wgblas 1.0.0 → 1.1.0

package/dist/wgblas.browser.js

@@ -1,4 +1,4 @@
-var wgblas=(()=>{var Je=Object.create;var or=Object.defineProperty;var rt=Object.getOwnPropertyDescriptor;var et=Object.getOwnPropertyNames;var tt=Object.getPrototypeOf,at=Object.prototype.hasOwnProperty;var nr=(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 W=(a,r,e)=>()=>{if(e)throw e[0];try{return a&&(r=a(a=0)),r}catch(i){throw e=[i],i}};var hr=(a,r)=>{for(var e in r)or(a,e,{get:r[e],enumerable:!0})},xr=(a,r,e,i)=>{if(r&&typeof r=="object"||typeof r=="function")for(let t of et(r))!at.call(a,t)&&t!==e&&or(a,t,{get:()=>r[t],enumerable:!(i=rt(r,t))||i.enumerable});return a};var sr=(a,r,e)=>(e=a!=null?Je(tt(a)):{},xr(r||!a||!a.__esModule?or(e,"default",{value:a,enumerable:!0}):e,a)),it=a=>xr(or({},"__esModule",{value:!0}),a);var jr,Ir=W(()=>{jr=`// amax reduction: collapses 2*WGS (value, index) pairs into one index.
+var wgblas=(()=>{var ct=Object.create;var ur=Object.defineProperty;var mt=Object.getOwnPropertyDescriptor;var dt=Object.getOwnPropertyNames;var pt=Object.getPrototypeOf,gt=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 dt(r))!gt.call(a,t)&&t!==e&&ur(a,t,{get:()=>r[t],enumerable:!(o=mt(r,t))||o.enumerable});return a};var fr=(a,r,e)=>(e=a!=null?ct(pt(a)):{},_r(r||!a||!a.__esModule?ur(e,"default",{value:a,enumerable:!0}):e,a)),wt=a=>_r(ur({},"__esModule",{value:!0}),a);var Mr,Ir=I(()=>{Mr=`// 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 Wr,Nr=W(()=>{Wr=`// sum reduction: collapses 2*WGS partials into one scalar.
+`});var Tr,Ur=I(()=>{Tr=`// sum reduction: collapses 2*WGS partials into one scalar.
 // dispatch: 1 workgroup of WGS threads.
 // partials must have exactly 2*WGS entries.
 
@@ -67,7 +67,7 @@ fn reduce(
 
   if (i == 0u) { result[0] = tile[0]; }
 }
-`});var Ur,Mr=W(()=>{Ur=`// sum reduction (f64): collapses 2*WGS partial [main, aux] pairs into one,
+`});var Rr,Vr=I(()=>{Rr=`// 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
@@ -116,7 +116,7 @@ fn reduce_f64(
     resultAux[0] = tile[0].aux;
   }
 }
-`});var Rr,Tr=W(()=>{Rr=`// sscal: x = alpha * x
+`});var Dr,Hr=I(()=>{Dr=`// sscal: x = alpha * x
 
 @group(0) @binding(0) var<storage, read_write> x: array<f32>;
 
@@ -139,7 +139,7 @@ fn main(
     x[id * params.x_inc] = params.alpha * x[id * params.x_inc];
   }
 }
-`});var Hr,Vr=W(()=>{Hr=`// sswap: x <-> y
+`});var Or,Cr=I(()=>{Or=`// 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 +165,7 @@ fn main(
     y[id * params.y_inc] = temp;
   }
 }
-`});var Cr,Dr=W(()=>{Cr=`// saxpy: y = alpha * x + y
+`});var Qr,zr=I(()=>{Qr=`// 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 +190,7 @@ fn main(
     y[id * params.y_inc] = params.alpha * x[id * params.x_inc] + y[id * params.y_inc];
   }
 }
-`});var zr,Or=W(()=>{zr=`// scopy: y = x
+`});var Zr,qr=I(()=>{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 +214,7 @@ fn main(
     y[id * params.y_inc] = x[id * params.x_inc];
   }
 }
-`});var qr,Qr=W(()=>{qr=`// sdot: result = sum(x[i] * y[i])
+`});var Xr,Kr=I(()=>{Xr=`// 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 +267,7 @@ fn main(
 
   if (lid.x == 0u) { partials[wgid.x] = tile[0]; }
 }
-`});var Kr,Zr=W(()=>{Kr=`// sasum: result = sum(|x[i]|)
+`});var Yr,$r=I(()=>{Yr=`// 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 +318,7 @@ fn main(
 
   if (lid.x == 0u) { partials[wgid.x] = tile[0]; }
 }
-`});var $r,Xr=W(()=>{$r=`// snrm2: result = sqrt(sum(x[i] * x[i]))
+`});var re,Jr=I(()=>{re=`// snrm2: result = sqrt(sum(x[i] * x[i]))
 // pass 1 dispatches exactly 2 * WGS workgroups; pass 2 uses reduction/sqsum.wgsl.
 
 @group(0) @binding(0) var<storage, read>       x:        array<f32>;
@@ -374,7 +374,7 @@ fn main(
 
