceil(t/64),Y=z*K>=36,$=await P(r,Y?"sgemmtr_large":"sgemmtr_small"),J=y?n._buf:h(r,n,"sgemmtr-A",!1),nr=b?l._buf:h(r,l,"sgemmtr-B",!1),er=x?c._buf:h(r,c,"sgemmtr-C",!0),Q=q(r,[{value:t,type:"u32"},{value:i,type:"u32"},{value:s,type:"u32"},{value:u,type:"f32"},{value:w,type:"f32"},{value:d,type:"u32"},{value:m,type:"u32"},{value:p,type:"u32"},{value:e==="transpose"?1:0,type:"u32"},{value:a==="transpose"?1:0,type:"u32"},{value:o==="upper"?1:0,type:"u32"}],"sgemmtr-params");try{let rr=D(r,$.getBindGroupLayout(0),[J,nr,er,Q]),ar=Y?{x:or(r,z,"sgemmtr","x"),y:or(r,K,"sgemmtr","y")}:{x:or(r,Math.ceil(i/32),"sgemmtr","x"),y:or(r,Math.ceil(t/32),"sgemmtr","y")},{commandEncoder:sr,ts:lr}=O(r,$,rr,ar),ur=x?null:E(r,sr,er);F(r,sr);let mr=await j(lr);if(x)return mr!==void 0?{gpuTimeMs:mr}:{};let yr=await G(ur,Float32Array);return mr!==void 0?{C:yr,gpuTimeMs:mr}:{C:yr}}finally{y||f(J),b||f(nr),x||f(er),f(Q)}}async function Va(r,o,e,a,t,i,s,u,n,d,l,m="row-major"){let w=s instanceof X,c=d instanceof X;if(H(r),T(r,"ssyrk",{A:s,C:d}),o!=="lower"&&o!=="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(a)||!Number.isInteger(t)||!Number.isInteger(u)||!Number.isInteger(l))throw new Error("n, k, lda, and ldc must be integers.");if(!w&&!(s instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!c&&!(d instanceof Float32Array))throw new Error("C must be a Float32Array or GpuMatrix.");if(w&&!c)throw new Error("C must be a GpuMatrix when A is a GpuMatrix.");if(c&&!w)throw new Error("A must be a GpuMatrix when C is a GpuMatrix.");if(a<0||t<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(a===0)return c?{}:{C:d};let p=w?s.layout:m,g=c?d.layout:m,y=p==="column-major"?t:a,b=p==="column-major"?a:t,x=e==="no-transpose"?y:b,v=e==="no-transpose"?b:y;if(u<v)throw new Error(`lda must be >= ${p==="column-major"?"rows":"cols"} of A as stored.`);if(w){if(u!==s.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");let[C,z]=e==="no-transpose"?[a,t]:[t,a];if(s.rows<C||s.cols<z)throw new Error("A is too small for the given n, k, and trans.")}else if(s.length<(x-1)*u+v)throw new Error("A does not have enough elements for the given dimensions and lda.");if(l<a)throw new Error("ldc must be >= n.");if(c){if(l!==d.lda)throw new Error("ldc must match C.lda when C is a GpuMatrix.");if(d.rows<a||d.cols<a)throw new Error("C is too small for the given n.")