umap-web 0.2.0

package/dist/webgpu.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"webgpu.js","names":["device","ctx: GpuContext","ctx","knnWgsl","dist: Float32Array","idx: Int32Array","normsBuf: GPUBuffer","qnormsBuf: GPUBuffer","cache","sgdWgsl","alpha","edgeBindGroups: GPUBindGroup[]","info: { epoch: number; nEpochs: number; phase: string; embedding?: Float32Array }","b","spmvWgsl","emb"],"sources":["../src/backends/webgpu/device.ts","../src/backends/webgpu/kernels/knn.wgsl","../src/backends/webgpu/knn.ts","../src/backends/webgpu/kernels/sgd.wgsl","../src/backends/webgpu/layout.ts","../src/backends/webgpu/kernels/spmv.wgsl","../src/backends/webgpu/spectral.ts","../src/backends/webgpu/index.ts"],"sourcesContent":["/**\n * WebGPU device acquisition (§5.4): injected device/provider or navigator.gpu,\n * elevated limits up to adapter limits. Device loss is tracked on the context\n * (kernels check it at loop boundaries and raise a typed UmapBackendError; the\n * estimator falls back to CPU under backend 'auto' — see DECISIONS D-017), and\n * every kernel runs inside validation + out-of-memory error scopes so GPU\n * failures surface as typed errors instead of silent zero-filled readbacks.\n */\n\nimport { UmapBackendError } from '../../types.js';\n\nexport interface GpuContext {\n  device: GPUDevice;\n  adapterInfo: { vendor: string; architecture: string; device: string; description: string };\n  limits: { maxStorageBufferBindingSize: number; maxBufferSize: number; maxWorkgroupsX: number };\n  timestamps: boolean;\n  injected: boolean;\n  /** set when the device reports loss; kernels must stop using the context */\n  lost: { reason: string; message: string } | null;\n}\n\nexport type DeviceProvider = GPUDevice | (() => Promise<GPUDevice>) | Promise<GPUDevice>;\n\nfunction trackLost(ctx: GpuContext): void {\n  ctx.device.lost\n    .then((info) => {\n      ctx.lost = { reason: String(info.reason), message: info.message };\n    })\n    .catch(() => {\n      ctx.lost = { reason: 'unknown', message: 'device.lost rejected' };\n    });\n}\n\n/** Throw (typed) if the device reported loss. */\nexport function assertUsable(ctx: GpuContext, stage: string): void {\n  if (ctx.lost) {\n    throw new UmapBackendError(\n      `GPU device lost during ${stage} (${ctx.lost.reason}): ${ctx.lost.message}`,\n    );\n  }\n}\n\n/**\n * Validate a planned buffer against the granted limits BEFORE creating it —\n * an oversized storage binding is an async validation error that would\n * otherwise surface as silently zero-filled results on low-limit adapters.\n */\nexport function ensureBufferFits(\n  ctx: GpuContext,\n  stage: string,\n  name: string,\n  byteLength: number,\n): void {\n  if (byteLength > ctx.limits.maxBufferSize) {\n    throw new UmapBackendError(\n      `${stage}: buffer '${name}' (${byteLength} bytes) exceeds this device's ` +\n        `maxBufferSize (${ctx.limits.maxBufferSize}); the dataset is too large for this GPU`,\n    );\n  }\n  if (byteLength > ctx.limits.maxStorageBufferBindingSize) {\n    throw new UmapBackendError(\n      `${stage}: buffer '${name}' (${byteLength} bytes) exceeds this device's ` +\n        `maxStorageBufferBindingSize (${ctx.limits.maxStorageBufferBindingSize}); ` +\n        'the dataset is too large for this GPU',\n    );\n  }\n}\n\n/**\n * Push validation + out-of-memory scopes; the returned closer pops them and\n * throws a typed error when either scope caught anything. The closer is\n * idempotent so kernels can also invoke it from a finally block (keeping the\n * device's scope stack balanced when the body throws first).\n */\nexport function beginErrorScopes(device: GPUDevice): (stage: string) => Promise<void> {\n  device.pushErrorScope('out-of-memory');\n  device.pushErrorScope('validation');\n  let popped = false;\n  return async (stage: string) => {\n    if (popped) return;\n    popped = true;\n    const validation = await device.popErrorScope();\n    const oom = await device.popErrorScope();\n    if (oom) {\n      throw new UmapBackendError(`${stage}: GPU out of memory: ${oom.message}`);\n    }\n    if (validation) {\n      throw new UmapBackendError(`${stage}: GPU validation error: ${validation.message}`);\n    }\n  };\n}\n\nexport async function acquireGpu(provided?: DeviceProvider): Promise<GpuContext> {\n  if (provided) {\n    const device =\n      typeof provided === 'function'\n        ? await provided()\n        : await Promise.resolve(provided as GPUDevice | Promise<GPUDevice>);\n    const ctx: GpuContext = {\n      device,\n      adapterInfo: {\n        vendor: 'injected',\n        architecture: 'injected',\n        device: '',\n        description: 'user-provided GPUDevice',\n      },\n      limits: {\n        maxStorageBufferBindingSize: device.limits.maxStorageBufferBindingSize,\n        maxBufferSize: device.limits.maxBufferSize,\n        maxWorkgroupsX: device.limits.maxComputeWorkgroupsPerDimension,\n      },\n      timestamps: device.features.has('timestamp-query'),\n      injected: true,\n      lost: null,\n    };\n    trackLost(ctx);\n    return ctx;\n  }\n  const gpu = (globalThis.navigator as Navigator & { gpu?: GPU })?.gpu;\n  if (!gpu) {\n    throw new UmapBackendError('WebGPU is not available (navigator.gpu undefined)');\n  }\n  const adapter = await gpu.requestAdapter();\n  if (!adapter) {\n    throw new UmapBackendError('WebGPU adapter unavailable (requestAdapter returned null)');\n  }\n  const wantTimestamps = adapter.features.has('timestamp-query');\n  const device = await adapter.requestDevice({\n    requiredLimits: {\n      maxStorageBufferBindingSize: adapter.limits.maxStorageBufferBindingSize,\n      maxBufferSize: adapter.limits.maxBufferSize,\n    },\n    requiredFeatures: wantTimestamps ? (['timestamp-query'] as GPUFeatureName[]) : [],\n  });\n  const info = adapter.info;\n  const ctx: GpuContext = {\n    device,\n    adapterInfo: {\n      vendor: info?.vendor ?? 'unknown',\n      architecture: info?.architecture ?? 'unknown',\n      device: info?.device ?? '',\n      description: info?.description ?? '',\n    },\n    limits: {\n      maxStorageBufferBindingSize: device.limits.maxStorageBufferBindingSize,\n      maxBufferSize: device.limits.maxBufferSize,\n      maxWorkgroupsX: device.limits.maxComputeWorkgroupsPerDimension,\n    },\n    timestamps: device.features.has('timestamp-query'),\n    injected: false,\n    lost: null,\n  };\n  trackLost(ctx);\n  return ctx;\n}\n\nexport function createBuffer(\n  device: GPUDevice,\n  data: Float32Array | Int32Array | Uint32Array,\n  usage: GPUBufferUsageFlags,\n): GPUBuffer {\n  const buf = device.createBuffer({ size: Math.ceil(data.byteLength / 4) * 4, usage });\n  device.queue.writeBuffer(buf, 0, data.buffer, data.byteOffset, data.byteLength);\n  return buf;\n}\n\nexport async function readBuffer(\n  device: GPUDevice,\n  src: GPUBuffer,\n  byteLength: number,\n): Promise<ArrayBuffer> {\n  const staging = device.createBuffer({\n    size: Math.ceil(byteLength / 4) * 4,\n    usage: GPUBufferUsage.COPY_DST | GPUBufferUsage.MAP_READ,\n  });\n  try {\n    const enc = device.createCommandEncoder();\n    enc.copyBufferToBuffer(src, 0, staging, 0, staging.size);\n    device.queue.submit([enc.finish()]);\n    await staging.mapAsync(GPUMapMode.READ);\n    const out = staging.getMappedRange().slice(0);\n    staging.unmap();\n    return out;\n  } catch (e) {\n    throw e instanceof UmapBackendError\n      ? e\n      : new UmapBackendError(`GPU readback failed: ${(e as Error).message}`);\n  } finally {\n    staging.destroy();\n  }\n}\n","\"// Tiled brute-force kNN (§A3): 256 queries per workgroup (one per thread),\\n// candidate tiles of 64 rows staged through workgroup memory as vec4 chunks,\\n// per-thread private top-k (k <= 32). No n×n buffer ever exists; fits the\\n// default 16 KB workgroup-storage and 8-storage-buffer limits.\\n//\\n// Data layout: rows padded to d4 = ceil(d/4) vec4s (zero padding is neutral for\\n// every supported metric). metricKind: 0 = squared euclidean, 1 = alternative\\n// cosine (log2(sqrt(nx*ny)/dot)), 2 = manhattan, 3 = hamming (raw count; /d on\\n// CPU), 4 = alternative jaccard. correlation = kind 1 on row-centered data.\\n\\nstruct Params {\\n  n: u32,          // candidate rows\\n  d4: u32,         // vec4s per row\\n  k: u32,          // neighbors (<= 32)\\n  metricKind: u32,\\n  nQueries: u32,\\n  candStart: u32,\\n  candEnd: u32,\\n  pad0: u32,\\n}\\n\\n@group(0) @binding(0) var<storage, read> data: array<vec4<f32>>;    // n x d4\\n@group(0) @binding(1) var<storage, read> queries: array<vec4<f32>>; // nQueries x d4\\n@group(0) @binding(2) var<storage, read_write> outDist: array<f32>;\\n@group(0) @binding(3) var<storage, read_write> outIdx: array<i32>;\\n@group(0) @binding(4) var<uniform> params: Params;\\n@group(0) @binding(5) var<storage, read> norms: array<f32>;\\n@group(0) @binding(6) var<storage, read> qnorms: array<f32>;\\n\\nconst TQ: u32 = 256u;  // queries per workgroup\\nconst TC: u32 = 64u;   // candidate tile rows\\nconst DC4: u32 = 8u;   // vec4s per dimension chunk (32 floats)\\nconst KMAX: u32 = 32u;\\nconst FLOAT32_MAX: f32 = 3.4028234663852886e38;\\n\\nvar<workgroup> cTile: array<vec4<f32>, 512>; // TC x DC4\\n\\n@compute @workgroup_size(256)\\nfn knnQueryTile(\\n  @builtin(workgroup_id) wg: vec3<u32>,\\n  @builtin(local_invocation_id) lid: vec3<u32>,\\n) {\\n  let qLocal = lid.x;\\n  let qRow = wg.x * TQ + qLocal;\\n  let isActive = qRow < params.nQueries;\\n\\n  var heapDist: array<f32, KMAX>;\\n  var heapIdx: array<i32, KMAX>;\\n  let k = params.k;\\n  for (var i = 0u; i < KMAX; i = i + 1u) {\\n    heapDist[i] = FLOAT32_MAX;\\n    heapIdx[i] = -1;\\n  }\\n  if (isActive && params.candStart > 0u) {\\n    for (var i = 0u; i < k; i = i + 1u) {\\n      heapDist[i] = outDist[qRow * k + i];\\n      heapIdx[i] = outIdx[qRow * k + i];\\n    }\\n  }\\n  var worst = 0u;\\n  var worstVal = -FLOAT32_MAX;\\n  for (var i = 0u; i < k; i = i + 1u) {\\n    if (heapDist[i] > worstVal) { worstVal = heapDist[i]; worst = i; }\\n  }\\n\\n  var acc: array<f32, TC>;\\n  var acc2: array<f32, TC>;\\n\\n  var cBase = params.candStart;\\n  loop {\\n    if (cBase >= params.candEnd) { break; }\\n    let tileLen = min(TC, params.candEnd - cBase);\\n\\n    for (var t = 0u; t < TC; t = t + 1u) { acc[t] = 0.0; acc2[t] = 0.0; }\\n\\n    var dBase = 0u; // vec4 offset within the row\\n    loop {\\n      if (dBase >= params.d4) { break; }\\n      let chunk = min(DC4, params.d4 - dBase);\\n      