@jax-js/jax 0.1.3 → 0.1.5

package/dist/webgpu-Db2JrNBr.cjs

@@ -0,0 +1,1261 @@
+const require_backend = require('./backend-DziQSaoQ.cjs');
+
+//#region src/backend/webgpu/builtins.ts
+const threefrySrc = `
+fn threefry2x32(key: vec2<u32>, ctr: vec2<u32>) -> vec2<u32> {
+  let ks0: u32 = key.x;
+  let ks1: u32 = key.y;
+  let ks2: u32 = ks0 ^ ks1 ^ 0x1BD11BDAu;
+
+  var x0: u32 = ctr.x + ks0;
+  var x1: u32 = ctr.y + ks1;
+
+  x0 += x1; x1 = (x1 << 13u) | (x1 >> 19u); x1 ^= x0;
+  x0 += x1; x1 = (x1 << 15u) | (x1 >> 17u); x1 ^= x0;
+  x0 += x1; x1 = (x1 << 26u) | (x1 >> 6u); x1 ^= x0;
+  x0 += x1; x1 = (x1 << 6u) | (x1 >> 26u); x1 ^= x0;
+  x0 += ks1;
+  x1 += ks2 + 1u;
+
+  x0 += x1; x1 = (x1 << 17u) | (x1 >> 15u); x1 ^= x0;
+  x0 += x1; x1 = (x1 << 29u) | (x1 >> 3u); x1 ^= x0;
+  x0 += x1; x1 = (x1 << 16u) | (x1 >> 16u); x1 ^= x0;
+  x0 += x1; x1 = (x1 << 24u) | (x1 >> 8u); x1 ^= x0;
+  x0 += ks2;
+  x1 += ks0 + 2u;
+
+  x0 += x1; x1 = (x1 << 13u) | (x1 >> 19u); x1 ^= x0;
+  x0 += x1; x1 = (x1 << 15u) | (x1 >> 17u); x1 ^= x0;
+  x0 += x1; x1 = (x1 << 26u) | (x1 >> 6u); x1 ^= x0;
+  x0 += x1; x1 = (x1 << 6u) | (x1 >> 26u); x1 ^= x0;
+  x0 += ks0;
+  x1 += ks1 + 3u;
+
+  x0 += x1; x1 = (x1 << 17u) | (x1 >> 15u); x1 ^= x0;
+  x0 += x1; x1 = (x1 << 29u) | (x1 >> 3u); x1 ^= x0;
+  x0 += x1; x1 = (x1 << 16u) | (x1 >> 16u); x1 ^= x0;
+  x0 += x1; x1 = (x1 << 24u) | (x1 >> 8u); x1 ^= x0;
+  x0 += ks1;
+  x1 += ks2 + 4u;
+
+  x0 += x1; x1 = (x1 << 13u) | (x1 >> 19u); x1 ^= x0;
+  x0 += x1; x1 = (x1 << 15u) | (x1 >> 17u); x1 ^= x0;
+  x0 += x1; x1 = (x1 << 26u) | (x1 >> 6u); x1 ^= x0;
+  x0 += x1; x1 = (x1 << 6u) | (x1 >> 26u); x1 ^= x0;
+  x0 += ks2;
+  x1 += ks0 + 5u;
+
+  return vec2<u32>(x0, x1);
+}`;
+const erfSrc = `
+const _erf_p: f32 = 0.3275911;
+const _erf_a1: f32 = 0.254829592;
+const _erf_a2: f32 = -0.284496736;
+const _erf_a3: f32 = 1.421413741;
+const _erf_a4: f32 = -1.453152027;
+const _erf_a5: f32 = 1.061405429;
+fn erf(x: f32) -> f32 {
+  let t = 1.0 / (1.0 + _erf_p * abs(x));
+  let P_t = fma(fma(fma(fma(_erf_a5, t, _erf_a4), t, _erf_a3), t, _erf_a2), t, _erf_a1) * t;
+  return sign(x) * (1.0 - P_t * exp(-x * x));
+}
+fn erfc(x: f32) -> f32 {
+  let t = 1.0 / (1.0 + _erf_p * abs(x));
+  let P_t = fma(fma(fma(fma(_erf_a5, t, _erf_a4), t, _erf_a3), t, _erf_a2), t, _erf_a1) * t;
+  let E = P_t * exp(-x * x);
+  return select(2.0 - E, E, x >= 0.0);
+}`;
+
+//#endregion
+//#region src/backend/webgpu/codegen.ts
+const headerWgsl = String.raw`
+fn nan() -> f32 { let bits = 0xffffffffu; return bitcast<f32>(bits); }
+fn inf() -> f32 { let bits = 0x7f800000u; return bitcast<f32>(bits); }
+`.trim();
+function dtypeToWgsl(dtype, storage = false) {
+	switch (dtype) {
+		case require_backend.DType.Bool: return storage ? "i32" : "bool";
+		case require_backend.DType.Int32: return "i32";
+		case require_backend.DType.Uint32: return "u32";
+		case require_backend.DType.Float32: return "f32";
+		case require_backend.DType.Float16: return "f16";
+		default: throw new Error(`Unsupported dtype for WebGPU: ${dtype}`);
+	}
+}
+function maxValueWgsl(dtype) {
+	switch (dtype) {
+		case require_backend.DType.Bool: return "1";
+		case require_backend.DType.Int32: return "2147483647";
+		case require_backend.DType.Uint32: return "4294967295u";
+		case require_backend.DType.Float32: return "inf()";
+		case require_backend.DType.Float16: return "f16(inf())";
+		default: throw new Error(`Unsupported dtype for WebGPU: ${dtype}`);
+	}
+}
+function constToWgsl(dtype, value) {
+	if (dtype === require_backend.DType.Bool) return value ? "true" : "false";
+	if (dtype === require_backend.DType.Int32) return value.toString();
+	if (dtype === require_backend.DType.Uint32) return value.toString() + "u";
+	if (dtype === require_backend.DType.Float32) {
+		if (Number.isNaN(value)) return "nan()";
+		if (!Number.isFinite(value)) return value > 0 ? "inf()" : "-inf()";
+		return "f32(" + value.toString() + ")";
+	}
+	if (dtype === require_backend.DType.Float16) {
+		if (Number.isNaN(value)) return "f16(nan())";
+		if (!Number.isFinite(value)) return value > 0 ? "f16(inf())" : "f16(-inf())";
+		return "f16(" + value.toString() + ")";
+	}
+	throw new Error(`Unsupported const dtype: ${dtype}`);
+}
+const gridOffsetY = 16384;
+function calculateGrid(gridSize) {
+	let gridX = gridSize;
+	let gridY = 1;
+	if (gridSize > 65535) {
+		gridX = gridOffsetY;
+		gridY = Math.ceil(gridSize / gridOffsetY);
+	}
+	return [gridX, gridY];
+}
+
+//#endregion
+//#region src/backend/webgpu/reader.ts
+/**
+* Graphics state used to synchronously read data from WebGPU buffers.
+*
+* This trick is borrowed from TensorFlow.js. Basically, the idea is to create
+* an offscreen canvas with one pixel for every 4 bytes ("device storage"), then
+* configure it with a WebGPU context. Copy the buffer to a texture, then draw
+* the canvas onto another offscreen canvas with '2d' context ("host storage").
+*
+* Once it's on host storage, we can use `getImageData()` to read the pixels
+* from the image directly.
+*
+* We use 256x256 canvases here (256 KiB). The performance of this is bad
+* because it involves multiple data copies, but it still works. We also
+* actually need to copy the image twice: once in "opaque" mode for the RGB
+* values, and once in "premultiplied" mode for the alpha channel.
