@jax-js/jax 0.1.9 → 0.1.11

package/dist/{backend-BId79r5b.js → backend-DZvR7mZV.js}

@@ -313,6 +313,16 @@ function runWithCache(cache, key, thunk) {
 		return value;
 	}
 }
+/** Async version of `runWithCache`. */
+async function runWithCacheAsync(cache, key, thunk) {
+	const keyStr = JSON.stringify(key);
+	if (cache.has(keyStr)) return cache.get(keyStr);
+	else {
+		const value = await thunk();
+		cache.set(keyStr, value);
+		return value;
+	}
+}
 
 //#endregion
 //#region src/alu.ts
@@ -415,8 +425,23 @@ var AluExp = class AluExp {
 		this.src = src;
 		this.arg = arg;
 		if (AluGroup.RequiredFloat.has(op) && !isFloatDtype(dtype)) throw new TypeError(`Unsupported dtype for ${op}: ${dtype}`);
-		if (op === AluOp.Bitcast && (dtype === DType.Bool || src[0].dtype === DType.Bool || byteWidth(dtype) !== byteWidth(src[0].dtype))) throw new TypeError(`Bitcast from ${src[0].dtype} -> ${dtype}`);
-		if (op === AluOp.Threefry2x32 && (dtype !== DType.Uint32 || src.some((x) => x.dtype !== DType.Uint32))) throw new TypeError("Threefry2x32 requires uint32 types");
+		switch (op) {
+			case AluOp.Bitcast:
+				if (dtype === DType.Bool || src[0].dtype === DType.Bool || byteWidth(dtype) !== byteWidth(src[0].dtype)) throw new TypeError(`Bitcast from ${src[0].dtype} -> ${dtype}`);
+				break;
+			case AluOp.Threefry2x32:
+				if (dtype !== DType.Uint32 || src.some((x) => x.dtype !== DType.Uint32)) throw new TypeError("Threefry2x32 requires uint32 types");
+				break;
+			case AluOp.BitCombine:
+				if (src[0].dtype !== src[1].dtype || isFloatDtype(src[0].dtype)) throw new TypeError(`BitCombine[${arg}] requires matching integral dtype, got ${src[0].dtype} and ${src[1].dtype}`);
+				break;
+			case AluOp.BitShift:
+				if (src[0].dtype === DType.Bool || src[1].dtype === DType.Bool || isFloatDtype(src[0].dtype) || isFloatDtype(src[1].dtype)) throw new TypeError(`BitShift[${arg}] requires two integral, non-bool dtypes, got ${src[0].dtype} and ${src[1].dtype}`);
+				break;
+			case AluOp.BitInvert:
+				if (isFloatDtype(src[0].dtype)) throw new TypeError(`BitInvert requires an integral dtype, got ${src[0].dtype}`);
+				break;
+		}
 	}
 	static add(a, b) {
 		return new AluExp(AluOp.Add, a.dtype, [a, b]);
@@ -493,6 +518,12 @@ var AluExp = class AluExp {
 			c1
 		], mode);
 	}
+	static bitCombine(a, b, mode) {
+		return new AluExp(AluOp.BitCombine, a.dtype, [a, b], mode);
+	}
+	static bitShift(a, b, mode) {
+		return new AluExp(AluOp.BitShift, a.dtype, [a, b], mode);
+	}
 	static cmplt(a, b) {
 		return new AluExp(AluOp.Cmplt, DType.Bool, [a, b]);
 	}
@@ -965,6 +996,16 @@ var AluExp = class AluExp {
 				case AluOp.Mod: return x % y;
 				case AluOp.Min: return Math.min(x, y);
 				case AluOp.Max: return Math.max(x, y);
+				case AluOp.BitCombine: {
+					let r;
+					if (this.arg === "and") r = x & y;
+					else if (this.arg === "or") r = x | y;
+					else r = x ^ y;
+					return this.dtype === DType.Int32 ? r | 0 : r >>> 0;
+				}
+				case AluOp.BitShift:
+					if (this.arg === "shl") return this.dtype === DType.Int32 ? x << y | 0 : x << y >>> 0;
+					return x >>> y;
 				case AluOp.Cmplt: return Number(x < y);
 				case AluOp.Cmpne: return Number(x != y);
 				default: throw new Error(`Missing implemementation for ${this.op}`);
@@ -1086,6 +1127,18 @@ var AluExp = class AluExp {
 			}
 			if (BIN_SYM[node.op]) return `(${parts[0]} ${BIN_SYM[node.op]} ${parts[1]})`;
 			if (CMP_SYM[node.op]) return `(${parts[0]} ${CMP_SYM[node.op]} ${parts[1]})`;
+			if (node.op === AluOp.BitCombine) {
+				const sym = {
+					and: "&",
+					or: "|",
+					xor: "^"
+				}[node.arg];
+				return `(${parts[0]} ${sym} ${parts[1]})`;
+			}
+			if (node.op === AluOp.BitShift) {
+				const sym = node.arg === "shl" ? "<<" : ">>";
+				return `(${parts[0]} ${sym} ${parts[1]})`;
+			}
 			if (UNARY_SYM[node.op]) return `${UNARY_SYM[node.op]}${parts[0]}`;
 			if (node.op === AluOp.Cast) return `Cast<${node.dtype}>(${strip1(parts[0])})`;
 			if (node.op === AluOp.Bitcast) return `Bitcast<${node.dtype}>(${strip1(parts[0])})`;
@@ -1178,6 +1231,9 @@ let AluOp = /* @__PURE__ */ function(AluOp$1) {
 	AluOp$1["Reciprocal"] = "Reciprocal";
 	AluOp$1["Cast"] = "Cast";
 	AluOp$1["Bitcast"] = "Bitcast";
+	AluOp$1["BitCombine"] = "BitCombine";
+	AluOp$1["BitInvert"] = "BitInvert";
+	AluOp$1["BitShift"] = "BitShift";
 	AluOp$1["Cmplt"] = "Cmplt";
 	AluOp$1["Cmpne"] = "Cmpne";
 	AluOp$1["Where"] = "Where";
@@ -1197,7 +1253,9 @@ const AluGroup = {
 		AluOp.Idiv,
 		AluOp.Mod,
 		AluOp.Min,
-		AluOp.Max
+		AluOp.Max,
+		AluOp.BitCombine,
+		AluOp.BitShift
 	]),
 	Unary: new Set([
 		AluOp.Sin,
@@ -1312,6 +1370,10 @@ var Reduction = class {
 		this.epilogue = epilogue;
 		if (!AluGroup.Reduce.has(op)) throw new TypeError(`Unsupported reduction: ${op}`);
 		this.epilogue = epilogue.simplify();
+		if (this.dtype === DType.Float16 && this.op === AluOp.Add) {
+			this.epilogue = this.epilogue.substitute({ acc: AluExp.cast(this.dtype, AluVar.acc(DType.Float32)) });