}else if(d.length<(a-1)*l+a)throw new Error("C does not have enough elements for the given dimensions and ldc.");let B=e;p==="column-major"&&(B=B==="no-transpose"?"transpose":"no-transpose");let A=B==="no-transpose"?"transpose":"no-transpose",S=o;g==="column-major"&&([B,A]=[A==="no-transpose"?"transpose":"no-transpose",B==="no-transpose"?"transpose":"no-transpose"],S=S==="lower"?"upper":"lower");let _=Math.ceil(a/64),N=Math.ceil(a/64),I=_*N>=36,k=await P(r,I?"sgemmtr_large":"sgemmtr_small"),L=w?s._buf:h(r,s,"ssyrk-A",!1),R=c?d._buf:h(r,d,"ssyrk-C",!0),W=w?fr(r,L.size,"ssyrk-B",GPUBufferUsage.COPY_DST):h(r,s,"ssyrk-B",!1),V=q(r,[{value:a,type:"u32"},{value:a,type:"u32"},{value:t,type:"u32"},{value:i,type:"f32"},{value:n,type:"f32"},{value:u,type:"u32"},{value:u,type:"u32"},{value:l,type:"u32"},{value:B==="transpose"?1:0,type:"u32"},{value:A==="transpose"?1:0,type:"u32"},{value:S==="upper"?1:0,type:"u32"}],"ssyrk-params");try{let C=D(r,k.getBindGroupLayout(0),[L,W,R,V]),z=I?{x:or(r,_,"ssyrk","x"),y:or(r,N,"ssyrk","y")}:{x:or(r,Math.ceil(a/32),"ssyrk","x"),y:or(r,Math.ceil(a/32),"ssyrk","y")},{commandEncoder:K,querySet:Y,passDescriptor:$}=Mr(r);w&&K.copyBufferToBuffer(L,0,W,0,L.size),hr(K,k,C,z,$);let J=kr(r,K,Y),nr=c?null:E(r,K,R);F(r,K);let er=await j(J);if(c)return er!==void 0?{gpuTimeMs:er}:{};let Q=await G(nr,Float32Array);return er!==void 0?{C:Q,gpuTimeMs:er}:{C:Q}}finally{w||f(L),f(W),c||f(R),f(V)}}async function Oa(r,o,e,a,t,i,s,u,n,d,l,m,w,c="row-major"){let p=s instanceof X,g=n instanceof X,y=m instanceof X;if(H(r),T(r,"ssyr2k",{A:s,B:n,C:m}),o!=="lower"&&o!=="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(a)||!Number.isInteger(t)||!Number.isInteger(u)||!Number.isInteger(d)||!Number.isInteger(w))throw new Error("n, k, lda, ldb, and ldc must be integers.");if(!p&&!(s instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!g&&!(n instanceof Float32Array))throw new Error("B must be a Float32Array or GpuMatrix.");if(!y&&!(m instanceof Float32Array))throw new Error("C must be a Float32Array or GpuMatrix.");if((p||g)&&!y)throw new Error("C must be a GpuMatrix when A or B is a GpuMatrix.");if(y&&(!p||!g))throw new Error("A and B must be GpuMatrix when C is a GpuMatrix.");if(a<0||t<0)throw new Error("n and k must be non-negative.");if(u<=0||d<=0||w<=0)throw new Error("lda, ldb, and ldc must be positive.");if(a===0)return y?{}:{C:m};let b=p?s.layout:c,x=g?n.layout:c,v=y?m.layout:c,B=b==="column-major"?t:a,A=b==="column-major"?a:t,S=e==="no-transpose"?B:A,_=e==="no-transpose"?A:B;if(u<_)throw new Error(`lda must be >= ${b==="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[lr,ur]=e==="no-transpose"?[a,t]:[t,a];if(s.rows<lr||s.cols<ur)throw new Error("A is too small for the given n, k, and trans.")}else if(s.length<(S-1)*u+_)throw new Error("A does not have enough elements for the given dimensions and lda.");let N=x==="column-major"?t:a,I=x==="column-major"?a:t,k=e==="no-transpose"?N:I,L=e==="no-transpose"?I:N;if(d<L)throw new Error(`ldb must be >= ${x==="column-major"?"rows":"cols"} of B as stored.