+*
+* https://github.com/tensorflow/tfjs/blob/tfjs-v4.22.0/tfjs-backend-webgpu/src/backend_webgpu.ts#L379
+*/
+var SyncReader = class SyncReader {
+	static alphaModes = ["opaque", "premultiplied"];
+	static width = 256;
+	static height = 256;
+	initialized = false;
+	deviceStorage;
+	deviceContexts;
+	hostStorage;
+	hostContext;
+	constructor(device) {
+		this.device = device;
+	}
+	#init() {
+		const makeCanvas = () => new OffscreenCanvas(SyncReader.width, SyncReader.height);
+		this.deviceStorage = SyncReader.alphaModes.map(makeCanvas);
+		this.deviceContexts = this.deviceStorage.map((canvas, i) => {
+			const context = canvas.getContext("webgpu");
+			context.configure({
+				device: this.device,
+				format: "bgra8unorm",
+				usage: GPUTextureUsage.COPY_DST,
+				alphaMode: SyncReader.alphaModes[i]
+			});
+			return context;
+		});
+		this.hostStorage = makeCanvas();
+		this.hostContext = this.hostStorage.getContext("2d", { willReadFrequently: true });
+		this.initialized = true;
+	}
+	read(buffer, start, count) {
+		if (!this.initialized) this.#init();
+		const deviceStorage = this.deviceStorage;
+		const deviceContexts = this.deviceContexts;
+		const hostContext = this.hostContext;
+		const pixelsSize = Math.ceil(count / 4);
+		const bytesPerRow = SyncReader.width * 4;
+		const valsGPU = /* @__PURE__ */ new ArrayBuffer(pixelsSize * 4);
+		for (let i = 0; i < deviceContexts.length; i++) {
+			const texture = deviceContexts[i].getCurrentTexture();
+			const readData = (width, height, offset$1) => {
+				const encoder = this.device.createCommandEncoder();
+				encoder.copyBufferToTexture({
+					buffer,
+					bytesPerRow,
+					offset: offset$1 + start
+				}, { texture }, {
+					width,
+					height,
+					depthOrArrayLayers: 1
+				});
+				const commandBuffer = encoder.finish();
+				this.device.queue.submit([commandBuffer]);
+				hostContext.clearRect(0, 0, width, height);
+				hostContext.drawImage(deviceStorage[i], 0, 0);
+				const values = hostContext.getImageData(0, 0, width, height).data;
+				const span = new Uint8ClampedArray(valsGPU, offset$1, 4 * width * height);
+				const alphaMode = SyncReader.alphaModes[i];
+				for (let k = 0; k < span.length; k += 4) if (alphaMode === "premultiplied") span[k + 3] = values[k + 3];
+				else {
+					span[k] = values[k + 2];
+					span[k + 1] = values[k + 1];
+					span[k + 2] = values[k];
+				}
+			};
+			const pixelsPerCanvas = SyncReader.width * SyncReader.height;
+			const wholeChunks = Math.floor(pixelsSize / pixelsPerCanvas);
+			let remainder = pixelsSize % pixelsPerCanvas;
+			const remainderRows = Math.floor(remainder / SyncReader.width);
+			remainder = remainder % SyncReader.width;
+			let offset = 0;
+			for (let j = 0; j < wholeChunks; j++) {
+				readData(SyncReader.width, SyncReader.height, offset);
+				offset += pixelsPerCanvas * 4;
+			}
+			if (remainderRows > 0) {
+				readData(SyncReader.width, remainderRows, offset);
+				offset += remainderRows * SyncReader.width * 4;
+			}
+			if (remainder > 0) readData(remainder, 1, offset);
+		}
+		return new Uint8Array(valsGPU, 0, count);
+	}
+};
+
+//#endregion
+//#region src/backend/webgpu/routines.ts
+function bitonicSortUniform(pass) {
+	const ar = new Uint32Array(3);
+	ar[0] = pass.kind === "sort" ? 0 : 1;
+	ar[1] = pass.mergeStep ?? 0;
+	ar[2] = pass.mergeStage ?? 0;
+	return new Uint8Array(ar.buffer);
+}
+/**
+* Generate a bitonic sort shader.
+*
+* We implement a variant of bitonic sort that [only has forward comparators](
+* <https://sortingalgos.miraheze.org/wiki/Bitonic_Sort#Bitonic_Sort_using_Forward_Comparators>),
+* so we don't need to allocate memory for power-of-two padding.
+*
+* This uses workgroup shared memory up to `2*workgroupSize` elements, for each
+* array in `batches`. For larger arrays, multiple passes are done:
+*
+* - Initial "sort" pass: each workgroup sorts its `2*workgroupSize` elements.
+* - Subsequent "merge" passes: each pass merges sorted sequences of size
+*   `2^(step+1)` with multiple workgroups. This doesn't use shared memory.
+*
+* The total number of passes is roughly `log2(n / workgroupSize)^2 / 2`.
+*/
+function bitonicSortShader(device, dtype, n, batches, outputIndices) {
+	const ty = dtypeToWgsl(dtype, true);
+	const paddedN = 1 << Math.ceil(Math.log2(n || 1));
+	const numThreads = Math.ceil(paddedN / 2);
+	const workgroupSize = require_backend.findPow2(numThreads, device.limits.maxComputeWorkgroupSizeX);
+	const workgroupsPerBatch = numThreads / workgroupSize;
+	const numStages = Math.log2(paddedN);
+	const numLocalStages = Math.min(numStages, Math.log2(workgroupSize * 2));
+	const needsF16 = dtype === require_backend.DType.Float16;
+	const padValue = require_backend.isFloatDtype(dtype) ? `${ty}(nan())` : maxValueWgsl(dtype);
+	const code = `
+${needsF16 ? "enable f16;" : ""}
+${headerWgsl}
+
+struct Uniforms {
+  kind: u32, // 0 = sort, 1 = merge
+  merge_step: u32, // half_block = 2^step
+  merge_stage: u32, // only used for merge
+}
+
+@group(0) @binding(0) var<storage, read> input: array<${ty}>;
+@group(0) @binding(1) var<storage, read_write> output: array<${ty}>;
+${outputIndices ? `@group(0) @binding(2) var<storage, read_write> output_idx: array<i32>;` : ""}
+
+@group(1) @binding(0) var<uniform> uniforms: Uniforms;
+
+var<workgroup> shared_vals: array<${ty}, ${workgroupSize * 2}>;
+${outputIndices ? `var<workgroup> shared_idx: array<i32, ${workgroupSize * 2}>;` : ""}
+
+fn compare(a: ${ty}, b: ${ty}) -> bool {
+${require_backend.isFloatDtype(dtype) ? `
+  let min_value = min(a, b);
+  return a == min_value && b != min_value;` : "  return a < b;"}
+}
+
+fn compare_and_swap(i: u32, j: u32) {
+  let val_i = shared_vals[i];
+  let val_j = shared_vals[j];
+  if (compare(val_j, val_i)) {
+    shared_vals[i] = val_j;
+    shared_vals[j] = val_i;
+${outputIndices ? `
+    let tmp_idx = shared_idx[i];
+    shared_idx[i] = shared_idx[j];
+    shared_idx[j] = tmp_idx;` : ""}
+  }
+}
+
+@compute @workgroup_size(${workgroupSize})
+fn main(
+  @builtin(workgroup_id) wg_id: vec3<u32>,
+  @builtin(local_invocation_id) local_id: vec3<u32>,
+) {
+  let blockid = wg_id.x + wg_id.y * ${gridOffsetY}u;
+  let batch = blockid / ${workgroupsPerBatch}u;
+  let wg_in_batch = blockid % ${workgroupsPerBatch}u;
+
+  let tid = local_id.x;
+  let base = batch * ${n}u;
+
+  if (uniforms.kind == 0u || (uniforms.kind == 1u && uniforms.merge_step == ${numLocalStages - 1}u)) {
+    let wg_base = wg_in_batch * ${workgroupSize * 2}u;
+
+    // Load data into shared memory (2 elements per thread)
+    let idx0 = tid * 2u;
+    let idx1 = tid * 2u + 1u;
+    // Load from input for initial 'sort' pass, then from output (read-write) for 'merge' passes.