+			this.dtype = DType.Float32;
+		}
 	}
 	hash(state) {
 		state.update(this.dtype).update(this.op).update(this.size).update(this.epilogue);
@@ -2266,11 +2328,15 @@ var TuneDims = class {
 };
 /** Tuning step that does not apply any optimization. */
 function tuneNullopt(kernel) {
+	let exp = kernel.exp;
 	const vars = {};
 	vars.gidx = AluExp.special(DType.Int32, "gidx", kernel.size);
-	if (kernel.reduction) vars.ridx = AluExp.special(DType.Int32, "ridx", kernel.reduction.size);
+	if (kernel.reduction) {
+		vars.ridx = AluExp.special(DType.Int32, "ridx", kernel.reduction.size);
+		if (exp.dtype !== kernel.reduction.dtype) exp = AluExp.cast(kernel.reduction.dtype, exp);
+	}
 	return {
-		exp: kernel.exp.substitute(vars).rewriteGlobalViews().simplify(),
+		exp: exp.substitute(vars).rewriteGlobalViews().simplify(),
 		epilogue: kernel.reduction?.epilogue.substitute({ gidx: vars.gidx }).rewriteGlobalViews().simplify(),
 		outputIdxExp: vars.gidx,
 		threadCount: kernel.size,
@@ -2279,8 +2345,9 @@ function tuneNullopt(kernel) {
 }
 /** Tuning for WebGPU kernels. */
 function tuneWebgpu(kernel) {
-	const { exp, reduction } = kernel;
+	const reduction = kernel.reduction;
 	if (!reduction) return tuneNullopt(kernel);
+	const exp = AluExp.cast(reduction.dtype, kernel.exp);
 	const globalIndexes = exp.collect((exp$1) => exp$1.op === AluOp.GlobalIndex);
 	if (globalIndexes.length > 0) {
 		if (DEBUG >= 4) console.info("Tuning: Found GlobalIndex ops, skipping opt.");
@@ -2508,6 +2575,85 @@ var CpuBackend = class {
 	}
 };
 
+//#endregion
+//#region src/tracing.ts
+let traceEnabled = false;
+const flushCallbacks = [];
+/**
+* Start collecting kernel traces.
+*
+* Traces appear in developer tools under the "Performance" tab, and they are
+* useful for measuring fine-grained kernel execution time.
+*/
+function startTrace() {
+	traceEnabled = true;
+}
+/**
+* Stop collecting kernel traces.
+*
+* Traces appear in developer tools under the "Performance" tab, and they are
+* useful for measuring fine-grained kernel execution time.
+*/
+function stopTrace() {
+	traceEnabled = false;
+	for (const cb of flushCallbacks) cb();
+}
+/** Check if tracing is currently enabled. */
+function isTracing() {
+	return traceEnabled;
+}
+/** Register a callback to flush pending trace data when tracing stops. */
+function onFlushTrace(cb) {
+	flushCallbacks.push(cb);
+}
+function humanSize(n) {
+	if (n >= 1e9) return `${(n / 1e9).toPrecision(3)}B`;
+	if (n >= 1e6) return `${(n / 1e6).toPrecision(3)}M`;
+	if (n >= 1e3) return `${(n / 1e3).toPrecision(3)}K`;
+	return `${n}`;
+}
+/** Build a trace label, properties, and color from a kernel or routine source. */
+function traceSourceInfo(source) {
+	const properties = [];
+	let label;
+	let color;
+	if (source instanceof Kernel) {
+		label = `Kernel[${humanSize(source.size)}]`;
+		properties.push(["exp", `${source.exp}`]);
+		properties.push(["size", `${source.size}`]);
+		properties.push(["nargs", `${source.nargs}`]);
+		if (!source.reduction) color = "primary";
+		else {
+			color = "secondary";
+			properties.push(["reduction", `${source.reduction.op}:${source.reduction.size}`]);
+		}
+	} else {
+		color = "tertiary";
+		label = source.name;
+		properties.push(["inputShapes", source.type.inputShapes.map((s) => `[${s}]`).join(", ")]);
+		properties.push(["outputShapes", source.type.outputShapes.map((s) => `[${s}]`).join(", ")]);
+		properties.push(["dtype", source.type.inputDtypes.join(", ")]);
+	}
+	return {
+		label,
+		color,
+		properties
+	};
+}
+/** Emit a trace entry as a `performance.measure` with devtools metadata. */
+function emitTrace(track, info, start, end) {
+	performance.measure(info.label, {
+		detail: { devtools: {
+			trackGroup: "JAX Profiler",
+			track,
+			color: info.color,
+			properties: info.properties
+		} },
+		start,
+		end
+	});
+}
+
 //#endregion
 //#region src/backend/wasm/allocator.ts
 /** Simple tensor memory allocator for WebAssembly linear memory. */
@@ -3070,6 +3216,147 @@ function wasm_threefry2x32(cg) {
 	});
 }
 
+//#endregion
+//#region src/backend/wasm/parallel.ts
+/** Check if SharedArrayBuffer is available. */
+function hasSharedArrayBuffer() {
+	return typeof SharedArrayBuffer !== "undefined" && typeof Worker !== "undefined";
+}
+const MIN_ELEMS_PER_THREAD = 256;
+const WORKER_SOURCE = `
+let memory = null;
+let cachedModule = null;
+let cachedFunc = null;
+
+self.onmessage = (e) => {
+  const msg = e.data;
+  if (msg.type === "init") {
+    memory = msg.memory;
+    postMessage({ type: "ready" });
+    return;
+  }
+  try {
+    const { module, ptrs, begin, end } = msg;
+    if (module !== cachedModule) {
+      cachedModule = module;
+      const instance = new WebAssembly.Instance(module, { env: { memory } });
+      cachedFunc = instance.exports.kernel;
+    }
+    cachedFunc(...ptrs, begin, end);
+    postMessage({ type: "done", ok: true });
+  } catch (err) {
