wgblas 1.2.1 → 2.0.0

package/dist/wgblas.browser.js

@@ -1,4 +1,4 @@
-var wgblas=(()=>{var ft=Object.create;var ur=Object.defineProperty;var ct=Object.getOwnPropertyDescriptor;var mt=Object.getOwnPropertyNames;var dt=Object.getPrototypeOf,pt=Object.prototype.hasOwnProperty;var lr=(a=>typeof require<"u"?require:typeof Proxy<"u"?new Proxy(a,{get:(e,r)=>(typeof require<"u"?require:e)[r]}):a)(function(a){if(typeof require<"u")return require.apply(this,arguments);throw Error('Dynamic require of "'+a+'" is not supported')});var W=(a,e,r)=>()=>{if(r)throw r[0];try{return a&&(e=a(a=0)),e}catch(o){throw r=[o],o}};var vr=(a,e)=>{for(var r in e)ur(a,r,{get:e[r],enumerable:!0})},yr=(a,e,r,o)=>{if(e&&typeof e=="object"||typeof e=="function")for(let t of mt(e))!pt.call(a,t)&&t!==r&&ur(a,t,{get:()=>e[t],enumerable:!(o=ct(e,t))||o.enumerable});return a};var fr=(a,e,r)=>(r=a!=null?ft(dt(a)):{},yr(e||!a||!a.__esModule?ur(r,"default",{value:a,enumerable:!0}):r,a)),gt=a=>yr(ur({},"__esModule",{value:!0}),a);var Ir,Fr=W(()=>{Ir=`// amax reduction: collapses 2*WGS (value, index) pairs into one index.
+var wgblas=(()=>{var Do=Object.create;var Ur=Object.defineProperty;var Ro=Object.getOwnPropertyDescriptor;var To=Object.getOwnPropertyNames;var jo=Object.getPrototypeOf,Co=Object.prototype.hasOwnProperty;var Or=(a=>typeof require<"u"?require:typeof Proxy<"u"?new Proxy(a,{get:(e,r)=>(typeof require<"u"?require:e)[r]}):a)(function(a){if(typeof require<"u")return require.apply(this,arguments);throw Error('Dynamic require of "'+a+'" is not supported')});var O=(a,e,r)=>()=>{if(r)throw r[0];try{return a&&(e=a(a=0)),e}catch(o){throw r=[o],o}};var te=(a,e)=>{for(var r in e)Ur(a,r,{get:e[r],enumerable:!0})},oe=(a,e,r,o)=>{if(e&&typeof e=="object"||typeof e=="function")for(let t of To(e))!Co.call(a,t)&&t!==r&&Ur(a,t,{get:()=>e[t],enumerable:!(o=Ro(e,t))||o.enumerable});return a};var Vr=(a,e,r)=>(r=a!=null?Do(jo(a)):{},oe(e||!a||!a.__esModule?Ur(r,"default",{value:a,enumerable:!0}):r,a)),Wo=a=>oe(Ur({},"__esModule",{value:!0}),a);var we,pe=O(()=>{we=`// 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.
 
@@ -41,7 +41,57 @@ fn reduce(
 
   if (i == 0u) { result[0] = tile_idx[0]; }
 }
-`});var Wr,Nr=W(()=>{Wr=`// sum reduction: collapses 2*WGS partials into one scalar.
+`});var be,ge=O(()=>{be=`// 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/
+// partialsIdx must have exactly 2*WGS entries each. Concatenated after
+// f64/dekker.wgsl (DD struct), f64/utils/greater.wgsl (ddGreater), and
+// f64/utils/equal.wgsl (ddEqual).
+
+@group(0) @binding(0) var<storage, read>       partialsValHi: array<f32>;
+@group(0) @binding(1) var<storage, read>       partialsValLo: array<f32>;
+@group(0) @binding(2) var<storage, read>       partialsIdx:   array<u32>;
+@group(0) @binding(3) var<storage, read_write> result:        array<u32>;
+
+const WGS: u32 = 64;
+
+var<workgroup> tile_val: array<DD, 64>;
+var<workgroup> tile_idx: array<u32, 64>;
+
+@compute @workgroup_size(64)
+fn reduce_f64(
+  @builtin(local_invocation_id) lid: vec3u,
+) {
+  let i = lid.x;
+  let a_val = DD(partialsValHi[i], partialsValLo[i]);
+  let b_val = DD(partialsValHi[i + WGS], partialsValLo[i + WGS]);
+  if (ddGreater(b_val, a_val) ||
+      (ddEqual(b_val, a_val) && partialsIdx[i + WGS] < partialsIdx[i])) {
+    tile_val[i] = b_val;
+    tile_idx[i] = partialsIdx[i + WGS];
+  } else {
+    tile_val[i] = a_val;
+    tile_idx[i] = partialsIdx[i];
+  }
+  workgroupBarrier();
+
+  for (var s = WGS / 2u; s > 0u; s >>= 1u) {
+    if (i < s) {
+      let c_val = tile_val[i];
+      let d_val = tile_val[i + s];
+      if (ddGreater(d_val, c_val) ||
+          (ddEqual(d_val, c_val) && tile_idx[i + s] < tile_idx[i])) {
+        tile_val[i] = d_val;
+        tile_idx[i] = tile_idx[i + s];
+      }
+    }
+    workgroupBarrier();
+  }
+
+  if (i == 0u) { result[0] = tile_idx[0]; }
+}
+`});var xe,he=O(()=>{xe=`// sum reduction: collapses 2*WGS partials into one scalar.
 // dispatch: 1 workgroup of WGS threads.
 // partials must have exactly 2*WGS entries.
 
@@ -67,12 +117,12 @@ fn reduce(
 
   if (i == 0u) { result[0] = tile[0]; }
 }
-`});var Mr,Dr=W(()=>{Mr=`// sum reduction (f64, double-double): collapses 2*WGS partial (hi, lo) pairs
+`});var ye,ve=O(()=>{ye=`// 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
-// exactly 2*WGS entries each. Concatenated after f64/dekker.wgsl for
-// DD/ddAddProtected (see it for why plain ddAdd isn't safe).
+// exactly 2*WGS entries each. Concatenated after f64/dekker.wgsl (DD struct)
+// and f64/utils/add.wgsl (ddAddProtected \u2014 see it for why plain ddAdd isn't safe).
 
 @group(0) @binding(0) var<storage, read>       partialsHi: array<f32>;
 @group(0) @binding(1) var<storage, read>       partialsLo: array<f32>;
@@ -107,7 +157,7 @@ fn reduce_f64(
     resultLo[0] = tile[0].lo;
   }
 }
-`});var Ur,Tr=W(()=>{Ur=`// sscal: x = alpha * x
+`});var Be,_e=O(()=>{Be=`// sscal: x = alpha * x
 
 @group(0) @binding(0) var<storage, read_write> x: array<f32>;
 
@@ -130,7 +180,7 @@ fn main(
     x[id * params.x_inc] = params.alpha * x[id * params.x_inc];
   }
 }
-`});var Vr,Hr=W(()=>{Vr=`// sswap: x <-> y
+`});var Ae,Ee=O(()=>{Ae=`// 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>;
@@ -156,7 +206,7 @@ fn main(
     y[id * params.y_inc] = temp;
   }
 }
-`});var Cr,Rr=W(()=>{Cr=`// saxpy: y = alpha * x + y
+`});var ke,Ge=O(()=>{ke=`// 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>;
@@ -181,7 +231,7 @@ fn main(
     y[id * params.y_inc] = params.alpha * x[id * params.x_inc] + y[id * params.y_inc];
   }
 }
-`});var zr,Or=W(()=>{zr=`// scopy: y = x
+`});var Ne,Pe=O(()=>{Ne=`// scopy: y = x
 
 @group(0) @binding(0) var<storage, read>       x: array<f32>;
 @group(0) @binding(1) var<storage, read_write> y: array<f32>;
@@ -205,7 +255,7 @@ fn main(
     y[id * params.y_inc] = x[id * params.x_inc];
   }
 }
-`});var qr,Qr=W(()=>{qr=`// sdot: result = sum(x[i] * y[i])
+`});var Me,Se=O(()=>{Me=`// 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>;
@@ -258,7 +308,7 @@ fn main(
 
   if (lid.x == 0u) { partials[wgid.x] = tile[0]; }
 }
-`});var Kr,Zr=W(()=>{Kr=`// sasum: result = sum(|x[i]|)
+`});var Le,Ie=O(()=>{Le=`// 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>;
@@ -309,7 +359,7 @@ fn main(
 
   if (lid.x == 0u) { partials[wgid.x] = tile[0]; }
 }
-`});var $r,Xr=W(()=>{$r=`// snrm2: result = sqrt(sum(x[i] * x[i]))
+`});var Re,De=O(()=>{Re=`// snrm2: result = sqrt(sum(x[i] * x[i]))
 // pass 1 dispatches exactly 2 * WGS workgroups; pass 2 uses reduction/sqsum.wgsl.
 
 @group(0) @binding(0) var<storage, read>       x:        array<f32>;
@@ -365,7 +415,7 @@ fn main(
 
   if (lid.x == 0u) { partials[wgid.x] = tile[0]; }
 }
-`});var Jr,Yr=W(()=>{Jr=`// srot: x = c*x + s*y,  y = -s*x + c*y
+`});var je,Te=O(()=>{je=`// 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>;
@@ -394,7 +444,7 @@ fn main(
     y[id * params.y_inc] = -params.s * xi + params.c * yi;
   }
 }
-`});var ee,re=W(()=>{ee=`// srotm: applies modified Givens rotation H to vectors x and y.
+`});var We,Ce=O(()=>{We=`// 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.
@@ -444,7 +494,7 @@ fn main(
     y[id * params.y_inc] = h21 * xi + h22 * yi;
   }
 }
-`});var ae,te=W(()=>{ae=`// isamax: returns index of element with largest absolute value
+`});var He,Fe=O(()=>{He=`// 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>;
@@ -526,7 +576,7 @@ fn main(
     partials_idx[wgid.x] = tile_idx[0];
   }
 }
-`});var ie,oe=W(()=>{ie=`// sgemv_n: y = alpha * A * x + beta * y  (A is m\xD7n row-major, no-transpose)
+`});var Oe,Ue=O(()=>{Oe=`// 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.
@@ -601,7 +651,7 @@ fn main(
     workgroupBarrier();
   }
 }
-`});var se,ne=W(()=>{se=`// sgemv_t: y = alpha * A^T * x + beta * y  (A is m\xD7n row-major, transposed)
+`});var Ke,Ve=O(()=>{Ke=`// 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)
@@ -666,7 +716,7 @@ fn main(
     y[yi] = params.alpha * (acc0 + acc1 + acc2 + acc3) + params.beta * y[yi];
   }
 }
-`});var le,ue=W(()=>{le=`// ssymv: y = alpha * A * x + beta * y
+`});var qe,ze=O(()=>{qe=`// 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
@@ -735,7 +785,7 @@ fn main(
     }
   }
 }
-`});var ce,fe=W(()=>{ce=`// strmv: y = op(A) * x
+`});var Xe,Ye=O(()=>{Xe=`// 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.
@@ -838,7 +888,7 @@ fn main(
     }
   }
 }
-`});var de,me=W(()=>{de=`// sger: A := alpha * x * y^T + A  (rank-1 update, A is m\xD7n general/dense)
+`});var Ze,Qe=O(()=>{Ze=`// 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>;
@@ -886,7 +936,7 @@ fn main(
     }
   }
 }
-`});var ge,pe=W(()=>{ge=`// ssyr: A := alpha * x * x^T + A  (symmetric rank-1 update)
+`});var Je,$e=O(()=>{Je=`// 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).
 
@@ -946,7 +996,7 @@ fn main(
     }
   }
 }
-`});var be,we=W(()=>{be=`// ssyr2: A := alpha * x * y^T + alpha * y * x^T + A  (symmetric rank-2 update)
+`});var et,rt=O(()=>{et=`// 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).
 
@@ -1009,7 +1059,7 @@ fn main(
     }
   }
 }
-`});var xe,he=W(()=>{xe=`// f64add: adds two doubles, each packed as a [main, aux] pair (main: f32
+`});var ot,tt=O(()=>{ot=`// f64add: adds two doubles, each packed as a [main, aux] pair (main: f32
 // value, aux: raw u32 bits \u2014 see src/util/f64pack.mjs; decode()/encode()
 // below are the WGSL mirror of that file's packedToFields()/fieldsToPacked()),
 // producing the sum as another [main, aux] pair.
@@ -1290,7 +1340,7 @@ fn computeSum(a: Fields, b: Fields) -> Packed {
   let f = addFields(a, b);
   return encode(f.sign, f.rawExp, f.mantissaHi, f.lo);
 }
-`});var ye,ve=W(()=>{ye=`// Double-double arithmetic via Dekker's algorithm \u2014 an alternative to
+`});var it,at=O(()=>{it=`// 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
@@ -1299,11 +1349,16 @@ fn computeSum(a: Fields, b: Fields) -> Packed {
 //
 // No bindings, no entry point \u2014 a helper library, concatenated with a
 // consumer's own bindings/entry point by getPipeline (WGSL has no #include).
+// The DD struct lives here \u2014 abs.wgsl/add.wgsl/greater.wgsl/equal.wgsl all
+// use it but don't redefine it (WGSL errors on duplicate struct definitions
+// once concatenated), so any consumer using those must concatenate this
+// file too, first.
 
 struct DD {
   hi: f32,
   lo: f32,
 }
+`});var nt,st=O(()=>{nt=`// 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.
@@ -1313,6 +1368,7 @@ fn ddAbs(a: DD) -> DD {
   }
   return a;
 }
+`});var lt,ut=O(()=>{lt=`// 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
 //
@@ -1389,9 +1445,29 @@ fn ddAddProtected(a: DD, b: DD, threadSlot: u32) -> DD {
   let loSum = a.lo + b.lo;
   return fastTwoSumProtected(s.hi, s.lo + loSum, threadSlot);
 }
-`});var Ee,_e=W(()=>{Ee=`// dasum: sum(|x[i]|), double-double (Dekker). Same ILP=4 shape as sasum.wgsl;
-// see f64/dekker.wgsl for ddAddProtected and why plain ddAdd isn't safe.
-// GpuVector input isn't pre-abs'd, so ddAbs() applies unconditionally below.
+`});var mt,ft=O(()=>{mt=`// 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.
+// A plain comparison, not a rounding-identity subtraction \u2014 no reassociation
+// risk, so unlike twoSum/fastTwoSum this needs no protection.
+fn ddGreater(a: DD, b: DD) -> bool {
+  if (a.hi != b.hi) {
+    return a.hi > b.hi;
+  }
+  return a.lo > b.lo;
+}
+`});var dt,ct=O(()=>{dt=`// 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 wt,pt=O(()=>{wt=`// 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.
 