+    if (uniforms.kind == 0u) {
+      shared_vals[idx0] = select(${padValue}, input[base + wg_base + idx0], wg_base + idx0 < ${n}u);
+      shared_vals[idx1] = select(${padValue}, input[base + wg_base + idx1], wg_base + idx1 < ${n}u);
+${outputIndices ? `
+      shared_idx[idx0] = i32(wg_base + idx0);
+      shared_idx[idx1] = i32(wg_base + idx1);` : ""}
+    } else {
+      shared_vals[idx0] = select(${padValue}, output[base + wg_base + idx0], wg_base + idx0 < ${n}u);
+      shared_vals[idx1] = select(${padValue}, output[base + wg_base + idx1], wg_base + idx1 < ${n}u);
+${outputIndices ? `
+      shared_idx[idx0] = select(${n}, output_idx[base + wg_base + idx0], wg_base + idx0 < ${n}u);
+      shared_idx[idx1] = select(${n}, output_idx[base + wg_base + idx1], wg_base + idx1 < ${n}u);` : ""}
+    }
+    workgroupBarrier();
+
+    let initial_stage = select(0u, ${numLocalStages - 1}u, uniforms.kind != 0u);
+    for (var stage = initial_stage; stage < ${numLocalStages}u; stage++) {
+      for (var step1 = stage + 1u; step1 > 0u; step1--) {
+        let step = step1 - 1u;
+        let half_block = 1u << step;
+        let is_first_step = uniforms.kind == 0u && step == stage;
+
+        let block_offset = (tid / half_block) * half_block;
+        let local_offset = tid % half_block;
+        let i = block_offset * 2u + local_offset;
+        let j = select(i + half_block, i ^ (half_block * 2u - 1u), is_first_step);
+        compare_and_swap(i, j);
+
+        workgroupBarrier();
+      }
+    }
+
+    if (wg_base + idx0 < ${n}u) {
+      output[base + wg_base + idx0] = shared_vals[idx0];
+      ${outputIndices ? `output_idx[base + wg_base + idx0] = shared_idx[idx0];` : ""}
+    }
+    if (wg_base + idx1 < ${n}u) {
+      output[base + wg_base + idx1] = shared_vals[idx1];
+      ${outputIndices ? `output_idx[base + wg_base + idx1] = shared_idx[idx1];` : ""}
+    }
+  } else {
+    // Execute single merge pass for a step >= numLocalStages.
+    let half_block = 1u << uniforms.merge_step;  // half_block >= workgroupSize * 2
+    let thread_in_batch = wg_in_batch * ${workgroupSize} + tid;
+    let is_first_step = uniforms.merge_step == uniforms.merge_stage;
+
+    let block_offset = (thread_in_batch / half_block) * half_block;
+    let local_offset = thread_in_batch % half_block;
+    let i = block_offset * 2u + local_offset;
+    let j = select(i + half_block, i ^ (half_block * 2u - 1u), is_first_step);
+
+    // Global version of compare_and_swap()
+    if (j < ${n}u) {
+      let val_i = output[base + i];
+      let val_j = output[base + j];
+      if (compare(val_j, val_i)) {
+        output[base + i] = val_j;
+        output[base + j] = val_i;
+${outputIndices ? `
+        let tmp_idx = output_idx[base + i];
+        output_idx[base + i] = output_idx[base + j];
+        output_idx[base + j] = tmp_idx;` : ""}
+      }
+    }
+  }
+}
+`.trim();
+	const grid = calculateGrid(batches * workgroupsPerBatch);
+	const passes = [{ kind: "sort" }];
+	for (let mergeStage = numLocalStages; mergeStage < numStages; mergeStage++) for (let mergeStep = mergeStage; mergeStep >= numLocalStages - 1; mergeStep--) passes.push({
+		kind: "merge",
+		mergeStep,
+		mergeStage
+	});
+	return [{
+		code,
+		numInputs: 1,
+		numOutputs: outputIndices ? 2 : 1,
+		hasUniform: true,
+		passes: passes.map((pass) => ({
+			grid,
+			uniform: bitonicSortUniform(pass)
+		}))
+	}];
+}
+function createSort(device, type) {
+	const dtype = type.inputDtypes[0];
+	const shape = type.inputShapes[0];
+	const n = shape[shape.length - 1];
+	const batches = require_backend.prod(shape.slice(0, -1));
+	return bitonicSortShader(device, dtype, n, batches, false);
+}
+function createArgsort(device, type) {
+	const dtype = type.inputDtypes[0];
+	const shape = type.inputShapes[0];
+	const n = shape[shape.length - 1];
+	const batches = require_backend.prod(shape.slice(0, -1));
+	return bitonicSortShader(device, dtype, n, batches, true);
+}
+/**
+* Generate a triangular solve shader.
+*
+* Solves A @ X.T = B.T for X, where A is upper-triangular.
+* Uses a parallelized back-substitution:
+*   1. Copy b to x
+*   2. For j = n-1 down to 0:
+*      - Divide x[j] by a[j,j] (single thread)
+*      - All threads subtract x[j] * a[i,j] from x[i] for i < j in parallel
+*/
+function createTriangularSolve(device, type, params) {
+	const dtype = type.inputDtypes[0];
+	const aShape = type.inputShapes[0];
+	const bShape = type.inputShapes[1];
+	const n = aShape[aShape.length - 1];
+	const numRhs = bShape[bShape.length - 2];
+	const numMatrices = require_backend.prod(aShape.slice(0, -2));
+	const needsF16 = dtype === require_backend.DType.Float16;
+	const ty = dtypeToWgsl(dtype, true);
+	const workgroupSize = require_backend.findPow2(n, device.limits.maxComputeWorkgroupSizeX);
+	const code = `
+${needsF16 ? "enable f16;" : ""}
+${headerWgsl}
+
+@group(0) @binding(0) var<storage, read> a: array<${ty}>;
+@group(0) @binding(1) var<storage, read> b: array<${ty}>;
+@group(0) @binding(2) var<storage, read_write> x: array<${ty}>;
+
+// Shared memory for the current pivot value x[j]
+var<workgroup> x_j: ${ty};
+
+@compute @workgroup_size(${workgroupSize})
+fn main(
+  @builtin(workgroup_id) wg_id: vec3<u32>,
+  @builtin(local_invocation_id) local_id: vec3<u32>,
+) {
+  let wg_idx = wg_id.x + wg_id.y * ${gridOffsetY}u;
+  let mat_idx = wg_idx / ${numRhs}u;
+  let rhs_idx = wg_idx % ${numRhs}u;
+
+  if (mat_idx >= ${numMatrices}u) {
+    return;
+  }
+
+  let a_base = mat_idx * ${n * n}u;
+  let bx_base = (mat_idx * ${numRhs}u + rhs_idx) * ${n}u;
+  let tid = local_id.x;
+
+  // Step 1: Copy b to x (threads collaborate)
+  for (var idx = tid; idx < ${n}u; idx += ${workgroupSize}u) {
+    x[bx_base + idx] = b[bx_base + idx];
+  }
+  storageBarrier();
+
+  // Step 2: Back-substitution from j = n-1 down to 0
+  for (var jj = 0u; jj < ${n}u; jj++) {
+    let j = ${n - 1}u - jj;
+
+    // Thread 0 computes x[j] = x[j] / a[j,j]
+    if (tid == 0u) {
+      ${params.unitDiagonal ? `x_j = x[bx_base + j];` : `x_j = x[bx_base + j] / a[a_base + j * ${n}u + j];`}
+      x[bx_base + j] = x_j;
+    }
+    workgroupBarrier();  // Sync shared memory x_j
+
+    // All threads subtract x[j] * a[i,j] from x[i] for i < j
+    for (var i = tid; i < j; i += ${workgroupSize}u) {
+      x[bx_base + i] -= x_j * a[a_base + i * ${n}u + j];
+    }
+    workgroupBarrier();
+    storageBarrier();
+  }
+}
+`.trim();
+	const totalWorkgroups = numMatrices * numRhs;
+	const grid = calculateGrid(totalWorkgroups);
+	return [{
+		code,
+		numInputs: 2,
+		numOutputs: 1,
+		hasUniform: false,
+		passes: [{ grid }]
+	}];
+}
+/**
+* Generate a Cholesky decomposition shader.
