@jax-js/jax 0.1.3 → 0.1.4

package/README.md

@@ -3,7 +3,8 @@
 <p align="center"><strong>
   <a href="https://jax-js.com">Website</a> |
   <a href="https://jax-js.com/docs/">API Reference</a> |
-  <a href="./FEATURES.md">Compatibility Table</a>
+  <a href="./FEATURES.md">Compatibility Table</a> |
+  <a href="https://discord.gg/BW6YsCd4Tf">Discord</a>
 </strong></p>
 
 **jax-js** is a machine learning framework for the browser. It aims to bring
@@ -323,7 +324,7 @@ pnpm -C website dev
 
 ## Future work / help wanted
 
-Contributions are welcomed! Especially in:
+Contributions are welcomed! Some fruitful areas to look into:
 
 - Adding support for more JAX functions and operations, see [compatibility table](./FEATURES.md).
 - Improving performance of the WebGPU and Wasm runtimes, generating better kernels, and using SIMD
@@ -334,3 +335,5 @@ Contributions are welcomed! Especially in:
 - Helping the JIT compiler to fuse operations in more cases, like `tanh` branches.
 - Adding WebGL runtime for older browsers that don't support WebGPU.
 - Making a fast transformer inference engine, comparing against onnxruntime-web.
+
+You may join our [Discord server](https://discord.gg/BW6YsCd4Tf) and chat with the community.

package/dist/{backend-CmaidnkQ.cjs → backend-Bu9GY6sK.cjs}

@@ -69,6 +69,9 @@ function zipn(...arrays) {
 	const minLength = Math.min(...arrays.map((x) => x.length));
 	return Array.from({ length: minLength }, (_, i) => arrays.map((arr) => arr[i]));
 }
+function sorted(arr) {
+	return [...arr].sort((a, b) => a - b);
+}
 function rep(length, value) {
 	if (value instanceof Function) return new Array(length).fill(0).map((_, i) => value(i));
 	return new Array(length).fill(value);
@@ -145,7 +148,7 @@ function normalizeAxis(axis, ndim) {
 			if (seen.has(ca)) throw new Error(`Duplicate axis ${ca} passed to function`);
 			seen.add(ca);
 		}
-		return [...seen].sort();
+		return sorted(seen);
 	}
 }
 function range(start, stop, step = 1) {
@@ -1328,7 +1331,7 @@ var Reduction = class {
 	/** Evaluate this operation on CPU. */
 	evaluate(...values) {
 		if (this.dtype === DType.Bool) {
-			if (this.op === AluOp.Add || this.op === AluOp.Max) return values.reduce((a, b) => a || b, true);
+			if (this.op === AluOp.Add || this.op === AluOp.Max) return values.reduce((a, b) => a || b, false);
 			else if (this.op === AluOp.Mul || this.op === AluOp.Min) return values.reduce((a, b) => a && b, true);
 		} else if (this.dtype === DType.Int32) {
 			if (this.op === AluOp.Add) return values.reduce((a, b) => a + b | 0, 0);
@@ -1438,6 +1441,112 @@ function erfc(x) {
 	else return 2 - _erfapprox$1(-x);
 }
 