`);if(g){if(d!==n.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");let[lr,ur]=e==="no-transpose"?[a,t]:[t,a];if(n.rows<lr||n.cols<ur)throw new Error("B is too small for the given n, k, and trans.")}else if(n.length<(k-1)*d+L)throw new Error("B does not have enough elements for the given dimensions and ldb.");if(w<a)throw new Error("ldc must be >= n.");if(y){if(w!==m.lda)throw new Error("ldc must match C.lda when C is a GpuMatrix.");if(m.rows<a||m.cols<a)throw new Error("C is too small for the given n.")}else if(m.length<(a-1)*w+a)throw new Error("C does not have enough elements for the given dimensions and ldc.");let R=e;b==="column-major"&&(R=R==="no-transpose"?"transpose":"no-transpose");let W=e;x==="column-major"&&(W=W==="no-transpose"?"transpose":"no-transpose");let V=v==="column-major"?o==="lower"?"upper":"lower":o,C=lr=>lr==="no-transpose"?"transpose":"no-transpose";function z(lr,ur,mr,yr,wr,tr){let gr=lr,Gr=C(yr);return v!=="column-major"?{transX:gr,X:ur,ldX:mr,transY:Gr,Y:wr,ldY:tr}:{transX:C(Gr),X:wr,ldX:tr,transY:C(gr),Y:ur,ldY:mr}}let K=Math.ceil(a/64),Y=Math.ceil(a/64),$=K*Y>=36,J=await P(r,$?"sgemmtr_large":"sgemmtr_small"),nr=$?{x:or(r,K,"ssyr2k","x"),y:or(r,Y,"ssyr2k","y")}:{x:or(r,Math.ceil(a/32),"ssyr2k","x"),y:or(r,Math.ceil(a/32),"ssyr2k","y")},er=p?s._buf:h(r,s,"ssyr2k-A",!1),Q=g?n._buf:h(r,n,"ssyr2k-B",!1),rr=y?m._buf:h(r,m,"ssyr2k-C",!0),ar=null,sr=null;try{let lr=z(R,er,u,W,Q,d),ur=z(W,Q,d,R,er,u),mr=(Er,_r)=>q(r,[{value:a,type:"u32"},{value:a,type:"u32"},{value:t,type:"u32"},{value:i,type:"f32"},{value:_r,type:"f32"},{value:Er.ldX,type:"u32"},{value:Er.ldY,type:"u32"},{value:w,type:"u32"},{value:Er.transX==="transpose"?1:0,type:"u32"},{value:Er.transY==="transpose"?1:0,type:"u32"},{value:V==="upper"?1:0,type:"u32"}],"ssyr2k-params");ar=mr(lr,l),sr=mr(ur,1);let yr=D(r,J.getBindGroupLayout(0),[lr.X,lr.Y,rr,ar]),wr=D(r,J.getBindGroupLayout(0),[ur.X,ur.Y,rr,sr]),{commandEncoder:tr,querySet:gr}=Mr(r),Gr=gr?{timestampWrites:{querySet:gr,beginningOfPassWriteIndex:0}}:void 0,Nr=gr?{timestampWrites:{querySet:gr,endOfPassWriteIndex:1}}:void 0;hr(tr,J,yr,nr,Gr),hr(tr,J,wr,nr,Nr);let Tr=kr(r,tr,gr),Sr=y?null:E(r,tr,rr);F(r,tr);let xr=await j(Tr);if(y)return xr!==void 0?{gpuTimeMs:xr}:{};let vr=await G(Sr,Float32Array);return xr!==void 0?{C:vr,gpuTimeMs:xr}:{C:vr}}finally{p||f(er),g||f(Q),y||f(rr),ar&&f(ar),sr&&f(sr)}}async function Ka(r,o,e,a,t,i,s,u,n,d,l,m,w,c="row-major"){let p=s instanceof X,g=n instanceof X,y=m instanceof X;if(H(r),T(r,"ssymm",{A:s,B:n,C:m}),o!=="left"&&o!=="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(a)||!Number.isInteger(t)||!Number.isInteger(u)||!Number.isInteger(d)||!Number.isInteger(w))throw new Error("m, n, lda, ldb, and ldc must be integers.");if(!p&&!