+*
+* Computes the lower triangular matrix L such that A = L * L^T for each
+* positive semi-definite matrix in the batch. Uses the Cholesky-Crout
+* algorithm which processes column-by-column.
+*
+* For each column j:
+*   1. All threads compute their row's sum in parallel and store to output
+*   2. Thread 0 computes L[j][j] = sqrt(output[j][j]) and stores to shared memory
+*   3. All threads divide their output[i][j] by L[j][j] in parallel
+*/
+function createCholesky(device, type) {
+	const dtype = type.inputDtypes[0];
+	const shape = type.inputShapes[0];
+	const n = shape[shape.length - 1];
+	const batches = require_backend.prod(shape.slice(0, -2));
+	const needsF16 = dtype === require_backend.DType.Float16;
+	const ty = dtypeToWgsl(dtype, true);
+	const workgroupSize = require_backend.findPow2(n, device.limits.maxComputeWorkgroupSizeX);
+	const code = `
+${needsF16 ? "enable f16;" : ""}
+${headerWgsl}
+
+@group(0) @binding(0) var<storage, read> input: array<${ty}>;
+@group(0) @binding(1) var<storage, read_write> output: array<${ty}>;
+
+// Shared memory for the diagonal element
+var<workgroup> L_jj: ${ty};
+
+@compute @workgroup_size(${workgroupSize})
+fn main(
+  @builtin(workgroup_id) wg_id: vec3<u32>,
+  @builtin(local_invocation_id) local_id: vec3<u32>,
+) {
+  let batch = wg_id.x + wg_id.y * ${gridOffsetY}u;
+  if (batch >= ${batches}u) {
+    return;
+  }
+
+  let base = batch * ${n * n}u;
+  let tid = local_id.x;
+
+  // Zero out output and copy lower triangle from input (threads collaborate)
+  for (var idx = tid; idx < ${n * n}u; idx += ${workgroupSize}u) {
+    let row = idx / ${n}u;
+    let col = idx % ${n}u;
+    output[base + idx] = select(0, input[base + idx], col <= row);
+  }
+  storageBarrier();
+
+  // Cholesky-Crout algorithm: process column by column
+  for (var j = 0u; j < ${n}u; j++) {
+    // Step 1: All threads compute sum for their rows i >= j in parallel
+    // sum = A[i][j] - sum(L[i][k] * L[j][k] for k < j)
+    for (var i = j + tid; i < ${n}u; i += ${workgroupSize}u) {
+      var sum = output[base + i * ${n}u + j];
+      for (var k = 0u; k < j; k++) {
+        sum -= output[base + i * ${n}u + k] * output[base + j * ${n}u + k];
+      }
+      output[base + i * ${n}u + j] = sum;
+    }
+    storageBarrier();
+
+    // Step 2: Thread 0 computes L[j][j] = sqrt(output[j][j])
+    if (tid == 0u) {
+      L_jj = sqrt(output[base + j * ${n}u + j]);
+      output[base + j * ${n}u + j] = L_jj;
+    }
+    workgroupBarrier();
+
+    // Step 3: All threads divide output[i][j] by L[j][j] for i > j
+    for (var i = j + 1u + tid; i < ${n}u; i += ${workgroupSize}u) {
+      output[base + i * ${n}u + j] /= L_jj;
+    }
+    storageBarrier();
+  }
+}
+`.trim();
+	const grid = calculateGrid(batches);
+	return [{
+		code,
+		numInputs: 1,
+		numOutputs: 1,
+		hasUniform: false,
+		passes: [{ grid }]
+	}];
+}
+/**
+* Generate an LU decomposition shader with partial pivoting.
+*
+* Computes PA = LU where P is a permutation matrix, L is lower triangular
+* with unit diagonal, and U is upper triangular.
+*
+* For each column j:
+*   1. Find pivot row (max absolute value in column j, rows >= j)
+*   2. Swap rows j and pivot row
+*   3. Compute L[i][j] = A[i][j] / A[j][j] for i > j
+*   4. Update submatrix: A[i][k] -= L[i][j] * A[j][k] for i > j, k > j
+*/
+function createLU(device, type) {
+	const dtype = type.inputDtypes[0];
+	const shape = type.inputShapes[0];
+	const m = shape[shape.length - 2];
+	const n = shape[shape.length - 1];
+	const r = Math.min(m, n);
+	const batches = require_backend.prod(shape.slice(0, -2));
+	const needsF16 = dtype === require_backend.DType.Float16;
+	const ty = dtypeToWgsl(dtype, true);
+	const workgroupSize = require_backend.findPow2(Math.max(m, n), device.limits.maxComputeWorkgroupSizeX);
+	const code = `
+${needsF16 ? "enable f16;" : ""}
+${headerWgsl}
+
+@group(0) @binding(0) var<storage, read> input: array<${ty}>;
+@group(0) @binding(1) var<storage, read_write> lu: array<${ty}>;
+@group(0) @binding(2) var<storage, read_write> pivots: array<i32>;
+@group(0) @binding(3) var<storage, read_write> perm: array<i32>;
+
+var<workgroup> pivot_row: u32;
+var<workgroup> pivot_val: ${ty};
+
+@compute @workgroup_size(${workgroupSize})
+fn main(
+  @builtin(workgroup_id) wg_id: vec3<u32>,
+  @builtin(local_invocation_id) local_id: vec3<u32>,
+) {
+  let batch = wg_id.x + wg_id.y * ${gridOffsetY}u;
+  if (batch >= ${batches}u) {
+    return;
+  }
+
+  let lu_base = batch * ${m * n}u;
+  let piv_base = batch * ${r}u;
+  let perm_base = batch * ${m}u;
+  let tid = local_id.x;
+
+  // Copy input to lu
+  for (var idx = tid; idx < ${m * n}u; idx += ${workgroupSize}u) {
+    lu[lu_base + idx] = input[lu_base + idx];
+  }
+  // Initialize permutation
+  for (var idx = tid; idx < ${m}u; idx += ${workgroupSize}u) {
+    perm[perm_base + idx] = i32(idx);
+  }
+  storageBarrier();
+
+  // LU decomposition with partial pivoting
+  for (var j = 0u; j < ${r}u; j++) {
+    // Step 1: Thread 0 finds pivot (max abs value in column j, rows >= j)
+    if (tid == 0u) {
+      var max_val = abs(lu[lu_base + j * ${n}u + j]);
+      var max_row = j;
+      for (var i = j + 1u; i < ${m}u; i++) {
+        let val = abs(lu[lu_base + i * ${n}u + j]);
+        if (val > max_val) {
+          max_val = val;
+          max_row = i;
+        }
+      }
+      pivot_row = max_row;
+      pivot_val = lu[lu_base + max_row * ${n}u + j];
+      pivots[piv_base + j] = i32(max_row);
+    }
+    workgroupBarrier();
+
+    // Step 2: Swap rows j and pivot_row (threads collaborate)
+    let pr = pivot_row;
+    if (pr != j) {
+      for (var col = tid; col < ${n}u; col += ${workgroupSize}u) {
+        let tmp = lu[lu_base + j * ${n}u + col];
+        lu[lu_base + j * ${n}u + col] = lu[lu_base + pr * ${n}u + col];
+        lu[lu_base + pr * ${n}u + col] = tmp;
+      }
+      if (tid == 0u) {
+        let tmp_p = perm[perm_base + j];
+        perm[perm_base + j] = perm[perm_base + pr];
+        perm[perm_base + pr] = tmp_p;
+      }
+    }
+    storageBarrier();
+
+    // Step 3: Compute L[i][j] and update submatrix
+    // Each thread handles one row i > j
+    for (var i = j + 1u + tid; i < ${m}u; i += ${workgroupSize}u) {
+      let factor = lu[lu_base + i * ${n}u + j] / pivot_val;
+      lu[lu_base + i * ${n}u + j] = factor; // L[i][j]
