wgblas 1.1.1 → 1.2.0

package/dist/wgblas.browser.js

@@ -1,4 +1,4 @@
-var wgblas=(()=>{var mt=Object.create;var ur=Object.defineProperty;var dt=Object.getOwnPropertyDescriptor;var pt=Object.getOwnPropertyNames;var gt=Object.getPrototypeOf,wt=Object.prototype.hasOwnProperty;var lr=(a=>typeof require<"u"?require:typeof Proxy<"u"?new Proxy(a,{get:(r,e)=>(typeof require<"u"?require:r)[e]}):a)(function(a){if(typeof require<"u")return require.apply(this,arguments);throw Error('Dynamic require of "'+a+'" is not supported')});var I=(a,r,e)=>()=>{if(e)throw e[0];try{return a&&(r=a(a=0)),r}catch(o){throw e=[o],o}};var yr=(a,r)=>{for(var e in r)ur(a,e,{get:r[e],enumerable:!0})},_r=(a,r,e,o)=>{if(r&&typeof r=="object"||typeof r=="function")for(let t of pt(r))!wt.call(a,t)&&t!==e&&ur(a,t,{get:()=>r[t],enumerable:!(o=dt(r,t))||o.enumerable});return a};var fr=(a,r,e)=>(e=a!=null?mt(gt(a)):{},_r(r||!a||!a.__esModule?ur(e,"default",{value:a,enumerable:!0}):e,a)),bt=a=>_r(ur({},"__esModule",{value:!0}),a);var Ur,Mr=I(()=>{Ur=`// amax reduction: collapses 2*WGS (value, index) pairs into one index.
+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.
 // dispatch: 1 workgroup of WGS threads.
 // partials_val and partials_idx must have exactly 2*WGS entries.
 
@@ -41,7 +41,7 @@ fn reduce(
 
   if (i == 0u) { result[0] = tile_idx[0]; }
 }
-`});var Vr,Tr=I(()=>{Vr=`// sum reduction: collapses 2*WGS partials into one scalar.
+`});var Wr,Nr=W(()=>{Wr=`// sum reduction: collapses 2*WGS partials into one scalar.
 // dispatch: 1 workgroup of WGS threads.
 // partials must have exactly 2*WGS entries.
 
@@ -67,56 +67,47 @@ fn reduce(
 
   if (i == 0u) { result[0] = tile[0]; }
 }
-`});var Hr,Rr=I(()=>{Hr=`// sum reduction (f64): collapses 2*WGS partial [main, aux] pairs into one,
-// using computeSum instead of plain f32 \`+\` (see reduction/sum.wgsl for the
-// f32 original this mirrors).
-// dispatch: 1 workgroup of WGS threads. partialsMain/partialsAux must have
-// exactly 2*WGS entries each.
-//
-// Concatenated after f64add.wgsl by getPipeline (WGSL has no #include),
-// reusing its decode/encode/computeSum and Packed struct \u2014 f64add.wgsl
-// declares no bindings and no entry point of its own (just helper functions),
-// so bindings here start at 0 and the entry point is simply \`reduce_f64\`.
-//
-// partialsAux/result's aux slot are array<u32>, not array<f32> \u2014 aux's bits
-// must never pass through an f32-typed storage slot (NaN-bit-pattern
-// corruption risk, see f64pack.mjs and the Packed struct comment above
-// decode()/encode() in f64add.wgsl).
-
-@group(0) @binding(0) var<storage, read>       partialsMain: array<f32>;
-@group(0) @binding(1) var<storage, read>       partialsAux:  array<u32>;
-@group(0) @binding(2) var<storage, read_write> resultMain:   array<f32, 1>;
-@group(0) @binding(3) var<storage, read_write> resultAux:    array<u32, 1>;
+`});var Mr,Dr=W(()=>{Mr=`// 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).
+
+@group(0) @binding(0) var<storage, read>       partialsHi: array<f32>;
+@group(0) @binding(1) var<storage, read>       partialsLo: array<f32>;
+@group(0) @binding(2) var<storage, read_write> resultHi:   array<f32, 1>;
+@group(0) @binding(3) var<storage, read_write> resultLo:   array<f32, 1>;
 
 const WGS: u32 = 64;
 
-var<workgroup> tile: array<Packed, 64>;
-
-fn addPair(a: Packed, b: Packed) -> Packed {
-  return computeSum(decode(bitcast<u32>(a.main), a.aux), decode(bitcast<u32>(b.main), b.aux));
-}
+var<workgroup> tile: array<DD, 64>;
 
 @compute @workgroup_size(64)
 fn reduce_f64(
   @builtin(local_invocation_id) lid: vec3u,
 ) {
   let i = lid.x;
-  let a = Packed(partialsMain[i], partialsAux[i]);
-  let b = Packed(partialsMain[i + WGS], partialsAux[i + WGS]);
-  tile[i] = addPair(a, b);
+  let a = DD(partialsHi[i], partialsLo[i]);
+  let b = DD(partialsHi[i + WGS], partialsLo[i + WGS]);
+  tile[i] = ddAddProtected(a, b, i);
   workgroupBarrier();
 
+  // ddAddProtected must be called unconditionally by every thread.
   for (var s = WGS / 2u; s > 0u; s >>= 1u) {
-    if (i < s) { tile[i] = addPair(tile[i], tile[i + s]); }
+    let partner = select(i, i + s, i < s);
+    let combined = ddAddProtected(tile[i], tile[partner], i);
+    workgroupBarrier();
+    if (i < s) { tile[i] = combined; }
     workgroupBarrier();
   }
 
   if (i == 0u) {
-    resultMain[0] = tile[0].main;
-    resultAux[0] = tile[0].aux;
+    resultHi[0] = tile[0].hi;
+    resultLo[0] = tile[0].lo;
   }
 }
-`});var Cr,Dr=I(()=>{Cr=`// sscal: x = alpha * x
+`});var Ur,Tr=W(()=>{Ur=`// sscal: x = alpha * x
 
 @group(0) @binding(0) var<storage, read_write> x: array<f32>;
 
@@ -139,7 +130,7 @@ fn main(
     x[id * params.x_inc] = params.alpha * x[id * params.x_inc];
   }
 }
-`});var zr,Or=I(()=>{zr=`// sswap: x <-> y
+`});var Vr,Hr=W(()=>{Vr=`// 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>;
@@ -165,7 +156,7 @@ fn main(
     y[id * params.y_inc] = temp;
   }
 }
-`});var qr,Qr=I(()=>{qr=`// saxpy: y = alpha * x + y
+`});var Cr,Rr=W(()=>{Cr=`// 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>;
@@ -190,7 +181,7 @@ fn main(
     y[id * params.y_inc] = params.alpha * x[id * params.x_inc] + y[id * params.y_inc];
   }
 }
-`});var Kr,Zr=I(()=>{Kr=`// scopy: y = x
+`});var zr,Or=W(()=>{zr=`// scopy: y = x
 
 @group(0) @binding(0) var<storage, read>       x: array<f32>;
 @group(0) @binding(1) var<storage, read_write> y: array<f32>;
@@ -214,7 +205,7 @@ fn main(
     y[id * params.y_inc] = x[id * params.x_inc];
   }
 }
-`});var $r,Xr=I(()=>{$r=`// sdot: result = sum(x[i] * y[i])
+`});var qr,Qr=W(()=>{qr=`// 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>;
@@ -267,7 +258,7 @@ fn main(
 
   if (lid.x == 0u) { partials[wgid.x] = tile[0]; }
 }
-`});var Jr,Yr=I(()=>{Jr=`// sasum: result = sum(|x[i]|)
+`});var Kr,Zr=W(()=>{Kr=`// 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>;
@@ -318,7 +309,7 @@ fn main(
 
   if (lid.x == 0u) { partials[wgid.x] = tile[0]; }
 }
-`});var ee,re=I(()=>{ee=`// snrm2: result = sqrt(sum(x[i] * x[i]))
+`});var $r,Xr=W(()=>{$r=`// 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>;
@@ -374,7 +365,7 @@ fn main(
 