   if (lid.x == 0u) { partials[wgid.x] = tile[0]; }
 }
-`});var Jr,Yr=W(()=>{Jr=`// srot: x = c*x + s*y,  y = -s*x + c*y
+`});var te,ee=I(()=>{te=`// 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 +403,7 @@ fn main(
     y[id * params.y_inc] = -params.s * xi + params.c * yi;
   }
 }
-`});var ee,re=W(()=>{ee=`// srotm: applies modified Givens rotation H to vectors x and y.
+`});var oe,ae=I(()=>{oe=`// 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 +453,7 @@ fn main(
     y[id * params.y_inc] = h21 * xi + h22 * yi;
   }
 }
-`});var ae,te=W(()=>{ae=`// isamax: returns index of element with largest absolute value
+`});var ne,ie=I(()=>{ne=`// 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 +535,7 @@ fn main(
     partials_idx[wgid.x] = tile_idx[0];
   }
 }
-`});var oe,ie=W(()=>{oe=`// sgemv_n: y = alpha * A * x + beta * y  (A is m\xD7n row-major, no-transpose)
+`});var ue,se=I(()=>{ue=`// 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 +610,7 @@ fn main(
     workgroupBarrier();
   }
 }
-`});var se,ne=W(()=>{se=`// sgemv_t: y = alpha * A^T * x + beta * y  (A is m\xD7n row-major, transposed)
+`});var fe,le=I(()=>{fe=`// 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 +675,7 @@ fn main(
     y[yi] = params.alpha * (acc0 + acc1 + acc2 + acc3) + params.beta * y[yi];
   }
 }
-`});var le,ue=W(()=>{le=`// ssymv: y = alpha * A * x + beta * y
+`});var me,ce=I(()=>{me=`// 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 +744,7 @@ fn main(
     }
   }
 }
-`});var fe,ce=W(()=>{fe=`// strmv: y = op(A) * x
+`});var pe,de=I(()=>{pe=`// 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 +847,178 @@ fn main(
     }
   }
 }
-`});var me,de=W(()=>{me=`// f64add: adds two doubles, each packed as a [main, aux] pair (main: f32
+`});var we,ge=I(()=>{we=`// 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>;
+@group(0) @binding(2) var<storage, read_write> A: array<f32>;
+
+struct Params {
+  m:     u32,
+  n:     u32,
+  alpha: f32,
+  incx:  u32,
+  incy:  u32,
+  lda:   u32,
+}
+
+@group(0) @binding(3) var<uniform> params: Params;
+
+const WGS: u32 = 64u;
+
+@compute @workgroup_size(64)
+fn main(
+  @builtin(workgroup_id)        wgid: vec3u,
+  @builtin(local_invocation_id) lid:  vec3u,
+  @builtin(num_workgroups)      nwg:  vec3u,
+) {
+  for (var row = wgid.x; row < params.m; row += nwg.x) {
+    let xi = params.alpha * x[row * params.incx];
+    let row_base = row * params.lda;
+
+    // 4-unrolled loop: each iteration issues 4 independent A/y accesses.
+    let n4_floor = (params.n / (4u * WGS)) * (4u * WGS);
+    for (var col: u32 = lid.x; col < n4_floor; col += 4u * WGS) {
+      let idx0 = row_base + col;
+      let idx1 = row_base + col + WGS;
+      let idx2 = row_base + col + 2u * WGS;
+      let idx3 = row_base + col + 3u * WGS;
+      A[idx0] = xi * y[ col            * params.incy] + A[idx0];
+      A[idx1] = xi * y[(col +     WGS) * params.incy] + A[idx1];
+      A[idx2] = xi * y[(col + 2u * WGS) * params.incy] + A[idx2];
+      A[idx3] = xi * y[(col + 3u * WGS) * params.incy] + A[idx3];
+    }
+    // Scalar tail: at most 3*WGS elements left after the unrolled block.
+    for (var col: u32 = n4_floor + lid.x; col < params.n; col += WGS) {
+      let idx = row_base + col;
+      A[idx] = xi * y[col * params.incy] + A[idx];
+    }
+  }
+}
+`});var he,be=I(()=>{he=`// 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).
+
+@group(0) @binding(0) var<storage, read>       x: array<f32>;
+@group(0) @binding(1) var<storage, read_write> A: array<f32>;
+
+struct Params {
+  n:     u32,
+  alpha: f32,
+  incx:  u32,
+  lda:   u32,
+  uplo:  u32,  // 0 = lower, 1 = upper
+}
+
+@group(0) @binding(2) var<uniform> params: Params;
+
+const WGS: u32 = 64u;
+
+@compute @workgroup_size(64)
+fn main(
+  @builtin(workgroup_id)        wgid: vec3u,
+  @builtin(local_invocation_id) lid:  vec3u,
+  @builtin(num_workgroups)      nwg:  vec3u,
+) {
+  for (var row = wgid.x; row < params.n; row += nwg.x) {
+    let xi = params.alpha * x[row * params.incx];
+    let row_base = row * params.lda;
+
+    // Stored-triangle column range for this row: lower [0,row], upper [row,n).
+    var colStart: u32;
+    var colEnd: u32;
+    if params.uplo == 1u {
+      colStart = row;
+      colEnd = params.n;
+    } else {
+      colStart = 0u;
+      colEnd = row + 1u;
+    }
+
+    // 4-unrolled loop over the stored range.
+    let rangeLen = colEnd - colStart;
+    let n4_floor = colStart + (rangeLen / (4u * WGS)) * (4u * WGS);
+    for (var col: u32 = colStart + lid.x; col < n4_floor; col += 4u * WGS) {
+      let idx0 = row_base + col;
+      let idx1 = row_base + col + WGS;
+      let idx2 = row_base + col + 2u * WGS;
+      let idx3 = row_base + col + 3u * WGS;
+      A[idx0] = xi * x[ col             * params.incx] + A[idx0];
+      A[idx1] = xi * x[(col +     WGS)  * params.incx] + A[idx1];
+      A[idx2] = xi * x[(col + 2u * WGS) * params.incx] + A[idx2];
+      A[idx3] = xi * x[(col + 3u * WGS) * params.incx] + A[idx3];
+    }
+    // Scalar tail: at most 3*WGS elements left after the unrolled block.
+    for (var col: u32 = n4_floor + lid.x; col < colEnd; col += WGS) {
+      let idx = row_base + col;
+      A[idx] = xi * x[col * params.incx] + A[idx];
+    }
+  }
+}
+`});var ve,xe=I(()=>{ve=`// 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).
+
+@group(0) @binding(0) var<storage, read>       x: array<f32>;
+@group(0) @binding(1) var<storage, read>       y: array<f32>;
+@group(0) @binding(2) var<storage, read_write> A: array<f32>;
+
+struct Params {
+  n:     u32,
+  alpha: f32,
+  incx:  u32,
+  incy:  u32,
+  lda:   u32,
+  uplo:  u32,  // 0 = lower, 1 = upper
+}
+
+@group(0) @binding(3) var<uniform> params: Params;
+
+const WGS: u32 = 64u;
+
+@compute @workgroup_size(64)
+fn main(
+  @builtin(workgroup_id)        wgid: vec3u,
+  @builtin(local_invocation_id) lid:  vec3u,
+  @builtin(num_workgroups)      nwg:  vec3u,
+) {
+  for (var row = wgid.x; row < params.n; row += nwg.x) {
+    let xi = params.alpha * x[row * params.incx];
+    let yi = params.alpha * y[row * params.incy];
+    let row_base = row * params.lda;
+
+    // Stored-triangle column range for this row: lower [0,row], upper [row,n).
+    var colStart: u32;
+    var colEnd: u32;
+    if params.uplo == 1u {
+      colStart = row;
+      colEnd = params.n;
+    } else {
+      colStart = 0u;
+      colEnd = row + 1u;
+    }
+
+    // 4-unrolled loop over the stored range.
+    let rangeLen = colEnd - colStart;
+    let n4_floor = colStart + (rangeLen / (4u * WGS)) * (4u * WGS);
+    for (var col: u32 = colStart + lid.x; col < n4_floor; col += 4u * WGS) {
+      let idx0 = row_base + col;
+      let idx1 = row_base + col + WGS;
+      let idx2 = row_base + col + 2u * WGS;
+      let idx3 = row_base + col + 3u * WGS;
+      A[idx0] = xi * y[ col             * params.incy] + yi * x[ col             * params.incx] + A[idx0];
+      A[idx1] = xi * y[(col +     WGS)  * params.incy] + yi * x[(col +     WGS)  * params.incx] + A[idx1];
+      A[idx2] = xi * y[(col + 2u * WGS) * params.incy] + yi * x[(col + 2u * WGS) * params.incx] + A[idx2];
+      A[idx3] = xi * y[(col + 3u * WGS) * params.incy] + yi * x[(col + 3u * WGS) * params.incx] + A[idx3];
+    }
+    // Scalar tail: at most 3*WGS elements left after the unrolled block.
+    for (var col: u32 = n4_floor + lid.x; col < colEnd; col += WGS) {
+      let idx = row_base + col;
+      A[idx] = xi * y[col * params.incy] + yi * x[col * params.incx] + A[idx];
+    }
+  }
+}
+`});var _e,ye=I(()=>{_e=`// 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.
@@ -1128,7 +1299,7 @@ fn computeSum(a: Fields, b: Fields) -> Packed {
   let f = addFields(a, b);
   return encode(f.sign, f.rawExp, f.mantissaHi, f.lo);
 }
-`});var ge,pe=W(()=>{ge=`// dasum: result = sum(|x[i]|)
+`});var Ge,Ee=I(()=>{Ge=`// 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
@@ -1226,7 +1397,7 @@ fn dasum_main(
     partialsAux[wgid.x] = tile[0].aux;
   }
 }
-`});var be,we=W(()=>{be=`// strsv_invert_block: computes ONE column (workgroup_id.x) of ONE block's
+`});var Be,Ae=I(()=>{Be=`// 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
@@ -1335,7 +1506,7 @@ fn strsv_invert_block_main(
     workgroupBarrier();
   }
 }
-`});var xe,he=W(()=>{xe=`// strsv_apply_inverse: given a precomputed block inverse (from