+      for (var k = j + 1u; k < ${n}u; k++) {
+        lu[lu_base + i * ${n}u + k] -= factor * lu[lu_base + j * ${n}u + k];
+      }
+    }
+    storageBarrier();
+  }
+}
+`.trim();
+	const grid = calculateGrid(batches);
+	return [{
+		code,
+		numInputs: 1,
+		numOutputs: 3,
+		hasUniform: false,
+		passes: [{ grid }]
+	}];
+}
+function createRoutineShader(device, routine) {
+	switch (routine.name) {
+		case require_backend.Routines.Sort: return createSort(device, routine.type);
+		case require_backend.Routines.Argsort: return createArgsort(device, routine.type);
+		case require_backend.Routines.TriangularSolve: return createTriangularSolve(device, routine.type, routine.params);
+		case require_backend.Routines.Cholesky: return createCholesky(device, routine.type);
+		case require_backend.Routines.LU: return createLU(device, routine.type);
+		default: throw new require_backend.UnsupportedRoutineError(routine.name, "webgpu");
+	}
+}
+
+//#endregion
+//#region src/backend/webgpu.ts
+/** Implementation of `Backend` that uses WebGPU in browsers. */
+var WebGPUBackend = class {
+	type = "webgpu";
+	maxArgs;
+	pipelines;
+	syncReader;
+	buffers;
+	nextSlot;
+	#cachedShaderMap = /* @__PURE__ */ new Map();
+	#reusableZsb;
+	constructor(device) {
+		this.device = device;
+		if (require_backend.DEBUG >= 3 && device.adapterInfo) console.info("webgpu adapter:", device.adapterInfo.vendor, device.adapterInfo.architecture);
+		this.maxArgs = this.device.limits.maxStorageBuffersPerShaderStage - 1;
+		this.pipelines = new ShaderPipelineCache(device);
+		this.syncReader = new SyncReader(device);
+		this.buffers = /* @__PURE__ */ new Map();
+		this.nextSlot = 1;
+		this.#reusableZsb = this.#createBuffer(4);
+		device.addEventListener("uncapturederror", (event) => {
+			console.error("Uncaptured error in WebGPU backend:", event.error.message);
+		});
+	}
+	malloc(size, initialData) {
+		let buffer;
+		const paddedSize = Math.ceil(size / 4) * 4;
+		if (size === 0) buffer = this.#reusableZsb;
+		else if (initialData) {
+			if (initialData.byteLength !== size) throw new Error("initialData size does not match buffer size");
+			if (initialData.byteLength < 4096) {
+				buffer = this.#createBuffer(paddedSize, { mapped: true });
+				new Uint8Array(buffer.getMappedRange(), 0, size).set(initialData);
+				buffer.unmap();
+			} else {
+				buffer = this.#createBuffer(paddedSize);
+				if (initialData.byteLength % 4 === 0) this.device.queue.writeBuffer(buffer, 0, initialData);
+				else {
+					const aligned = initialData.byteLength - initialData.byteLength % 4;
+					this.device.queue.writeBuffer(buffer, 0, initialData, 0, aligned);
+					const remainder = new Uint8Array(4);
+					remainder.set(initialData.subarray(aligned));
+					this.device.queue.writeBuffer(buffer, aligned, remainder);
+				}
+			}
+		} else buffer = this.#createBuffer(paddedSize);
+		const slot = this.nextSlot++;
+		this.buffers.set(slot, {
+			buffer,
+			size,
+			ref: 1
+		});
+		return slot;
+	}
+	incRef(slot) {
+		const buffer = this.buffers.get(slot);
+		if (!buffer) throw new require_backend.SlotError(slot);
+		buffer.ref++;
+	}
+	decRef(slot) {
+		const buffer = this.buffers.get(slot);
+		if (!buffer) throw new require_backend.SlotError(slot);
+		buffer.ref--;
+		if (buffer.ref === 0) {
+			this.buffers.delete(slot);
+			if (buffer.buffer !== this.#reusableZsb) buffer.buffer.destroy();
+		}
+	}
+	async read(slot, start, count) {
+		const { buffer, size } = this.#getBuffer(slot);
+		if (buffer === this.#reusableZsb) return new Uint8Array();
+		if (start === void 0) start = 0;
+		if (count === void 0) count = size - start;
+		const paddedSize = Math.ceil(count / 4) * 4;
+		const staging = this.#createBuffer(paddedSize, { read: true });
+		try {
+			const commandEncoder = this.device.createCommandEncoder();
+			commandEncoder.copyBufferToBuffer(buffer, start, staging, 0, paddedSize);
+			this.device.queue.submit([commandEncoder.finish()]);
+			await staging.mapAsync(GPUMapMode.READ);
+			const arrayBuffer = staging.getMappedRange();
+			return new Uint8Array(arrayBuffer.slice(), 0, count);
+		} finally {
+			staging.destroy();
+		}
+	}
+	readSync(slot, start, count) {
+		const { buffer, size } = this.#getBuffer(slot);
+		if (buffer === this.#reusableZsb) return new Uint8Array();
+		if (start === void 0) start = 0;
+		if (count === void 0) count = size - start;
+		return this.syncReader.read(buffer, start, count);
+	}
+	#cachedShader(kernel) {
+		const cacheKey = require_backend.FpHash.hash(kernel);
+		let result = this.#cachedShaderMap.get(cacheKey);
+		if (!result) {
+			result = pipelineSource(this.device, kernel);
+			this.#cachedShaderMap.set(cacheKey, result);
+		}
+		return result;
+	}
+	async prepareKernel(kernel) {
+		const shader = this.#cachedShader(kernel);
+		const pipeline = await this.pipelines.prepare(shader);
+		return new require_backend.Executable(kernel, [{
+			...shader,
+			pipeline
+		}]);
+	}
+	prepareKernelSync(kernel) {
+		const shader = this.#cachedShader(kernel);
+		const pipeline = this.pipelines.prepareSync(shader);
+		return new require_backend.Executable(kernel, [{
+			...shader,
+			pipeline
+		}]);
+	}
+	async prepareRoutine(routine) {
+		const shaders = createRoutineShader(this.device, routine);
+		const dispatches = await Promise.all(shaders.map(async (shader) => {
+			const pipeline = await this.pipelines.prepare(shader);
+			return {
+				...shader,
+				pipeline
+			};
+		}));
+		return new require_backend.Executable(routine, dispatches);
+	}
+	prepareRoutineSync(routine) {
+		const shaders = createRoutineShader(this.device, routine);
+		const dispatches = shaders.map((shader) => {
+			const pipeline = this.pipelines.prepareSync(shader);
+			return {
+				...shader,
+				pipeline
+			};
+		});
+		return new require_backend.Executable(routine, dispatches);
+	}
+	dispatch(exe, inputs, outputs) {
+		const inputBuffers = inputs.map((slot) => this.#getBuffer(slot).buffer);
+		const outputBuffers = outputs.map((slot) => this.#getBuffer(slot).buffer);
+		pipelineSubmit(this.device, exe.data, inputBuffers, outputBuffers);
+	}
+	#getBuffer(slot) {
+		const buffer = this.buffers.get(slot);
+		if (!buffer) throw new require_backend.SlotError(slot);
+		return {
+			buffer: buffer.buffer,
+			size: buffer.size
+		};
+	}
+	/**
+	* Create a GPU buffer.