 @group(0) @binding(0) var<storage, read>       xHi:        array<f32>;
 @group(0) @binding(1) var<storage, read>       xLo:        array<f32>;
@@ -1473,7 +1549,103 @@ fn dasum_main(
     partialsLo[wgid.x] = tile[0].lo;
   }
 }
-`});var Ae,Ge=W(()=>{Ae=`// strsv_invert_block: computes ONE column (workgroup_id.x) of ONE block's
+`});var bt,gt=O(()=>{bt=`// 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).
+
+@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> partialsValHi:  array<f32>;
+@group(0) @binding(3) var<storage, read_write> partialsValLo:  array<f32>;
+@group(0) @binding(4) var<storage, read_write> partialsIdx:    array<u32>;
+@group(0) @binding(5) var<uniform>             params:         Params;
+
+struct Params {
+  n:     u32,
+  x_inc: u32,
+}
+
+const WGS: u32 = 64;
+
+var<workgroup> tile_val: array<DD, 64>;
+var<workgroup> tile_idx: array<u32, 64>;
+
+@compute @workgroup_size(64)
+fn idamax_main(
+  @builtin(global_invocation_id) gid:    vec3u,
+  @builtin(local_invocation_id)  lid:    vec3u,
+  @builtin(workgroup_id)         wgid:   vec3u,
+  @builtin(num_workgroups)       num_wg: vec3u,
+) {
+  // DD(-1.0, 0.0) is a safe sentinel: any |x[i]| >= 0 beats it,
+  // so workgroups with no elements lose gracefully in the epilogue.
+  var best_val0 = DD(-1.0, 0.0); var best_idx0: u32 = 0u;
+  var best_val1 = DD(-1.0, 0.0); var best_idx1: u32 = 0u;
+  var best_val2 = DD(-1.0, 0.0); var best_idx2: u32 = 0u;
+  var best_val3 = DD(-1.0, 0.0); var best_idx3: u32 = 0u;
+
+  let stride   = num_wg.x * WGS;
+  let n4_floor = (params.n / (4u * stride)) * (4u * stride);
+
+  for (var id = gid.x; id < n4_floor; id += 4u * stride) {
+    let i0 =  id                * params.x_inc;
+    let i1 = (id +      stride) * params.x_inc;
+    let i2 = (id + 2u * stride) * params.x_inc;
+    let i3 = (id + 3u * stride) * params.x_inc;
+    let v0 = ddAbs(DD(xHi[i0], xLo[i0]));
+    let v1 = ddAbs(DD(xHi[i1], xLo[i1]));
+    let v2 = ddAbs(DD(xHi[i2], xLo[i2]));
+    let v3 = ddAbs(DD(xHi[i3], xLo[i3]));
+    if (ddGreater(v0, best_val0)) { best_val0 = v0; best_idx0 = id; }
+    if (ddGreater(v1, best_val1)) { best_val1 = v1; best_idx1 = id +      stride; }
+    if (ddGreater(v2, best_val2)) { best_val2 = v2; best_idx2 = id + 2u * stride; }
+    if (ddGreater(v3, best_val3)) { best_val3 = v3; best_idx3 = id + 3u * stride; }
+  }
+  for (var id = n4_floor + gid.x; id < params.n; id += stride) {
+    let i = id * params.x_inc;
+    let v = ddAbs(DD(xHi[i], xLo[i]));
+    if (ddGreater(v, best_val0)) { best_val0 = v; best_idx0 = id; }
+  }
+
+  // merge 4 independent lanes; prefer lower index on tie (first occurrence wins)
+  if (ddGreater(best_val1, best_val0) ||
+      (ddEqual(best_val1, best_val0) && best_idx1 < best_idx0)) {
+    best_val0 = best_val1; best_idx0 = best_idx1;
+  }
+  if (ddGreater(best_val2, best_val0) ||
+      (ddEqual(best_val2, best_val0) && best_idx2 < best_idx0)) {
+    best_val0 = best_val2; best_idx0 = best_idx2;
+  }
+  if (ddGreater(best_val3, best_val0) ||
+      (ddEqual(best_val3, best_val0) && best_idx3 < best_idx0)) {
+    best_val0 = best_val3; best_idx0 = best_idx3;
+  }
+
+  tile_val[lid.x] = best_val0;
+  tile_idx[lid.x] = best_idx0;
+  workgroupBarrier();
+
+  for (var s = WGS / 2u; s > 0u; s >>= 1u) {
+    if (lid.x < s) {
+      let a_val = tile_val[lid.x];
+      let b_val = tile_val[lid.x + s];
+      if (ddGreater(b_val, a_val) ||
+          (ddEqual(b_val, a_val) && tile_idx[lid.x + s] < tile_idx[lid.x])) {
+        tile_val[lid.x] = b_val;
+        tile_idx[lid.x] = tile_idx[lid.x + s];
+      }
+    }
+    workgroupBarrier();
+  }
+
+  if (lid.x == 0u) {
+    partialsValHi[wgid.x] = tile_val[0].hi;
+    partialsValLo[wgid.x] = tile_val[0].lo;
+    partialsIdx[wgid.x]   = tile_idx[0];
+  }
+}
+`});var xt,ht=O(()=>{xt=`// 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
@@ -1582,7 +1754,7 @@ fn strsv_invert_block_main(
     workgroupBarrier();
   }
 }
-`});var ke,Be=W(()=>{ke=`// strsv_apply_inverse: given a precomputed block inverse (from
+`});var yt,vt=O(()=>{yt=`// 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,
@@ -1629,7 +1801,7 @@ fn strsv_apply_inverse_main(@builtin(local_invocation_id) lid: vec3u) {
   }
   x[(params.blockStart + lid.x) * params.incx] = acc;
 }
-`});var Pe,Se=W(()=>{Pe=`// strsv_update: subtracts a solved block's contribution from every
+`});var Bt,_t=O(()=>{Bt=`// 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.
@@ -1704,7 +1876,580 @@ fn strsv_update_main(
     workgroupBarrier();
   }
 }
-`});var Le={};vr(Le,{shaderSources:()=>Ct});var Ct,je=W(()=>{Fr();Nr();Dr();Tr();Hr();Rr();Or();Qr();Zr();Xr();Yr();re();te();oe();ne();ue();fe();me();pe();we();he();ve();_e();Ge();Be();Se();Ct={"reduction/argmax":Ir,"reduction/sum":Wr,"reduction/sumF64":Mr,sscal:Ur,sswap:Vr,saxpy:Cr,scopy:zr,sdot:qr,sasum:Kr,snrm2:$r,srot:Jr,srotm:ee,isamax:ae,sgemv_n:ie,sgemv_t:se,ssymv:le,strmv:ce,sger:de,ssyr:ge,ssyr2:be,f64add:xe,"f64/dekker":ye,dasum:Ee,strsv_invert_block:Ae,strsv_apply_inverse:ke,strsv_update:Pe}});var Zt={};vr(Zt,{GpuMatrix:()=>V,GpuVector:()=>x,cleanup:()=>kr,dasum:()=>Ve,gpuName:()=>Sr,init:()=>Br,isamax:()=>Oe,randomFloat32Array:()=>Pr,randomFloat64Array:()=>Lr,randomTriangularFloat32Array:()=>jr,sasum:()=>He,saxpy:()=>We,scopy:()=>De,sdot:()=>Te,sgemv:()=>qe,sger:()=>Je,snrm2:()=>Ce,srot:()=>ze,srotm:()=>Qe,sscal:()=>Ie,sswap:()=>Ne,ssymv:()=>Ze,ssyr:()=>rt,ssyr2:()=>et,strmv:()=>Ke,strsv:()=>Ye});function _r(a,e){return e?a.features.has("timestamp-query")?{requiredFeatures:["timestamp-query"]}:(console.warn("timestamp-query not supported on this device \u2014 benchmark mode disabled."),{}):{}}function Er(){if(!Gr())return{querySet:null,passDescriptor:void 0};let e=U().createQuerySet({type:"timestamp",count:2});return{querySet:e,passDescriptor:{timestampWrites:{querySet:e,beginningOfPassWriteIndex:0,endOfPassWriteIndex:1}}}}function cr(a,e){if(!e)return null;let r=U(),o=r.createBuffer({label:"timestamp-resolve",size:16,usage:GPUBufferUsage.QUERY_RESOLVE|GPUBufferUsage.COPY_SRC});a.resolveQuerySet(e,0,2,o,0);let t=r.createBuffer({label:"timestamp-readback",size:16,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});return a.copyBufferToBuffer(o,0,t,0,16),{tsReadBuffer:t,resolveBuffer:o,querySet:e}}async function P(a){if(!a)return;let{tsReadBuffer:e,resolveBuffer:r,querySet:o}=a;await e.mapAsync(GPUMapMode.READ);let t=new BigInt64Array(e.getMappedRange().slice());return e.unmap(),e.destroy(),r.destroy(),o.destroy(),Math.max(0,Number(t[1]-t[0]))/1e6}var K=null,J=null,Ar=null,mr=!1;async function Br({powerPreference:a="high-performance",benchmark:e=!1}={}){if(K)return K;let r;if(typeof window>"u"){let{create:i,globals:n}=await import("webgpu");Object.assign(globalThis,n),r=i([]),Ar=r}else r=navigator.gpu;if(!r)throw new Error("WebGPU not supported in this environment.");if(J=await r.requestAdapter({powerPreference:a})??await r.requestAdapter(),!J)throw new Error("No WebGPU adapter found.");mr=e;let t=[..._r(J,e).requiredFeatures??[]];return K=await J.requestDevice({requiredFeatures:t}),K.addEventListener("uncapturederror",i=>{console.error("Uncaptured GPU error:",i.error.message)}),K}function kr(){K&&(K.destroy(),K=null),J=null,Ar=null,mr=!1}function Sr(){if(!J)throw new Error("WebGPU adapter not initialized \u2014 call init() first.");let{device:a,description:e}=J.info;return{description:e||"unknown",device:a||"unknown"}}function Gr(){return mr}function U(){if(!K)throw new Error("WebGPU device not initialized \u2014 call init() first.");return K}function m(...a){a.flat().forEach(e=>e.destroy())}function b(a,e="blas-input",r=!1){let o=U(),t=o.limits.maxStorageBufferBindingSize,i=a.byteLength;if(i>t)throw new Error(`Buffer size ${i} bytes exceeds device limit of ${t} bytes.`);let n=r?GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC:GPUBufferUsage.STORAGE,u=o.createBuffer({label:e,size:i,usage:n,mappedAtCreation:!0}),s=a.constructor;return new s(u.getMappedRange()).set(a),u.unmap(),u}function C(a,e="blas-storage"){return U().createBuffer({label:e,size:a,usage:GPUBufferUsage.STORAGE})}function q(a,e="blas-result"){return U().createBuffer({label:e,size:a,usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC})}function _(a,e){let o=U().createBuffer({label:"blas-readback",size:e.size,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});return a.copyBufferToBuffer(e,0,o,0,e.size),o}function I(a,e="blas-params"){let r=U(),o=a.length*4,t=Math.ceil(o/16)*16,i=new ArrayBuffer(t),n=new DataView(i);a.forEach(({value:s,type:l},f)=>{let c=f*4;if(l==="u32")n.setUint32(c,s,!0);else if(l==="i32")n.setInt32(c,s,!0);else if(l==="f32")n.setFloat32(c,s,!0);else throw new Error(`Unknown param type "${l}". 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 y(a,e=Float32Array){try{await a.mapAsync(GPUMapMode.READ);let r=new e(a.getMappedRange().slice());return a.unmap(),r}finally{a.destroy()}}function er(a){let e=a.length,r=new Float32Array(e),o=new Float32Array(e);for(let t=0;t<e;t++){let i=Math.fround(a[t]);r[t]=i,o[t]=Math.fround(a[t]-i)}return{hi:r,lo:o}}function tr(a,e){let r=a.length,o=new Float64Array(r);for(let t=0;t<r;t++)o[t]=a[t]+e[t];return o}var x=class a{constructor(e,r,o=Float32Array,t=null){this._buf=e,this._loBuf=t,this.length=r,this.dtype=o}static from(e){if(e instanceof Float64Array){let{hi:o,lo:t}=er(e),i=b(o,"gpu-vector-f64-hi",!0),n=b(t,"gpu-vector-f64-lo",!0);return new a(i,e.length,Float64Array,n)}if(!(e instanceof Float32Array))throw new Error("GpuVector.from expects a Float32Array or Float64Array.");let r=b(e,"gpu-vector",!0);return new a(r,e.length,e.constructor)}async read(){let e=U(),r=e.createCommandEncoder(),o=_(r,this._buf);if(e.queue.submit([r.finish()]),!this._loBuf)return y(o,this.dtype);let t=e.createCommandEncoder(),i=_(t,this._loBuf);e.queue.submit([t.finish()]);let[n,u]=await Promise.all([y(o,Float32Array),y(i,Float32Array)]);return tr(n,u)}destroy(){this._buf.destroy(),this._loBuf&&this._loBuf.destroy()}};var V=class a{constructor(e,r,o,t,i=null,n="row-major"){this._buf=e,this._loBuf=i,this.rows=r,this.cols=o,this.lda=t,this.layout=n}static from(e,r,o,t,i="row-major"){if(i!=="row-major"&&i!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");let n=i==="row-major";if(t===void 0&&(t=n?o:r),!(e instanceof Float32Array)&&!(e instanceof Float64Array))throw new Error("GpuMatrix.from expects a Float32Array or Float64Array.");if(!Number.isInteger(r)||r<=0)throw new Error("rows must be a positive integer.");if(!Number.isInteger(o)||o<=0)throw new Error("cols must be a positive integer.");let u=n?o:r;if(!Number.isInteger(t)||t<u)throw new Error(`lda must be an integer >= ${n?"cols":"rows"}.`);let s=n?r:o;if(e.length<s*t)throw new Error("data does not have enough elements for the given rows, cols, and lda.");if(e instanceof Float64Array){let f=s*t,{hi:c,lo:p}=er(e.subarray(0,f)),d=b(c,"gpu-matrix-f64-hi",!0),g=b(p,"gpu-matrix-f64-lo",!0);return new a(d,r,o,t,g,i)}let l=b(e.subarray(0,s*t),"gpu-matrix",!0);return new a(l,r,o,t,null,i)}async read(){let e=U(),r=e.createCommandEncoder(),o=_(r,this._buf);e.queue.submit([r.finish()]);let t=this.layout!=="column-major",i=t?this.rows:this.cols,n=t?this.cols:this.rows;if(this._loBuf){let l=e.createCommandEncoder(),f=_(l,this._loBuf);e.queue.submit([l.finish()]);let[c,p]=await Promise.all([y(o,Float32Array),y(f,Float32Array)]),d=tr(c,p);if(this.lda===n)return d;let g=new Float64Array(i*n);for(let w=0;w<i;w++)g.set(d.subarray(w*this.lda,w*this.lda+n),w*n);return g}let u=await y(o,Float32Array);if(this.lda===n)return u;let s=new Float32Array(i*n);for(let l=0;l<i;l++)s.set(u.subarray(l*this.lda,l*this.lda+n),l*n);return s}destroy(){this._buf.destroy(),this._loBuf&&this._loBuf.destroy()}};function Pr(a,e=-1,r=1){let o=new Float32Array(a);for(let t=0;t<a;t++)o[t]=e+Math.random()*(r-e);return o}function Lr(a,e=-1,r=1){let o=new Float64Array(a);for(let t=0;t<a;t++)o[t]=e+Math.random()*(r-e);return o}function jr(a,e,r="lower",o=-1,t=1,i=5,n=15){if(r!=="lower"&&r!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(e<a)throw new Error("lda must be >= n.");let u=new Float32Array(a*e);for(let s=0;s<a;s++){for(let l=0;l<a;l++){if(s===l)continue;(r==="lower"?l<s:l>s)&&(u[s*e+l]=o+Math.random()*(t-o))}u[s*e+s]=i+Math.random()*(n-i)}return u}function A(a,e,r=0){let o=U(),t=e.map((i,n)=>({binding:r+n,resource:i instanceof GPUBuffer?{buffer:i}:i}));return o.createBindGroup({layout:a,entries:t})}var wt=new WeakMap;function L(a){U().queue.submit([a.finish()])}function dr(){let a=U(),{querySet:e,passDescriptor:r}=Er();return{commandEncoder:a.createCommandEncoder(),querySet:e,passDescriptor:r}}function ir(a,e,r,o,t){let i=a.beginComputePass(t);i.setPipeline(e),i.setBindGroup(0,r),typeof o=="number"?i.dispatchWorkgroups(o):i.dispatchWorkgroups(o.x,o.y),i.end(),wt.set(a,i)}function F(a,e,r){let{commandEncoder:o,querySet:t,passDescriptor:i}=dr();ir(o,a,e,r,i);let n=cr(o,t);return{commandEncoder:o,ts:n}}var Qt={},pr=new WeakMap;async function B(a,e,r="main"){pr.has(a)||pr.set(a,new Map);let o=pr.get(a),t=Array.isArray(e)?e:[e],i=`${t.join("+")}::${r}`;return o.has(i)||o.set(i,await zt(t,r)),o.get(i)}async function Ot(a){if(typeof process>"u"||!process.versions?.node){let{shaderSources:e}=await Promise.resolve().then(()=>(je(),Le)),r=e[a];if(!r)throw new Error(`Shader "${a}" not found in browser bundle.`);return r}else{let{readFileSync:e}=await import("fs"),{fileURLToPath:r}=await import("url"),{dirname:o,join:t}=await import("path"),i=o(r(Qt.url));return e(t(i,`../shaders/${a}.wgsl`),"utf8")}}async function zt(a,e="main"){let r=U(),o=a.join("+"),t=(await Promise.all(a.map(Ot))).join(`
-`),i=r.createShaderModule({label:o,code:t}),u=(await i.getCompilationInfo()).messages.filter(f=>f.type==="error");if(u.length>0)throw new Error(`Shader "${o}" compilation failed:
-${u.map(f=>`  line ${f.lineNum}: ${f.message}`).join(`
-`)}`);let s=e==="main"?{module:i}:{module:i,entryPoint:e},l=r.createComputePipeline({label:o,layout:"auto",compute:s});return l._shaderModule=i,l}var qt=64,Fe=8;function O(a,e){let r=U().limits.maxComputeWorkgroupsPerDimension;return e===void 0?Math.min(Math.ceil(a/qt),r):{x:Math.min(Math.ceil(e/Fe),r),y:Math.min(Math.ceil(a/Fe),r)}}async function Ie(a,e,r,o,t){let i=o instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!Number.isInteger(t))throw new Error("n and incx must be integers.");if(typeof r!="number")throw new Error("alpha must be a number.");if(Number.isNaN(r))throw new Error("alpha must not be NaN.");if(!Number.isFinite(r))throw new Error("alpha must be finite.");if(t<=0)throw new Error("incx must be positive.");if(!(o instanceof Float32Array)&&!(o instanceof x))throw new Error("x must be a Float32Array or GpuVector.");if(e<=0)return i?{}:o;if(o.length<(e-1)*t+1)throw new Error("x does not have enough elements for the given n and incx.");let n=await B(a,"sscal"),u=null,s=null,l=null;try{u=i?o._buf:b(o,"sscal-x",!0),s=I([{value:e,type:"u32"},{value:r,type:"f32"},{value:t,type:"u32"}],"sscal-params");let f=A(n.getBindGroupLayout(0),[u,s]),{commandEncoder:c,ts:p}=F(n,f,O(e));l=i?null:_(c,u),L(c);let d=await P(p);if(i)return d!==void 0?{gpuTimeMs:d}:{};let g=await y(l,Float32Array);return l=null,d!==void 0?{x:g,gpuTimeMs:d}:g}finally{!i&&u&&m(u),s&&m(s),l&&m(l)}}async function Ne(a,e,r,o,t,i){let n=r instanceof x,u=t instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!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(!(r instanceof Float32Array)&&!(r instanceof x))throw new Error("x must be a Float32Array or GpuVector.");if(!(t instanceof Float32Array)&&!(t instanceof x))throw new Error("y must be a Float32Array or GpuVector.");if(r.constructor!==t.constructor)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(e<=0)return n?{}:{x:r,y:t};if(r.length<(e-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(e-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let s=await B(a,"sswap"),l=null,f=null,c=null,p=null,d=null;try{l=n?r._buf:b(r,"sswap-x",!0),f=u?t._buf:b(t,"sswap-y",!0),c=I([{value:e,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"sswap-params");let g=A(s.getBindGroupLayout(0),[l,f,c]),{commandEncoder:w,ts:E}=F(s,g,O(e));p=n?null:_(w,l),d=u?null:_(w,f),L(w);let h=await P(E);if(n&&u)return h!==void 0?{gpuTimeMs:h}:{};let G=await y(p,Float32Array);p=null;let v=await y(d,Float32Array);return d=null,h!==void 0?{x:G,y:v,gpuTimeMs:h}:{x:G,y:v}}finally{!n&&l&&m(l),!u&&f&&m(f),c&&m(c),p&&m(p),d&&m(d)}}async function We(a,e,r,o,t,i,n){let u=o instanceof x,s=i instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!Number.isInteger(t)||!Number.isInteger(n))throw new Error("n, incx, and incy must be integers.");if(typeof r!="number")throw new Error("alpha must be a number.");if(Number.isNaN(r))throw new Error("alpha must not be NaN.");if(!Number.isFinite(r))throw new Error("alpha must be finite.");if(t<=0||n<=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(!s&&!(i instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(u!==s)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(e<=0)return s?{}:{y:i};if(o.length<(e-1)*t+1)throw new Error("x does not have enough elements for the given n and incx.");if(i.length<(e-1)*n+1)throw new Error("y does not have enough elements for the given n and incy.");let l=await B(a,"saxpy"),f=null,c=null,p=null,d=null;try{f=u?o._buf:b(o,"saxpy-x",!1),c=s?i._buf:b(i,"saxpy-y",!0),p=I([{value:e,type:"u32"},{value:r,type:"f32"},{value:t,type:"u32"},{value:n,type:"u32"}],"saxpy-params");let g=A(l.getBindGroupLayout(0),[f,c,p]),{commandEncoder:w,ts:E}=F(l,g,O(e));d=s?null:_(w,c),L(w);let h=await P(E);if(s&&u)return h!==void 0?{gpuTimeMs:h}:{};let G=await y(d,Float32Array);return d=null,h!==void 0?{y:G,gpuTimeMs:h}:{y:G}}finally{!u&&f&&m(f),!s&&c&&m(c),p&&m(p),d&&m(d)}}async function De(a,e,r,o,t,i){let n=r instanceof x,u=t instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!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(!n&&!(r 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(n!