+`});var Pe,ke=I(()=>{Pe=`// 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,
@@ -1382,7 +1553,7 @@ fn strsv_apply_inverse_main(@builtin(local_invocation_id) lid: vec3u) {
   }
   x[(params.blockStart + lid.x) * params.incx] = acc;
 }
-`});var ye,ve=W(()=>{ye=`// strsv_update: subtracts a solved block's contribution from every
+`});var Fe,Se=I(()=>{Fe=`// 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.
@@ -1457,7 +1628,7 @@ fn strsv_update_main(
     workgroupBarrier();
   }
 }
-`});var _e={};hr(_e,{shaderSources:()=>Nt});var Nt,Ee=W(()=>{Ir();Nr();Mr();Tr();Vr();Dr();Or();Qr();Zr();Xr();Yr();re();te();ie();ne();ue();ce();de();pe();we();he();ve();Nt={"reduction/argmax":jr,"reduction/sum":Wr,"reduction/sumF64":Ur,sscal:Rr,sswap:Hr,saxpy:Cr,scopy:zr,sdot:qr,sasum:Kr,snrm2:$r,srot:Jr,srotm:ee,isamax:ae,sgemv_n:oe,sgemv_t:se,ssymv:le,strmv:fe,f64add:me,dasum:ge,strsv_invert_block:be,strsv_apply_inverse:xe,strsv_update:ye}});var Rt={};hr(Rt,{GpuMatrix:()=>C,GpuVector:()=>h,cleanup:()=>Br,dasum:()=>je,gpuName:()=>Ar,init:()=>kr,isamax:()=>Me,randomFloat32Array:()=>Fr,randomFloat64Array:()=>Lr,sasum:()=>Ie,saxpy:()=>Ge,scopy:()=>Pe,sdot:()=>Fe,sgemv:()=>Re,snrm2:()=>We,srot:()=>Ue,srotm:()=>Te,sscal:()=>Be,sswap:()=>Ae,ssymv:()=>Ve,strmv:()=>He,strsv:()=>Oe});function vr(a,r){return r?a.features.has("timestamp-query")?{requiredFeatures:["timestamp-query"]}:(console.warn("timestamp-query not supported on this device \u2014 benchmark mode disabled."),{}):{}}function yr(){if(!_r())return{querySet:null,passDescriptor:void 0};let r=M().createQuerySet({type:"timestamp",count:2});return{querySet:r,passDescriptor:{timestampWrites:{querySet:r,beginningOfPassWriteIndex:0,endOfPassWriteIndex:1}}}}function ur(a,r){if(!r)return null;let e=M(),i=e.createBuffer({label:"timestamp-resolve",size:16,usage:GPUBufferUsage.QUERY_RESOLVE|GPUBufferUsage.COPY_SRC});a.resolveQuerySet(r,0,2,i,0);let t=e.createBuffer({label:"timestamp-readback",size:16,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});return a.copyBufferToBuffer(i,0,t,0,16),{tsReadBuffer:t,resolveBuffer:i,querySet:r}}async function G(a){if(!a)return;let{tsReadBuffer:r,resolveBuffer:e,querySet:i}=a;await r.mapAsync(GPUMapMode.READ);let t=new BigInt64Array(r.getMappedRange().slice());return r.unmap(),r.destroy(),e.destroy(),i.destroy(),Math.max(0,Number(t[1]-t[0]))/1e6}var Q=null,$=null,Er=null,cr=!1;async function kr({powerPreference:a="high-performance",benchmark:r=!1}={}){if(Q)return Q;let e;if(typeof window>"u"){let{create:o,globals:u}=await import("webgpu");Object.assign(globalThis,u),e=o([]),Er=e}else e=navigator.gpu;if(!e)throw new Error("WebGPU not supported in this environment.");if($=await e.requestAdapter({powerPreference:a})??await e.requestAdapter(),!$)throw new Error("No WebGPU adapter found.");cr=r;let t=[...vr($,r).requiredFeatures??[]];return Q=await $.requestDevice({requiredFeatures:t}),Q.addEventListener("uncapturederror",o=>{console.error("Uncaptured GPU error:",o.error.message)}),Q}function Br(){Q&&(Q.destroy(),Q=null),$=null,Er=null,cr=!1}function Ar(){if(!$)throw new Error("WebGPU adapter not initialized \u2014 call init() first.");let{device:a,description:r}=$.info;return{description:r||"unknown",device:a||"unknown"}}function _r(){return cr}function M(){if(!Q)throw new Error("WebGPU device not initialized \u2014 call init() first.");return Q}function m(...a){a.flat().forEach(r=>r.destroy())}function b(a,r="blas-input",e=!1){let i=M(),t=i.limits.maxStorageBufferBindingSize,o=a.byteLength;if(o>t)throw new Error(`Buffer size ${o} bytes exceeds device limit of ${t} bytes.`);let u=e?GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC:GPUBufferUsage.STORAGE,n=i.createBuffer({label:r,size:o,usage:u,mappedAtCreation:!0}),l=a.constructor;return new l(n.getMappedRange()).set(a),n.unmap(),n}function H(a,r="blas-storage"){return M().createBuffer({label:r,size:a,usage:GPUBufferUsage.STORAGE})}function O(a,r="blas-result"){return M().createBuffer({label:r,size:a,usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC})}function _(a,r){let i=M().createBuffer({label:"blas-readback",size:r.size,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});return a.copyBufferToBuffer(r,0,i,0,r.size),i}function L(a,r="blas-params"){let e=M(),i=a.length*4,t=Math.ceil(i/16)*16,o=new ArrayBuffer(t),u=new DataView(o);a.forEach(({value:l,type:s},f)=>{let c=f*4;if(s==="u32")u.setUint32(c,l,!0);else if(s==="i32")u.setInt32(c,l,!0);else if(s==="f32")u.setFloat32(c,l,!0);else throw new Error(`Unknown param type "${s}". Use "f32", "u32", or "i32".`)});let n=e.createBuffer({label:r,size:t,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});return e.queue.writeBuffer(n,0,o),n}async function y(a,r=Float32Array){try{await a.mapAsync(GPUMapMode.READ);let e=new r(a.getMappedRange().slice());return a.unmap(),e}finally{a.destroy()}}var ot=new ArrayBuffer(8),J=new DataView(ot),Gr=new ArrayBuffer(4),Pr=new Uint32Array(Gr),Sr=new Float32Array(Gr);function nt(a){return Pr[0]=a>>>0,Sr[0]}function st(a){return Sr[0]=a,Pr[0]}function ut(a,r,e,i){let t=r>>>3,o=r&7,u=i>>>29,n=e<<3|u,l=i&536870911,s=a<<31|t<<23|n,f=o>>>2&1,c=o&3,d=l>>>23,p=l&8388607,g=c<<6|d,w=(f<<31|g<<23|p)>>>0;return[nt(s),w]}function lt(a,r){let e=st(a);r=r>>>0;let i=e>>>31,t=e>>>23&255,o=e&8388607,u=r>>>31,n=r>>>23&255,l=r&8388607,s=u<<2|n>>>6,f=(n&63)<<23|l,c=t<<3|s,d=o>>>3,g=((o&7)<<29|f)>>>0;return{sign:i,rawExp:c,mantissaHi:d,lo:g}}var ct=2040;function lr(a){J.setFloat64(0,a,!1);let r=J.getUint32(0,!1),e=J.getUint32(4,!1),i=r>>>31,t=r>>>20&2047,o=r&1048575;if(t>=ct)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 ut(i,t,o,e)}function rr(a,r){let{sign:e,rawExp:i,mantissaHi:t,lo:o}=lt(a,r),u=(e<<31|i<<20|t)>>>0;return J.setUint32(0,u,!1),J.setUint32(4,o,!1),J.getFloat64(0,!1)}var h=class a{constructor(r,e,i=Float32Array,t=null){this._buf=r,this._auxBuf=t,this.length=e,this.dtype=i}static from(r){if(r instanceof Float64Array){let i=new Float32Array(r.length),t=new Uint32Array(r.length);for(let n=0;n<r.length;n++){let l=lr(r[n]);i[n]=l[0],t[n]=l[1]}let o=b(i,"gpu-vector-f64-main",!0),u=b(t,"gpu-vector-f64-aux",!0);return new a(o,r.length,Float64Array,u)}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=M(),e=r.createCommandEncoder(),i=_(e,this._buf);if(r.queue.submit([e.finish()]),!this._auxBuf)return y(i,this.dtype);let t=r.createCommandEncoder(),o=_(t,this._auxBuf);r.queue.submit([t.finish()]);let[u,n]=await Promise.all([y(i,Float32Array),y(o,Uint32Array)]),l=new Float64Array(this.length);for(let s=0;s<this.length;s++)l[s]=rr(u[s],n[s]);return l}destroy(){this._buf.destroy(),this._auxBuf&&this._auxBuf.destroy()}};var C=class a{constructor(r,e,i,t,o=null){this._buf=r,this._auxBuf=o,this.rows=e,this.cols=i,this.lda=t}static from(r,e,i,t=i){if(!(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(i)||i<=0)throw new Error("cols must be a positive integer.");if(!Number.isInteger(t)||t<i)throw new Error("lda must be an integer >= cols.");if(r.length<e*t)throw new Error("data does not have enough elements for the given rows and lda.");if(r instanceof Float64Array){let u=e*t,n=new Float32Array(u),l=new Uint32Array(u);for(let c=0;c<u;c++){let d=lr(r[c]);n[c]=d[0],l[c]=d[1]}let s=b(n,"gpu-matrix-f64-main",!0),f=b(l,"gpu-matrix-f64-aux",!0);return new a(s,e,i,t,f)}let o=b(r.subarray(0,e*t),"gpu-matrix",!0);return new a(o,e,i,t)}async read(){let r=M(),e=r.createCommandEncoder(),i=_(e,this._buf);if(r.queue.submit([e.finish()]),this._auxBuf){let u=r.createCommandEncoder(),n=_(u,this._auxBuf);r.queue.submit([u.finish()]);let[l,s]=await Promise.all([y(i,Float32Array),y(n,Uint32Array)]),f=new Float64Array(this.rows*this.lda);for(let d=0;d<f.length;d++)f[d]=rr(l[d],s[d]);if(this.lda===this.cols)return f;let c=new Float64Array(this.rows*this.cols);for(let d=0;d<this.rows;d++)c.set(f.subarray(d*this.lda,d*this.lda+this.cols),d*this.cols);return c}let t=await y(i,Float32Array);if(this.lda===this.cols)return t;let o=new Float32Array(this.rows*this.cols);for(let u=0;u<this.rows;u++)o.set(t.subarray(u*this.lda,u*this.lda+this.cols),u*this.cols);return o}destroy(){this._buf.destroy(),this._auxBuf&&this._auxBuf.destroy()}};function Fr(a,r=-1,e=1){let i=new Float32Array(a);for(let t=0;t<a;t++)i[t]=r+Math.random()*(e-r);return i}function Lr(a,r=-1,e=1){let i=new Float64Array(a);for(let t=0;t<a;t++)i[t]=r+Math.random()*(e-r);return i}function B(a,r,e=0){let i=M(),t=r.map((o,u)=>({binding:e+u,resource:o instanceof GPUBuffer?