workgroupBarrier();\\n      // cooperative tile load: 512 vec4 slots over 256 threads (2 each)\\n      for (var li = qLocal; li < TC * DC4; li = li + 256u) {\\n        let row = li / DC4;\\n        let slot = li % DC4;\\n        if (row < tileLen && slot < chunk) {\\n          cTile[li] = data[(cBase + row) * params.d4 + dBase + slot];\\n        } else {\\n          cTile[li] = vec4<f32>(0.0);\\n        }\\n      }\\n      workgroupBarrier();\\n      if (isActive) {\\n        // fully-unrolled registers (dynamic private-array indexing spills to local\\n        // memory on Metal); zero vec4s are neutral for every metric here\\n        let qb = qRow * params.d4 + dBase;\\n        var q0 = vec4<f32>(0.0);\\n        var q1 = vec4<f32>(0.0);\\n        var q2 = vec4<f32>(0.0);\\n        var q3 = vec4<f32>(0.0);\\n        var q4 = vec4<f32>(0.0);\\n        var q5 = vec4<f32>(0.0);\\n        var q6 = vec4<f32>(0.0);\\n        var q7 = vec4<f32>(0.0);\\n        q0 = queries[qb];\\n        if (chunk > 1u) { q1 = queries[qb + 1u]; }\\n        if (chunk > 2u) { q2 = queries[qb + 2u]; }\\n        if (chunk > 3u) { q3 = queries[qb + 3u]; }\\n        if (chunk > 4u) { q4 = queries[qb + 4u]; }\\n        if (chunk > 5u) { q5 = queries[qb + 5u]; }\\n        if (chunk > 6u) { q6 = queries[qb + 6u]; }\\n        if (chunk > 7u) { q7 = queries[qb + 7u]; }\\n        switch params.metricKind {\\n          case 0u: {\\n            for (var t = 0u; t < tileLen; t = t + 1u) {\\n              let tb = t * DC4;\\n              let d0 = q0 - cTile[tb];\\n              let d1 = q1 - cTile[tb + 1u];\\n              let d2 = q2 - cTile[tb + 2u];\\n              let d3 = q3 - cTile[tb + 3u];\\n              let d4v = q4 - cTile[tb + 4u];\\n              let d5 = q5 - cTile[tb + 5u];\\n              let d6 = q6 - cTile[tb + 6u];\\n              let d7 = q7 - cTile[tb + 7u];\\n              let s = d0 * d0 + d1 * d1 + d2 * d2 + d3 * d3 + d4v * d4v + d5 * d5 + d6 * d6 + d7 * d7;\\n              acc[t] = acc[t] + s.x + s.y + s.z + s.w;\\n            }\\n          }\\n          case 1u: {\\n            for (var t = 0u; t < tileLen; t = t + 1u) {\\n              let tb = t * DC4;\\n              let s = q0 * cTile[tb] + q1 * cTile[tb + 1u] + q2 * cTile[tb + 2u] + q3 * cTile[tb + 3u]\\n                + q4 * cTile[tb + 4u] + q5 * cTile[tb + 5u] + q6 * cTile[tb + 6u] + q7 * cTile[tb + 7u];\\n              acc[t] = acc[t] + s.x + s.y + s.z + s.w;\\n            }\\n          }\\n          case 2u: {\\n            for (var t = 0u; t < tileLen; t = t + 1u) {\\n              let tb = t * DC4;\\n              let s = abs(q0 - cTile[tb]) + abs(q1 - cTile[tb + 1u]) + abs(q2 - cTile[tb + 2u]) + abs(q3 - cTile[tb + 3u])\\n                + abs(q4 - cTile[tb + 4u]) + abs(q5 - cTile[tb + 5u]) + abs(q6 - cTile[tb + 6u]) + abs(q7 - cTile[tb + 7u]);\\n              acc[t] = acc[t] + s.x + s.y + s.z + s.w;\\n            }\\n          }\\n          case 3u: {\\n            // hamming: padded zero lanes compare equal on both sides — neutral\\n            for (var t = 0u; t < tileLen; t = t + 1u) {\\n              let tb = t * DC4;\\n              let s = select(vec4<f32>(0.0), vec4<f32>(1.0), q0 != cTile[tb]) +\\n                select(vec4<f32>(0.0), vec4<f32>(1.0), q1 != cTile[tb + 1u]) +\\n                select(vec4<f32>(0.0), vec4<f32>(1.0), q2 != cTile[tb + 2u]) +\\n                select(vec4<f32>(0.0), vec4<f32>(1.0), q3 != cTile[tb + 3u]) +\\n                select(vec4<f32>(0.0), vec4<f32>(1.0), q4 != cTile[tb + 4u]) +\\n                select(vec4<f32>(0.0), vec4<f32>(1.0), q5 != cTile[tb + 5u]) +\\n                select(vec4<f32>(0.0), vec4<f32>(1.0), q6 != cTile[tb + 6u]) +\\n                select(vec4<f32>(0.0), vec4<f32>(1.0), q7 != cTile[tb + 7u]);\\n              acc[t] = acc[t] + s.x + s.y + s.z + s.w;\\n            }\\n          }\\n          default: {\\n            for (var t = 0u; t < tileLen; t = t + 1u) {\\n              let tb = t * DC4;\\n              var eq = vec4<f32>(0.0);\\n              var nz = vec4<f32>(0.0);\\n              let c0 = cTile[tb];\\n              let c1 = cTile[tb + 1u];\\n              let c2 = cTile[tb + 2u];\\n              let c3 = cTile[tb + 3u];\\n              eq = eq + select(vec4<f32>(0.0), vec4<f32>(1.0), (q0 != vec4<f32>(0.0)) & (c0 != vec4<f32>(0.0)));\\n              eq = eq + select(vec4<f32>(0.0), vec4<f32>(1.0), (q1 != vec4<f32>(0.0)) & (c1 != vec4<f32>(0.0)));\\n              eq = eq + select(vec4<f32>(0.0), vec4<f32>(1.0), (q2 != vec4<f32>(0.0)) & (c2 != vec4<f32>(0.0)));\\n              eq = eq + select(vec4<f32>(0.0), vec4<f32>(1.0), (q3 != vec4<f32>(0.0)) & (c3 != vec4<f32>(0.0)));\\n              let c4 = cTile[tb + 4u];\\n              let c5 = cTile[tb + 5u];\\n              let c6 = cTile[tb + 6u];\\n              let c7 = cTile[tb + 7u];\\n              eq = eq + select(vec4<f32>(0.0), vec4<f32>(1.0), (q4 != vec4<f32>(0.0)) & (c4 != vec4<f32>(0.0)));\\n              eq = eq + select(vec4<f32>(0.0), vec4<f32>(1.0), (q5 != vec4<f32>(0.0)) & (c5 != vec4<f32>(0.0)));\\n              eq = eq + select(vec4<f32>(0.0), vec4<f32>(1.0), (q6 != vec4<f32>(0.0)) & (c6 != vec4<f32>(0.0)));\\n              eq = eq + select(vec4<f32>(0.0), vec4<f32>(1.0), (q7 != vec4<f32>(0.0)) & (c7 != vec4<f32>(0.0)));\\n              nz = nz + select(vec4<f32>(0.0), vec4<f32>(1.0), (q0 != vec4<f32>(0.0)) | (c0 != vec4<f32>(0.0)));\\n              nz = nz + select(vec4<f32>(0.0), vec4<f32>(1.0), (q1 != vec4<f32>(0.0)) | (c1 != vec4<f32>(0.0)));\\n              nz = nz + select(vec4<f32>(0.0), vec4<f32>(1.0), (q2 != vec4<f32>(0.0)) | (c2 != vec4<f32>(0.0)));\\n              nz = nz + select(vec4<f32>(0.0), vec4<f32>(1.0), (q3 != vec4<f32>(0.0)) | (c3 != vec4<f32>(0.0)));\\n              nz = nz + select(vec4<f32>(0.0), vec4<f32>(1.0), (q4 != vec4<f32>(0.0)) | (c4 != vec4<f32>(0.0)));\\n              nz = nz + select(vec4<f32>(0.0), vec4<f32>(1.0), (q5 != vec4<f32>(0.0)) | (c5 != vec4<f32>(0.0)));\\n              nz = nz + select(vec4<f32>(0.0), vec4<f32>(1.0), (q6 != vec4<f32>(0.0)) | (c6 != vec4<f32>(0.0)));\\n              nz = nz + select(vec4<f32>(0.0), vec4<f32>(1.0), (q7 != vec4<f32>(0.0)) | (c7 != vec4<f32>(0.0)));\\n              acc[t] = acc[t] + eq.x + eq.y + eq.z + eq.w;\\n              acc2[t] = acc2[t] + nz.x + nz.y + nz.z + nz.w;\\n            }\\n          }\\n        }\\n      }\\n      dBase = dBase + DC4;\\n    }\\n\\n    if (isActive) {\\n      for (var t = 0u; t < tileLen; t = t + 1u) {\\n        var dist: f32;\\n        switch params.metricKind {\\n          case 0u, 2u, 3u: { dist = acc[t]; }\\n          case 1u: {\\n            let nx = qnorms[qRow];\\n            let ny = norms[cBase + t];\\n            if (nx == 0.0 && ny == 0.0) {\\n              dist = 0.0;\\n            } else if (nx == 0.0 || ny == 0.0 || acc[t] <= 0.0) {\\n              dist = FLOAT32_MAX;\\n            } else {\\n              dist = log2(sqrt(nx * ny) / acc[t]);\\n            }\\n          }\\n          default: {\\n            if (acc2[t] == 0.0) {\\n              dist = 0.0;\\n            } else if (acc[t] == 0.0) {\\n              dist = FLOAT32_MAX;\\n            } else {\\n              dist = -log2(acc[t] / acc2[t]);\\n            }\\n          }\\n        }\\n        if (dist < worstVal) {\\n          heapDist[worst] = dist;\\n          heapIdx[worst] = i32(cBase + t);\\n          worstVal = -FLOAT32_MAX;\\n          for (var i = 0u; i < k; i = i + 1u) {\\n            if (heapDist[i] > worstVal) { worstVal = heapDist[i]; worst = i; }\\n          }\\n        }\\n      }\\n    }\\n    cBase = cBase + TC;\\n  }\\n\\n  if (isActive) {\\n    for (var i = 0u; i < k; i = i + 1u) {\\n      outDist[qRow * k + i] = heapDist[i];\\n      outIdx[qRow * k + i] = heapIdx[i];\\n    }\\n  }\\n}\\n\\n// squared norms (cosine family); outDist reused as the norms output\\n@compute @workgroup_size(256)\\nfn rowNorms(@builtin(global_invocation_id) gid: vec3<u32>) {\\n  let row = gid.x;\\n  if (row >= params.n) { return; }\\n  var s = vec4<f32>(0.0);\\n  for (var c = 0u; c < params.d4; c = c + 1u) {\\n    let v = data[row * params.d4 + c];\\n    s = s + v * v;\\n  }\\n  outDist[row] = s.x + s.y + s.z + s.w;\\n}\\n\"","/**\n * GPU brute-force kNN driver (§A3): streams candidate ranges (watchdog-safe\n * dispatch sizes), per-row ascending sort + distance correction on readback.\n */\n\nimport type { KnnGraph, Matrix } from '../../types.js';\nimport { throwIfAborted, UmapBackendError } from '../../types.js';\nimport {\n  assertUsable,\n  beginErrorScopes,\n  createBuffer,\n  ensureBufferFits,\n  type GpuContext,\n  readBuffer,\n} from './device.js';\nimport knnWgsl from './kernels/knn.wgsl';\n\nexport type GpuMetricKind = 0 | 1 | 2 | 3 | 4; // sqeuclidean | altcosine | manhattan | hamming | altjaccard\n\nexport interface GpuMetricPlan {\n  kind: GpuMetricKind;\n  /** center rows before upload (correlation) */\n  center: boolean;\n  correction: (d: number) => number;\n}\n\n/** GPU support matrix (§4.3). Returns null for unsupported metrics. */\nexport function gpuMetricPlan(name: string): GpuMetricPlan | null {\n  switch (name) {\n    case 'euclidean':\n    case 'l2':\n      return { kind: 0, center: false, correction: (d) => Math.sqrt(d) };\n    case 'sqeuclidean':\n      return { kind: 0, center: false, correction: (d) => d };\n    case 'cosine':\n      return { kind: 1, center: false, correction: (d) => 1.0 - 2.0 ** -d };\n    case 'correlation':\n      return { kind: 1, center: true, correction: (d) => 1.0 - 2.0 ** -d };\n    case 'manhattan':\n    case 'taxicab':\n    case 'l1':\n      return { kind: 2, center: false, correction: (d) => d };\n    case 'hamming':\n      return { kind: 3, center: false, correction: (d) => d }; // ÷d applied by the driver\n    case 'jaccard':\n      return { kind: 4, center: false, correction: (d) => 1.0 - 2.0 ** -d };\n    default:\n      return null;\n  }\n}\n\nconst PARAMS_SIZE = 32;\n\ninterface KnnPipelines {\n  module: GPUShaderModule;\n  tile: GPUComputePipeline;\n  norms: GPUComputePipeline;\n}\n\nconst pipelineCache = new WeakMap<GPUDevice, KnnPipelines>();\n\nfunction getPipelines(device: GPUDevice): KnnPipelines {\n  let p = pipelineCache.get(device);\n  if (!p) {\n    const module = device.createShaderModule({ code: knnWgsl });\n    p = {\n      module,\n      tile: device.createComputePipeline({\n        layout: 'auto',\n        compute: { module, entryPoint: 'knnQueryTile' },\n      }),\n      norms: device.createComputePipeline({\n        layout: 'auto',\n        compute: { module, entryPoint: 'rowNorms' },\n      }),\n    };\n    pipelineCache.set(device, p);\n  }\n  return p;\n}\n\nexport interface GpuKnnOptions {\n  /** candidates per dispatch (watchdog safety); tuned by data size */\n  candChunk?: number;\n  onProgress?: (done: number, total: number) => void;\n  signal?: AbortSignal;\n}\n\n/**\n * kNN of `queries` against `data` (self-kNN when they alias) in INTERNAL metric\n * space; returns corrected, ascending-sorted results.