+//#endregion
+//#region src/routine.ts
+/**
+* Advanced operations that don't fit into the `AluExp` compiler representation.
+*
+* Some routines like iterative matrix algorithms, FFTs, or sorting may not be
+* easy to express efficiently as a `Kernel` object. These also tend to be
+* somewhat expensive, so the benefit of kernel fusion and inlining is less
+* relevant.
+*
+* For these operations, we dispatch them as a custom operation on the backend,
+* which each backend implements in a specific way. These are listed in the
+* `Routines` enum below.
+*
+* Routines cannot be fused into other kernels and always operate on contiguous
+* arrays (default `ShapeTracker`).
+*/
+var Routine = class {
+	constructor(name, type, params) {
+		this.name = name;
+		this.type = type;
+		this.params = params;
+	}
+};
+/** One of the valid `Routine` that can be dispatched to backend. */
+let Routines = /* @__PURE__ */ function(Routines$1) {
+	/** Stable sorting algorithm along the last axis. */
+	Routines$1["Sort"] = "Sort";
+	/** Returns `int32` indices of the stably sorted array. */
+	Routines$1["Argsort"] = "Argsort";
+	/** Solve a triangular system of questions. */
+	Routines$1["TriangularSolve"] = "TriangularSolve";
+	/** Cholesky decomposition of 2D positive semi-definite matrices. */
+	Routines$1["Cholesky"] = "Cholesky";
+	return Routines$1;
+}({});
+function runCpuRoutine(routine, inputs, outputs) {
+	const { name, type } = routine;
+	const inputAr = inputs.map((buf, i) => dtypedArray(type.inputDtypes[i], buf));
+	const outputAr = outputs.map((buf, i) => dtypedArray(type.outputDtypes[i], buf));
+	switch (name) {
+		case Routines.Sort: return runSort(type, inputAr, outputAr);
+		case Routines.Argsort: return runArgsort(type, inputAr, outputAr);
+		case Routines.TriangularSolve: return runTriangularSolve(type, inputAr, outputAr, routine.params);
+		case Routines.Cholesky: return runCholesky(type, inputAr, outputAr);
+		default:
+	}
+}
+function runSort(type, [x], [y]) {
+	const xs = type.inputShapes[0];
+	if (xs.length === 0) throw new Error("sort: cannot sort a scalar");
+	const n = xs[xs.length - 1];
+	y.set(x);
+	for (let i = 0; i < y.length; i += n) y.subarray(i, i + n).sort();
+}
+function runArgsort(type, [x], [y, yi]) {
+	const xs = type.inputShapes[0];
+	if (xs.length === 0) throw new Error("argsort: cannot sort a scalar");
+	const n = xs[xs.length - 1];
+	for (let offset = 0; offset < y.length; offset += n) {
+		const ar = x.subarray(offset, offset + n);
+		const out = y.subarray(offset, offset + n);
+		const outi = yi.subarray(offset, offset + n);
+		for (let i = 0; i < n; i++) outi[i] = i;
+		outi.sort((a, b) => ar[a] - ar[b]);
+		for (let i = 0; i < n; i++) out[i] = ar[outi[i]];
+	}
+}
+function runTriangularSolve(type, [a, b], [x], { unitDiagonal }) {
+	const as = type.inputShapes[0];
+	const bs = type.inputShapes[1];
+	if (as.length < 2) throw new Error(`triangular_solve: a must be at least 2D, got ${as}`);
+	if (bs.length < 2) throw new Error(`triangular_solve: b must be at least 2D, got ${bs}`);
+	const n = as[as.length - 2];
+	if (n !== as[as.length - 1] || n !== bs[bs.length - 1]) throw new Error(`triangular_solve: incompatible shapes a=${as}, b=${bs}`);
+	const batch = bs[bs.length - 2];
+	for (let counter = 0; counter < a.length / (n * n); counter++) {
+		const a1 = a.subarray(counter * n * n, (counter + 1) * n * n);
+		for (let t = 0; t < batch; t++) {
+			const b1 = b.subarray((counter * batch + t) * n, (counter * batch + t + 1) * n);
+			const x1 = x.subarray((counter * batch + t) * n, (counter * batch + t + 1) * n);
+			for (let i = n - 1; i >= 0; i--) {
+				let sum = b1[i];
+				for (let j = i + 1; j < n; j++) sum -= a1[i * n + j] * x1[j];
+				x1[i] = unitDiagonal ? sum : sum / a1[i * n + i];
+			}
+		}
+	}
+}
+function runCholesky(type, [x], [y]) {
+	const xs = type.inputShapes[0];
+	if (xs.length < 2) throw new Error("cholesky: input must be at least 2D");
+	const n = xs[xs.length - 2];
+	const m = xs[xs.length - 1];
+	if (n !== m) throw new Error(`cholesky: input must be square, got [${n}, ${m}]`);
+	for (let offset = 0; offset < y.length; offset += n * n) {
+		const ar = x.subarray(offset, offset + n * n);
+		const out = y.subarray(offset, offset + n * n);
+		for (let i = 0; i < n; i++) for (let j = 0; j <= i; j++) {
+			let sum = ar[i * n + j];
+			for (let k = 0; k < j; k++) sum -= out[i * n + k] * out[j * n + k];
+			out[i * n + j] = i === j ? Math.sqrt(sum) : sum / out[j * n + j];