(s instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!g&&!(n instanceof Float32Array))throw new Error("B must be a Float32Array or GpuMatrix.");if(!y&&!(m instanceof Float32Array))throw new Error("C must be a Float32Array or GpuMatrix.");if((p||g)&&!y)throw new Error("C must be a GpuMatrix when A or B is a GpuMatrix.");if(y&&(!p||!g))throw new Error("A and B must be GpuMatrix when C is a GpuMatrix.");if(a<0||t<0)throw new Error("m and n must be non-negative.");if(a===0||t===0)return y?{}:{C:m};let b=p?s.layout:c,x=g?n.layout:c,v=y?m.layout:c,B=o==="left"?a:t;if(u<B)throw new Error("lda must be >= "+(o==="left"?"m":"n")+".");if(p){if(u!==s.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(s.rows<B||s.cols<B)throw new Error("A is too small for the given m/n and side.")}else if(s.length<(B-1)*u+B)throw new Error("A does not have enough elements for the given dimensions and lda.");let A=x==="column-major"?t:a,S=x==="column-major"?a:t;if(d<S)throw new Error(`ldb must be >= ${x==="column-major"?"rows":"cols"} of B as stored.`);if(g){if(d!==n.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");if(n.rows<a||n.cols<t)throw new Error("B is too small for the given m and n.")}else if(n.length<(A-1)*d+S)throw new Error("B does not have enough elements for the given dimensions and ldb.");let _=v==="column-major"?t:a,N=v==="column-major"?a:t;if(w<N)throw new Error(`ldc must be >= ${v==="column-major"?"rows":"cols"} of C as stored.`);if(y){if(w!==m.lda)throw new Error("ldc must match C.lda when C is a GpuMatrix.");if(m.rows<a||m.cols<t)throw new Error("C is too small for the given m and n.")}else if(m.length<(_-1)*w+N)throw new Error("C does not have enough elements for the given dimensions and ldc.");let I=b==="column-major"?e==="lower"?"upper":"lower":e,k=x==="column-major"?"transpose":"no-transpose",L="no-transpose",R=a,W=t,V=B,C=o==="left"?L:k,z=o==="left"?k:L,K=tr=>tr==="no-transpose"?"transpose":"no-transpose",Y=o==="right";v==="column-major"&&([C,z]=[K(z),K(C)],Y=!Y,[R,W]=[W,R]);let $=B,J=Math.ceil(W/64),nr=Math.ceil(R/64),er=J*nr>=36,Q=await P(r,er?"sgemm_large":"sgemm_small"),rr=await P(r,"symmetrize"),ar=er?{x:or(r,J,"ssymm","x"),y:or(r,nr,"ssymm","y")}:{x:or(r,Math.ceil(W/32),"ssymm","x"),y:or(r,Math.ceil(R/32),"ssymm","y")},sr=p?s._buf:h(r,s,"ssymm-A",!1),lr=g?n._buf:h(r,n,"ssymm-B",!1),ur=y?m._buf:h(r,m,"ssymm-C",!0),mr=fr(r,B*$*4,"ssymm-Adense"),yr=null,wr=null;try{yr=q(r,[{value:B,type:"u32"},{value:u,type:"u32"},{value:$,type:"u32"},{value:I==="upper"?1:0,type:"u32"}],"ssymm-sym-params");let tr=D(r,rr.getBindGroupLayout(0),[sr,mr,yr]),gr=Y?lr:mr,Gr=Y?d:$,Nr=Y?mr:lr;wr=q(r,[{value:R,type:"u32"},{value:W,type:"u32"},{value:V,type:"u32"},{value:i,type:"f32"},{value:l,type:"f32"},{value:Gr,type:"u32"},{value:Y?