\n */\nexport async function gpuKnn(\n  ctx: GpuContext,\n  X: Matrix,\n  queries: Matrix | null,\n  k: number,\n  plan: GpuMetricPlan,\n  { candChunk, onProgress, signal }: GpuKnnOptions = {},\n): Promise<KnnGraph> {\n  if (k > 32) throw new UmapBackendError('GPU kNN supports k <= 32');\n  const device = ctx.device;\n  const n = X.rows;\n  const d = X.cols;\n  const d4 = Math.ceil(d / 4);\n  const dPad = d4 * 4;\n  const nQueries = queries === null ? n : queries.rows;\n\n  // validate planned allocations against the granted limits up front (§A12):\n  // an oversized storage binding would otherwise fail asynchronously and read\n  // back as silent zeros on low-limit adapters\n  assertUsable(ctx, 'gpuKnn');\n  ensureBufferFits(ctx, 'gpuKnn', 'data', n * dPad * 4);\n  if (queries !== null) ensureBufferFits(ctx, 'gpuKnn', 'queries', nQueries * dPad * 4);\n  ensureBufferFits(ctx, 'gpuKnn', 'outDistances', nQueries * k * 4);\n  ensureBufferFits(ctx, 'gpuKnn', 'outIndices', nQueries * k * 4);\n\n  // pad rows to vec4 boundaries (zero padding is metric-neutral), center for correlation\n  const padRows = (src: Float32Array, rows: number, center: boolean): Float32Array => {\n    if (dPad === d && !center) return src; // already vec4-aligned — zero-copy\n    const out = new Float32Array(rows * dPad);\n    for (let i = 0; i < rows; i++) {\n      let mu = 0;\n      if (center) {\n        for (let c = 0; c < d; c++) mu += src[i * d + c]!;\n        mu /= d;\n      }\n      for (let c = 0; c < d; c++) out[i * dPad + c] = src[i * d + c]! - mu;\n    }\n    return out;\n  };\n  const hostData = padRows(X.data, n, plan.center);\n  const sameQuery = queries === null;\n  const nq = sameQuery ? n : queries!.rows;\n  const hostQ = sameQuery ? hostData : padRows(queries!.data, nq, plan.center);\n\n  const pipes = getPipelines(device);\n  // every buffer is tracked and destroyed in the finally below — abort or a\n  // validation failure must not leak GPU memory (injected devices accumulate)\n  const scratch = new Set<GPUBuffer>();\n  const track = (buf: GPUBuffer): GPUBuffer => {\n    scratch.add(buf);\n    return buf;\n  };\n  const endScopes = beginErrorScopes(device);\n  let dist: Float32Array;\n  let idx: Int32Array;\n  try {\n    const dataBuf = track(\n      createBuffer(device, hostData, GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST),\n    );\n    const qBuf = sameQuery\n      ? dataBuf\n      : track(createBuffer(device, hostQ, GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST));\n    const outDist = track(\n      device.createBuffer({\n        size: nq * k * 4,\n        usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC,\n      }),\n    );\n    const outIdx = track(\n      device.createBuffer({\n        size: nq * k * 4,\n        usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC,\n      }),\n    );\n    const paramsBuf = track(\n      device.createBuffer({\n        size: PARAMS_SIZE,\n        usage: GPUBufferUsage.UNIFORM | GPUBufferUsage.COPY_DST,\n      }),\n    );\n    const dummy = track(device.createBuffer({ size: 4, usage: GPUBufferUsage.STORAGE }));\n\n    // norms (cosine family)\n    let normsBuf: GPUBuffer = dummy;\n    let qnormsBuf: GPUBuffer = dummy;\n    if (plan.kind === 1) {\n      normsBuf = track(\n        device.createBuffer({\n          size: n * 4,\n          usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC,\n        }),\n      );\n      const np = track(\n        device.createBuffer({\n          size: PARAMS_SIZE,\n          usage: GPUBufferUsage.UNIFORM | GPUBufferUsage.COPY_DST,\n        }),\n      );\n      device.queue.writeBuffer(np, 0, new Uint32Array([n, d4, k, plan.kind, n, 0, n, 0]));\n      const bg = device.createBindGroup({\n        layout: pipes.norms.getBindGroupLayout(0),\n        entries: [\n          { binding: 0, resource: { buffer: dataBuf } },\n          { binding: 2, resource: { buffer: normsBuf } },\n          { binding: 4, resource: { buffer: np } },\n        ],\n      });\n      const enc = device.createCommandEncoder();\n      const pass = enc.beginComputePass();\n      pass.setPipeline(pipes.norms);\n      pass.setBindGroup(0, bg);\n      pass.dispatchWorkgroups(Math.ceil(n / 256));\n      pass.end();\n      device.queue.submit([enc.finish()]);\n      if (sameQuery) {\n        qnormsBuf = normsBuf;\n      } else {\n        qnormsBuf = track(\n          device.createBuffer({\n            size: nq * 4,\n            usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC,\n          }),\n        );\n        const qp = track(\n          device.createBuffer({\n            size: PARAMS_SIZE,\n            usage: GPUBufferUsage.UNIFORM | GPUBufferUsage.COPY_DST,\n          }),\n        );\n        device.queue.writeBuffer(qp, 0, new Uint32Array([nq, d4, k, plan.kind, nq, 0, nq, 0]));\n        const bgq = device.createBindGroup({\n          layout: pipes.norms.getBindGroupLayout(0),\n          entries: [\n            { binding: 0, resource: { buffer: qBuf } },\n            { binding: 2, resource: { buffer: qnormsBuf } },\n            { binding: 4, resource: { buffer: qp } },\n          ],\n        });\n        const enc2 = device.createCommandEncoder();\n        const pass2 = enc2.beginComputePass();\n        pass2.setPipeline(pipes.norms);\n        pass2.setBindGroup(0, bgq);\n        pass2.dispatchWorkgroups(Math.ceil(nq / 256));\n        pass2.end();\n        device.queue.submit([enc2.finish()]);\n      }\n    }\n\n    const bindGroup = device.createBindGroup({\n      layout: pipes.tile.getBindGroupLayout(0),\n      entries: [\n        { binding: 0, resource: { buffer: dataBuf } },\n        { binding: 1, resource: { buffer: qBuf } },\n        { binding: 2, resource: { buffer: outDist } },\n        { binding: 3, resource: { buffer: outIdx } },\n        { binding: 4, resource: { buffer: paramsBuf } },\n        { binding: 5, resource: { buffer: normsBuf } },\n        { binding: 6, resource: { buffer: qnormsBuf } },\n      ],\n    });\n\n    // watchdog-safe candidate streaming: budget ≈ 6e10 madd per dispatch\n    const chunk = candChunk ?? Math.max(4096, Math.min(n, Math.floor(6e10 / Math.max(1, nq * d))));\n    const wgCount = Math.ceil(nq / 256);\n    for (let cs = 0; cs < n; cs += chunk) {\n      throwIfAborted(signal);\n      assertUsable(ctx, 'gpuKnn');\n      const ce = Math.min(n, cs + chunk);\n      device.queue.writeBuffer(paramsBuf, 0, new Uint32Array([n, d4, k, plan.kind, nq, cs, ce, 0]));\n      const enc = device.createCommandEncoder();\n      const pass = enc.beginComputePass();\n      pass.setPipeline(pipes.tile);\n      pass.setBindGroup(0, bindGroup);\n      pass.dispatchWorkgroups(wgCount);\n      pass.end();\n      device.queue.submit([enc.finish()]);\n      await device.queue.onSubmittedWorkDone();\n      onProgress?.(ce, n);\n    }\n    await endScopes('gpuKnn');\n    assertUsable(ctx, 'gpuKnn');\n\n    dist = new Float32Array(await readBuffer(device, outDist, nq * k * 4));\n    idx = new Int32Array(await readBuffer(device, outIdx, nq * k * 4));\n  } finally {\n    // balance the scope stack even when the body threw first\n    await endScopes('gpuKnn').catch(() => {});\n    for (const b of scratch) b.destroy();\n  }\n\n  // per-row ascending sort + correction (final output space)\n  const indices = new Int32Array(nq * k);\n  const distances = new Float32Array(nq * k);\n  const order = new Array<number>(k);\n  for (let i = 0; i < nq; i++) {\n    for (let j = 0; j < k; j++) order[j] = j;\n    order.sort((a, b) => dist[i * k + a]! - dist[i * k + b]!);\n    for (let j = 0; j < k; j++) {\n      const o = order[j]!;\n      let raw = dist[i * k + o]!;\n      if (plan.kind === 3) raw /= d; // hamming: kernel returns the mismatch count\n      indices[i * k + j] = idx[i * k + o]!;\n      distances[i * k + j] = raw >= 3.4e38 ? Infinity : plan.correction(raw);\n    }\n  }\n  return { indices, distances, k };\n}\n","\"// SGD layout kernels (§A2). Two strategies, both implemented and measured\\n// (DECISIONS D-011):\\n//   (A) vertexGather — one thread per head vertex walking its CSR edge range;\\n//       writes only its own row (conflict-free by ownership; Hogwild reads).\\n//   (B) edgeAtomic — one thread per edge; fixed-point atomicAdd on i32 embedding\\n//       (scale 2^20), converted to/from f32 by pack/unpack kernels. Epochs run\\n//       as ordered sub-batches; reads come from the previous sub-batch's\\n//       snapshot (binding 11), making the result bit-deterministic per seed.\\n\\nstruct LayoutParams {\\n  nVertices: u32,\\n  nEdges: u32,\\n  dim: u32,\\n  epoch: u32,\\n  alpha: f32,\\n  a: f32,\\n  b: f32,\\n  gamma: f32,\\n  negativeSampleRate: f32,\\n  moveOther: u32,\\n  scale: f32,      // fixed-point scale for edgeAtomic\\n  pad0: u32,\\n}\\n\\n// Ordered sub-batch of the edge list for one edgeAtomic dispatch. Epochs are\\n// split into sequential sub-batches so an edge sees the committed updates of\\n// every earlier sub-batch — approximating the CPU's in-order (Gauss–Seidel)\\n// update visibility instead of one epoch-wide racy dispatch.\\nstruct BatchParams {\\n  edgeOffset: u32,\\n  edgeCount: u32,\\n}\\n\\n@group(0) @binding(0) var<storage, read_write> embedding: array<f32>;\\n@group(0) @binding(1) var<storage, read> head: array<i32>;\\n@group(0) @binding(2) var<storage, read> tail: array<i32>;\\n@group(0) @binding(3) var<storage, read> epochsPerSample: array<f32>;\\n// schedule[e] = epoch_of_next_sample; schedule[nEdges + e] = epoch_of_next_negative\\n@group(0) @binding(4) var<storage, read_write> schedule: array<f32>;\\n@group(0) @binding(6) var<storage, read_write> rngStates: array<u32>; // 3 per vertex (A) / per edge (B)\\n@group(0) @binding(7) var<uniform> P: LayoutParams;\\n@group(0) @binding(8) var<storage, read> rowPtr: array<u32>;          // vertexGather CSR\\n@group(0) @binding(9) var<storage, read_write> embeddingFx: array<atomic<i32>>; // edgeAtomic\\n@group(0) @binding(10) var<uniform> B: BatchParams;\\n// Fixed-point coordinate snapshot: the committed state after the previous\\n// sub-batch. edgeAtomic reads ONLY from here and atomically accumulates into\\n// embeddingFx, so every thread sees identical inputs regardless of scheduling\\n// and i32 adds commute exactly — the layout is bit-deterministic per seed.