+		}
+	}
+}
+
 //#endregion
 //#region src/shape.ts
 const jstr = JSON.stringify;
@@ -1908,7 +2017,7 @@ var ShapeTracker = class ShapeTracker {
 		let st = this;
 		if (axis.length > 0) {
 			const unsqueezed = [...st.shape];
-			for (const i of axis.toSorted()) unsqueezed.splice(i, 0, 1);
+			for (const i of sorted(axis)) unsqueezed.splice(i, 0, 1);
 			st = st.reshape(unsqueezed);
 		}
 		return st.expand(newShape);
@@ -2133,7 +2242,7 @@ function tuneWebgpu(kernel) {
 			break;
 		}
 	}
-	for (const ax of Array.from(upcastedAxis).sort()) {
+	for (const ax of sorted(upcastedAxis)) {
 		const s = dim.st.shape[ax];
 		for (const amount of [8, 4]) if (s % amount === 0) {
 			dim.applyLocal(ax, amount);
@@ -2251,13 +2360,21 @@ var CpuBackend = class {
 		if (count === void 0) count = buffer.byteLength - start;
 		return buffer.slice(start, start + count);
 	}
-	async prepare(kernel) {
-		return this.prepareSync(kernel);
+	async prepareKernel(kernel) {
+		return this.prepareKernelSync(kernel);
 	}
-	prepareSync(kernel) {
+	prepareKernelSync(kernel) {
 		return new Executable(kernel, void 0);
 	}
-	dispatch({ kernel }, inputs, outputs) {
+	async prepareRoutine(routine) {
+		return this.prepareRoutineSync(routine);
+	}
+	prepareRoutineSync(routine) {
+		return new Executable(routine, void 0);
+	}
+	dispatch(exe, inputs, outputs) {
+		if (exe.source instanceof Routine) return runCpuRoutine(exe.source, inputs.map((slot) => this.#getBuffer(slot)), outputs.map((slot) => this.#getBuffer(slot)));
+		const kernel = exe.source;
 		const { exp, epilogue } = tuneNullopt(kernel);
 		const inputBuffers = inputs.map((slot) => this.#getBuffer(slot));
 		const outputBuffers = outputs.map((slot) => this.#getBuffer(slot));
@@ -2315,8 +2432,10 @@ var WasmAllocator = class {
 		const sizeClass = this.#findSizeClass(size);
 		const freeList = this.#freeLists.get(sizeClass);
 		let ptr;
-		if (freeList && freeList.length > 0) ptr = freeList.pop();
-		else ptr = this.#bumpAlloc(sizeClass);
+		if (freeList && freeList.length > 0) {
+			ptr = freeList.pop();
+			new Uint8Array(this.#memory.buffer, ptr, sizeClass).fill(0);
+		} else ptr = this.#bumpAlloc(sizeClass);
 		this.#allocatedBuffers.set(ptr, sizeClass);
 		return ptr;
 	}
@@ -3682,10 +3801,10 @@ var WasmBackend = class {
 		if (count === void 0) count = buffer.byteLength - start;
 		return buffer.slice(start, start + count);
 	}
-	async prepare(kernel) {
-		return this.prepareSync(kernel);
+	async prepareKernel(kernel) {
+		return this.prepareKernelSync(kernel);
 	}
-	prepareSync(kernel) {
+	prepareKernelSync(kernel) {
 		const kernelHash = FpHash.hash(kernel);
 		const module$1 = runWithCache(moduleCache, kernelHash.toString(), () => {
 			const bytes = codegenWasm(kernel);
@@ -3693,7 +3812,14 @@ var WasmBackend = class {
 		});
 		return new Executable(kernel, { module: module$1 });
 	}
+	async prepareRoutine(routine) {
+		return this.prepareRoutineSync(routine);
+	}
+	prepareRoutineSync(routine) {
+		return new Executable(routine, void 0);
+	}
 	dispatch(exe, inputs, outputs) {
+		if (exe.source instanceof Routine) return runCpuRoutine(exe.source, inputs.map((slot) => this.#getBuffer(slot)), outputs.map((slot) => this.#getBuffer(slot)));
 		const instance = new WebAssembly.Instance(exe.data.module, { env: { memory: this.#memory } });
 		const func = instance.exports.kernel;
 		const ptrs = [...inputs, ...outputs].map((slot) => this.#buffers.get(slot).ptr);
@@ -4041,7 +4167,7 @@ async function createBackend(device) {
 		if (!navigator.gpu) return null;
 		const adapter = await navigator.gpu.requestAdapter({ powerPreference: "high-performance" });
 		if (!adapter) return null;
-		const { WebGPUBackend } = await Promise.resolve().then(() => require("./webgpu-BVns4DbI.cjs"));
+		const { WebGPUBackend } = await Promise.resolve().then(() => require("./webgpu-Oj3Kd-kd.cjs"));
 		const importantLimits = [
 			"maxBufferSize",
 			"maxComputeInvocationsPerWorkgroup",
@@ -4075,8 +4201,8 @@ function getBackend(device) {
 	return backend;
 }
 var Executable = class {
-	constructor(kernel, data) {
-		this.kernel = kernel;
+	constructor(source, data) {
+		this.source = source;
 		this.data = data;
 	}
 };
@@ -4092,6 +4218,11 @@ var UnsupportedOpError = class extends Error {
 		super(msg);
 	}
 };
+var UnsupportedRoutineError = class extends Error {
+	constructor(name, device) {
+		super(`routine '${name}' is not supported in ${device} backend`);
+	}
+};
 