   if (lid.x == 0u) { partials[wgid.x] = tile[0]; }
 }
-`});var ae,te=I(()=>{ae=`// srot: x = c*x + s*y,  y = -s*x + c*y
+`});var Jr,Yr=W(()=>{Jr=`// 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>;
@@ -403,7 +394,7 @@ fn main(
     y[id * params.y_inc] = -params.s * xi + params.c * yi;
   }
 }
-`});var ie,oe=I(()=>{ie=`// srotm: applies modified Givens rotation H to vectors x and y.
+`});var ee,re=W(()=>{ee=`// 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.
@@ -453,7 +444,7 @@ fn main(
     y[id * params.y_inc] = h21 * xi + h22 * yi;
   }
 }
-`});var se,ne=I(()=>{se=`// isamax: returns index of element with largest absolute value
+`});var ae,te=W(()=>{ae=`// 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>;
@@ -535,7 +526,7 @@ fn main(
     partials_idx[wgid.x] = tile_idx[0];
   }
 }
-`});var le,ue=I(()=>{le=`// sgemv_n: y = alpha * A * x + beta * y  (A is m\xD7n row-major, no-transpose)
+`});var ie,oe=W(()=>{ie=`// 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.
@@ -610,7 +601,7 @@ fn main(
     workgroupBarrier();
   }
 }
-`});var ce,fe=I(()=>{ce=`// sgemv_t: y = alpha * A^T * x + beta * y  (A is m\xD7n row-major, transposed)
+`});var se,ne=W(()=>{se=`// 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)
@@ -675,7 +666,7 @@ fn main(
     y[yi] = params.alpha * (acc0 + acc1 + acc2 + acc3) + params.beta * y[yi];
   }
 }
-`});var de,me=I(()=>{de=`// ssymv: y = alpha * A * x + beta * y
+`});var le,ue=W(()=>{le=`// 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
@@ -744,7 +735,7 @@ fn main(
     }
   }
 }
-`});var ge,pe=I(()=>{ge=`// strmv: y = op(A) * x
+`});var ce,fe=W(()=>{ce=`// 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.
@@ -847,7 +838,7 @@ fn main(
     }
   }
 }
-`});var be,we=I(()=>{be=`// sger: A := alpha * x * y^T + A  (rank-1 update, A is m\xD7n general/dense)
+`});var de,me=W(()=>{de=`// 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>;
@@ -895,7 +886,7 @@ fn main(
     }
   }
 }
-`});var xe,he=I(()=>{xe=`// ssyr: A := alpha * x * x^T + A  (symmetric rank-1 update)
+`});var ge,pe=W(()=>{ge=`// 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).
 
@@ -955,7 +946,7 @@ fn main(
     }
   }
 }
-`});var ye,ve=I(()=>{ye=`// ssyr2: A := alpha * x * y^T + alpha * y * x^T + A  (symmetric rank-2 update)
+`});var be,we=W(()=>{be=`// 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).
 
@@ -1018,7 +1009,7 @@ fn main(
     }
   }
 }
-`});var Ee,_e=I(()=>{Ee=`// f64add: adds two doubles, each packed as a [main, aux] pair (main: f32
+`});var xe,he=W(()=>{xe=`// 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.
@@ -1299,63 +1290,124 @@ fn computeSum(a: Fields, b: Fields) -> Packed {
   let f = addFields(a, b);
   return encode(f.sign, f.rawExp, f.mantissaHi, f.lo);
 }
-`});var Ae,Ge=I(()=>{Ae=`// dasum: result = sum(|x[i]|)
-// pass 1 dispatches exactly 2 * WGS workgroups; pass 2 uses reduction/sumF64.wgsl.
-// Same structure as sasum.wgsl \u2014 every value is now a [main, aux] pair
-// (see src/util/f64pack.mjs) and every \`+\`/\`+=\` is computeSum via addPair
-// instead of plain f32 addition. Concatenated after f64add.wgsl by
-// getPipeline (WGSL has no #include), reusing its decode/encode/computeSum/
-// addFields and Packed struct \u2014 f64add.wgsl declares no bindings and no entry
-// point of its own (just helper functions), so bindings here start at 0 and
-// the entry point is simply \`dasum_main\`.
+`});var ye,ve=W(()=>{ye=`// Double-double arithmetic via Dekker's algorithm \u2014 an alternative to
+// f64add.wgsl's bit-exact IEEE-754 emulation. Doesn't touch that path.
 //
-// xAux/partialsAux are array<u32>, not array<f32> \u2014 aux's bits must never
-// pass through an f32-typed storage slot (NaN-bit-pattern corruption risk,
-// see f64pack.mjs and the Packed struct comment above decode()/encode() in
-// f64add.wgsl); Packed (from f64add.wgsl) keeps aux as u32 in registers/
-// workgroup memory too.
+// A double-double number is a pair (hi, lo) of f32 with hi+lo approximating
+// a higher-precision value, hi holding the leading bits and lo the rounding
+// error hi lost. ~48 bits of mantissa vs f32's 24, less than real f64's 52.
 //
-// Per-thread accumulation (acc0..acc3) stays in DECODED Fields form for the
-// entire strided loop below, via addFields \u2014 not re-encoded to Packed and
-// re-decoded on every single element like a naive version would. Only the
-// freshly-loaded x[idx] needs decoding each iteration (unavoidable, it's new
-// data every time); the running total never leaves Fields form until the
-// four accumulators are combined and encoded exactly once, right before
-// writing into workgroup-shared \`tile\`. The cross-thread reduction tree
-// after that still goes through Packed per level (unavoidable \u2014 each level
-// combines values that live in different threads' registers via shared
-// memory), but that's a fixed 6 levels regardless of n, unlike the strided
-// loop above whose iteration count scales with n.
-
-@group(0) @binding(0) var<storage, read>       xMain:        array<f32>;
-@group(0) @binding(1) var<storage, read>       xAux:         array<u32>;
-@group(0) @binding(2) var<storage, read_write> partialsMain: array<f32>;
-@group(0) @binding(3) var<storage, read_write> partialsAux:  array<u32>;
-@group(0) @binding(4) var<uniform>             params:       Params;
+// No bindings, no entry point \u2014 a helper library, concatenated with a
+// consumer's own bindings/entry point by getPipeline (WGSL has no #include).
 
-struct Params {
-  n:     u32,
-  x_inc: u32,
+struct DD {
+  hi: f32,
+  lo: f32,
 }
 
-const WGS: u32 = 64;
+// |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.
+fn ddAbs(a: DD) -> DD {
+  if (a.hi < 0.0) {
+    return DD(-a.hi, -a.lo);
+  }
+  return a;
+}
+
+// \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
+//
+// twoSum/fastTwoSum's error term \`e\` should be nonzero (that's the point \u2014
+// \`s\` is rounded). A front-end-optimizer bug zeros it anyway, on both NVIDIA
+// and Mesa (ANV + llvmpipe), via two different mechanisms needing two fixes:
+// bitcast-based subtraction (\`fsub\`/\`negf\`, fixes NVIDIA) and materializing
+// the sum through workgroup memory + workgroupBarrier() (fixes Mesa). Only
+// both together (ddAddProtected) is verified correct everywhere \u2014 the plain
+// twoSum/fastTwoSum/ddAdd below are reference-only, not safe to use.
+fn negf(x: f32) -> f32 {
+  return bitcast<f32>(bitcast<u32>(x) ^ 0x80000000u);
+}
+fn fsub(a: f32, b: f32) -> f32 {
+  return a + negf(b);
+}
+
+// Knuth/M\xF8ller's TwoSum: s = fl(a+b), e = exact rounding error, a+b == s+e.
+// Works for any a, b. UNPROTECTED \u2014 see header above.
+fn twoSum(a: f32, b: f32) -> DD {
+  let s = a + b;
+  let v = s - a;
+  let e = (a - (s - v)) + (b - v);
+  return DD(s, e);
+}
+
+// Dekker's Fast-Two-Sum: same contract, but only correct when |a| >= |b|.
+// UNPROTECTED \u2014 see header above.
+fn fastTwoSum(a: f32, b: f32) -> DD {
+  let s = a + b;
+  let e = b - (s - a);
+  return DD(s, e);
+}
 
-var<workgroup> tile: array<Packed, 64>;
+// Double-double addition (Dekker's Add2). UNPROTECTED \u2014 see header above.
+fn ddAdd(a: DD, b: DD) -> DD {
+  let s = twoSum(a.hi, b.hi);
+  let loSum = a.lo + b.lo;
+  return fastTwoSum(s.hi, s.lo + loSum);
+}
+
+// \u2500\u2500 Protected variants \u2014 use these \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
+//
+// Bitcast subtraction + workgroup-barrier materialization, verified correct
+// on all three backends tested. Costs a real barrier: fine for O(1)-per-
+// thread or O(log n) reduction use, not a long per-element loop. A
+// workgroupBarrier() requires uniform control flow, so:
+//   - \`threadSlot\` must be unique per concurrent caller (e.g. local_invocation_index).
+//   - Every thread in the workgroup must call this the same number of times
+//     \u2014 including ones whose result gets discarded. Compute unconditionally;
+//     only the write-back should be conditional.
+var<workgroup> dekkerScratch: array<f32, 64>;
+
+fn twoSumProtected(a: f32, b: f32, threadSlot: u32) -> DD {
+  dekkerScratch[threadSlot] = a + b;
+  workgroupBarrier();
+  let s = dekkerScratch[threadSlot];
+  let v = fsub(s, a);
+  let e = fsub(a, fsub(s, v)) + fsub(b, v);
+  return DD(s, e);
+}
 
-// a + b, where a/b are [main, aux] pairs \u2014 computeSum takes decoded Fields.
-// Only used for the cross-thread reduction tree below; the per-thread
-// strided loop uses addFields directly instead (see module comment).
-fn addPair(a: Packed, b: Packed) -> Packed {
-  return computeSum(decode(bitcast<u32>(a.main), a.aux), decode(bitcast<u32>(b.main), b.aux));
+fn fastTwoSumProtected(a: f32, b: f32, threadSlot: u32) -> DD {
+  dekkerScratch[threadSlot] = a + b;
+  workgroupBarrier();
+  let s = dekkerScratch[threadSlot];
+  let e = fsub(b, fsub(s, a));
+  return DD(s, e);
+}
+
+// Protected double-double addition \u2014 same contract as ddAdd, but exact.
+fn ddAddProtected(a: DD, b: DD, threadSlot: u32) -> DD {
+  let s = twoSumProtected(a.hi, b.hi, threadSlot);
+  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.
 