==u)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(e<=0)return u?{}:{y:t};if(r.length<(e-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(e-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let s=await B(a,"scopy"),l=null,f=null,c=null,p=null;try{l=n?r._buf:b(r,"scopy-x",!1),f=u?t._buf:b(t,"scopy-y",!0),c=I([{value:e,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"scopy-params");let d=A(s.getBindGroupLayout(0),[l,f,c]),{commandEncoder:g,ts:w}=F(s,d,O(e));p=u?null:_(g,f),L(g);let E=await P(w);if(u&&n)return E!==void 0?{gpuTimeMs:E}:{};let h=await y(p,Float32Array);return p=null,E!==void 0?{y:h,gpuTimeMs:E}:{y:h}}finally{!n&&l&&m(l),!u&&f&&m(f),c&&m(c),p&&m(p)}}var Me=64;async function Te(a,e,r,o,t,i){let n=r instanceof x,u=t instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!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(!n&&!(r 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(n!==u)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(e<=0)return{dot:0};if(r.length<(e-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(e-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let s=await B(a,"sdot"),l=await B(a,"reduction/sum"),f=null,c=null,p=null,d=null,g=null,w=null;try{f=n?r._buf:b(r,"sdot-x",!1),c=u?t._buf:b(t,"sdot-y",!1),p=C(2*Me*4,"sdot-partials"),d=q(4,"sdot-result"),g=I([{value:e,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"sdot-params");let E=A(s.getBindGroupLayout(0),[f,c,p,g]),{commandEncoder:h,ts:G}=F(s,E,2*Me);L(h);let v=A(l.getBindGroupLayout(0),[p,d]),{commandEncoder:k,ts:S}=F(l,v,1);w=_(k,d),L(k);let j=y(w,Float32Array);w=null;let[N,D,M]=await Promise.all([P(G),P(S),j]);return N!==void 0&&D!==void 0?{dot:M[0],gpuTimeMs:N+D}:{dot:M[0]}}finally{!n&&f&&m(f),!u&&c&&m(c),p&&m(p),d&&m(d),g&&m(g),w&&m(w)}}var Ue=64;async function He(a,e,r,o){let t=r instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!t&&!(r instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(e<=0)return{asum:0};if(r.length<(e-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=await B(a,"sasum"),n=await B(a,"reduction/sum"),u=null,s=null,l=null,f=null,c=null;try{u=t?r._buf:b(r,"sasum-x",!1),s=C(2*Ue*4,"sasum-partials"),l=q(4,"sasum-result"),f=I([{value:e,type:"u32"},{value:o,type:"u32"}],"sasum-params");let p=A(i.getBindGroupLayout(0),[u,s,f]),{commandEncoder:d,ts:g}=F(i,p,2*Ue);L(d);let w=A(n.getBindGroupLayout(0),[s,l]),{commandEncoder:E,ts:h}=F(n,w,1);c=_(E,l),L(E);let G=y(c,Float32Array);c=null;let[v,k,S]=await Promise.all([P(g),P(h),G]);return v!==void 0&&k!==void 0?{asum:S[0],gpuTimeMs:v+k}:{asum:S[0]}}finally{!t&&u&&m(u),s&&m(s),l&&m(l),f&&m(f),c&&m(c)}}var gr=64;async function Ve(a,e,r,o){let t=r instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!t&&!(r instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(t&&r.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(e<=0)return{asum:0};if(r.length<(e-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=await B(a,["f64/dekker","dasum"]),n=await B(a,["f64/dekker","reduction/sumF64"]),u=null,s=null,l=null,f=null,c=null,p=null,d=null,g=null,w=null;try{if(t)u=r._buf,s=r._loBuf;else{let{hi:Z,lo:z}=er(r.map(Math.abs));u=b(Z,"dasum-xHi",!1),s=b(z,"dasum-xLo",!1)}l=C(2*gr*4,"dasum-partialsHi"),f=C(2*gr*4,"dasum-partialsLo"),c=q(4,"dasum-result-hi"),p=q(4,"dasum-result-lo"),d=I([{value:e,type:"u32"},{value:o,type:"u32"}],"dasum-params");let E=A(i.getBindGroupLayout(0),[u,s,l,f,d]),{commandEncoder:h,ts:G}=F(i,E,2*gr);L(h);let v=A(n.getBindGroupLayout(0),[l,f,c,p]),{commandEncoder:k,ts:S}=F(n,v,1);g=_(k,c),w=_(k,p),L(k);let j=y(g,Float32Array),N=y(w,Float32Array);g=null,w=null;let[D,M,T,H]=await Promise.all([P(G),P(S),j,N]),R=tr(T,H)[0];return D!==void 0&&M!==void 0?{asum:R,gpuTimeMs:D+M}:{asum:R}}finally{!t&&u&&m(u),!t&&s&&m(s),l&&m(l),f&&m(f),c&&m(c),p&&m(p),d&&m(d),g&&m(g),w&&m(w)}}var Re=64;async function Ce(a,e,r,o){let t=r instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!t&&!(r instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(e<=0)return{nrm2:0};if(r.length<(e-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=await B(a,"snrm2"),n=await B(a,"reduction/sum"),u=null,s=null,l=null,f=null,c=null;try{u=t?r._buf:b(r,"snrm2-x",!1),s=C(2*Re*4,"snrm2-partials"),l=q(4,"snrm2-result"),f=I([{value:e,type:"u32"},{value:o,type:"u32"}],"snrm2-params");let p=A(i.getBindGroupLayout(0),[u,s,f]),{commandEncoder:d,ts:g}=F(i,p,2*Re);L(d);let w=A(n.getBindGroupLayout(0),[s,l]),{commandEncoder:E,ts:h}=F(n,w,1);c=_(E,l),L(E);let G=y(c,Float32Array);c=null;let[v,k,S]=await Promise.all([P(g),P(h),G]),j=Math.sqrt(S[0]);return v!==void 0&&k!==void 0?{nrm2:j,gpuTimeMs:v+k}:{nrm2:j}}finally{!t&&u&&m(u),s&&m(s),l&&m(l),f&&m(f),c&&m(c)}}var wr=64;async function Oe(a,e,r,o){let t=r instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!t&&!(r instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(e<=0)return{index:0};if(r.length<(e-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=await B(a,"isamax"),n=await B(a,"reduction/argmax"),u=null,s=null,l=null,f=null,c=null,p=null;try{u=t?r._buf:b(r,"isamax-x",!1),s=C(2*wr*4,"isamax-partials-val"),l=C(2*wr*4,"isamax-partials-idx"),f=q(4,"isamax-result"),c=I([{value:e,type:"u32"},{value:o,type:"u32"}],"isamax-params");let d=A(i.getBindGroupLayout(0),[u,s,l,c]),{commandEncoder:g,ts:w}=F(i,d,2*wr);L(g);let E=A(n.getBindGroupLayout(0),[s,l,f]),{commandEncoder:h,ts:G}=F(n,E,1);p=_(h,f),L(h);let v=y(p,Uint32Array);p=null;let[k,S,j]=await Promise.all([P(w),P(G),v]),N=j[0];return k!==void 0&&S!==void 0?{index:N,gpuTimeMs:k+S}:{index:N}}finally{!t&&u&&m(u),s&&m(s),l&&m(l),f&&m(f),c&&m(c),p&&m(p)}}async function ze(a,e,r,o,t,i,n,u){let s=r instanceof x,l=t instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(typeof n!="number")throw new Error("c must be a number.");if(typeof u!="number")throw new Error("s must be a number.");if(Number.isNaN(n)||Number.isNaN(u))throw new Error("c and s must not be NaN.");if(!Number.isFinite(n))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(!s&&!(r instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!l&&!(t instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(s!==l)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(e<=0)return s?{}:{x:r,y:t};if(r.length<(e-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(e-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let f=await B(a,"srot"),c=null,p=null,d=null,g=null,w=null;try{c=s?r._buf:b(r,"srot-x",!0),p=l?t._buf:b(t,"srot-y",!0),d=I([{value:e,type:"u32"},{value:n,type:"f32"},{value:u,type:"f32"},{value:o,type:"u32"},{value:i,type:"u32"}],"srot-params");let E=A(f.getBindGroupLayout(0),[c,p,d]),{commandEncoder:h,ts:G}=F(f,E,O(e));g=s?null:_(h,c),w=l?null:_(h,p),L(h);let v=await P(G);if(s&&l)return v!==void 0?{gpuTimeMs:v}:{};let k=y(g,Float32Array),S=y(w,Float32Array);g=null,w=null;let[j,N]=await Promise.all([k,S]);return v!==void 0?{x:j,y:N,gpuTimeMs:v}:{x:j,y:N}}finally{!s&&c&&m(c),!l&&p&&m(p),d&&m(d),g&&m(g),w&&m(w)}}async function Qe(a,e,r,o,t,i,n){let u=r instanceof x,s=t instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(!(n instanceof Float32Array)||n.length!==5)throw new Error("param must be a Float32Array of length 5.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!u&&!(r instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!s&&!(t instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(u!==s)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(e<=0||n[0]===-2)return u?{}:{x:r,y:t};if(r.length<(e-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(e-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let l=await B(a,"srotm"),f=null,c=null,p=null,d=null,g=null,w=null;try{f=u?r._buf:b(r,"srotm-x",!0),c=s?t._buf:b(t,"srotm-y",!0),p=b(n,"srotm-param",!1),d=I([{value:e,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"srotm-params");let E=A(l.getBindGroupLayout(0),[f,c,p,d]),{commandEncoder:h,ts:G}=F(l,E,O(e));g=u?null:_(h,f),w=s?null:_(h,c),L(h);let v=await P(G);if(u&&s)return v!==void 0?{gpuTimeMs:v}:{};let k=y(g,Float32Array),S=y(w,Float32Array);g=null,w=null;let[j,N]=await Promise.all([k,S]);return v!==void 0?{x:j,y:N,gpuTimeMs:v}:{x:j,y:N}}finally{!u&&f&&m(f),!s&&c&&m(c),p&&m(p),d&&m(d),g&&m(g),w&&m(w)}}async function qe(a,e,r,o,t,i,n,u,s,l,f,c,p="row-major"){let d=i instanceof V,g=u instanceof x,w=f instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(e!=="no-transpose"&&e!=="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 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 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(r)||!Number.isInteger(o)||!Number.isInteger(s)||!Number.isInteger(c)||!Number.isInteger(n))throw new Error("m, n, incx, incy, and lda must be integers.");if(s<=0||c<=0)throw new Error("incx and incy must be positive.");if(!d&&!(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&&!(f 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&&!d)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(g&&u._buf===f._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(d&&n!==i.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(d&&(i.rows<r||i.cols<o))throw new Error("A is too small for the given m and n.");if(r<0||o<0)throw new Error("m and n must be non-negative.");if(r===0||o===0)return w?{}:{y:f};(d?i.layout:p)==="column-major"&&([r,o]=[o,r],e=e==="no-transpose"?"transpose":"no-transpose");let h=e==="no-transpose",G=h?o:r,v=h?r:o;if(n<o)throw new Error("lda must be >= n.");if(!d&&i.length<(r-1)*n+o)throw new Error("A does not have enough elements for the given m, n, and lda.");if(u.length<(G-1)*s+1)throw new Error("x does not have enough elements for the given dimensions and incx.");if(f.length<(v-1)*c+1)throw new Error("y does not have enough elements for the given dimensions and incy.");let S=await B(a,h?"sgemv_n":"sgemv_t"),j=d?i._buf:b(i,"sgemv-A",!1),N=g?u._buf:b(u,"sgemv-x",!1),D=w?f._buf:b(f,"sgemv-y",!0),M=I([{value:r,type:"u32"},{value:o,type:"u32"},{value:t,type:"f32"},{value:l,type:"f32"},{value:s,type:"u32"},{value:c,type:"u32"},{value:n,type:"u32"}],"sgemv-params");try{let T=A(S.getBindGroupLayout(0),[j,N,D,M]),H=h?Math.min(r,a.limits.maxComputeWorkgroupsPerDimension):O(v),{commandEncoder:R,ts:Z}=F(S,T,H),z=w?null:_(R,D);L(R);let Q=await P(Z);if(w)return Q!==void 0?{gpuTimeMs:Q}:{};let nr=await y(z,Float32Array);return Q!==void 0?{y:nr,gpuTimeMs:Q}:{y:nr}}finally{d||m(j),g||m(N),w||m(D),m(M)}}async function Ze(a,e,r,o,t,i,n,u,s,l,f,c="row-major"){let p=n instanceof x,d=l instanceof x,g=t instanceof V;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");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(!Number.isInteger(r)||!Number.isInteger(u)||!Number.isInteger(f)||!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 s!="number")throw new Error("beta must be a number.");if(Number.isNaN(s))throw new Error("beta must not be NaN.");if(!Number.isFinite(s))throw new Error("beta must be finite.");if(u<=0||f<=0)throw new Error("incx and incy must be positive.");if(i<r)throw new Error("lda must be >= n.");if(!g&&!(t instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!p&&!(n instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!d&&!(l instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(p!==d)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(p&&n._buf===l._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(g&&i!==t.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(g&&(t.rows<r||t.cols<r))throw new Error("A is too small for the given n.");if(r<0)throw new Error("n must be non-negative.");if(r===0)return d?{}:{y:l};if(!g&&t.length<(r-1)*i+r)throw new Error("A does not have enough elements for the given n and lda.");if(n.length<(r-1)*u+1)throw new Error("x does not have enough elements for the given n and incx.");if(l.length<(r-1)*f+1)throw new Error("y does not have enough elements for the given n and incy.");let E=(g?t.layout:c)==="column-major"?e==="upper":e==="lower",h=await B(a,"ssymv"),G=null,v=null,k=null,S=null;try{G=g?t._buf:b(t,"ssymv-A",!1),v=p?n._buf:b(n,"ssymv-x",!1),k=d?l._buf:b(l,"ssymv-y",!0),S=I([{value:r,type:"u32"},{value:o,type:"f32"},{value:s,type:"f32"},{value:u,type:"u32"},{value:f,type:"u32"},{value:i,type:"u32"},{value:E?0:1,type:"u32"}],"ssymv-params");let j=A(h.getBindGroupLayout(0),[G,v,k,S]),N=Math.min(r,a.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:D,ts:M}=F(h,j,N),T=d?null:_(D,k);L(D);let H=await P(M);if(d)return H!==void 0?{gpuTimeMs:H}:{};let R=await y(T,Float32Array);return H!==void 0?{y:R,gpuTimeMs:H}:{y:R}}finally{!g&&G&&m(G),!p&&v&&m(v),!d&&k&&m(k),S&&m(S)}}async function Ke(a,e,r,o,t,i,n,u,s,l,f,c="row-major"){let p=u instanceof x,d=l instanceof x,g=i instanceof V,w=o==="unit";if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(e!=="lower"&&e!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(r!=="no-transpose"&&r!=="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(c!=="row-major"&&c!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(!Number.isInteger(t)||!Number.isInteger(s)||!Number.isInteger(f)||!Number.isInteger(n))throw new Error("n, incx, incy, and lda must be integers.");if(s<=0||f<=0)throw new Error("incx and incy must be positive.");if(n<t)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(!d&&!(l instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(p!==d)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(p&&u._buf===l._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&&d&&i._buf===l._buf)throw new Error("A and y must not reference the same GPU buffer.");if(g&&n!==i.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(g&&(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 d?{}:{y:l};if(!g&&i.length<(t-1)*n+t)throw new Error("A does not have enough elements for the given n and lda.");if(u.length<(t-1)*s+1)throw new Error("x does not have enough elements for the given n and incx.");if(l.length<(t-1)*f+1)throw new Error("y does not have enough elements for the given n and incy.");let h=(g?i.layout:c)==="column-major",G=h?e==="upper":e==="lower",v=h?r==="transpose":r==="no-transpose",k=await B(a,"strmv"),S=null,j=null,N=null,D=null;try{S=g?i._buf:b(i,"strmv-A",!1),j=p?u._buf:b(u,"strmv-x",!1),N=d?l._buf:b(l,"strmv-y",!0),D=I([{value:t,type:"u32"},{value:s,type:"u32"},{value:f,type:"u32"},{value:n,type:"u32"},{value:v?0:1,type:"u32"},{value:G?0:1,type:"u32"},{value:w?1:0,type:"u32"}],"strmv-params");let M=A(k.getBindGroupLayout(0),[S,j,N,D]),T=Math.min(t,a.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:H,ts:R}=F(k,M,T),Z=d?null:_(H,N);L(H);let z=await P(R);if(d)return z!==void 0?{gpuTimeMs:z}:{};let Q=await y(Z,Float32Array);return z!==void 0?{y:Q,gpuTimeMs:z}:{y:Q}}finally{!g&&S&&m(S),!p&&j&&m(j),!d&&N&&m(N),D&&m(D)}}var X=64;function Xe(a,e,r){let o=new ArrayBuffer(a*e),t=new DataView(o);for(let i=0;i<a;i++){let n=r(i),u=i*e;n.forEach((s,l)=>t.setUint32(u+l*4,s,!0))}return o}function $e(a,e,r){let o=a.createBuffer({label:r,size:e.byteLength,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});return a.queue.writeBuffer(o,0,e),o}async function Ye(a,e,r,o,t,i,n,u,s,l="row-major"){let f=u instanceof x,c=i instanceof V,p=o==="unit";if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(e!=="lower"&&e!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(r!=="no-transpose"&&r!=="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(l!=="row-major"&&l!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(!Number.isInteger(t)||!Number.isInteger(s)||!Number.isInteger(n))throw new Error("n, incx, and lda must be integers.");if(s<=0)throw new Error("incx must be positive.");if(n<t)throw new Error("lda must be >= n.");if(!c&&!(i instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!f&&!(u instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(f&&!c)throw new Error("A must be a GpuMatrix when x is a GpuVector.");if(c&&n!==i.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(c&&(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 f?{}:{x:u};if(!c&&i.length<(t-1)*n+t)throw new Error("A does not have enough elements for the given n and lda.");if(u.length<(t-1)*s+1)throw new Error("x does not have enough elements for the given n and incx.");let g=(c?i.layout:l)==="column-major",w=g?e==="upper":e==="lower",E=g?r==="transpose":r==="no-transpose",h=await B(a,"strsv_invert_block"),G=await B(a,"strsv_apply_inverse"),v=await B(a,"strsv_update"),k=E===w,S=[];for(let Q=0;Q<t;Q+=X)S.push(Q);k||S.reverse();let j=S.length,N=a.limits.maxComputeWorkgroupsPerDimension,D=a.limits.minUniformBufferOffsetAlignment,M=null,T=null,H=null,R=null,Z=null,z=null;try{M=c?i._buf:b(i,"strsv-A",!1),T=f?u._buf:b(u,"strsv-x",!0),H=C(j*X*X*4,"strsv-Ainv");let Q=Xe(j,D,$=>{let Y=$*X,or=Math.min(Y+X,t);return[s,$,Y,or]});R=$e(a,Q,"strsv-apply-params");let nr=Xe(j,D,$=>{let Y=$*X,or=Math.min(Y+X,t);return[t,s,n,E?0:1,w?0:1,Y,or]});Z=$e(a,nr,"strsv-update-params");let{commandEncoder:rr,querySet:ar}=dr();z=I([{value:t,type:"u32"},{value:n,type:"u32"},{value:E?0:1,type:"u32"},{value:w?0:1,type:"u32"},{value:p?1:0,type:"u32"}],"strsv-invert-params");let tt=A(h.getBindGroupLayout(0),[M,H,z]);ir(rr,h,tt,{x:X,y:j},ar?{timestampWrites:{querySet:ar,beginningOfPassWriteIndex:0}}:void 0);for(let $=0;$<S.length;$++){let Y=S[$],or=Math.min(Y+X,t),it=Y/X,nt=$===S.length-1,hr=it*D,st=A(G.getBindGroupLayout(0),[H,T,{buffer:R,offset:hr,size:16}]);ir(rr,G,st,1,nt&&ar?{timestampWrites:{querySet:ar,endOfPassWriteIndex:1}}:void 0);let xr=k?t-or:Y;if(xr===0)continue;let ut=A(v.getBindGroupLayout(0),[M,T,{buffer:Z,offset:hr,size:32}]),lt=Math.min(xr,N);ir(rr,v,ut,lt)}let at=cr(rr,ar),ot=f?null:_(rr,T);L(rr);let sr=await P(at);if(f)return sr!==void 0?{gpuTimeMs:sr}:{};let br=await y(ot,Float32Array);return sr!==void 0?{x:br,gpuTimeMs:sr}:{x:br}}finally{!c&&M&&m(M),!f&&T&&m(T),H&&m(H),R&&m(R),Z&&m(Z),z&&m(z)}}async function Je(a,e,r,o,t,i,n,u,s,l,f="row-major"){let c=s instanceof V;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(f!=="row-major"&&f!