 //#endregion
 Object.defineProperty(exports, 'AluExp', {
@@ -4160,6 +4291,18 @@ Object.defineProperty(exports, 'Reduction', {
     return Reduction;
   }
 });
+Object.defineProperty(exports, 'Routine', {
+  enumerable: true,
+  get: function () {
+    return Routine;
+  }
+});
+Object.defineProperty(exports, 'Routines', {
+  enumerable: true,
+  get: function () {
+    return Routines;
+  }
+});
 Object.defineProperty(exports, 'ShapeTracker', {
   enumerable: true,
   get: function () {
@@ -4178,6 +4321,12 @@ Object.defineProperty(exports, 'UnsupportedOpError', {
     return UnsupportedOpError;
   }
 });
+Object.defineProperty(exports, 'UnsupportedRoutineError', {
+  enumerable: true,
+  get: function () {
+    return UnsupportedRoutineError;
+  }
+});
 Object.defineProperty(exports, 'accessorAluExp', {
   enumerable: true,
   get: function () {
@@ -4387,5 +4536,4 @@ Object.defineProperty(exports, 'zipn', {
   get: function () {
     return zipn;
   }
-});
-//# sourceMappingURL=backend-CmaidnkQ.cjs.map
+});

package/dist/{backend-BY8wlLEl.js → backend-tngXtWe4.js}

@@ -68,6 +68,9 @@ function zipn(...arrays) {
 	const minLength = Math.min(...arrays.map((x) => x.length));
 	return Array.from({ length: minLength }, (_, i) => arrays.map((arr) => arr[i]));
 }
+function sorted(arr) {
+	return [...arr].sort((a, b) => a - b);
+}
 function rep(length, value) {
 	if (value instanceof Function) return new Array(length).fill(0).map((_, i) => value(i));
 	return new Array(length).fill(value);
@@ -144,7 +147,7 @@ function normalizeAxis(axis, ndim) {
 			if (seen.has(ca)) throw new Error(`Duplicate axis ${ca} passed to function`);
 			seen.add(ca);
 		}
-		return [...seen].sort();
+		return sorted(seen);
 	}
 }
 function range(start, stop, step = 1) {
@@ -1327,7 +1330,7 @@ var Reduction = class {
 	/** Evaluate this operation on CPU. */
 	evaluate(...values) {
 		if (this.dtype === DType.Bool) {
-			if (this.op === AluOp.Add || this.op === AluOp.Max) return values.reduce((a, b) => a || b, true);
+			if (this.op === AluOp.Add || this.op === AluOp.Max) return values.reduce((a, b) => a || b, false);
 			else if (this.op === AluOp.Mul || this.op === AluOp.Min) return values.reduce((a, b) => a && b, true);
 		} else if (this.dtype === DType.Int32) {
 			if (this.op === AluOp.Add) return values.reduce((a, b) => a + b | 0, 0);
@@ -1437,6 +1440,112 @@ function erfc(x) {
 	else return 2 - _erfapprox$1(-x);
 }
 