-// |x| for a packed double is abs(main) with aux untouched \u2014 only main's
-// sign bit carries the double's sign (see fieldsToPacked() in f64pack.mjs).
-// Returns decoded Fields directly (not Packed) for the per-thread loop.
-fn absFields(idx: u32) -> Fields {
-  return decode(bitcast<u32>(abs(xMain[idx])), xAux[idx]);
+@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> partialsHi: array<f32>;
+@group(0) @binding(3) var<storage, read_write> partialsLo: array<f32>;
+@group(0) @binding(4) var<uniform>             params:     Params;
+
+struct Params {
+  n:     u32,
+  x_inc: u32,
 }
 
+const WGS: u32 = 64;
+
+var<workgroup> tile: array<DD, 64>;
+
 @compute @workgroup_size(64)
 fn dasum_main(
   @builtin(global_invocation_id) gid:    vec3u,
@@ -1363,41 +1415,65 @@ fn dasum_main(
   @builtin(workgroup_id)         wgid:   vec3u,
   @builtin(num_workgroups)       num_wg: vec3u,
 ) {
-  var acc0: Fields = Fields(0u, 0u, 0u, 0u);
-  var acc1: Fields = Fields(0u, 0u, 0u, 0u);
-  var acc2: Fields = Fields(0u, 0u, 0u, 0u);
-  var acc3: Fields = Fields(0u, 0u, 0u, 0u);
+  var acc0 = DD(0.0, 0.0);
+  var acc1 = DD(0.0, 0.0);
+  var acc2 = DD(0.0, 0.0);
+  var acc3 = DD(0.0, 0.0);
 
   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) {
-    acc0 = addFields(acc0, absFields( id                * params.x_inc));
-    acc1 = addFields(acc1, absFields((id +      stride) * params.x_inc));
-    acc2 = addFields(acc2, absFields((id + 2u * stride) * params.x_inc));
-    acc3 = addFields(acc3, absFields((id + 3u * stride) * params.x_inc));
-  }
-  for (var id = n4_floor + gid.x; id < params.n; id += stride) {
-    acc0 = addFields(acc0, absFields(id * params.x_inc));
-  }
-
-  // Combine the 4 per-thread accumulators in Fields form too \u2014 still no
-  // encode/decode needed, since none of them have touched Packed yet.
-  let combined = addFields(addFields(acc0, acc1), addFields(acc2, acc3));
-  tile[lid.x] = encode(combined.sign, combined.rawExp, combined.mantissaHi, combined.lo);
+  // Same trip count for every thread, but driven by a counter, not \`id\`
+  // itself (ddAddProtected's barrier needs a provably-uniform loop bound).
+  let mainIters = n4_floor / (4u * stride);
+  for (var iter = 0u; iter < mainIters; iter++) {
+    let id =  gid.x + iter * 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;
+    acc0 = ddAddProtected(acc0, ddAbs(DD(xHi[i0], xLo[i0])), lid.x);
+    acc1 = ddAddProtected(acc1, ddAbs(DD(xHi[i1], xLo[i1])), lid.x);
+    acc2 = ddAddProtected(acc2, ddAbs(DD(xHi[i2], xLo[i2])), lid.x);
+    acc3 = ddAddProtected(acc3, ddAbs(DD(xHi[i3], xLo[i3])), lid.x);
+  }
+
+  // Tail is ragged (0-3 extra per thread) \u2014 pad to this workgroup's worst case.
+  let wgBaseGid = wgid.x * WGS;
+  var tailIters = 0u;
+  if (n4_floor + wgBaseGid < params.n) {
+    tailIters = (params.n - 1u - n4_floor - wgBaseGid) / stride + 1u;
+  }
+  for (var iter = 0u; iter < tailIters; iter++) {
+    let id    = n4_floor + gid.x + iter * stride;
+    let valid = id < params.n;
+    let i     = select(0u, id * params.x_inc, valid);
+    let loaded = ddAbs(DD(xHi[i], xLo[i])); // select() has no DD overload
+    let contribution = DD(select(0.0, loaded.hi, valid), select(0.0, loaded.lo, valid));
+    acc0 = ddAddProtected(acc0, contribution, lid.x);
+  }
+
+  let combined01 = ddAddProtected(acc0, acc1, lid.x);
+  let combined23 = ddAddProtected(acc2, acc3, lid.x);
+  tile[lid.x] = ddAddProtected(combined01, combined23, lid.x);
   workgroupBarrier();
 
+  // Inactive threads combine against a throwaway partner and discard it
+  // (ddAddProtected must be called unconditionally by every thread).
   for (var s = WGS / 2u; s > 0u; s >>= 1u) {
-    if (lid.x < s) { tile[lid.x] = addPair(tile[lid.x], tile[lid.x + s]); }
+    let partner = select(lid.x, lid.x + s, lid.x < s);
+    let combined = ddAddProtected(tile[lid.x], tile[partner], lid.x);
+    workgroupBarrier(); // all threads must read tile[] above before any write below
+    if (lid.x < s) { tile[lid.x] = combined; }
     workgroupBarrier();
   }
 