=="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(e)||!Number.isInteger(r)||!Number.isInteger(i)||!Number.isInteger(u)||!Number.isInteger(l))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(!c&&!(s instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(c&&l!==s.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(c&&(s.rows<e||s.cols<r))throw new Error("A is too small for the given m and n.");(c?s.layout:f)==="column-major"&&([e,r]=[r,e],[t,n]=[n,t],[i,u]=[u,i]);let d=t instanceof x,g=n instanceof x;if(l<r)throw new Error("lda must be >= n.");if(!d&&!(t instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!g&&!(n instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(d!==g)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(d&&!c)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(c&&d&&s._buf===t._buf)throw new Error("A and x must not reference the same GPU buffer.");if(c&&g&&s._buf===n._buf)throw new Error("A and y must not reference the same GPU buffer.");if(e<0||r<0)throw new Error("m and n must be non-negative.");if(e===0||r===0)return c?{}:{A:s};if(!c&&s.length<(e-1)*l+r)throw new Error("A does not have enough elements for the given m, n, and lda.");if(t.length<(e-1)*i+1)throw new Error("x does not have enough elements for the given m and incx.");if(n.length<(r-1)*u+1)throw new Error("y does not have enough elements for the given n and incy.");let w=await B(a,"sger"),E=null,h=null,G=null,v=null;try{E=d?t._buf:b(t,"sger-x",!1),h=g?n._buf:b(n,"sger-y",!1),G=c?s._buf:b(s,"sger-A",!0),v=I([{value:e,type:"u32"},{value:r,type:"u32"},{value:o,type:"f32"},{value:i,type:"u32"},{value:u,type:"u32"},{value:l,type:"u32"}],"sger-params");let k=A(w.getBindGroupLayout(0),[E,h,G,v]),S=Math.min(e,a.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:j,ts:N}=F(w,k,S),D=c?null:_(j,G);L(j);let M=await P(N);if(c)return M!==void 0?{gpuTimeMs:M}:{};let T=await y(D,Float32Array);return M!==void 0?{A:T,gpuTimeMs:M}:{A:T}}finally{!d&&E&&m(E),!g&&h&&m(h),!c&&G&&m(G),v&&m(v)}}async function rt(a,e,r,o,t,i,n,u,s="row-major"){let l=t instanceof x,f=n instanceof V;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(e!=="lower"&&e!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(s!=="row-major"&&s!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(!Number.isInteger(r)||!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<r)throw new Error("lda must be >= n.");if(!f&&!(n instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!l&&!(t instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(l&&!f)throw new Error("A must be a GpuMatrix when x is a GpuVector.");if(f&&l&&n._buf===t._buf)throw new Error("A and x must not reference the same GPU buffer.");if(f&&u!==n.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(f&&(n.rows<r||n.cols<r))throw new Error("A is too small for the given n.");if(r<0)throw new Error("n must be non-negative.");if(r===0)return f?{}:{A:n};if(!f&&n.length<(r-1)*u+r)throw new Error("A does not have enough elements for the given n and lda.");if(t.length<(r-1)*i+1)throw new Error("x does not have enough elements for the given n and incx.");let p=(f?n.layout:s)==="column-major"?e==="upper":e==="lower",d=await B(a,"ssyr"),g=null,w=null,E=null;try{g=l?t._buf:b(t,"ssyr-x",!1),w=f?n._buf:b(n,"ssyr-A",!0),E=I([{value:r,type:"u32"},{value:o,type:"f32"},{value:i,type:"u32"},{value:u,type:"u32"},{value:p?0:1,type:"u32"}],"ssyr-params");let h=A(d.getBindGroupLayout(0),[g,w,E]),G=Math.min(r,a.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:v,ts:k}=F(d,h,G),S=f?null:_(v,w);L(v);let j=await P(k);if(f)return j!==void 0?{gpuTimeMs:j}:{};let N=await y(S,Float32Array);return j!==void 0?{A:N,gpuTimeMs:j}:{A:N}}finally{!l&&g&&m(g),!f&&w&&m(w),E&&m(E)}}async function et(a,e,r,o,t,i,n,u,s,l,f="row-major"){let c=t instanceof x,p=n instanceof x,d=s instanceof V;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(e!=="lower"&&e!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(f!=="row-major"&&f!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(!Number.isInteger(r)||!Number.isInteger(i)||!Number.isInteger(u)||!Number.isInteger(l))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(l<r)throw new Error("lda must be >= n.");if(!d&&!(s instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!c&&!(t instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!p&&!(n 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&&!d)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(d&&c&&s._buf===t._buf)throw new Error("A and x must not reference the same GPU buffer.");if(d&&p&&s._buf===n._buf)throw new Error("A and y must not reference the same GPU buffer.");if(c&&t._buf===n._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(d&&l!==s.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(d&&(s.rows<r||s.cols<r))throw new Error("A is too small for the given n.");if(r<0)throw new Error("n must be non-negative.");if(r===0)return d?{}:{A:s};if(!d&&s.length<(r-1)*l+r)throw new Error("A does not have enough elements for the given n and lda.");if(t.length<(r-1)*i+1)throw new Error("x does not have enough elements for the given n and incx.");if(n.length<(r-1)*u+1)throw new Error("y does not have enough elements for the given n and incy.");let w=(d?s.layout:f)==="column-major"?e==="upper":e==="lower",E=await B(a,"ssyr2"),h=null,G=null,v=null,k=null;try{h=c?t._buf:b(t,"ssyr2-x",!1),G=p?n._buf:b(n,"ssyr2-y",!1),v=d?s._buf:b(s,"ssyr2-A",!0),k=I([{value:r,type:"u32"},{value:o,type:"f32"},{value:i,type:"u32"},{value:u,type:"u32"},{value:l,type:"u32"},{value:w?0:1,type:"u32"}],"ssyr2-params");let S=A(E.getBindGroupLayout(0),[h,G,v,k]),j=Math.min(r,a.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:N,ts:D}=F(E,S,j),M=d?null:_(N,v);L(N);let T=await P(D);if(d)return T!==void 0?{gpuTimeMs:T}:{};let H=await y(M,Float32Array);return T!==void 0?{A:H,gpuTimeMs:T}:{A:H}}finally{!c&&h&&m(h),!p&&G&&m(G),!d&&v&&m(v),k&&m(k)}}return gt(Zt);})();
+`});var At,Et=O(()=>{At=`// 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
+// workgroups to fill the GPU. Same structure as sgemm_large.wgsl (2D
+// register-blocked, shared-memory-tiled), just smaller.
+//
+// col mapped to gid.x for coalesced B/C access (row-major: col contiguous).
+
+const BM: u32 = 32u;
+const BN: u32 = 32u;
+const BK: u32 = 8u;
+const TM: u32 = 2u;
+const TN: u32 = 2u;
+const THREADS_X: u32 = BN / TN;
+const THREADS_Y: u32 = BM / TM;
+const NUM_THREADS: u32 = THREADS_X * THREADS_Y; // 256
+const STRIDE_A: u32 = NUM_THREADS / BK;
+const STRIDE_B: u32 = NUM_THREADS / BN;
+
+@group(0) @binding(0) var<storage, read>       A: array<f32>;
+@group(0) @binding(1) var<storage, read>       B: array<f32>;
+@group(0) @binding(2) var<storage, read_write> C: array<f32>;
+
+struct Params {
+  m:      u32,
+  n:      u32,
+  k:      u32,
+  alpha:  f32,
+  beta:   f32,
+  lda:    u32,
+  ldb:    u32,
+  ldc:    u32,
+  transA: u32, // 0 = no-transpose, 1 = transpose
+  transB: u32,
+}
+
+@group(0) @binding(3) var<uniform> params: Params;
+
+var<workgroup> As: array<f32, BM * BK>;
+var<workgroup> Bs: array<f32, BK * BN>;
+
+@compute @workgroup_size(THREADS_X, THREADS_Y)
+fn main(
+  @builtin(workgroup_id) wid: vec3u,
+  @builtin(local_invocation_id) lid: vec3u,
+  @builtin(local_invocation_index) tid: u32,
+) {
+  let blockRow = wid.y * BM;
+  let blockCol = wid.x * BN;
+  let threadCol = lid.x;
+  let threadRow = lid.y;
+
+  let innerRowA = tid / BK;
+  let innerColA = tid % BK;
+  let innerRowB = tid / BN;
+  let innerColB = tid % BN;
+
+  var threadResults: array<f32, TM * TN>;
+  for (var i = 0u; i < TM * TN; i++) {
+    threadResults[i] = 0.0;
+  }
+  var regM: array<f32, TM>;
+  var regN: array<f32, TN>;
+
+  let numTiles = (params.k + BK - 1u) / BK;
+  for (var t = 0u; t < numTiles; t++) {
+    for (var loadOffset = 0u; loadOffset < BM; loadOffset += STRIDE_A) {
+      let gRowA = blockRow + innerRowA + loadOffset;
+      let gColA = t * BK + innerColA;
+      let aIdx = select(gRowA * params.lda + gColA, gColA * params.lda + gRowA, params.transA != 0u);
+      As[(innerRowA + loadOffset) * BK + innerColA] = select(0.0, A[aIdx], gRowA < params.m && gColA < params.k);
+    }
+    for (var loadOffset = 0u; loadOffset < BK; loadOffset += STRIDE_B) {
+      let gRowB = t * BK + innerRowB + loadOffset;
+      let gColB = blockCol + innerColB;
+      let bIdx = select(gRowB * params.ldb + gColB, gColB * params.ldb + gRowB, params.transB != 0u);
+      Bs[(innerRowB + loadOffset) * BN + innerColB] = select(0.0, B[bIdx], gRowB < params.k && gColB < params.n);
+    }
+
+    workgroupBarrier();
+
+    for (var dotIdx = 0u; dotIdx < BK; dotIdx++) {
+      for (var i = 0u; i < TM; i++) {
+        regM[i] = As[(threadRow * TM + i) * BK + dotIdx];
+      }
+      for (var i = 0u; i < TN; i++) {
+        regN[i] = Bs[dotIdx * BN + threadCol * TN + i];
+      }
+      for (var resIdxM = 0u; resIdxM < TM; resIdxM++) {
+        for (var resIdxN = 0u; resIdxN < TN; resIdxN++) {
+          threadResults[resIdxM * TN + resIdxN] += regM[resIdxM] * regN[resIdxN];
+        }
+      }
+    }
+
+    workgroupBarrier();
+  }
+
+  for (var resIdxM = 0u; resIdxM < TM; resIdxM++) {
+    let row = blockRow + threadRow * TM + resIdxM;
+    if (row < params.m) {
+      for (var resIdxN = 0u; resIdxN < TN; resIdxN++) {
+        let col = blockCol + threadCol * TN + resIdxN;
+        if (col < params.n) {
+          let cIdx = row * params.ldc + col;
+          C[cIdx] = params.alpha * threadResults[resIdxM * TN + resIdxN] + params.beta * C[cIdx];
+        }
+      }
+    }
+  }
+}
+`});var kt,Gt=O(()=>{kt=`// 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,
+// swept BM/BN/BK/TM/TN and warp-tiled variants), +69% over the old BM=32
+// single-tier baseline at n=512, +84% at n=1024. But BM=64 loses to BM=32
+// below a 6x6=36 workgroup grid (not enough workgroups to fill the GPU at
+// that tile size), hence the two-tier split rather than one global config.
+// Neither vectorized loads (kernel 6) nor warp-tiling (kernel 10) beat this
+// at the sizes tried, including warp-tiled variants in the same sweep at
+// BM=64/128.
+
+const BM: u32 = 64u;
+const BN: u32 = 64u;
+const BK: u32 = 8u;
+const TM: u32 = 8u;
+const TN: u32 = 4u;
+const THREADS_X: u32 = BN / TN;
+const THREADS_Y: u32 = BM / TM;
+const NUM_THREADS: u32 = THREADS_X * THREADS_Y; // 128
+const STRIDE_A: u32 = NUM_THREADS / BK; // rows of As covered per load-loop step
+const STRIDE_B: u32 = NUM_THREADS / BN; // rows of Bs covered per load-loop step
+
+@group(0) @binding(0) var<storage, read>       A: array<f32>;
+@group(0) @binding(1) var<storage, read>       B: array<f32>;
+@group(0) @binding(2) var<storage, read_write> C: array<f32>;
+
+struct Params {
+  m:      u32,
+  n:      u32,
+  k:      u32,
+  alpha:  f32,
+  beta:   f32,
+  lda:    u32,
+  ldb:    u32,
+  ldc:    u32,
+  transA: u32, // 0 = no-transpose, 1 = transpose
+  transB: u32,
+}
+
+@group(0) @binding(3) var<uniform> params: Params;
+
+var<workgroup> As: array<f32, BM * BK>;
+var<workgroup> Bs: array<f32, BK * BN>;
+
+@compute @workgroup_size(THREADS_X, THREADS_Y)
+fn main(
+  @builtin(workgroup_id) wid: vec3u,
+  @builtin(local_invocation_id) lid: vec3u,
+  @builtin(local_invocation_index) tid: u32,
+) {
+  let blockRow = wid.y * BM;
+  let blockCol = wid.x * BN;
+  let threadCol = lid.x;
+  let threadRow = lid.y;
+
+  // Load indices, independent of the compute thread shape \u2014 a loop since
+  // NUM_THREADS doesn't match the tile size 1:1 at this config.
+  let innerRowA = tid / BK;
+  let innerColA = tid % BK;
+  let innerRowB = tid / BN;
+  let innerColB = tid % BN;
+
+  var threadResults: array<f32, TM * TN>;
+  for (var i = 0u; i < TM * TN; i++) {
+    threadResults[i] = 0.0;
+  }
+  var regM: array<f32, TM>;
+  var regN: array<f32, TN>;
+
+  let numTiles = (params.k + BK - 1u) / BK;
+  for (var t = 0u; t < numTiles; t++) {
+    for (var loadOffset = 0u; loadOffset < BM; loadOffset += STRIDE_A) {
+      let gRowA = blockRow + innerRowA + loadOffset;
+      let gColA = t * BK + innerColA;
+      let aIdx = select(gRowA * params.lda + gColA, gColA * params.lda + gRowA, params.transA != 0u);
+      As[(innerRowA + loadOffset) * BK + innerColA] = select(0.0, A[aIdx], gRowA < params.m && gColA < params.k);
+    }
+    for (var loadOffset = 0u; loadOffset < BK; loadOffset += STRIDE_B) {
+      let gRowB = t * BK + innerRowB + loadOffset;
+      let gColB = blockCol + innerColB;
+      let bIdx = select(gRowB * params.ldb + gColB, gColB * params.ldb + gRowB, params.transB != 0u);
+      Bs[(innerRowB + loadOffset) * BN + innerColB] = select(0.0, B[bIdx], gRowB < params.k && gColB < params.n);
+    }
+
+    workgroupBarrier();
+
+    for (var dotIdx = 0u; dotIdx < BK; dotIdx++) {
+      for (var i = 0u; i < TM; i++) {
+        regM[i] = As[(threadRow * TM + i) * BK + dotIdx];
+      }
+      for (var i = 0u; i < TN; i++) {
+        regN[i] = Bs[dotIdx * BN + threadCol * TN + i];
+      }
+      for (var resIdxM = 0u; resIdxM < TM; resIdxM++) {
+        for (var resIdxN = 0u; resIdxN < TN; resIdxN++) {
+          threadResults[resIdxM * TN + resIdxN] += regM[resIdxM] * regN[resIdxN];
+        }
+      }
+    }
+
+    workgroupBarrier();
+  }
+
+  for (var resIdxM = 0u; resIdxM < TM; resIdxM++) {
+    let row = blockRow + threadRow * TM + resIdxM;
+    if (row < params.m) {
+      for (var resIdxN = 0u; resIdxN < TN; resIdxN++) {
+        let col = blockCol + threadCol * TN + resIdxN;
+        if (col < params.n) {
+          let cIdx = row * params.ldc + col;
+          C[cIdx] = params.alpha * threadResults[resIdxM * TN + resIdxN] + params.beta * C[cIdx];
+        }
+      }
+    }
+  }
+}
+`});var Nt,Pt=O(()=>{Nt=`// 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).
+
+const BM: u32 = 32u;
+const BN: u32 = 32u;
+const BK: u32 = 8u;
+const TM: u32 = 2u;
+const TN: u32 = 2u;
+const THREADS_X: u32 = BN / TN;
+const THREADS_Y: u32 = BM / TM;
+const NUM_THREADS: u32 = THREADS_X * THREADS_Y; // 256
+const STRIDE_A: u32 = NUM_THREADS / BK;
+const STRIDE_B: u32 = NUM_THREADS / BN;
+
+@group(0) @binding(0) var<storage, read>       A: array<f32>;
+@group(0) @binding(1) var<storage, read>       B: array<f32>;
+@group(0) @binding(2) var<storage, read_write> C: array<f32>;
+
+struct Params {
+  m:      u32,
+  n:      u32,
+  k:      u32,
+  alpha:  f32,
+  beta:   f32,
+  lda:    u32,
+  ldb:    u32,
+  ldc:    u32,
+  transA: u32, // 0 = no-transpose, 1 = transpose
+  transB: u32,
+  uplo:   u32, // 0 = lower (col <= row), 1 = upper (col >= row)
+}
+
+@group(0) @binding(3) var<uniform> params: Params;
+
+var<workgroup> As: array<f32, BM * BK>;
+var<workgroup> Bs: array<f32, BK * BN>;
+
+@compute @workgroup_size(THREADS_X, THREADS_Y)
+fn main(
+  @builtin(workgroup_id) wid: vec3u,
+  @builtin(local_invocation_id) lid: vec3u,
+  @builtin(local_invocation_index) tid: u32,
+) {
+  let blockRow = wid.y * BM;
+  let blockCol = wid.x * BN;
+  let threadCol = lid.x;
+  let threadRow = lid.y;
+
+  let innerRowA = tid / BK;
+  let innerColA = tid % BK;
+  let innerRowB = tid / BN;
+  let innerColB = tid % BN;
+
+  var threadResults: array<f32, TM * TN>;
+  for (var i = 0u; i < TM * TN; i++) {
+    threadResults[i] = 0.0;
+  }
+  var regM: array<f32, TM>;
+  var regN: array<f32, TN>;
+
+  let numTiles = (params.k + BK - 1u) / BK;
+  for (var t = 0u; t < numTiles; t++) {
+    for (var loadOffset = 0u; loadOffset < BM; loadOffset += STRIDE_A) {
+      let gRowA = blockRow + innerRowA + loadOffset;
+      let gColA = t * BK + innerColA;
+      let aIdx = select(gRowA * params.lda + gColA, gColA * params.lda + gRowA, params.transA != 0u);
+      As[(innerRowA + loadOffset) * BK + innerColA] = select(0.0, A[aIdx], gRowA < params.m && gColA < params.k);
+    }
+    for (var loadOffset = 0u; loadOffset < BK; loadOffset += STRIDE_B) {
+      let gRowB = t * BK + innerRowB + loadOffset;
+      let gColB = blockCol + innerColB;
+      let bIdx = select(gRowB * params.ldb + gColB, gColB * params.ldb + gRowB, params.transB != 0u);
+      Bs[(innerRowB + loadOffset) * BN + innerColB] = select(0.0, B[bIdx], gRowB < params.k && gColB < params.n);
+    }
+
+    workgroupBarrier();
+
+    for (var dotIdx = 0u; dotIdx < BK; dotIdx++) {
+      for (var i = 0u; i < TM; i++) {
+        regM[i] = As[(threadRow * TM + i) * BK + dotIdx];
+      }
+      for (var i = 0u; i < TN; i++) {
+        regN[i] = Bs[dotIdx * BN + threadCol * TN + i];
+      }
+      for (var resIdxM = 0u; resIdxM < TM; resIdxM++) {
+        for (var resIdxN = 0u; resIdxN < TN; resIdxN++) {
+          threadResults[resIdxM * TN + resIdxN] += regM[resIdxM] * regN[resIdxN];
+        }
+      }
+    }
+
+    workgroupBarrier();
+  }
+
+  for (var resIdxM = 0u; resIdxM < TM; resIdxM++) {
+    let row = blockRow + threadRow * TM + resIdxM;
+    if (row < params.m) {
+      for (var resIdxN = 0u; resIdxN < TN; resIdxN++) {
+        let col = blockCol + threadCol * TN + resIdxN;
+        let inTriangle = select(col >= row, col <= row, params.uplo == 0u);
+        if (col < params.n && inTriangle) {
+          let cIdx = row * params.ldc + col;
+          C[cIdx] = params.alpha * threadResults[resIdxM * TN + resIdxN] + params.beta * C[cIdx];
+        }
+      }
+    }
+  }
+}
+`});var Mt,St=O(()=>{Mt=`// 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
+// use (0 = lower: col <= row, 1 = upper: col >= row). Every other element of
+// C \u2014 including inside the compute loop, where the full tile is still
+// computed regardless of uplo, only the write is masked \u2014 is left untouched.
+// gemmtr's uplo(C) test is a plain row/col comparison over the full m\xD7n
+// grid, well-defined even when m != n (not restricted to square C).
+
+const BM: u32 = 64u;
+const BN: u32 = 64u;
+const BK: u32 = 8u;
+const TM: u32 = 8u;
+const TN: u32 = 4u;
+const THREADS_X: u32 = BN / TN;
+const THREADS_Y: u32 = BM / TM;
+const NUM_THREADS: u32 = THREADS_X * THREADS_Y; // 128
+const STRIDE_A: u32 = NUM_THREADS / BK; // rows of As covered per load-loop step
+const STRIDE_B: u32 = NUM_THREADS / BN; // rows of Bs covered per load-loop step