+    postMessage({ type: "done", ok: false, error: String(err) });
+  }
+};
+`;
+/** Pool of Web Workers for parallel WASM kernel dispatch. */
+var WasmWorkerPool = class {
+	#memory;
+	#numWorkers;
+	#workers = [];
+	#ready = Promise.resolve();
+	/** Serializes dispatches so concurrent read() calls don't clobber onmessage. */
+	#queue = Promise.resolve();
+	#epoch = 0n;
+	#epochEnd = 0n;
+	#hooks = /* @__PURE__ */ new Map();
+	constructor(memory, numWorkers) {
+		if (numWorkers <= 0) throw new Error("numWorkers must be positive");
+		this.#memory = memory;
+		this.#numWorkers = numWorkers;
+	}
+	get epoch() {
+		return this.#epoch;
+	}
+	waitForEpoch(target) {
+		if (target <= this.#epoch) return Promise.resolve();
+		return new Promise((resolve) => {
+			if (target <= this.#epoch) return resolve();
+			const hooks = this.#hooks.get(target);
+			if (hooks) hooks.push(resolve);
+			else this.#hooks.set(target, [resolve]);
+		});
+	}
+	#ensureInit() {
+		if (this.#workers.length > 0) return;
+		const blob = new Blob([WORKER_SOURCE], { type: "application/javascript" });
+		const url = URL.createObjectURL(blob);
+		this.#workers = [];
+		const readyPromises = [];
+		for (let i = 0; i < this.#numWorkers; i++) {
+			const worker = new Worker(url, { type: "module" });
+			this.#workers.push(worker);
+			readyPromises.push(new Promise((resolve, reject) => {
+				worker.onmessage = () => resolve();
+				worker.onerror = (e) => reject(new Error(e.message || "Worker failed to load"));
+			}));
+			worker.postMessage({
+				type: "init",
+				memory: this.#memory
+			});
+		}
+		this.#ready = Promise.all(readyPromises).then(() => {
+			URL.revokeObjectURL(url);
+		});
+		this.#queue = this.#ready;
+	}
+	/**
+	* Dispatch a kernel across multiple workers.
+	*
+	* Returns an epoch that can be used to wait for the ongoing work to complete,
+	* which is guaranteed to be monotonically increasing.
+	*/
+	dispatch(module, ptrs, size) {
+		this.#ensureInit();
+		this.#epochEnd++;
+		const result = this.#queue.then(() => this.#dispatchNow(module, ptrs, size));
+		this.#queue = result.then(() => {}, () => {}).then(() => {
+			this.#epoch++;
+			const hooks = this.#hooks.get(this.#epoch);
+			if (hooks) {
+				for (const hook of hooks) hook();
+				this.#hooks.delete(this.#epoch);
+			}
+		});
+		return this.#epochEnd;
+	}
+	async #dispatchNow(module, ptrs, size) {
+		if (size === 0) return;
+		const n = Math.min(this.#workers.length, Math.ceil(size / MIN_ELEMS_PER_THREAD));
+		const chunkSize = Math.ceil(size / n / 16) * 16;
+		const promises = [];
+		for (let i = 0; i < n; i++) {
+			const begin = i * chunkSize;
+			const end = Math.min(begin + chunkSize, size);
+			if (begin >= size) break;
+			const worker = this.#workers[i];
+			promises.push(new Promise((resolve, reject) => {
+				worker.onmessage = (e) => {
+					if (e.data.ok) resolve();
+					else reject(/* @__PURE__ */ new Error(`Worker error: ${e.data.error}`));
+				};
+				worker.postMessage({
+					module,
+					ptrs,
+					begin,
+					end
+				});
+			}));
+		}
+		await Promise.all(promises);
+	}
+};
+/** Try to create a worker pool. Returns null if workers are unavailable. */
+function createWorkerPool(memory) {
+	if (!hasSharedArrayBuffer()) return null;
+	try {
+		const numWorkers = Math.max(1, typeof navigator !== "undefined" && navigator.hardwareConcurrency || 4);
+		return new WasmWorkerPool(memory, numWorkers);
+	} catch {
+		return null;
+	}
+}
+
 //#endregion
 //#region src/backend/wasm/wasmblr.ts
 /**
@@ -3407,7 +3694,7 @@ var CodeGenerator = class {
 				concat(importSectionBytes, encodeString(this.memory.aString));
 				concat(importSectionBytes, encodeString(this.memory.bString));
 				importSectionBytes.push(2);
-				if (this.memory.min && this.memory.max) {
+				if (this.memory.max) {
 					if (this.memory.isShared) importSectionBytes.push(3);
 					else importSectionBytes.push(1);
 					concat(importSectionBytes, encodeUnsigned(this.memory.min));
@@ -3814,6 +4101,8 @@ var I32x4 = class extends V128 {
 	min_u = VECTOR_OP("min_u", 183, ["v128", "v128"], "v128");
 	max_s = VECTOR_OP("max_s", 184, ["v128", "v128"], "v128");
 	max_u = VECTOR_OP("max_u", 185, ["v128", "v128"], "v128");
+	trunc_sat_f32x4_s = VECTOR_OP("trunc_sat_f32x4_s", 248, ["v128"], "v128");
+	trunc_sat_f32x4_u = VECTOR_OP("trunc_sat_f32x4_u", 249, ["v128"], "v128");
 };
 var F32x4 = class extends V128 {
 	splat = VECTOR_OP("splat", 19, ["f32"], "v128");
@@ -3840,10 +4129,333 @@ var F32x4 = class extends V128 {
 	max = VECTOR_OP("max", 233, ["v128", "v128"], "v128");
 	pmin = VECTOR_OP("pmin", 234, ["v128", "v128"], "v128");
 	pmax = VECTOR_OP("pmax", 235, ["v128", "v128"], "v128");
+	convert_i32x4_s = VECTOR_OP("convert_i32x4_s", 250, ["v128"], "v128");
+	convert_i32x4_u = VECTOR_OP("convert_i32x4_u", 251, ["v128"], "v128");
 };
 
 //#endregion
 //#region src/backend/wasm.ts
+/**
+* SIMD version of translateExp: emits v128 (f32x4 or i32x4) instructions instead of scalar.