   if (lid.x == 0u) {
-    partialsMain[wgid.x] = tile[0].main;
-    partialsAux[wgid.x] = tile[0].aux;
+    partialsHi[wgid.x] = tile[0].hi;
+    partialsLo[wgid.x] = tile[0].lo;
   }
 }
-`});var ke,Be=I(()=>{ke=`// strsv_invert_block: computes ONE column (workgroup_id.x) of ONE block's
+`});var Ae,Ge=W(()=>{Ae=`// 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
@@ -1506,7 +1582,7 @@ fn strsv_invert_block_main(
     workgroupBarrier();
   }
 }
-`});var Se,Pe=I(()=>{Se=`// strsv_apply_inverse: given a precomputed block inverse (from
+`});var ke,Be=W(()=>{ke=`// 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,
@@ -1553,7 +1629,7 @@ fn strsv_apply_inverse_main(@builtin(local_invocation_id) lid: vec3u) {
   }
   x[(params.blockStart + lid.x) * params.incx] = acc;
 }
-`});var je,Fe=I(()=>{je=`// strsv_update: subtracts a solved block's contribution from every
+`});var Pe,Se=W(()=>{Pe=`// 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.
@@ -1628,7 +1704,7 @@ fn strsv_update_main(
     workgroupBarrier();
   }
 }
-`});var Le={};yr(Le,{shaderSources:()=>Xt});var Xt,Ne=I(()=>{Mr();Tr();Rr();Dr();Or();Qr();Zr();Xr();Yr();re();te();oe();ne();ue();fe();me();pe();we();he();ve();_e();Ge();Be();Pe();Fe();Xt={"reduction/argmax":Ur,"reduction/sum":Vr,"reduction/sumF64":Hr,sscal:Cr,sswap:zr,saxpy:qr,scopy:Kr,sdot:$r,sasum:Jr,snrm2:ee,srot:ae,srotm:ie,isamax:se,sgemv_n:le,sgemv_t:ce,ssymv:de,strmv:ge,sger:be,ssyr:xe,ssyr2:ye,f64add:Ee,dasum:Ae,strsv_invert_block:ke,strsv_apply_inverse:Se,strsv_update:je}});var ea={};yr(ea,{GpuMatrix:()=>H,GpuVector:()=>x,cleanup:()=>Pr,dasum:()=>Ce,gpuName:()=>Sr,init:()=>kr,isamax:()=>Qe,randomFloat32Array:()=>Nr,randomFloat64Array:()=>Wr,randomTriangularFloat32Array:()=>Ir,sasum:()=>De,saxpy:()=>Ue,scopy:()=>Te,sdot:()=>Re,sgemv:()=>Ke,sger:()=>et,snrm2:()=>ze,srot:()=>qe,srotm:()=>Ze,sscal:()=>Ie,sswap:()=>Me,ssymv:()=>Xe,ssyr:()=>tt,ssyr2:()=>at,strmv:()=>$e,strsv:()=>rt});function Er(a,r){return r?a.features.has("timestamp-query")?{requiredFeatures:["timestamp-query"]}:(console.warn("timestamp-query not supported on this device \u2014 benchmark mode disabled."),{}):{}}function Gr(){if(!Ar())return{querySet:null,passDescriptor:void 0};let r=V().createQuerySet({type:"timestamp",count:2});return{querySet:r,passDescriptor:{timestampWrites:{querySet:r,beginningOfPassWriteIndex:0,endOfPassWriteIndex:1}}}}function cr(a,r){if(!r)return null;let e=V(),o=e.createBuffer({label:"timestamp-resolve",size:16,usage:GPUBufferUsage.QUERY_RESOLVE|GPUBufferUsage.COPY_SRC});a.resolveQuerySet(r,0,2,o,0);let t=e.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:r}}async function P(a){if(!a)return;let{tsReadBuffer:r,resolveBuffer:e,querySet:o}=a;await r.mapAsync(GPUMapMode.READ);let t=new BigInt64Array(r.getMappedRange().slice());return r.unmap(),r.destroy(),e.destroy(),o.destroy(),Math.max(0,Number(t[1]-t[0]))/1e6}var Z=null,J=null,Br=null,dr=!1;async function kr({powerPreference:a="high-performance",benchmark:r=!1}={}){if(Z)return Z;let e;if(typeof window>"u"){let{create:i,globals:s}=await import("webgpu");Object.assign(globalThis,s),e=i([]),Br=e}else e=navigator.gpu;if(!e)throw new Error("WebGPU not supported in this environment.");if(J=await e.requestAdapter({powerPreference:a})??await e.requestAdapter(),!J)throw new Error("No WebGPU adapter found.");dr=r;let t=[...Er(J,r).requiredFeatures??[]];return Z=await J.requestDevice({requiredFeatures:t}),Z.addEventListener("uncapturederror",i=>{console.error("Uncaptured GPU error:",i.error.message)}),Z}function Pr(){Z&&(Z.destroy(),Z=null),J=null,Br=null,dr=!1}function Sr(){if(!J)throw new Error("WebGPU adapter not initialized \u2014 call init() first.");let{device:a,description:r}=J.info;return{description:r||"unknown",device:a||"unknown"}}function Ar(){return dr}function V(){if(!Z)throw new Error("WebGPU device not initialized \u2014 call init() first.");return Z}function d(...a){a.flat().forEach(r=>r.destroy())}function b(a,r="blas-input",e=!1){let o=V(),t=o.limits.maxStorageBufferBindingSize,i=a.byteLength;if(i>t)throw new Error(`Buffer size ${i} bytes exceeds device limit of ${t} bytes.`);let s=e?GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC:GPUBufferUsage.STORAGE,u=o.createBuffer({label:r,size:i,usage:s,mappedAtCreation:!0}),n=a.constructor;return new n(u.getMappedRange()).set(a),u.unmap(),u}function D(a,r="blas-storage"){return V().createBuffer({label:r,size:a,usage:GPUBufferUsage.STORAGE})}function q(a,r="blas-result"){return V().createBuffer({label:r,size:a,usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC})}function E(a,r){let o=V().createBuffer({label:"blas-readback",size:r.size,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});return a.copyBufferToBuffer(r,0,o,0,r.size),o}function N(a,r="blas-params"){let e=V(),o=a.length*4,t=Math.ceil(o/16)*16,i=new ArrayBuffer(t),s=new DataView(i);a.forEach(({value:n,type:l},f)=>{let c=f*4;if(l==="u32")s.setUint32(c,n,!0);else if(l==="i32")s.setInt32(c,n,!0);else if(l==="f32")s.setFloat32(c,n,!0);else throw new Error(`Unknown param type "${l}". Use "f32", "u32", or "i32".`)});let u=e.createBuffer({label:r,size:t,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});return e.queue.writeBuffer(u,0,i),u}async function _(a,r=Float32Array){try{await a.mapAsync(GPUMapMode.READ);let e=new r(a.getMappedRange().slice());return a.unmap(),e}finally{a.destroy()}}var ht=new ArrayBuffer(8),er=new DataView(ht),Fr=new ArrayBuffer(4),jr=new Uint32Array(Fr),Lr=new Float32Array(Fr);function xt(a){return jr[0]=a>>>0,Lr[0]}function vt(a){return Lr[0]=a,jr[0]}function yt(a,r,e,o){let t=r>>>3,i=r&7,s=o>>>29,u=e<<3|s,n=o&536870911,l=a<<31|t<<23|u,f=i>>>2&1,c=i&3,p=n>>>23,m=n&8388607,g=c<<6|p,w=(f<<31|g<<23|m)>>>0;return[xt(l),w]}function _t(a,r){let e=vt(a);r=r>>>0;let o=e>>>31,t=e>>>23&255,i=e&8388607,s=r>>>31,u=r>>>23&255,n=r&8388607,l=s<<2|u>>>6,f=(u&63)<<23|n,c=t<<3|l,p=i>>>3,g=((i&7)<<29|f)>>>0;return{sign:o,rawExp:c,mantissaHi:p,lo:g}}var Et=2040;function mr(a){er.setFloat64(0,a,!1);let r=er.getUint32(0,!1),e=er.getUint32(4,!1),o=r>>>31,t=r>>>20&2047,i=r&1048575;if(t>=Et)throw new RangeError(`packF64: |${a}| is too large to pack safely (must be finite with magnitude below ~1.4e306); main's bit pattern would itself be NaN/Infinity-shaped and get silently corrupted by any real float32 round-trip`);return yt(o,t,i,e)}function tr(a,r){let{sign:e,rawExp:o,mantissaHi:t,lo:i}=_t(a,r),s=(e<<31|o<<20|t)>>>0;return er.setUint32(0,s,!1),er.setUint32(4,i,!1),er.getFloat64(0,!1)}var x=class a{constructor(r,e,o=Float32Array,t=null){this._buf=r,this._auxBuf=t,this.length=e,this.dtype=o}static from(r){if(r instanceof Float64Array){let o=new Float32Array(r.length),t=new Uint32Array(r.length);for(let u=0;u<r.length;u++){let n=mr(r[u]);o[u]=n[0],t[u]=n[1]}let i=b(o,"gpu-vector-f64-main",!0),s=b(t,"gpu-vector-f64-aux",!0);return new a(i,r.length,Float64Array,s)}if(!(r instanceof Float32Array))throw new Error("GpuVector.from expects a Float32Array or Float64Array.");let e=b(r,"gpu-vector",!0);return new a(e,r.length,r.constructor)}async read(){let r=V(),e=r.createCommandEncoder(),o=E(e,this._buf);if(r.queue.submit([e.finish()]),!this._auxBuf)return _(o,this.dtype);let t=r.createCommandEncoder(),i=E(t,this._auxBuf);r.queue.submit([t.finish()]);let[s,u]=await Promise.all([_(o,Float32Array),_(i,Uint32Array)]),n=new Float64Array(this.length);for(let l=0;l<this.length;l++)n[l]=tr(s[l],u[l]);return n}destroy(){this._buf.destroy(),this._auxBuf&&this._auxBuf.destroy()}};var H=class a{constructor(r,e,o,t,i=null,s="row-major"){this._buf=r,this._auxBuf=i,this.rows=e,this.cols=o,this.lda=t,this.layout=s}static from(r,e,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:e),!