+//#endregion
+//#region src/routine.ts
+/**
+* Advanced operations that don't fit into the `AluExp` compiler representation.
+*
+* Some routines like iterative matrix algorithms, FFTs, or sorting may not be
+* easy to express efficiently as a `Kernel` object. These also tend to be
+* somewhat expensive, so the benefit of kernel fusion and inlining is less
+* relevant.
+*
+* For these operations, we dispatch them as a custom operation on the backend,
+* which each backend implements in a specific way. These are listed in the
+* `Routines` enum below.
+*
+* Routines cannot be fused into other kernels and always operate on contiguous
+* arrays (default `ShapeTracker`).
+*/
+var Routine = class {
+	constructor(name, type, params) {
+		this.name = name;
+		this.type = type;
+		this.params = params;
+	}
+};
+/** One of the valid `Routine` that can be dispatched to backend. */
+let Routines = /* @__PURE__ */ function(Routines$1) {
+	/** Stable sorting algorithm along the last axis. */
+	Routines$1["Sort"] = "Sort";
+	/** Returns `int32` indices of the stably sorted array. */
+	Routines$1["Argsort"] = "Argsort";
+	/** Solve a triangular system of questions. */
+	Routines$1["TriangularSolve"] = "TriangularSolve";
+	/** Cholesky decomposition of 2D positive semi-definite matrices. */
+	Routines$1["Cholesky"] = "Cholesky";
+	return Routines$1;
+}({});
+function runCpuRoutine(routine, inputs, outputs) {
+	const { name, type } = routine;
+	const inputAr = inputs.map((buf, i) => dtypedArray(type.inputDtypes[i], buf));
+	const outputAr = outputs.map((buf, i) => dtypedArray(type.outputDtypes[i], buf));
+	switch (name) {
+		case Routines.Sort: return runSort(type, inputAr, outputAr);
+		case Routines.Argsort: return runArgsort(type, inputAr, outputAr);
+		case Routines.TriangularSolve: return runTriangularSolve(type, inputAr, outputAr, routine.params);
+		case Routines.Cholesky: return runCholesky(type, inputAr, outputAr);
+		default:
+	}
+}
+function runSort(type, [x], [y]) {
+	const xs = type.inputShapes[0];
+	if (xs.length === 0) throw new Error("sort: cannot sort a scalar");
+	const n = xs[xs.length - 1];
+	y.set(x);
+	for (let i = 0; i < y.length; i += n) y.subarray(i, i + n).sort();
+}
+function runArgsort(type, [x], [y, yi]) {
+	const xs = type.inputShapes[0];
+	if (xs.length === 0) throw new Error("argsort: cannot sort a scalar");
+	const n = xs[xs.length - 1];
+	for (let offset = 0; offset < y.length; offset += n) {
+		const ar = x.subarray(offset, offset + n);
+		const out = y.subarray(offset, offset + n);
+		const outi = yi.subarray(offset, offset + n);
+		for (let i = 0; i < n; i++) outi[i] = i;
+		outi.sort((a, b) => ar[a] - ar[b]);
+		for (let i = 0; i < n; i++) out[i] = ar[outi[i]];
+	}
+}
+function runTriangularSolve(type, [a, b], [x], { unitDiagonal }) {
+	const as = type.inputShapes[0];
+	const bs = type.inputShapes[1];
+	if (as.length < 2) throw new Error(`triangular_solve: a must be at least 2D, got ${as}`);
+	if (bs.length < 2) throw new Error(`triangular_solve: b must be at least 2D, got ${bs}`);
+	const n = as[as.length - 2];
+	if (n !== as[as.length - 1] || n !== bs[bs.length - 1]) throw new Error(`triangular_solve: incompatible shapes a=${as}, b=${bs}`);
+	const batch = bs[bs.length - 2];
+	for (let counter = 0; counter < a.length / (n * n); counter++) {
+		const a1 = a.subarray(counter * n * n, (counter + 1) * n * n);
+		for (let t = 0; t < batch; t++) {
+			const b1 = b.subarray((counter * batch + t) * n, (counter * batch + t + 1) * n);
+			const x1 = x.subarray((counter * batch + t) * n, (counter * batch + t + 1) * n);
+			for (let i = n - 1; i >= 0; i--) {
+				let sum = b1[i];
+				for (let j = i + 1; j < n; j++) sum -= a1[i * n + j] * x1[j];
+				x1[i] = unitDiagonal ? sum : sum / a1[i * n + i];
+			}
+		}
+	}
+}
+function runCholesky(type, [x], [y]) {
+	const xs = type.inputShapes[0];
+	if (xs.length < 2) throw new Error("cholesky: input must be at least 2D");
+	const n = xs[xs.length - 2];
+	const m = xs[xs.length - 1];
+	if (n !== m) throw new Error(`cholesky: input must be square, got [${n}, ${m}]`);
+	for (let offset = 0; offset < y.length; offset += n * n) {
+		const ar = x.subarray(offset, offset + n * n);
+		const out = y.subarray(offset, offset + n * n);
+		for (let i = 0; i < n; i++) for (let j = 0; j <= i; j++) {
+			let sum = ar[i * n + j];
+			for (let k = 0; k < j; k++) sum -= out[i * n + k] * out[j * n + k];