+* gidx always steps by 4. strideMap classifies each GlobalIndex as broadcast/contiguous/gather.
+*/
+function translateExpSimd(cg, funcs, exp, ctx, strideMap) {
+	const references = /* @__PURE__ */ new Map();
+	const seen = /* @__PURE__ */ new Set();
+	const countReferences = (exp$1) => {
+		references.set(exp$1, (references.get(exp$1) ?? 0) + 1);
+		if (!seen.has(exp$1)) {
+			seen.add(exp$1);
+			for (const src of exp$1.src) countReferences(src);
+		}
+	};
+	const expContext = /* @__PURE__ */ new Map();
+	const gen = (exp$1) => {
+		if (expContext.has(exp$1)) return cg.local.get(expContext.get(exp$1));
+		const { op, src, arg, dtype } = exp$1;
+		const isInt = dtype === DType.Int32 || dtype === DType.Uint32 || dtype === DType.Bool;
+		const isSigned = dtype === DType.Int32;
+		if (op === AluOp.Add) {
+			gen(src[0]);
+			gen(src[1]);
+			if (isInt) cg.i32x4.add();
+			else cg.f32x4.add();
+		} else if (op === AluOp.Sub) {
+			gen(src[0]);
+			gen(src[1]);
+			if (isInt) cg.i32x4.sub();
+			else cg.f32x4.sub();
+		} else if (op === AluOp.Mul) {
+			gen(src[0]);
+			gen(src[1]);
+			if (isInt) cg.i32x4.mul();
+			else cg.f32x4.mul();
+		} else if (op === AluOp.Min) {
+			gen(src[0]);
+			gen(src[1]);
+			if (isInt) if (isSigned) cg.i32x4.min_s();
+			else cg.i32x4.min_u();
+			else cg.f32x4.min();
+		} else if (op === AluOp.Max) {
+			gen(src[0]);
+			gen(src[1]);
+			if (isInt) if (isSigned) cg.i32x4.max_s();
+			else cg.i32x4.max_u();
+			else cg.f32x4.max();
+		} else if (op === AluOp.Sqrt) {
+			gen(src[0]);
+			cg.f32x4.sqrt();
+		} else if (op === AluOp.Floor) {
+			gen(src[0]);
+			cg.f32x4.floor();
+		} else if (op === AluOp.Ceil) {
+			gen(src[0]);
+			cg.f32x4.ceil();
+		} else if (op === AluOp.Const) if (isInt) {
+			cg.i32.const(arg);
+			cg.i32x4.splat();
+		} else {
+			cg.f32.const(arg);
+			cg.f32x4.splat();
+		}
+		else if (op === AluOp.Cast) {
+			gen(src[0]);
+			const dtype0 = src[0].dtype;
+			const src0IsInt = dtype0 === DType.Int32 || dtype0 === DType.Uint32 || dtype0 === DType.Bool;
+			if (isInt && !src0IsInt) if (isSigned) cg.i32x4.trunc_sat_f32x4_s();
+			else cg.i32x4.trunc_sat_f32x4_u();
+			else if (!isInt && src0IsInt) if (dtype0 === DType.Int32 || dtype0 === DType.Bool) cg.f32x4.convert_i32x4_s();
+			else cg.f32x4.convert_i32x4_u();
+		} else if (op === AluOp.Cmplt) {
+			gen(src[0]);
+			gen(src[1]);
+			const srcDtype = src[0].dtype;
+			if (srcDtype === DType.Float32) cg.f32x4.lt();
+			else if (srcDtype === DType.Int32) cg.i32x4.lt_s();
+			else if (srcDtype === DType.Uint32) cg.i32x4.lt_u();
+			else throw new UnsupportedOpError(op, dtype, "wasm");
+			cg.i32.const(1);
+			cg.i32x4.splat();
+			cg.v128.and();
+		} else if (op === AluOp.Cmpne) {
+			gen(src[0]);
+			gen(src[1]);
+			const srcDtype = src[0].dtype;
+			if (srcDtype === DType.Float32) cg.f32x4.ne();
+			else cg.i32x4.ne();
+			cg.i32.const(1);
+			cg.i32x4.splat();
+			cg.v128.and();
+		} else if (op === AluOp.Where) {
+			gen(src[1]);
+			gen(src[2]);
+			gen(src[0]);
+			cg.i32.const(0);
+			cg.i32x4.splat();
+			cg.i32x4.ne();
+			cg.v128.bitselect();
+		} else if (op === AluOp.Variable || op === AluOp.Special) throw new Error(`translateExpSimd: unexpected ${op}(${arg})`);
+		else if (op === AluOp.GlobalIndex) {
+			const [gid, len] = arg;
+			const indexSubtree = src[0];
+			const stride = strideMap.get(exp$1) ?? GATHER;
+			if (stride.kind === "contiguous") {
+				translateExp(cg, funcs, indexSubtree, ctx);
+				{
+					const maxIdx = Math.max(len - SIMD_LANES, 0);
+					const wideIdx = cg.local.declare(cg.i32);
+					cg.local.set(wideIdx);
+					cg.local.get(wideIdx);
+					cg.i32.const(maxIdx);
+					cg.local.get(wideIdx);
+					cg.i32.const(maxIdx);
+					cg.i32.lt_u();
+					cg.select();
+				}
+				cg.i32.const(byteWidth(dtype));
+				cg.i32.mul();
+				cg.local.get(gid);
+				cg.i32.add();
+				if (isInt) cg.i32x4.load(4);
+				else cg.f32x4.load(4);
+			} else if (stride.kind === "broadcast") {
+				translateExp(cg, funcs, indexSubtree, ctx);
+				const local = cg.local.declare(cg.i32);
+				cg.local.tee(local);
+				cg.i32.const(0);
+				cg.local.get(local), cg.i32.const(len), cg.i32.lt_u();
+				cg.select();
+				cg.i32.const(byteWidth(dtype));
+				cg.i32.mul();
+				cg.local.get(gid);
+				cg.i32.add();
+				if (isInt) {
+					cg.i32.load(2);
+					cg.i32x4.splat();
+				} else {
+					cg.f32.load(2);
+					cg.f32x4.splat();
+				}