{buffer:o}:o}));return i.createBindGroup({layout:a,entries:t})}var ft=new WeakMap;function P(a){M().queue.submit([a.finish()])}function fr(){let a=M(),{querySet:r,passDescriptor:e}=yr();return{commandEncoder:a.createCommandEncoder(),querySet:r,passDescriptor:e}}function ar(a,r,e,i,t){let o=a.beginComputePass(t);o.setPipeline(r),o.setBindGroup(0,e),typeof i=="number"?o.dispatchWorkgroups(i):o.dispatchWorkgroups(i.x,i.y),o.end(),ft.set(a,o)}function S(a,r,e){let{commandEncoder:i,querySet:t,passDescriptor:o}=fr();ar(i,a,r,e,o);let u=ur(i,t);return{commandEncoder:i,ts:u}}var Ut={},dr=new WeakMap;async function A(a,r,e="main"){dr.has(a)||dr.set(a,new Map);let i=dr.get(a),t=Array.isArray(r)?r:[r],o=`${t.join("+")}::${e}`;return i.has(o)||i.set(o,await Mt(t,e)),i.get(o)}async function Wt(a){if(typeof process>"u"||!process.versions?.node){let{shaderSources:r}=await Promise.resolve().then(()=>(Ee(),_e)),e=r[a];if(!e)throw new Error(`Shader "${a}" not found in browser bundle.`);return e}else{let{readFileSync:r}=await import("fs"),{fileURLToPath:e}=await import("url"),{dirname:i,join:t}=await import("path"),o=i(e(Ut.url));return r(t(o,`../shaders/${a}.wgsl`),"utf8")}}async function Mt(a,r="main"){let e=M(),i=a.join("+"),t=(await Promise.all(a.map(Wt))).join(`
-`),o=e.createShaderModule({label:i,code:t}),n=(await o.getCompilationInfo()).messages.filter(f=>f.type==="error");if(n.length>0)throw new Error(`Shader "${i}" compilation failed:
-${n.map(f=>`  line ${f.lineNum}: ${f.message}`).join(`
-`)}`);let l=r==="main"?{module:o}:{module:o,entryPoint:r},s=e.createComputePipeline({label:i,layout:"auto",compute:l});return s._shaderModule=o,s}var Tt=64,ke=8;function D(a,r){let e=M().limits.maxComputeWorkgroupsPerDimension;return r===void 0?Math.min(Math.ceil(a/Tt),e):{x:Math.min(Math.ceil(r/ke),e),y:Math.min(Math.ceil(a/ke),e)}}async function Be(a,r,e,i,t){let o=i instanceof h;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(!(i instanceof Float32Array)&&!(i instanceof h))throw new Error("x must be a Float32Array or GpuVector.");if(r<=0)return o?{}:i;if(i.length<(r-1)*t+1)throw new Error("x does not have enough elements for the given n and incx.");let u=await A(a,"sscal"),n=null,l=null,s=null;try{n=o?i._buf:b(i,"sscal-x",!0),l=L([{value:r,type:"u32"},{value:e,type:"f32"},{value:t,type:"u32"}],"sscal-params");let f=B(u.getBindGroupLayout(0),[n,l]),{commandEncoder:c,ts:d}=S(u,f,D(r));s=o?null:_(c,n),P(c);let p=await G(d);if(o)return p!==void 0?{gpuTimeMs:p}:{};let g=await y(s,Float32Array);return s=null,p!==void 0?{x:g,gpuTimeMs:p}:g}finally{!o&&n&&m(n),l&&m(l),s&&m(s)}}async function Ae(a,r,e,i,t,o){let u=e instanceof h,n=t instanceof h;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(r)||!Number.isInteger(i)||!Number.isInteger(o))throw new Error("n, incx, and incy must be integers.");if(i<=0||o<=0)throw new Error("incx and incy must be positive.");if(!(e instanceof Float32Array)&&!(e instanceof h))throw new Error("x must be a Float32Array or GpuVector.");if(!(t instanceof Float32Array)&&!(t instanceof h))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 u?{}:{x:e,y:t};if(e.length<(r-1)*i+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(r-1)*o+1)throw new Error("y does not have enough elements for the given n and incy.");let l=await A(a,"sswap"),s=null,f=null,c=null,d=null,p=null;try{s=u?e._buf:b(e,"sswap-x",!0),f=n?t._buf:b(t,"sswap-y",!0),c=L([{value:r,type:"u32"},{value:i,type:"u32"},{value:o,type:"u32"}],"sswap-params");let g=B(l.getBindGroupLayout(0),[s,f,c]),{commandEncoder:w,ts:k}=S(l,g,D(r));d=u?null:_(w,s),p=n?null:_(w,f),P(w);let x=await G(k);if(u&&n)return x!==void 0?{gpuTimeMs:x}:{};let E=await y(d,Float32Array);d=null;let v=await y(p,Float32Array);return p=null,x!==void 0?{x:E,y:v,gpuTimeMs:x}:{x:E,y:v}}finally{!u&&s&&m(s),!n&&f&&m(f),c&&m(c),d&&m(d),p&&m(p)}}async function Ge(a,r,e,i,t,o,u){let n=i instanceof h,l=o instanceof h;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(r)||!Number.isInteger(t)||!Number.isInteger(u))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||u<=0)throw new Error("incx and incy must be positive.");if(!n&&!(i instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!l&&!(o 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 l?{}:{y:o};if(i.length<(r-1)*t+1)throw new Error("x does not have enough elements for the given n and incx.");if(o.length<(r-1)*u+1)throw new Error("y does not have enough elements for the given n and incy.");let s=await A(a,"saxpy"),f=null,c=null,d=null,p=null;try{f=n?i._buf:b(i,"saxpy-x",!1),c=l?o._buf:b(o,"saxpy-y",!0),d=L([{value:r,type:"u32"},{value:e,type:"f32"},{value:t,type:"u32"},{value:u,type:"u32"}],"saxpy-params");let g=B(s.getBindGroupLayout(0),[f,c,d]),{commandEncoder:w,ts:k}=S(s,g,D(r));p=l?null:_(w,c),P(w);let x=await G(k);if(l&&n)return x!==void 0?{gpuTimeMs:x}:{};let E=await y(p,Float32Array);return p=null,x!==void 0?{y:E,gpuTimeMs:x}:{y:E}}finally{!n&&f&&m(f),!l&&c&&m(c),d&&m(d),p&&m(p)}}async function Pe(a,r,e,i,t,o){let u=e instanceof h,n=t instanceof h;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(r)||!Number.isInteger(i)||!Number.isInteger(o))throw new Error("n, incx, and incy must be integers.");if(i<=0||o<=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)return n?{}:{y:t};if(e.length<(r-1)*i+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(r-1)*o+1)throw new Error("y does not have enough elements for the given n and incy.");let l=await A(a,"scopy"),s=null,f=null,c=null,d=null;try{s=u?e._buf:b(e,"scopy-x",!1),f=n?t._buf:b(t,"scopy-y",!0),c=L([{value:r,type:"u32"},{value:i,type:"u32"},{value:o,type:"u32"}],"scopy-params");let p=B(l.getBindGroupLayout(0),[s,f,c]),{commandEncoder:g,ts:w}=S(l,p,D(r));d=n?null:_(g,f),P(g);let k=await G(w);if(n&&u)return k!==void 0?{gpuTimeMs:k}:{};let x=await y(d,Float32Array);return d=null,k!==void 0?{y:x,gpuTimeMs:k}:{y:x}}finally{!u&&s&&m(s),!n&&f&&m(f),c&&m(c),d&&m(d)}}var Se=64;async function Fe(a,r,e,i,t,o){let u=e instanceof h,n=t instanceof h;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(r)||!Number.isInteger(i)||!Number.isInteger(o))throw new Error("n, incx, and incy must be integers.");if(i<=0||o<=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)return{dot:0};if(e.length<(r-1)*i+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(r-1)*o+1)throw new Error("y does not have enough elements for the given n and incy.");let l=await A(a,"sdot"),s=await A(a,"reduction/sum"),f=null,c=null,d=null,p=null,g=null,w=null;try{f=u?e._buf:b(e,"sdot-x",!1),c=n?t._buf:b(t,"sdot-y",!1),d=H(2*Se*4,"sdot-partials"),p=O(4,"sdot-result"),g=L([{value:r,type:"u32"},{value:i,type:"u32"},{value:o,type:"u32"}],"sdot-params");let k=B(l.getBindGroupLayout(0),[f,c,d,g]),{commandEncoder:x,ts:E}=S(l,k,2*Se);P(x);let v=B(s.getBindGroupLayout(0),[d,p]),{commandEncoder:F,ts:I}=S(s,v,1);w=_(F,p),P(F);let j=y(w,Float32Array);w=null;let[N,U,T]=await Promise.all([G(E),G(I),j]);return N!==void 0&&U!==void 0?{dot:T[0],gpuTimeMs:N+U}:{dot:T[0]}}finally{!u&&f&&m(f),!n&&c&&m(c),d&&m(d),p&&m(p),g&&m(g),w&&m(w)}}var Le=64;async function Ie(a,r,e,i){let t=e instanceof h;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(r)||!Number.isInteger(i))throw new Error("n and incx must be integers.");if(i<=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)*i+1)throw new Error("x does not have enough elements for the given n and incx.");let o=await A(a,"sasum"),u=await A(a,"reduction/sum"),n=null,l=null,s=null,f=null,c=null;try{n=t?e._buf:b(e,"sasum-x",!1),l=H(2*Le*4,"sasum-partials"),s=O(4,"sasum-result"),f=L([{value:r,type:"u32"},{value:i,type:"u32"}],"sasum-params");let d=B(o.getBindGroupLayout(0),[n,l,f]),{commandEncoder:p,ts:g}=S(o,d,2*Le);P(p);let w=B(u.getBindGroupLayout(0),[l,s]),{commandEncoder:k,ts:x}=S(u,w,1);c=_(k,s),P(k);let E=y(c,Float32Array);c=null;let[v,F,I]=await Promise.all([G(g),G(x),E]);return v!==void 0&&F!==void 0?