+			out[i * n + j] = i === j ? Math.sqrt(sum) : sum / out[j * n + j];
+		}
+	}
+}
+
 //#endregion
 //#region src/shape.ts
 const jstr = JSON.stringify;
@@ -1907,7 +2016,7 @@ var ShapeTracker = class ShapeTracker {
 		let st = this;
 		if (axis.length > 0) {
 			const unsqueezed = [...st.shape];
-			for (const i of axis.toSorted()) unsqueezed.splice(i, 0, 1);
+			for (const i of sorted(axis)) unsqueezed.splice(i, 0, 1);
 			st = st.reshape(unsqueezed);
 		}
 		return st.expand(newShape);
@@ -2132,7 +2241,7 @@ function tuneWebgpu(kernel) {
 			break;
 		}
 	}
-	for (const ax of Array.from(upcastedAxis).sort()) {
+	for (const ax of sorted(upcastedAxis)) {
 		const s = dim.st.shape[ax];
 		for (const amount of [8, 4]) if (s % amount === 0) {
 			dim.applyLocal(ax, amount);
@@ -2250,13 +2359,21 @@ var CpuBackend = class {
 		if (count === void 0) count = buffer.byteLength - start;
 		return buffer.slice(start, start + count);
 	}
-	async prepare(kernel) {
-		return this.prepareSync(kernel);
+	async prepareKernel(kernel) {
+		return this.prepareKernelSync(kernel);
 	}
-	prepareSync(kernel) {
+	prepareKernelSync(kernel) {
 		return new Executable(kernel, void 0);
 	}
-	dispatch({ kernel }, inputs, outputs) {
+	async prepareRoutine(routine) {
+		return this.prepareRoutineSync(routine);
+	}
+	prepareRoutineSync(routine) {
+		return new Executable(routine, void 0);
+	}
+	dispatch(exe, inputs, outputs) {
+		if (exe.source instanceof Routine) return runCpuRoutine(exe.source, inputs.map((slot) => this.#getBuffer(slot)), outputs.map((slot) => this.#getBuffer(slot)));
+		const kernel = exe.source;
 		const { exp, epilogue } = tuneNullopt(kernel);
 		const inputBuffers = inputs.map((slot) => this.#getBuffer(slot));
 		const outputBuffers = outputs.map((slot) => this.#getBuffer(slot));
@@ -2314,8 +2431,10 @@ var WasmAllocator = class {
 		const sizeClass = this.#findSizeClass(size);
 		const freeList = this.#freeLists.get(sizeClass);
 		let ptr;
-		if (freeList && freeList.length > 0) ptr = freeList.pop();
-		else ptr = this.#bumpAlloc(sizeClass);
+		if (freeList && freeList.length > 0) {
+			ptr = freeList.pop();
+			new Uint8Array(this.#memory.buffer, ptr, sizeClass).fill(0);
+		} else ptr = this.#bumpAlloc(sizeClass);
 		this.#allocatedBuffers.set(ptr, sizeClass);
 		return ptr;
 	}
@@ -3681,10 +3800,10 @@ var WasmBackend = class {
 		if (count === void 0) count = buffer.byteLength - start;
 		return buffer.slice(start, start + count);
 	}
-	async prepare(kernel) {
-		return this.prepareSync(kernel);
+	async prepareKernel(kernel) {
+		return this.prepareKernelSync(kernel);
 	}
-	prepareSync(kernel) {
+	prepareKernelSync(kernel) {
 		const kernelHash = FpHash.hash(kernel);
 		const module = runWithCache(moduleCache, kernelHash.toString(), () => {
 			const bytes = codegenWasm(kernel);
@@ -3692,7 +3811,14 @@ var WasmBackend = class {
 		});
 		return new Executable(kernel, { module });
 	}
+	async prepareRoutine(routine) {
+		return this.prepareRoutineSync(routine);
+	}
+	prepareRoutineSync(routine) {
+		return new Executable(routine, void 0);
+	}
 	dispatch(exe, inputs, outputs) {
+		if (exe.source instanceof Routine) return runCpuRoutine(exe.source, inputs.map((slot) => this.#getBuffer(slot)), outputs.map((slot) => this.#getBuffer(slot)));
 		const instance = new WebAssembly.Instance(exe.data.module, { env: { memory: this.#memory } });
 		const func = instance.exports.kernel;
 		const ptrs = [...inputs, ...outputs].map((slot) => this.#buffers.get(slot).ptr);
@@ -4040,7 +4166,7 @@ async function createBackend(device) {
 		if (!navigator.gpu) return null;
 		const adapter = await navigator.gpu.requestAdapter({ powerPreference: "high-performance" });
 		if (!adapter) return null;
-		const { WebGPUBackend } = await import("./webgpu-C9iAP5h5.js");
+		const { WebGPUBackend } = await import("./webgpu-ChVgx3b6.js");
 		const importantLimits = [
 			"maxBufferSize",
 			"maxComputeInvocationsPerWorkgroup",
@@ -4074,8 +4200,8 @@ function getBackend(device) {
 	return backend;
 }
 var Executable = class {
-	constructor(kernel, data) {
-		this.kernel = kernel;
+	constructor(source, data) {
+		this.source = source;
 		this.data = data;
 	}
 };
@@ -4091,7 +4217,11 @@ var UnsupportedOpError = class extends Error {
 		super(msg);
 	}
 };
+var UnsupportedRoutineError = class extends Error {
+	constructor(name, device) {
+		super(`routine '${name}' is not supported in ${device} backend`);
+	}
+};
 