\\n@group(0) @binding(11) var<storage, read> snapFx: array<i32>;\\n\\nconst CLIP: f32 = 4.0;\\n\\nfn tau_step(s: ptr<function, vec3<u32>>) -> i32 {\\n  var s0 = (*s).x;\\n  var s1 = (*s).y;\\n  var s2 = (*s).z;\\n  s0 = ((s0 & 0xfffffffeu) << 12u) ^ (((s0 << 13u) ^ s0) >> 19u);\\n  s1 = ((s1 & 0xfffffff8u) << 4u) ^ (((s1 << 2u) ^ s1) >> 25u);\\n  s2 = ((s2 & 0xfffffff0u) << 17u) ^ (((s2 << 3u) ^ s2) >> 11u);\\n  (*s) = vec3<u32>(s0, s1, s2);\\n  return i32(s0 ^ s1 ^ s2);\\n}\\n\\nfn clipv(v: f32) -> f32 {\\n  return clamp(v, -CLIP, CLIP);\\n}\\n\\n// ---------------------------------------------------------------------------\\n// (A) vertexGather: thread = head vertex; edges rowPtr[v]..rowPtr[v+1]\\n// ---------------------------------------------------------------------------\\n\\n@compute @workgroup_size(256)\\nfn vertexGather(@builtin(global_invocation_id) gid: vec3<u32>) {\\n  let v = gid.x;\\n  if (v >= P.nVertices) { return; }\\n  let dim = P.dim;\\n  let n = f32(P.epoch);\\n  var rng = vec3<u32>(rngStates[v * 3u], rngStates[v * 3u + 1u], rngStates[v * 3u + 2u]);\\n\\n  for (var e = rowPtr[v]; e < rowPtr[v + 1u]; e = e + 1u) {\\n    if (schedule[e] > n) { continue; }\\n    let j = u32(head[e]);\\n    let kv = u32(tail[e]);\\n    var d2 = 0.0;\\n    for (var c = 0u; c < dim; c = c + 1u) {\\n      let diff = embedding[j * dim + c] - embedding[kv * dim + c];\\n      d2 = d2 + diff * diff;\\n    }\\n    var gradCoeff = 0.0;\\n    if (d2 > 0.0) {\\n      gradCoeff = (-2.0 * P.a * P.b * pow(d2, P.b - 1.0)) / (P.a * pow(d2, P.b) + 1.0);\\n    }\\n    for (var c = 0u; c < dim; c = c + 1u) {\\n      let diff = embedding[j * dim + c] - embedding[kv * dim + c];\\n      let g = clipv(gradCoeff * diff);\\n      // ownership: each endpoint gets one pull per undirected edge (the reverse\\n      // directed edge pulls the other endpoint)\\n      embedding[j * dim + c] = embedding[j * dim + c] + g * P.alpha;\\n    }\\n    schedule[e] = schedule[e] + epochsPerSample[e];\\n\\n    let epn = epochsPerSample[e] / P.negativeSampleRate;\\n    let nNeg = i32((n - schedule[P.nEdges + e]) / epn);\\n    for (var p = 0; p < nNeg; p = p + 1) {\\n      let r = tau_step(&rng);\\n      var kk = r % i32(P.nVertices);\\n      if (kk < 0) { kk = kk + i32(P.nVertices); }\\n      if (u32(kk) == j) { continue; }\\n      var dn = 0.0;\\n      for (var c = 0u; c < dim; c = c + 1u) {\\n        let diff = embedding[j * dim + c] - embedding[u32(kk) * dim + c];\\n        dn = dn + diff * diff;\\n      }\\n      if (dn > 0.0) {\\n        let rep = (2.0 * P.gamma * P.b) / ((0.001 + dn) * (P.a * pow(dn, P.b) + 1.0));\\n        for (var c = 0u; c < dim; c = c + 1u) {\\n          let diff = embedding[j * dim + c] - embedding[u32(kk) * dim + c];\\n          embedding[j * dim + c] = embedding[j * dim + c] + clipv(rep * diff) * P.alpha;\\n        }\\n      }\\n    }\\n    schedule[P.nEdges + e] = schedule[P.nEdges + e] + f32(nNeg) * epn;\\n  }\\n  rngStates[v * 3u] = rng.x;\\n  rngStates[v * 3u + 1u] = rng.y;\\n  rngStates[v * 3u + 2u] = rng.z;\\n}\\n\\n// ---------------------------------------------------------------------------\\n// (B) edgeAtomic: thread = directed edge; fixed-point atomics\\n// ---------------------------------------------------------------------------\\n\\n@compute @workgroup_size(256)\\nfn packFx(@builtin(global_invocation_id) gid: vec3<u32>) {\\n  let i = gid.x;\\n  if (i >= P.nVertices * P.dim) { return; }\\n  atomicStore(&embeddingFx[i], i32(round(embedding[i] * P.scale)));\\n}\\n\\n@compute @workgroup_size(256)\\nfn unpackFx(@builtin(global_invocation_id) gid: vec3<u32>) {\\n  let i = gid.x;\\n  if (i >= P.nVertices * P.dim) { return; }\\n  embedding[i] = f32(atomicLoad(&embeddingFx[i])) / P.scale;\\n}\\n\\n@compute @workgroup_size(256)\\nfn edgeAtomic(@builtin(global_invocation_id) gid: vec3<u32>) {\\n  if (gid.x >= B.edgeCount) { return; }\\n  let e = B.edgeOffset + gid.x;\\n  if (e >= P.nEdges) { return; }\\n  if (schedule[e] > f32(P.epoch)) { return; }\\n  let dim = P.dim;\\n  let j = u32(head[e]);\\n  let kv = u32(tail[e]);\\n  var rng = vec3<u32>(rngStates[e * 3u], rngStates[e * 3u + 1u], rngStates[e * 3u + 2u]);\\n\\n  // Thread-local running coordinates for j (the CPU kernel's within-edge\\n  // update chain); every cross-vertex read comes from the sub-batch snapshot.\\n  var cj: array<f32, 8>;\\n  var net: array<f32, 8>;\\n  for (var c = 0u; c < dim; c = c + 1u) {\\n    cj[c] = f32(snapFx[j * dim + c]) / P.scale;\\n    net[c] = 0.0;\\n  }\\n\\n  var d2 = 0.0;\\n  for (var c = 0u; c < dim; c = c + 1u) {\\n    let diff = cj[c] - f32(snapFx[kv * dim + c]) / P.scale;\\n    d2 = d2 + diff * diff;\\n  }\\n  if (d2 > 0.0) {\\n    let gradCoeff = (-2.0 * P.a * P.b * pow(d2, P.b - 1.0)) / (P.a * pow(d2, P.b) + 1.0);\\n    for (var c = 0u; c < dim; c = c + 1u) {\\n      let diff = cj[c] - f32(snapFx[kv * dim + c]) / P.scale;\\n      let g = clipv(gradCoeff * diff) * P.alpha;\\n      cj[c] = cj[c] + g;\\n      net[c] = net[c] + g;\\n      if (P.moveOther == 1u) {\\n        atomicAdd(&embeddingFx[kv * dim + c], i32(round(-g * P.scale)));\\n      }\\n    }\\n  }\\n  schedule[e] = schedule[e] + epochsPerSample[e];\\n\\n  let epn = epochsPerSample[e] / P.negativeSampleRate;\\n  let nNeg = i32((f32(P.epoch) - schedule[P.nEdges + e]) / epn);\\n  for (var p = 0; p < nNeg; p = p + 1) {\\n    let r = tau_step(&rng);\\n    var kk = r % i32(P.nVertices);\\n    if (kk < 0) { kk = kk + i32(P.nVertices); }\\n    if (u32(kk) == j) { continue; }\\n    var dn = 0.0;\\n    for (var c = 0u; c < dim; c = c + 1u) {\\n      let diff = cj[c] - f32(snapFx[u32(kk) * dim + c]) / P.scale;\\n      dn = dn + diff * diff;\\n    }\\n    if (dn > 0.0) {\\n      let rep = (2.0 * P.gamma * P.b) / ((0.001 + dn) * (P.a * pow(dn, P.b) + 1.0));\\n      for (var c = 0u; c < dim; c = c + 1u) {\\n        let diff = cj[c] - f32(snapFx[u32(kk) * dim + c]) / P.scale;\\n        let g = clipv(rep * diff) * P.alpha;\\n        cj[c] = cj[c] + g;\\n        net[c] = net[c] + g;\\n      }\\n    }\\n  }\\n  schedule[P.nEdges + e] = schedule[P.nEdges + e] + f32(nNeg) * epn;\\n  // one rounded net update per component — less quantization than per-step adds\\n  for (var c = 0u; c < dim; c = c + 1u) {\\n    atomicAdd(&embeddingFx[j * dim + c], i32(round(net[c] * P.scale)));\\n  }\\n  rngStates[e * 3u] = rng.x;\\n  rngStates[e * 3u + 1u] = rng.y;\\n  rngStates[e * 3u + 2u] = rng.z;\\n}\\n\"","/**\n * GPU SGD layout driver (§A2). Two strategies (DECISIONS D-011):\n * 'vertexGather' (conflict-free by ownership) and 'edgeAtomic' (fixed-point\n * atomicAdd, scale 2^20). edgeAtomic epochs run as ordered sub-batches that\n * read a coordinate snapshot and atomically accumulate — bit-deterministic\n * per seed (D-013); alpha via uniform.\n */\n\nimport type { Pcg32 } from '../../rng.js';\nimport { throwIfAborted, UmapBackendError } from '../../types.js';\nimport {\n  assertUsable,\n  beginErrorScopes,\n  createBuffer,\n  ensureBufferFits,\n  type GpuContext,\n  readBuffer,\n} from './device.js';\nimport sgdWgsl from './kernels/sgd.wgsl';\n\nexport type GpuLayoutStrategy = 'vertexGather' | 'edgeAtomic';\n\ninterface SgdPipelines {\n  vertexGather: GPUComputePipeline;\n  edgeAtomic: GPUComputePipeline;\n  packFx: GPUComputePipeline;\n  unpackFx: GPUComputePipeline;\n}\n\nconst cache = new WeakMap<GPUDevice, SgdPipelines>();\n\ninterface SgdPipelineSet extends SgdPipelines {\n  bindLayout: GPUBindGroupLayout;\n  bindLayoutEdge: GPUBindGroupLayout;\n}\n\nfunction pipelines(device: GPUDevice): SgdPipelineSet {\n  let p = cache.get(device) as SgdPipelineSet | undefined;\n  if (!p) {\n    const module = device.createShaderModule({ code: sgdWgsl });\n    const storage = (binding: number, type: GPUBufferBindingType): GPUBindGroupLayoutEntry => ({\n      binding,\n      visibility: GPUShaderStage.COMPUTE,\n      buffer: { type },\n    });\n    // shared by vertexGather/packFx/unpackFx (single bind group works for all)\n    const bindLayout = device.createBindGroupLayout({\n      entries: [\n        storage(0, 'storage'),\n        storage(1, 'read-only-storage'),\n        storage(2, 'read-only-storage'),\n        storage(3, 'read-only-storage'),\n        storage(4, 'storage'),\n        storage(6, 'storage'),\n        storage(7, 'uniform'),\n        storage(8, 'read-only-storage'),\n        storage(9, 'storage'),\n      ],\n    });\n    // edgeAtomic needs the batch uniform + coordinate snapshot; dropping the\n    // f32 embedding and rowPtr keeps it at 7 storage buffers (default limit 8)\n    const bindLayoutEdge = device.createBindGroupLayout({\n      entries: [\n        storage(1, 'read-only-storage'),\n        storage(2, 'read-only-storage'),\n        storage(3, 'read-only-storage'),\n        storage(4, 'storage'),\n        storage(6, 'storage'),\n        storage(7, 'uniform'),\n        storage(9, 'storage'),\n        storage(10, 'uniform'),\n        storage(11, 'read-only-storage'),\n      ],\n    });\n    const layout = device.createPipelineLayout({ bindGroupLayouts: [bindLayout] });\n    const layoutEdge = device.createPipelineLayout({ bindGroupLayouts: [bindLayoutEdge] });\n    const mk = (entryPoint: string, pl: GPUPipelineLayout = layout) =>\n      device.createComputePipeline({ layout: pl, compute: { module, entryPoint } });\n    p = {\n      vertexGather: mk('vertexGather'),\n      edgeAtomic: mk('edgeAtomic', layoutEdge),\n      packFx: mk('packFx'),\n      unpackFx: mk('unpackFx'),\n      bindLayout,\n      bindLayoutEdge,\n    };\n    cache.set(device, p);\n  }\n  return p;\n}\n\nexport interface GpuLayoutOptions {\n  gamma: number;\n  initialAlpha: number;\n  negativeSampleRate: number;\n  moveOther: boolean;\n  strategy?: GpuLayoutStrategy;\n  onEpoch?: (info: {\n    epoch: number;\n    nEpochs: number;\n    phase: string;\n    embedding?: Float32Array;\n  }) => void;\n  snapshotEvery?: number;\n  signal?: AbortSignal;\n}\n\nconst UNIFORM_SIZE = 48;\n\nexport async function gpuOptimizeLayout(\n  ctx: GpuContext,\n  embedding: Float32Array, // mutated with the final result\n  dim: number,\n  head: Int32Array,\n  tail: Int32Array,\n  nEpochs: number,\n  nVertices: number,\n  epochsPerSample: Float64Array,\n  a: number,\n  b: number,\n  rng: Pcg32,\n  opts: GpuLayoutOptions,\n): Promise<Float32Array> {\n  const device = ctx.device;\n  const strategy = opts.strategy ?? 