(r instanceof Float32Array)&&!(r instanceof Float64Array))throw new Error("GpuMatrix.from expects a Float32Array or Float64Array.");if(!Number.isInteger(e)||e<=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=s?o:e;if(!Number.isInteger(t)||t<u)throw new Error(`lda must be an integer >= ${s?"cols":"rows"}.`);let n=s?e:o;if(r.length<n*t)throw new Error("data does not have enough elements for the given rows, cols, and lda.");if(r instanceof Float64Array){let f=n*t,c=new Float32Array(f),p=new Uint32Array(f);for(let w=0;w<f;w++){let y=mr(r[w]);c[w]=y[0],p[w]=y[1]}let m=b(c,"gpu-matrix-f64-main",!0),g=b(p,"gpu-matrix-f64-aux",!0);return new a(m,e,o,t,g,i)}let l=b(r.subarray(0,n*t),"gpu-matrix",!0);return new a(l,e,o,t,null,i)}async read(){let r=V(),e=r.createCommandEncoder(),o=E(e,this._buf);r.queue.submit([e.finish()]);let t=this.layout!=="column-major",i=t?this.rows:this.cols,s=t?this.cols:this.rows;if(this._auxBuf){let l=r.createCommandEncoder(),f=E(l,this._auxBuf);r.queue.submit([l.finish()]);let[c,p]=await Promise.all([_(o,Float32Array),_(f,Uint32Array)]),m=new Float64Array(i*this.lda);for(let w=0;w<m.length;w++)m[w]=tr(c[w],p[w]);if(this.lda===s)return m;let g=new Float64Array(i*s);for(let w=0;w<i;w++)g.set(m.subarray(w*this.lda,w*this.lda+s),w*s);return g}let u=await _(o,Float32Array);if(this.lda===s)return u;let n=new Float32Array(i*s);for(let l=0;l<i;l++)n.set(u.subarray(l*this.lda,l*this.lda+s),l*s);return n}destroy(){this._buf.destroy(),this._auxBuf&&this._auxBuf.destroy()}};function Nr(a,r=-1,e=1){let o=new Float32Array(a);for(let t=0;t<a;t++)o[t]=r+Math.random()*(e-r);return o}function Wr(a,r=-1,e=1){let o=new Float64Array(a);for(let t=0;t<a;t++)o[t]=r+Math.random()*(e-r);return o}function Ir(a,r,e="lower",o=-1,t=1,i=5,s=15){if(e!=="lower"&&e!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(r<a)throw new Error("lda must be >= n.");let u=new Float32Array(a*r);for(let n=0;n<a;n++){for(let l=0;l<a;l++){if(n===l)continue;(e==="lower"?l<n:l>n)&&(u[n*r+l]=o+Math.random()*(t-o))}u[n*r+n]=i+Math.random()*(s-i)}return u}function A(a,r,e=0){let o=V(),t=r.map((i,s)=>({binding:e+s,resource:i instanceof GPUBuffer?{buffer:i}:i}));return o.createBindGroup({layout:a,entries:t})}var Gt=new WeakMap;function S(a){V().queue.submit([a.finish()])}function pr(){let a=V(),{querySet:r,passDescriptor:e}=Gr();return{commandEncoder:a.createCommandEncoder(),querySet:r,passDescriptor:e}}function ir(a,r,e,o,t){let i=a.beginComputePass(t);i.setPipeline(r),i.setBindGroup(0,e),typeof o=="number"?i.dispatchWorkgroups(o):i.dispatchWorkgroups(o.x,o.y),i.end(),Gt.set(a,i)}function L(a,r,e){let{commandEncoder:o,querySet:t,passDescriptor:i}=pr();ir(o,a,r,e,i);let s=cr(o,t);return{commandEncoder:o,ts:s}}var Jt={},gr=new WeakMap;async function B(a,r,e="main"){gr.has(a)||gr.set(a,new Map);let o=gr.get(a),t=Array.isArray(r)?r:[r],i=`${t.join("+")}::${e}`;return o.has(i)||o.set(i,await Yt(t,e)),o.get(i)}async function $t(a){if(typeof process>"u"||!process.versions?.node){let{shaderSources:r}=await Promise.resolve().then(()=>(Ne(),Le)),e=r[a];if(!e)throw new Error(`Shader "${a}" not found in browser bundle.`);return e}else{let{readFileSync:r}=await import("fs"),{fileURLToPath:e}=await import("url"),{dirname:o,join:t}=await import("path"),i=o(e(Jt.url));return r(t(i,`../shaders/${a}.wgsl`),"utf8")}}async function Yt(a,r="main"){let e=V(),o=a.join("+"),t=(await Promise.all(a.map($t))).join(`
-`),i=e.createShaderModule({label:o,code:t}),u=(await i.getCompilationInfo()).messages.filter(f=>f.type==="error");if(u.length>0)throw new Error(`Shader "${o}" compilation failed:
+`});var Le={};vr(Le,{shaderSources:()=>Ct});var Ct,je=W(()=>{Fr();Nr();Dr();Tr();Hr();Rr();Or();Qr();Zr();Xr();Yr();re();te();oe();ne();ue();fe();me();pe();we();he();ve();_e();Ge();Be();Se();Ct={"reduction/argmax":Ir,"reduction/sum":Wr,"reduction/sumF64":Mr,sscal:Ur,sswap:Vr,saxpy:Cr,scopy:zr,sdot:qr,sasum:Kr,snrm2:$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 n=r==="main"?{module:i}:{module:i,entryPoint:r},l=e.createComputePipeline({label:o,layout:"auto",compute:n});return l._shaderModule=i,l}var ra=64,We=8;function C(a,r){let e=V().limits.maxComputeWorkgroupsPerDimension;return r===void 0?Math.min(Math.ceil(a/ra),e):{x:Math.min(Math.ceil(r/We),e),y:Math.min(Math.ceil(a/We),e)}}async function Ie(a,r,e,o,t){let i=o instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(r)||!Number.isInteger(t))throw new Error("n and incx must be integers.");if(typeof e!="number")throw new Error("alpha must be a number.");if(Number.isNaN(e))throw new Error("alpha must not be NaN.");if(!Number.isFinite(e))throw new Error("alpha must be finite.");if(t<=0)throw new Error("incx must be positive.");if(!(o instanceof Float32Array)&&!(o instanceof x))throw new Error("x must be a Float32Array or GpuVector.");if(r<=0)return i?{}:o;if(o.length<(r-1)*t+1)throw new Error("x does not have enough elements for the given n and incx.");let s=await B(a,"sscal"),u=null,n=null,l=null;try{u=i?o._buf:b(o,"sscal-x",!0),n=N([{value:r,type:"u32"},{value:e,type:"f32"},{value:t,type:"u32"}],"sscal-params");let f=A(s.getBindGroupLayout(0),[u,n]),{commandEncoder:c,ts:p}=L(s,f,C(r));l=i?null:E(c,u),S(c);let m=await P(p);if(i)return m!==void 0?{gpuTimeMs:m}:{};let g=await _(l,Float32Array);return l=null,m!==void 0?{x:g,gpuTimeMs:m}:g}finally{!i&&u&&d(u),n&&d(n),l&&d(l)}}async function Me(a,r,e,o,t,i){let s=e instanceof x,u=t instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(r)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!(e instanceof Float32Array)&&!(e instanceof 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(e.constructor!==t.constructor)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(r<=0)return s?{}:{x:e,y:t};if(e.length<(r-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(r-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=await B(a,"sswap"),l=null,f=null,c=null,p=null,m=null;try{l=s?e._buf:b(e,"sswap-x",!0),f=u?t._buf:b(t,"sswap-y",!0),c=N([{value:r,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"sswap-params");let g=A(n.getBindGroupLayout(0),[l,f,c]),{commandEncoder:w,ts:y}=L(n,g,C(r));p=s?null:E(w,l),m=u?null:E(w,f),S(w);let h=await P(y);if(s&&u)return h!==void 0?{gpuTimeMs:h}:{};let G=await _(p,Float32Array);p=null;let v=await _(m,Float32Array);return m=null,h!==void 0?{x:G,y:v,gpuTimeMs:h}:{x:G,y:v}}finally{!s&&l&&d(l),!u&&f&&d(f),c&&d(c),p&&d(p),m&&d(m)}}async function Ue(a,r,e,o,t,i,s){let u=o instanceof x,n=i instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(r)||!Number.isInteger(t)||!Number.isInteger(s))throw new Error("n, incx, and incy must be integers.");if(typeof e!="number")throw new Error("alpha must be a number.");if(Number.isNaN(e))throw new Error("alpha must not be NaN.");if(!Number.isFinite(e))throw new Error("alpha must be finite.");if(t<=0||s<=0)throw new Error("incx and incy must be positive.");if(!u&&!(o instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!n&&!(i instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(u!==n)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(r<=0)return n?{}:{y:i};if(o.length<(r-1)*t+1)throw new Error("x does not have enough elements for the given n and incx.");if(i.length<(r-1)*s+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,m=null;try{f=u?o._buf:b(o,"saxpy-x",!1),c=n?i._buf:b(i,"saxpy-y",!0),p=N([{value:r,type:"u32"},{value:e,type:"f32"},{value:t,type:"u32"},{value:s,type:"u32"}],"saxpy-params");let g=A(l.getBindGroupLayout(0),[f,c,p]),{commandEncoder:w,ts:y}=L(l,g,C(r));m=n?null:E(w,c),S(w);let h=await P(y);if(n&&u)return h!==void 0?{gpuTimeMs:h}:{};let G=await _(m,Float32Array);return m=null,h!