+			} else {
+				const steppingLocal = ctx["gidx"];
+				const origValue = cg.local.declare(cg.i32);
+				cg.local.get(steppingLocal);
+				cg.local.set(origValue);
+				if (isInt) {
+					cg.i32.const(0);
+					cg.i32x4.splat();
+				} else {
+					cg.f32.const(0);
+					cg.f32x4.splat();
+				}
+				const vec = cg.local.declare(isInt ? cg.i32x4 : cg.f32x4);
+				cg.local.set(vec);
+				const idx = cg.local.declare(cg.i32);
+				const scalarVal = cg.local.declare(isInt ? cg.i32 : cg.f32);
+				for (let lane = 0; lane < SIMD_LANES; lane++) {
+					cg.local.get(origValue);
+					if (lane > 0) {
+						cg.i32.const(lane);
+						cg.i32.add();
+					}
+					cg.local.set(steppingLocal);
+					translateExp(cg, funcs, indexSubtree, ctx);
+					cg.local.tee(idx);
+					cg.i32.const(0);
+					cg.local.get(idx), cg.i32.const(len), cg.i32.lt_u();
+					cg.select();
+					cg.i32.const(byteWidth(dtype));
+					cg.i32.mul();
+					cg.local.get(gid);
+					cg.i32.add();
+					if (isInt) cg.i32.load(2);
+					else cg.f32.load(2);
+					cg.local.set(scalarVal);
+					cg.local.get(vec);
+					cg.local.get(scalarVal);
+					if (isInt) cg.i32x4.replace_lane(lane);
+					else cg.f32x4.replace_lane(lane);
+					cg.local.set(vec);
+				}
+				cg.local.get(origValue);
+				cg.local.set(steppingLocal);
+				cg.local.get(vec);
+			}
+		} else throw new Error(`translateExpSimd: unsupported op ${op}`);
+		if ((references.get(exp$1) ?? 0) > 1) {
+			const local = cg.local.declare(isInt ? cg.i32x4 : cg.f32x4);
+			cg.local.tee(local);
+			expContext.set(exp$1, local);
+		}
+	};
+	countReferences(exp);
+	gen(exp);
+}
+/** Number of SIMD lanes (f32x4 / i32x4 = 4 lanes). */
+const SIMD_LANES = 4;
+function referencesGidx(exp) {
+	if (exp.op === AluOp.Special && exp.arg[0] === "gidx") return true;
+	return exp.src.some(referencesGidx);
+}
+/** When tileSize > N but doesn't divide evenly, the last group before the
+*  inner reset is shorter than N — a SIMD group could straddle it. */
+function hasFragmentRisk(tileSize, N) {
+	return isFinite(tileSize) && tileSize > N && tileSize % N !== 0;
+}
+const GATHER = { kind: "gather" };
+/**
+* Classify how a GlobalIndex's index expression behaves as gidx increments.
+*/
+function analyzeStride(exp) {
+	if (!referencesGidx(exp)) return {
+		kind: "broadcast",
+		tileSize: Infinity
+	};
+	if (exp.op === AluOp.Special && exp.arg[0] === "gidx") return {
+		kind: "contiguous",
+		tileSize: Infinity
+	};
+	if (exp.op === AluOp.Idiv && exp.src[1].op === AluOp.Const) {
+		const N = exp.src[1].arg;
+		const inner = analyzeStride(exp.src[0]);
+		if (inner.kind === "broadcast") return inner;
+		if (inner.kind !== "contiguous") return GATHER;
+		if (hasFragmentRisk(inner.tileSize, N)) return GATHER;
+		return {
+			kind: "broadcast",
+			tileSize: Math.min(inner.tileSize, N)
+		};
+	}
+	if (exp.op === AluOp.Mod && exp.src[1].op === AluOp.Const) {
+		const N = exp.src[1].arg;
+		const inner = analyzeStride(exp.src[0]);
+		if (inner.kind === "broadcast") return inner;
+		if (inner.kind !== "contiguous") return GATHER;
+		if (hasFragmentRisk(inner.tileSize, N)) return GATHER;
+		return {
+			kind: "contiguous",
+			tileSize: Math.min(inner.tileSize, N)
+		};
+	}
+	if (exp.op === AluOp.Mul) {
+		for (let i = 0; i < 2; i++) if (exp.src[i].op === AluOp.Const) {
+			const inner = analyzeStride(exp.src[1 - i]);
+			if (inner.kind === "broadcast") return inner;
+			return GATHER;
+		}
+	}
+	if (exp.op === AluOp.Add) {
+		const lhsHasGidx = referencesGidx(exp.src[0]);
+		const rhsHasGidx = referencesGidx(exp.src[1]);
+		if (lhsHasGidx && !rhsHasGidx) return analyzeStride(exp.src[0]);
+		if (!lhsHasGidx && rhsHasGidx) return analyzeStride(exp.src[1]);
+	}
+	return GATHER;
+}
+/** Ops that have direct SIMD (f32x4) instruction variants. */
+const simdF32Ops = new Set([
+	AluOp.Add,
+	AluOp.Sub,
+	AluOp.Mul,
+	AluOp.Floor,
+	AluOp.Ceil,
+	AluOp.Min,
+	AluOp.Max,
+	AluOp.Sqrt,
+	AluOp.Cast,
+	AluOp.Where,
+	AluOp.Const,
+	AluOp.GlobalIndex
+]);
+/** Ops that have direct SIMD (i32x4) instruction variants. */
+const simdI32Ops = new Set([
+	AluOp.Add,
+	AluOp.Sub,
+	AluOp.Mul,
+	AluOp.Min,
+	AluOp.Max,
+	AluOp.Cast,
+	AluOp.Where,
+	AluOp.Const,
+	AluOp.GlobalIndex
+]);
+/** Ops that produce Bool (i32x4 bitmask) in SIMD. */
+const simdBoolOps = new Set([
+	AluOp.Cmplt,
+	AluOp.Cmpne,
+	AluOp.Const,
+	AluOp.GlobalIndex
+]);
+/**
+* Check if a kernel is eligible for SIMD codegen.