'edgeAtomic'; // D-011: measured winner\n  const nEdges = head.length;\n  if (strategy === 'edgeAtomic' && dim > 8) {\n    throw new UmapBackendError(`webgpu layout supports dim <= 8 (got ${dim})`);\n  }\n  const pipes = pipelines(device);\n\n  // validate planned allocations against the granted limits up front (§A12)\n  assertUsable(ctx, 'gpuOptimizeLayout');\n  const fxSizePlanned = strategy === 'edgeAtomic' ? nVertices * dim * 4 : 4;\n  ensureBufferFits(ctx, 'gpuOptimizeLayout', 'embedding', embedding.byteLength);\n  ensureBufferFits(ctx, 'gpuOptimizeLayout', 'edges', nEdges * 4);\n  ensureBufferFits(ctx, 'gpuOptimizeLayout', 'schedule', nEdges * 2 * 4);\n  ensureBufferFits(\n    ctx,\n    'gpuOptimizeLayout',\n    'rngStates',\n    (strategy === 'vertexGather' ? nVertices : nEdges) * 3 * 4,\n  );\n  ensureBufferFits(ctx, 'gpuOptimizeLayout', 'fixedPoint', fxSizePlanned);\n\n  const epsF32 = Float32Array.from(epochsPerSample);\n  const scheduleInit = new Float32Array(nEdges * 2);\n  scheduleInit.set(epsF32, 0);\n  for (let i = 0; i < nEdges; i++) {\n    scheduleInit[nEdges + i] = epsF32[i]! / opts.negativeSampleRate;\n  }\n\n  // every buffer is tracked and destroyed in the finally below (abort-safe)\n  const scratch = new Set<GPUBuffer>();\n  const track = (buf: GPUBuffer): GPUBuffer => {\n    scratch.add(buf);\n    return buf;\n  };\n  const endScopes = beginErrorScopes(device);\n  try {\n    const embBuf = track(\n      createBuffer(\n        device,\n        embedding,\n        GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST | GPUBufferUsage.COPY_SRC,\n      ),\n    );\n    const headBuf = track(\n      createBuffer(device, head, GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST),\n    );\n    const tailBuf = track(\n      createBuffer(device, tail, GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST),\n    );\n    const epsBuf = track(\n      createBuffer(device, epsF32, GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST),\n    );\n    const scheduleBuf = track(\n      createBuffer(device, scheduleInit, GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST),\n    );\n\n    // rng streams: per vertex (gather) or per edge (atomic)\n    const nStreams = strategy === 'vertexGather' ? nVertices : nEdges;\n    const states = new Uint32Array(nStreams * 3);\n    for (let i = 0; i < states.length; i++) states[i] = rng.nextU32();\n    const rngBuf = track(\n      createBuffer(device, states, GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST),\n    );\n\n    // vertexGather CSR rowPtr over head-sorted edges\n    const rowPtr = new Uint32Array(nVertices + 1);\n    for (let e = 0; e < nEdges; e++) rowPtr[head[e]! + 1]!++;\n    for (let v = 0; v < nVertices; v++) rowPtr[v + 1]! += rowPtr[v]!;\n    const rowPtrBuf = track(\n      createBuffer(device, rowPtr, GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST),\n    );\n\n    const scale = 1048576.0; // 2^20 fixed-point: |coord| < 2048, resolution ~1e-6 (D-011)\n    const fxSize = fxSizePlanned;\n    const fxBuf = track(\n      device.createBuffer({\n        size: fxSize,\n        usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC,\n      }),\n    );\n    // committed-state snapshot read by each edgeAtomic sub-batch (D-013)\n    const snapBuf = track(\n      device.createBuffer({\n        size: fxSize,\n        usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST,\n      }),\n    );\n\n    const uni = track(\n      device.createBuffer({\n        size: UNIFORM_SIZE,\n        usage: GPUBufferUsage.UNIFORM | GPUBufferUsage.COPY_DST,\n      }),\n    );\n    const writeUniform = (epoch: number, alpha: number) => {\n      const buf = new ArrayBuffer(UNIFORM_SIZE);\n      const u = new Uint32Array(buf);\n      const f = new Float32Array(buf);\n      u[0] = nVertices;\n      u[1] = nEdges;\n      u[2] = dim;\n      u[3] = epoch;\n      f[4] = alpha;\n      f[5] = a;\n      f[6] = b;\n      f[7] = opts.gamma;\n      f[8] = opts.negativeSampleRate;\n      u[9] = opts.moveOther ? 1 : 0;\n      f[10] = scale;\n      device.queue.writeBuffer(uni, 0, buf);\n    };\n\n    // Sequential sub-batches per epoch (edgeAtomic): edges of batch s only start\n    // after batch s-1's atomics are committed, approximating the CPU's in-order\n    // edge visibility. Slot stride obeys minUniformBufferOffsetAlignment.\n    const SUBBATCHES = strategy === 'edgeAtomic' ? Math.min(32, Math.max(1, nEdges)) : 1;\n    const chunk = Math.ceil(nEdges / SUBBATCHES);\n    const slotStride = Math.max(256, device.limits.minUniformBufferOffsetAlignment);\n    const batchUni = track(\n      device.createBuffer({\n        size: slotStride * SUBBATCHES,\n        usage: GPUBufferUsage.UNIFORM | GPUBufferUsage.COPY_DST,\n      }),\n    );\n    for (let s = 0; s < SUBBATCHES; s++) {\n      const off = s * chunk;\n      const cnt = Math.max(0, Math.min(chunk, nEdges - off));\n      device.queue.writeBuffer(batchUni, s * slotStride, new Uint32Array([off, cnt]));\n    }\n\n    const bindGroup = device.createBindGroup({\n      layout: pipes.bindLayout,\n      entries: [\n        { binding: 0, resource: { buffer: embBuf } },\n        { binding: 1, resource: { buffer: headBuf } },\n        { binding: 2, resource: { buffer: tailBuf } },\n        { binding: 3, resource: { buffer: epsBuf } },\n        { binding: 4, resource: { buffer: scheduleBuf } },\n        { binding: 6, resource: { buffer: rngBuf } },\n        { binding: 7, resource: { buffer: uni } },\n        { binding: 8, resource: { buffer: rowPtrBuf } },\n        { binding: 9, resource: { buffer: fxBuf } },\n      ],\n    });\n    const edgeBindGroups: GPUBindGroup[] = [];\n    for (let s = 0; s < SUBBATCHES; s++) {\n      edgeBindGroups.push(\n        device.createBindGroup({\n          layout: pipes.bindLayoutEdge,\n          entries: [\n            { binding: 1, resource: { buffer: headBuf } },\n            { binding: 2, resource: { buffer: tailBuf } },\n            { binding: 3, resource: { buffer: epsBuf } },\n            { binding: 4, resource: { buffer: scheduleBuf } },\n            { binding: 6, resource: { buffer: rngBuf } },\n            { binding: 7, resource: { buffer: uni } },\n            { binding: 9, resource: { buffer: fxBuf } },\n            { binding: 10, resource: { buffer: batchUni, offset: s * slotStride, size: 8 } },\n            { binding: 11, resource: { buffer: snapBuf } },\n          ],\n        }),\n      );\n    }\n\n    const run = (pipeline: GPUComputePipeline, bg: GPUBindGroup, threads: number) => {\n      const enc = device.createCommandEncoder();\n      const pass = enc.beginComputePass();\n      pass.setPipeline(pipeline);\n      pass.setBindGroup(0, bg);\n      pass.dispatchWorkgroups(Math.ceil(threads / 256));\n      pass.end();\n      device.queue.submit([enc.finish()]);\n    };\n\n    if (strategy === 'edgeAtomic') {\n      writeUniform(0, opts.initialAlpha);\n      const enc = device.createCommandEncoder();\n      const pass = enc.beginComputePass();\n      pass.setPipeline(pipes.packFx);\n      pass.setBindGroup(0, bindGroup);\n      pass.dispatchWorkgroups(Math.ceil((nVertices * dim) / 256));\n      pass.end();\n      enc.copyBufferToBuffer(fxBuf, 0, snapBuf, 0, fxSize);\n      device.queue.submit([enc.finish()]);\n    }\n\n    let alpha = opts.initialAlpha;\n    const SYNC_EVERY = 24; // bound queued work; keeps the device responsive\n    for (let epoch = 0; epoch < nEpochs; epoch++) {\n      throwIfAborted(opts.signal);\n      assertUsable(ctx, 'gpuOptimizeLayout');\n      writeUniform(epoch, alpha);\n      if (strategy === 'vertexGather') {\n        run(pipes.vertexGather, bindGroup, nVertices);\n      } else {\n        // one submit, SUBBATCHES ordered passes; after each pass the committed\n        // coordinates are copied into the snapshot the next pass reads (D-013)\n        const enc = device.createCommandEncoder();\n        for (let s = 0; s < SUBBATCHES; s++) {\n          const off = s * chunk;\n          const cnt = Math.min(chunk, nEdges - off);\n          if (cnt <= 0) break;\n          const pass = enc.beginComputePass();\n          pass.setPipeline(pipes.edgeAtomic);\n          pass.setBindGroup(0, edgeBindGroups[s]!);\n          pass.dispatchWorkgroups(Math.ceil(cnt / 256));\n          pass.end();\n          enc.copyBufferToBuffer(fxBuf, 0, snapBuf, 0, fxSize);\n        }\n        device.queue.submit([enc.finish()]);\n      }\n      if (epoch % SYNC_EVERY === SYNC_EVERY - 1 || epoch === nEpochs - 1) {\n        await device.queue.onSubmittedWorkDone();\n      }\n      alpha = opts.initialAlpha * (1.0 - epoch / nEpochs);\n      if (opts.onEpoch) {\n        const info: { epoch: number; nEpochs: number; phase: string; embedding?: Float32Array } = {\n          epoch,\n          nEpochs,\n          phase: 'layout',\n        };\n        if (opts.snapshotEvery && (epoch % opts.snapshotEvery === 0 || epoch === nEpochs - 1)) {\n          if (strategy === 'edgeAtomic') run(pipes.unpackFx, bindGroup, nVertices * dim);\n          const snap = new Float32Array(await readBuffer(device, embBuf, embedding.length * 4));\n          info.embedding = snap;\n        }\n        opts.onEpoch(info);\n      }\n    }\n    if (strategy === 'edgeAtomic') {\n      writeUniform(nEpochs, alpha);\n      run(pipes.unpackFx, bindGroup, nVertices * dim);\n    }\n    await device.queue.onSubmittedWorkDone();\n    await endScopes('gpuOptimizeLayout');\n    assertUsable(ctx, 'gpuOptimizeLayout');\n    const out = new Float32Array(await readBuffer(device, embBuf, embedding.length * 4));\n    embedding.set(out);\n    return embedding;\n  } finally {\n    // balance the scope stack even when the body threw first\n    await endScopes('gpuOptimizeLayout').catch(() => {});\n    for (const b of scratch) b.destroy();\n  }\n}\n","\"// Blocked normalized-Laplacian SpMV for the spectral init (D-014):\\n// Y[c*n+i] = X[c*n+i] − dinvsqrt[i] · Σ_p scaled[p]·X[c*n+indices[p]]\\n// One thread per row; each thread walks its CSR row once and accumulates all\\n// block columns. Reduction order per (row, column) is fixed → deterministic.