==void 0?{y:G,gpuTimeMs:h}:{y:G}}finally{!u&&f&&d(f),!n&&c&&d(c),p&&d(p),m&&d(m)}}async function Te(a,r,e,o,t,i){let s=e instanceof x,u=t instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(r)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!s&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!u&&!(t instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(s!==u)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(r<=0)return u?{}:{y:t};if(e.length<(r-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(r-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=await B(a,"scopy"),l=null,f=null,c=null,p=null;try{l=s?e._buf:b(e,"scopy-x",!1),f=u?t._buf:b(t,"scopy-y",!0),c=N([{value:r,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"scopy-params");let m=A(n.getBindGroupLayout(0),[l,f,c]),{commandEncoder:g,ts:w}=L(n,m,C(r));p=u?null:E(g,f),S(g);let y=await P(w);if(u&&s)return y!==void 0?{gpuTimeMs:y}:{};let h=await _(p,Float32Array);return p=null,y!==void 0?{y:h,gpuTimeMs:y}:{y:h}}finally{!s&&l&&d(l),!u&&f&&d(f),c&&d(c),p&&d(p)}}var Ve=64;async function Re(a,r,e,o,t,i){let s=e instanceof x,u=t instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(r)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!s&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!u&&!(t instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(s!==u)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(r<=0)return{dot:0};if(e.length<(r-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(r-1)*i+1)throw new Error("y does not have enough elements for the given n and incy.");let n=await B(a,"sdot"),l=await B(a,"reduction/sum"),f=null,c=null,p=null,m=null,g=null,w=null;try{f=s?e._buf:b(e,"sdot-x",!1),c=u?t._buf:b(t,"sdot-y",!1),p=D(2*Ve*4,"sdot-partials"),m=q(4,"sdot-result"),g=N([{value:r,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"sdot-params");let y=A(n.getBindGroupLayout(0),[f,c,p,g]),{commandEncoder:h,ts:G}=L(n,y,2*Ve);S(h);let v=A(l.getBindGroupLayout(0),[p,m]),{commandEncoder:F,ts:k}=L(l,v,1);w=E(F,m),S(F);let j=_(w,Float32Array);w=null;let[W,M,U]=await Promise.all([P(G),P(k),j]);return W!==void 0&&M!==void 0?{dot:U[0],gpuTimeMs:W+M}:{dot:U[0]}}finally{!s&&f&&d(f),!u&&c&&d(c),p&&d(p),m&&d(m),g&&d(g),w&&d(w)}}var He=64;async function De(a,r,e,o){let t=e instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(r)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!t&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(r<=0)return{asum:0};if(e.length<(r-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=await B(a,"sasum"),s=await B(a,"reduction/sum"),u=null,n=null,l=null,f=null,c=null;try{u=t?e._buf:b(e,"sasum-x",!1),n=D(2*He*4,"sasum-partials"),l=q(4,"sasum-result"),f=N([{value:r,type:"u32"},{value:o,type:"u32"}],"sasum-params");let p=A(i.getBindGroupLayout(0),[u,n,f]),{commandEncoder:m,ts:g}=L(i,p,2*He);S(m);let w=A(s.getBindGroupLayout(0),[n,l]),{commandEncoder:y,ts:h}=L(s,w,1);c=E(y,l),S(y);let G=_(c,Float32Array);c=null;let[v,F,k]=await Promise.all([P(g),P(h),G]);return v!==void 0&&F!==void 0?{asum:k[0],gpuTimeMs:v+F}:{asum:k[0]}}finally{!t&&u&&d(u),n&&d(n),l&&d(l),f&&d(f),c&&d(c)}}var wr=64;async function Ce(a,r,e,o){let t=e instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(r)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!t&&!(e instanceof Float64Array))throw new Error("x must be a Float64Array or GpuVector.");if(t&&e.dtype!==Float64Array)throw new Error("x must be a Float64Array-backed GpuVector.");if(r<=0)return{asum:0};if(e.length<(r-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=await B(a,["f64add","dasum"]),s=await B(a,["f64add","reduction/sumF64"]),u=null,n=null,l=null,f=null,c=null,p=null,m=null,g=null;try{u=t?e:x.from(e),n=D(2*wr*4,"dasum-partialsMain"),l=D(2*wr*4,"dasum-partialsAux"),f=q(4,"dasum-result-main"),c=q(4,"dasum-result-aux"),p=N([{value:r,type:"u32"},{value:o,type:"u32"}],"dasum-params");let w=A(i.getBindGroupLayout(0),[u._buf,u._auxBuf,n,l,p]),{commandEncoder:y,ts:h}=L(i,w,2*wr);S(y);let G=A(s.getBindGroupLayout(0),[n,l,f,c]),{commandEncoder:v,ts:F}=L(s,G,1);m=E(v,f),g=E(v,c),S(v);let k=_(m,Float32Array),j=_(g,Uint32Array);m=null,g=null;let[W,M,U,T]=await Promise.all([P(h),P(F),k,j]),R=tr(U[0],T[0]);return W!==void 0&&M!==void 0?{asum:R,gpuTimeMs:W+M}:{asum:R}}finally{!t&&u&&u.destroy(),n&&d(n),l&&d(l),f&&d(f),c&&d(c),p&&d(p),m&&d(m),g&&d(g)}}var Oe=64;async function ze(a,r,e,o){let t=e instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(r)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!t&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(r<=0)return{nrm2:0};if(e.length<(r-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=await B(a,"snrm2"),s=await B(a,"reduction/sum"),u=null,n=null,l=null,f=null,c=null;try{u=t?e._buf:b(e,"snrm2-x",!1),n=D(2*Oe*4,"snrm2-partials"),l=q(4,"snrm2-result"),f=N([{value:r,type:"u32"},{value:o,type:"u32"}],"snrm2-params");let p=A(i.getBindGroupLayout(0),[u,n,f]),{commandEncoder:m,ts:g}=L(i,p,2*Oe);S(m);let w=A(s.getBindGroupLayout(0),[n,l]),{commandEncoder:y,ts:h}=L(s,w,1);c=E(y,l),S(y);let G=_(c,Float32Array);c=null;let[v,F,k]=await Promise.all([P(g),P(h),G]),j=Math.sqrt(k[0]);return v!==void 0&&F!==void 0?{nrm2:j,gpuTimeMs:v+F}:{nrm2:j}}finally{!t&&u&&d(u),n&&d(n),l&&d(l),f&&d(f),c&&d(c)}}var br=64;async function Qe(a,r,e,o){let t=e instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(r)||!Number.isInteger(o))throw new Error("n and incx must be integers.");if(o<=0)throw new Error("incx must be positive.");if(!t&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(r<=0)return{index:0};if(e.length<(r-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");let i=await B(a,"isamax"),s=await B(a,"reduction/argmax"),u=null,n=null,l=null,f=null,c=null,p=null;try{u=t?e._buf:b(e,"isamax-x",!1),n=D(2*br*4,"isamax-partials-val"),l=D(2*br*4,"isamax-partials-idx"),f=q(4,"isamax-result"),c=N([{value:r,type:"u32"},{value:o,type:"u32"}],"isamax-params");let m=A(i.getBindGroupLayout(0),[u,n,l,c]),{commandEncoder:g,ts:w}=L(i,m,2*br);S(g);let y=A(s.getBindGroupLayout(0),[n,l,f]),{commandEncoder:h,ts:G}=L(s,y,1);p=E(h,f),S(h);let v=_(p,Uint32Array);p=null;let[F,k,j]=await Promise.all([P(w),P(G),v]),W=j[0];return F!==void 0&&k!==void 0?{index:W,gpuTimeMs:F+k}:{index:W}}finally{!t&&u&&d(u),n&&d(n),l&&d(l),f&&d(f),c&&d(c),p&&d(p)}}async function qe(a,r,e,o,t,i,s,u){let n=e instanceof x,l=t instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(r)||!Number.isInteger(o)||!Number.isInteger(i))throw new Error("n, incx, and incy must be integers.");if(typeof s!="number")throw new Error("c must be a number.");if(typeof u!="number")throw new Error("s must be a number.");if(Number.isNaN(s)||Number.isNaN(u))throw new Error("c and s must not be NaN.");if(!Number.isFinite(s))throw new Error("c must be finite.");if(!Number.isFinite(u))throw new Error("s must be finite.");if(o<=0||i<=0)throw new Error("incx and incy must be positive.");if(!n&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!l&&!(t instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(n!==l)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(r<=0)return n?{}:{x:e,y:t};if(e.length<(r-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(r-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,m=null,g=null,w=null;try{c=n?e._buf:b(e,"srot-x",!0),p=l?t._buf:b(t,"srot-y",!0),m=N([{value:r,type:"u32"},{value:s,type:"f32"},{value:u,type:"f32"},{value:o,type:"u32"},{value:i,type:"u32"}],"srot-params");let y=A(f.getBindGroupLayout(0),[c,p,m]),{commandEncoder:h,ts:G}=L(f,y,C(r));g=n?null:E(h,c),w=l?null:E(h,p),S(h);let v=await P(G);if(n&&l)return v!==void 0?