+	*
+	* By default, this creates a general-purpose buffer with the given size.
+	*
+	* - If `mapped` is true, initialize the buffer in mapped mode so that it can
+	*   be populated with data from the CPU. (Call `.unmap()` later.)
+	* - If `read` is true, create a staging buffer for returning data to CPU.
+	*   (Call `.mapAsync()` later.)
+	*/
+	#createBuffer(size, { mapped = false, read = false } = {}) {
+		if (read && mapped) throw new Error("mapped and read cannot both be true");
+		const buffer = this.device.createBuffer({
+			size,
+			usage: read ? GPUBufferUsage.MAP_READ | GPUBufferUsage.COPY_DST : GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC | GPUBufferUsage.COPY_DST,
+			mappedAtCreation: mapped
+		});
+		return buffer;
+	}
+};
+/**
+* Compiles an expression into WebGPU shader source code.
+*
+* Returns the shader source and the number of workgroups to dispatch along x
+* and y axes, to run the kernel.
+*/
+function pipelineSource(device, kernel) {
+	const tune = require_backend.tuneWebgpu(kernel);
+	if (require_backend.DEBUG >= 3) console.info(`kernel.exp: ${kernel.exp}\ntune.exp: ${tune.exp}`);
+	const { nargs, reduction: re } = kernel;
+	const args = Array.from({ length: nargs }, (_, i) => `in${i}`);
+	const shader = [];
+	let indent = "";
+	const pushIndent = Symbol("pushIndent");
+	const popIndent = Symbol("popIndent");
+	const emit = (...lines) => {
+		for (const line of lines) if (line === pushIndent) indent += "  ";
+		else if (line === popIndent) indent = indent.slice(0, -2);
+		else shader.push(line ? indent + line : line);
+	};
+	if (tune.exp.some((exp) => exp.dtype === require_backend.DType.Float16) || tune.epilogue?.some((exp) => exp.dtype === require_backend.DType.Float16)) {
+		if (!device.features.has("shader-f16")) throw new Error("WebGPU device does not support shader-f16 feature");
+		emit("enable f16;");
+	}
+	emit(headerWgsl);
+	const distinctOps = require_backend.mapSetUnion(tune.exp.distinctOps(), tune.epilogue?.distinctOps());
+	if (distinctOps.has(require_backend.AluOp.Threefry2x32)) emit(threefrySrc);
+	if (distinctOps.has(require_backend.AluOp.Erf) || distinctOps.has(require_backend.AluOp.Erfc)) emit(erfSrc);
+	emit("");
+	const usedArgs = Array.from({ length: nargs }, () => null);
+	tune.exp.fold((exp) => {
+		if (exp.op === require_backend.AluOp.GlobalIndex) usedArgs[exp.arg[0]] = exp.dtype;
+	});
+	tune.epilogue?.fold((exp) => {
+		if (exp.op === require_backend.AluOp.GlobalIndex) usedArgs[exp.arg[0]] = exp.dtype;
+	});
+	for (let i = 0; i < nargs; i++) {
+		const ty = dtypeToWgsl(usedArgs[i] ?? require_backend.DType.Float32, true);
+		emit(`@group(0) @binding(${i}) var<storage, read> ${args[i]} : array<${ty}>;`);
+	}
+	const resultTy = dtypeToWgsl(kernel.dtype, true);
+	emit(`@group(0) @binding(${nargs}) var<storage, read_write> result : array<${resultTy}>;`);
+	const workgroupSize = require_backend.findPow2(tune.threadCount, 256);
+	const gridSize = Math.ceil(tune.threadCount / workgroupSize);
+	const [gridX, gridY] = calculateGrid(gridSize);
+	emit("", `@compute @workgroup_size(${workgroupSize})`, "fn main(@builtin(global_invocation_id) id : vec3<u32>) {", pushIndent);
+	if (gridY === 1) emit(`if (id.x >= ${tune.threadCount}) { return; }`, "let gidx: i32 = i32(id.x);");
+	else {
+		const sizeX = gridX * workgroupSize;
+		emit(`if (${sizeX} * id.y + id.x >= ${tune.threadCount}) { return; }`, `let gidx: i32 = i32(${sizeX} * id.y + id.x);`);
+	}
+	let gensymCount = 0;
+	const gensym = () => `alu${gensymCount++}`;
+	const isGensym = (text) => text.match(/^alu[0-9]+$/);
+	if (args.length > 0) emit(args.map((arg) => `_ = &${arg};`).join(" "));
+	const references = /* @__PURE__ */ new Map();
+	const seen = /* @__PURE__ */ new Set();
+	const countReferences = (exp) => {
+		references.set(exp, (references.get(exp) ?? 0) + 1);
+		if (!seen.has(exp)) {
+			seen.add(exp);
+			for (const src of exp.src) countReferences(src);
+		}
+	};
+	const expContext = /* @__PURE__ */ new Map();
+	const gen = (exp) => {
+		if (expContext.has(exp)) return expContext.get(exp);
+		const { op, src, dtype, arg } = exp;
+		let source = "";
+		if (require_backend.AluGroup.Binary.has(op) || require_backend.AluGroup.Compare.has(op)) {
+			const a = gen(src[0]);
+			const b = gen(src[1]);
+			if (op === require_backend.AluOp.Add) if (dtype === require_backend.DType.Bool) source = `(${a} || ${b})`;
+			else source = `(${a} + ${b})`;
+			else if (op === require_backend.AluOp.Sub) source = `(${a} - ${b})`;
+			else if (op === require_backend.AluOp.Mul) if (dtype === require_backend.DType.Bool) source = `(${a} && ${b})`;
+			else source = `(${a} * ${b})`;
+			else if (op === require_backend.AluOp.Idiv) source = require_backend.isFloatDtype(dtype) ? `trunc(${a} / ${b})` : `(${a} / ${b})`;
+			else if (op === require_backend.AluOp.Mod) source = `(${a} % ${b})`;
+			else if (op === require_backend.AluOp.Min) if (dtype === require_backend.DType.Bool) source = `(${a} && ${b})`;
+			else source = `min(${require_backend.strip1(a)}, ${require_backend.strip1(b)})`;
+			else if (op === require_backend.AluOp.Max) if (dtype === require_backend.DType.Bool) source = `(${a} || ${b})`;
+			else source = `max(${require_backend.strip1(a)}, ${require_backend.strip1(b)})`;
+			else if (op === require_backend.AluOp.Cmplt) source = `(${a} < ${b})`;
+			else if (op === require_backend.AluOp.Cmpne) if (require_backend.isFloatDtype(src[0].dtype)) {
+				const x = isGensym(a) ? a : gensym();
+				if (x !== a) emit(`let ${x} = ${a};`);
+				source = `(${x} != ${b} || min(${x}, ${dtypeToWgsl(src[0].dtype)}(inf())) != ${x})`;
+			} else source = `(${a} != ${b})`;
+		} else if (require_backend.AluGroup.Unary.has(op)) if (op === require_backend.AluOp.Reciprocal && src[0].op === require_backend.AluOp.Sqrt) {
+			const a = gen(src[0].src[0]);
+			source = `inverseSqrt(${a})`;
+		} else {
+			const a = gen(src[0]);
+			if (op === require_backend.AluOp.Sin) source = `sin(${require_backend.strip1(a)})`;
+			else if (op === require_backend.AluOp.Cos) source = `cos(${require_backend.strip1(a)})`;
+			else if (op === require_backend.AluOp.Asin) source = `asin(${require_backend.strip1(a)})`;
+			else if (op === require_backend.AluOp.Atan) source = `atan(${require_backend.strip1(a)})`;