{asum:I[0],gpuTimeMs:v+F}:{asum:I[0]}}finally{!t&&n&&m(n),l&&m(l),s&&m(s),f&&m(f),c&&m(c)}}var mr=64;async function je(a,r,e,i){let t=e instanceof h;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(r)||!Number.isInteger(i))throw new Error("n and incx must be integers.");if(i<=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)*i+1)throw new Error("x does not have enough elements for the given n and incx.");let o=await A(a,["f64add","dasum"]),u=await A(a,["f64add","reduction/sumF64"]),n=null,l=null,s=null,f=null,c=null,d=null,p=null,g=null;try{n=t?e:h.from(e),l=H(2*mr*4,"dasum-partialsMain"),s=H(2*mr*4,"dasum-partialsAux"),f=O(4,"dasum-result-main"),c=O(4,"dasum-result-aux"),d=L([{value:r,type:"u32"},{value:i,type:"u32"}],"dasum-params");let w=B(o.getBindGroupLayout(0),[n._buf,n._auxBuf,l,s,d]),{commandEncoder:k,ts:x}=S(o,w,2*mr);P(k);let E=B(u.getBindGroupLayout(0),[l,s,f,c]),{commandEncoder:v,ts:F}=S(u,E,1);p=_(v,f),g=_(v,c),P(v);let I=y(p,Float32Array),j=y(g,Uint32Array);p=null,g=null;let[N,U,T,R]=await Promise.all([G(x),G(F),I,j]),V=rr(T[0],R[0]);return N!==void 0&&U!==void 0?{asum:V,gpuTimeMs:N+U}:{asum:V}}finally{!t&&n&&n.destroy(),l&&m(l),s&&m(s),f&&m(f),c&&m(c),d&&m(d),p&&m(p),g&&m(g)}}var Ne=64;async function We(a,r,e,i){let t=e instanceof h;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(r)||!Number.isInteger(i))throw new Error("n and incx must be integers.");if(i<=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)*i+1)throw new Error("x does not have enough elements for the given n and incx.");let o=await A(a,"snrm2"),u=await A(a,"reduction/sum"),n=null,l=null,s=null,f=null,c=null;try{n=t?e._buf:b(e,"snrm2-x",!1),l=H(2*Ne*4,"snrm2-partials"),s=O(4,"snrm2-result"),f=L([{value:r,type:"u32"},{value:i,type:"u32"}],"snrm2-params");let d=B(o.getBindGroupLayout(0),[n,l,f]),{commandEncoder:p,ts:g}=S(o,d,2*Ne);P(p);let w=B(u.getBindGroupLayout(0),[l,s]),{commandEncoder:k,ts:x}=S(u,w,1);c=_(k,s),P(k);let E=y(c,Float32Array);c=null;let[v,F,I]=await Promise.all([G(g),G(x),E]),j=Math.sqrt(I[0]);return v!==void 0&&F!==void 0?{nrm2:j,gpuTimeMs:v+F}:{nrm2:j}}finally{!t&&n&&m(n),l&&m(l),s&&m(s),f&&m(f),c&&m(c)}}var pr=64;async function Me(a,r,e,i){let t=e instanceof h;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(r)||!Number.isInteger(i))throw new Error("n and incx must be integers.");if(i<=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)*i+1)throw new Error("x does not have enough elements for the given n and incx.");let o=await A(a,"isamax"),u=await A(a,"reduction/argmax"),n=null,l=null,s=null,f=null,c=null,d=null;try{n=t?e._buf:b(e,"isamax-x",!1),l=H(2*pr*4,"isamax-partials-val"),s=H(2*pr*4,"isamax-partials-idx"),f=O(4,"isamax-result"),c=L([{value:r,type:"u32"},{value:i,type:"u32"}],"isamax-params");let p=B(o.getBindGroupLayout(0),[n,l,s,c]),{commandEncoder:g,ts:w}=S(o,p,2*pr);P(g);let k=B(u.getBindGroupLayout(0),[l,s,f]),{commandEncoder:x,ts:E}=S(u,k,1);d=_(x,f),P(x);let v=y(d,Uint32Array);d=null;let[F,I,j]=await Promise.all([G(w),G(E),v]),N=j[0];return F!==void 0&&I!==void 0?{index:N,gpuTimeMs:F+I}:{index:N}}finally{!t&&n&&m(n),l&&m(l),s&&m(s),f&&m(f),c&&m(c),d&&m(d)}}async function Ue(a,r,e,i,t,o,u,n){let l=e instanceof h,s=t instanceof h;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(r)||!Number.isInteger(i)||!Number.isInteger(o))throw new Error("n, incx, and incy must be integers.");if(typeof u!="number")throw new Error("c must be a number.");if(typeof n!="number")throw new Error("s must be a number.");if(Number.isNaN(u)||Number.isNaN(n))throw new Error("c and s must not be NaN.");if(!Number.isFinite(u))throw new Error("c must be finite.");if(!Number.isFinite(n))throw new Error("s must be finite.");if(i<=0||o<=0)throw new Error("incx and incy must be positive.");if(!l&&!(e 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(l!==s)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(r<=0)return l?{}:{x:e,y:t};if(e.length<(r-1)*i+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(r-1)*o+1)throw new Error("y does not have enough elements for the given n and incy.");let f=await A(a,"srot"),c=null,d=null,p=null,g=null,w=null;try{c=l?e._buf:b(e,"srot-x",!0),d=s?t._buf:b(t,"srot-y",!0),p=L([{value:r,type:"u32"},{value:u,type:"f32"},{value:n,type:"f32"},{value:i,type:"u32"},{value:o,type:"u32"}],"srot-params");let k=B(f.getBindGroupLayout(0),[c,d,p]),{commandEncoder:x,ts:E}=S(f,k,D(r));g=l?null:_(x,c),w=s?null:_(x,d),P(x);let v=await G(E);if(l&&s)return v!==void 0?{gpuTimeMs:v}:{};let F=y(g,Float32Array),I=y(w,Float32Array);g=null,w=null;let[j,N]=await Promise.all([F,I]);return v!==void 0?{x:j,y:N,gpuTimeMs:v}:{x:j,y:N}}finally{!l&&c&&m(c),!s&&d&&m(d),p&&m(p),g&&m(g),w&&m(w)}}async function Te(a,r,e,i,t,o,u){let n=e instanceof h,l=t instanceof h;if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!Number.isInteger(r)||!Number.isInteger(i)||!Number.isInteger(o))throw new Error("n, incx, and incy must be integers.");if(!(u instanceof Float32Array)||u.length!==5)throw new Error("param must be a Float32Array of length 5.");if(i<=0||o<=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||u[0]===-2)return n?{}:{x:e,y:t};if(e.length<(r-1)*i+1)throw new Error("x does not have enough elements for the given n and incx.");if(t.length<(r-1)*o+1)throw new Error("y does not have enough elements for the given n and incy.");let s=await A(a,"srotm"),f=null,c=null,d=null,p=null,g=null,w=null;try{f=n?e._buf:b(e,"srotm-x",!0),c=l?t._buf:b(t,"srotm-y",!0),d=b(u,"srotm-param",!1),p=L([{value:r,type:"u32"},{value:i,type:"u32"},{value:o,type:"u32"}],"srotm-params");let k=B(s.getBindGroupLayout(0),[f,c,d,p]),{commandEncoder:x,ts:E}=S(s,k,D(r));g=n?null:_(x,f),w=l?null:_(x,c),P(x);let v=await G(E);if(n&&l)return v!==void 0?{gpuTimeMs:v}:{};let F=y(g,Float32Array),I=y(w,Float32Array);g=null,w=null;let[j,N]=await Promise.all([F,I]);return v!==void 0?{x:j,y:N,gpuTimeMs:v}:{x:j,y:N}}finally{!n&&f&&m(f),!l&&c&&m(c),d&&m(d),p&&m(p),g&&m(g),w&&m(w)}}async function Re(a,r,e,i,t,o,u,n,l,s,f,c){let d=n instanceof h,p=f instanceof h,g=o instanceof C,w=r==="no-transpose";if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!w&&r!=="transpose")throw new Error("trans must be 'no-transpose' or 'transpose'.");if(!Number.isInteger(e)||!Number.isInteger(i)||!Number.isInteger(l)||!Number.isInteger(c)||!Number.isInteger(u))throw new Error("m, n, incx, incy, and lda must be integers.");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 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(l<=0||c<=0)throw new Error("incx and incy must be positive.");if(u<i)throw new Error("lda must be >= n.");if(!g&&!(o instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!d&&!(n instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!p&&!(f instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(d!==p)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(d&&!g)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(d&&n._buf===f._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(g&&u!==o.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(g&&(o.rows<e||o.cols<i))throw new Error("A is too small for the given m and n.");if(e<0||i<0)throw new Error("m and n must be non-negative.");if(e===0||i===0)return p?{}:{y:f};let k=w?i:e,x=w?e:i;if(!g&&o.length<(e-1)*u+i)throw new Error("A does not have enough elements for the given m, n, and lda.");if(n.length<(k-1)*l+1)throw new Error("x does not have enough elements for the given dimensions and incx.");if(f.length<(x-1)*c+1)throw new Error("y does not have enough elements for the given dimensions and incy.");let v=await A(a,w?"sgemv_n":"sgemv_t"),F=g?o._buf:b(o,"sgemv-A",!1),I=d?n._buf:b(n,"sgemv-x",!1),j=p?f._buf:b(f,"sgemv-y",!0),N=L([{value:e,type:"u32"},{value:i,type:"u32"},{value:t,type:"f32"},{value:s,type:"f32"},{value:l,type:"u32"},{value:c,type:"u32"},{value:u,type:"u32"}],"sgemv-params");try{let U=B(v.getBindGroupLayout(0),[F,I,j,N]),T=w?Math.min(e,a.limits.maxComputeWorkgroupsPerDimension):D(x),{commandEncoder:R,ts:V}=S(v,U,T),z=p?null:_(R,j);P(R);let Y=await G(V);if(p)return Y!==void 0?{gpuTimeMs:Y}:{};let Z=await y(z,Float32Array);return Y!==void 0?{y:Z,gpuTimeMs:Y}:{y:Z}}finally{g||m(F),d||m(I),p||m(j),m(N)}}async function Ve(a,r,e,i,t,o,u,n,l,s,f){let c=u instanceof h,d=s instanceof h,p=t instanceof C,g=r==="lower";if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!g&&r!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(!Number.isInteger(e)||!Number.isInteger(n)||!Number.isInteger(f)||!Number.isInteger(o))throw new Error("n, incx, incy, and lda must be integers.");if(typeof i!