{gpuTimeMs:v}:{};let F=_(g,Float32Array),k=_(w,Float32Array);g=null,w=null;let[j,W]=await Promise.all([F,k]);return v!==void 0?{x:j,y:W,gpuTimeMs:v}:{x:j,y:W}}finally{!n&&c&&d(c),!l&&p&&d(p),m&&d(m),g&&d(g),w&&d(w)}}async function Ze(a,r,e,o,t,i,s){let u=e instanceof x,n=t instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(r)||!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(!u&&!(e instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!n&&!(t instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(u!==n)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(r<=0||s[0]===-2)return u?{}:{x:e,y:t};if(e.length<(r-1)*o+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(r-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,m=null,g=null,w=null;try{f=u?e._buf:b(e,"srotm-x",!0),c=n?t._buf:b(t,"srotm-y",!0),p=b(s,"srotm-param",!1),m=N([{value:r,type:"u32"},{value:o,type:"u32"},{value:i,type:"u32"}],"srotm-params");let y=A(l.getBindGroupLayout(0),[f,c,p,m]),{commandEncoder:h,ts:G}=L(l,y,C(r));g=u?null:E(h,f),w=n?null:E(h,c),S(h);let v=await P(G);if(u&&n)return v!==void 0?{gpuTimeMs:v}:{};let F=_(g,Float32Array),k=_(w,Float32Array);g=null,w=null;let[j,W]=await Promise.all([F,k]);return v!==void 0?{x:j,y:W,gpuTimeMs:v}:{x:j,y:W}}finally{!u&&f&&d(f),!n&&c&&d(c),p&&d(p),m&&d(m),g&&d(g),w&&d(w)}}async function Ke(a,r,e,o,t,i,s,u,n,l,f,c,p="row-major"){let m=i instanceof H,g=u instanceof x,w=f instanceof x;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(r!=="no-transpose"&&r!=="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(e)||!Number.isInteger(o)||!Number.isInteger(n)||!Number.isInteger(c)||!Number.isInteger(s))throw new Error("m, n, incx, incy, and lda must be integers.");if(n<=0||c<=0)throw new Error("incx and incy must be positive.");if(!m&&!(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&&!m)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(m&&s!==i.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(m&&(i.rows<e||i.cols<o))throw new Error("A is too small for the given m and n.");if(e<0||o<0)throw new Error("m and n must be non-negative.");if(e===0||o===0)return w?{}:{y:f};(m?i.layout:p)==="column-major"&&([e,o]=[o,e],r=r==="no-transpose"?"transpose":"no-transpose");let h=r==="no-transpose",G=h?o:e,v=h?e:o;if(s<o)throw new Error("lda must be >= n.");if(!m&&i.length<(e-1)*s+o)throw new Error("A does not have enough elements for the given m, n, and lda.");if(u.length<(G-1)*n+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 k=await B(a,h?"sgemv_n":"sgemv_t"),j=m?i._buf:b(i,"sgemv-A",!1),W=g?u._buf:b(u,"sgemv-x",!1),M=w?f._buf:b(f,"sgemv-y",!0),U=N([{value:e,type:"u32"},{value:o,type:"u32"},{value:t,type:"f32"},{value:l,type:"f32"},{value:n,type:"u32"},{value:c,type:"u32"},{value:s,type:"u32"}],"sgemv-params");try{let T=A(k.getBindGroupLayout(0),[j,W,M,U]),R=h?Math.min(e,a.limits.maxComputeWorkgroupsPerDimension):C(v),{commandEncoder:z,ts:Y}=L(k,T,R),Q=w?null:E(z,M);S(z);let O=await P(Y);if(w)return O!==void 0?{gpuTimeMs:O}:{};let nr=await _(Q,Float32Array);return O!==void 0?{y:nr,gpuTimeMs:O}:{y:nr}}finally{m||d(j),g||d(W),w||d(M),d(U)}}async function Xe(a,r,e,o,t,i,s,u,n,l,f,c="row-major"){let p=s instanceof x,m=l instanceof x,g=t instanceof H;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");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(!Number.isInteger(e)||!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 n!="number")throw new Error("beta must be a number.");if(Number.isNaN(n))throw new Error("beta must not be NaN.");if(!Number.isFinite(n))throw new Error("beta must be finite.");if(u<=0||f<=0)throw new Error("incx and incy must be positive.");if(i<e)throw new Error("lda must be >= n.");if(!g&&!(t instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!p&&!(s instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!m&&!(l instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(p!==m)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&&s._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<e||t.cols<e))throw new Error("A is too small for the given n.");if(e<0)throw new Error("n must be non-negative.");if(e===0)return m?{}:{y:l};if(!g&&t.length<(e-1)*i+e)throw new Error("A does not have enough elements for the given n and lda.");if(s.length<(e-1)*u+1)throw new Error("x does not have enough elements for the given n and incx.");if(l.length<(e-1)*f+1)throw new Error("y does not have enough elements for the given n and incy.");let y=(g?t.layout:c)==="column-major"?r==="upper":r==="lower",h=await B(a,"ssymv"),G=null,v=null,F=null,k=null;try{G=g?t._buf:b(t,"ssymv-A",!1),v=p?s._buf:b(s,"ssymv-x",!1),F=m?l._buf:b(l,"ssymv-y",!0),k=N([{value:e,type:"u32"},{value:o,type:"f32"},{value:n,type:"f32"},{value:u,type:"u32"},{value:f,type:"u32"},{value:i,type:"u32"},{value:y?0:1,type:"u32"}],"ssymv-params");let j=A(h.getBindGroupLayout(0),[G,v,F,k]),W=Math.min(e,a.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:M,ts:U}=L(h,j,W),T=m?null:E(M,F);S(M);let R=await P(U);if(m)return R!==void 0?{gpuTimeMs:R}:{};let z=await _(T,Float32Array);return R!==void 0?{y:z,gpuTimeMs:R}:{y:z}}finally{!g&&G&&d(G),!p&&v&&d(v),!m&&F&&d(F),k&&d(k)}}async function $e(a,r,e,o,t,i,s,u,n,l,f,c="row-major"){let p=u instanceof x,m=l instanceof x,g=i instanceof H,w=o==="unit";if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(r!=="lower"&&r!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(e!=="no-transpose"&&e!=="transpose")throw new Error("trans must be 'no-transpose' or 'transpose'.");if(!w&&o!=="non-unit")throw new Error("diag must be 'unit' or 'non-unit'.");if(c!=="row-major"&&c!=="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(!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(!m&&!(l instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(p!==m)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&&m&&i._buf===l._buf)throw new Error("A and y must not reference the same GPU buffer.");if(g&&s!==i.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(g&&(i.rows<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 m?{}:{y:l};if(!g&&i.length<(t-1)*s+t)throw new Error("A does not have enough elements for the given n and lda.");if(u.length<(t-1)*n+1)throw new Error("x does not have enough elements for the given n and incx.");if(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?r==="upper":r==="lower",v=h?e==="transpose":e==="no-transpose",F=await B(a,"strmv"),k=null,j=null,W=null,M=null;try{k=g?i._buf:b(i,"strmv-A",!1),j=p?u._buf:b(u,"strmv-x",!1),W=m?l._buf:b(l,"strmv-y",!0),M=N([{value:t,type:"u32"},{value:n,type:"u32"},{value:f,type:"u32"},{value:s,type:"u32"},{value:v?0:1,type:"u32"},{value:G?0:1,type:"u32"},{value:w?1:0,type:"u32"}],"strmv-params");let U=A(F.getBindGroupLayout(0),[k,j,W,M]),T=Math.min(t,a.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:R,ts:z}=L(F,U,T),Y=m?null:E(R,W);S(R);let Q=await P(z);if(m)return Q!==void 0?{gpuTimeMs:Q}:{};let O=await _(Y,Float32Array);return Q!==void 0?{y:O,gpuTimeMs:Q}:{y:O}}finally{!g&&k&&d(k),!p&&j&&d(j),!m&&W&&d(W),M&&d(M)}}var K=64;function Ye(a,r,e){let o=new ArrayBuffer(a*r),t=new DataView(o);for(let i=0;i<a;i++){let s=e(i),u=i*r;s.forEach((n,l)=>t.setUint32(u+l*4,n,!0))}return o}function Je(a,r,e){let o=a.createBuffer({label:e,size:r.byteLength,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});return a.queue.writeBuffer(o,0,r),o}async function rt(a,r,e,o,t,i,s,u,n,l="row-major"){let f=u instanceof x,c=i instanceof H,p=o==="unit";if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(r!=="lower"&&r!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(e!=="no-transpose"&&e!=="transpose")throw new Error("trans must be 'no-transpose' or 'transpose'.");if(!p&&o!