+			else if (op === require_backend.AluOp.Exp) source = `exp(${require_backend.strip1(a)})`;
+			else if (op === require_backend.AluOp.Log) source = `log(${require_backend.strip1(a)})`;
+			else if (op === require_backend.AluOp.Erf || op === require_backend.AluOp.Erfc) {
+				const funcName = op === require_backend.AluOp.Erf ? "erf" : "erfc";
+				if (dtype !== require_backend.DType.Float32) source = `${dtypeToWgsl(dtype)}(${funcName}(f32(${require_backend.strip1(a)})))`;
+				else source = `${funcName}(${require_backend.strip1(a)})`;
+			} else if (op === require_backend.AluOp.Sqrt) source = `sqrt(${require_backend.strip1(a)})`;
+			else if (op === require_backend.AluOp.Reciprocal) source = `(1.0 / ${a})`;
+			else if (op === require_backend.AluOp.Floor) source = `floor(${require_backend.strip1(a)})`;
+			else if (op === require_backend.AluOp.Ceil) source = `ceil(${require_backend.strip1(a)})`;
+			else if (op === require_backend.AluOp.Cast) source = `${dtypeToWgsl(dtype)}(${require_backend.strip1(a)})`;
+			else if (op === require_backend.AluOp.Bitcast) source = `bitcast<${dtypeToWgsl(dtype)}>(${require_backend.strip1(a)})`;
+		}
+		else if (op === require_backend.AluOp.Where) source = `select(${require_backend.strip1(gen(src[2]))}, ${require_backend.strip1(gen(src[1]))}, ${require_backend.strip1(gen(src[0]))})`;
+		else if (op === require_backend.AluOp.Threefry2x32) {
+			const x = gensym();
+			const [k0, k1, c0, c1] = src.map((x$1) => require_backend.strip1(gen(x$1)));
+			emit(`let ${x} = threefry2x32(vec2(${k0}, ${k1}), vec2(${c0}, ${c1}));`);
+			if (arg === "xor") source = `(${x}.x ^ ${x}.y)`;
+			else if (arg === 0) source = `${x}.x`;
+			else if (arg === 1) source = `${x}.y`;
+			else throw new require_backend.UnsupportedOpError(op, dtype, "webgpu", arg);
+		} else if (op === require_backend.AluOp.Const) return constToWgsl(dtype, arg);
+		else if (op === require_backend.AluOp.Special) return arg[0];
+		else if (op === require_backend.AluOp.Variable) return arg;
+		else if (op === require_backend.AluOp.GlobalIndex) {
+			source = `${args[arg[0]]}[${require_backend.strip1(gen(src[0]))}]`;
+			if (dtype === require_backend.DType.Bool) source = `(${source} != 0)`;
+		}
+		if (!source) throw new require_backend.UnsupportedOpError(op, dtype, "webgpu", arg);
+		const typeName = dtypeToWgsl(dtype);
+		if ((references.get(exp) ?? 0) > 1) {
+			const name = gensym();
+			expContext.set(exp, name);
+			emit(`let ${name}: ${typeName} = ${require_backend.strip1(source)};`);
+			return name;
+		} else {
+			expContext.set(exp, source);
+			return source;
+		}
+	};
+	if (!re) {
+		countReferences(tune.exp);
+		let rhs = require_backend.strip1(gen(tune.exp));
+		if (resultTy !== dtypeToWgsl(tune.exp.dtype)) rhs = `${resultTy}(${rhs})`;
+		emit(`result[gidx] = ${rhs};`);
+	} else {
+		if ((tune.size.groups ?? 1) > 1) throw new Error("WebGPU backend does not support group optimization yet");
+		const unroll = tune.size.unroll ?? 1;
+		const upcast = tune.size.upcast ?? 1;
+		const acc = [...Array(upcast)].map((_, i) => `acc${i}`);
+		for (let i = 0; i < upcast; i++) emit(`var ${acc[i]}: ${dtypeToWgsl(re.dtype)} = ${constToWgsl(re.dtype, re.identity)};`);
+		emit(`for (var ridx: i32 = 0; ridx < ${tune.size.reduce}; ridx++) {`, pushIndent);
+		const exps = [];
+		const cache = /* @__PURE__ */ new Map();
+		for (let up = 0; up < upcast; up++) {
+			exps.push([]);
+			for (let un = 0; un < unroll; un++) {
+				const exp = tune.exp.substitute({
+					upcast: require_backend.AluExp.i32(up),
+					unroll: require_backend.AluExp.i32(un)
+				});
+				exps[up].push(exp.simplify(cache));
+				countReferences(exps[up][un]);
+			}
+		}
+		const items = exps.map((ar) => ar.map(gen).map(require_backend.strip1));
+		for (let i = 0; i < upcast; i++) {
+			let rhs = items[i][0];
+			for (let j = 1; j < unroll; j++) if (re.op === require_backend.AluOp.Add) rhs = `${rhs} + ${items[i][j]}`;
+			else if (re.op === require_backend.AluOp.Mul) rhs = `${rhs} * ${items[i][j]}`;
+			else if (re.op === require_backend.AluOp.Min) rhs = re.dtype === require_backend.DType.Bool ? `(${rhs} && ${items[i][j]})` : `min(${rhs}, ${items[i][j]})`;
+			else if (re.op === require_backend.AluOp.Max) rhs = re.dtype === require_backend.DType.Bool ? `(${rhs} || ${items[i][j]})` : `max(${rhs}, ${items[i][j]})`;
+			else throw new Error(`Unsupported reduction op: ${re.op}`);
+			if (re.op === require_backend.AluOp.Add) emit(`${acc[i]} += ${rhs};`);
+			else if (re.op === require_backend.AluOp.Mul) emit(`${acc[i]} *= ${rhs};`);
+			else if (re.op === require_backend.AluOp.Min) if (re.dtype === require_backend.DType.Bool) emit(`${acc[i]} = ${acc[i]} && ${rhs};`);
+			else emit(`${acc[i]} = min(${acc[i]}, ${rhs});`);
+			else if (re.op === require_backend.AluOp.Max) if (re.dtype === require_backend.DType.Bool) emit(`${acc[i]} = ${acc[i]} || ${rhs};`);
+			else emit(`${acc[i]} = max(${acc[i]}, ${rhs});`);
+			else throw new Error(`Unsupported reduction op: ${re.op}`);
+		}
+		emit(popIndent, "}");
+		expContext.clear();
+		references.clear();
+		seen.clear();
+		const outputIdxExps = [];
+		const fusionExps = [];
+		for (let i = 0; i < upcast; i++) {
+			const exp = tune.outputIdxExp.substitute({ upcast: require_backend.AluExp.i32(i) });
+			outputIdxExps.push(exp.simplify(cache));
+			countReferences(outputIdxExps[i]);
+			fusionExps.push(tune.epilogue.substitute({
+				acc: require_backend.AluExp.variable(re.dtype, acc[i]),
+				upcast: require_backend.AluExp.i32(i)
+			}).simplify(cache));
+			countReferences(fusionExps[i]);
+		}
+		for (let i = 0; i < upcast; i++) {
+			const index = require_backend.strip1(gen(outputIdxExps[i]));
+			let rhs = require_backend.strip1(gen(fusionExps[i]));
+			if (resultTy !== dtypeToWgsl(fusionExps[i].dtype)) rhs = `${resultTy}(${rhs})`;
+			emit(`result[${index}] = ${rhs};`);
+		}
+	}
+	emit(popIndent, "}");
+	return {
+		code: shader.join("\n"),
+		numInputs: nargs,
+		numOutputs: 1,
+		hasUniform: false,
+		passes: [{ grid: [gridX, gridY] }]
+	};
+}
+function pipelineSubmit(device, pipelines, inputs, outputs) {
+	const commandEncoder = device.createCommandEncoder();
+	for (const { pipeline,...shader } of pipelines) {