+
+@group(0) @binding(0) var<storage, read>       A: array<f32>;
+@group(0) @binding(1) var<storage, read>       B: array<f32>;
+@group(0) @binding(2) var<storage, read_write> C: array<f32>;
+
+struct Params {
+  m:      u32,
+  n:      u32,
+  k:      u32,
+  alpha:  f32,
+  beta:   f32,
+  lda:    u32,
+  ldb:    u32,
+  ldc:    u32,
+  transA: u32, // 0 = no-transpose, 1 = transpose
+  transB: u32,
+  uplo:   u32, // 0 = lower (col <= row), 1 = upper (col >= row)
+}
+
+@group(0) @binding(3) var<uniform> params: Params;
+
+var<workgroup> As: array<f32, BM * BK>;
+var<workgroup> Bs: array<f32, BK * BN>;
+
+@compute @workgroup_size(THREADS_X, THREADS_Y)
+fn main(
+  @builtin(workgroup_id) wid: vec3u,
+  @builtin(local_invocation_id) lid: vec3u,
+  @builtin(local_invocation_index) tid: u32,
+) {
+  let blockRow = wid.y * BM;
+  let blockCol = wid.x * BN;
+  let threadCol = lid.x;
+  let threadRow = lid.y;
+
+  // Load indices, independent of the compute thread shape \u2014 a loop since
+  // NUM_THREADS doesn't match the tile size 1:1 at this config.
+  let innerRowA = tid / BK;
+  let innerColA = tid % BK;
+  let innerRowB = tid / BN;
+  let innerColB = tid % BN;
+
+  var threadResults: array<f32, TM * TN>;
+  for (var i = 0u; i < TM * TN; i++) {
+    threadResults[i] = 0.0;
+  }
+  var regM: array<f32, TM>;
+  var regN: array<f32, TN>;
+
+  let numTiles = (params.k + BK - 1u) / BK;
+  for (var t = 0u; t < numTiles; t++) {
+    for (var loadOffset = 0u; loadOffset < BM; loadOffset += STRIDE_A) {
+      let gRowA = blockRow + innerRowA + loadOffset;
+      let gColA = t * BK + innerColA;
+      let aIdx = select(gRowA * params.lda + gColA, gColA * params.lda + gRowA, params.transA != 0u);
+      As[(innerRowA + loadOffset) * BK + innerColA] = select(0.0, A[aIdx], gRowA < params.m && gColA < params.k);
+    }
+    for (var loadOffset = 0u; loadOffset < BK; loadOffset += STRIDE_B) {
+      let gRowB = t * BK + innerRowB + loadOffset;
+      let gColB = blockCol + innerColB;
+      let bIdx = select(gRowB * params.ldb + gColB, gColB * params.ldb + gRowB, params.transB != 0u);
+      Bs[(innerRowB + loadOffset) * BN + innerColB] = select(0.0, B[bIdx], gRowB < params.k && gColB < params.n);
+    }
+
+    workgroupBarrier();
+
+    for (var dotIdx = 0u; dotIdx < BK; dotIdx++) {
+      for (var i = 0u; i < TM; i++) {
+        regM[i] = As[(threadRow * TM + i) * BK + dotIdx];
+      }
+      for (var i = 0u; i < TN; i++) {
+        regN[i] = Bs[dotIdx * BN + threadCol * TN + i];
+      }
+      for (var resIdxM = 0u; resIdxM < TM; resIdxM++) {
+        for (var resIdxN = 0u; resIdxN < TN; resIdxN++) {
+          threadResults[resIdxM * TN + resIdxN] += regM[resIdxM] * regN[resIdxN];
+        }
+      }
+    }
+
+    workgroupBarrier();
+  }
+
+  for (var resIdxM = 0u; resIdxM < TM; resIdxM++) {
+    let row = blockRow + threadRow * TM + resIdxM;
+    if (row < params.m) {
+      for (var resIdxN = 0u; resIdxN < TN; resIdxN++) {
+        let col = blockCol + threadCol * TN + resIdxN;
+        let inTriangle = select(col >= row, col <= row, params.uplo == 0u);
+        if (col < params.n && inTriangle) {
+          let cIdx = row * params.ldc + col;
+          C[cIdx] = params.alpha * threadResults[resIdxM * TN + resIdxN] + params.beta * C[cIdx];
+        }
+      }
+    }
+  }
+}
+`});var Lt,It=O(()=>{Lt=`// 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
+// that read a symmetric matrix as a normal dense gemm input (e.g. ssymm),
+// rather than teaching the tiled gemm kernel itself to mirror-read.
+
+@group(0) @binding(0) var<storage, read>       A: array<f32>;
+@group(0) @binding(1) var<storage, read_write> Adense: array<f32>;
+
+struct Params {
+  n:    u32,
+  lda:  u32,
+  ldd:  u32, // leading dimension of Adense
+  uplo: u32, // 0 = lower (stored where col <= row), 1 = upper (col >= row)
+}
+
+@group(0) @binding(2) var<uniform> params: Params;
+
+@compute @workgroup_size(8, 8)
+fn main(@builtin(global_invocation_id) gid: vec3u) {
+  let row = gid.y;
+  let col = gid.x;
+  if (row >= params.n || col >= params.n) {
+    return;
+  }
+
+  let isStored = select(col >= row, col <= row, params.uplo == 0u);
+  let srcIdx = select(col * params.lda + row, row * params.lda + col, isStored);
+  Adense[row * params.ldd + col] = A[srcIdx];
+}
+`});var Rt,Dt=O(()=>{Rt=`// 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.
+
+@group(0) @binding(0) var<storage, read>       A: array<f32>;
+@group(0) @binding(1) var<storage, read_write> Adense: array<f32>;
+
+struct Params {
+  n:     u32,
+  lda:   u32,
+  ldd:   u32, // leading dimension of Adense
+  uplo:  u32, // 0 = lower, 1 = upper
+  trans: u32, // 0 = no-transpose (op(A) = A), 1 = transpose (op(A) = A^T)
+  diag:  u32, // 0 = non-unit, 1 = unit
+}
+
+@group(0) @binding(2) var<uniform> params: Params;
+
+@compute @workgroup_size(8, 8)
+fn main(@builtin(global_invocation_id) gid: vec3u) {
+  let row = gid.y;
+  let col = gid.x;
+  if (row >= params.n || col >= params.n) {
+    return;
+  }
+
+  if (row == col) {
+    Adense[row * params.ldd + col] = select(A[row * params.lda + row], 1.0, params.diag == 1u);
+    return;
+  }
+
+  var isMeaningful: bool;
+  var srcRow: u32;
+  var srcCol: u32;
+  if (params.trans == 0u) {
+    isMeaningful = select(col >= row, col <= row, params.uplo == 0u);
+    srcRow = row; srcCol = col;
+  } else {
+    isMeaningful = select(col <= row, col >= row, params.uplo == 0u);
+    srcRow = col; srcCol = row;
+  }
+
+  Adense[row * params.ldd + col] = select(0.0, A[srcRow * params.lda + srcCol], isMeaningful);
+}
+`});var jt,Tt=O(()=>{jt=`// 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.
+
+@group(0) @binding(0) var<storage, read_write> block:   array<f32>; // blockLen x otherLen, block[i*otherLen+j]
+@group(0) @binding(1) var<storage, read_write> strided: array<f32>; // B's or A's own buffer
+
+struct Params {
+  blockStart: u32,
+  blockLen:   u32,
+  otherStart: u32,
+  otherLen:   u32,
+  ld:         u32,
+  isColMajor: u32, // 0 = row-major addressing, 1 = column-major (row/col swapped)
+  blockIsRow: u32, // 1 = blockStart indexes strided's rows, 0 = its columns
+  mode:       u32, // 0 = scatter (strided := block), 1 = scatter_sub (strided -= block), 2 = gather (block := strided)
+}
+
+@group(0) @binding(2) var<uniform> params: Params;
+
+@compute @workgroup_size(8, 8)
+fn main(@builtin(global_invocation_id) gid: vec3u) {
+  let i = gid.y; // index along the blocked axis, within the block
+  let j = gid.x; // index along the other axis, within the block
+  if (i >= params.blockLen || j >= params.otherLen) {
+    return;
+  }
+
+  let row = select(params.otherStart + j, params.blockStart + i, params.blockIsRow == 1u);
+  let col = select(params.blockStart + i, params.otherStart + j, params.blockIsRow == 1u);
+  let stridedIdx = select(row * params.ld + col, col * params.ld + row, params.isColMajor == 1u);
+  let blockIdx = i * params.otherLen + j;
+
+  if (params.mode == 2u) {
+    block[blockIdx] = strided[stridedIdx];
+  } else if (params.mode == 1u) {
+    strided[stridedIdx] -= block[blockIdx];
+  } else {
+    strided[stridedIdx] = block[blockIdx];
+  }
+}
+`});var Ct={};te(Ct,{shaderSources:()=>Ga});var Ga,Wt=O(()=>{pe();ge();he();ve();_e();Ee();Ge();Pe();Se();Ie();De();Te();Ce();Fe();Ue();Ve();ze();Ye();Qe();$e();rt();tt();at();st();ut();ft();ct();pt();gt();ht();vt();_t();Et();Gt();Pt();St();It();Dt();Tt();Ga={"reduction/argmax":we,"reduction/argmaxF64":be,"reduction/sum":xe,"reduction/sumF64":ye,sscal:Be,sswap:Ae,saxpy:ke,scopy:Ne,sdot:Me,sasum:Le,snrm2:Re,srot:je,srotm:We,isamax:He,sgemv_n:Oe,sgemv_t:Ke,ssymv:qe,strmv:Xe,sger:Ze,ssyr:Je,ssyr2:et,f64add:ot,"f64/dekker":it,"f64/utils/abs":nt,"f64/utils/add":lt,"f64/utils/greater":mt,"f64/utils/equal":dt,dasum:wt,idamax:bt,strsv_invert_block:xt,strsv_apply_inverse:yt,strsv_update:Bt,sgemm_small:At,sgemm_large:kt,sgemmtr_small:Nt,sgemmtr_large:Mt,symmetrize:Lt,triangularize:Rt,block_transfer:jt}});var ni={};te(ni,{GpuMatrix:()=>H,GpuVector:()=>I,cleanup:()=>le,dasum:()=>Xt,gpuName:()=>fe,idamax:()=>Jt,init:()=>ue,isamax:()=>$t,randomFloat32Array:()=>me,randomFloat64Array:()=>ce,randomTriangularFloat32Array:()=>de,sasum:()=>Yt,saxpy:()=>Ot,scopy:()=>Vt,sdot:()=>zt,sgemm:()=>mo,sgemmtr:()=>co,sgemv:()=>to,sger:()=>uo,snrm2:()=>Zt,srot:()=>ro,srotm:()=>eo,sscal:()=>Ht,sswap:()=>Ut,ssymm:()=>bo,ssymv:()=>oo,ssyr:()=>lo,ssyr2:()=>fo,ssyr2k:()=>wo,ssyrk:()=>po,strmm:()=>xo,strmv:()=>ao,strsm:()=>Ao,strsv:()=>no});function ae(a,e){return e?a.features.has("timestamp-query")?{requiredFeatures:["timestamp-query"]}:(console.warn("timestamp-query not supported on this device \u2014 benchmark mode disabled."),{}):{}}function ie(){if(!se())return{querySet:null,passDescriptor:void 0};let e=lr().createQuerySet({type:"timestamp",count:2});return{querySet:e,passDescriptor:{timestampWrites:{querySet:e,beginningOfPassWriteIndex:0,endOfPassWriteIndex:1}}}}function br(a,e){if(!e)return null;let r=lr(),o=r.createBuffer({label:"timestamp-resolve",size:16,usage:GPUBufferUsage.QUERY_RESOLVE|GPUBufferUsage.COPY_SRC});a.resolveQuerySet(e,0,2,o,0);let t=r.createBuffer({label:"timestamp-readback",size:16,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});return a.copyBufferToBuffer(o,0,t,0,16),{tsReadBuffer:t,resolveBuffer:o,querySet:e}}async function S(a){if(!a)return;let{tsReadBuffer:e,resolveBuffer:r,querySet:o}=a;await e.mapAsync(GPUMapMode.READ);let t=new BigInt64Array(e.getMappedRange().slice());return e.unmap(),e.destroy(),r.destroy(),o.destroy(),Math.max(0,Number(t[1]-t[0]))/1e6}var Ar=null,Ir=null,ne=null,Qr=!1;async function ue({powerPreference:a="high-performance",benchmark:e=!1,dumpShaders:r=!1}={}){if(Ar)return Ar;let o;if(typeof window>"u"){let{create:s,globals:l}=await import("webgpu");Object.assign(globalThis,l),o=s(r?["enable-dawn-features=dump_shaders,disable_symbol_renaming"]:[]),ne=o}else r&&console.warn("dumpShaders has no effect in the browser \u2014 see init()'s docs."),o=navigator.gpu;if(!o)throw new Error("WebGPU not supported in this environment.");if(Ir=await o.requestAdapter({powerPreference:a})??await o.requestAdapter(),!Ir)throw new Error("No WebGPU adapter found.");Qr=e;let i=[...ae(Ir,e).requiredFeatures??[]];return Ar=await Ir.requestDevice({requiredFeatures:i}),Ar.addEventListener("uncapturederror",s=>{console.error("Uncaptured GPU error:",s.error.message)}),Ar}function le(){Ar&&(Ar.destroy(),Ar=null),Ir=null,ne=null,Qr=!1}function fe(){if(!Ir)throw new Error("WebGPU adapter not initialized \u2014 call init() first.");let{device:a,description:e}=Ir.info;return{description:e||"unknown",device:a||"unknown"}}function se(){return Qr}function lr(){if(!Ar)throw new Error("WebGPU device not initialized \u2014 call init() first.");return Ar}function d(...a){a.flat().forEach(e=>e.destroy())}function v(a,e="blas-input",r=!1){let o=lr(),t=o.limits.maxStorageBufferBindingSize,i=a.byteLength;if(i>t)throw new Error(`Buffer size ${i} bytes exceeds device limit of ${t} bytes.`);let s=r?GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC:GPUBufferUsage.STORAGE,l=o.createBuffer({label:e,size:i,usage:s,mappedAtCreation:!0}),n=a.constructor;return new n(l.getMappedRange()).set(a),l.unmap(),l}function er(a,e="blas-storage",r=0){return lr().createBuffer({label:e,size:a,usage:GPUBufferUsage.STORAGE|r})}function xr(a,e="blas-result"){return lr().createBuffer({label:e,size:a,usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC})}function N(a,e){let o=lr().createBuffer({label:"blas-readback",size:e.size,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});return a.copyBufferToBuffer(e,0,o,0,e.size),o}function L(a,e="blas-params"){let r=lr(),o=a.length*4,t=Math.ceil(o/16)*16,i=new ArrayBuffer(t),s=new DataView(i);a.forEach(({value:n,type:u},f)=>{let m=f*4;if(u==="u32")s.setUint32(m,n,!0);else if(u==="i32")s.setInt32(m,n,!0);else if(u==="f32")s.setFloat32(m,n,!0);else throw new Error(`Unknown param type "${u}". Use "f32", "u32", or "i32".`)});let l=r.createBuffer({label:e,size:t,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});return r.queue.writeBuffer(l,0,i),l}async function k(a,e=Float32Array){try{await a.mapAsync(GPUMapMode.READ);let r=new e(a.getMappedRange().slice());return a.unmap(),r}finally{a.destroy()}}function Pr(a){let e=a.length,r=new Float32Array(e),o=new Float32Array(e);for(let t=0;t<e;t++){let i=Math.fround(a[t]);r[t]=i,o[t]=Math.fround(a[t]-i)}return{hi:r,lo:o}}function Dr(a,e){let r=a.length,o=new Float64Array(r);for(let t=0;t<r;t++)o[t]=a[t]+e[t];return o}var I=class a{constructor(e,r,o=Float32Array,t=null){this._buf=e,this._loBuf=t,this.length=r,this.dtype=o}static from(e){if(e instanceof Float64Array){let{hi:o,lo:t}=Pr(e),i=v(o,"gpu-vector-f64-hi",!0),s=v(t,"gpu-vector-f64-lo",!0);return new a(i,e.length,Float64Array,s)}if(!(e instanceof Float32Array))throw new Error("GpuVector.from expects a Float32Array or Float64Array.");let r=v(e,"gpu-vector",!0);return new a(r,e.length,e.constructor)}async read(){let e=lr(),r=e.createCommandEncoder(),o=N(r,this._buf);if(e.queue.submit([r.finish()]),!this._loBuf)return k(o,this.dtype);let t=e.createCommandEncoder(),i=N(t,this._loBuf);e.queue.submit([t.finish()]);let[s,l]=await Promise.all([k(o,Float32Array),k(i,Float32Array)]);return Dr(s,l)}destroy(){this._buf.destroy(),this._loBuf&&this._loBuf.destroy()}};var H=class a{constructor(e,r,o,t,i=null,s="row-major"){this._buf=e,this._loBuf=i,this.rows=r,this.cols=o,this.lda=t,this.layout=s}static from(e,r,o,t,i="row-major"){if(i!=="row-major"&&i!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");let s=i==="row-major";if(t===void 0&&(t=s?o:r),!(e instanceof Float32Array)&&!(e instanceof Float64Array))throw new Error("GpuMatrix.from expects a Float32Array or Float64Array.");if(!Number.isInteger(r)||r<=0)throw new Error("rows must be a positive integer.");if(!Number.isInteger(o)||o<=0)throw new Error("cols must be a positive integer.");let l=s?o:r;if(!Number.isInteger(t)||t<l)throw new Error(`lda must be an integer >= ${s?"cols":"rows"}.`);let n=s?r:o;if(e.length<n*t)throw new Error("data does not have enough elements for the given rows, cols, and lda.");if(e instanceof Float64Array){let f=n*t,{hi:m,lo:w}=Pr(e.subarray(0,f)),c=v(m,"gpu-matrix-f64-hi",!0),p=v(w,"gpu-matrix-f64-lo",!0);return new a(c,r,o,t,p,i)}let u=v(e.subarray(0,n*t),"gpu-matrix",!0);return new a(u,r,o,t,null,i)}async read(){let e=lr(),r=e.createCommandEncoder(),o=N(r,this._buf);e.queue.submit([r.finish()]);let t=this.layout!=="column-major",i=t?this.rows:this.cols,s=t?this.cols:this.rows;if(this._loBuf){let u=e.createCommandEncoder(),f=N(u,this._loBuf);e.queue.submit([u.finish()]);let[m,w]=await Promise.all([k(o,Float32Array),k(f,Float32Array)]),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 l=await k(o,Float32Array);if(this.lda===s)return l;let n=new Float32Array(i*s);for(let u=0;u<i;u++)n.set(l.subarray(u*this.lda,u*this.lda+s),u*s);return n}destroy(){this._buf.destroy(),this._loBuf&&this._loBuf.destroy()}};function me(a,e=-1,r=1){let o=new Float32Array(a);for(let t=0;t<a;t++)o[t]=e+Math.random()*(r-e);return o}function ce(a,e=-1,r=1){let o=new Float64Array(a);for(let t=0;t<a;t++)o[t]=e+Math.random()*(r-e);return o}function de(a,e,r="lower",o=-1,t=1,i=5,s=15){if(r!=="lower"&&r!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(e<a)throw new Error("lda must be >= n.");let l=new Float32Array(a*e);for(let n=0;n<a;n++){for(let u=0;u<a;u++){if(n===u)continue;(r==="lower"?u<n:u>n)&&(l[n*e+u]=o+Math.random()*(t-o))}l[n*e+n]=i+Math.random()*(s-i)}return l}function B(a,e,r=0){let o=lr(),t=e.map((i,s)=>({binding:r+s,resource:i instanceof GPUBuffer?{buffer:i}:i}));return o.createBindGroup({layout:a,entries:t})}var Fo=new WeakMap;function M(a){lr().queue.submit([a.finish()])}function vr(){let a=lr(),{querySet:e,passDescriptor:r}=ie();return{commandEncoder:a.createCommandEncoder(),querySet:e,passDescriptor:r}}function ar(a,e,r,o,t){let i=a.beginComputePass(t);i.setPipeline(e),i.setBindGroup(0,r),typeof o=="number"?i.dispatchWorkgroups(o):i.dispatchWorkgroups(o.x,o.y,o.z??1),i.end(),Fo.set(a,i)}function C(a,e,r){let{commandEncoder:o,querySet:t,passDescriptor:i}=vr();ar(o,a,e,r,i);let s=br(o,t);return{commandEncoder:o,ts:s}}var Na={},Zr=new WeakMap;async function G(a,e,r="main"){Zr.has(a)||Zr.set(a,new Map);let o=Zr.get(a),t=Array.isArray(e)?e:[e],i=`${t.join("+")}::${r}`;return o.has(i)||o.set(i,await Pa(t,r)),o.get(i)}async function ka(a){if(typeof process>"u"||!process.versions?.node){let{shaderSources:e}=await Promise.resolve().then(()=>(Wt(),Ct)),r=e[a];if(!r)throw new Error(`Shader "${a}" not found in browser bundle.`);return r}else{let{readFileSync:e}=await import("fs"),{fileURLToPath:r}=await import("url"),{dirname:o,join:t}=await import("path"),i=o(r(Na.url));return e(t(i,`../shaders/${a}.wgsl`),"utf8")}}async function Pa(a,e="main"){let r=lr(),o=a.join("+"),t=(await Promise.all(a.map(ka))).join(`
+`),i=r.createShaderModule({label:o,code:t}),l=(await i.getCompilationInfo()).messages.filter(f=>f.type==="error");if(l.length>0)throw new Error(`Shader "${o}" compilation failed:
+${l.map(f=>`  line ${f.lineNum}: ${f.message}`).join(`