="number")throw new Error("alpha must be a number.");if(Number.isNaN(i))throw new Error("alpha must not be NaN.");if(!Number.isFinite(i))throw new Error("alpha must be finite.");if(typeof l!="number")throw new Error("beta must be a number.");if(Number.isNaN(l))throw new Error("beta must not be NaN.");if(!Number.isFinite(l))throw new Error("beta must be finite.");if(n<=0||f<=0)throw new Error("incx and incy must be positive.");if(o<e)throw new Error("lda must be >= n.");if(!p&&!(t instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!c&&!(u instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!d&&!(s instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(c!==d)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(c&&!p)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(c&&u._buf===s._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(p&&o!==t.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(p&&(t.rows<e||t.cols<e))throw new Error("A is too small for the given n.");if(e<0)throw new Error("n must be non-negative.");if(e===0)return d?{}:{y:s};if(!p&&t.length<(e-1)*o+e)throw new Error("A does not have enough elements for the given n and lda.");if(u.length<(e-1)*n+1)throw new Error("x does not have enough elements for the given n and incx.");if(s.length<(e-1)*f+1)throw new Error("y does not have enough elements for the given n and incy.");let w=await A(a,"ssymv"),k=null,x=null,E=null,v=null;try{k=p?t._buf:b(t,"ssymv-A",!1),x=c?u._buf:b(u,"ssymv-x",!1),E=d?s._buf:b(s,"ssymv-y",!0),v=L([{value:e,type:"u32"},{value:i,type:"f32"},{value:l,type:"f32"},{value:n,type:"u32"},{value:f,type:"u32"},{value:o,type:"u32"},{value:g?0:1,type:"u32"}],"ssymv-params");let F=B(w.getBindGroupLayout(0),[k,x,E,v]),I=Math.min(e,a.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:j,ts:N}=S(w,F,I),U=d?null:_(j,E);P(j);let T=await G(N);if(d)return T!==void 0?{gpuTimeMs:T}:{};let R=await y(U,Float32Array);return T!==void 0?{y:R,gpuTimeMs:T}:{y:R}}finally{!p&&k&&m(k),!c&&x&&m(x),!d&&E&&m(E),v&&m(v)}}async function He(a,r,e,i,t,o,u,n,l,s,f){let c=n instanceof h,d=s instanceof h,p=o instanceof C,g=r==="lower",w=e==="no-transpose",k=i==="unit";if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!g&&r!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(!w&&e!=="transpose")throw new Error("trans must be 'no-transpose' or 'transpose'.");if(!k&&i!=="non-unit")throw new Error("diag must be 'unit' or 'non-unit'.");if(!Number.isInteger(t)||!Number.isInteger(l)||!Number.isInteger(f)||!Number.isInteger(u))throw new Error("n, incx, incy, and lda must be integers.");if(l<=0||f<=0)throw new Error("incx and incy must be positive.");if(u<t)throw new Error("lda must be >= n.");if(!p&&!(o instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!c&&!(n instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(!d&&!(s instanceof Float32Array))throw new Error("y must be a Float32Array or GpuVector.");if(c!==d)throw new Error("x and y must be the same type (both Float32Array or both GpuVector).");if(c&&n._buf===s._buf)throw new Error("x and y must not reference the same GPU buffer when both are GpuVectors.");if(c&&!p)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(p&&d&&o._buf===s._buf)throw new Error("A and y must not reference the same GPU buffer.");if(p&&u!==o.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(p&&(o.rows<t||o.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:s};if(!p&&o.length<(t-1)*u+t)throw new Error("A does not have enough elements for the given n and lda.");if(n.length<(t-1)*l+1)throw new Error("x does not have enough elements for the given n and incx.");if(s.length<(t-1)*f+1)throw new Error("y does not have enough elements for the given n and incy.");let x=await A(a,"strmv"),E=null,v=null,F=null,I=null;try{E=p?o._buf:b(o,"strmv-A",!1),v=c?n._buf:b(n,"strmv-x",!1),F=d?s._buf:b(s,"strmv-y",!0),I=L([{value:t,type:"u32"},{value:l,type:"u32"},{value:f,type:"u32"},{value:u,type:"u32"},{value:w?0:1,type:"u32"},{value:g?0:1,type:"u32"},{value:k?1:0,type:"u32"}],"strmv-params");let j=B(x.getBindGroupLayout(0),[E,v,F,I]),N=Math.min(t,a.limits.maxComputeWorkgroupsPerDimension),{commandEncoder:U,ts:T}=S(x,j,N),R=d?null:_(U,F);P(U);let V=await G(T);if(d)return V!==void 0?{gpuTimeMs:V}:{};let z=await y(R,Float32Array);return V!==void 0?{y:z,gpuTimeMs:V}:{y:z}}finally{!p&&E&&m(E),!c&&v&&m(v),!d&&F&&m(F),I&&m(I)}}var q=64;function De(a,r,e){let i=new ArrayBuffer(a*r),t=new DataView(i);for(let o=0;o<a;o++){let u=e(o),n=o*r;u.forEach((l,s)=>t.setUint32(n+s*4,l,!0))}return i}function Ce(a,r,e){let i=a.createBuffer({label:e,size:r.byteLength,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});return a.queue.writeBuffer(i,0,r),i}async function Oe(a,r,e,i,t,o,u,n,l){let s=n instanceof h,f=o instanceof C,c=r==="lower",d=e==="no-transpose",p=i==="unit";if(!(a instanceof GPUDevice))throw new Error("device must be a GPUDevice.");if(!c&&r!=="upper")throw new Error("uplo must be 'lower' or 'upper'.");if(!d&&e!=="transpose")throw new Error("trans must be 'no-transpose' or 'transpose'.");if(!p&&i!=="non-unit")throw new Error("diag must be 'unit' or 'non-unit'.");if(!Number.isInteger(t)||!Number.isInteger(l)||!Number.isInteger(u))throw new Error("n, incx, and lda must be integers.");if(l<=0)throw new Error("incx must be positive.");if(u<t)throw new Error("lda must be >= n.");if(!f&&!(o instanceof Float32Array))throw new Error("A must be a Float32Array or GpuMatrix.");if(!s&&!(n instanceof Float32Array))throw new Error("x must be a Float32Array or GpuVector.");if(s&&!f)throw new Error("A must be a GpuMatrix when x is a GpuVector.");if(f&&u!==o.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(f&&(o.rows<t||o.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 s?{}:{x:n};if(!f&&o.length<(t-1)*u+t)throw new Error("A does not have enough elements for the given n and lda.");if(n.length<(t-1)*l+1)throw new Error("x does not have enough elements for the given n and incx.");let g=await A(a,"strsv_invert_block"),w=await A(a,"strsv_apply_inverse"),k=await A(a,"strsv_update"),x=d===c,E=[];for(let z=0;z<t;z+=q)E.push(z);x||E.reverse();let v=E.length,F=a.limits.maxComputeWorkgroupsPerDimension,I=a.limits.minUniformBufferOffsetAlignment,j=null,N=null,U=null,T=null,R=null,V=null;try{j=f?o._buf:b(o,"strsv-A",!1),N=s?n._buf:b(n,"strsv-x",!0),U=H(v*q*q*4,"strsv-Ainv");let z=De(v,I,K=>{let X=K*q,tr=Math.min(X+q,t);return[l,K,X,tr]});T=Ce(a,z,"strsv-apply-params");let Y=De(v,I,K=>{let X=K*q,tr=Math.min(X+q,t);return[t,l,u,d?0:1,c?0:1,X,tr]});R=Ce(a,Y,"strsv-update-params");let{commandEncoder:Z,querySet:er}=fr();V=L([{value:t,type:"u32"},{value:u,type:"u32"},{value:d?0:1,type:"u32"},{value:c?0:1,type:"u32"},{value:p?1:0,type:"u32"}],"strsv-invert-params");let ze=B(g.getBindGroupLayout(0),[j,U,V]);ar(Z,g,ze,{x:q,y:v},er?{timestampWrites:{querySet:er,beginningOfPassWriteIndex:0}}:void 0);for(let K=0;K<E.length;K++){let X=E[K],tr=Math.min(X+q,t),Ze=X/q,Ke=K===E.length-1,wr=Ze*I,Xe=B(w.getBindGroupLayout(0),[U,N,{buffer:T,offset:wr,size:16}]);ar(Z,w,Xe,1,Ke&&er?{timestampWrites:{querySet:er,endOfPassWriteIndex:1}}:void 0);let br=x?t-tr:X;if(br===0)continue;let $e=B(k.getBindGroupLayout(0),[j,N,{buffer:R,offset:wr,size:32}]),Ye=Math.min(br,F);ar(Z,k,$e,Ye)}let Qe=ur(Z,er),qe=s?null:_(Z,N);P(Z);let ir=await G(Qe);if(s)return ir!==void 0?{gpuTimeMs:ir}:{};let gr=await y(qe,Float32Array);return ir!==void 0?{x:gr,gpuTimeMs:ir}:{x:gr}}finally{!f&&j&&m(j),!s&&N&&m(N),U&&m(U),T&&m(T),R&&m(R),V&&m(V)}}return it(Rt);})();