+		if (inputs.length !== shader.numInputs || outputs.length !== shader.numOutputs) throw new Error(`webgpu: expected ${shader.numInputs} inputs and ${shader.numOutputs} outputs, got ${inputs.length} inputs and ${outputs.length} outputs`);
+		const filteredPasses = shader.passes.filter(({ grid }) => require_backend.prod(grid) > 0);
+		if (filteredPasses.length === 0) continue;
+		const bindGroup = device.createBindGroup({
+			layout: pipeline.getBindGroupLayout(0),
+			entries: [...inputs.map((buffer, i) => ({
+				binding: i,
+				resource: { buffer }
+			})), ...outputs.map((buffer, i) => ({
+				binding: inputs.length + i,
+				resource: { buffer }
+			}))]
+		});
+		let uniformBindGroup = null;
+		let uniformAlignment = 0;
+		if (shader.hasUniform) {
+			const uniforms = filteredPasses.map(({ uniform }) => uniform);
+			const [uniformBuffer, alignment] = combineUniforms(device, uniforms);
+			uniformAlignment = alignment;
+			uniformBindGroup = device.createBindGroup({
+				layout: pipeline.getBindGroupLayout(1),
+				entries: [{
+					binding: 0,
+					resource: {
+						buffer: uniformBuffer,
+						size: alignment
+					}
+				}]
+			});
+		}
+		for (let i = 0; i < filteredPasses.length; i++) {
+			const { grid } = filteredPasses[i];
+			const passEncoder = commandEncoder.beginComputePass();
+			passEncoder.setPipeline(pipeline);
+			passEncoder.setBindGroup(0, bindGroup);
+			if (uniformBindGroup) passEncoder.setBindGroup(1, uniformBindGroup, [i * uniformAlignment]);
+			passEncoder.dispatchWorkgroups(grid[0], grid[1]);
+			passEncoder.end();
+		}
+	}
+	device.queue.submit([commandEncoder.finish()]);
+}
+function combineUniforms(device, uniforms) {
+	for (const buf of uniforms) if (!buf || buf.byteLength === 0 || buf.byteLength !== uniforms[0].byteLength) throw new Error("webgpu: Uniform mismatch between shader passes");
+	const minAlign = device.limits.minUniformBufferOffsetAlignment;
+	const alignment = Math.ceil(uniforms[0].byteLength / minAlign) * minAlign;
+	const buffer = device.createBuffer({
+		size: alignment * uniforms.length,
+		usage: GPUBufferUsage.UNIFORM,
+		mappedAtCreation: true
+	});
+	const bufferMapped = new Uint8Array(buffer.getMappedRange());
+	for (let i = 0; i < uniforms.length; i++) bufferMapped.set(uniforms[i], i * alignment);
+	buffer.unmap();
+	return [buffer, alignment];
+}
+/**
+* A cache for compiled GPU compute pipelines, keyed by the shader source.
+*
+* This supports both async compilation (recommended) and a synchronous variant.
+* If the pipeline is not in the cache, it will be compiled and added. For async
+* compilation, only one compilation will be in progress at a time for a given
+* shader source.
+*/
+var ShaderPipelineCache = class {
+	cache;
+	inProgress;
+	constructor(device) {
+		this.device = device;
+		this.cache = /* @__PURE__ */ new Map();
+		this.inProgress = /* @__PURE__ */ new Map();
+	}
+	#getLayout(shader) {
+		if (shader.numInputs + shader.numOutputs > this.device.limits.maxStorageBuffersPerShaderStage) {
+			const actual = shader.numInputs + shader.numOutputs;
+			const max = this.device.limits.maxStorageBuffersPerShaderStage;
+			throw new Error(`Too many buffers (${actual}) for WebGPU pipeline (max: ${max})`);
+		}
+		const bindGroupLayouts = [this.device.createBindGroupLayout({ entries: require_backend.range(shader.numInputs + shader.numOutputs).map((i) => ({
+			binding: i,
+			visibility: GPUShaderStage.COMPUTE,
+			buffer: { type: i < shader.numInputs ? "read-only-storage" : "storage" }
+		})) })];
+		if (shader.hasUniform) bindGroupLayouts.push(this.device.createBindGroupLayout({ entries: [{
+			binding: 0,
+			visibility: GPUShaderStage.COMPUTE,
+			buffer: {
+				type: "uniform",
+				hasDynamicOffset: true
+			}
+		}] }));
+		return this.device.createPipelineLayout({ bindGroupLayouts });
+	}
+	async prepare(shader) {
+		const existingPipeline = this.cache.get(shader.code);
+		if (existingPipeline) return existingPipeline;
+		const existingPromise = this.inProgress.get(shader.code);
+		if (existingPromise) return await existingPromise;
+		if (require_backend.DEBUG >= 2) console.info("=========== WebGPU shader ===========\n" + shader.code);
+		const shaderModule = this.device.createShaderModule({ code: shader.code });
+		const promise = (async () => {
+			this.device.pushErrorScope("validation");
+			try {
+				const pipeline$1 = await this.device.createComputePipelineAsync({
+					layout: this.#getLayout(shader),
+					compute: {
+						module: shaderModule,
+						entryPoint: "main"
+					}
+				});
+				await this.device.popErrorScope();
+				return pipeline$1;
+			} catch (_error) {
+				const scope = await this.device.popErrorScope();
+				const emsg = await compileError(shaderModule, scope, shader.code);
+				throw new Error(emsg);
+			}
+		})();
+		this.inProgress.set(shader.code, promise);
+		const pipeline = await promise;
+		this.cache.set(shader.code, pipeline);
+		return pipeline;
+	}
+	prepareSync(shader) {
+		const existingPipeline = this.cache.get(shader.code);
+		if (existingPipeline) return existingPipeline;
+		if (require_backend.DEBUG >= 2) console.info("=========== WebGPU shader ===========\n" + shader.code);
+		const shaderModule = this.device.createShaderModule({ code: shader.code });
+		this.device.pushErrorScope("validation");
+		const pipeline = this.device.createComputePipeline({
+			layout: this.#getLayout(shader),
+			compute: {
+				module: shaderModule,
+				entryPoint: "main"
+			}
+		});
+		this.device.popErrorScope().then(async (scope) => {
+			if (scope !== null) {
+				const emsg = await compileError(shaderModule, scope, shader.code);
+				console.error(emsg);
+			}
+		});
+		this.cache.set(shader.code, pipeline);
+		return pipeline;
+	}
+};
+/** Gather information about a compilation error and format it. */
+async function compileError(shaderModule, scope, code) {
+	let message = `Failed to compile shader: ${scope ? scope.message : "(no error scope)"}`;
+	const info = await shaderModule.getCompilationInfo();
+	for (const msg of info.messages) message += `\n  [${msg.type} at ${msg.lineNum}:${msg.linePos}] ${msg.message}`;
+	if (code) message += `\n\n${code}`;
+	return message;
+}
+
+//#endregion
+exports.WebGPUBackend = WebGPUBackend;