 //#endregion
-export { AluExp, AluGroup, AluOp, AluVar, DEBUG, DType, Executable, FpHash, Kernel, PPrint, Reduction, ShapeTracker, SlotError, UnsupportedOpError, accessorAluExp, accessorGlobal, assertNonNull, byteWidth, checkAxis, deepEqual, defaultDevice, devices, dtypedArray, dtypedJsArray, findPow2, generalBroadcast, getBackend, init, invertPermutation, isFloatDtype, isNumberPair, isPermutation, mapSetUnion, normalizeAxis, partitionList, prod, promoteTypes, range, recursiveFlatten, rep, runWithCache, setDebug, strip1, toposort, tuneWebgpu, unravelAlu, unzip2, zip, zipn };
-//# sourceMappingURL=backend-BY8wlLEl.js.map
+export { AluExp, AluGroup, AluOp, AluVar, DEBUG, DType, Executable, FpHash, Kernel, PPrint, Reduction, Routine, Routines, ShapeTracker, SlotError, UnsupportedOpError, UnsupportedRoutineError, accessorAluExp, accessorGlobal, assertNonNull, byteWidth, checkAxis, deepEqual, defaultDevice, devices, dtypedArray, dtypedJsArray, findPow2, generalBroadcast, getBackend, init, invertPermutation, isFloatDtype, isNumberPair, isPermutation, mapSetUnion, normalizeAxis, partitionList, prod, promoteTypes, range, recursiveFlatten, rep, runWithCache, setDebug, strip1, toposort, tuneWebgpu, unravelAlu, unzip2, zip, zipn };