\\n\\nstruct SpmvParams {\\n  n: u32,\\n  cols: u32,\\n}\\n\\n@group(0) @binding(0) var<storage, read> indptr: array<u32>;\\n@group(0) @binding(1) var<storage, read> indices: array<u32>;\\n@group(0) @binding(2) var<storage, read> scaled: array<f32>;\\n@group(0) @binding(3) var<storage, read> dinv: array<f32>;\\n@group(0) @binding(4) var<storage, read> X: array<f32>;\\n@group(0) @binding(5) var<storage, read_write> Y: array<f32>;\\n@group(0) @binding(6) var<uniform> P: SpmvParams;\\n\\n@compute @workgroup_size(256)\\nfn spmv(@builtin(global_invocation_id) gid: vec3<u32>) {\\n  let i = gid.x;\\n  if (i >= P.n) { return; }\\n  let n = P.n;\\n  let start = indptr[i];\\n  let end = indptr[i + 1u];\\n  var acc: array<f32, 18>; // up to 3·(dim+1) block columns, dim ≤ 5 on this path\\n  for (var c = 0u; c < P.cols; c = c + 1u) { acc[c] = 0.0; }\\n  for (var p = start; p < end; p = p + 1u) {\\n    let v = scaled[p];\\n    let j = indices[p];\\n    for (var c = 0u; c < P.cols; c = c + 1u) {\\n      acc[c] = acc[c] + v * X[c * n + j];\\n    }\\n  }\\n  let di = dinv[i];\\n  for (var c = 0u; c < P.cols; c = c + 1u) {\\n    Y[c * n + i] = X[c * n + i] - di * acc[c];\\n  }\\n}\\n\"","/**\n * GPU-assisted spectral initialization for the WebGPU backend (D-014): the\n * LOBPCG driver stays on the CPU in f64, but the dominant cost — the blocked\n * normalized-Laplacian SpMV — runs on the GPU in f32 (deterministic\n * row-per-thread kernel). Single-component graphs only; callers fall back to\n * the CPU `initializeEmbedding` when this returns null. The CPU spectral path\n * is untouched — this module exists so the webgpu backend's end-to-end time\n * reflects the GPU, not a single-threaded eigensolver.\n */\n\nimport { finalInitRescale, noisyScaleCoords } from '../../init/index.js';\nimport { connectedComponents, degreeInvSqrt, lobpcgSmallestAsync } from '../../init/spectral.js';\nimport type { Pcg32 } from '../../rng.js';\nimport type { Csr } from '../../sparse.js';\nimport type { Logger } from '../../types.js';\nimport { silentLogger, throwIfAborted } from '../../types.js';\nimport {\n  assertUsable,\n  beginErrorScopes,\n  createBuffer,\n  ensureBufferFits,\n  type GpuContext,\n  readBuffer,\n} from './device.js';\nimport spmvWgsl from './kernels/spmv.wgsl';\n\ninterface SpmvPipeline {\n  pipeline: GPUComputePipeline;\n  bindLayout: GPUBindGroupLayout;\n}\n\nconst cache = new WeakMap<GPUDevice, SpmvPipeline>();\n\nfunction spmvPipeline(device: GPUDevice): SpmvPipeline {\n  let p = cache.get(device);\n  if (!p) {\n    const module = device.createShaderModule({ code: spmvWgsl });\n    const entry = (binding: number, type: GPUBufferBindingType): GPUBindGroupLayoutEntry => ({\n      binding,\n      visibility: GPUShaderStage.COMPUTE,\n      buffer: { type },\n    });\n    const bindLayout = device.createBindGroupLayout({\n      entries: [\n        entry(0, 'read-only-storage'),\n        entry(1, 'read-only-storage'),\n        entry(2, 'read-only-storage'),\n        entry(3, 'read-only-storage'),\n        entry(4, 'read-only-storage'),\n        entry(5, 'storage'),\n        entry(6, 'uniform'),\n      ],\n    });\n    const layout = device.createPipelineLayout({ bindGroupLayouts: [bindLayout] });\n    p = {\n      pipeline: device.createComputePipeline({\n        layout,\n        compute: { module, entryPoint: 'spmv' },\n      }),\n      bindLayout,\n    };\n    cache.set(device, p);\n  }\n  return p;\n}\n\nexport interface GpuInitResult {\n  embedding: Float32Array; // n × dim, after the final [0,10] rescale\n  spectralFallback: boolean;\n}\n\n/**\n * Spectral init with GPU SpMV. Returns null for multi-component graphs (the\n * caller must then run the CPU init path; no rng draws are consumed here in\n * that case) and for dim > 5 (kernel block bound).\n */\nexport async function gpuSpectralInit(\n  ctx: GpuContext,\n  graph: Csr,\n  dim: number,\n  rng: Pcg32,\n  logger: Logger = silentLogger,\n  signal?: AbortSignal,\n): Promise<GpuInitResult | null> {\n  if (dim > 5) return null;\n  const { count } = connectedComponents(graph);\n  if (count > 1) return null;\n\n  const device = ctx.device;\n  const n = graph.rows;\n  const m = dim + 1;\n  const maxCols = 3 * m;\n\n  // validate planned allocations against the granted limits up front (§A12)\n  assertUsable(ctx, 'gpuSpectralInit');\n  ensureBufferFits(ctx, 'gpuSpectralInit', 'graph', graph.data.length * 4);\n  ensureBufferFits(ctx, 'gpuSpectralInit', 'block', n * maxCols * 4);\n\n  const { deg, dinvsqrt } = degreeInvSqrt(graph);\n\n  // X0 exactly as the CPU single-component path builds it\n  const X0 = new Float64Array(n * m);\n  for (let c = 0; c < m; c++) {\n    for (let r = 0; r < n; r++) X0[c * n + r] = rng.normal();\n  }\n  let norm = 0;\n  for (let r = 0; r < n; r++) norm += deg[r]!;\n  norm = Math.sqrt(norm);\n  for (let r = 0; r < n; r++) X0[r] = Math.sqrt(deg[r]!) / norm;\n\n  // pre-scaled off-diagonals A_ij·dinvsqrt[j] in f32\n  const nnz = graph.data.length;\n  const scaled = new Float32Array(nnz);\n  for (let i = 0; i < n; i++) {\n    for (let p = graph.indptr[i]!; p < graph.indptr[i + 1]!; p++) {\n      scaled[p] = Math.fround(graph.data[p]! * dinvsqrt[graph.indices[p]!]!);\n    }\n  }\n  const indptrU32 = new Uint32Array(n + 1);\n  for (let i = 0; i <= n; i++) indptrU32[i] = graph.indptr[i]!;\n  const indicesU32 = new Uint32Array(nnz);\n  for (let p = 0; p < nnz; p++) indicesU32[p] = graph.indices[p]!;\n  const dinvF32 = Float32Array.from(dinvsqrt);\n\n  const { pipeline, bindLayout } = spmvPipeline(device);\n  const ro = GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST;\n  // every buffer is tracked and destroyed in the finally below (abort-safe)\n  const scratch = new Set<GPUBuffer>();\n  const track = (buf: GPUBuffer): GPUBuffer => {\n    scratch.add(buf);\n    return buf;\n  };\n  const endScopes = beginErrorScopes(device);\n  try {\n    const indptrBuf = track(createBuffer(device, indptrU32, ro));\n    const indicesBuf = track(createBuffer(device, indicesU32, ro));\n    const scaledBuf = track(createBuffer(device, scaled, ro));\n    const dinvBuf = track(createBuffer(device, dinvF32, ro));\n    const xBuf = track(\n      device.createBuffer({\n        size: n * maxCols * 4,\n        usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST,\n      }),\n    );\n    const yBuf = track(\n      device.createBuffer({\n        size: n * maxCols * 4,\n        usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC,\n      }),\n    );\n    const uni = track(\n      device.createBuffer({\n        size: 8,\n        usage: GPUBufferUsage.UNIFORM | GPUBufferUsage.COPY_DST,\n      }),\n    );\n    const bindGroup = device.createBindGroup({\n      layout: bindLayout,\n      entries: [\n        { binding: 0, resource: { buffer: indptrBuf } },\n        { binding: 1, resource: { buffer: indicesBuf } },\n        { binding: 2, resource: { buffer: scaledBuf } },\n        { binding: 3, resource: { buffer: dinvBuf } },\n        { binding: 4, resource: { buffer: xBuf } },\n        { binding: 5, resource: { buffer: yBuf } },\n        { binding: 6, resource: { buffer: uni } },\n      ],\n    });\n\n    const xStage = new Float32Array(n * maxCols);\n    const applyL = async (\n      S: Float64Array,\n      cols: number,\n      out: Float64Array,\n    ): Promise<Float64Array> => {\n      throwIfAborted(signal);\n      assertUsable(ctx, 'gpuSpectralInit');\n      const len = n * cols;\n      for (let i = 0; i < len; i++) xStage[i] = Math.fround(S[i]!);\n      device.queue.writeBuffer(uni, 0, new Uint32Array([n, cols]));\n      device.queue.writeBuffer(xBuf, 0, xStage.buffer, 0, len * 4);\n      const enc = device.createCommandEncoder();\n      const pass = enc.beginComputePass();\n      pass.setPipeline(pipeline);\n      pass.setBindGroup(0, bindGroup);\n      pass.dispatchWorkgroups(Math.ceil(n / 256));\n      pass.end();\n      device.queue.submit([enc.finish()]);\n      const y = new Float32Array(await readBuffer(device, yBuf, len * 4));\n      for (let i = 0; i < len; i++) out[i] = y[i]!;\n      return out;\n    };\n\n    const res = await lobpcgSmallestAsync(\n      graph,\n      dinvsqrt,\n      X0,\n      m,\n      1e-4,\n      5 * Math.min(n, 60),\n      applyL,\n    );\n    await endScopes('gpuSpectralInit');\n    if (!res.converged) {\n      logger.warn(\n        'UmapConvergenceWarning: Spectral initialisation failed! The eigenvector solver failed. This is likely due to too small an eigengap. Consider adding some noise or jitter to your data. Falling back to random initialisation!',\n      );\n      const emb = new Float64Array(n * dim);\n      for (let i = 0; i < emb.length; i++) emb[i] = rng.uniform(-10, 10);\n      return { embedding: finalInitRescale(emb, n, dim), spectralFallback: true };\n    }\n    // drop the trivial eigenvector; columns 1..m → row-major n×dim\n    const emb = new Float64Array(n * dim);\n    for (let d = 0; d < dim; d++) {\n      const c = d + 1;\n      for (let r = 0; r < n; r++) emb[r * dim + d] = res.vectors[c * n + r]!;\n    }\n    noisyScaleCoords(emb, rng, 10, 0.0001);\n    return { embedding: finalInitRescale(emb, n, dim), spectralFallback: false };\n  } finally {\n    // balance the scope stack even when the body threw first\n    await endScopes('gpuSpectralInit').catch(() => {});\n    for (const b of scratch) b.destroy();\n  }\n}\n","/**\n * WebGPU backend entry (§5.4) — separate `umap-web/webgpu` export so browser-only\n * code stays out of the main bundle. The estimator loads this lazily when\n * backend 'webgpu' (or 'auto' with an adapter) is selected.\n */\n\nexport { acquireGpu, type DeviceProvider, type GpuContext } from './device.js';\nexport { type GpuMetricPlan, gpuKnn, gpuMetricPlan } from './knn.js';\nexport {\n  type GpuLayoutOptions,\n  type GpuLayoutStrategy,\n  gpuOptimizeLayout,\n} from './layout.js';\nexport { type GpuInitResult, gpuSpectralInit } from './spectral.js';\n\n/** Which fit configurations the GPU backend supports end-to-end (§4.3). */\nexport function gpuSupportsConfig(opts: {\n  metric?: string | ((a: Float32Array, b: Float32Array) => number);\n  outputMetric?: string;\n  densMap?: boolean;\n  nNeighbors?: number;\n  nComponents?: number;\n}): { knn: boolean; layout: boolean; reason?: string } {\n  const metric = opts.metric ?? 'euclidean';\n  const knnOk =\n    typeof metric === 'string' &&\n    [\n      'euclidean',\n      'l2',\n      'sqeuclidean',\n      'cosine',\n      'correlation',\n      'manhattan',\n      'taxicab',\n      'l1',\n      'hamming',\n      'jaccard',\n    ].includes(metric) &&\n    (opts.nNeighbors ?? 15) <= 32;\n  const outputMetric = opts.outputMetric ?? 