$:d,type:"u32"},{value:w,type:"u32"},{value:C==="transpose"?1:0,type:"u32"},{value:z==="transpose"?1:0,type:"u32"}],"ssymm-gemm-params");let Sr=D(r,Q.getBindGroupLayout(0),[gr,Dr(r,gr),Nr,Dr(r,Nr),ur,wr]),{commandEncoder:xr,querySet:vr}=Mr(r),Er=vr?{timestampWrites:{querySet:vr,beginningOfPassWriteIndex:0}}:void 0,_r=vr?{timestampWrites:{querySet:vr,endOfPassWriteIndex:1}}:void 0;hr(xr,rr,tr,{x:Math.ceil(B/8),y:Math.ceil(B/8)},Er),hr(xr,Q,Sr,ar,_r);let Wr=kr(r,xr,vr),Ur=y?null:E(r,xr,ur);F(r,xr);let Zr=await j(Wr);if(y)return Zr!==void 0?{gpuTimeMs:Zr}:{};let fe=await G(Ur,Float32Array);return Zr!==void 0?{C:fe,gpuTimeMs:Zr}:{C:fe}}finally{p||f(sr),g||f(lr),y||f(ur),f(mr),yr&&f(yr),wr&&f(wr)}}async function Ua(r,o,e,a,t,i,s,u,n,d,l,m,w="row-major"){let c=n instanceof X,p=l instanceof X,g=t==="unit";if(H(r),T(r,"strmm",{A:n,B:l}),o!=="left"&&o!=="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(a!=="no-transpose"&&a!=="transpose")throw new Error("transA must be 'no-transpose' or 'transpose'.");if(!g&&t!=="non-unit")throw new Error("diag must be 'unit' or 'non-unit'.");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(!Number.isInteger(i)||!Number.isInteger(s)||!Number.isInteger(d)||!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(!p&&!(l instanceof Float32Array))throw new Error("B must be a Float32Array or GpuMatrix.");if(c!==p)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 p?{}:{B:l};let y=c?n.layout:w,b=p?l.layout:w,x=o==="left"?i:s;if(d<x)throw new Error("lda must be >= "+(o==="left"?"m":"n")+".");if(c){if(d!==n.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(n.rows<x||n.cols<x)throw new Error("A is too small for the given m/n and side.")}else if(n.length<(x-1)*d+x)throw new Error("A does not have enough elements for the given dimensions and lda.");let v=b==="column-major"?s:i,B=b==="column-major"?i:s;if(m<B)throw new Error(`ldb must be >= ${b==="column-major"?"rows":"cols"} of B as stored.`);if(p){if(m!==l.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");if(l.rows<i||l.cols<s)throw new Error("B is too small for the given m and n.")}else if(l.length<(v-1)*m+B)throw new Error("B does not have enough elements for the given dimensions and ldb.");let A=y==="column-major"?e==="lower"?"upper":"lower":e,S=y==="column-major"?a==="no-transpose"?"transpose":"no-transpose":a,_=b==="column-major"?"transpose":"no-transpose",N="no-transpose",I=i,k=s,L=x,R=o==="left"?N:_,W=o==="left"?_:N,V=yr=>yr==="no-transpose"?"transpose":"no-transpose",C=o==="right";b==="column-major"&&([R,W]=[V(W),V(R)],C=!C,[I,k]=[k,I]);let z=x,K=Math.ceil(k/64),Y=Math.ceil(I/64),$=K*Y>=36,J=await P(r,$?"sgemm_large":"sgemm_small"),nr=await P(r,"triangularize"),er=$?