=="non-unit")throw new Error("diag must be 'unit' or 'non-unit'.");if(l!=="row-major"&&l!=="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(!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&&s!==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)*s+t)throw new Error("A does not have enough elements for the given n and lda.");if(u.length<(t-1)*n+1)throw new Error("x does not have enough elements for the given n and incx.");let g=(c?i.layout:l)==="column-major",w=g?r==="upper":r==="lower",y=g?e==="transpose":e==="no-transpose",h=await B(a,"strsv_invert_block"),G=await B(a,"strsv_apply_inverse"),v=await B(a,"strsv_update"),F=y===w,k=[];for(let O=0;O<t;O+=K)k.push(O);F||k.reverse();let j=k.length,W=a.limits.maxComputeWorkgroupsPerDimension,M=a.limits.minUniformBufferOffsetAlignment,U=null,T=null,R=null,z=null,Y=null,Q=null;try{U=c?i._buf:b(i,"strsv-A",!1),T=f?u._buf:b(u,"strsv-x",!0),R=D(j*K*K*4,"strsv-Ainv");let O=Ye(j,M,X=>{let $=X*K,or=Math.min($+K,t);return[n,X,$,or]});z=Je(a,O,"strsv-apply-params");let nr=Ye(j,M,X=>{let $=X*K,or=Math.min($+K,t);return[t,n,s,y?0:1,w?0:1,$,or]});Y=Je(a,nr,"strsv-update-params");let{commandEncoder:rr,querySet:ar}=pr();Q=N([{value:t,type:"u32"},{value:s,type:"u32"},{value:y?0:1,type:"u32"},{value:w?0:1,type:"u32"},{value:p?1:0,type:"u32"}],"strsv-invert-params");let ot=A(h.getBindGroupLayout(0),[U,R,Q]);ir(rr,h,ot,{x:K,y:j},ar?{timestampWrites:{querySet:ar,beginningOfPassWriteIndex:0}}:void 0);for(let X=0;X<k.length;X++){let $=k[X],or=Math.min($+K,t),st=$/K,ut=X===k.length-1,xr=st*M,lt=A(G.getBindGroupLayout(0),[R,T,{buffer:z,offset:xr,size:16}]);ir(rr,G,lt,1,ut&&ar?{timestampWrites:{querySet:ar,endOfPassWriteIndex:1}}:void 0);let vr=F?t-or:$;if(vr===0)continue;let ft=A(v.getBindGroupLayout(0),[U,T,{buffer:Y,offset:xr,size:32}]),ct=Math.min(vr,W);ir(rr,v,ft,ct)}let it=cr(rr,ar),nt=f?null:E(rr,T);S(rr);let sr=await P(it);if(f)return sr!==void 0?{gpuTimeMs:sr}:{};let hr=await _(nt,Float32Array);return sr!==void 0?{x:hr,gpuTimeMs:sr}:{x:hr}}finally{!c&&U&&d(U),!f&&T&&d(T),R&&d(R),z&&d(z),Y&&d(Y),Q&&d(Q)}}async function et(a,r,e,o,t,i,s,u,n,l,f="row-major"){let c=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(r)||!Number.isInteger(e)||!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&&!(n instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(c&&l!==n.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(c&&(n.rows<r||n.cols<e))throw new Error("A is too small for the given m and n.");(c?n.layout:f)==="column-major"&&([r,e]=[e,r],[t,s]=[s,t],[i,u]=[u,i]);let m=t instanceof x,g=s instanceof x;if(l<e)throw new Error("lda must be >= n.");if(!m&&!(t instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!g&&!(s instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(m!==g)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&&n._buf===t._buf)throw new Error("A and x must not reference the same GPU buffer.");if(c&&g&&n._buf===s._buf)throw new Error("A and y must not reference the same GPU buffer.");if(r<0||e<0)throw new Error("m and n must be non-negative.");if(r===0||e===0)return c?{}:{A:n};if(!c&&n.length<(r-1)*l+e)throw new Error("A does not have enough elements for the given m, n, and lda.");if(t.length<(r-1)*i+1)throw new Error("x does not have enough elements for the given m and incx.");if(s.length<(e-1)*u+1)throw new Error("y does not have enough elements for the given n and incy.");let w=await B(a,"sger"),y=null,h=null,G=null,v=null;try{y=m?t._buf:b(t,"sger-x",!1),h=g?s._buf:b(s,"sger-y",!1),G=c?n._buf:b(n,"sger-A",!0),v=N([{value:r,type:"u32"},{value:e,type:"u32"},{value:o,type:"f32"},{value:i,type:"u32"},{value:u,type:"u32"},{value:l,type:"u32"}],"sger-params");let F=A(w.getBindGroupLayout(0),[y,h,G,v]),k=Math.min(r,a.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:j,ts:W}=L(w,F,k),M=c?null:E(j,G);S(j);let U=await P(W);if(c)return U!==void 0?{gpuTimeMs:U}:{};let T=await _(M,Float32Array);return U!==void 0?{A:T,gpuTimeMs:U}:{A:T}}finally{!m&&y&&d(y),!g&&h&&d(h),!c&&G&&d(G),v&&d(v)}}async function tt(a,r,e,o,t,i,s,u,n="row-major"){let l=t instanceof x,f=s instanceof H;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(r!=="lower"&&r!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(n!=="row-major"&&n!=="column-major")throw new Error("layout must be 'row-major' or 'column-major'.");if(!Number.isInteger(e)||!Number.isInteger(i)||!Number.isInteger(u))throw new Error("n, incx, and lda must be integers.");if(typeof o!="number")throw new Error("alpha must be a number.");if(Number.isNaN(o))throw new Error("alpha must not be NaN.");if(!Number.isFinite(o))throw new Error("alpha must be finite.");if(i<=0)throw new Error("incx must be positive.");if(u<e)throw new Error("lda must be >= n.");if(!f&&!(s 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&&s._buf===t._buf)throw new Error("A and x must not reference the same GPU buffer.");if(f&&u!==s.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(f&&(s.rows<e||s.cols<e))throw new Error("A is too small for the given n.");if(e<0)throw new Error("n must be non-negative.");if(e===0)return f?{}:{A:s};if(!f&&s.length<(e-1)*u+e)throw new Error("A does not have enough elements for the given n and lda.");if(t.length<(e-1)*i+1)throw new Error("x does not have enough elements for the given n and incx.");let p=(f?s.layout:n)==="column-major"?r==="upper":r==="lower",m=await B(a,"ssyr"),g=null,w=null,y=null;try{g=l?t._buf:b(t,"ssyr-x",!1),w=f?s._buf:b(s,"ssyr-A",!0),y=N([{value:e,type:"u32"},{value:o,type:"f32"},{value:i,type:"u32"},{value:u,type:"u32"},{value:p?0:1,type:"u32"}],"ssyr-params");let h=A(m.getBindGroupLayout(0),[g,w,y]),G=Math.min(e,a.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:v,ts:F}=L(m,h,G),k=f?null:E(v,w);S(v);let j=await P(F);if(f)return j!==void 0?{gpuTimeMs:j}:{};let W=await _(k,Float32Array);return j!==void 0?{A:W,gpuTimeMs:j}:{A:W}}finally{!l&&g&&d(g),!f&&w&&d(w),y&&d(y)}}async function at(a,r,e,o,t,i,s,u,n,l,f="row-major"){let c=t instanceof x,p=s instanceof x,m=n instanceof H;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(r!=="lower"&&r!=="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(e)||!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<e)throw new Error("lda must be >= n.");if(!m&&!(n 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&&!(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&&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(c&&t._buf===s._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(m&&l!==n.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(m&&(n.rows<e||n.cols<e))throw new Error("A is too small for the given n.");if(e<0)throw new Error("n must be non-negative.");if(e===0)return m?{}:{A:n};if(!m&&n.length<(e-1)*l+e)throw new Error("A does not have enough elements for the given n and lda.");if(t.length<(e-1)*i+1)throw new Error("x does not have enough elements for the given n and incx.");if(s.length<(e-1)*u+1)throw new Error("y does not have enough elements for the given n and incy.");let w=(m?n.layout:f)==="column-major"?r==="upper":r==="lower",y=await B(a,"ssyr2"),h=null,G=null,v=null,F=null;try{h=c?t._buf:b(t,"ssyr2-x",!1),G=p?s._buf:b(s,"ssyr2-y",!1),v=m?n._buf:b(n,"ssyr2-A",!0),F=N([{value:e,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 k=A(y.getBindGroupLayout(0),[h,G,v,F]),j=Math.min(e,a.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:W,ts:M}=L(y,k,j),U=m?null:E(W,v);S(W);let T=await P(M);if(m)return T!==void 0?{gpuTimeMs:T}:{};let R=await _(U,Float32Array);return T!==void 0?{A:R,gpuTimeMs:T}:{A:R}}finally{!c&&h&&d(h),!p&&G&&d(G),!m&&v&&d(v),F&&d(F)}}return bt(ea);})();
+`)}`);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);})();