+*
+* A kernel qualifies when:
+* - size >= 4 (need at least 4 elements for a SIMD group)
+* - For reductions: the reduction op has a SIMD variant for its dtype
+* - All nodes have a supported dtype (f32, i32, u32, bool) with SIMD variants
+*/
+function isSimdEligible(tunedExp, kernel) {
+	if (kernel.size < SIMD_LANES) return false;
+	if (kernel.reduction) {
+		if (!simdSupportedOpsForDtype(kernel.reduction.dtype)?.has(kernel.reduction.op)) return false;
+	}
+	const check = (exp, visited) => {
+		if (visited.has(exp)) return true;
+		visited.add(exp);
+		const supportedOps = simdSupportedOpsForDtype(exp.dtype);
+		if (!supportedOps || !supportedOps.has(exp.op)) return false;
+		if (exp.op === AluOp.GlobalIndex) return true;
+		for (const child of exp.src) if (!check(child, visited)) return false;
+		return true;
+	};
+	return check(tunedExp, /* @__PURE__ */ new Set());
+}
+function simdSupportedOpsForDtype(dtype) {
+	if (dtype === DType.Float32) return simdF32Ops;
+	if (dtype === DType.Int32 || dtype === DType.Uint32) return simdI32Ops;
+	if (dtype === DType.Bool) return simdBoolOps;
+	return null;
+}
 const moduleCache = /* @__PURE__ */ new Map();
 /** Backend that compiles into WebAssembly bytecode for immediate execution. */
 var WasmBackend = class {
@@ -3853,11 +4465,18 @@ var WasmBackend = class {
 	#nextSlot;
 	#allocator;
 	#buffers;
+	#workerPool;
+	#pendingWork = /* @__PURE__ */ new Map();
 	constructor() {
-		this.#memory = new WebAssembly.Memory({ initial: 0 });
+		this.#memory = hasSharedArrayBuffer() ? new WebAssembly.Memory({
+			initial: 0,
+			maximum: 65536,
+			shared: true
+		}) : new WebAssembly.Memory({ initial: 0 });
 		this.#allocator = new WasmAllocator(this.#memory);
 		this.#nextSlot = 1;
 		this.#buffers = /* @__PURE__ */ new Map();
+		this.#workerPool = createWorkerPool(this.#memory);
 	}
 	malloc(size, initialData) {
 		const ptr = this.#allocator.malloc(size);
@@ -3888,37 +4507,70 @@ var WasmBackend = class {
 		}
 	}
 	async read(slot, start, count) {
-		return this.readSync(slot, start, count);
+		const epoch = this.#pendingWork.get(slot);
+		if (epoch) await this.#workerPool.waitForEpoch(epoch);
+		return this.#readData(slot, start, count);
 	}
 	readSync(slot, start, count) {
+		const epoch = this.#pendingWork.get(slot);
+		if (epoch && this.#workerPool.epoch < epoch) throw new Error("cannot read synchronously from a slot with async work");
+		return this.#readData(slot, start, count);
+	}
+	#readData(slot, start, count) {
 		const buffer = this.#getBuffer(slot);
 		if (start === void 0) start = 0;
 		if (count === void 0) count = buffer.byteLength - start;
-		return buffer.slice(start, start + count);
+		if (buffer.buffer instanceof SharedArrayBuffer) return new Uint8Array(buffer.slice(start, start + count));
+		else return buffer.slice(start, start + count);
 	}
 	async prepareKernel(kernel) {
-		return this.prepareKernelSync(kernel);
+		const kernelHash = FpHash.hash(kernel);
+		const module = await runWithCacheAsync(moduleCache, kernelHash.toString(), () => WebAssembly.compile(codegenWasm(kernel)));
+		return new Executable(kernel, {
+			module,
+			parallel: this.#workerPool !== null
+		});
 	}
 	prepareKernelSync(kernel) {
 		const kernelHash = FpHash.hash(kernel);
-		const module = runWithCache(moduleCache, kernelHash.toString(), () => {
-			const bytes = codegenWasm(kernel);
-			return new WebAssembly.Module(bytes);
+		const module = runWithCache(moduleCache, kernelHash.toString(), () => new WebAssembly.Module(codegenWasm(kernel)));
+		return new Executable(kernel, {
+			module,
+			parallel: false
 		});
-		return new Executable(kernel, { module });
 	}
 	async prepareRoutine(routine) {
 		return this.prepareRoutineSync(routine);
 	}
 	prepareRoutineSync(routine) {
-		return new Executable(routine, void 0);
+		return new Executable(routine, {
+			module: void 0,
+			parallel: false
+		});
 	}
 	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);
-		func(...ptrs);
+		const tracing = isTracing();
+		const start = tracing ? performance.now() : 0;
+		if (exe.source instanceof Routine) runCpuRoutine(exe.source, inputs.map((slot) => this.#getBuffer(slot)), outputs.map((slot) => this.#getBuffer(slot)));
+		else {
+			const ptrs = [...inputs, ...outputs].map((slot) => this.#buffers.get(slot).ptr);
+			if (exe.data.parallel && this.#workerPool) {
+				const epoch = this.#workerPool.dispatch(exe.data.module, ptrs, exe.source.size);
+				for (const slot of outputs) this.#pendingWork.set(slot, epoch);