{x:or(r,K,"strmm","x"),y:or(r,Y,"strmm","y")}:{x:or(r,Math.ceil(k/32),"strmm","x"),y:or(r,Math.ceil(I/32),"strmm","y")},Q=null,rr=null,ar=null,sr=null,lr=null,ur=null,mr=!1;try{Q=c?n._buf:h(r,n,"strmm-A",!1),rr=p?l._buf:h(r,l,"strmm-B",!0),ar=fr(r,x*z*4,"strmm-Adense"),sr=fr(r,v*m*4,"strmm-out",GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST),lr=q(r,[{value:x,type:"u32"},{value:d,type:"u32"},{value:z,type:"u32"},{value:A==="upper"?1:0,type:"u32"},{value:S==="transpose"?1:0,type:"u32"},{value:g?1:0,type:"u32"}],"strmm-tri-params");let yr=D(r,nr.getBindGroupLayout(0),[Q,ar,lr]),wr=C?rr:ar,tr=C?m:z,gr=C?ar:rr;ur=q(r,[{value:I,type:"u32"},{value:k,type:"u32"},{value:L,type:"u32"},{value:u,type:"f32"},{value:0,type:"f32"},{value:tr,type:"u32"},{value:C?z:m,type:"u32"},{value:m,type:"u32"},{value:R==="transpose"?1:0,type:"u32"},{value:W==="transpose"?1:0,type:"u32"}],"strmm-gemm-params");let Nr=D(r,J.getBindGroupLayout(0),[wr,Dr(r,wr),gr,Dr(r,gr),sr,ur]),{commandEncoder:Tr,querySet:Sr}=Mr(r);Tr.copyBufferToBuffer(rr,0,sr,0,Math.min(rr.size,sr.size));let xr=Sr?{timestampWrites:{querySet:Sr,beginningOfPassWriteIndex:0}}:void 0,vr=Sr?{timestampWrites:{querySet:Sr,endOfPassWriteIndex:1}}:void 0;hr(Tr,nr,yr,{x:Math.ceil(x/8),y:Math.ceil(x/8)},xr),hr(Tr,J,Nr,er,vr);let Er=kr(r,Tr,Sr),_r=p?null:E(r,Tr,sr);F(r,Tr);let Wr=await j(Er);if(p)return f(l._buf),l._buf=sr,mr=!0,Wr!==void 0?{gpuTimeMs:Wr}:{};let Ur=await G(_r,Float32Array);return Wr!==void 0?{B:Ur,gpuTimeMs:Wr}:{B:Ur}}finally{!c&&Q&&f(Q),!p&&rr&&f(rr),ar&&f(ar),sr&&!mr&&f(sr),lr&&f(lr),ur&&f(ur)}}async function za(r,o,e,a,t,i,s,u,n,d,l,m,w="row-major"){let c=n instanceof X,p=l instanceof X,g=t==="unit";if(H(r),T(r,"strsm",{A:n,B:l}),o!=="left"&&o!=="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(a!=="no-transpose"&&a!=="transpose")throw new Error("transA must be 'no-transpose' or 'transpose'.");if(!g&&t!=="non-unit")throw new Error("diag must be 'unit' or 'non-unit'.");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(!Number.isInteger(i)||!Number.isInteger(s)||!Number.isInteger(d)||!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(!p&&!(l instanceof Float32Array))throw new Error("B must be a Float32Array or GpuMatrix.");if(c!==p)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 p?{}:{B:l};let y=c?n.layout:w,b=p?l.layout:w,x=o==="left"?i:s;if(d<x)throw new Error("lda must be >= "+(o==="left"?"m":"n")+".");if(c){if(d!==n.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(n.rows<x||n.cols<x)throw new Error("A is too small for the given m/n and side.")