+`)}`);let n=e==="main"?{module:i}:{module:i,entryPoint:e},u=r.createComputePipeline({label:o,layout:"auto",compute:n});return u._shaderModule=i,u}var Sa=64,Ft=8;function mr(a,e){let r=lr().limits.maxComputeWorkgroupsPerDimension;return e===void 0?Math.min(Math.ceil(a/Sa),r):{x:Math.min(Math.ceil(e/Ft),r),y:Math.min(Math.ceil(a/Ft),r)}}async function Ht(a,e,r,o,t){let i=o instanceof I;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!Number.isInteger(t))throw new Error("n and incx must be integers.");if(typeof r!="number")throw new Error("alpha must be a number.");if(Number.isNaN(r))throw new Error("alpha must not be NaN.");if(!Number.isFinite(r))throw new Error("alpha must be finite.");if(t<=0)throw new Error("incx must be positive.");if(!(o instanceof Float32Array)&&!(o instanceof I))throw new Error("x must be a Float32Array or GpuVector.");if(e<=0)return i?{}:o;if(o.length<(e-1)*t+1)throw new Error("x does not have enough elements for the given n and incx.");let s=await G(a,"sscal"),l=null,n=null,u=null;try{l=i?o._buf:v(o,"sscal-x",!0),n=L([{value:e,type:"u32"},{value:r,type:"f32"},{value:t,type:"u32"}],"sscal-params");let f=B(s.getBindGroupLayout(0),[l,n]),{commandEncoder:m,ts:w}=C(s,f,mr(e));u=i?null:N(m,l),M(m);let c=await S(w);if(i)return c!==void 0?{gpuTimeMs:c}:{};let p=await k(u,Float32Array);return u=null,c!==void 0?{x:p,gpuTimeMs:c}:p}finally{!i&&l&&d(l),n&&d(n),u&&d(u)}}async function Ut(a,e,r,o,t,i){let s=r instanceof I,l=t instanceof I;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!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(!(r instanceof Float32Array)&&!(r instanceof I))throw new Error("x must be a Float32Array or GpuVector.");if(!(t instanceof Float32Array)&&!(t instanceof I))throw new Error("y must be a Float32Array or GpuVector.");if(r.constructor!==t.constructor)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(e<=0)return s?{}:{x:r,y:t};if(r.length<(e-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(e-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=await G(a,"sswap"),u=null,f=null,m=null,w=null,c=null;try{u=s?r._buf:v(r,"sswap-x",!0),f=l?t._buf:v(t,"sswap-y",!0),m=L([{value:e,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"sswap-params");let p=B(n.getBindGroupLayout(0),[u,f,m]),{commandEncoder:g,ts:h}=C(n,p,mr(e));w=s?null:N(g,u),c=l?null:N(g,f),M(g);let b=await S(h);if(s&&l)return b!==void 0?{gpuTimeMs:b}:{};let x=await k(w,Float32Array);w=null;let _=await k(c,Float32Array);return c=null,b!==void 0?{x,y:_,gpuTimeMs:b}:{x,y:_}}finally{!s&&u&&d(u),!l&&f&&d(f),m&&d(m),w&&d(w),c&&d(c)}}async function Ot(a,e,r,o,t,i,s){let l=o instanceof I,n=i instanceof I;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!Number.isInteger(t)||!Number.isInteger(s))throw new Error("n, incx, and incy must be integers.");if(typeof r!="number")throw new Error("alpha must be a number.");if(Number.isNaN(r))throw new Error("alpha must not be NaN.");if(!Number.isFinite(r))throw new Error("alpha must be finite.");if(t<=0||s<=0)throw new Error("incx and incy must be positive.");if(!l&&!(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(l!==n)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(e<=0)return n?{}:{y:i};if(o.length<(e-1)*t+1)throw new Error("x does not have enough elements for the given n and incx.");if(i.length<(e-1)*s+1)throw new Error("y does not have enough elements for the given n and incy.");let u=await G(a,"saxpy"),f=null,m=null,w=null,c=null;try{f=l?o._buf:v(o,"saxpy-x",!1),m=n?i._buf:v(i,"saxpy-y",!0),w=L([{value:e,type:"u32"},{value:r,type:"f32"},{value:t,type:"u32"},{value:s,type:"u32"}],"saxpy-params");let p=B(u.getBindGroupLayout(0),[f,m,w]),{commandEncoder:g,ts:h}=C(u,p,mr(e));c=n?null:N(g,m),M(g);let b=await S(h);if(n&&l)return b!==void 0?{gpuTimeMs:b}:{};let x=await k(c,Float32Array);return c=null,b!==void 0?{y:x,gpuTimeMs:b}:{y:x}}finally{!l&&f&&d(f),!n&&m&&d(m),w&&d(w),c&&d(c)}}async function Vt(a,e,r,o,t,i){let s=r instanceof I,l=t instanceof I;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!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&&!(r instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!l&&!(t instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(s!==l)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(e<=0)return l?{}:{y:t};if(r.length<(e-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(e-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=await G(a,"scopy"),u=null,f=null,m=null,w=null;try{u=s?r._buf:v(r,"scopy-x",!1),f=l?t._buf:v(t,"scopy-y",!0),m=L([{value:e,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"scopy-params");let c=B(n.getBindGroupLayout(0),[u,f,m]),{commandEncoder:p,ts:g}=C(n,c,mr(e));w=l?null:N(p,f),M(p);let h=await S(g);if(l&&s)return h!==void 0?{gpuTimeMs:h}:{};let b=await k(w,Float32Array);return w=null,h!==void 0?{y:b,gpuTimeMs:h}:{y:b}}finally{!s&&u&&d(u),!l&&f&&d(f),m&&d(m),w&&d(w)}}var Kt=64;async function zt(a,e,r,o,t,i){let s=r instanceof I,l=t instanceof I;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!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&&!(r instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!l&&!(t instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(s!==l)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(e<=0)return{dot:0};if(r.length<(e-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(e-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=await G(a,"sdot"),u=await G(a,"reduction/sum"),f=null,m=null,w=null,c=null,p=null,g=null;try{f=s?r._buf:v(r,"sdot-x",!1),m=l?t._buf:v(t,"sdot-y",!1),w=er(2*Kt*4,"sdot-partials"),c=xr(4,"sdot-result"),p=L([{value:e,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"sdot-params");let h=B(n.getBindGroupLayout(0),[f,m,w,p]),{commandEncoder:b,ts:x}=C(n,h,2*Kt);M(b);let _=B(u.getBindGroupLayout(0),[w,c]),{commandEncoder:y,ts:A}=C(u,_,1);g=N(y,c),M(y);let P=k(g,Float32Array);g=null;let[E,T,D]=await Promise.all([S(x),S(A),P]);return E!==void 0&&T!==void 0?{dot:D[0],gpuTimeMs:E+T}:{dot:D[0]}}finally{!s&&f&&d(f),!l&&m&&d(m),w&&d(w),c&&d(c),p&&d(p),g&&d(g)}}var qt=64;async function Yt(a,e,r,o){let t=r instanceof I;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!t&&!(r instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(e<=0)return{asum:0};if(r.length<(e-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=await G(a,"sasum"),s=await G(a,"reduction/sum"),l=null,n=null,u=null,f=null,m=null;try{l=t?r._buf:v(r,"sasum-x",!1),n=er(2*qt*4,"sasum-partials"),u=xr(4,"sasum-result"),f=L([{value:e,type:"u32"},{value:o,type:"u32"}],"sasum-params");let w=B(i.getBindGroupLayout(0),[l,n,f]),{commandEncoder:c,ts:p}=C(i,w,2*qt);M(c);let g=B(s.getBindGroupLayout(0),[n,u]),{commandEncoder:h,ts:b}=C(s,g,1);m=N(h,u),M(h);let x=k(m,Float32Array);m=null;let[_,y,A]=await Promise.all([S(p),S(b),x]);return _!==void 0&&y!==void 0?{asum:A[0],gpuTimeMs:_+y}:{asum:A[0]}}finally{!t&&l&&d(l),n&&d(n),u&&d(u),f&&d(f),m&&d(m)}}var $r=64;async function Xt(a,e,r,o){let t=r instanceof I;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!t&&!(r instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(t&&r.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(e<=0)return{asum:0};if(r.length<(e-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=["f64/dekker","f64/utils/abs","f64/utils/add"],s=await G(a,[...i,"dasum"]),l=await G(a,[...i,"reduction/sumF64"]),n=null,u=null,f=null,m=null,w=null,c=null,p=null,g=null,h=null;try{if(t)n=r._buf,u=r._loBuf;else{let{hi:K,lo:W}=Pr(r.map(Math.abs));n=v(K,"dasum-xHi",!1),u=v(W,"dasum-xLo",!1)}f=er(2*$r*4,"dasum-partialsHi"),m=er(2*$r*4,"dasum-partialsLo"),w=xr(4,"dasum-result-hi"),c=xr(4,"dasum-result-lo"),p=L([{value:e,type:"u32"},{value:o,type:"u32"}],"dasum-params");let b=B(s.getBindGroupLayout(0),[n,u,f,m,p]),{commandEncoder:x,ts:_}=C(s,b,2*$r);M(x);let y=B(l.getBindGroupLayout(0),[f,m,w,c]),{commandEncoder:A,ts:P}=C(l,y,1);g=N(A,w),h=N(A,c),M(A);let E=k(g,Float32Array),T=k(h,Float32Array);g=null,h=null;let[D,R,j,F]=await Promise.all([S(_),S(P),E,T]),V=Dr(j,F)[0];return D!==void 0&&R!==void 0?{asum:V,gpuTimeMs:D+R}:{asum:V}}finally{!t&&n&&d(n),!t&&u&&d(u),f&&d(f),m&&d(m),w&&d(w),c&&d(c),p&&d(p),g&&d(g),h&&d(h)}}var Qt=64;async function Zt(a,e,r,o){let t=r instanceof I;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!t&&!(r instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(e<=0)return{nrm2:0};if(r.length<(e-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=await G(a,"snrm2"),s=await G(a,"reduction/sum"),l=null,n=null,u=null,f=null,m=null;try{l=t?r._buf:v(r,"snrm2-x",!1),n=er(2*Qt*4,"snrm2-partials"),u=xr(4,"snrm2-result"),f=L([{value:e,type:"u32"},{value:o,type:"u32"}],"snrm2-params");let w=B(i.getBindGroupLayout(0),[l,n,f]),{commandEncoder:c,ts:p}=C(i,w,2*Qt);M(c);let g=B(s.getBindGroupLayout(0),[n,u]),{commandEncoder:h,ts:b}=C(s,g,1);m=N(h,u),M(h);let x=k(m,Float32Array);m=null;let[_,y,A]=await Promise.all([S(p),S(b),x]),P=Math.sqrt(A[0]);return _!==void 0&&y!==void 0?{nrm2:P,gpuTimeMs:_+y}:{nrm2:P}}finally{!t&&l&&d(l),n&&d(n),u&&d(u),f&&d(f),m&&d(m)}}var Jr=64;async function $t(a,e,r,o){let t=r instanceof I;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!t&&!(r instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(e<=0)return{index:0};if(r.length<(e-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=await G(a,"isamax"),s=await G(a,"reduction/argmax"),l=null,n=null,u=null,f=null,m=null,w=null;try{l=t?r._buf:v(r,"isamax-x",!1),n=er(2*Jr*4,"isamax-partials-val"),u=er(2*Jr*4,"isamax-partials-idx"),f=xr(4,"isamax-result"),m=L([{value:e,type:"u32"},{value:o,type:"u32"}],"isamax-params");let c=B(i.getBindGroupLayout(0),[l,n,u,m]),{commandEncoder:p,ts:g}=C(i,c,2*Jr);M(p);let h=B(s.getBindGroupLayout(0),[n,u,f]),{commandEncoder:b,ts:x}=C(s,h,1);w=N(b,f),M(b);let _=k(w,Uint32Array);w=null;let[y,A,P]=await Promise.all([S(g),S(x),_]),E=P[0];return y!==void 0&&A!==void 0?{index:E,gpuTimeMs:y+A}:{index:E}}finally{!t&&l&&d(l),n&&d(n),u&&d(u),f&&d(f),m&&d(m),w&&d(w)}}var Kr=64;async function Jt(a,e,r,o){let t=r instanceof I;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!t&&!(r instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(t&&r.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(e<=0)return{index:0};if(r.length<(e-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 G(a,[...i,"idamax"],"idamax_main"),l=await G(a,[...i,"reduction/argmaxF64"],"reduce_f64"),n=null,u=null,f=null,m=null,w=null,c=null,p=null,g=null;try{if(t)n=r._buf,u=r._loBuf;else{let{hi:j,lo:F}=Pr(r);n=v(j,"idamax-xHi",!1),u=v(F,"idamax-xLo",!1)}f=er(2*Kr*4,"idamax-partials-val-hi"),m=er(2*Kr*4,"idamax-partials-val-lo"),w=er(2*Kr*4,"idamax-partials-idx"),c=xr(4,"idamax-result"),p=L([{value:e,type:"u32"},{value:o,type:"u32"}],"idamax-params");let h=B(s.getBindGroupLayout(0),[n,u,f,m,w,p]),{commandEncoder:b,ts:x}=C(s,h,2*Kr);M(b);let _=B(l.getBindGroupLayout(0),[f,m,w,c]),{commandEncoder:y,ts:A}=C(l,_,1);g=N(y,c),M(y);let P=k(g,Uint32Array);g=null;let[E,T,D]=await Promise.all([S(x),S(A),P]),R=D[0];return E!==void 0&&T!==void 0?{index:R,gpuTimeMs:E+T}:{index:R}}finally{!t&&n&&d(n),!t&&u&&d(u),f&&d(f),m&&d(m),w&&d(w),c&&d(c),p&&d(p),g&&d(g)}}async function ro(a,e,r,o,t,i,s,l){let n=r instanceof I,u=t instanceof I;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!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 l!="number")throw new Error("s must be a number.");if(Number.isNaN(s)||Number.isNaN(l))throw new Error("c and s must not be NaN.");if(!Number.isFinite(s))throw new Error("c must be finite.");if(!Number.isFinite(l))throw new Error("s must be finite.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!n&&!(r 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(n!==u)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(e<=0)return n?{}:{x:r,y:t};if(r.length<(e-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(e-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let f=await G(a,"srot"),m=null,w=null,c=null,p=null,g=null;try{m=n?r._buf:v(r,"srot-x",!0),w=u?t._buf:v(t,"srot-y",!0),c=L([{value:e,type:"u32"},{value:s,type:"f32"},{value:l,type:"f32"},{value:o,type:"u32"},{value:i,type:"u32"}],"srot-params");let h=B(f.getBindGroupLayout(0),[m,w,c]),{commandEncoder:b,ts:x}=C(f,h,mr(e));p=n?null:N(b,m),g=u?null:N(b,w),M(b);let _=await S(x);if(n&&u)return _!==void 0?{gpuTimeMs:_}:{};let y=k(p,Float32Array),A=k(g,Float32Array);p=null,g=null;let[P,E]=await Promise.all([y,A]);return _!==void 0?{x:P,y:E,gpuTimeMs:_}:{x:P,y:E}}finally{!n&&m&&d(m),!u&&w&&d(w),c&&d(c),p&&d(p),g&&d(g)}}async function eo(a,e,r,o,t,i,s){let l=r instanceof I,n=t instanceof I;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(e)||!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(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!l&&!(r 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(l!==n)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(e<=0||s[0]===-2)return l?{}:{x:r,y:t};if(r.length<(e-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(e-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let u=await G(a,"srotm"),f=null,m=null,w=null,c=null,p=null,g=null;try{f=l?r._buf:v(r,"srotm-x",!0),m=n?t._buf:v(t,"srotm-y",!0),w=v(s,"srotm-param",!1),c=L([{value:e,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"srotm-params");let h=B(u.getBindGroupLayout(0),[f,m,w,c]),{commandEncoder:b,ts:x}=C(u,h,mr(e));p=l?null:N(b,f),g=n?null:N(b,m),M(b);let _=await S(x);if(l&&n)return _!==void 0?{gpuTimeMs:_}:{};let y=k(p,Float32Array),A=k(g,Float32Array);p=null,g=null;let[P,E]=await Promise.all([y,A]);return _!==void 0?{x:P,y:E,gpuTimeMs:_}:{x:P,y:E}}finally{!l&&f&&d(f),!n&&m&&d(m),w&&d(w),c&&d(c),p&&d(p),g&&d(g)}}async function to(a,e,r,o,t,i,s,l,n,u,f,m,w="row-major"){let c=i instanceof H,p=l instanceof I,g=f instanceof I;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(e!=="no-transpose"&&e!=="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 u!="number")throw new Error("beta must be a number.");if(Number.isNaN(u))throw new Error("beta must not be NaN.");if(!Number.isFinite(u))throw new Error("beta must be finite.");if(!Number.isInteger(r)||!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(!p&&!(l instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!g&&!(f 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&&l._buf===f._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(c&&s!==i.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(c&&(i.rows<r||i.cols<o))throw new Error("A is too small for the given m and n.");if(r<0||o<0)throw new Error("m and n must be non-negative.");if(r===0||o===0)return g?{}:{y:f};(c?i.layout:w)==="column-major"&&([r,o]=[o,r],e=e==="no-transpose"?"transpose":"no-transpose");let b=e==="no-transpose",x=b?o:r,_=b?r:o;if(s<o)throw new Error("lda must be >= n.");if(!c&&i.length<(r-1)*s+o)throw new Error("A does not have enough elements for the given m, n, and lda.");if(l.length<(x-1)*n+1)throw new Error("x does not have enough elements for the given dimensions and incx.");if(f.length<(_-1)*m+1)throw new Error("y does not have enough elements for the given dimensions and incy.");let A=await G(a,b?"sgemv_n":"sgemv_t"),P=c?i._buf:v(i,"sgemv-A",!1),E=p?l._buf:v(l,"sgemv-x",!1),T=g?f._buf:v(f,"sgemv-y",!0),D=L([{value:r,type:"u32"},{value:o,type:"u32"},{value:t,type:"f32"},{value:u,type:"f32"},{value:n,type:"u32"},{value:m,type:"u32"},{value:s,type:"u32"}],"sgemv-params");try{let R=B(A.getBindGroupLayout(0),[P,E,T,D]),j=b?Math.min(r,a.limits.maxComputeWorkgroupsPerDimension):mr(_),{commandEncoder:F,ts:V}=C(A,R,j),K=g?null:N(F,T);M(F);let W=await S(V);if(g)return W!==void 0?{gpuTimeMs:W}:{};let $=await k(K,Float32Array);return W!==void 0?{y:$,gpuTimeMs:W}:{y:$}}finally{c||d(P),p||d(E),g||d(T),d(D)}}async function oo(a,e,r,o,t,i,s,l,n,u,f,m="row-major"){let w=s instanceof I,c=u instanceof I,p=t instanceof H;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(e!=="lower"&&e!=="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(r)||!Number.isInteger(l)||!Number.isInteger(f)||!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(l<=0||f<=0)throw new Error("incx and incy must be positive.");if(i<r)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&&!(u 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===u._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<r||t.cols<r))throw new Error("A is too small for the given n.");if(r<0)throw new Error("n must be non-negative.");if(r===0)return c?{}:{y:u};if(!p&&t.length<(r-1)*i+r)throw new Error("A does not have enough elements for the given n and lda.");if(s.length<(r-1)*l+1)throw new Error("x does not have enough elements for the given n and incx.");if(u.length<(r-1)*f+1)throw new Error("y does not have enough elements for the given n and incy.");let h=(p?t.layout:m)==="column-major"?e==="upper":e==="lower",b=await G(a,"ssymv"),x=null,_=null,y=null,A=null;try{x=p?t._buf:v(t,"ssymv-A",!1),_=w?s._buf:v(s,"ssymv-x",!1),y=c?u._buf:v(u,"ssymv-y",!0),A=L([{value:r,type:"u32"},{value:o,type:"f32"},{value:n,type:"f32"},{value:l,type:"u32"},{value:f,type:"u32"},{value:i,type:"u32"},{value:h?0:1,type:"u32"}],"ssymv-params");let P=B(b.getBindGroupLayout(0),[x,_,y,A]),E=Math.min(r,a.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:T,ts:D}=C(b,P,E),R=c?null:N(T,y);M(T);let j=await S(D);if(c)return j!==void 0?{gpuTimeMs:j}:{};let F=await k(R,Float32Array);return j!==void 0?{y:F,gpuTimeMs:j}:{y:F}}finally{!p&&x&&d(x),!w&&_&&d(_),!c&&y&&d(y),A&&d(A)}}async function ao(a,e,r,o,t,i,s,l,n,u,f,m="row-major"){let w=l instanceof I,c=u instanceof I,p=i instanceof H,g=o==="unit";if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(e!=="lower"&&e!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(r!=="no-transpose"&&r!=="transpose")throw new Error("trans must be 'no-transpose' or 'transpose'.");if(!g&&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(t)||!Number.isInteger(n)||!Number.isInteger(f)||!Number.isInteger(s))throw new Error("n, incx, incy, and lda must be integers.");if(n<=0||f<=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&&!(l instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!c&&!(u 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&&l._buf===u._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===u._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:u};if(!p&&i.length<(t-1)*s+t)throw new Error("A does not have enough elements for the given n and lda.");if(l.length<(t-1)*n+1)throw new Error("x does not have enough elements for the given n and incx.");if(u.length<(t-1)*f+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?e==="upper":e==="lower",_=b?r==="transpose":r==="no-transpose",y=await G(a,"strmv"),A=null,P=null,E=null,T=null;try{A=p?i._buf:v(i,"strmv-A",!1),P=w?l._buf:v(l,"strmv-x",!1),E=c?u._buf:v(u,"strmv-y",!0),T=L([{value:t,type:"u32"},{value:n,type:"u32"},{value:f,type:"u32"},{value:s,type:"u32"},{value:_?0:1,type:"u32"},{value:x?0:1,type:"u32"},{value:g?1:0,type:"u32"}],"strmv-params");let D=B(y.getBindGroupLayout(0),[A,P,E,T]),R=Math.min(t,a.