+`});var je={};yr(je,{shaderSources:()=>Kt});var Kt,Le=I(()=>{Ir();Ur();Vr();Hr();Cr();zr();qr();Kr();$r();Jr();ee();ae();ie();se();le();ce();de();ge();be();xe();ye();Ee();Ae();ke();Se();Kt={"reduction/argmax":Mr,"reduction/sum":Tr,"reduction/sumF64":Rr,sscal:Dr,sswap:Or,saxpy:Qr,scopy:Zr,sdot:Xr,sasum:Yr,snrm2:re,srot:te,srotm:oe,isamax:ne,sgemv_n:ue,sgemv_t:fe,ssymv:me,strmv:pe,sger:we,ssyr:he,ssyr2:ve,f64add:_e,dasum:Ge,strsv_invert_block:Be,strsv_apply_inverse:Pe,strsv_update:Fe}});var ra={};yr(ra,{GpuMatrix:()=>H,GpuVector:()=>x,cleanup:()=>Pr,dasum:()=>De,gpuName:()=>Sr,init:()=>kr,isamax:()=>ze,randomFloat32Array:()=>Nr,randomFloat64Array:()=>Wr,sasum:()=>He,saxpy:()=>Me,scopy:()=>Ue,sdot:()=>Ve,sgemv:()=>Ze,sger:()=>rt,snrm2:()=>Oe,srot:()=>Qe,srotm:()=>qe,sscal:()=>We,sswap:()=>Ie,ssymv:()=>Ke,ssyr:()=>et,ssyr2:()=>tt,strmv:()=>Xe,strsv:()=>Je});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:n}=await import("webgpu");Object.assign(globalThis,n),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 ${t} bytes.`);let n=e?GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC:GPUBufferUsage.STORAGE,u=o.createBuffer({label:r,size:i,usage:n,mappedAtCreation:!0}),s=a.constructor;return new s(u.getMappedRange()).set(a),u.unmap(),u}function D(a,r="blas-storage"){return V().createBuffer({label:r,size:a,usage:GPUBufferUsage.STORAGE})}function q(a,r="blas-result"){return V().createBuffer({label:r,size:a,usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC})}function E(a,r){let o=V().createBuffer({label:"blas-readback",size:r.size,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});return a.copyBufferToBuffer(r,0,o,0,r.size),o}function N(a,r="blas-params"){let e=V(),o=a.length*4,t=Math.ceil(o/16)*16,i=new ArrayBuffer(t),n=new DataView(i);a.forEach(({value:s,type:l},f)=>{let c=f*4;if(l==="u32")n.setUint32(c,s,!0);else if(l==="i32")n.setInt32(c,s,!0);else if(l==="f32")n.setFloat32(c,s,!0);else throw new Error(`Unknown param type "${l}". Use "f32", "u32", or "i32".`)});let u=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 bt=new ArrayBuffer(8),er=new DataView(bt),Fr=new ArrayBuffer(4),jr=new Uint32Array(Fr),Lr=new Float32Array(Fr);function ht(a){return jr[0]=a>>>0,Lr[0]}function xt(a){return Lr[0]=a,jr[0]}function vt(a,r,e,o){let t=r>>>3,i=r&7,n=o>>>29,u=e<<3|n,s=o&536870911,l=a<<31|t<<23|u,f=i>>>2&1,c=i&3,p=s>>>23,m=s&8388607,g=c<<6|p,w=(f<<31|g<<23|m)>>>0;return[ht(l),w]}function yt(a,r){let e=xt(a);r=r>>>0;let o=e>>>31,t=e>>>23&255,i=e&8388607,n=r>>>31,u=r>>>23&255,s=r&8388607,l=n<<2|u>>>6,f=(u&63)<<23|s,c=t<<3|l,p=i>>>3,g=((i&7)<<29|f)>>>0;return{sign:o,rawExp:c,mantissaHi:p,lo:g}}var _t=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>=_t)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 vt(o,t,i,e)}function tr(a,r){let{sign:e,rawExp:o,mantissaHi:t,lo:i}=yt(a,r),n=(e<<31|o<<20|t)>>>0;return er.setUint32(0,n,!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 s=mr(r[u]);o[u]=s[0],t[u]=s[1]}let i=b(o,"gpu-vector-f64-main",!0),n=b(t,"gpu-vector-f64-aux",!0);return new a(i,r.length,Float64Array,n)}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[n,u]=await Promise.all([_(o,Float32Array),_(i,Uint32Array)]),s=new Float64Array(this.length);for(let l=0;l<this.length;l++)s[l]=tr(n[l],u[l]);return s}destroy(){this._buf.destroy(),this._auxBuf&&this._auxBuf.destroy()}};var H=class a{constructor(r,e,o,t,i=null,n="row-major"){this._buf=r,this._auxBuf=i,this.rows=e,this.cols=o,this.lda=t,this.layout=n}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 n=i==="row-major";if(t===void 0&&(t=n?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=n?o:e;if(!Number.isInteger(t)||t<u)throw new Error(`lda must be an integer >= ${n?"cols":"rows"}.`);let s=n?e:o;if(r.length<s*t)throw new Error("data does not have enough elements for the given rows, cols, and lda.");if(r instanceof Float64Array){let f=s*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,s*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,n=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===n)return m;let g=new Float64Array(i*n);for(let w=0;w<i;w++)g.set(m.subarray(w*this.lda,w*this.lda+n),w*n);return g}let u=await _(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._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 A(a,r,e=0){let o=V(),t=r.map((i,n)=>({binding:e+n,resource:i instanceof GPUBuffer?{buffer:i}:i}));return o.createBindGroup({layout:a,entries:t})}var Et=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(),Et.set(a,i)}function L(a,r,e){let{commandEncoder:o,querySet:t,passDescriptor:i}=pr();ir(o,a,r,e,i);let n=cr(o,t);return{commandEncoder:o,ts:n}}var Yt={},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 $t(t,e)),o.get(i)}async function Xt(a){if(typeof process>"u"||!process.versions?.node){let{shaderSources:r}=await Promise.resolve().then(()=>(Le(),je)),e=r[a];if(!e)throw new Error(`Shader "${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(Yt.url));return r(t(i,`../shaders/${a}.wgsl`),"utf8")}}async function $t(a,r="main"){let e=V(),o=a.join("+"),t=(await Promise.all(a.map(Xt))).join(`
+`),i=e.createShaderModule({label:o,code:t}),u=(await i.getCompilationInfo()).messages.filter(f=>f.type==="error");if(u.length>0)throw new Error(`Shader "${o}" compilation failed:
+${u.map(f=>`  line ${f.lineNum}: ${f.message}`).join(`
+`)}`);let s=r==="main"?{module:i}:{module:i,entryPoint:r},l=e.createComputePipeline({label:o,layout:"auto",compute:s});return l._shaderModule=i,l}var Jt=64,Ne=8;function C(a,r){let e=V().limits.maxComputeWorkgroupsPerDimension;return r===void 0?Math.min(Math.ceil(a/Jt),e):{x:Math.min(Math.ceil(r/Ne),e),y:Math.min(Math.ceil(a/Ne),e)}}async function We(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 n=await B(a,"sscal"),u=null,s=null,l=null;try{u=i?o._buf:b(o,"sscal-x",!0),s=N([{value:r,type:"u32"},{value:e,type:"f32"},{value:t,type:"u32"}],"sscal-params");let f=A(n.getBindGroupLayout(0),[u,s]),{commandEncoder:c,ts:p}=L(n,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),s&&d(s),l&&d(l)}}async function Ie(a,r,e,o,t,i){let n=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 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 s=await B(a,"sswap"),l=null,f=null,c=null,p=null,m=null;try{l=n?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(s.getBindGroupLayout(0),[l,f,c]),{commandEncoder:w,ts:y}=L(s,g,C(r));p=n?null:E(w,l),m=u?null:E(w,f),S(w);let h=await P(y);if(n&&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{!n&&l&&d(l),!u&&f&&d(f),c&&d(c),p&&d(p),m&&d(m)}}async function Me(a,r,e,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(r)||!Number.isInteger(t)||!Number.isInteger(n))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||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(r<=0)return s?{}:{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)*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,m=null;try{f=u?o._buf:b(o,"saxpy-x",!1),c=s?i._buf:b(i,"saxpy-y",!0),p=N([{value:r,type:"u32"},{value:e,type:"f32"},{value:t,type:"u32"},{value:n,type:"u32"}],"saxpy-params");let g=A(l.getBindGroupLayout(0),[f,c,p]),{commandEncoder:w,ts:y}=L(l,g,C(r));m=s?null:E(w,c),S(w);let h=await P(y);if(s&&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),!s&&c&&d(c),p&&d(p),m&&d(m)}}async function Ue(a,r,e,o,t,i){let n=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(!n&&!(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(n!==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 s=await B(a,"scopy"),l=null,f=null,c=null,p=null;try{l=n?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(s.getBindGroupLayout(0),[l,f,c]),{commandEncoder:g,ts:w}=L(s,m,C(r));p=u?null:E(g,f),S(g);let y=await P(w);if(u&&n)return y!==void 0?{gpuTimeMs:y}:{};let h=await _(p,Float32Array);return p=null,y!==void 0?{y:h,gpuTimeMs:y}:{y:h}}finally{!n&&l&&d(l),!u&&f&&d(f),c&&d(c),p&&d(p)}}var Te=64;async function Ve(a,r,e,o,t,i){let n=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(!n&&!(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(n!