}else if(n.length<(x-1)*d+x)throw new Error("A does not have enough elements for the given dimensions and lda.");let v=b==="column-major"?s:i,B=b==="column-major"?i:s;if(m<B)throw new Error(`ldb must be >= ${b==="column-major"?"rows":"cols"} of B as stored.`);if(p){if(m!==l.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");if(l.rows<i||l.cols<s)throw new Error("B is too small for the given m and n.")}else if(l.length<(v-1)*m+B)throw new Error("B does not have enough elements for the given dimensions and ldb.");let A=y==="column-major"?e==="lower"?"upper":"lower":e,S=y==="column-major"?a==="no-transpose"?"transpose":"no-transpose":a,_=o==="left"?s:i,N=o==="left",I=S==="no-transpose"==(A==="lower"),k=o==="left"?I:!I,L=[];for(let rr=0;rr<x;rr+=64)L.push(rr);k||L.reverse();let R=L.length,W=await P(r,"strsv_invert_block"),V=await P(r,"block_transfer"),C=await P(r,"sscal"),z=null,K=null,Y=null,$=[],J=[];function nr(rr,ar){let sr=fr(r,rr,ar);return J.push(sr),sr}function er(rr,ar){let sr=q(r,rr,ar);return $.push(sr),sr}let Q=(v-1)*m+B;try{z=c?n._buf:h(r,n,"strsm-A",!1),K=p?l._buf:h(r,l,"strsm-B",!0),Y=fr(r,R*64*64*4,"strsm-Ainv");let rr=null;if(u!==1&&u!==0){let Sr=er([{value:Q,type:"u32"},{value:u,type:"f32"},{value:1,type:"u32"}],"strsm-scale-params");rr=D(r,C.getBindGroupLayout(0),[K,Sr])}let ar=er([{value:x,type:"u32"},{value:d,type:"u32"},{value:S==="transpose"?1:0,type:"u32"},{value:A==="upper"?1:0,type:"u32"},{value:g?1:0,type:"u32"}],"strsm-invert-params"),sr=D(r,W.getBindGroupLayout(0),[z,Y,ar]),lr=nr(64*_*4,"strsm-Bblock"),ur=nr(64*_*4,"strsm-Xblock"),mr=nr(x*64*4,"strsm-Aoff"),yr=nr(x*_*4,"strsm-delta"),{commandEncoder:wr,querySet:tr}=Mr(r);if(u===0){let Sr=Math.ceil(B/64),xr=Math.ceil(v/64),vr=Sr*xr>=36,Er=await P(r,vr?"sgemm_large":"sgemm_small"),_r=er([{value:v,type:"u32"},{value:B,type:"u32"},{value:0,type:"u32"},{value:0,type:"f32"},{value:0,type:"f32"},{value:1,type:"u32"},{value:1,type:"u32"},{value:m,type:"u32"},{value:0,type:"u32"},{value:0,type:"u32"}],"strsm-zero-params"),Wr=D(r,Er.getBindGroupLayout(0),[Y,Dr(r,Y),Y,Dr(r,Y),K,_r]),Ur=vr?{x:or(r,Sr,"strsm","x"),y:or(r,xr,"strsm","y")}:{x:or(r,Math.ceil(B/32),"strsm","x"),y:or(r,Math.ceil(v/32),"strsm","y")};hr(wr,Er,Wr,Ur,tr?{timestampWrites:{querySet:tr,beginningOfPassWriteIndex:0,endOfPassWriteIndex:1}}:void 0)}else{rr&&hr(wr,C,rr,br(r,Q)),hr(wr,W,sr,{x:64,y:R},tr?