+			} else {
+				if (inputs.some((slot) => {
+					const epoch = this.#pendingWork.get(slot);
+					return epoch && this.#workerPool.epoch < epoch;
+				})) throw new Error("cannot dispatch synchronously with pending async work");
+				const instance = new WebAssembly.Instance(exe.data.module, { env: { memory: this.#memory } });
+				const func = instance.exports.kernel;
+				func(...ptrs, 0, exe.source.size);
+			}
+		}
+		if (tracing) {
+			const info = traceSourceInfo(exe.source);
+			emitTrace("wasm", info, start, performance.now());
+		}
 	}
 	#getBuffer(slot) {
 		const buffer = this.#buffers.get(slot);
@@ -3926,12 +4578,36 @@ var WasmBackend = class {
 		return new Uint8Array(this.#memory.buffer, buffer.ptr, buffer.size);
 	}
 };
+/** Emit a runtime guard: enter the if-block only when [begin, end) is SIMD-aligned. */
+function emitAlignmentGuard(cg, paramBegin, paramEnd) {
+	const mask = SIMD_LANES - 1;
+	cg.local.get(paramEnd);
+	cg.local.get(paramBegin);
+	cg.i32.sub();
+	cg.i32.const(mask);
+	cg.i32.and();
+	cg.i32.eqz();
+	cg.local.get(paramBegin);
+	cg.i32.const(mask);
+	cg.i32.and();
+	cg.i32.eqz();
+	cg.i32.and();
+	cg.if(cg.void);
+}
 function codegenWasm(kernel) {
 	const tune = tuneNullopt(kernel);
 	const re = kernel.reduction;
 	if (DEBUG >= 3) console.info(`kernel.exp: ${kernel.exp}\ntune.exp: ${tune.exp}`);
+	const useSimd = isSimdEligible(tune.exp, kernel);
+	const bufferStrides = /* @__PURE__ */ new Map();
+	if (useSimd) tune.exp.collect((e) => e.op === AluOp.GlobalIndex).forEach((gi) => {
+		const result = analyzeStride(gi.src[0]);
+		if (result.kind !== "gather" && (result.tileSize < SIMD_LANES || isFinite(result.tileSize) && result.tileSize % SIMD_LANES !== 0)) bufferStrides.set(gi, GATHER);
+		else bufferStrides.set(gi, result);
+	});
 	const cg = new CodeGenerator();
 	cg.memory.import("env", "memory");
+	if (hasSharedArrayBuffer()) cg.memory.pages(0, 65536).shared(true);
 	const distinctOps = mapSetUnion(tune.exp.distinctOps(), tune.epilogue?.distinctOps());
 	const funcs = {};
 	if (distinctOps.has(AluOp.Sin)) funcs.sin = wasm_sin(cg);
@@ -3943,12 +4619,127 @@ function codegenWasm(kernel) {
 	if (distinctOps.has(AluOp.Erf)) funcs.erf = wasm_erf(cg, funcs.exp);
 	if (distinctOps.has(AluOp.Erfc)) funcs.erfc = wasm_erfc(cg, funcs.exp);
 	if (distinctOps.has(AluOp.Threefry2x32)) funcs.threefry2x32 = wasm_threefry2x32(cg);
-	const kernelFunc = cg.function(rep(kernel.nargs + 1, cg.i32), [], () => {
+	const paramBegin = kernel.nargs + 1;
+	const paramEnd = kernel.nargs + 2;
+	const kernelFunc = cg.function(rep(kernel.nargs + 3, cg.i32), [], () => {
 		const gidx = cg.local.declare(cg.i32);
+		cg.local.get(paramBegin);
+		cg.local.set(gidx);
+		if (useSimd) {
+			emitAlignmentGuard(cg, paramBegin, paramEnd);
+			cg.loop(cg.void);
+			if (!re) {
+				cg.block(cg.void);
+				cg.local.get(gidx);
+				cg.local.get(paramEnd);
+				cg.i32.ge_u();
+				cg.br_if(0);
+				cg.local.get(kernel.nargs);
+				cg.local.get(gidx);
+				cg.i32.const(byteWidth(kernel.dtype));
+				cg.i32.mul();
+				cg.i32.add();
+				translateExpSimd(cg, funcs, tune.exp, { gidx }, bufferStrides);
+				cg.v128.store(4);
+				cg.local.get(gidx);
+				cg.i32.const(SIMD_LANES);
+				cg.i32.add();
+				cg.local.set(gidx);
+				cg.br(1);
+				cg.end();
+			} else {
+				const reIsInt = kernel.exp.dtype === DType.Int32 || kernel.exp.dtype === DType.Uint32;
+				cg.block(cg.void);
+				cg.local.get(gidx);
+				cg.local.get(paramEnd);
+				cg.i32.ge_u();
+				cg.br_if(0);
+				const vecAcc = cg.local.declare(reIsInt ? cg.i32x4 : cg.f32x4);
+				if (reIsInt) {
+					cg.i32.const(re.identity);
+					cg.i32x4.splat();
+				} else {
+					cg.f32.const(re.identity);
+					cg.f32x4.splat();
+				}
+				cg.local.set(vecAcc);
+				const ridx = cg.local.declare(cg.i32);
+				cg.i32.const(0);
+				cg.local.set(ridx);
+				cg.loop(cg.void);
+				cg.block(cg.void);
+				cg.local.get(ridx);
+				cg.i32.const(re.size);
+				cg.i32.ge_u();
+				cg.br_if(0);
+				translateExpSimd(cg, funcs, tune.exp, {
+					gidx,
+					ridx
+				}, bufferStrides);
+				cg.local.get(vecAcc);
+				if (reIsInt) if (re.op === AluOp.Add) cg.i32x4.add();
+				else if (re.op === AluOp.Mul) cg.i32x4.mul();
+				else if (re.op === AluOp.Min) if (re.dtype === DType.Int32) cg.i32x4.min_s();
+				else cg.i32x4.min_u();
+				else if (re.op === AluOp.Max) if (re.dtype === DType.Int32) cg.i32x4.max_s();
+				else cg.i32x4.max_u();