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:j,ts:F}=C(y,D,R),V=c?null:N(j,E);M(j);let K=await S(F);if(c)return K!==void 0?{gpuTimeMs:K}:{};let W=await k(V,Float32Array);return K!==void 0?{y:W,gpuTimeMs:K}:{y:W}}finally{!p&&A&&d(A),!w&&P&&d(P),!c&&E&&d(E),T&&d(T)}}var Gr=64;function io(a,e,r){let o=new ArrayBuffer(a*e),t=new DataView(o);for(let i=0;i<a;i++){let s=r(i),l=i*e;s.forEach((n,u)=>t.setUint32(l+u*4,n,!0))}return o}function so(a,e,r){let o=a.createBuffer({label:r,size:e.byteLength,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});return a.queue.writeBuffer(o,0,e),o}async function no(a,e,r,o,t,i,s,l,n,u="row-major"){let f=l instanceof I,m=i instanceof H,w=o==="unit";if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(e!=="lower"&&e!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(r!=="no-transpose"&&r!=="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(u!=="row-major"&&u!=="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(!f&&!(l instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(f&&!m)throw new Error("A must be a GpuMatrix when x is a GpuVector.");if(m&&!f)throw new Error("x must be a GpuVector when A is a GpuMatrix.");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 f?{}:{x:l};if(!m&&i.length<(t-1)*s+t)throw new Error("A does not have enough elements for the given n and lda.");if(l.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:u)==="column-major",g=p?e==="upper":e==="lower",h=p?r==="transpose":r==="no-transpose",b=await G(a,"strsv_invert_block"),x=await G(a,"strsv_apply_inverse"),_=await G(a,"strsv_update"),y=h===g,A=[];for(let W=0;W<t;W+=Gr)A.push(W);y||A.reverse();let P=A.length,E=a.limits.maxComputeWorkgroupsPerDimension,T=a.limits.minUniformBufferOffsetAlignment,D=null,R=null,j=null,F=null,V=null,K=null;try{D=m?i._buf:v(i,"strsv-A",!1),R=f?l._buf:v(l,"strsv-x",!0),j=er(P*Gr*Gr*4,"strsv-Ainv");let W=io(P,T,q=>{let z=q*Gr,X=Math.min(z+Gr,t);return[n,q,z,X]});F=so(a,W,"strsv-apply-params");let $=io(P,T,q=>{let z=q*Gr,X=Math.min(z+Gr,t);return[t,n,s,h?0:1,g?0:1,z,X]});V=so(a,$,"strsv-update-params");let{commandEncoder:Y,querySet:J}=vr();K=L([{value:t,type:"u32"},{value:s,type:"u32"},{value:h?0:1,type:"u32"},{value:g?0:1,type:"u32"},{value:w?1:0,type:"u32"}],"strsv-invert-params");let nr=B(b.getBindGroupLayout(0),[D,j,K]);ar(Y,b,nr,{x:Gr,y:P},J?{timestampWrites:{querySet:J,beginningOfPassWriteIndex:0}}:void 0);for(let q=0;q<A.length;q++){let z=A[q],X=Math.min(z+Gr,t),Q=z/Gr,tr=q===A.length-1,fr=Q*T,or=B(x.getBindGroupLayout(0),[j,R,{buffer:F,offset:fr,size:16}]);ar(Y,x,or,1,tr&&J?{timestampWrites:{querySet:J,endOfPassWriteIndex:1}}:void 0);let dr=y?t-X:z;if(dr===0)continue;let Br=B(_.getBindGroupLayout(0),[D,R,{buffer:V,offset:fr,size:32}]),yr=Math.min(dr,E);ar(Y,_,Br,yr)}let sr=br(Y,J),Z=f?null:N(Y,R);M(Y);let rr=await S(sr);if(f)return rr!==void 0?{gpuTimeMs:rr}:{};let U=await k(Z,Float32Array);return rr!==void 0?{x:U,gpuTimeMs:rr}:{x:U}}finally{!m&&D&&d(D),!f&&R&&d(R),j&&d(j),F&&d(F),V&&d(V),K&&d(K)}}async function uo(a,e,r,o,t,i,s,l,n,u,f="row-major"){let m=n instanceof H;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(f!=="row-major"&&f!=="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(e)||!Number.isInteger(r)||!Number.isInteger(i)||!Number.isInteger(l)||!Number.isInteger(u))throw new Error("m, n, incx, incy, and lda must be integers.");if(i<=0||l<=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&&u!==n.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(m&&(n.rows<e||n.cols<r))throw new Error("A is too small for the given m and n.");(m?n.layout:f)==="column-major"&&([e,r]=[r,e],[t,s]=[s,t],[i,l]=[l,i]);let c=t instanceof I,p=s instanceof I;if(u<r)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(e<0||r<0)throw new Error("m and n must be non-negative.");if(e===0||r===0)return m?{}:{A:n};if(!m&&n.length<(e-1)*u+r)throw new Error("A does not have enough elements for the given m, n, and lda.");if(t.length<(e-1)*i+1)throw new Error("x does not have enough elements for the given m and incx.");if(s.length<(r-1)*l+1)throw new Error("y does not have enough elements for the given n and incy.");let g=await G(a,"sger"),h=null,b=null,x=null,_=null;try{h=c?t._buf:v(t,"sger-x",!1),b=p?s._buf:v(s,"sger-y",!1),x=m?n._buf:v(n,"sger-A",!0),_=L([{value:e,type:"u32"},{value:r,type:"u32"},{value:o,type:"f32"},{value:i,type:"u32"},{value:l,type:"u32"},{value:u,type:"u32"}],"sger-params");let y=B(g.getBindGroupLayout(0),[h,b,x,_]),A=Math.min(e,a.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:P,ts:E}=C(g,y,A),T=m?null:N(P,x);M(P);let D=await S(E);if(m)return D!==void 0?{gpuTimeMs:D}:{};let R=await k(T,Float32Array);return D!==void 0?{A:R,gpuTimeMs:D}:{A:R}}finally{!c&&h&&d(h),!p&&b&&d(b),!m&&x&&d(x),_&&d(_)}}async function lo(a,e,r,o,t,i,s,l,n="row-major"){let u=t instanceof I,f=s instanceof H;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(e!=="lower"&&e!=="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(r)||!Number.isInteger(i)||!Number.isInteger(l))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(l<r)throw new Error("lda must be >= n.");if(!f&&!(s instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!u&&!(t instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(u&&!f)throw new Error("A must be a GpuMatrix when x is a GpuVector.");if(f&&!u)throw new Error("x must be a GpuVector when A is a GpuMatrix.");if(f&&u&&s._buf===t._buf)throw new Error("A and x must not reference the same GPU buffer.");if(f&&l!==s.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(f&&(s.rows<r||s.cols<r))throw new Error("A is too small for the given n.");if(r<0)throw new Error("n must be non-negative.");if(r===0)return f?{}:{A:s};if(!f&&s.length<(r-1)*l+r)throw new Error("A does not have enough elements for the given n and lda.");if(t.length<(r-1)*i+1)throw new Error("x does not have enough elements for the given n and incx.");let w=(f?s.layout:n)==="column-major"?e==="upper":e==="lower",c=await G(a,"ssyr"),p=null,g=null,h=null;try{p=u?t._buf:v(t,"ssyr-x",!1),g=f?s._buf:v(s,"ssyr-A",!0),h=L([{value:r,type:"u32"},{value:o,type:"f32"},{value:i,type:"u32"},{value:l,type:"u32"},{value:w?0:1,type:"u32"}],"ssyr-params");let b=B(c.getBindGroupLayout(0),[p,g,h]),x=Math.min(r,a.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:_,ts:y}=C(c,b,x),A=f?null:N(_,g);M(_);let P=await S(y);if(f)return P!==void 0?{gpuTimeMs:P}:{};let E=await k(A,Float32Array);return P!==void 0?{A:E,gpuTimeMs:P}:{A:E}}finally{!u&&p&&d(p),!f&&g&&d(g),h&&d(h)}}async function fo(a,e,r,o,t,i,s,l,n,u,f="row-major"){let m=t instanceof I,w=s instanceof I,c=n instanceof H;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(e!=="lower"&&e!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(f!=="row-major"&&f!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(!Number.isInteger(r)||!Number.isInteger(i)||!Number.isInteger(l)||!Number.isInteger(u))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||l<=0)throw new Error("incx and incy must be positive.");if(u<r)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&&u!==n.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(c&&(n.rows<r||n.cols<r))throw new Error("A is too small for the given n.");if(r<0)throw new Error("n must be non-negative.");if(r===0)return c?{}:{A:n};if(!c&&n.length<(r-1)*u+r)throw new Error("A does not have enough elements for the given n and lda.");if(t.length<(r-1)*i+1)throw new Error("x does not have enough elements for the given n and incx.");if(s.length<(r-1)*l+1)throw new Error("y does not have enough elements for the given n and incy.");let g=(c?n.layout:f)==="column-major"?e==="upper":e==="lower",h=await G(a,"ssyr2"),b=null,x=null,_=null,y=null;try{b=m?t._buf:v(t,"ssyr2-x",!1),x=w?s._buf:v(s,"ssyr2-y",!1),_=c?n._buf:v(n,"ssyr2-A",!0),y=L([{value:r,type:"u32"},{value:o,type:"f32"},{value:i,type:"u32"},{value:l,type:"u32"},{value:u,type:"u32"},{value:g?0:1,type:"u32"}],"ssyr2-params");let A=B(h.getBindGroupLayout(0),[b,x,_,y]),P=Math.min(r,a.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:E,ts:T}=C(h,A,P),D=c?null:N(E,_);M(E);let R=await S(T);if(c)return R!==void 0?{gpuTimeMs:R}:{};let j=await k(D,Float32Array);return R!==void 0?{A:j,gpuTimeMs:R}:{A:j}}finally{!m&&b&&d(b),!w&&x&&d(x),!c&&_&&d(_),y&&d(y)}}var Ma=32,Ia=32,La=64,Da=64,Ra=36;async function mo(a,e,r,o,t,i,s,l,n,u,f,m,w,c,p="row-major"){let g=l instanceof H,h=u instanceof H,b=w instanceof H;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(e!=="no-transpose"&&e!=="transpose")throw new Error("transA must be 'no-transpose' or 'transpose'.");if(r!=="no-transpose"&&r!=="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(o)||!Number.isInteger(t)||!Number.isInteger(i)||!Number.isInteger(n)||!Number.isInteger(f)||!Number.isInteger(c))throw new Error("m, n, k, lda, ldb, and ldc must be integers.");if(!g&&!(l instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!h&&!(u 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||h)&&!b)throw new Error("C must be a GpuMatrix when A or B is a GpuMatrix.");if(b&&(!g||!h))throw new Error("A and B must be GpuMatrix when C is a GpuMatrix.");if(o<0||t<0||i<0)throw new Error("m, n, and k must be non-negative.");if(o===0||t===0)return b?{}:{C:w};let x=g?l.layout:p,_=h?u.layout:p,y=b?w.layout:p,A=x==="column-major"?i:o,P=x==="column-major"?o:i,E=e==="no-transpose"?A:P,T=e==="no-transpose"?P:A;if(n<T)throw new Error(`lda must be >= ${x==="column-major"?"rows":"cols"} of A as stored.`);if(g){if(n!==l.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");let[rr,U]=e==="no-transpose"?[o,i]:[i,o];if(l.rows<rr||l.cols<U)throw new Error("A is too small for the given m, k, and transA.")}else if(l.length<(E-1)*n+T)throw new Error("A does not have enough elements for the given dimensions and lda.");let D=_==="column-major"?t:i,R=_==="column-major"?i:t,j=r==="no-transpose"?D:R,F=r==="no-transpose"?R:D;if(f<F)throw new Error(`ldb must be >= ${_==="column-major"?"rows":"cols"} of B as stored.`);if(h){if(f!==u.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");let[rr,U]=r==="no-transpose"?[i,t]:[t,i];if(u.rows<rr||u.cols<U)throw new Error("B is too small for the given n, k, and transB.")}else if(u.length<(j-1)*f+F)throw new Error("B does not have enough elements for the given dimensions and ldb.");let V=y==="column-major"?t:o,K=y==="column-major"?o:t;if(c<K)throw new Error(`ldc must be >= ${y==="column-major"?"rows":"cols"} of C as stored.`);if(b){if(c!==w.lda)throw new Error("ldc must match C.lda when C is a GpuMatrix.");if(w.rows<o||w.cols<t)throw new Error("C is too small for the given m and n.")}else if(w.length<(V-1)*c+K)throw new Error("C does not have enough elements for the given dimensions and ldc.");x==="column-major"&&(e=e==="no-transpose"?"transpose":"no-transpose"),_==="column-major"&&(r=r==="no-transpose"?"transpose":"no-transpose"),y==="column-major"&&([l,u]=[u,l],[g,h]=[h,g],[n,f]=[f,n],[e,r]=[r==="no-transpose"?"transpose":"no-transpose",e==="no-transpose"?"transpose":"no-transpose"],[o,t]=[t,o]);let W=Math.ceil(t/Da),$=Math.ceil(o/La),Y=W*$>=Ra,J=await G(a,Y?"sgemm_large":"sgemm_small"),nr=g?l._buf:v(l,"sgemm-A",!1),ur=h?u._buf:v(u,"sgemm-B",!1),sr=b?w._buf:v(w,"sgemm-C",!0),Z=L([{value:o,type:"u32"},{value:t,type:"u32"},{value:i,type:"u32"},{value:s,type:"f32"},{value:m,type:"f32"},{value:n,type:"u32"},{value:f,type:"u32"},{value:c,type:"u32"},{value:e==="transpose"?1:0,type:"u32"},{value:r==="transpose"?1:0,type:"u32"}],"sgemm-params");try{let rr=B(J.getBindGroupLayout(0),[nr,ur,sr,Z]),U=Y?{x:Math.min(W,a.limits.maxComputeWorkgroupsPerDimension),y:Math.min($,a.limits.maxComputeWorkgroupsPerDimension)}:{x:Math.min(Math.ceil(t/Ia),a.limits.maxComputeWorkgroupsPerDimension),y:Math.min(Math.ceil(o/Ma),a.limits.maxComputeWorkgroupsPerDimension)},{commandEncoder:q,ts:z}=C(J,rr,U),X=b?null:N(q,sr);M(q);let Q=await S(z);if(b)return Q!==void 0?{gpuTimeMs:Q}:{};let tr=await k(X,Float32Array);return Q!==void 0?{C:tr,gpuTimeMs:Q}:{C:tr}}finally{g||d(nr),h||d(ur),b||d(sr),d(Z)}}var Ta=32,ja=32,Ca=64,Wa=64,Fa=36;async function co(a,e,r,o,t,i,s,l,n,u,f,m,w,c,p,g="row-major"){let h=n instanceof H,b=f instanceof H,x=c instanceof H;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(e!=="lower"&&e!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(r!=="no-transpose"&&r!=="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(g!=="row-major"&&g!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(typeof l!="number")throw new Error("alpha must be a number.");if(Number.isNaN(l))throw new Error("alpha must not be NaN.");if(!Number.isFinite(l))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(u)||!Number.isInteger(m)||!Number.isInteger(p))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&&!(f 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((h||b)&&!x)throw new Error("C must be a GpuMatrix when A or B is a GpuMatrix.");if(x&&(!h||!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(t===0||i===0)return x?{}:{C:c};let _=h?n.layout:g,y=b?f.layout:g,A=x?c.layout:g,P=_==="column-major"?s:t,E=_==="column-major"?t:s,T=r==="no-transpose"?P:E,D=r==="no-transpose"?E:P;if(u<D)throw new Error(`lda must be >= ${_==="column-major"?"rows":"cols"} of A as stored.`);if(h){if(u!==n.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");let[U,q]=r==="no-transpose"?[t,s]:[s,t];if(n.rows<U||n.cols<q)throw new Error("A is too small for the given m, k, and transA.")}else if(n.length<(T-1)*u+D)throw new Error("A does not have enough elements for the given dimensions and lda.");let R=y==="column-major"?i:s,j=y==="column-major"?s:i,F=o==="no-transpose"?R:j,V=o==="no-transpose"?j:R;if(m<V)throw new Error(`ldb must be >= ${y==="column-major"?"rows":"cols"} of B as stored.`);if(b){if(m!==f.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");let[U,q]=o==="no-transpose"?[s,i]:[i,s];if(f.rows<U||f.cols<q)throw new Error("B is too small for the given n, k, and transB.")}else if(f.length<(F-1)*m+V)throw new Error("B does not have enough elements for the given dimensions and ldb.");let K=A==="column-major"?i:t,W=A==="column-major"?t:i;if(p<W)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<(K-1)*p+W)throw new Error("C does not have enough elements for the given dimensions and ldc.");_==="column-major"&&(r=r==="no-transpose"?"transpose":"no-transpose"),y==="column-major"&&(o=o==="no-transpose"?"transpose":"no-transpose"),A==="column-major"&&([n,f]=[f,n],[h,b]=[b,h],[u,m]=[m,u],[r,o]=[o==="no-transpose"?"transpose":"no-transpose",r==="no-transpose"?"transpose":"no-transpose"],[t,i]=[i,t],e=e==="lower"?"upper":"lower");let $=Math.ceil(i/Wa),Y=Math.ceil(t/Ca),J=$*Y>=Fa,nr=await G(a,J?"sgemmtr_large":"sgemmtr_small"),ur=h?n._buf:v(n,"sgemmtr-A",!1),sr=b?f._buf:v(f,"sgemmtr-B",!1),Z=x?c._buf:v(c,"sgemmtr-C",!0),rr=L([{value:t,type:"u32"},{value:i,type:"u32"},{value:s,type:"u32"},{value:l,type:"f32"},{value:w,type:"f32"},{value:u,type:"u32"},{value:m,type:"u32"},{value:p,type:"u32"},{value:r==="transpose"?1:0,type:"u32"},{value:o==="transpose"?1:0,type:"u32"},{value:e==="upper"?1:0,type:"u32"}],"sgemmtr-params");try{let U=B(nr.getBindGroupLayout(0),[ur,sr,Z,rr]),q=J?{x:Math.min($,a.limits.maxComputeWorkgroupsPerDimension),y:Math.min(Y,a.limits.maxComputeWorkgroupsPerDimension)}:{x:Math.min(Math.ceil(i/ja),a.limits.maxComputeWorkgroupsPerDimension),y:Math.min(Math.ceil(t/Ta),a.limits.maxComputeWorkgroupsPerDimension)},{commandEncoder:z,ts:X}=C(nr,U,q),Q=x?null:N(z,Z);M(z);let tr=await S(X);if(x)return tr!==void 0?{gpuTimeMs:tr}:{};let fr=await k(Q,Float32Array);return tr!==void 0?{C:fr,gpuTimeMs:tr}:{C:fr}}finally{h||d(ur),b||d(sr),x||d(Z),d(rr)}}var Ha=32,Ua=32,Oa=64,Va=64,Ka=36;async function po(a,e,r,o,t,i,s,l,n,u,f,m="row-major"){let w=s instanceof H,c=u instanceof H;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(e!=="lower"&&e!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(r!=="no-transpose"&&r!=="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(t)||!Number.isInteger(l)||!Number.isInteger(f))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&&!(u 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(o<0||t<0)throw new Error("n and k must be non-negative.");if(o===0)return c?{}:{C:u};let p=w?s.layout:m,g=c?u.layout:m,h=p==="column-major"?t:o,b=p==="column-major"?o:t,x=r==="no-transpose"?h:b,_=r==="no-transpose"?b:h;if(l<_)throw new Error(`lda must be >= ${p==="column-major"?"rows":"cols"} of A as stored.`);if(w){if(l!==s.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");let[W,$]=r==="no-transpose"?[o,t]:[t,o];if(s.rows<W||s.cols<$)throw new Error("A is too small for the given n, k, and trans.")}else if(s.length<(x-1)*l+_)throw new Error("A does not have enough elements for the given dimensions and lda.");if(f<o)throw new Error("ldc must be >= n.");if(c){if(f!==u.lda)throw new Error("ldc must match C.lda when C is a GpuMatrix.");if(u.rows<o||u.cols<o)throw new Error("C is too small for the given n.")}else if(u.length<(o-1)*f+o)throw new Error("C does not have enough elements for the given dimensions and ldc.");let y=r;p==="column-major"&&(y=y==="no-transpose"?"transpose":"no-transpose");let A=y==="no-transpose"?"transpose":"no-transpose",P=e;g==="column-major"&&([y,A]=[A==="no-transpose"?"transpose":"no-transpose",y==="no-transpose"?"transpose":"no-transpose"],P=P==="lower"?"upper":"lower");let E=Math.ceil(o/Va),T=Math.ceil(o/Oa),D=E*T>=Ka,R=await G(a,D?"sgemmtr_large":"sgemmtr_small"),j=w?s._buf:v(s,"ssyrk-A",!1),F=c?u._buf:v(u,"ssyrk-C",!0),V=w?er(j.size,"ssyrk-B",GPUBufferUsage.COPY_DST):v(s,"ssyrk-B",!1),K=L([{value:o,type:"u32"},{value:o,type:"u32"},{value:t,type:"u32"},{value:i,type:"f32"},{value:n,type:"f32"},{value:l,type:"u32"},{value:l,type:"u32"},{value:f,type:"u32"},{value:y==="transpose"?1:0,type:"u32"},{value:A==="transpose"?1:0,type:"u32"},{value:P==="upper"?1:0,type:"u32"}],"ssyrk-params");try{let W=B(R.getBindGroupLayout(0),[j,V,F,K]),$=D?