'euclidean';\n  const layoutOk =\n    (outputMetric === 'euclidean' || outputMetric === 'l2') &&\n    !opts.densMap &&\n    (opts.nComponents ?? 2) <= 8;\n  const reason = !knnOk\n    ? `metric '${String(metric)}' runs on CPU (GPU subset: euclidean, sqeuclidean, cosine, correlation, manhattan, hamming, jaccard; k<=32)`\n    : !layoutOk\n      ? 'non-euclidean output / densMAP / dim>8 layout runs on CPU'\n      : undefined;\n  return { knn: knnOk, layout: layoutOk, ...(reason ? { reason } : {}) };\n}\n"],"mappings":";;;AAuBA,SAAS,UAAU,KAAuB;AACxC,KAAI,OAAO,KACR,MAAM,SAAS;AACd,MAAI,OAAO;GAAE,QAAQ,OAAO,KAAK,OAAO;GAAE,SAAS,KAAK;GAAS;GACjE,CACD,YAAY;AACX,MAAI,OAAO;GAAE,QAAQ;GAAW,SAAS;GAAwB;GACjE;;;AAIN,SAAgB,aAAa,KAAiB,OAAqB;AACjE,KAAI,IAAI,KACN,OAAM,IAAI,iBACR,0BAA0B,MAAM,IAAI,IAAI,KAAK,OAAO,KAAK,IAAI,KAAK,UACnE;;;;;;;AASL,SAAgB,iBACd,KACA,OACA,MACA,YACM;AACN,KAAI,aAAa,IAAI,OAAO,cAC1B,OAAM,IAAI,iBACR,GAAG,MAAM,YAAY,KAAK,KAAK,WAAW,+CACtB,IAAI,OAAO,cAAc,0CAC9C;AAEH,KAAI,aAAa,IAAI,OAAO,4BAC1B,OAAM,IAAI,iBACR,GAAG,MAAM,YAAY,KAAK,KAAK,WAAW,6DACR,IAAI,OAAO,4BAA4B,0CAE1E;;;;;;;;AAUL,SAAgB,iBAAiB,QAAqD;AACpF,QAAO,eAAe,gBAAgB;AACtC,QAAO,eAAe,aAAa;CACnC,IAAI,SAAS;AACb,QAAO,OAAO,UAAkB;AAC9B,MAAI,OAAQ;AACZ,WAAS;EACT,MAAM,aAAa,MAAM,OAAO,eAAe;EAC/C,MAAM,MAAM,MAAM,OAAO,eAAe;AACxC,MAAI,IACF,OAAM,IAAI,iBAAiB,GAAG,MAAM,uBAAuB,IAAI,UAAU;AAE3E,MAAI,WACF,OAAM,IAAI,iBAAiB,GAAG,MAAM,0BAA0B,WAAW,UAAU;;;AAKzF,eAAsB,WAAW,UAAgD;AAC/E,KAAI,UAAU;EACZ,MAAMA,WACJ,OAAO,aAAa,aAChB,MAAM,UAAU,GAChB,MAAM,QAAQ,QAAQ,SAA2C;EACvE,MAAMC,QAAkB;GACtB;GACA,aAAa;IACX,QAAQ;IACR,cAAc;IACd,QAAQ;IACR,aAAa;IACd;GACD,QAAQ;IACN,6BAA6BD,SAAO,OAAO;IAC3C,eAAeA,SAAO,OAAO;IAC7B,gBAAgBA,SAAO,OAAO;IAC/B;GACD,YAAYA,SAAO,SAAS,IAAI,kBAAkB;GAClD,UAAU;GACV,MAAM;GACP;AACD,YAAUE,MAAI;AACd,SAAOA;;CAET,MAAM,MAAO,WAAW,WAAyC;AACjE,KAAI,CAAC,IACH,OAAM,IAAI,iBAAiB,oDAAoD;CAEjF,MAAM,UAAU,MAAM,IAAI,gBAAgB;AAC1C,KAAI,CAAC,QACH,OAAM,IAAI,iBAAiB,4DAA4D;CAEzF,MAAM,iBAAiB,QAAQ,SAAS,IAAI,kBAAkB;CAC9D,MAAM,SAAS,MAAM,QAAQ,cAAc;EACzC,gBAAgB;GACd,6BAA6B,QAAQ,OAAO;GAC5C,eAAe,QAAQ,OAAO;GAC/B;EACD,kBAAkB,iBAAkB,CAAC,kBAAkB,GAAwB,EAAE;EAClF,CAAC;CACF,MAAM,OAAO,QAAQ;CACrB,MAAMD,MAAkB;EACtB;EACA,aAAa;GACX,QAAQ,MAAM,UAAU;GACxB,cAAc,MAAM,gBAAgB;GACpC,QAAQ,MAAM,UAAU;GACxB,aAAa,MAAM,eAAe;GACnC;EACD,QAAQ;GACN,6BAA6B,OAAO,OAAO;GAC3C,eAAe,OAAO,OAAO;GAC7B,gBAAgB,OAAO,OAAO;GAC/B;EACD,YAAY,OAAO,SAAS,IAAI,kBAAkB;EAClD,UAAU;EACV,MAAM;EACP;AACD,WAAU,IAAI;AACd,QAAO;;AAGT,SAAgB,aACd,QACA,MACA,OACW;CACX,MAAM,MAAM,OAAO,aAAa;EAAE,MAAM,KAAK,KAAK,KAAK,aAAa,EAAE,GAAG;EAAG;EAAO,CAAC;AACpF,QAAO,MAAM,YAAY,KAAK,GAAG,KAAK,QAAQ,KAAK,YAAY,KAAK,WAAW;AAC/E,QAAO;;AAGT,eAAsB,WACpB,QACA,KACA,YACsB;CACtB,MAAM,UAAU,OAAO,aAAa;EAClC,MAAM,KAAK,KAAK,aAAa,EAAE,GAAG;EAClC,OAAO,eAAe,WAAW,eAAe;EACjD,CAAC;AACF,KAAI;EACF,MAAM,MAAM,OAAO,sBAAsB;AACzC,MAAI,mBAAmB,KAAK,GAAG,SAAS,GAAG,QAAQ,KAAK;AACxD,SAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;AACnC,QAAM,QAAQ,SAAS,WAAW,KAAK;EACvC,MAAM,MAAM,QAAQ,gBAAgB,CAAC,MAAM,EAAE;AAC7C,UAAQ,OAAO;AACf,SAAO;UACA,GAAG;AACV,QAAM,aAAa,mBACf,IACA,IAAI,iBAAiB,wBAAyB,EAAY,UAAU;WAChE;AACR,UAAQ,SAAS;;;;;;kBC5LrB;;;;;AC2BA,SAAgB,cAAc,MAAoC;AAChE,SAAQ,MAAR;EACE,KAAK;EACL,KAAK,KACH,QAAO;GAAE,MAAM;GAAG,QAAQ;GAAO,aAAa,MAAM,KAAK,KAAK,EAAE;GAAE;EACpE,KAAK,cACH,QAAO;GAAE,MAAM;GAAG,QAAQ;GAAO,aAAa,MAAM;GAAG;EACzD,KAAK,SACH,QAAO;GAAE,MAAM;GAAG,QAAQ;GAAO,aAAa,MAAM,IAAM,KAAO,CAAC;GAAG;EACvE,KAAK,cACH,QAAO;GAAE,MAAM;GAAG,QAAQ;GAAM,aAAa,MAAM,IAAM,KAAO,CAAC;GAAG;EACtE,KAAK;EACL,KAAK;EACL,KAAK,KACH,QAAO;GAAE,MAAM;GAAG,QAAQ;GAAO,aAAa,MAAM;GAAG;EACzD,KAAK,UACH,QAAO;GAAE,MAAM;GAAG,QAAQ;GAAO,aAAa,MAAM;GAAG;EACzD,KAAK,UACH,QAAO;GAAE,MAAM;GAAG,QAAQ;GAAO,aAAa,MAAM,IAAM,KAAO,CAAC;GAAG;EACvE,QACE,QAAO;;;AAIb,MAAM,cAAc;AAQpB,MAAM,gCAAgB,IAAI,SAAkC;AAE5D,SAAS,aAAa,QAAiC;CACrD,IAAI,IAAI,cAAc,IAAI,OAAO;AACjC,KAAI,CAAC,GAAG;EACN,MAAM,SAAS,OAAO,mBAAmB,EAAE,MAAME,aAAS,CAAC;AAC3D,MAAI;GACF;GACA,MAAM,OAAO,sBAAsB;IACjC,QAAQ;IACR,SAAS;KAAE;KAAQ,YAAY;KAAgB;IAChD,CAAC;GACF,OAAO,OAAO,sBAAsB;IAClC,QAAQ;IACR,SAAS;KAAE;KAAQ,YAAY;KAAY;IAC5C,CAAC;GACH;AACD,gBAAc,IAAI,QAAQ,EAAE;;AAE9B,QAAO;;;;;;AAcT,eAAsB,OACpB,KACA,GACA,SACA,GACA,MACA,EAAE,WAAW,YAAY,WAA0B,EAAE,EAClC;AACnB,KAAI,IAAI,GAAI,OAAM,IAAI,iBAAiB,2BAA2B;CAClE,MAAM,SAAS,IAAI;CACnB,MAAM,IAAI,EAAE;CACZ,MAAM,IAAI,EAAE;CACZ,MAAM,KAAK,KAAK,KAAK,IAAI,EAAE;CAC3B,MAAM,OAAO,KAAK;CAClB,MAAM,WAAW,YAAY,OAAO,IAAI,QAAQ;AAKhD,cAAa,KAAK,SAAS;AAC3B,kBAAiB,KAAK,UAAU,QAAQ,IAAI,OAAO,EAAE;AACrD,KAAI,YAAY,KAAM,kBAAiB,KAAK,UAAU,WAAW,WAAW,OAAO,EAAE;AACrF,kBAAiB,KAAK,UAAU,gBAAgB,WAAW,IAAI,EAAE;AACjE,kBAAiB,KAAK,UAAU,cAAc,WAAW,IAAI,EAAE;CAG/D,MAAM,WAAW,KAAmB,MAAc,WAAkC;AAClF,MAAI,SAAS,KAAK,CAAC,OAAQ,QAAO;EAClC,MAAM,MAAM,IAAI,aAAa,OAAO,KAAK;AACzC,OAAK,IAAI,IAAI,GAAG,IAAI,MAAM,KAAK;GAC7B,IAAI,KAAK;AACT,OAAI,QAAQ;AACV,SAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IAAK,OAAM,IAAI,IAAI,IAAI;AAC9C,UAAM;;AAER,QAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IAAK,KAAI,IAAI,OAAO,KAAK,IAAI,IAAI,IAAI,KAAM;;AAEpE,SAAO;;CAET,MAAM,WAAW,QAAQ,EAAE,MAAM,GAAG,KAAK,OAAO;CAChD,MAAM,YAAY,YAAY;CAC9B,MAAM,KAAK,YAAY,IAAI,QAAS;CACpC,MAAM,QAAQ,YAAY,WAAW,QAAQ,QAAS,MAAM,IAAI,KAAK,OAAO;CAE5E,MAAM,QAAQ,aAAa,OAAO;CAGlC,MAAM,0BAAU,IAAI,KAAgB;CACpC,MAAM,SAAS,QAA8B;AAC3C,UAAQ,IAAI,IAAI;AAChB,SAAO;;CAET,MAAM,YAAY,iBAAiB,OAAO;CAC1C,IAAIC;CACJ,IAAIC;AACJ,KAAI;EACF,MAAM,UAAU,MACd,aAAa,QAAQ,UAAU,eAAe,UAAU,eAAe,SAAS,CACjF;EACD,MAAM,OAAO,YACT,UACA,MAAM,aAAa,QAAQ,OAAO,eAAe,UAAU,eAAe,SAAS,CAAC;EACxF,MAAM,UAAU,MACd,OAAO,aAAa;GAClB,MAAM,KAAK,IAAI;GACf,OAAO,eAAe,UAAU,eAAe;GAChD,CAAC,CACH;EACD,MAAM,SAAS,MACb,OAAO,aAAa;GAClB,MAAM,KAAK,IAAI;GACf,OAAO,eAAe,UAAU,eAAe;GAChD,CAAC,CACH;EACD,MAAM,YAAY,MAChB,OAAO,aAAa;GAClB,MAAM;GACN,OAAO,eAAe,UAAU,eAAe;GAChD,CAAC,CACH;EACD,MAAM,QAAQ,MAAM,OAAO,aAAa;GAAE,MAAM;GAAG,OAAO,eAAe;GAAS,CAAC,CAAC;EAGpF,IAAIC,WAAsB;EAC1B,IAAIC,YAAuB;AAC3B,MAAI,KAAK,SAAS,GAAG;AACnB,cAAW,MACT,OAAO,aAAa;IAClB,MAAM,IAAI;IACV,OAAO,eAAe,UAAU,eAAe;IAChD,CAAC,CACH;GACD,MAAM,KAAK,MACT,OAAO,aAAa;IAClB,MAAM;IACN,OAAO,eAAe,UAAU,eAAe;IAChD,CAAC,CACH;AACD,UAAO,MAAM,YAAY,IAAI,GAAG,IAAI,YAAY;IAAC;IAAG;IAAI;IAAG,KAAK;IAAM;IAAG;IAAG;IAAG;IAAE,CAAC,CAAC;GACnF,MAAM,KAAK,OAAO,gBAAgB;IAChC,QAAQ,MAAM,MAAM,mBAAmB,EAAE;IACzC,SAAS;KACP;MAAE,SAAS;MAAG,UAAU,EAAE,QAAQ,SAAS;MAAE;KAC7C;MAAE,SAAS;MAAG,UAAU,EAAE,QAAQ,UAAU;MAAE;KAC9C;MAAE,SAAS;MAAG,UAAU,EAAE,QAAQ,IAAI;MAAE;KACzC;IACF,CAAC;GACF,MAAM,MAAM,OAAO,sBAAsB;GACzC,MAAM,OAAO,IAAI,kBAAkB;AACnC,QAAK,YAAY,MAAM,MAAM;AAC7B,QAAK,aAAa,GAAG,GAAG;AACxB,QAAK,mBAAmB,KAAK,KAAK,IAAI,IAAI,CAAC;AAC3C,QAAK,KAAK;AACV,UAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;AACnC,OAAI,UACF,aAAY;QACP;AACL,gBAAY,MACV,OAAO,aAAa;KAClB,MAAM,KAAK;KACX,OAAO,eAAe,UAAU,eAAe;KAChD,CAAC,CACH;IACD,MAAM,KAAK,MACT,OAAO,aAAa;KAClB,MAAM;KACN,OAAO,eAAe,UAAU,eAAe;KAChD,CAAC,CACH;AACD,WAAO,MAAM,YAAY,IAAI,GAAG,IAAI,YAAY;KAAC;KAAI;KAAI;KAAG,KAAK;KAAM;KAAI;KAAG;KAAI;KAAE,CAAC,CAAC;IACtF,MAAM,MAAM,OAAO,gBAAgB;KACjC,QAAQ,MAAM,MAAM,mBAAmB,EAAE;KACzC,SAAS;MACP;OAAE,SAAS;OAAG,UAAU,EAAE,QAAQ,MAAM;OAAE;MAC1C;OAAE,SAAS;OAAG,UAAU,EAAE,QAAQ,WAAW;OAAE;MAC/C;OAAE,SAAS;OAAG,UAAU,EAAE,QAAQ,IAAI;OAAE;MACzC;KACF,CAAC;IACF,MAAM,OAAO,OAAO,sBAAsB;IAC1C,MAAM,QAAQ,KAAK,kBAAkB;AACrC,UAAM,YAAY,MAAM,MAAM;AAC9B,UAAM,aAAa,GAAG,IAAI;AAC1B,UAAM,mBAAmB,KAAK,KAAK,KAAK,IAAI,CAAC;AAC7C,UAAM,KAAK;AACX,WAAO,MAAM,OAAO,CAAC,KAAK,QAAQ,CAAC,CAAC;;;EAIxC,MAAM,YAAY,OAAO,gBAAgB;GACvC,QAAQ,MAAM,KAAK,mBAAmB,EAAE;GACxC,SAAS;IACP;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,SAAS;KAAE;IAC7C;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,MAAM;KAAE;IAC1C;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,SAAS;KAAE;IAC7C;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,QAAQ;KAAE;IAC5C;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,WAAW;KAAE;IAC/C;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,UAAU;KAAE;IAC9C;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,WAAW;KAAE;IAChD;GACF,CAAC;EAGF,MAAM,QAAQ,aAAa,KAAK,IAAI,MAAM,KAAK,IAAI,GAAG,KAAK,MAAM,OAAO,KAAK,IAAI,GAAG,KAAK,EAAE,CAAC,CAAC,CAAC;EAC9F,MAAM,UAAU,KAAK,KAAK,KAAK,IAAI;AACnC,OAAK,IAAI,KAAK,GAAG,KAAK,GAAG,MAAM,OAAO;AACpC,kBAAe,OAAO;AACtB,gBAAa,KAAK,SAAS;GAC3B,MAAM,KAAK,KAAK,IAAI,GAAG,KAAK,MAAM;AAClC,UAAO,MAAM,YAAY,WAAW,GAAG,IAAI,YAAY;IAAC;IAAG;IAAI;IAAG,KAAK;IAAM;IAAI;IAAI;IAAI;IAAE,CAAC,CAAC;GAC7F,MAAM,MAAM,OAAO,sBAAsB;GACzC,MAAM,OAAO,IAAI,kBAAkB;AACnC,QAAK,YAAY,MAAM,KAAK;AAC5B,QAAK,aAAa,GAAG,UAAU;AAC/B,QAAK,mBAAmB,QAAQ;AAChC,QAAK,KAAK;AACV,UAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;AACnC,SAAM,OAAO,MAAM,qBAAqB;AACxC,gBAAa,IAAI,EAAE;;AAErB,QAAM,UAAU,SAAS;AACzB,eAAa,KAAK,SAAS;AAE3B,SAAO,IAAI,aAAa,MAAM,WAAW,QAAQ,SAAS,KAAK,IAAI,EAAE,CAAC;AACtE,QAAM,IAAI,WAAW,MAAM,WAAW,QAAQ,QAAQ,KAAK,IAAI,EAAE,CAAC;WAC1D;AAER,QAAM,UAAU,SAAS,CAAC,YAAY,GAAG;AACzC,OAAK,MAAM,KAAK,QAAS,GAAE,SAAS;;CAItC,MAAM,UAAU,IAAI,WAAW,KAAK,EAAE;CACtC,MAAM,YAAY,IAAI,aAAa,KAAK,EAAE;CAC1C,MAAM,QAAQ,IAAI,MAAc,EAAE;AAClC,MAAK,IAAI,IAAI,GAAG,IAAI,IAAI,KAAK;AAC3B,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IAAK,OAAM,KAAK;AACvC,QAAM,MAAM,GAAG,MAAM,KAAK,IAAI,IAAI,KAAM,KAAK,IAAI,IAAI,GAAI;AACzD,OAAK,IAAI,IAAI,GAAG,IAAI,GAAG,KAAK;GAC1B,MAAM,IAAI,MAAM;GAChB,IAAI,MAAM,KAAK,IAAI,IAAI;AACvB,OAAI,KAAK,SAAS,EAAG,QAAO;AAC5B,WAAQ,IAAI,IAAI,KAAK,IAAI,IAAI,IAAI;AACjC,aAAU,IAAI,IAAI,KAAK,OAAO,QAAS,WAAW,KAAK,WAAW,IAAI;;;AAG1E,QAAO;EAAE;EAAS;EAAW;EAAG;;;;;kBCzSlC;;;;AC6BA,MAAMC,0BAAQ,IAAI,SAAkC;AAOpD,SAAS,UAAU,QAAmC;CACpD,IAAI,IAAIA,QAAM,IAAI,OAAO;AACzB,KAAI,CAAC,GAAG;EACN,MAAM,SAAS,OAAO,mBAAmB,EAAE,MAAMC,aAAS,CAAC;EAC3D,MAAM,WAAW,SAAiB,UAAyD;GACzF;GACA,YAAY,eAAe;GAC3B,QAAQ,EAAE,MAAM;GACjB;EAED,MAAM,aAAa,OAAO,sBAAsB,EAC9C,SAAS;GACP,QAAQ,GAAG,UAAU;GACrB,QAAQ,GAAG,oBAAoB;GAC/B,QAAQ,GAAG,oBAAoB;GAC/B,QAAQ,GAAG,oBAAoB;GAC/B,QAAQ,GAAG,UAAU;GACrB,QAAQ,GAAG,UAAU;GACrB,QAAQ,GAAG,UAAU;GACrB,QAAQ,GAAG,oBAAoB;GAC/B,QAAQ,GAAG,UAAU;GACtB,EACF,CAAC;EAGF,MAAM,iBAAiB,OAAO,sBAAsB,EAClD,SAAS;GACP,QAAQ,GAAG,oBAAoB;GAC/B,QAAQ,GAAG,oBAAoB;GAC/B,QAAQ,GAAG,oBAAoB;GAC/B,QAAQ,GAAG,UAAU;GACrB,QAAQ,GAAG,UAAU;GACrB,QAAQ,GAAG,UAAU;GACrB,QAAQ,GAAG,UAAU;GACrB,QAAQ,IAAI,UAAU;GACtB,QAAQ,IAAI,oBAAoB;GACjC,EACF,CAAC;EACF,MAAM,SAAS,OAAO,qBAAqB,EAAE,kBAAkB,CAAC,WAAW,EAAE,CAAC;EAC9E,MAAM,aAAa,OAAO,qBAAqB,EAAE,kBAAkB,CAAC,eAAe,EAAE,CAAC;EACtF,MAAM,MAAM,YAAoB,KAAwB,WACtD,OAAO,sBAAsB;GAAE,QAAQ;GAAI,SAAS;IAAE;IAAQ;IAAY;GAAE,CAAC;AAC/E,MAAI;GACF,cAAc,GAAG,eAAe;GAChC,YAAY,GAAG,cAAc,WAAW;GACxC,QAAQ,GAAG,SAAS;GACpB,UAAU,GAAG,WAAW;GACxB;GACA;GACD;AACD,UAAM,IAAI,QAAQ,EAAE;;AAEtB,QAAO;;AAmBT,MAAM,eAAe;AAErB,eAAsB,kBACpB,KACA,WACA,KACA,MACA,MACA,SACA,WACA,iBACA,GACA,GACA,KACA,MACuB;CACvB,MAAM,SAAS,IAAI;CACnB,MAAM,WAAW,KAAK,YAAY;CAClC,MAAM,SAAS,KAAK;AACpB,KAAI,aAAa,gBAAgB,MAAM,EACrC,OAAM,IAAI,iBAAiB,wCAAwC,IAAI,GAAG;CAE5E,MAAM,QAAQ,UAAU,OAAO;AAG/B,cAAa,KAAK,oBAAoB;CACtC,