+				else throw new Error(`invalid SIMD reduction op: ${re.op}`);
+				else if (re.op === AluOp.Add) cg.f32x4.add();
+				else if (re.op === AluOp.Mul) cg.f32x4.mul();
+				else if (re.op === AluOp.Min) cg.f32x4.min();
+				else if (re.op === AluOp.Max) cg.f32x4.max();
+				else throw new Error(`invalid SIMD reduction op: ${re.op}`);
+				cg.local.set(vecAcc);
+				cg.local.get(ridx);
+				cg.i32.const(1);
+				cg.i32.add();
+				cg.local.set(ridx);
+				cg.br(1);
+				cg.end();
+				cg.end();
+				for (let lane = 0; lane < SIMD_LANES; lane++) {
+					cg.local.get(kernel.nargs);
+					cg.local.get(gidx);
+					if (lane > 0) {
+						cg.i32.const(lane);
+						cg.i32.add();
+					}
+					cg.i32.const(byteWidth(kernel.dtype));
+					cg.i32.mul();
+					cg.i32.add();
+					const acc = cg.local.declare(reIsInt ? cg.i32 : cg.f32);
+					cg.local.get(vecAcc);
+					if (reIsInt) cg.i32x4.extract_lane(lane);
+					else cg.f32x4.extract_lane(lane);
+					cg.local.set(acc);
+					const laneGidx = cg.local.declare(cg.i32);
+					cg.local.get(gidx);
+					if (lane > 0) {
+						cg.i32.const(lane);
+						cg.i32.add();
+					}
+					cg.local.set(laneGidx);
+					translateExp(cg, funcs, tune.epilogue, {
+						acc,
+						gidx: laneGidx
+					});
+					dty(cg, null, kernel.dtype).store(Math.log2(byteWidth(kernel.dtype)));
+				}
+				cg.local.get(gidx);
+				cg.i32.const(SIMD_LANES);
+				cg.i32.add();
+				cg.local.set(gidx);
+				cg.br(1);
+				cg.end();
+			}
+			cg.end();
+			cg.return();
+			cg.end();
+		}
 		cg.loop(cg.void);
 		cg.block(cg.void);
 		cg.local.get(gidx);
-		cg.i32.const(kernel.size);
+		cg.local.get(paramEnd);
 		cg.i32.ge_u();
 		cg.br_if(0);
 		cg.local.get(kernel.nargs);
@@ -4087,6 +4878,11 @@ function translateExp(cg, funcs, exp, ctx) {
 				else cg.i32.gt_u();
 				cg.select();
 			} else throw new UnsupportedOpError(op, dtype, "wasm");
+			else if (op === AluOp.BitCombine) if (arg === "and") cg.i32.and();
+			else if (arg === "or") cg.i32.or();
+			else cg.i32.xor();
+			else if (op === AluOp.BitShift) if (arg === "shl") cg.i32.shl();
+			else cg.i32.shr_u();
 			else if (op === AluOp.Cmplt) {
 				const srcDtype = src[0].dtype;
 				if (isFloatDtype(srcDtype)) dtyF(cg, op, srcDtype).lt();
@@ -4263,7 +5059,7 @@ 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-AN0cG_nB.js");
+		const { WebGPUBackend } = await import("./webgpu-Dg8FpYrH.js");
 		const importantLimits = [
 			"maxBufferSize",
 			"maxComputeInvocationsPerWorkgroup",
@@ -4301,7 +5097,7 @@ async function createBackend(device) {
 		});
 		if (!gl) return null;
 		if (!gl.getExtension("EXT_color_buffer_float")) return null;
-		const { WebGLBackend } = await import("./webgl-DnGrclTz.js");
+		const { WebGLBackend } = await import("./webgl-D8-14NzA.js");
 		return new WebGLBackend(gl);
 	} else throw new Error(`Backend not found: ${device}`);
 }
@@ -4335,6 +5131,15 @@ var UnsupportedRoutineError = class extends Error {
 		super(`routine '${name}' is not supported in ${device} backend`);
 	}
 };
+/**
+* If the WebGPU backend has been initialized, return the `GPUDevice` that this
+* backend runs on. This is useful for sharing buffers.
+*/
+function getWebGPUDevice() {
+	const backend = initializedBackends.get("webgpu");
+	if (!backend) throw new Error("WebGPU backend not initialized, call init('webgpu') first");
+	return backend.device;
+}
 
 //#endregion
-export { AluExp, AluGroup, AluOp, AluVar, DEBUG, DType, Executable, FpHash, Kernel, PPrint, Reduction, Routine, Routines, ShapeTracker, SlotError, UnsupportedOpError, UnsupportedRoutineError, accessorAluExp, accessorGlobal, assertNonNull, byteWidth, checkAxis, checkInts, 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, tuneNullopt, tuneWebgpu, unravelAlu, unzip2, zip, zipn };
+export { AluExp, AluGroup, AluOp, AluVar, DEBUG, DType, Executable, FpHash, Kernel, PPrint, Reduction, Routine, Routines, ShapeTracker, SlotError, UnsupportedOpError, UnsupportedRoutineError, accessorAluExp, accessorGlobal, assertNonNull, byteWidth, checkAxis, checkInts, deepEqual, defaultDevice, devices, dtypedArray, dtypedJsArray, emitTrace, findPow2, generalBroadcast, getBackend, getWebGPUDevice, init, invertPermutation, isFloatDtype, isNumberPair, isPermutation, isTracing, mapSetUnion, normalizeAxis, onFlushTrace, partitionList, prod, promoteTypes, range, recursiveFlatten, rep, runWithCache, setDebug, startTrace, stopTrace, strip1, toposort, traceSourceInfo, tuneNullopt, tuneWebgpu, unravelAlu, unzip2, zip, zipn };