==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 s=await B(a,"sdot"),l=await B(a,"reduction/sum"),f=null,c=null,p=null,m=null,g=null,w=null;try{f=n?e._buf:b(e,"sdot-x",!1),c=u?t._buf:b(t,"sdot-y",!1),p=D(2*Te*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(s.getBindGroupLayout(0),[f,c,p,g]),{commandEncoder:h,ts:G}=L(s,y,2*Te);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{!n&&f&&d(f),!u&&c&&d(c),p&&d(p),m&&d(m),g&&d(g),w&&d(w)}}var Re=64;async function He(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"),n=await B(a,"reduction/sum"),u=null,s=null,l=null,f=null,c=null;try{u=t?e._buf:b(e,"sasum-x",!1),s=D(2*Re*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,s,f]),{commandEncoder:m,ts:g}=L(i,p,2*Re);S(m);let w=A(n.getBindGroupLayout(0),[s,l]),{commandEncoder:y,ts:h}=L(n,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),s&&d(s),l&&d(l),f&&d(f),c&&d(c)}}var wr=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 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"]),n=await B(a,["f64add","reduction/sumF64"]),u=null,s=null,l=null,f=null,c=null,p=null,m=null,g=null;try{u=t?e:x.from(e),s=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,s,l,p]),{commandEncoder:y,ts:h}=L(i,w,2*wr);S(y);let G=A(n.getBindGroupLayout(0),[s,l,f,c]),{commandEncoder:v,ts:F}=L(n,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(),s&&d(s),l&&d(l),f&&d(f),c&&d(c),p&&d(p),m&&d(m),g&&d(g)}}var Ce=64;async function Oe(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"),n=await B(a,"reduction/sum"),u=null,s=null,l=null,f=null,c=null;try{u=t?e._buf:b(e,"snrm2-x",!1),s=D(2*Ce*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,s,f]),{commandEncoder:m,ts:g}=L(i,p,2*Ce);S(m);let w=A(n.getBindGroupLayout(0),[s,l]),{commandEncoder:y,ts:h}=L(n,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),s&&d(s),l&&d(l),f&&d(f),c&&d(c)}}var br=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{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"),n=await B(a,"reduction/argmax"),u=null,s=null,l=null,f=null,c=null,p=null;try{u=t?e._buf:b(e,"isamax-x",!1),s=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,s,l,c]),{commandEncoder:g,ts:w}=L(i,m,2*br);S(g);let y=A(n.getBindGroupLayout(0),[s,l,f]),{commandEncoder:h,ts:G}=L(n,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),s&&d(s),l&&d(l),f&&d(f),c&&d(c),p&&d(p)}}async function Qe(a,r,e,o,t,i,n,u){let s=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 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&&!(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(s!==l)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 f=await B(a,"srot"),c=null,p=null,m=null,g=null,w=null;try{c=s?e._buf:b(e,"srot-x",!0),p=l?t._buf:b(t,"srot-y",!0),m=N([{value:r,type:"u32"},{value:n,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=s?null:E(h,c),w=l?null:E(h,p),S(h);let v=await P(G);if(s&&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{!s&&c&&d(c),!l&&p&&d(p),m&&d(m),g&&d(g),w&&d(w)}}async function qe(a,r,e,o,t,i,n){let u=e instanceof x,s=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(!(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&&!(e 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(r<=0||n[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=s?t._buf:b(t,"srotm-y",!0),p=b(n,"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=s?null:E(h,c),S(h);let v=await P(G);if(u&&s)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),!s&&c&&d(c),p&&d(p),m&&d(m),g&&d(g),w&&d(w)}}async function Ze(a,r,e,o,t,i,n,u,s,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(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(!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&&n!==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(n<o)throw new Error("lda must be >= n.");if(!m&&i.length<(e-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 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:s,type:"u32"},{value:c,type:"u32"},{value:n,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 Ke(a,r,e,o,t,i,n,u,s,l,f,c="row-major"){let p=n 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 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<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&&!(n 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&&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<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(n.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?n._buf:b(n,"ssymv-x",!1),F=m?l._buf:b(l,"ssymv-y",!0),k=N([{value:e,type:"u32"},{value:o,type:"f32"},{value:s,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 Xe(a,r,e,o,t,i,n,u,s,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(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(!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&&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 m?{}:{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?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: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 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 $e(a,r,e){let o=new ArrayBuffer(a*r),t=new DataView(o);for(let i=0;i<a;i++){let n=e(i),u=i*r;n.forEach((s,l)=>t.setUint32(u+l*4,s,!0))}return o}function Ye(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 Je(a,r,e,o,t,i,n,u,s,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(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?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=$e(j,M,X=>{let $=X*K,or=Math.min($+K,t);return[s,X,$,or]});z=Ye(a,O,"strsv-apply-params");let nr=$e(j,M,X=>{let $=X*K,or=Math.min($+K,t);return[t,s,n,y?0:1,w?0:1,$,or]});Y=Ye(a,nr,"strsv-update-params");let{commandEncoder:rr,querySet:ar}=pr();Q=N([{value:t,type:"u32"},{value:n,type:"u32"},{value:y?0:1,type:"u32"},{value:w?0:1,type:"u32"},{value:p?1:0,type:"u32"}],"strsv-invert-params");let at=A(h.getBindGroupLayout(0),[U,R,Q]);ir(rr,h,at,{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),nt=$/K,st=X===k.length-1,xr=nt*M,ut=A(G.getBindGroupLayout(0),[R,T,{buffer:z,offset:xr,size:16}]);ir(rr,G,ut,1,st&&ar?{timestampWrites:{querySet:ar,endOfPassWriteIndex:1}}:void 0);let vr=F?t-or:$;if(vr===0)continue;let lt=A(v.getBindGroupLayout(0),[U,T,{buffer:Y,offset:xr,size:32}]),ft=Math.min(vr,W);ir(rr,v,lt,ft)}let ot=cr(rr,ar),it=f?null:E(rr,T);S(rr);let sr=await P(ot);if(f)return sr!==void 0?{gpuTimeMs:sr}:{};let hr=await _(it,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 rt(a,r,e,o,t,i,n,u,s,l,f="row-major"){let c=s 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&&!(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<r||s.cols<e))throw new Error("A is too small for the given m and n.");(c?s.layout:f)==="column-major"&&([r,e]=[e,r],[t,n]=[n,t],[i,u]=[u,i]);let m=t instanceof x,g=n 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&&!(n 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&&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(r<0||e<0)throw new Error("m and n must be non-negative.");if(r===0||e===0)return c?{}:{A:s};if(!c&&s.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(n.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?n._buf:b(n,"sger-y",!1),G=c?s._buf:b(s,"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 et(a,r,e,o,t,i,n,u,s="row-major"){let l=t instanceof x,f=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(s!=="row-major"&&s!=="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&&!(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<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 f?{}:{A:n};if(!f&&n.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?n.layout:s)==="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?n._buf:b(n,"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 tt(a,r,e,o,t,i,n,u,s,l,f="row-major"){let c=t instanceof x,p=n instanceof x,m=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(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&&!(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&&!m)throw new Error("A must be a GpuMatrix when x and y are GpuVectors.");if(m&&c&&s._buf===t._buf)throw new Error("A and x must not reference the same GPU buffer.");if(m&&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(m&&l!==s.lda)throw new Error("lda must match A.lda when A is a GpuMatrix.");if(m&&(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 m?{}:{A:s};if(!m&&s.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(n.length<(e-1)*u+1)throw new Error("y does not have enough elements for the given n and incy.");let w=(m?s.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?n._buf:b(n,"ssyr2-y",!1),v=m?s._buf:b(s,"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 wt(ra);})();

package/index.d.mts

@@ -22,6 +22,9 @@ export { sgemv } from "./src/sgemv/sgemv.mjs";
 export { ssymv } from "./src/ssymv/ssymv.mjs";
 export { strmv } from "./src/strmv/strmv.mjs";
 export { strsv } from "./src/strsv/strsv.mjs";
+export { sger } from "./src/sger/sger.mjs";
+export { ssyr } from "./src/ssyr/ssyr.mjs";
+export { ssyr2 } from "./src/ssyr2/ssyr2.mjs";
 
 /**
  * Initializes the WebGPU device.

package/index.mjs

@@ -20,3 +20,6 @@ export { sgemv } from "./src/sgemv/sgemv.mjs";
 export { ssymv } from "./src/ssymv/ssymv.mjs";
 export { strmv } from "./src/strmv/strmv.mjs";
 export { strsv } from "./src/strsv/strsv.mjs";
+export { sger } from "./src/sger/sger.mjs";
+export { ssyr } from "./src/ssyr/ssyr.mjs";
+export { ssyr2 } from "./src/ssyr2/ssyr2.mjs";