{x:Math.min(E,a.limits.maxComputeWorkgroupsPerDimension),y:Math.min(T,a.limits.maxComputeWorkgroupsPerDimension)}:{x:Math.min(Math.ceil(o/Ua),a.limits.maxComputeWorkgroupsPerDimension),y:Math.min(Math.ceil(o/Ha),a.limits.maxComputeWorkgroupsPerDimension)},{commandEncoder:Y,querySet:J,passDescriptor:nr}=vr();w&&Y.copyBufferToBuffer(j,0,V,0,j.size),ar(Y,R,W,$,nr);let ur=br(Y,J),sr=c?null:N(Y,F);M(Y);let Z=await S(ur);if(c)return Z!==void 0?{gpuTimeMs:Z}:{};let rr=await k(sr,Float32Array);return Z!==void 0?{C:rr,gpuTimeMs:Z}:{C:rr}}finally{w||d(j),d(V),c||d(F),d(K)}}var za=32,qa=32,Ya=64,Xa=64,Qa=36;async function wo(a,e,r,o,t,i,s,l,n,u,f,m,w,c="row-major"){let p=s instanceof H,g=n instanceof H,h=m instanceof H;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(e!=="lower"&&e!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(r!=="no-transpose"&&r!=="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 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(o)||!Number.isInteger(t)||!Number.isInteger(l)||!Number.isInteger(u)||!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(!h&&!(m instanceof Float32Array))throw new Error("C must be a Float32Array or GpuMatrix.");if((p||g)&&!h)throw new Error("C must be a GpuMatrix when A or B is a GpuMatrix.");if(h&&(!p||!g))throw new Error("A and B must be GpuMatrix when C is a GpuMatrix.");if(o<0||t<0)throw new Error("n and k must be non-negative.");if(o===0)return h?{}:{C:m};let b=p?s.layout:c,x=g?n.layout:c,_=h?m.layout:c,y=b==="column-major"?t:o,A=b==="column-major"?o:t,P=r==="no-transpose"?y:A,E=r==="no-transpose"?A:y;if(l<E)throw new Error(`lda must be >= ${b==="column-major"?"rows":"cols"} of A as stored.`);if(p){if(l!==s.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");let[X,Q]=r==="no-transpose"?[o,t]:[t,o];if(s.rows<X||s.cols<Q)throw new Error("A is too small for the given n, k, and trans.")}else if(s.length<(P-1)*l+E)throw new Error("A does not have enough elements for the given dimensions and lda.");let T=x==="column-major"?t:o,D=x==="column-major"?o:t,R=r==="no-transpose"?T:D,j=r==="no-transpose"?D:T;if(u<j)throw new Error(`ldb must be >= ${x==="column-major"?"rows":"cols"} of B as stored.`);if(g){if(u!==n.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");let[X,Q]=r==="no-transpose"?[o,t]:[t,o];if(n.rows<X||n.cols<Q)throw new Error("B is too small for the given n, k, and trans.")}else if(n.length<(R-1)*u+j)throw new Error("B does not have enough elements for the given dimensions and ldb.");if(w<o)throw new Error("ldc must be >= n.");if(h){if(w!==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)*w+o)throw new Error("C does not have enough elements for the given dimensions and ldc.");let F=r;b==="column-major"&&(F=F==="no-transpose"?"transpose":"no-transpose");let V=r;x==="column-major"&&(V=V==="no-transpose"?"transpose":"no-transpose");let K=_==="column-major"?e==="lower"?"upper":"lower":e,W=X=>X==="no-transpose"?"transpose":"no-transpose";function $(X,Q,tr,fr,or,ir){let dr=X,Br=W(fr);return _!=="column-major"?{transX:dr,X:Q,ldX:tr,transY:Br,Y:or,ldY:ir}:{transX:W(Br),X:or,ldX:ir,transY:W(dr),Y:Q,ldY:tr}}let Y=Math.ceil(o/Xa),J=Math.ceil(o/Ya),nr=Y*J>=Qa,ur=await G(a,nr?"sgemmtr_large":"sgemmtr_small"),sr=nr?{x:Math.min(Y,a.limits.maxComputeWorkgroupsPerDimension),y:Math.min(J,a.limits.maxComputeWorkgroupsPerDimension)}:{x:Math.min(Math.ceil(o/qa),a.limits.maxComputeWorkgroupsPerDimension),y:Math.min(Math.ceil(o/za),a.limits.maxComputeWorkgroupsPerDimension)},Z=p?s._buf:v(s,"ssyr2k-A",!1),rr=g?n._buf:v(n,"ssyr2k-B",!1),U=h?m._buf:v(m,"ssyr2k-C",!0),q=null,z=null;try{let X=$(F,Z,l,V,rr,u),Q=$(V,rr,u,F,Z,l),tr=(Er,wr)=>L([{value:o,type:"u32"},{value:o,type:"u32"},{value:t,type:"u32"},{value:i,type:"f32"},{value:wr,type:"f32"},{value:Er.ldX,type:"u32"},{value:Er.ldY,type:"u32"},{value:w,type:"u32"},{value:Er.transX==="transpose"?1:0,type:"u32"},{value:Er.transY==="transpose"?1:0,type:"u32"},{value:K==="upper"?1:0,type:"u32"}],"ssyr2k-params");q=tr(X,f),z=tr(Q,1);let fr=B(ur.getBindGroupLayout(0),[X.X,X.Y,U,q]),or=B(ur.getBindGroupLayout(0),[Q.X,Q.Y,U,z]),{commandEncoder:ir,querySet:dr}=vr(),Br=dr?{timestampWrites:{querySet:dr,beginningOfPassWriteIndex:0}}:void 0,yr=dr?{timestampWrites:{querySet:dr,endOfPassWriteIndex:1}}:void 0;ar(ir,ur,fr,sr,Br),ar(ir,ur,or,sr,yr);let _r=br(ir,dr),gr=h?null:N(ir,U);M(ir);let pr=await S(_r);if(h)return pr!==void 0?{gpuTimeMs:pr}:{};let cr=await k(gr,Float32Array);return pr!==void 0?{C:cr,gpuTimeMs:pr}:{C:cr}}finally{p||d(Z),g||d(rr),h||d(U),q&&d(q),z&&d(z)}}var Za=32,$a=32,Ja=64,ri=64,ei=36,go=8;async function bo(a,e,r,o,t,i,s,l,n,u,f,m,w,c="row-major"){let p=s instanceof H,g=n instanceof H,h=m instanceof H;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(e!=="left"&&e!=="right")throw new Error("side must be 'left' or 'right'.");if(r!=="lower"&&r!=="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 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(o)||!Number.isInteger(t)||!Number.isInteger(l)||!Number.isInteger(u)||!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(!h&&!(m instanceof Float32Array))throw new Error("C must be a Float32Array or GpuMatrix.");if((p||g)&&!h)throw new Error("C must be a GpuMatrix when A or B is a GpuMatrix.");if(h&&(!p||!g))throw new Error("A and B must be GpuMatrix when C is a GpuMatrix.");if(o<0||t<0)throw new Error("m and n must be non-negative.");if(o===0||t===0)return h?{}:{C:m};let b=p?s.layout:c,x=g?n.layout:c,_=h?m.layout:c,y=e==="left"?o:t;if(l<y)throw new Error("lda must be >= "+(e==="left"?"m":"n")+".");if(p){if(l!==s.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(s.rows<y||s.cols<y)throw new Error("A is too small for the given m/n and side.")}else if(s.length<(y-1)*l+y)throw new Error("A does not have enough elements for the given dimensions and lda.");let A=x==="column-major"?t:o,P=x==="column-major"?o:t;if(u<P)throw new Error(`ldb must be >= ${x==="column-major"?"rows":"cols"} of B as stored.`);if(g){if(u!==n.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");if(n.rows<o||n.cols<t)throw new Error("B is too small for the given m and n.")}else if(n.length<(A-1)*u+P)throw new Error("B does not have enough elements for the given dimensions and ldb.");let E=_==="column-major"?t:o,T=_==="column-major"?o:t;if(w<T)throw new Error(`ldc must be >= ${_==="column-major"?"rows":"cols"} of C as stored.`);if(h){if(w!==m.lda)throw new Error("ldc must match C.lda when C is a GpuMatrix.");if(m.rows<o||m.cols<t)throw new Error("C is too small for the given m and n.")}else if(m.length<(E-1)*w+T)throw new Error("C does not have enough elements for the given dimensions and ldc.");let D=b==="column-major"?r==="lower"?"upper":"lower":r,R=x==="column-major"?"transpose":"no-transpose",j="no-transpose",F=o,V=t,K=y,W=e==="left"?j:R,$=e==="left"?R:j,Y=ir=>ir==="no-transpose"?"transpose":"no-transpose",J=e==="right";_==="column-major"&&([W,$]=[Y($),Y(W)],J=!J,[F,V]=[V,F]);let nr=y,ur=Math.ceil(V/ri),sr=Math.ceil(F/Ja),Z=ur*sr>=ei,rr=await G(a,Z?"sgemm_large":"sgemm_small"),U=await G(a,"symmetrize"),q=Z?{x:Math.min(ur,a.limits.maxComputeWorkgroupsPerDimension),y:Math.min(sr,a.limits.maxComputeWorkgroupsPerDimension)}:{x:Math.min(Math.ceil(V/$a),a.limits.maxComputeWorkgroupsPerDimension),y:Math.min(Math.ceil(F/Za),a.limits.maxComputeWorkgroupsPerDimension)},z=p?s._buf:v(s,"ssymm-A",!1),X=g?n._buf:v(n,"ssymm-B",!1),Q=h?m._buf:v(m,"ssymm-C",!0),tr=er(y*nr*4,"ssymm-Adense"),fr=null,or=null;try{fr=L([{value:y,type:"u32"},{value:l,type:"u32"},{value:nr,type:"u32"},{value:D==="upper"?1:0,type:"u32"}],"ssymm-sym-params");let ir=B(U.getBindGroupLayout(0),[z,tr,fr]),dr=J?X:tr,Br=J?u:nr,yr=J?tr:X;or=L([{value:F,type:"u32"},{value:V,type:"u32"},{value:K,type:"u32"},{value:i,type:"f32"},{value:f,type:"f32"},{value:Br,type:"u32"},{value:J?nr:u,type:"u32"},{value:w,type:"u32"},{value:W==="transpose"?1:0,type:"u32"},{value:$==="transpose"?1:0,type:"u32"}],"ssymm-gemm-params");let gr=B(rr.getBindGroupLayout(0),[dr,yr,Q,or]),{commandEncoder:pr,querySet:cr}=vr(),Er=cr?{timestampWrites:{querySet:cr,beginningOfPassWriteIndex:0}}:void 0,wr=cr?{timestampWrites:{querySet:cr,endOfPassWriteIndex:1}}:void 0;ar(pr,U,ir,{x:Math.ceil(y/go),y:Math.ceil(y/go)},Er),ar(pr,rr,gr,q,wr);let kr=br(pr,cr),Nr=h?null:N(pr,Q);M(pr);let Lr=await S(kr);if(h)return Lr!==void 0?{gpuTimeMs:Lr}:{};let Fr=await k(Nr,Float32Array);return Lr!==void 0?{C:Fr,gpuTimeMs:Lr}:{C:Fr}}finally{p||d(z),g||d(X),h||d(Q),d(tr),fr&&d(fr),or&&d(or)}}var ti=32,oi=32,ai=64,ii=64,si=36,ho=8;async function xo(a,e,r,o,t,i,s,l,n,u,f,m,w="row-major"){let c=n instanceof H,p=f instanceof H,g=t==="unit";if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(e!=="left"&&e!=="right")throw new Error("side must be 'left' or 'right'.");if(r!=="lower"&&r!=="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(!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 l!="number")throw new Error("alpha must be a number.");if(Number.isNaN(l))throw new Error("alpha must not be NaN.");if(!Number.isFinite(l))throw new Error("alpha must be finite.");if(!Number.isInteger(i)||!Number.isInteger(s)||!Number.isInteger(u)||!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&&!(f 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:f};let h=c?n.layout:w,b=p?f.layout:w,x=e==="left"?i:s;if(u<x)throw new Error("lda must be >= "+(e==="left"?"m":"n")+".");if(c){if(u!==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)*u+x)throw new Error("A does not have enough elements for the given dimensions and lda.");let _=b==="column-major"?s:i,y=b==="column-major"?i:s;if(m<y)throw new Error(`ldb must be >= ${b==="column-major"?"rows":"cols"} of B as stored.`);if(p){if(m!==f.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");if(f.rows<i||f.cols<s)throw new Error("B is too small for the given m and n.")}else if(f.length<(_-1)*m+y)throw new Error("B does not have enough elements for the given dimensions and ldb.");let A=h==="column-major"?r==="lower"?"upper":"lower":r,P=h==="column-major"?o==="no-transpose"?"transpose":"no-transpose":o,E=b==="column-major"?"transpose":"no-transpose",T="no-transpose",D=i,R=s,j=x,F=e==="left"?T:E,V=e==="left"?E:T,K=fr=>fr==="no-transpose"?"transpose":"no-transpose",W=e==="right";b==="column-major"&&([F,V]=[K(V),K(F)],W=!W,[D,R]=[R,D]);let $=x,Y=Math.ceil(R/ii),J=Math.ceil(D/ai),nr=Y*J>=si,ur=await G(a,nr?"sgemm_large":"sgemm_small"),sr=await G(a,"triangularize"),Z=nr?{x:Math.min(Y,a.limits.maxComputeWorkgroupsPerDimension),y:Math.min(J,a.limits.maxComputeWorkgroupsPerDimension)}:{x:Math.min(Math.ceil(R/oi),a.limits.maxComputeWorkgroupsPerDimension),y:Math.min(Math.ceil(D/ti),a.limits.maxComputeWorkgroupsPerDimension)},rr=c?n._buf:v(n,"strmm-A",!1),U=p?f._buf:v(f,"strmm-B",!0),q=er(x*$*4,"strmm-Adense"),z=er(_*m*4,"strmm-out",GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST),X=null,Q=null,tr=!1;try{X=L([{value:x,type:"u32"},{value:u,type:"u32"},{value:$,type:"u32"},{value:A==="upper"?1:0,type:"u32"},{value:P==="transpose"?1:0,type:"u32"},{value:g?1:0,type:"u32"}],"strmm-tri-params");let fr=B(sr.getBindGroupLayout(0),[rr,q,X]),or=W?U:q,ir=W?m:$,dr=W?q:U;Q=L([{value:D,type:"u32"},{value:R,type:"u32"},{value:j,type:"u32"},{value:l,type:"f32"},{value:0,type:"f32"},{value:ir,type:"u32"},{value:W?$:m,type:"u32"},{value:m,type:"u32"},{value:F==="transpose"?1:0,type:"u32"},{value:V==="transpose"?1:0,type:"u32"}],"strmm-gemm-params");let yr=B(ur.getBindGroupLayout(0),[or,dr,z,Q]),{commandEncoder:_r,querySet:gr}=vr();_r.copyBufferToBuffer(U,0,z,0,Math.min(U.size,z.size));let pr=gr?{timestampWrites:{querySet:gr,beginningOfPassWriteIndex:0}}:void 0,cr=gr?{timestampWrites:{querySet:gr,endOfPassWriteIndex:1}}:void 0;ar(_r,sr,fr,{x:Math.ceil(x/ho),y:Math.ceil(x/ho)},pr),ar(_r,ur,yr,Z,cr);let Er=br(_r,gr),wr=p?null:N(_r,z);M(_r);let kr=await S(Er);if(p)return d(f._buf),f._buf=z,tr=!0,kr!==void 0?{gpuTimeMs:kr}:{};let Nr=await k(wr,Float32Array);return kr!==void 0?{B:Nr,gpuTimeMs:kr}:{B:Nr}}finally{c||d(rr),p||d(U),d(q),tr||d(z),X&&d(X),Q&&d(Q)}}var hr=64,vo=32,yo=32,_o=64,Bo=64,Eo=36;async function Ao(a,e,r,o,t,i,s,l,n,u,f,m,w="row-major"){let c=n instanceof H,p=f instanceof H,g=t==="unit";if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(e!=="left"&&e!=="right")throw new Error("side must be 'left' or 'right'.");if(r!=="lower"&&r!=="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(!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 l!="number")throw new Error("alpha must be a number.");if(Number.isNaN(l))throw new Error("alpha must not be NaN.");if(!Number.isFinite(l))throw new Error("alpha must be finite.");if(!Number.isInteger(i)||!Number.isInteger(s)||!Number.isInteger(u)||!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&&!(f 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:f};let h=c?n.layout:w,b=p?f.layout:w,x=e==="left"?i:s;if(u<x)throw new Error("lda must be >= "+(e==="left"?"m":"n")+".");if(c){if(u!==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)*u+x)throw new Error("A does not have enough elements for the given dimensions and lda.");let _=b==="column-major"?s:i,y=b==="column-major"?i:s;if(m<y)throw new Error(`ldb must be >= ${b==="column-major"?"rows":"cols"} of B as stored.`);if(p){if(m!==f.lda)throw new Error("ldb must match B.lda when B is a GpuMatrix.");if(f.rows<i||f.cols<s)throw new Error("B is too small for the given m and n.")}else if(f.length<(_-1)*m+y)throw new Error("B does not have enough elements for the given dimensions and ldb.");let A=h==="column-major"?r==="lower"?"upper":"lower":r,P=h==="column-major"?o==="no-transpose"?"transpose":"no-transpose":o,E=e==="left"?s:i,T=e==="left",D=P==="no-transpose"==(A==="lower"),R=e==="left"?D:!D,j=[];for(let U=0;U<x;U+=hr)j.push(U);R||j.reverse();let F=j.length,V=await G(a,"strsv_invert_block"),K=await G(a,"block_transfer"),W=await G(a,"sscal"),$=c?n._buf:v(n,"strsm-A",!1),Y=p?f._buf:v(f,"strsm-B",!0),J=er(F*hr*hr*4,"strsm-Ainv"),nr=[],ur=[];function sr(U,q){let z=er(U,q);return ur.push(z),z}function Z(U,q){let z=L(U,q);return nr.push(z),z}let rr=(_-1)*m+y;try{let U=null;if(l!==1){let gr=Z([{value:rr,type:"u32"},{value:l,type:"f32"},{value:1,type:"u32"}],"strsm-scale-params");U=B(W.getBindGroupLayout(0),[Y,gr])}let q=Z([{value:x,type:"u32"},{value:u,type:"u32"},{value:P==="transpose"?1:0,type:"u32"},{value:A==="upper"?1:0,type:"u32"},{value:g?1:0,type:"u32"}],"strsm-invert-params"),z=B(V.getBindGroupLayout(0),[$,J,q]),X=sr(hr*E*4,"strsm-Bblock"),Q=sr(hr*E*4,"strsm-Xblock"),tr=sr(x*hr*4,"strsm-Aoff"),fr=sr(x*E*4,"strsm-delta"),{commandEncoder:or,querySet:ir}=vr();if(l===0){let gr=ir?{timestampWrites:{querySet:ir,beginningOfPassWriteIndex:0,endOfPassWriteIndex:1}}:void 0;ar(or,W,U,mr(rr),gr)}else{U&&ar(or,W,U,mr(rr)),ar(or,V,z,{x:hr,y:F},ir?{timestampWrites:{querySet:ir,beginningOfPassWriteIndex:0}}:void 0);for(let pr=0;pr<j.length;pr++){let cr=j[pr],Er=Math.min(cr+hr,x),wr=Er-cr,kr=cr/hr,Nr=pr===j.length-1,Lr=Z([{value:cr,type:"u32"},{value:wr,type:"u32"},{value:0,type:"u32"},{value:E,type:"u32"},{value:m,type:"u32"},{value:b==="column-major"?1:0,type:"u32"},{value:T?1:0,type:"u32"},{value:2,type:"u32"}],"strsm-gather-B-params"),Fr=B(K.getBindGroupLayout(0),[X,Y,Lr]);ar(or,K,Fr,mr(wr,E));{let Sr=wr,Mr=E,zr=wr,Tr=Math.ceil(Mr/Bo),jr=Math.ceil(Sr/_o),Cr=Tr*jr>=Eo,Wr=await G(a,Cr?"sgemm_large":"sgemm_small"),qr=Z([{value:Sr,type:"u32"},{value:Mr,type:"u32"},{value:zr,type:"u32"},{value:1,type:"f32"},{value:0,type:"f32"},{value:hr,type:"u32"},{value:E,type:"u32"},{value:E,type:"u32"},{value:e==="right"?1:0,type:"u32"},{value:0,type:"u32"}],"strsm-apply-params"),Yr=B(Wr.getBindGroupLayout(0),[{buffer:J,offset:kr*hr*hr*4,size:hr*hr*4},X,Q,qr]),Xr=Cr?{x:Math.min(Tr,a.limits.maxComputeWorkgroupsPerDimension),y:Math.min(jr,a.limits.maxComputeWorkgroupsPerDimension)}:{x:Math.min(Math.ceil(Mr/yo),a.limits.maxComputeWorkgroupsPerDimension),y:Math.min(Math.ceil(Sr/vo),a.limits.maxComputeWorkgroupsPerDimension)};ar(or,Wr,Yr,Xr)}let Hr=R?Er:0,re=R?x:cr,ee=Hr<re,Go=Z([{value:cr,type:"u32"},{value:wr,type:"u32"},{value:0,type:"u32"},{value:E,type:"u32"},{value:m,type:"u32"},{value:b==="column-major"?1:0,type:"u32"},{value:T?1:0,type:"u32"},{value:0,type:"u32"}],"strsm-scatter-params"),ko=B(K.getBindGroupLayout(0),[Q,Y,Go]),Po=Nr&&!ee&&ir?{timestampWrites:{querySet:ir,endOfPassWriteIndex:1}}:void 0;if(ar(or,K,ko,mr(wr,E),Po),!ee)continue;let Rr=re-Hr,No=Z([{value:Hr,type:"u32"},{value:Rr,type:"u32"},{value:cr,type:"u32"},{value:wr,type:"u32"},{value:u,type:"u32"},{value:P==="transpose"?1:0,type:"u32"},{value:T?1:0,type:"u32"},{value:2,type:"u32"}],"strsm-gather-A-params"),So=B(K.getBindGroupLayout(0),[tr,$,No]);ar(or,K,So,mr(Rr,wr));{let Sr=Rr,Mr=E,zr=wr,Tr=Math.ceil(Mr/Bo),jr=Math.ceil(Sr/_o),Cr=Tr*jr>=Eo,Wr=await G(a,Cr?"sgemm_large":"sgemm_small"),qr=Z([{value:Sr,type:"u32"},{value:Mr,type:"u32"},{value:zr,type:"u32"},{value:1,type:"f32"},{value:0,type:"f32"},{value:wr,type:"u32"},{value:E,type:"u32"},{value:E,type:"u32"},{value:0,type:"u32"},{value:0,type:"u32"}],"strsm-update-params"),Yr=B(Wr.getBindGroupLayout(0),[tr,Q,fr,qr]),Xr=Cr?{x:Math.min(Tr,a.limits.maxComputeWorkgroupsPerDimension),y:Math.min(jr,a.limits.maxComputeWorkgroupsPerDimension)}:{x:Math.min(Math.ceil(Mr/yo),a.limits.maxComputeWorkgroupsPerDimension),y:Math.min(Math.ceil(Sr/vo),a.limits.maxComputeWorkgroupsPerDimension)};ar(or,Wr,Yr,Xr)}let Mo=Z([{value:Hr,type:"u32"},{value:Rr,type:"u32"},{value:0,type:"u32"},{value:E,type:"u32"},{value:m,type:"u32"},{value:b==="column-major"?1:0,type:"u32"},{value:T?1:0,type:"u32"},{value:1,type:"u32"}],"strsm-scatter-sub-params"),Io=B(K.getBindGroupLayout(0),[fr,Y,Mo]),Lo=Nr&&ir?{timestampWrites:{querySet:ir,endOfPassWriteIndex:1}}:void 0;ar(or,K,Io,mr(Rr,E),Lo)}}let dr=br(or,ir),Br=p?null:N(or,Y);M(or);let yr=await S(dr);if(p)return yr!==void 0?{gpuTimeMs:yr}:{};let _r=await k(Br,Float32Array);return yr!==void 0?{B:_r,gpuTimeMs:yr}:{B:_r}}finally{c||d($),p||d(Y),d(J),d(ur),d(nr)}}return Wo(ni);})();