MAAM,gBAAgB,aAAa,eAAe,YAAY,MAAM,IAAI;AACxE,kBAAiB,KAAK,qBAAqB,aAAa,UAAU,WAAW;AAC7E,kBAAiB,KAAK,qBAAqB,SAAS,SAAS,EAAE;AAC/D,kBAAiB,KAAK,qBAAqB,YAAY,SAAS,IAAI,EAAE;AACtE,kBACE,KACA,qBACA,cACC,aAAa,iBAAiB,YAAY,UAAU,IAAI,EAC1D;AACD,kBAAiB,KAAK,qBAAqB,cAAc,cAAc;CAEvE,MAAM,SAAS,aAAa,KAAK,gBAAgB;CACjD,MAAM,eAAe,IAAI,aAAa,SAAS,EAAE;AACjD,cAAa,IAAI,QAAQ,EAAE;AAC3B,MAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,IAC1B,cAAa,SAAS,KAAK,OAAO,KAAM,KAAK;CAI/C,MAAM,0BAAU,IAAI,KAAgB;CACpC,MAAM,SAAS,QAA8B;AAC3C,UAAQ,IAAI,IAAI;AAChB,SAAO;;CAET,MAAM,YAAY,iBAAiB,OAAO;AAC1C,KAAI;EACF,MAAM,SAAS,MACb,aACE,QACA,WACA,eAAe,UAAU,eAAe,WAAW,eAAe,SACnE,CACF;EACD,MAAM,UAAU,MACd,aAAa,QAAQ,MAAM,eAAe,UAAU,eAAe,SAAS,CAC7E;EACD,MAAM,UAAU,MACd,aAAa,QAAQ,MAAM,eAAe,UAAU,eAAe,SAAS,CAC7E;EACD,MAAM,SAAS,MACb,aAAa,QAAQ,QAAQ,eAAe,UAAU,eAAe,SAAS,CAC/E;EACD,MAAM,cAAc,MAClB,aAAa,QAAQ,cAAc,eAAe,UAAU,eAAe,SAAS,CACrF;EAGD,MAAM,WAAW,aAAa,iBAAiB,YAAY;EAC3D,MAAM,SAAS,IAAI,YAAY,WAAW,EAAE;AAC5C,OAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,IAAK,QAAO,KAAK,IAAI,SAAS;EACjE,MAAM,SAAS,MACb,aAAa,QAAQ,QAAQ,eAAe,UAAU,eAAe,SAAS,CAC/E;EAGD,MAAM,SAAS,IAAI,YAAY,YAAY,EAAE;AAC7C,OAAK,IAAI,IAAI,GAAG,IAAI,QAAQ,IAAK,QAAO,KAAK,KAAM;AACnD,OAAK,IAAI,IAAI,GAAG,IAAI,WAAW,IAAK,QAAO,IAAI,MAAO,OAAO;EAC7D,MAAM,YAAY,MAChB,aAAa,QAAQ,QAAQ,eAAe,UAAU,eAAe,SAAS,CAC/E;EAED,MAAM,QAAQ;EACd,MAAM,SAAS;EACf,MAAM,QAAQ,MACZ,OAAO,aAAa;GAClB,MAAM;GACN,OAAO,eAAe,UAAU,eAAe;GAChD,CAAC,CACH;EAED,MAAM,UAAU,MACd,OAAO,aAAa;GAClB,MAAM;GACN,OAAO,eAAe,UAAU,eAAe;GAChD,CAAC,CACH;EAED,MAAM,MAAM,MACV,OAAO,aAAa;GAClB,MAAM;GACN,OAAO,eAAe,UAAU,eAAe;GAChD,CAAC,CACH;EACD,MAAM,gBAAgB,OAAe,YAAkB;GACrD,MAAM,MAAM,IAAI,YAAY,aAAa;GACzC,MAAM,IAAI,IAAI,YAAY,IAAI;GAC9B,MAAM,IAAI,IAAI,aAAa,IAAI;AAC/B,KAAE,KAAK;AACP,KAAE,KAAK;AACP,KAAE,KAAK;AACP,KAAE,KAAK;AACP,KAAE,KAAKC;AACP,KAAE,KAAK;AACP,KAAE,KAAK;AACP,KAAE,KAAK,KAAK;AACZ,KAAE,KAAK,KAAK;AACZ,KAAE,KAAK,KAAK,YAAY,IAAI;AAC5B,KAAE,MAAM;AACR,UAAO,MAAM,YAAY,KAAK,GAAG,IAAI;;EAMvC,MAAM,aAAa,aAAa,eAAe,KAAK,IAAI,IAAI,KAAK,IAAI,GAAG,OAAO,CAAC,GAAG;EACnF,MAAM,QAAQ,KAAK,KAAK,SAAS,WAAW;EAC5C,MAAM,aAAa,KAAK,IAAI,KAAK,OAAO,OAAO,gCAAgC;EAC/E,MAAM,WAAW,MACf,OAAO,aAAa;GAClB,MAAM,aAAa;GACnB,OAAO,eAAe,UAAU,eAAe;GAChD,CAAC,CACH;AACD,OAAK,IAAI,IAAI,GAAG,IAAI,YAAY,KAAK;GACnC,MAAM,MAAM,IAAI;GAChB,MAAM,MAAM,KAAK,IAAI,GAAG,KAAK,IAAI,OAAO,SAAS,IAAI,CAAC;AACtD,UAAO,MAAM,YAAY,UAAU,IAAI,YAAY,IAAI,YAAY,CAAC,KAAK,IAAI,CAAC,CAAC;;EAGjF,MAAM,YAAY,OAAO,gBAAgB;GACvC,QAAQ,MAAM;GACd,SAAS;IACP;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,QAAQ;KAAE;IAC5C;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,SAAS;KAAE;IAC7C;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,SAAS;KAAE;IAC7C;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,QAAQ;KAAE;IAC5C;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,aAAa;KAAE;IACjD;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,QAAQ;KAAE;IAC5C;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,KAAK;KAAE;IACzC;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,WAAW;KAAE;IAC/C;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,OAAO;KAAE;IAC5C;GACF,CAAC;EACF,MAAMC,iBAAiC,EAAE;AACzC,OAAK,IAAI,IAAI,GAAG,IAAI,YAAY,IAC9B,gBAAe,KACb,OAAO,gBAAgB;GACrB,QAAQ,MAAM;GACd,SAAS;IACP;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,SAAS;KAAE;IAC7C;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,SAAS;KAAE;IAC7C;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,QAAQ;KAAE;IAC5C;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,aAAa;KAAE;IACjD;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,QAAQ;KAAE;IAC5C;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,KAAK;KAAE;IACzC;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,OAAO;KAAE;IAC3C;KAAE,SAAS;KAAI,UAAU;MAAE,QAAQ;MAAU,QAAQ,IAAI;MAAY,MAAM;MAAG;KAAE;IAChF;KAAE,SAAS;KAAI,UAAU,EAAE,QAAQ,SAAS;KAAE;IAC/C;GACF,CAAC,CACH;EAGH,MAAM,OAAO,UAA8B,IAAkB,YAAoB;GAC/E,MAAM,MAAM,OAAO,sBAAsB;GACzC,MAAM,OAAO,IAAI,kBAAkB;AACnC,QAAK,YAAY,SAAS;AAC1B,QAAK,aAAa,GAAG,GAAG;AACxB,QAAK,mBAAmB,KAAK,KAAK,UAAU,IAAI,CAAC;AACjD,QAAK,KAAK;AACV,UAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;;AAGrC,MAAI,aAAa,cAAc;AAC7B,gBAAa,GAAG,KAAK,aAAa;GAClC,MAAM,MAAM,OAAO,sBAAsB;GACzC,MAAM,OAAO,IAAI,kBAAkB;AACnC,QAAK,YAAY,MAAM,OAAO;AAC9B,QAAK,aAAa,GAAG,UAAU;AAC/B,QAAK,mBAAmB,KAAK,KAAM,YAAY,MAAO,IAAI,CAAC;AAC3D,QAAK,KAAK;AACV,OAAI,mBAAmB,OAAO,GAAG,SAAS,GAAG,OAAO;AACpD,UAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;;EAGrC,IAAI,QAAQ,KAAK;EACjB,MAAM,aAAa;AACnB,OAAK,IAAI,QAAQ,GAAG,QAAQ,SAAS,SAAS;AAC5C,kBAAe,KAAK,OAAO;AAC3B,gBAAa,KAAK,oBAAoB;AACtC,gBAAa,OAAO,MAAM;AAC1B,OAAI,aAAa,eACf,KAAI,MAAM,cAAc,WAAW,UAAU;QACxC;IAGL,MAAM,MAAM,OAAO,sBAAsB;AACzC,SAAK,IAAI,IAAI,GAAG,IAAI,YAAY,KAAK;KACnC,MAAM,MAAM,IAAI;KAChB,MAAM,MAAM,KAAK,IAAI,OAAO,SAAS,IAAI;AACzC,SAAI,OAAO,EAAG;KACd,MAAM,OAAO,IAAI,kBAAkB;AACnC,UAAK,YAAY,MAAM,WAAW;AAClC,UAAK,aAAa,GAAG,eAAe,GAAI;AACxC,UAAK,mBAAmB,KAAK,KAAK,MAAM,IAAI,CAAC;AAC7C,UAAK,KAAK;AACV,SAAI,mBAAmB,OAAO,GAAG,SAAS,GAAG,OAAO;;AAEtD,WAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;;AAErC,OAAI,QAAQ,eAAe,aAAa,KAAK,UAAU,UAAU,EAC/D,OAAM,OAAO,MAAM,qBAAqB;AAE1C,WAAQ,KAAK,gBAAgB,IAAM,QAAQ;AAC3C,OAAI,KAAK,SAAS;IAChB,MAAMC,OAAoF;KACxF;KACA;KACA,OAAO;KACR;AACD,QAAI,KAAK,kBAAkB,QAAQ,KAAK,kBAAkB,KAAK,UAAU,UAAU,IAAI;AACrF,SAAI,aAAa,aAAc,KAAI,MAAM,UAAU,WAAW,YAAY,IAAI;AAE9E,UAAK,YADQ,IAAI,aAAa,MAAM,WAAW,QAAQ,QAAQ,UAAU,SAAS,EAAE,CAAC;;AAGvF,SAAK,QAAQ,KAAK;;;AAGtB,MAAI,aAAa,cAAc;AAC7B,gBAAa,SAAS,MAAM;AAC5B,OAAI,MAAM,UAAU,WAAW,YAAY,IAAI;;AAEjD,QAAM,OAAO,MAAM,qBAAqB;AACxC,QAAM,UAAU,oBAAoB;AACpC,eAAa,KAAK,oBAAoB;EACtC,MAAM,MAAM,IAAI,aAAa,MAAM,WAAW,QAAQ,QAAQ,UAAU,SAAS,EAAE,CAAC;AACpF,YAAU,IAAI,IAAI;AAClB,SAAO;WACC;AAER,QAAM,UAAU,oBAAoB,CAAC,YAAY,GAAG;AACpD,OAAK,MAAMC,OAAK,QAAS,KAAE,SAAS;;;;;;mBC9WxC;;;;AC+BA,MAAM,wBAAQ,IAAI,SAAkC;AAEpD,SAAS,aAAa,QAAiC;CACrD,IAAI,IAAI,MAAM,IAAI,OAAO;AACzB,KAAI,CAAC,GAAG;EACN,MAAM,SAAS,OAAO,mBAAmB,EAAE,MAAMC,cAAU,CAAC;EAC5D,MAAM,SAAS,SAAiB,UAAyD;GACvF;GACA,YAAY,eAAe;GAC3B,QAAQ,EAAE,MAAM;GACjB;EACD,MAAM,aAAa,OAAO,sBAAsB,EAC9C,SAAS;GACP,MAAM,GAAG,oBAAoB;GAC7B,MAAM,GAAG,oBAAoB;GAC7B,MAAM,GAAG,oBAAoB;GAC7B,MAAM,GAAG,oBAAoB;GAC7B,MAAM,GAAG,oBAAoB;GAC7B,MAAM,GAAG,UAAU;GACnB,MAAM,GAAG,UAAU;GACpB,EACF,CAAC;EACF,MAAM,SAAS,OAAO,qBAAqB,EAAE,kBAAkB,CAAC,WAAW,EAAE,CAAC;AAC9E,MAAI;GACF,UAAU,OAAO,sBAAsB;IACrC;IACA,SAAS;KAAE;KAAQ,YAAY;KAAQ;IACxC,CAAC;GACF;GACD;AACD,QAAM,IAAI,QAAQ,EAAE;;AAEtB,QAAO;;;;;;;AAaT,eAAsB,gBACpB,KACA,OACA,KACA,KACA,SAAiB,cACjB,QAC+B;AAC/B,KAAI,MAAM,EAAG,QAAO;CACpB,MAAM,EAAE,UAAU,oBAAoB,MAAM;AAC5C,KAAI,QAAQ,EAAG,QAAO;CAEtB,MAAM,SAAS,IAAI;CACnB,MAAM,IAAI,MAAM;CAChB,MAAM,IAAI,MAAM;CAChB,MAAM,UAAU,IAAI;AAGpB,cAAa,KAAK,kBAAkB;AACpC,kBAAiB,KAAK,mBAAmB,SAAS,MAAM,KAAK,SAAS,EAAE;AACxE,kBAAiB,KAAK,mBAAmB,SAAS,IAAI,UAAU,EAAE;CAElE,MAAM,EAAE,KAAK,aAAa,cAAc,MAAM;CAG9C,MAAM,KAAK,IAAI,aAAa,IAAI,EAAE;AAClC,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IAAK,IAAG,IAAI,IAAI,KAAK,IAAI,QAAQ;CAE1D,IAAI,OAAO;AACX,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IAAK,SAAQ,IAAI;AACxC,QAAO,KAAK,KAAK,KAAK;AACtB,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IAAK,IAAG,KAAK,KAAK,KAAK,IAAI,GAAI,GAAG;CAGzD,MAAM,MAAM,MAAM,KAAK;CACvB,MAAM,SAAS,IAAI,aAAa,IAAI;AACpC,MAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IACrB,MAAK,IAAI,IAAI,MAAM,OAAO,IAAK,IAAI,MAAM,OAAO,IAAI,IAAK,IACvD,QAAO,KAAK,KAAK,OAAO,MAAM,KAAK,KAAM,SAAS,MAAM,QAAQ,IAAM;CAG1E,MAAM,YAAY,IAAI,YAAY,IAAI,EAAE;AACxC,MAAK,IAAI,IAAI,GAAG,KAAK,GAAG,IAAK,WAAU,KAAK,MAAM,OAAO;CACzD,MAAM,aAAa,IAAI,YAAY,IAAI;AACvC,MAAK,IAAI,IAAI,GAAG,IAAI,KAAK,IAAK,YAAW,KAAK,MAAM,QAAQ;CAC5D,MAAM,UAAU,aAAa,KAAK,SAAS;CAE3C,MAAM,EAAE,UAAU,eAAe,aAAa,OAAO;CACrD,MAAM,KAAK,eAAe,UAAU,eAAe;CAEnD,MAAM,0BAAU,IAAI,KAAgB;CACpC,MAAM,SAAS,QAA8B;AAC3C,UAAQ,IAAI,IAAI;AAChB,SAAO;;CAET,MAAM,YAAY,iBAAiB,OAAO;AAC1C,KAAI;EACF,MAAM,YAAY,MAAM,aAAa,QAAQ,WAAW,GAAG,CAAC;EAC5D,MAAM,aAAa,MAAM,aAAa,QAAQ,YAAY,GAAG,CAAC;EAC9D,MAAM,YAAY,MAAM,aAAa,QAAQ,QAAQ,GAAG,CAAC;EACzD,MAAM,UAAU,MAAM,aAAa,QAAQ,SAAS,GAAG,CAAC;EACxD,MAAM,OAAO,MACX,OAAO,aAAa;GAClB,MAAM,IAAI,UAAU;GACpB,OAAO,eAAe,UAAU,eAAe;GAChD,CAAC,CACH;EACD,MAAM,OAAO,MACX,OAAO,aAAa;GAClB,MAAM,IAAI,UAAU;GACpB,OAAO,eAAe,UAAU,eAAe;GAChD,CAAC,CACH;EACD,MAAM,MAAM,MACV,OAAO,aAAa;GAClB,MAAM;GACN,OAAO,eAAe,UAAU,eAAe;GAChD,CAAC,CACH;EACD,MAAM,YAAY,OAAO,gBAAgB;GACvC,QAAQ;GACR,SAAS;IACP;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,WAAW;KAAE;IAC/C;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,YAAY;KAAE;IAChD;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,WAAW;KAAE;IAC/C;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,SAAS;KAAE;IAC7C;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,MAAM;KAAE;IAC1C;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,MAAM;KAAE;IAC1C;KAAE,SAAS;KAAG,UAAU,EAAE,QAAQ,KAAK;KAAE;IAC1C;GACF,CAAC;EAEF,MAAM,SAAS,IAAI,aAAa,IAAI,QAAQ;EAC5C,MAAM,SAAS,OACb,GACA,MACA,QAC0B;AAC1B,kBAAe,OAAO;AACtB,gBAAa,KAAK,kBAAkB;GACpC,MAAM,MAAM,IAAI;AAChB,QAAK,IAAI,IAAI,GAAG,IAAI,KAAK,IAAK,QAAO,KAAK,KAAK,OAAO,EAAE,GAAI;AAC5D,UAAO,MAAM,YAAY,KAAK,GAAG,IAAI,YAAY,CAAC,GAAG,KAAK,CAAC,CAAC;AAC5D,UAAO,MAAM,YAAY,MAAM,GAAG,OAAO,QAAQ,GAAG,MAAM,EAAE;GAC5D,MAAM,MAAM,OAAO,sBAAsB;GACzC,MAAM,OAAO,IAAI,kBAAkB;AACnC,QAAK,YAAY,SAAS;AAC1B,QAAK,aAAa,GAAG,UAAU;AAC/B,QAAK,mBAAmB,KAAK,KAAK,IAAI,IAAI,CAAC;AAC3C,QAAK,KAAK;AACV,UAAO,MAAM,OAAO,CAAC,IAAI,QAAQ,CAAC,CAAC;GACnC,MAAM,IAAI,IAAI,aAAa,MAAM,WAAW,QAAQ,MAAM,MAAM,EAAE,CAAC;AACnE,QAAK,IAAI,IAAI,GAAG,IAAI,KAAK,IAAK,KAAI,KAAK,EAAE;AACzC,UAAO;;EAGT,MAAM,MAAM,MAAM,oBAChB,OACA,UACA,IACA,GACA,MACA,IAAI,KAAK,IAAI,GAAG,GAAG,EACnB,OACD;AACD,QAAM,UAAU,kBAAkB;AAClC,MAAI,CAAC,IAAI,WAAW;AAClB,UAAO,KACL,gOACD;GACD,MAAMC,QAAM,IAAI,aAAa,IAAI,IAAI;AACrC,QAAK,IAAI,IAAI,GAAG,IAAIA,MAAI,QAAQ,IAAK,OAAI,KAAK,IAAI,QAAQ,KAAK,GAAG;AAClE,UAAO;IAAE,WAAW,iBAAiBA,OAAK,GAAG,IAAI;IAAE,kBAAkB;IAAM;;EAG7E,MAAM,MAAM,IAAI,aAAa,IAAI,IAAI;AACrC,OAAK,IAAI,IAAI,GAAG,IAAI,KAAK,KAAK;GAC5B,MAAM,IAAI,IAAI;AACd,QAAK,IAAI,IAAI,GAAG,IAAI,GAAG,IAAK,KAAI,IAAI,MAAM,KAAK,IAAI,QAAQ,IAAI,IAAI;;AAErE,mBAAiB,KAAK,KAAK,IAAI,KAAO;AACtC,SAAO;GAAE,WAAW,iBAAiB,KAAK,GAAG,IAAI;GAAE,kBAAkB;GAAO;WACpE;AAER,QAAM,UAAU,kBAAkB,CAAC,YAAY,GAAG;AAClD,OAAK,MAAM,KAAK,QAAS,GAAE,SAAS;;;;;;;AC9MxC,SAAgB,kBAAkB,MAMqB;CACrD,MAAM,SAAS,KAAK,UAAU;CAC9B,MAAM,QACJ,OAAO,WAAW,YAClB;EACE;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACD,CAAC,SAAS,OAAO,KACjB,KAAK,cAAc,OAAO;CAC7B,MAAM,eAAe,KAAK,gBAAgB;CAC1C,MAAM,YACH,iBAAiB,eAAe,iBAAiB,SAClD,CAAC,KAAK,YACL,KAAK,eAAe,MAAM;CAC7B,MAAM,SAAS,CAAC,QACZ,WAAW,OAAO,OAAO,CAAC,+GAC1B,CAAC,WACC,8DACA;AACN,QAAO;EAAE,KAAK;EAAO,QAAQ;EAAU,GAAI,SAAS,EAAE,QAAQ,GAAG,EAAE;EAAG"}

package/dist/worker.d.ts

@@ -0,0 +1 @@
+export { };