{timestampWrites:{querySet:tr,beginningOfPassWriteIndex:0}}:void 0);for(let xr=0;xr<L.length;xr++){let vr=L[xr],Er=Math.min(vr+64,x),_r=Er-vr,Wr=vr/64,Ur=xr===L.length-1,Zr=er([{value:vr,type:"u32"},{value:_r,type:"u32"},{value:0,type:"u32"},{value:_,type:"u32"},{value:m,type:"u32"},{value:b==="column-major"?1:0,type:"u32"},{value:N?1:0,type:"u32"},{value:2,type:"u32"}],"strsm-gather-B-params"),fe=D(r,V.getBindGroupLayout(0),[lr,K,Zr]);hr(wr,V,fe,Yr(r,"strsm",_r,_));{let Xr=_r,$r=_,_e=_r,ae=Math.ceil($r/64),ie=Math.ceil(Xr/64),se=ae*ie>=36,ne=await P(r,se?"sgemm_large":"sgemm_small"),Be=er([{value:Xr,type:"u32"},{value:$r,type:"u32"},{value:_e,type:"u32"},{value:1,type:"f32"},{value:0,type:"f32"},{value:64,type:"u32"},{value:_,type:"u32"},{value:_,type:"u32"},{value:o==="right"?1:0,type:"u32"},{value:0,type:"u32"}],"strsm-apply-params"),ce={buffer:Y,offset:Wr*64*64*4,size:4096*4},Ae=D(r,ne.getBindGroupLayout(0),[ce,Dr(r,ce),lr,Dr(r,lr),ur,Be]),ti=se?{x:or(r,ae,"strsm","x"),y:or(r,ie,"strsm","y")}:{x:or(r,Math.ceil($r/32),"strsm","x"),y:or(r,Math.ceil(Xr/32),"strsm","y")};hr(wr,ne,Ae,ti)}let me=k?Er:0,Re=k?x:vr,je=me<Re,Ya=er([{value:vr,type:"u32"},{value:_r,type:"u32"},{value:0,type:"u32"},{value:_,type:"u32"},{value:m,type:"u32"},{value:b==="column-major"?1:0,type:"u32"},{value:N?1:0,type:"u32"},{value:0,type:"u32"}],"strsm-scatter-params"),Za=D(r,V.getBindGroupLayout(0),[ur,K,Ya]),Xa=Ur&&!je&&tr?{timestampWrites:{querySet:tr,endOfPassWriteIndex:1}}:void 0;if(hr(wr,V,Za,Yr(r,"strsm",_r,_),Xa),!je)continue;let oe=Re-me,$a=er([{value:me,type:"u32"},{value:oe,type:"u32"},{value:vr,type:"u32"},{value:_r,type:"u32"},{value:d,type:"u32"},{value:S==="transpose"?1:0,type:"u32"},{value:N?1:0,type:"u32"},{value:2,type:"u32"}],"strsm-gather-A-params"),Ja=D(r,V.getBindGroupLayout(0),[mr,z,$a]);hr(wr,V,Ja,Yr(r,"strsm",oe,_r));{let Xr=oe,$r=_,_e=_r,ae=Math.ceil($r/64),ie=Math.ceil(Xr/64),se=ae*ie>=36,ne=await P(r,se?"sgemm_large":"sgemm_small"),Be=er([{value:Xr,type:"u32"},{value:$r,type:"u32"},{value:_e,type:"u32"},{value:1,type:"f32"},{value:0,type:"f32"},{value:_r,type:"u32"},{value:_,type:"u32"},{value:_,type:"u32"},{value:0,type:"u32"},{value:0,type:"u32"}],"strsm-update-params"),ce=D(r,ne.getBindGroupLayout(0),[mr,Dr(r,mr),ur,Dr(r,ur),yr,Be]),Ae=se?{x:or(r,ae,"strsm","x"),y:or(r,ie,"strsm","y")}:{x:or(r,Math.ceil($r/32),"strsm","x"),y:or(r,Math.ceil(Xr/32),"strsm","y")};hr(wr,ne,ce,Ae)}let Qa=er([{value:me,type:"u32"},{value:oe,type:"u32"},{value:0,type:"u32"},{value:_,type:"u32"},{value:m,type:"u32"},{value:b==="column-major"?1:0,type:"u32"},{value:N?1:0,type:"u32"},{value:1,type:"u32"}],"strsm-scatter-sub-params"),ri=D(r,V.getBindGroupLayout(0),[yr,K,Qa]),ei=Ur&&tr?{timestampWrites:{querySet:tr,endOfPassWriteIndex:1}}:void 0;hr(wr,V,ri,Yr(r,"strsm",oe,_),ei)}}let gr=kr(r,wr,tr),Gr=p?null:E(r,wr,K);F(r,wr);let Nr=await j(gr);if(p)return Nr!==void 0?{gpuTimeMs:Nr}:{};let Tr=await G(Gr,Float32Array);return Nr!==void 0?{B:Tr,gpuTimeMs:Nr}:{B:Tr}}finally{!c&&z&&f(z),!p&&K&&f(K),Y&&f(Y),f(J),f($)}}return li(_s);})();